0% found this document useful (0 votes)
403 views483 pages

Mathematical Simulation in Glass Technology PDF

Download as pdf or txt
Download as pdf or txt
Download as pdf or txt
You are on page 1/ 483

Mathematical Simulation

in Glass Technology

Springer-Verlag Berlin Heidelberg GmbH


Schott Series on Glass and Glass Ceramics
Science, Technology, and Applications

Low Thermal Expansion Glass Ceramics


ISBN 3-540-58598-2

Fibre Optics and Glass Integrated Optics


ISBN 3-540-58595-8

The Properties of Optical Glass


ISBN 3-540-58357-2

Thin Films on Glass


ISBN 3-540-58597-4

Electrochemistry of Glasses and Glass Melts,


Including Glass Electrodes
ISBN 3-540-58608-3

Surface Analysis of Glasses and Glass


Ceramics, and Coatings
ISBN 3-540-58609-1

Analysis of the Composition and Structure


of Glass and Glass Ceramics
ISBN 3-540-58610-52

Mathematical Simulation in Glass Technology


ISBN 3-540-43204-3
Horst Loch
Dieter Krause
Editors

Mathematical Simulation
in Glass Technology

With 302 Figures and 17 Tables


and 27 Videos on CD

Springer
Editors
Dr. Horst Loch
Prof. Dr. Dieter Krause
Schott Glas
Hattenbergstr. 10
D-55122 Mainz, Germany

Library of Congress Catalog-in-Publication Data

Mathematical simulation in glass technology ! Horst Loch, Dieter Krause, editors.


p. cm. -- (Schott series on glass ceramics)
Included biblographical references and index.
Additional material to this book can be downloaded from http://extras.springer.com.

ISBN 978-3-642-62782-8 ISBN 978-3-642-55966-2 (eBook)


DOI 10.1007/978-3-642-55966-2
1. Glass manufacture--Mathematical models. 2. Glass manufacture--Computer simulation
1. Loch, Horst. II. Krause, Dieter, 1933- III. Series
TP858.2.M37 M37 2002 666'.1'015118--dc21 2002021146

ISBN 978-3-642-62782-8

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specificallythe rights of translation, reprinting, reuse of illustrations, recitation, broadcasting,
reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9,
1965, in its current vers ion, and permis sion for use must always be obtained from Springer-Verlag.
Violations are liable for prosecution under the German Copyright Law.

http://www.springer.de
© Springer-Verlag Berlin Heidelberg 2002
Originally published by Springer-Verlag Berlin Heidelberg in 2002
The use of designations, trademarks, etc. in this publication does not imply, even in the absence of a
specific statement, that such names are exempt from the relevant protective laws and regulations and
therefore free for general use. The following trademarks used in this book are registered trademarks of
Schott Glas or of Schott Group companies, respectively: Ceran, Duran, Suprax. Other trademarks
mentioned in this book (e.g., Pyrex, ProfilExpert) are registered trademarks of other companies.
Please note: Before using the programs in this book, please consult the technical manuals provided by
the manufacturer of the computer - and of any additional plug-in boards - to be used. The authors and
the publisher accept no legal responsibility for any damage caused by improper use of the instructions
and programs contained here in. Although these programs have been tested with extreme care, we can
offer no formal guarantee that they will function correctly. The programs on the enclosed CD- ROM are
under copyright protection and may not be reproduced without written permission by Springer-Verlag.
One copy of the programs may be made as a back-up, but all further copies violate copyright law.
Cover production: design & production, Heidelberg
Typesetting: Computer to plate from editors data
Printed on acid-free paper SPIN 10866458 56/3142YL 543210
Foreword

This book, entitled Mathematical Simulation in Glass Technology, is one of


a series reporting on research and development activities on products and
processes conducted by the Schott Group.
The scientifically founded development of new products and technical pro-
cesses has traditionally been of vital importance to Schott and has always
been performed on a scale determined by the prospects for application of our
special glasses. Since the reconstruction of the Schott Glaswerke in Mainz,
the scale has increased enormously. The range of expert knowledge required
could never have been supplied by Schott alone. It is also a tradition in our
company to cultivate collaboration with customers, universities, and research
institutes. Publications in numerous technical journals, which since 1969 we
have edited to a regular schedule as Forschungsberichte - "research reports"
- describe the results of these cooperations. They contain up-to-date infor-
mation on various topics for the expert but are not suited as survey material
for those whose standpoint is more remote.
This is the point where we would like to place our series, to stimulate the
exchange of thoughts, so that we can consider from different points of view
the possibilities offered by those incredibly versatile materials, glass and glass
ceramics. We would like to share the knowledge won through our research
and development at Schott in cooperation with the users of our materials
with scientists and engineers, interested customers and friends, and with the
employees of our firm.
Though the results documented in the volumes of the Schott Series are, of
course, oriented to the tasks and targets of a company, we believe that readers
can nevertheless - or just for that very reason - find demanding challenges for
the development of process engineering, the characterization of measurement
practice, and for applied research. Besides realizability, the profitability of
solutions to customers' problems always plays a decisive role.
The first comprehensive presentation of research findings after the recon-
struction of the factory in Mainz was edited by Prof. Dr. Dr. h.c. Erich Schott
in 1959. It was entitled Beitriige zur angewandten Glasforschung - "Contri-
butions to Applied Glass Research" (Wissenschaftliche Verlagsgesellschaft
m.b.H., Stuttgart 1959). Since then, there has been an extraordinary world-
wide increase in the application of glass and glass ceramic materials. Glass
II Foreword

fibers and components manufactured from them for use in lighting and traffic
engineering or in telecommunications, high-purity and highly homogeneous
glasses for masks and projection lenses in electronics, or glass ceramics with
zero expansion in astronomy and in household-appliance technology are only
some examples. In many of these fields, Schott has made essential contribu-
tions.
Due to the breadth and complexity of the Schott activities, it takes several
volumes to describe the company's research and development results. Oth-
erwise it would be impossible to do full justice to the fundamental research
work and technological innovation that is indispensable for product develop-
ment, and to give an appropriate description of the methods of measurement
and analysis needed for the development and manufacture of new products.
Five volumes, entitled The Properties of Optical Glass, Low Thermal Ex-
pansion Glass Ceramics, Thin Films on Glass, Analysis of the Composition
and Structure of Glass and Glass Ceramics, and Electrochemistry of Glasses
and Glass Melts, Including Glass Electrodes, have already been published.
Another two volumes, entitled Surface Analysis of Glasses and Glass Ce-
ramics, and Coatings and Fibre Optics and Glass Integrated Optics, are in
preparation and will be published in the next few years. Glasses for various
applications in industry and science and their properties are being considered,
and melting and processing technologies described.
With the presentation - in part detailed - of the work required for the
development of successful products, Schott employees are giving all their in-
terested colleagues working in the field of science and technology an insight
into the special experiences and successes in material science, material de-
velopment, and the application of materials at Schott. Contributions from
scientists and engineers who work at universities and other research insti-
tutes, and who played an essential role in Schott developments, complete the
survey of what has been achieved and prove the usefulness of the collabora-
tions mentioned above.
In all the volumes of the series, the fundamental issues from chemistry,
physics, and engineering are dealt with, or at least studies are cited that en-
able or assist the reader to work his or her own way into the topics treated.
Thus, the series may serve to fill gaps between the basic knowledge imparted
by textbooks on material science and the product descriptions published by
Schott. We see this as the best way to enable all our potential business part-
ners who are not already familiar with glass and glass ceramics to compare
these materials with alternatives on a thoroughly scientific basis. We hope
that this will lead to intense technical discussions and collaborations on new
fields of applications of our materials and products, to our mutual advantage.
Every volume of the Schott Series will begin with a chapter providing
a general idea of the current problems, results, and trends relating to the
subjects treated. These introductory chapters and the reviews of the basic
principles are intended for readers dealing for the first time with the special
Foreword III

properties of glass and glass ceramic materials and their surface treatment in
engineering, science, and education.
Many of our German clients are accustomed to reading scientific and
technical publications in English, and most of our foreign customers are better
conversant with English than with German. We therefore decided to publish
the Schott Series in English.
The publication of the Schott Series has been substantially supported by
Springer-Verlag. We would like to express our special thanks to Dr. H.K.V.
Lotsch and Dr. H.J. K6lsch for advice and assistance in this project.
The investment of resources by Schott and its employees to produce the
Schott Series is, as already stated, necessary for the interdisciplinary dia-
log and collaboration that are traditional at Schott. A model we still find
exemplary today of a fruitful dialog between fundamental research, glass
research, and glass manufacture was achieved in the collaboration between
Ernst Abbe, Otto Schott, and Carl Zeiss. It resulted in the manufacturing
of optical microscopes that realized, in practice, the maximum theoretically
achievable resolution. It was such experiences, in particular, that shaped the
formulation of the founding statute of the Carl Zeiss Foundation, and the
initiative for the Schott Series is in accord with the commitment expressed
in the founding statute "to promote methodical scientific studies" .

Mainz, Dieter Krause


April 2002 Chief Scientist
Preface

Glass products and their manufacturing require continuous adjustments to


customer needs or wishes, as well as constant process optimizations to ensure
quality or improve yield. Therefore, glass specialists and scientists have en-
deavored to base their understanding of product behavior and processes on
fundamental physical principles or chemical reaction theories. Unfortunately,
the governing equations and their boundary conditions are, in most cases, so
complex that rough approximations or severe simplifications had to be made
in the past, which restricted the solutions to very special situations.
Modern numerical methods in combination with user-friendly commer-
cial software and powerful computer hardware have drastically changed this
situation. Strict numerical treatment, i.e., the mathematical simulation of
product behavior and all aspects of production processes, is a must for every
material-producing company. The specific capabilities required for such sim-
ulations and the results gained by them are essential know-how and therefore
often classified as confidential. Today, many elaborate mathematical tools are
commercially available and some are even public property, but good simula-
tions, which can definitely replace experimental work and save time, are still
not easy to produce. Professional and experienced application and continuous
development, adaptation, and completion of these tools are required, as well
as comprehensive knowledge of the material and process properties, which
is based on complex experimentation and measurement, especially at high
temperatures.
The most important materials for Schott are many types of oxide glasses
and all the ceramics or metals used in the production equipment. In this
volume of the Schott Series on Glass and Glass Ceramics we present a variety
of applications of the mathematical approach to Schott's technology: melting
and fining, forming, annealing, post-processing, and some special product-
related problems. These examples demonstrate the wide use and usefulness of
mathematical simulations on the basis of a solid knowledge of the properties of
"our" materials and processes - without disclosing confidential information,
of course.
Not all of the results were obtainable with commercially available soft-
ware packages. Special problems, for instance the contribution of radiation to
the thermal conductivity in semi-transparent melts, required fundamental re-
VI Preface

search and integration of new software modules into the commercial products.
These tasks were always solved in cooperation with external partners. We ap-
preciate, for example, the contributions from the Technical University Eind-
hoven and from the Fraunhofer Institute for Techno- and Econo-Mathematics
in Kaiserslautern. Many other scientists gave valuable comments Or advice,
often on the occasion of international conferences or meetings of the Technical
Committees of the International Commission on Glass (ICG).
In summary, the information in this book shows a snapshot of successful
transfer of research results via product and process development into pro-
duction. The literature cited will refer the interested reader to more detailed
presentations in specialist fields.
We wish, above all, to express our thanks to all the authors of this
book for their steady and pleasant cooperation. Special thanks are due to
Ms. Petra Schulz, Dipl.-Translator (address: see List of Contributors), who
translated the greater part of the manuscripts from German into English. We
are grateful to Dr. Hans J. K6lsch for helpful discussions in the early phases
of this volume and in planning the Schott Series in general.
Very special thanks go to Wiltrud Witan, M.A., and Karin Langner-
Bahmann, both of Schott Glass, for all the corrections of the manuscripts,
for the creation of the Macromedia Flash presentation on CD and for the
optimization of numerous graphics and images, and for their enthusiasm in
always doing all the hard work necessary to make this book ready for printing.

Mainz, April 2002 Dieter Krause, Horst Loch


Contents

1. Overview............................................... 1
Horst Loch
1.1 Introduction........................................... 1
1.2 Systematics and Boundary Conditions of This Book ........ 3
1.3 Some Important 3D Continuum Equations. . . . . . . . . . . . . . . . . 6
References ................................................. 15

2. Melting and Fining. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 17


2.1 Modeling of the Melting Process in Industrial Glass Furnaces
Ruud G. C. Beerkens. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 17
2.1.1 Application of Process Simulation Models
for Glass Furnaces. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 18
2.1.2 Modeling of Heat Transfer and Convection Flows
in Glass-Melting Tanks. . . . . . . . . . . . . . . . . . . . . . . . . . .. 23
2.1.3 Sand-Grain Dissolution, Behavior of Gas Bubbles
in Glass Melts, and Glass-Quality Index. . . . . . . . . . . .. 47
2.1.4 Models for Evaporation and Superstructure
Refractory Attack by Vapors. . . . . . . . . . . . . . . . . . . . . .. 61
2.1.5 Dynamic Modeling ............................... 71
2.1.6 Concluding Remarks. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 72
2.2 Mathematical Modeling of Batch Melting in Glass Tanks
Wolf S. Kuhn. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .. 73
2.2.1 Motivation and Requirements on Batch Modeling .... 74
2.2.2 Survey of Batch Melting .......................... 76
2.2.3 Theoretical Basis of Batch Modeling. . . . . . . . . . . . . . .. 93
2.2.4 Key Values and Non-Dimensional Numbers .......... 105
2.2.5 Batch Models .................................... 110
2.3 High-Frequency Melting of Glass in Crucibles
Frank- Thomas Lentes ................................... 126
2.3.1 Basics of Electrodynamics ......................... 126
2.3.2 Mathematical Formulation of the Simulation Model ... 129
2.3.3 Simulation Results ............................... 134
2.3.4 Conclusion and Outlook ........................... 136
VIII Contents

2.4 Model-Based Glass Melter Control


Ton Backx . ............................................ 137
2.4.1 Model Concepts .................................. 138
2.4.2 Model-Predictive Control .......................... 142
2.4.3 Extensions of the MPC Technology ................. 148
2.4.4 Application of MPC in the Glass Industry ........... 150
References ................................................. 155

3. Homogenizing and Conditioning .................... 165


3.1 The Intensity of Mixing Processes
Henry Eisermann, Ulrich Lange, Horst Loch,
Gunter Weidmann . ..................................... 165
3.1.1 Description and Quantification of Mixing Processes ... 165
3.1.2 Flows and Particle Paths in Stirrers ................ 169
3.1.3 Statistics of Residence Time and Dispersion ......... 173
3.1.4 Deformation of Infinitesimal Test Bodies
Along Particle Paths. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
3.1.5 Deformation Statistics ............................ 187
3.1.6 Example: a Simple Paddle Stirrer .................. 189
3.1.7 Outlook ......................................... 192
3.2 Instabilities and Stabilization of Glass Pipe Flows
Ulrich Lange, Horst Loch . ............................... 193
3.2.1 Stationary Temperature and Pressure Profiles
in the Pipe ...................................... 193
3.2.2 A Stability Phenomenon .......................... 197
3.2.3 Appendix: Derivation of Several Equations .......... 205
3.3 Shape Optimization of Flanges
Norbert Siedow, Horst Loch, Sandra Manservisi ............ 208
3.3.1 General Shape Optimization:
Continuously Varying Thicknesses and Contours ..... 209
3.3.2 Finite-Dimensional Shape Optimization:
the 3-Ring/Spoke Flange .......................... 230
References ................................................. 237

4. Shaping at Low Viscosities . .......................... 239


4.1 Heat Transfer Between Glass and Mold During Hot Forming
Matthias Brinkmann, Norbert Siedow ..................... 239
4.1.1 Heat Transfer Coefficient Between Glass and Mold .... 241
4.1.2 Physics and Mathematics of the Heat Transfer ....... 245
4.1.3 Sample Computations ............................. 252
4.1.4 Radiative Contributions to the Heat Transfer '" ..... 255
4.1.5 Laboratory Experiments .......................... 259
4.2 Remote Spectral Temperature Profile Sensing
Matthias Brinkmann, Norbert Siedow, Thomas Korb ........ 262
4.2.1 Thermal Radiation in Hot Glass .................... 263
Contents IX

4.2.2
The Inverse Problem
of Spectral Temperature Sensing ................... 266
4.2.3 Sample Computations ............................. 273
4.2.4 Laboratory Experiment ........................... 275
4.2.5 Spectral Imaging of Hot Glass ..................... 279
4.3 Heat Transfer During Casting Experiments
Matthias Brinkmann, Thomas Korb . ...................... 286
4.3.1 Experimental Set-Up ............................. 287
4.3.2 Comparison Between "Exact" Modeling
and Measurement ................................ 289
4.3.3 Alternative Modeling Using the
Active Thermal Conductivity ...................... 290
4.4 Thin-Layer Flows of Glass
Ulrich Lange. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
4.4.1 Example of a Thin-Layer Model .................... 294
4.4.2 Simplified Energy Balance ......................... 298
4.4.3 Validation of the Model ........................... 300
4.4.4 Fiber- and Tube-Drawing Models ................... 302
4.4.5 More Comprehensive Thin-Layer Flow Models ....... 305
4.5 Pressing of Drinking-Glass Stems
Gunter Weidmann, Kurt Holtberg, Henry Eisermann . ....... 306
4.5.1 Model 1: Finite-Element Modeling .................. 309
4.5.2 Model 2: Analytical Modeling ...................... 310
4.5.3 Comparison of Model 1 and Model 2 ................ 316
4.6 The Use of Remeshing Methods in Pressing Simulations
Ulrich Lange . .......................................... 317
4.6.1 Some Technical Aspects of the Method .............. 319
4.6.2 Example: Pressing of a Tumbler .................... 320
4.6.3 Example: Pressing of an "Axisymmetric TV Screen" .. 321
4.7 Chill Ripples in Pressing and Casting Processes
Ulrich Lange. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326
4.7.1 A Simple Casting Process ......................... 328
4.7.2 A Model for Kluge's Experimental Set-Up ........... 330
References ................................................. 335

5. Reshaping at High Viscosities ....................... 339


5.1 Temperature-Dependent Elasticity in Reshaping Simulations
Thoralf Johansson, Ulrich Fotheringham . ................. 339
5.1.1 Model .......................................... 339
5.1.2 Simulation Results ............................... 341
5.2 Sagging and Pressing of Glass Sheets
Thoralf Johansson . ..................................... 343
5.2.1 Model and Boundary Conditions ................... 344
5.2.2 Results of the Model Computations ................. 345
X Contents

5.3 Calibration of Glass Tubes


Thoralf Johansson . ..................................... 349
5.3.1 Model Description ................................ 349
5.3.2 Results of the Model Computations ................. 350

6. Thermal Treatment ................................... 359


6.1 Verification of Relaxation Models
Peter Thomas, Clemens K unisch ......................... 359
6.1.1 Mathematical Models ............................. 359
6.1.2 Experiments in the Lehr .......................... 362
6.1.3 Simulation ....................................... 363
6.1.4 Measuring Stress and Compaction .................. 363
6.1.5 Results .......................................... 363
6.2 Stresses and Crack Growth in Continuously Formed Slabs
Clemens K unisch ....................................... 367
6.2.1 Cooling a Continuous Strip ........................ 369
6.2.2 Crack Growth ................................... 370
6.2.3 Modified Temperature Program
in Order to Avoid Cracking ........................ 371
6.2.4 Cutting the Strip into Slabs ....................... 372
6.3 Thermal Tempering of Drinking Glasses
Volker Seibert, Andreas Bensberg ......................... 374
6.3.1 Principles of Thermal Tempering ................... 375
6.3.2 Results for Spatially Inhomogeneous Quenching ...... 376
6.3.3 Realization of a Quenching Process ................. 378

7. Post-Processing by Laser Cutting ................... 381


Kurt Nattermann, Christoph Hermanns
7.1 Rough Estimation of Process Parameters .................. 381
7.1.1 Stress Levels ..................................... 381
7.1.2 Laser-Beam Profiling ............................. 382
7.1.3 Selection of Laser ................................ 384
7.2 Numerical Analysis of Cutting Processes .................. 385
7.2.1 Calculation of Temperature Distributions ............ 386
7.2.2 Calculation of Stress Distributions .................. 391
7.2.3 Condition for Cut Elongation ...................... 396
7.2.4 Calculation of Stress Intensities for Laser Cutting .... 399
7.3 Practical Realization .................................... 404
7.4 Appendix: Fundamentals of Fracture Mechanics ............ 408
7.4.1 Fracture Mechanics for Brittle Solids ................ 408
7.4.2 FEA Calculation of Stress-Intensity Factors .......... 410
7.4.3 Prediction of the Crack Path ...................... 411
Contents XI

8. Glass Products Under Mechanical


and Thermal Loads . ................................... 413
8.1 Strength Optimization of Airbag Igniters
Kurt Nattermann, Harald Krummel, Ludwig Prank . ........ 413
8.1.1 FEA for Axial-Symmetric Models .................. 413
8.1.2 FEA of 3D Models ............................... 420
8.1.3 Pull-Out Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
8.1.4 Push-Out Tests .................................. 432
8.1.5 Pressure Tests ................................... 435
8.1.6 Appendix: Statistical Procedure .................... 437
8.2 Stiffness and Weight Optimization of a Reticle Stage
for Optical Lithography
Volker Seibert . ......................................... 438
8.2.1 Requirements for a (9 x 9)" Reticle Stage ........... 439
8.2.2 Design of a Prototype ............................. 440
8.2.3 FEM Optimization Without Additional Masses ...... 442
8.2.4 FEM Analysis With Additional Masses .............. 444
References ................................................. 446

9. Simulation and Test of the Spinning Process


Applied to Platinum Metals ......................... 449
Seth Nzahumunyurwa, Hans-Jiirgen Hoffmann, Hans Oechsner,
Joachim Disam
9.1 Necessity to Shape Materials ............................. 449
9.2 Qualitative Description of the Spinning Process ............ 449
9.3 Essential Assumptions for the Modeling
of the Spinning Process ................................. 451
9.4 General Relations for the Model of the
Spinning Process ....................................... 453
9.5 Approximations ........................................ 455
9.5.1 First Approximation:
Quasi-Homogeneous Deformation. . . . . . . . . . . . . . . . . . . 455
9.5.2 Second Approximation:
Linearly Decreasing Deformation ................... 458
9.6 A Practical Example
for the First and Second Approximations .................. 460
9.7 Experimental Observations and Discussion ................. 464
References ................................................. 465

List of Contributors . ...................................... 467


Sources of Figures and Tables ........................... 471
Index ........................................................ 473
1. Overview
Horst Loch

1.1 Introduction

In the second half of the 20th century, a new machine was invented and estab-
lished, which by now has influenced and fundamentally changed all parts of
human life: the computer. The crucial characteristic of this machine is its ca-
pacity to process large amounts of data at high velocity according to precisely
formulated conditions. Basically, there are two possibilities: the numerical
computation in the narrow sense, and the organization, search, manipula-
tion, and description of symbols. Initially, computation was the more impor-
tant aspect, a fact still conveyed by the word "computer". In the meantime,
however, particularly the combination of numerics and symbol processing is
the motor for the enormous progress concerning the application of computer
technology (data bases, graphics, internet, virtual reality, ... ).
The capacities of computers opened up a wide range of new possibilities
for mathematical sciences and engineering. In the previous centuries, many
fundamental natural phenomena were formulated by equations that, in prin-
ciple, described the relative processes exactly and thoroughly. Most of these
equations, however, were so complicated that their solution for arbitrary ini-
tial and boundary conditions was absolutely impossible. Whole generations
of scientists tried to find mainly analytic solutions for specifically simpli-
fied situations, in order to understand the structures behind the equations.
These methods led to deep and beautiful insights into the system of the laws
of nature, and their application to engineering techniques was the basis for
considerable progress. However, if the real problem could not be solved by
the existing simplifying approaches (disturbance methods, use of symmetries,
separation approaches, ... ), the limits quickly became clear. This was partic-
ularly the case for most nonlinear problems. The computer radically changed
this situation. Today, a numerical solution is possible for the entire com-
plex structure of the relevant equations of many real problems. The number
of problems that can now be solved is continuously increasing. Due to these
technical developments, enormous progress was also made with respect to the
search for algorithms, a necessary and worthwhile field of research because
the new technical applications require efficient algorithms for large amounts
of data and complicated numerical operations.

H. Loch et al. (eds.), Mathematical Simulation in Glass Technology


© Springer-Verlag Berlin Heidelberg 2002
2 1. Overview

The computer and the implemented algorithms together form a kind of


"mathematical machine". It is an attractive and very popular vision today
that we might be able to simulate the behavior of interesting real systems
virtually (i.e., without referring to the real world) on such mathematical ma-
chines in order to understand and optimize them. However, before we can
start analyzing a technically relevant system in a "virtual laboratory", it
must first be formulated by equations, i.e., apart from the usual conservation
laws, material and process parameters as well as initial and boundary condi-
tions are particularly necessary. This mathematical modeling, together with
the required validation of the models and model parameters by comparing
them with the real problems, is only possible by a combination of highly de-
veloped experimental and measuring techniques with mathematical methods,
which is virtually the modern definition of natural sciences and engineering.
A mathematical simulation is only possible on the solid basis of such models
and model parameters, which finally enable us to study and optimize the
system we are interested in in the virtual laboratory. Thus, a powerful new
tool has been developed and opens up possibilities that cannot be realized on
a purely experimental basis or, besides, are often too expensive and too slow.
For example, we can look into otherwise hidden internal processes, tryout
many variants under entirely different conditions and study critical situations
without being in danger. With respect to industrial applications, simulation
offers the possibility to take important decisions concerning the development
of new products early on a solid basis, in order to reduce the overall costs.
Simulation enables us to look at the "forward" behavior of a system, i.e.,
we can see what happens if precise initial and boundary conditions are given.
But we have no information about the structure of the global behavior of
the system. We do not know how it behaves in different situations and which
behavior is possible in principle. However, exactly this kind of information
is needed to solve "inverse" problems, i.e., if one searches for states of a sys-
tem with specific characteristics optimized for a particular purpose. Mainly,
there are two optimization methods: either one carries out numerous forward
simulations with intuitively or systematically chosen conditions and then de-
cides on the best result (numerical optimization algorithms represent such
systematic search strategies that can automate this method), or one applies
formula-analytic methods to the fundamental equations, which ideally leads
to a complete analytic survey of the possible system behavior. Logically, there
are advantages and disadvantages: in practice, the first method in most cases
only leads to local optima, and the second method often reaches its limits
due to reasons of time or general complexity (as mentioned at the begin-
ning of this overview). It is interesting that the computer has meanwhile also
led to considerable progress of the formula-analytical method because of its
symbol-processing capacity. Computer-algebra today leads to a wide range of
analyses in a closed, non-numerical form - previous generations of scientists
could only dream of these possibilities. Concerning the applications to the
1.2 Systematics and Boundary Conditions of This Book 3

field of engineering, the adequate answer to the question "mathematics or


experiment?" is "mathematics and experiment!"; analogously, the only cor-
rect answer today to the question "numerical or symbolic computation?" is
"numerical and symbolic computation!" .
As our discussion has revealed, the field of mathematical simulation really
consists of three (technically rather different) aspects:
• the modeling, describing "reality" by adequate equations;
• the forward simulation, solving these equations for given initial and bound-
ary conditions;
• the optimization, looking for initial and boundary conditions that lead to
a specifically optimized system behavior.
There are many prominent and partly spectacular applications of mathe-
matical simulation techniques in today's world: weather prediction, oil ex-
ploration, flight simulators, crash simulations, control of robots (including,
for example, space vehicles), medical tomography, earthquake safety of large
buildings, collision of galaxies, supernova explosions, and so on. Without ex-
aggeration, one can say that mathematical simulation has meanwhile become
a new scientific method that closes a large gap between the purely experi-
mental and the purely analytical methods. Apart from its application in the
field of engineering (which, of course, is the main subject of this book), sim-
ulation has also led to unexpected insights - for example into chaotic and
self-organizing systems - which have thoroughly revolutionized the entire
field of natural sciences.

1.2 Systematics and Boundary Conditions of This Book

This book deals with the application of mathematical simulation for the op-
timization of production processes and product qualities in glass technology
(with glass ceramics being always implicitly included). The intention is to
reach a stable and high output and quality of the production processes by
appropriate tools and machines and adapted process control. Naturally, ther-
mal aspects (temperature control, temperature homogeneity) are of central
importance for the glass production. Concerning the product quality, inter-
esting aspects are, for example, higher strength, lighter weight, improved op-
tical homogeneity, higher or less transparency, adapted thermal expansion,
low compaction.
The main focus of this book is on some important production steps of glass
technology. We therefore organized and categorized our material according to
the various technologies rather than according to the applied mathematical
approaches and formulated chapters on
• melting and fining,
• homogenizing and conditioning,
4 1. Overview

• shaping at low viscosities,


• reshaping at high viscosities,
• thermal treatment,
• post-processing.
Of course, this is a very rough classification that mainly suits our mathemat-
ical view on the subject; production staff would perhaps choose a different
perspective and sometimes also use a different terminology. Each particu-
lar production process has its own individual steps and devices. Moreover,
this list is far from being complete; for example, ceramization, grinding and
polishing, and coating are not treated in this book.
One additional chapter analyzes the behavior of
• glass products under mechanical and thermal loads
and demonstrates the optimization of important product properties.
The last chapter,
• cold forming of metal crucibles,
is devoted to a highly specialized aspect of glass manufacturing: the forming
of precious metal tools used, for example, in pipe systems for glass melts, in
glass-mixing systems, and as glass-forming tools.
A delicate problem concerning the publication of industrial simulation
results is the confidential know-how of an enterprise. Good simulation models
contain very much information about one's own knowledge, in particular
about optimization possibilities and the limits of processes and products.
Therefore, this book essentially describes the state-of-the-art.
As mentioned above, simulation is based on a previous modeling that
itself results from a combination of mathematics, experiment, and intuition
(=: science!). Modeling in the narrow sense, and particularly its complex
experimental aspects and questions of measuring techniques, are comprehen-
sively discussed in other volumes of the "Schott Series on Glass and Glass
Ceramics". The main focus of the present volume being a mathematical one,
the only extensive treatment of the subject is given in Sects. 3.1-3.3, where we
report on some latest developments in the field of temperature reconstruction
from spectral measurements.
Another restriction concerns the model scale: we deal exclusively with
simulations on a technical scale where microscopic processes are condensed
in phenomenological material parameters (viscosity, etc.) and where matter is
considered as a continuum. The field of molecular simulation is only beginning
to develop at the moment, and as yet has had no decisive practical impact on
the glass-process technology. The interested reader may refer to Sect. 3.7 in
Analysis of the Composition and Structure of Glass and Glass Ceramics [1.1]
for detailed information about this promising new technique.
We have also restricted the description of our methods concerning the
application of the mathematical machine computer. Industrial simulation to-
day means, above all, the use of large software packages (either commercial
1.2 Systematics and Boundary Conditions of This Book 5

or developed within the respective enterprise) for the modeling and solution
of problems by very different algorithms and techniques, as well as the use
of powerful computers with complex operating systems. Of course, it is quite
another problem to establish and operate such a system effectively and with-
out any faults. In this book, we will not describe this labor-consuming part
of the job, although it is an essential prerequisite for the methods and results
represented here.
Close at hand is the question of whether the simulation results are correct.
This is an extremely difficult subject that, in practice, usually cannot be dealt
with on a strictly mathematical basis, but must be considered pragmatically.
To start with, we must divide the term "correctness" into different aspects.
The first question is whether the basic model includes all relevant mecha-
nisms of the problem and represents them correctly. This question can only
be answered through a continuous scientific process over several years with
a large amount of experimental work. The second question is whether the
quantitative solutions for a (basically accepted) mathematical model are cor-
rect, i.e., whether all formula and numerical operations have been carried out
correctly, and whether discretizations and iterative solution processes have
been refined to such a degree that the deviation of the resulting solution from
the ideal one can be neglected. Unfortunately, there are only a few practicable
approaches. We can compare the result with reality (= experiment), which is
probably the most reassuring proof of correctness. However, such a compar-
ison is usually only possible for a few selected cases. The second approach is
to check the applied methods by referring to examples from textbooks where
there is consensus about the correct solution. Nevertheless, application to
more complicated real situations naturally leaves doubts. Apart from this,
there are necessary properties (which, unfortunately, are not sufficient) for a
correct solution method. For example, the result of the numerical continuum
simulation must not be influenced significantly by small changes of the se-
lected discretization. As mentioned above for the problem of modeling, we can
say that confidence in the correctness of the mathematical solution methods
only results from a process taking place over several years that includes many
different examinations and comparisons within an enterprise and within the
entire scientific community. For some cases relevant in practice, real proof of
the correctness in the mathematical sense will never be possible (at least not
as quickly as needed). Essentially, this is a special form of the known "re-
maining risk" of all technical methods (only the amateur will be astonished
that this is also the case for mathematics).
And when we have finally gained correct solutions of forward simulations,
we face another challenge: how do we analyze and "understand" these solu-
tions? Think of a three-dimensional transient-fluid flow computation. A typi-
cal result is a field that consists at every point in space and time of a velocity
vector, a pressure, and a temperature. So we have a field with five-scalar data
in a four-dimensional space. Obviously the human brain has trouble handling
6 1. Overview

such complex information! It really does not help very much to look only at
long lists of numbers. Color graphics is one of today's methods to present
such results in an understandable and intuitive way. (Actually, my own in-
terest in mathematical simulation was quickened when I became fascinated
with these graphical evaluations, and the fascination still holds.) A second
method is to condense higher-dimensional information into lower-dimensional
one. A typical example are residence-time distributions and glass-quality in-
dices, which are derived from the primary results (velocity, temperature, ... )
of a glass-tank simulation.
A final remark on the formal mathematical precision of the following
sections: this is not a book for mathematical purists. On the contrary, it is
an attempt to describe as compactly as possible technically relevant results
which, to a considerable extent, are based on mathematical methods.
In summary, we will show how typical industrially relevant problems of
glass technology can be analyzed and solved by methods of mathematical sim-
ulation. The models and verifications preceding these simulations and perma-
nently accompanying them are explained only by a few, mostly present-day
problems. The focus will be on forward simulations and on inverse problem
solving. Apart from purely numerical methods, we apply formula analytical
methods (sometimes extensively with full intention) in order to gain insights
into the global behavior of the systems and to facilitate their numerical de-
scription. The reader should have a basic knowledge of glass production and
the specific glass properties.

1.3 Some Important 3D Continuum Equations

In the following, we have to refer to several large fields of mathematical


physics (particularly thermodynamics, mechanics, and electrodynamics). To
make things easier, we will first describe the most important continuum equa-
tions that repeatedly occur in this book. This is done with a list without
detailed explanation or derivation. The equations and their parameters are
simply summarized in order to give a quick survey and help understand the
complexity of the problems. The notation corresponds to the usual conven-
tions (for example, V' is a "vector" that can be used to form scalar and cross
products, ... ). We refer the reader to the following books for a much more de-
tailed overview and insight into the equations: [1.2] for heat transport, [1.3]
and [1.4] for fluid dynamics, [1.5] for structural mechanics and relaxation,
and [1.6] for electromagnetism. Glass-specific details can, in part, be found
in the following chapters, as well as in the other volumes of the Schott Series
on Glass and Glass Ceramics. Here, we will concentrate on 3D continuum
equations because these are the greatest challenge for a simulation. The pa-
rameters occurring in the equations always depend on space and time. In
the most general case, this is also valid for the material parameters, where
1.3 Some Important 3D Continuum Equations 7

the space dependence is often created indirectly through the temperature de-
pendence. Of course, for the single applications, these equations must also be
completed by several mechanisms that, for example, only depend on the local
temperature (chemical reactions, ... ). Several chapters of this book deal with
the problem of simplifying the general equations as far as possible for special
cases, in order to make them more easily manageable and understandable.
All equations must, of course, always be completed with appropriate initial
and/or boundary conditions.
The most important state variable for glass production is temperature.
If we assume that heat conductivity is the exclusive transport mechanism
(which is at least the case for low enough temperatures or high enough ab-
sorptions within the glass), the respective equation of the convection/diffusion
type is:

pc (~~ + U\lT) = \l('\\lT) . (1.1 )

Here, T is the temperature, p the density, c the specific heat, ,\ the (phononic)
heat conductivity, and U the velocity field (which is zero in a solid body).
Equations of this type in general also describe the convection/diffusion of
chemical species in a basic material. In that case, the temperature must be
substituted by the species concentration.
For high temperatures or small absorptions, thermal radiation is dominat-
ing in semi-transparent glass. Here, mathematical modeling becomes much
more difficult because, instead of a simple diffusion equation, we have a sys-
tem of coupled differential and integral equations. The reason is that apart
from the "indirect" heat conduction, the "direct" heat radiation also becomes
important, which allows a (practically) instantaneous energy transfer between
distant points within the volume. The above equation is completed by a term
for the divergence of the radiation flow vector qr(r), which describes the heat
produced by radiation at each point of the volume:

pc (~~ + U\lT) = \l('\\lT) - \l qr . (1.2)

The new term can be computed in the following way. The radiation transport
equation for the spectral radiation intensity J(r, [1, v) at a point r with
direction in space [1 and frequency v is

[1 \l J(r, [1, v) = K(V, T) [B(T(r)) - J(r, [1, v)] , (1.3)

where B(T(r)) is the Planck function and K(V, T) the absorption coefficient.
The radiation flow vector results from

J J
00

qr(r) = dv dD [1J(r, [1,1/) , (1.4)


o D=47r
8 1. Overview

which finally leads to the divergence necessary for the above-completed heat
conduction equation:

J J
00

V' qr(r) = dv Ii dS? [B(T(r)) - I(r, [1, v)] . (1.5)


o f.I=47r

The computation of the heat radiation in semi-transparent glass at high tem-


peratures is one of the most complex problems in glass technology. Up to
now, the equations have mostly been reduced to the simple diffusion type
by the use of so-called "effective heat conductivities" (approximations of the
Rosseland type). This approach corresponds to the case of high absorption,
respectively, large technical proportions, and necessity was the mother of in-
vention. It results in more or less strongly approximated solutions, and is a
source of permanent uncertainty. By now, the focus is on the search for algo-
rithms that solve the problem "correctly", because simulation has meanwhile
become important for all fields of glass technology, and knowing the correct
temperature distribution is crucial for the correctness of (nearly) all other
mechanical, electrical, and chemical computations. This is a specific problem
of glass technology that seldom appears in other applications of commercial
simulation software, i.e., it is commercially not very attractive for the soft-
ware industry. Therefore, progress has unfortunately been relatively slow up
to now.
For the description of the large-scale flow of glass melts, the adequate
methods are those of fluid mechanics. Here, two special properties must be
emphasized. First, the viscosity TJ of glass depends exponentially on the tem-
perature (see, e.g., the well-known Vogel-Fulcher-Tamman equation), which
yields a strong nonlinearity of the equations. Second, glass flows are incom-
pressible and almost always laminar or even creeping, which can be used for
simplification in many situations. Flows of combustion gases, cooling air, or
cooling water are, of course, typically turbulent. The Navier-Stokes equa-
tions generally describe impulse conservation, and the continuity equation
describes the mass conservation in flowing media:

p (~~ + uV'u) = - V'p + V'(TJ(V'u + V'u T )) + pf , (1.6)

ap
at + V'(pu) = 0 . (1.7)

f is a volume force (e.g., gravitation), p is the pressure. The density p in the


molten glass is linearly dependent on temperature to a good approximation.
When we use these equations, we neglect the elastic effects that are, in prin-
ciple, possible for glass at high deformation velocities. We will not explain
here the models for turbulent flow derived from the Navier-Stokes equations
by appropriate averaging and additional assumptions.
1.3 Some Important 3D Continuum Equations 9

For the computation of deformations and stresses in highly viscous and


in solid glass, the methods of structural mechanics are the most practical
ones. In the so-called transition range (where the molten glass solidifies in a
continuous transition process induced by decreasing temperature, resulting
in an increase of viscosity), glass can be described, to a good approximation,
as a viscoelastic material with structural relaxation. For lower temperatures,
the linear elasticity theory can be applied. In this overview, we will only
describe static, respectively quasi-static, processes where there are no inertia
effects.
If we lay a plane element dA with normal vector n and tangential vector t
through a point of a body under stress, a force dF is effective on this plane
element. The respective stress vector is defined as S = dF IdA. For each
point of a body, the state of stress is determined completely by a symmetric
3 x 3 tensor, the so-called stress tensor S = (O"i,jkj=1, ... ,3. Because of the
symmetry, only six of the nine components are independent of each other. If
the stress tensor is known, the stress for arbitrary plane sections through the
selected point is s = S n. By projection, this yields the normal stress Sn and
the tangential stress St: Sn = (s n)n and St = (s t )t. The stress tensor can be
diagonalized, and the three diagonal elements are called principal (normal)
stresses at the respective point. The strength of glass is often limited by
normal surface stresses (which open existing microfractures); for this reason,
the evaluations of stress simulations focus mainly on the largest principal
stress at the surface.
If a force is effective on a body, this body is deformed. Deformations
of this kind can be summarized by a displacement vector field u( x) =
(Ul (x), U2 (x), U3 (x)) that describes by what amount each point x is displaced
by the deformation. The so-called strain tensor (ti,j kj=1, ... ,3 determines the
induced changes of length and angle. In the case of low strains, the following
linear relations apply between strain and displacement:

(1.8)

(1.9)

tii determine the percentage longitudinal expansion of an infinitesimal rod


in the direction Xi. tij quantify the deformation of a right angle towards
an angle ¢, according to tij = ~ - ¢. In the case of large strains, products
of derivatives and angle functions must additionally be accounted for. The
relations between strains and displacements are purely geometrical and in-
dependent of specific material behavior. Obviously, the strain tensor is also
symmetric. For the purpose of a compact notation, we (formally) define the
two following "vectors":
10 1. Overview

EU au
E22 a22
E33 a33
.-
E .- and .-
a .- (1.10)
E12 a12
E23 a23
E31 a31

These are the respective six independent components of the general strain
and stress tensor.
The simplest possible material behavior is that of a so-called Hooke or lin-
ear elastic solid body. For the isothermal case, it is described in the following
way:

a= GE, (1.11)

where
K+±G
3 K-~G K- ~G 000

K- ~G K+±G
3 K-~G o0 0

K- ~G K- ~G K+±G
3 000
G= (1.12)
0 0 0 GOO

0 0 0 oG 0

0 0 0 o0 G

K is the compression modulus and G the shear modulus. Hooke's law de-
scribes an instantaneous coupling between strain and stress.
For the non-isothermal case, there exist thermally as well as mechanically
induced strains. The linear elastic relation then is

a = G E = G (Etotal - Ethermal) . (1.13)

With a constant, homogeneous, and isotropic coefficient of thermal expansion


a and a homogeneous temperature difference D.T, this yields the following
isotropic thermal expansion:

aD.T
aD.T
aD.T
Ethermal := (1.14)
0
0
0
1.3 Some Important 3D Continuum Equations 11

Stresses occur only if the thermal expansion is obstructed (i.e., if Etotal "I-
Ethermal). For sufficiently low temperatures (below the transition range), the
typical stress-strain behavior of glass can be completely described in this
way. For the transition range, the physics is much more complicated. Shear
stresses tend to relax (viscoelasticity), and the density is also subject to re-
laxation (structural relaxation). The velocity of these processes depends on
the viscosity and thus on the temperature. The general formulas are very
complex, therefore we will only show a typical example of the new types
of equations occurring here. According to Hooke's law, the shear stress 0"12
simply results from the strain E12 , i.e., 0"12 = GE12. A strain that is instan-
taneously caused at time 0 and then maintained leads to an instantaneous
stress that remains constant. At a high (constant and homogeneous) tem-
perature, this shear stress will not remain constant in glass but will relax.
Formally, we can say that the shear modulus G is time dependent and relaxes
converging towards zero: 0"12(t) = G(t)E12(0). If one intends to describe not
only a jump of strain at time 0, but also arbitrary continuous changes of
strain, this results in the so-called Boltzmann convolution integral:

0"12 = i
o
t
G(t -
aE
t/)~(t/)dt'.
at
(1.15)

The instantaneous interaction of Hooke's law has become a retarded interac-


tion in the viscoelastic case. A typical time dependence of the shear modulus
for glass is
n n
G(t) = G(O) 2:wie-(t/T;) , 2: Wi = 1. (1.16)
i=l i=l

The weights Wi and the time constants Ti must be determined by experiment.


A similar theory can also be applied to the compression modulus K, which
does not relax converging towards zero, but instead towards a positive final
parameter. If we substitute the constants G and K in Hooke's law by time-
dependent parameters, and the occurring multiplications of G and K with the
components of E by convolutions according to Boltzmann's law, we get (for
constant and homogeneous temperatures) a linear stress-strain behavior also
for the transition range. This can be generalized for arbitrary temperature
behaviours. However, we will not go into detail here.
For the transition range, apart from Hooke's law, the formula for thermal
expansions must also be modified. These expansions no longer depend only
on temperature change and the coefficient of thermal expansion through a
simple linear relation. We must add an additional term that describes the
structural changes of glass after a jump of temperature has occurred. Apart
from temperature T, we introduce the so-called fictive temperature Tf as a
further state variable, which has proved to be very useful. The fictive tem-
perature is a (very abstract) description for the structural state of the glass.
In the case of stabilized glass (where all structural relaxation processes are
12 1. Overview

finished), Tr = T. This relation particularly applies to molten glass. The


structural relaxation can be computed by differential equations of the follow-
ing type:
aTf
7it(t) = -)..(Tf(t) - T(t)) . (1.17)

The differential equation yields that Tf tries to approximate T. The velocity


of this approximation is determined by the parameter ).. (which must be
found by measurements). We can see that in the transition range, ).. converges
very quickly towards zero with decreasing temperature (i.e., with increasing
viscosity), so that at a certain time there are no more changes of Tf. This
smallest parameter for T f is the so-called glass temperature T g • If we define
the coefficient of thermal expansion of solid glass by a g and that of liquid
glass by aI, this yields the following equation for the computation of thermal
expansion:

aEthermal _ aT () ( _ ) aTf ( )
at - a g at t + al a g at t . (1.18)

The thermal expansions computed in this way can now be applied to the
mechanical stress-strain relations as before.
If temperatures and stress are not too high, a linear elastic behavior can
also be assumed for metals. The formulas for plastic deformations at high
temperatures or for high stresses are left out here.
In glass technology, glass and platinum tools are usually heated not only
by flames, but also using electricity. For the computation of the Joule heat
thereby produced, the universal basis are the Maxwell equations of electro-
magnetism:

\7 D = p, \7xD=--
aB (1.19)
at'
H . aD
\7 B = 0, \7x =J+ at' (1.20)

Here, E is the electric field intensity, D the electric flux density, H the
magnetic field intensity, B the magnetic flux density, j the total current
density, and p the electric charge density. We have j = jc + jj, where jc
is the current density directly caused by the applied electrical sources, and
ji the current density of the induced eddy currents. The Maxwell equations
must be completed by the description of the interaction of electromagnetic .
fields and matter:

(1.21 )

B = J-loJ-lrH, (1.22)
j=O"E. (1.23)
1.3 Some Important 3D Continuum Equations 13

Here, EO is the electric permeability of vacuum, Er the relative electric per-


meability of the material, f-Lo the magnetic permeability of vacuum, f-Lr the
relative magnetic permeability of the material, and a the electrical conductiv-
ity. In the so-called quasi-static approximation (which can often be applied),
the displacement current density (aD) / (at) is assumed as zero.
If the electric field E is known, the power input P through the Joule effect
within the volume can easily be computed:

(1.24)

For the determination of E, the scalar electric potential <P and the mag-
netic vector potential A are often applied, which are defined by the following
relations:

B = V x A, VA =0; (1.25)

aA
E = -at - V<P, ~A = -f-Lf-Loj . (1.26)

There are two different heating methods: the direct electric heating method
conducts an electric current through the glass or a platinum tool. If frequen-
cies are low enough, the eddy currents induced through electromagnetism can
be neglected with respect to the directly applied currents. Only one single
partial differential equation for <P must now be solved:

V (aV<P) =0 . (1.27)

If a remains constant (i.e., if it does not depend on temperature in partic-


ular), this equation can be reduced to the classic potential equation for the
respective region. Inductive heating creates an electromagnetic field outside
of the glass, which causes eddy currents in the interior of the glass by induc-
tion and thus generates the Joule effect. In this case, the above-mentioned
equations for ¢ and A must be solved in general form.
This short introduction can, of course, only give a hint of the manifold
mathematical problems occurring in glass technology. For the numerical so-
lution of the equations described above, in most cases we can use commercial
software today. These solvers are usually based on the finite-volume or finite-
element method. Unfortunately, there is no software that completely covers
all the problems occurring in glass technology. Therefore, we have to use sev-
eral software packages and combine them if necessary. The reader will see
many examples of the application of commercial packages in this book.

The Problem of the Parameters in the Model Equations

In the previous sections, we repeatedly indicated that mathematical mod-


eling essentially consists of two equally important and difficult parts: the
14 1. Overview

formulation of the model equations (which are derived in most cases from
fundamental physical and chemical principles), and the experimental deter-
mination of the parameters therein, which definitely characterize the special
situation to be simulated (in mathematical terms: which lead to a definite
solution of the equations). For well-defined fundamental equations these pa-
rameters quantitatively describe the behavior of the involved materials as
well as the initial and boundary values of the sought solution variables (e.g.,
temperature, stress, deformation, ... ). The material parameters are usually
measured in a lab environment, whereas the initial and boundary values have
to be measured directly at the process to be simulated.
In Sect. 1.3 several equations were presented that describe many impor-
tant phenomena in glass technology. (Of course there are difficult problems
for which, up to now, no satisfying equations could be formulated, e.g., for a
self-consistent description of batch layers, which only contains real material
parameters and initial/boundary values of the solution variables.) We indi-
cated also that today it is possible to get standardized commercial software
for the mathematical solution of many of the interesting equations.
In this section, we want to emphasize that the situation is completely dif-
ferent concerning many of the parameters in the well-known equations! There
are, of course, established measuring procedures for typical mechanical and
thermal parameters of glass near room temperature such as, for example,
Young's modulus of elasticity, Poisson number, heat capacity, heat conduc-
tivity, density, and so on. These are usually parameters needed for the char-
acterization of the final glass products under mechanical and thermal loads,
for example. But the most important objective of our simulation efforts is to
analyze and improve the production processes, which particularly includes
(very) high temperatures. The experimental equipment necessary for mea-
suring, for example, the material properties of glass in this high-temperature
region is definitely non-standard and much more complicated than the equip-
ment required for measurements near room temperature. Moreover, in the
high-temperature region the material itself often changes during the mea-
surement (evaporation, crystallization, decomposition, ... ), so that even the
definition of the parameter in question may become doubtful! Today, it is a
true challenge to get values for
• reaction and diffusion rates in glass melts,
• heat transmission and absorption in glass melts,
• temperature distributions in glass melts,
• shear thinning in the transition region,
• viscoelastic and structural relaxation in the transition region,
to give just a few examples.
Each glass company that really relies on mathematical simulations has
its own methods to measure (and to guess) the required parameters, and to
verify their range of applicability by comparing the theoretical results with
References 15

many practical results. This is the essential knowledge without which all
simulations will only produce nice pictures but no real profit.

References
1.1 H. Bach, D. Krause (Eds.): Analysis of the Composition and Structure
of Glass and Glass Ceramics, Schott Series on Glass and Glass Ceramics
(Springer, Berlin, Heidelberg 1999)
1.2 M.F. Modest: Radiative Heat Transfer (McGraw-Hill, New York 1993)
1.3 R. Peyret (Ed.): Handbook of Computational Fluid Mechanics (Academic
Press, New York 2000)
1.4 R.W. Fox, A.T. McDonald: Introduction to Fluid Mechanics (Wiley, Chich-
ester 1985)
1.5 G.W. Scherer: Relaxation in Glass and Composites (Wiley, Chichester 1986)
1.6 J.D. Jackson: Classical Electrodynamics, 2nd edn. (Wiley, Chichester 1975)
2. Melting and Fining

2.1 Modeling of the Melting Process


in Industrial Glass Furnaces

Ruud G. C. Beerkens
Introduction

Industrial glass melting in tank furnaces involves transfer of heat required


for the fusion of the batch materials and to obtain a low-viscous melt. The
heat transfer process to the melt, in fossil-fuel-fired furnaces, is dominated by
the radiation from the combustion chamber to the batch blanket and to the
glass-melt surface. Another possible source of energy is an electrode system
immersed in the melt. Temperature differences will cause density gradients,
driving free convection flows in the melt tank.
The convection of mass in the melt is essential for mixing and heat trans-
port, particularly for melts with low heat transmission. The mass and heat-
transfer processes have been described in the last two decades of the 20th cen-
tury by several models based on fluid-dynamic laws. From the calculated
temperature field and flow conditions, residence-time distribution and chem-
ical conversions can be derived. Melting kinetics and the removal of gaseous
inclusions are the most critical processes in a melting tank, depending on
temperature, velocity gradients, and glass-melt or batch composition. Today,
there exist models, describing the kinetics of sand-grain or aluminum-oxide-
grain dissolution and the exchange of gases between melts and bubbles. In this
chapter we describe the use of CFD models and post-process melting and fin-
ing models for simulating glass-melt processes. The combustion process and
the heat radiation are briefly discussed and additional models for predicting
evaporation and superstructure corrosion by alkali vapors are shown.
Simulation models offer the possibility of optimizing furnace design and
operating conditions, and, in the near future, fast dynamic modeling will be
applied for so-called model-based control of glass-furnace operations.
Simulation of the most essential processes in glass-melting furnaces is a
valuable tool that becomes increasingly important for the:
• design of glass furnaces,
• optimization of the process operation parameters,

H. Loch et al. (eds.), Mathematical Simulation in Glass Technology


© Springer-Verlag Berlin Heidelberg 2002
18 2. Melting and Fining

• understanding of the physics and chemistry of glass melting and combus-


tion,
• controlling of glass-melting operations.
The first section addresses the possible application of simulation models,
mainly based on computational fluid dynamics, in the glass industry and the
development of such models over the last decades. Section 2.1.2 discusses the
modeling of heat transfer in the melt and combustion space and the modeling
of glass-melt flows in melting tanks of glass furnaces. Illustrative examples of
the application of such modeling will be presented. Section 2.1.3 describes the
submodels on sand-grain dissolution and gas-bubble removal, often used to
model the melting and fining kinetics of glass-melting furnaces. We will also
demonstrate the use of so-called quality-index calculations to characterize the
fining or melting quality of a glass-melting process with specific furnace design
and given process conditions. Section 2.1.4 treats new models currently ap-
plied to estimate evaporation kinetics, emissions, and refractory attack in the
superstructure by using combinations of thermodynamic and mass-transfer
modeling. Section 2.1.5 briefly addresses the development of dynamic models
for predicting the time-dependent response of the glass-furnace state vari-
ables or glass quality to disturbances or changes in the process conditions or
input parameters. These models lay the foundations for model-based predic-
tive control of glass-melting processes. Section 2.1.6 contains conclusions and
prospects for further improvements and the future use of simulation models.

2.1.1 Application of Process Simulation Models


for Glass Furnaces
Before the first mathematical models had been developed for predicting the
glass-melt flows and the heat transfer in industrial glass-melting furnaces,
physical models were applied.
Nolet and Murmane [2.1] presented a paper in 1988 on the development
of physical and mathematical models for the simulation of glass-furnace oper-
ation, predicting or estimating temperature fields and flow patterns. Studies
using physical scale models date back to 1914 [2.2,3]. The first mathematical
models were already presented in the period 1970-1975 [2.4,5]. Since 1980,
improved and detailed 3D models [2.6-9] have been developed. Modeling of
chemical processes such as sand-grain dissolution, fining, and redox state in
industrial tanks [2.10-14] started at the end of the 1980s.
Chaudhary [2.15,16] discussed the present status of mathematical simula-
tion models for glass furnaces and the application of such models for expert
systems and process control. Mathematical models developed by different
groups are presented in the literature references [2.17-22]. Special models for
the simulation of heat transfer and the movements of the batch blanket are
discussed in [2.23,24].
With scale models [2.25-27]' fabricated from transparent polymer mate-
rials such as polycarbonate, the heat transfer and flow patterns, the effect of
2.1 Modeling of the Melting Process in Industrial Glass Furnaces 19

bubbling on the flows in glass-melt tanks, as well as feeders can be exper-


imentally simulated. In these models, the tank or furnace design is almost
completely fixed, and a viscous liquid (oil or liquid polymer) can be used
to simulate the glass melt. The properties of the selected liquid should re-
sult in the same values of the Reynolds, Froude, Prandtl, and Grashof guide
numbers. The scaling of the model and the liquid properties should agree ac-
cording to certain scaling rules. By obeying these scaling rules (for instance
the ratio of RelGr is the same in the scale model as in the real process),
the contribution of free and forced convection and the effect of the scaled
dimensions of the scale model on heat transfer and convection will be similar
(taking the scale into account) to the real furnace.
Today, application of the physical models is mainly limited to processes in
glass melting that cannot be properly simulated by mathematical modeling
because of the complexity of the flow patterns, for instance in the vicinity of
stirrers or bubbling. In most of the mathematical models, the simulation of
bubbling and stirring in the high-viscous melts is based on simplified assump-
tions. Often, physical models are used for the simulation of these processes
or to provide empirical information for the mathematical model [2.28], for
instance the effect of bubbling on the momentum and on the convection in
the melt.
Physical models have the drawback of limited flexibility in modifying the
furnace design; for each major design change a new model has to be con-
structed. The scaling of furnace dimensions and model-liquid properties will
always be a compromise - completely correct scaling of all dimensions and
properties is hardly possible. There are no satisfactory solutions to simu-
late the melting kinetics, rheology, and heat transfer in the batch blanket or
to simulate the batch islands floating on top of the melt. Heat transfer by
radiation can be partly simulated in the physical models by using infrared ra-
diating tiles to simulate the radiation from the flames and the crown, but the
temperature and glass-composition dependency [2.29] of the heat radiation
in the melt will not be correctly simulated.
Today's mathematical models certainly still have shortcomings in simu-
lating melting and heat-transfer processes or calculating the flow patterns
and the transfer of heat and mass in the vicinity of bubbles and stirrers, but
they are becoming more accurate and are more flexible in use than physical
models. Adequate validation [2.30-32] of the results gained by mathemati-
cal models (temperatures, residence-time distribution, flow patterns, energy
balances, redox state) is an essential issue for future improvements.
The performance of glass-melting processes is decisively influenced by the
heat-transfer distribution from the combustion space to and into the molten
glass and the flow patterns of the materials charged into the doghouse and
following different trajectories in the tank. Each process step - melting of raw
materials, complete dissolution of sand grains or refractory grains, degassing
and bubble removal, homogenization, seed dissolution and conditioning -
20 2. Melting and Fining

basically requires a certain residence time, an adequate temperature, and


a suitable flow regime for each trajectory in the tank. To understand the
quality of a glass melt, it is important to investigate the temperatures, the
residence time, and the flow conditions the melt will be subjected to on
its way through the tank. The trajectories through the tank, the prevailing
temperatures, the flows, and the time course along such a path depend on
the furnace design, pull rate, glass properties (viscosity, thermal expansion,
density, surface tension, heat capacity, heat conductivity or extinction of the
heat radiation in the melt), and heat transfer from the combustion space to
the molten glass.

Furnace Design

Mathematical models based on computational fluid dynamics [2.17-22] are


used to optimize the design of the furnace. They help in investigating the
effect of:
• boosting and the electrode arrangement on the heating and flow patterns
in the melt;
• the length-width-depth ratio of the tank;
• the presence of a weir or dam and the position of the weir;
• application of forced bubbling;
• segmentation of the melting tank in a melting, fining and refining or con-
ditioning zone;
• the shapes and dimensions of neck and throat;
• the type of charging machines and doghouse design;
• the deep refiner zone and of
• stirrers and cooling bars,
on the temperature field and the flow patterns, in particular on the shortcut
flow path or the value of the minimum residence time. Figure 2.1 shows
the critical path of the material from the doghouse to the throat at minimum
residence time for a container glass-melting tank with very high pull. Particle
tracing is a technique in CFD models that can be applied after calculating the
glass-melt flow vectors in each grid cell. An infinitesimal volume of material
(called the "particle") follows the directions of the vectors from the doghouse
to the throat. This will lead to a path from the doghouse to the throat.
Different initial positions of the material in the doghouse result in different
paths through the tank. The path of the "particle" with the lowest residence
time is called the minimum residence time path. A shortcut flow is indicated
and identified by this path.
Models combining the descriptions of the melting tank, the combustion
space, and the batch blanket [2.17,33,34] are particularly suitable for design-
ing glass-melting furnaces. Such models can predict the effect of the combus-
tion space design (height, location of flue ports) and the number and positions
2.1 Modeling of the Melting Process in Industrial Glass Furnaces 21

l_ __-- Side view


cross section

Top view

0.0
-
0.5
"
.~,: '.

1.1
Trace time I h
1.6 2.2 2.7

Fig. 2.1. Shortcut flow in a melting tank for container glass production, determined
by CFD modeling and "particle" tracing, time in hours

of the burners on the local heat transfer to the melt. The heat flux distri-
bution is very important for the development of free-convection flows in the
melting tank.

Identification of Critical Process Conditions in Glass Melting


Another application of these models is the identification of the glass-melt
flow patterns that will show the poorest fining or melting conditions. The
optimum process conditions for improving the glass quality and/or for ob-
taining an increased minimum residence time can be sought by using the
CFD (computational fluid dynamics) model in combination with submodels
that describe degassing and melting or homogenization.

Furnace Control

Current developments for application of mathematical simulation models are


the prediction of the dynamic behavior of the melt in a tank during or after
22 2. Melting and Fining

changes in process conditions or changes in the batch composition or after


disturbances in the process. Fast dynamic models can be coupled to glass-
furnace control systems to achieve feed-forward control. Model-based predic-
tive control (MPC) [2.35] requires fast models that can accurately predict
temperatures, flow patterns, or chemical changes in the melt within reason-
able time (at least 100 times faster than the real process) for time periods
of a few days. For MPC, the model calculation time should be orders of
magnitude faster than the real process. Model-based predictive controllers
can be used for adjusting the fuel input, the power distribution among the
electrodes, the batch composition, or the air-fuel ratio so as to stabilize the
conditions in the melting tank or to follow predefined temperatures or glass
composition changes during alterations in pull, glass color, or glass type.

G lass Defect Source Tracking

By modeling the furnace and tracking the glass-melt trajectories in the tank,
bubbles or stones formed at a certain location in the tank can be related to
the final position and composition these defects will have in the glass prod-
uct [2.11,14,17,36-38]. The change of a bubble size and composition or the
stretching and chemical changes of cord or stones in the melt, following the
trajectories calculated by CFD requires submodels for diffusion of gases or
chemical reaction, or, respectively, the diffusion coefficient of Si0 2 or refrac-
tory components in the melt. Unfortunately, the diffusion coefficients and
solubility values of gases and refractory oxides in the glass melt are hardly
known or only available for a limited number of glass compositions. The mod-
els and the required data on these properties are often not reliable or they are
not always sufficiently validated for a reliable identification of the cause and
source of defects in the glass products, using the mathematical glass defect
tracking models.
Several glass research institutes and universities have launched studies
to gain reliable data on diffusion coefficients and solubility. But the large
number of glass compositions and components make systematic investigations
extremely extensive and tedious, and therefore unattractive to researchers.

Limitations for Use of CFD Simulation Models for Glass Furnaces

Unfortunately, not all of the chemical and physical or hydrodynamic processes


can be accurately modeled with first-principle descriptions. The complexity
of the batch blanket limits the possibilities for simulating the melting and
heat-transfer processes in the batch. Foaming in glass furnaces is still very
poorly understood. Recent studies [2.39] yield improved knowledge about
foam formation and the conditions that influence foaming. However, a quan-
titative description of the foam decay rate and a model for the prediction of
foam-bubble sizes in the foam layer are not yet available. To model fining and
2.1 Heat Transfer and Convection Flows in Glass Melting Tanks 23

melting kinetics, the diffusion coefficients and solubility coefficients of gases


or sand (Si0 2) or other refractory raw materials in the melt must be known.
Gas diffusion and gas solubility in molten glass are difficult to measure. So
far, most investigations have not provided sufficiently accurate results. In
practice, the validation of the results of the modeling of industrial glass-melt
tanks is limited to a comparison of predicted temperatures and measured
temperatures at only a few accessible locations in the tank. The residence-
time distribution can be analyzed with tracer studies [2.30,31] and compared
to distributions calculated from the CFD models with the so-called "particle-
tracing" technique. The redox state of the melt can be calculated [2.36] and
measured with potentiometric techniques [2.40]. The partial pressure of oxy-
gen in equilibrium with the dissolved oxygen in the melt can be measured in
the melt in the doghouse area and in the feeders. In order to be able to use the
mathematical simulation model for furnace control and glass defect tracing,
future validation of simulation models is absolutely vital. Reliable models are
also required for finding better furnace designs and for use of such dynamic
simulation models in process-control systems.

2.1.2 Modeling of Heat Transfer and Convection Flows


in Glass-Melting Tanks

The glass-melt tank can be considered as a chemical reactor. In the first stage
of the process, a large number of chemical reactions take place at different
temperature levels and in different phases or between different phases. This
stage is mainly concentrated in or around the batch blanket. After these
fusion reactions, the glass melt contains incompletely dissolved sand grains
or grains of other refractory oxides (Ah03, Zr02, iron chromate), dissolved
gases, and gas bubbles. The next stage in the melting process is the complete
dissolution of the residual grains in the melt. Stirring of the melt and high
temperatures will enhance this dissolution process.
The freshly molten glass contains dissolved gases, smaller gas inclusions
(seeds), large bubbles (blisters), and dissolved polyvalent ions present in a
certain redox state (depending on the level of the oxygen-gas concentration
dissolved in the melt), which is determined by the amount of reducing or
oxidizing components in the batch and the composition of the furnace atmo-
sphere [2.41,42]. At high temperatures, above a certain fining onset tempera-
ture level, the polyvalent ions are converted into a lower redox state and gases
such as oxygen or S02 are generated in the melt. Owing to their low physi-
cal solubility, these fining gases diffuse towards the bubbles or seeds already
present in the melt. The composition of the seeds changes: the original gas
is diluted by the fining gases. The seeds grow, and the initial equilibrium be-
tween the gases in the seeds and the gases dissolved in the melt is disturbed.
Fining gases and other dissolved gases start diffusing into the growing seeds,
driven by the difference in the equilibrium pressure of these gases in the melt
and the changing partial pressure in the seed. The growing bubbles ascend to
24 2. Melting and Fining

the glass-melt surface with a velocity proportional to the square of the diam-
eter and proportional to the reciprocal value of the glass-melt viscosity. This
bubble-removal and gas-stripping process is called fining. It requires temper-
atures above the fining onset temperature and almost no mixing in the melt
to avoid the dispersion of the ascending bubbles throughout the whole melt.
The last stages of fining (refining) involve the slow cooling of the melt.
Now, the residual tiny bubbles that have not reached the glass-melt sur-
face and the bubbles that have hardly grown, can completely dissolve in the
melt. This dissolution (also called secondary fining or refining) requires an
increasing solubility as the melt cools down. Bubbles that contain gases dis-
solving chemically in the melt (e.g., oxygen, CO 2, S02) can be re-absorbed
because the solubility of these gases in the melt will increase during temper-
ature reduction. The dissolution rates of gases and bubbles also depend on
the diffusion coefficients of the gases in the melt, and these values decrease
with temperature reduction. This means that the dissolution of seeds mainly
takes place in a certain temperature range. Below these temperature levels,
the dissolution rates become to small and the seeds "freeze-in" in the rigid
glass.
After this stage, the melt is conditioned to the viscosity level required for
forming of the glass product. The viscosity should be uniform prior to form-
ing in order to obtain a uniform glass distribution during blowing, pressing,
drawing, or floating of the glass. Proper control of the glass-melt viscosity
and the uniformity of the viscosity in the feeder are essential prerequisites
for an optimum forming process.
All the above-mentioned processes require optimum process conditions in
the different sections of the furnace. CFD models have been developed to
estimate these conditions in the melting tank, because direct measurements
of temperatures, dissolved gases in the melt, or glass-melt flows [2.43] in an
industrial melting tank are very difficult to perform and access to the furnace
is limited.
The production of specified glass qualities (in terms of seeds, blisters,
stones, cords, or other inclusions), at highest possible pull, requires optimized
melting, fining, and homogenization processes.

Process Requirements for Melting or Fusion, Fining,


and Homogenizing

In continuous glass-melting tanks, where> 98% of the world's glass produc-


tion is carried out, the essential process steps - melting, homogenizing, fining,
and refining - all take place in one tank. Separation of these processes is often
difficult to achieve without segmenting the melter, and optimum conditions
for homogenization (strong mixing) may be disadvantageous for removal of
gas bubbles during primary fining. The following section will discuss the dif-
ferent process steps and the process requirements for each step.
2.1 Heat Transfer and Convection Flows in Glass Melting Tanks 25

• Melting kinetics, fusion, and dissolution of mw materials into the melt:


Complete fusion of the raw materials to form a glass melt needs a minimum
residence time in the melting zone, depending on the local tempemtures,
the batch composition (chemical composition and grain sizes) and the local
mixing and convection by glass-melt flows. With a high fusion temperature
and strong convection currents the required residence time may be short.
However, the temperature must not be too high because then the refrac-
tory linings of the tank are damaged or the fining process is prematurely
initiated. For Na2S04 fining, the fusion temperatures as a rule should stay
below 1350 DC to avoid early sulfate decomposition.
• Fining, refining:
A properly degassed glass melt requires a minimum residence time within
the fining zone. The melt should pass through a certain tempemture profile
(the whole melt should be heated at least to the fining onset temperature),
depending on the type, the amount of fining agent, and the oxidation state
of the melt. First, the glass temperature should be increased in order to
release the fining gas (an increase of about 50-150 DC relative to the fusion
temperature) and thus induce bubble growth, and in order to decrease the
viscosity. In this way, the growing bubbles are able to rise easily to the
surface. In the secondary fining or refining stage, the occasionally present
residual small bubbles should then be absorbed by a gradual, controlled
temperature decrease. Mixing by convection currents should be limited so
that the bubbles are not dispersed into the melt too much.
• Homogenization:
The glass melt should be mixed sufficiently within the total residence time
by convection currents (macro-mixing) stretching the cord to decrease the
diffusion distance of the main components in the cord (micro-mixing),
whose concentration profiles have to be leveled out [2.44,45]. This cord-
stretching process, which is caused by the velocity gradient in the melt,
can enhance the homogenization process by one order of magnitude or
even more.

The time required for obtaining a melt without non-dissolved sand grains,
for the homogenization of a cord-rich glass, or for the removal of bub-
bles/seeds during fining can be estimated by applying the basic equations
describing the dissolution of sand grains or the bubble growth in static or
flowing glass melts. Table 2.1 gives the estimation rules derived from the de-
scription of dissolution of sand-grain or bubble behavior involving Si0 2 or
gas dissolution and diffusion processes in the melt (derived from equations
such as (2.15)-(2.16ac) and (2.18)-(2.21)).
The minimum time required for these intermediate processes and a given
set of raw-material parameters (grain size R, glass composition, raw material
component concentration in the batch, concentration fining agent) depends
on the viscosity, diffusion coefficient (of gases, Si0 2, or cord components in
26 2. Melting and Fining

Table 2.1. Estimation rules for prediction of required melting, homogenization,


and fining times

Minimum residence time (s) Relation required t


required for
Fusion/dissolution of sand
grains
• stationary melt (P Si0 2/Ps)R 2 D- 1 ~C-1
(no convection)
• forced convection 1.33 (P Si0 2 / Ps) R 4 / 3 Ilgradv 11- 1/ 3 D- 2 / 3 ~C-1
(bubbling)
• free convection 1.07 (P Si0 2/ Ps) g-1/4 R 5 / 4 'T}1/4 D- 3 / 4 ~p-1/4 ~C-1

Homogenization to 1% of
the original concentration
difference cordi glass melt
for a flat slab-shaped cord
• with convection 0.75 L 2/3 D-1/3Igradvl-2/3
• no convection 0.5L2 D- 1

Primary fining (bubble 4 .9'T} 5/9H1/3 -5/9 D- 4/ 9 :-5/9( "C/C )-2/3


finer 9 g ps Ll b
growth and ascension)

R radius of cord or dissolving grain


'T}(T) dynamic glass-melt viscosity
D(T) diffusion coefficient in the glass melt (Si0 2 for dissolution of sand or
diffusion coefficient of critical cord components in glass melt)
Dg(T) diffusion coefficient of most critical gas in the glass melt
~C(T) concentration difference in the melt (dissolved gas concentration or Si0 2
concentration)
9 gravity
Cb gas concentration in the bubble
ps density of the glass melt
PSi0 2 density of grain or inhomogeneity
~P difference in density between melt and inhomogeneity
Ilgradvil absolute value of the glass-melt flow velocity gradient
H finer height of the melt in the fining zone in the melting tank
L initial thickness of the cord (in this case a slab-shaped cord)

the melt) and the local velocity gradients: Ilgradvil (= absolute value of the
gradient (in all directions) of velocity vector v).
The viscosity values and diffusion coefficients, the concentration differ-
ences (difference between the concentration in the bulk of the melt and at
the bubble or sand-grain interface), as well as the velocity gradients as such,
are also strongly temperature-dependent. The residence time of the glass melt
within the furnace should be such that all the minimum time requirements
for each intermediate process are being satisfied. The residence time in each
section of the tank or refiner depends on the glass pull and on the tempera-
ture profile within the melt (free-convection flow). Therefore, there are four
important criteria the melting process has to comply with:
2.1 Heat Transfer and Convection Flows in Glass Melting Tanks 27

• proper temperature profile within the melt,


• sufficient residence time of the glass within the melting tank,
• adequate velocity gradients, but avoidance of shortcut flows,
• low mixing of the melt in the primary fining zone to avoid dispersion of
the bubbles ascending in the melt.

These four aspects are linked to each other by the flow characteristics of
the melt.

Modeling of Convection Flows

Flow induced within the melt by glass pull or bubbling is called forced convec-
tion. Free convection is caused by density differences present within the melt
(due to differences in temperature or composition). Free convection of the
melt in melting tanks ~ outside the bubbling zone and outside the throat sec-
tion ~ is generally much stronger than forced convection (by a factor of about
10). The local convection flows can be intensified by blowing air, oxygen, or
steam into the melt through tubes (bubbling) or by locally introducing extra
heat within the glass, for instance by electric boosting. The convection flow
has a major influence on the temperature distribution (heat transfer) within
the melt, the mixing, and the residence-time distribution. Both the flows and
the temperature distribution in the melting tank can be described by the
physical conservation laws for energy, mass, and momentum.
General form of the conservation laws:
Non-stationary term = convection divergence + diffusion divergence
+ source/sink term
(for stationary conditions the non-stationary term = 0).
The above-presented general equation (for energy, momentum, total mass,
mass of a certain component, or electric charge) states, for instance, that
an increase of the energy level in a small-volume element (imaginary small-
volume cell) in the melting tank is determined by:
• the energy (heat) transported into the volume element by convection and
diffusion (the radiation is formally included in the diffusion term in this
approach),
• minus the energy transported away from this cell,
• plus the energy generated by a chemical exothermic reaction or electric
boosting,
• minus the energy consumed in the volume element by endothermic reac-
tions or evaporation/decomposition processes.
If all the energy input and output plus the energy sources together in the
volume cells compensate each other, then the sustainable energy level in the
volume cells will not change with time and the conditions are stationary: the
temperature at a certain position in the tank will not change with time.
28 2. Melting and Fining

The temperature can be calculated via the law of conservation of energy.


For the glass-melt flow, the relevant conservation laws are the equations for
momentum (Navier-Stokes equations for the velocities or momentum in three
directions) and for mass (equation of continuity). All these equations are
strongly coupled by temperature-dependent properties such as viscosity and
density.
Energy equation, conservation law for energy:
&(pC p T)
-&-t- -div (pcp Tv) + div (,\ grad T) + q
I I I (2.1)
increase sensible convection heat local boosting
heat of heat conduction or cooling

For the "generation" of heat (for instance by electric boosting) q has a


positive sign, for cooling or endothermic reactions q is negative. The value of
the effective heat conductivity ,\ is determined by phonon and photon con-
duction and depends on the temperature and on the transmission spectrum
of the molten glass in the wavelength range between 600 nm and 4000 nm.
The glass melt in the tank absorbs and emits radiation, depending on the
respective temperature and extinction coefficient (wavelength dependent). In
most glass-tank models, the Rosseland approach (radiation and conductive
heat transport is considered as diffusion of heat) is applied, assuming gray ra-
diation (not taking into account the different radiation intensities for different
wavelengths). In the feeders or during the forming process, this approach will
lead to inaccurate results so that other - often more complicated - methods
have to be used [2.46,47]. The first term of the equation is the energy accu-
mulation term, which is determined by the net convection heat transfer, the
total heat conduction, and the generation of heat by reactions or electrodes.
Momentum equation, momentum conservation law:

&~;) = -div (pvv) + div (7] grad v) -grad p + pg .


I I I I I (2.2)
increase convection of shear due to pressure gravity
of flow momentum viscous forces term
Continuity (mass balance) equation for a melt, assumed to behave as non-
compressible fluid:

~~ = -div p v = 0 .
I (2.3)
convection of mass

In these equations, v is the velocity vector of the glass melt.


In the case of electric boosting, the equation describing the conservation
of charge (or no net accumulation of electric charge) also has to be solved
2.1 Heat Transfer and Convection Flows in Glass Melting Tanks 29

for each volume element. The modeling of electric boosting will briefly be
presented in the subsection "Electric Boosting", p. 41 ff.
Using the proper boundary conditions (e.g., the heat-flux distribution
from the combustion space to the melt, batch-blanket position and batch
coverage on top of the melt, heat losses through the tank walls, etc.), the so-
lution of these equations by numerical models will give the local temperatures
and glass-melt velocities in the melting tank. These boundary conditions can
be calculated by dedicated submodels for combustion or batch-blanket melt-
ing. Often, such submodels are strongly coupled to (via continuity in heat
and mass transfer at the boundary) or integrated within the glass-tank CFD
models. Two different situations are distinguished when modeling the flow
conditions and temperature field in the melt tank:

(a) Steady-state behavior of a glass-melt tank. The temperatures and flow


patterns do not change with time and depend only on the position in the
tank (x, y, z). In this case, the term at the left-hand sides of (2.1)-(2.3)
equals O.
(b) Dynamic behavior of the glass-melting process. Knowledge and calcula-
tions of this behavior is required for modeling the glass-melting process
during job changes and color changes or for model-based control strate-
gies.

The dynamic behavior of a glass-melting tank becomes important as pro-


cess conditions become time dependent. In real practice, there will never
be a steady-state situation. Temporary disturbances or fluctuations in input
parameters (e.g., fuel input, fuel composition, cold-air leakage, pressure fluc-
tuations in the combustion space, batch contamination, batch segregation,
etc.) will make processes, flows, and temperatures time dependent. The most
frequently encountered situations where dynamic behavior becomes relevant
are
• the periodic alternating of firing sides in regenerative-fired furnaces, where
every 15-25 min the firing and exhaust sides of the furnace are reversed;
• pull changes;
• glass-composition changes, such as conversion from one glass type to an-
other glass type or from one color to another color;
• model-based process control, to eliminate, at an early stage, the effects of
disturbances or fluctuations in the input parameters on the output param-
eters and on the operation of the furnace.
When, for example, the batch composition is modified in order to change
the color of the glass product, some properties of the batch and the melt
are changed too. For instance, the "effective" heat conductivity, which al-
ways includes photon conductivity, then becomes time dependent, because
the photon heat conductivity depends very much on the spectral properties
of the glass melt. Owing to the time dependency of the glass color in the
30 2. Melting and Fining

melting tank, the value of Aeffective is both position and time dependent, and
the left-hand side terms in (2.1)- (2.3) are not equal to zero. The tempera-
ture field and, consequently, the glass-melt flows will change during the color
change.
The conservation laws have to be met in each position of the tank. Most
CFD models divide the glass-melt tank into a large number of volume cells.
In most simulation models the grid cells are of a rectangular shape. Soft-
ware codes using complex or non-structured grids with non-rectangular grid
cells are particularly suitable for simulating glass-melting processes in non-
rectangular-shaped tanks or refiners [2.49]. The relevant differential equa-
tions, based on mass, energy, momentum, and electric-charge conservation,
are solved for each grid cell and time step. Numerical procedures are used
for solving the differential equations because purely analytical methods for
the complete 3D domain are not applicable. Figure 2.2 shows a grid of small-
volume elements used for CFD modeling of a glass-melt tank. In each volume
element the conservation equations apply. Numerical methods, such as the
solvers described by Patanker [2.48], can be used for solving all these equa-
tions for all the individual volume cells using the proper boundary and initial
conditions. The solution of the conservation equation for curved cells be-

Fig. 2.2. Example of a grid of volume elements for solving the conservation equa-
tions for each cell in the entire tank volume. Local grid refinement has been applied
at the positions of the electrodes
2.1 Heat Transfer and Convection Flows in Glass Melting Tanks 31

comes more complicated, but it offers the possibility of following exactly the
glass-melt tank design in the case of non-rectangular basins.

Boundary Conditions

Boundary conditions must be defined at the refractory walls, the surface of


the melt, and the interface between batch blanket and molten-glass phase.
It is assumed that there is no slip of the glass-melt flow at the surface of
the inner walls and at the batch-blanket-glass-melt interface. The tempera-
ture profile will be continuous through the boundaries. But the slope of the
temperature profile at the interface will change abruptly, due to differences
in the convection and diffusion of the heat transport in the melt, compared
to the heat conductivity or diffusion in the refractory material, in the batch
blanket, or in the furnace atmosphere.
Batch Blanket. The coverage of the melt by a relatively cold batch blanket
strongly influences the temperatures at the glass-melt surface. The temper-
ature differences between the hot spot and locations underneath the batch
blanket drive the free convection in a melting tank. This free convection in-
fluences the heat-transfer processes, the residence-time distribution, and the
mixing of the melt. The batch blanket consumes about 85% of the total en-
ergy needed for fusing the batch, and temperature gradients in the blanket
can be very high. It is important to predict the total energy consumption
of the batch blanket, the temperature distribution, the batch blanket-glass
melt interface temperature, and the batch-blanket flow pattern.
The mass transport (flow) of material from the batch blanket entering
the glass-melt phase (underneath the batch blanket) should be equal to the
material molten in the batch blanket itself. The temperature at the interface
batch blanket-glass melt, the length and thickness of the batch blanket, and
the local release of material from the batch blanket into the glass-melt phase
depends on:
• the heat transfer in the batch blanket itself;
• the kinetics of the fusion reactions in the batch blanket, which mainly
depend on the local temperature in the blanket;
• the flow pattern and movements of the batch blanket (batches are often
charged by screw feeders or batch pushers/scrapers. In the last case, the
batch forms a blanket, whereas use of screw feeders frequently results in
formation of batch heaps floating as islands on top of the glass melt. For
the two types of batch shapes in the furnace, different models are used. A
model description of the batch-blanket movement and the batch coverage
area of the melt is given by Wang [2.17]);
• the energy demand for chemical (mostly endothermic) reactions (Verheijen
[2.50] describes the batch-to-glass-melt conversion and the theoretical heat
demand during batch melting).
32 2. Melting and Fining

Heat is mainly transferred from the combustion space to the top of the
batch blanket by radiation. The bottom side of the batch is heated by radia-
tion and conduction of heat from the glass melt flowing underneath the batch
blanket to the batch. The release of gases at this hot batch-bottom side heats
the inner parts of the batch blanket because hot dissociation gases ascend
through the batch. The penetration of heat into the batch material depends
on the heat conductivity of the batch, the endothermic or exothermic reac-
tions, and on the release of batch gases. During these thermal processes in
the batch blanket, the batch phases change due to chemical reactions and
formation of liquid or gas phases. Composition and density are changed by
the disappearance of the original raw materials, the formation of melts, and
by the changing batch porosity. The batch-to-melt conversion is built up by
a series of sequential and parallel chemical reactions. The conversion rate is
mainly limited by mass transfer (bringing the reacting phases together) and
heat transfer. Therefore, the degree of batch-to-melt conversion is both de-
pendent on the local temperature Tb and on the temperature history of the
glass batch. As mentioned above, batch parameters, such as porosity, density,
conductivity, and reactivity, are temperature as well as time dependent. The
energy balance for the batch blanket can be given by
8Tb
Pb (T, t) Cp,b (T, t) at = '9 [(Ab (T, t) '9Tt)]+S (T, t)+h f (Tg - Tt) , (2.4)

where Pb is the density of the glass batch (kg/m 3 ); Cp,b is the heat capacity of
the glass batch (J I (kg K)); Ab is the temperature-dependent effective batch
heat conductivity (W/(mK)); S is the net heat sink rate due to all chemical
reactions (J I (m 3 s)), h is the heat transfer coefficient between the batch and
enclosed gases (WI (m 2 K)); f is the specific heat transfer area between the
batch and enclosed gases (m 2 /m 3 ); Tg is the temperature of the gases enclosed
in the batch (K); Tt is the temperature of the glass batch (K); and t is
time (s). The last term in (2.4) is often neglected and the effect of released
gases on the heat transfer is taken into account in an effective Ab(T, t) value,
which can be determined by experiments.
In this equation, convection terms are omitted. The term S as a function
of time and temperature has to be derived from laboratory tests using the
complete batch [2.50]. At low temperatures, the batch density is determined
by the particle sizes (distributions) of the individual raw materials. At higher
temperatures, both batch sintering and batch expansion due to the formation
of batch gases determine the batch density. If batch temperature and batch
composition are known, the heat capacity of the glass batch can be estimated
by thermodynamics.
According to the literature [2.50-55], the effective batch heat conductiv-
ity Ab strongly depends on temperature and on the volume fraction of liquid
phase present in the partly molten glass batch. An accurate description of the
energy balance of the heating glass batch therefore requires knowledge con-
cerning the amount of liquid phase present during batch heating. Detailed
2.1 Heat Transfer and Convection Flows in Glass Melting Tanks 33

first-principle models describing all the batch transformations and, conse-


quently, the changes in thermal properties are not yet available. Laboratory
tests and estimates of A and cp values, as functions of temperature and the
course of the melting process, must be experimentally derived for industrial
batches. These values are used in (2.4). Hrrna [2.56] shows the complexity and
the different stages in the conversion process during batch melting. Estima-
tions of temperature-dependent (effective) heat conductivity values, taking
the radiation heat transfer into account can be derived from temperature
gradients measured during laboratory batch-melting experiments.
Because of the high content of silica in glass-forming batches, the amount
of dissolved sand grains can be regarded as a measure for the total conver-
sion of batch into melt for a cullet-free batch. In the simplest batch-blanket
model, the melting rate qrnelt (kg/ (m 3 s)) is calculated by estimation rules
such as: qmelt = (1 - c) AAexp[(T - To)/CC]; in this formula the values of
To, AA, and CC have to be determined for each batch type by melting ex-
periments (To: onset temperature for batch melting, AA: melting starts at
T = To, CC: melting rate is doubled every CC degrees). Better and more
detailed models for the calculation of the melting rate are developed based
on time- and temperature-dependent conversion rates measured in labora-
tory experiments [2.50]. The flow or movement of the batch blanket or the
pattern of batch islands floating on the melt can be simulated by assuming
Newtonian flow behavior, using an effective batch viscosity, and by assuming
forces between the batch islands imposed on the batch by the chargers or
side walls or by assuming a certain predestined batch path [2.17].

Combustion Space

The flux of heat from the combustion space to the top of the batch blanket
or batch heaps and to the surface of the molten glass depends on the position
in the furnace. Heat from the combustion space can be transferred by radi-
ation, in particular between 600 and 4500 nm wavelength, or by convection
of the hot combustion gases and by conduction. The radiation process is the
most important heat-transfer process in the combustion space of fossil-fuel-
fired furnaces. Local heat flux from the superstructure space (combustion
chamber) to the melt or batch blanket is mainly composed of the radiation
from the flames (function of flame thickness, flame emission coefficient, flame
temperature) and the radiation from the crown (emission of radiation and
reflection of flame radiation). The highest heat fluxes to the melt in the tank
are observed at the locations with the highest flame temperatures, and to the
relatively cold batch-blanket surface.
Today (2000), combinations of combustion chamber-batch blanket-glass
tank models exist [2.17,33,34]. CFD calculations are performed separately
for these three sections. The continuity of the heat flux and temperatures at
the interfaces combustion space-glass melt, combustion space-batch blanket,
34 2. Melting and Fining

and batch blanket-glass melt couples the models for the three sections in the
furnace.
Combustion-chamber and glass-tank models are coupled in the following
way: The interface heat fluxes from one section to another section (e.g., from
the combustion chamber to the melt) are used as boundary conditions for the
glass melt. Assuming a certain temperature field at the glass-melt surface,
the heat flux from the combustion chamber to the melt can be calculated
by the combustion-space model. This heat flux and other boundary condi-
tions will determine the glass-melt temperatures. A new temperature field
in the melt and at the glass-melt surface is calculated. This new surface-
temperature distribution is used for combustion-space modeling, and a new
heat flux distribution to the melt is calculated. The same calculation processes
are performed for the batch blanket. The iterative calculation process pro-
ceeds till the heat fluxes and temperature values converge to a final solution
and the values no longer change in the calculation process. Other modeling
approaches use an integrated solution of the differential equations in all grid
cells for the whole glass-melting furnace (combustion chamber, melt basin,
batch blanket).
Simple Combustion Space Radiation Model. In the early days of 3D model-
ing of heat- and mass-transport processes in glass furnaces, the heat fluxes
from the combustion space or superstructure to the glass melt and the batch
blanket were estimated from measured crown temperatures.
Figure 2.3 shows a simplified representation of the heat exchange within
the combustion chamber. The flame radiates heat directly to the glass bath,
and indirectly via absorption and reflection at the crown.

Crown

1 r Mci

Flame

Mg; Mgo

Fig. 2.3. Schematic presentation


of radiation heat transfer within
the combustion chamber
2.1 Heat Transfer and Convection Flows in Glass Melting Tanks 35

The radiation exchange (most important process in the heat-exchange


processes) is described by the following (simplified) relation given by (2.5).
The expression for net heat flux (from flames and crown minus the radiation
heat loss from the melt surface) to the glass-melt bath can be derived from
a simplified balance of radiation heat fluxes. In a simple one-dimensional
model, the net flux Qheat(X) at position x in the furnace is given by:

AO" (Tf4 (X)-Ti(x))+BO" (T~(x)-Ti(x)) 2


Qheat (x) = C (W jm ) , (2.5)

where T is the temperature (K); 0" = 5.67 X 10- 8 [W j(m 2 K4)] (Stefan-
Boltzmann constant); and
A = Eg Ef(2 - Ef - Ec + Ec Ed, The indices mean:
B = Eg Ec(1 - Ed, g = glass bath, f = flame,
C = 1- (1- Er)2(1- Ec)(l- Eg) . c = crown.
E is the emission coefficient (here assumed to be independent of the wave-
length).
This simplified model for local heat flux Qheat(X) from the combustion
space to the melting tank gives only a heat-flux distribution in the x direction.
The effect of flame length and shape and details of the flame spectrum On
the net radiation heat transfer are not taken into account.
Detailed Combustion and Radiation Modeling in the Combustion Space. So-
phisticated combustion models calculate the local composition of the fuel-air
or fuel-oxygen mixture (often taking into account statistically determined
fluctuations in the mixture composition with time) as well as gas expansion
and flows. The combustion reactions (assuming local equilibrium or certain
reaction kinetics) release energy. This energy is transported by radiation and
convection to the environment (walls, glass melt, batch blanket, local gas in
the combustion space); the exhaust gases transport heat from the combus-
tion space to the heat exchangers of the furnace. Typical combustion model
results are time-averaged gas flows (vectors), time-averaged temperatures of
the flames in the combustion space, emission coefficients, crown temperatures,
and radiative and convective heat flux values. Figure 2.4 shows the calculated
temperature field in a horizontal plane at the burner level in a combustion
chamber of a glass furnace. Future models will also address the spectral prop-
erties of the flames and the calculation of the radiation intensity depending
On the wavelength bands, but such calculations are time-consuming. Models
exist for estimating also the formation of nitrogen oxides in the combustion
space [2.57]. In the following we give a short description of the combustion
processes occurring in glass-melting furnaces and describe the current status
of combustion models applied for simulation.
In combustion chambers of glass-melting furnaces, complex chemical re-
actions occur in highly turbulent flows at high temperatures. The level of tur-
bulence determines the rate of mixing between fuel, oxidant, and products,
and thereby the flame temperature. Heat transfer from combustion products
36 2. Melting and Fining

- :> _ _ _ I _ _ •
1000 1200 1400 1600 1800 2000 2200

Fig. 2.4. Temperature field (0C) in a horizontal plane of the combustion chamber
of a natural-gas-fired float-glass furnace

to furnace walls and glass melts mainly takes place by thermal radiation. The
radiant emission of the combustion atmosphere depends on the temperature
and the levels of H 2 0, CO 2 , and soot. Various submodels are used to ac-
curately describe these heat-transfer processes and combustion in the CFD
model "TNO-WISH3D":
Turbulence Model. Turbulence is accounted for by the well-known two-
equation k-c turbulence model by Launder and Spalding [2.58]. Turbulent
fluctuations of the variables (for instance fuel-air ratio) in the grid cells are de-
scribed by means of statistical methods. The instantaneous value of a variable
is decomposed into a mean value and a randomly fluctuating part. Successive
time-averaging of the equations leads to the introduction of unknown terms:
the Reynolds stress tensor and the turbulent heat flux by the fluctuating ve-
locities. Empirical relations for these turbulent transport terms are required
to solve this problem (turbulence closure). In the k-c turbulence model both
transport terms are estimated by making use of a turbulent diffusion for mo-
mentum, the eddy viscosity. In the eddy-viscosity concept, first postulated
by Boussinesque, the turbulent momentum flux is related to the velocity gra-
dient of the time-mean flow. In a similar way, the turbulent heat transport
is thought to be proportional to the mean temperature gradient. In the k-c
model, the mean kinetic energy of the turbulent fluctuations (k) and the dis-
sipation rate of turbulent kinetic energy (c) are used to estimate the value of
the eddy viscosity. In analogy with the kinematic viscosity derived from the
kinetic theory of gases, in which this viscosity is proportional to the product
of the average molecular velocity and the mean free molecular path, in the
k-c model the eddy viscosity is estimated as the product of a characteristic
velocity scale of the turbulent velocity fluctuations and an appropriate length
scale of the turbulent eddies. In the k-c model, k is used for determining the
velocity scale and c for determining the length scale. For both variables, k
and c, a convection-diffusion differential equation is solved. The eddy viscos-
ity calculated from local values of k and c is added to the molecular viscosity
2.1 Heat Transfer and Convection Flows in Glass Melting Tanks 37

in all differential equations describing conservation of mass, momentum, and


energy.
Radiation Model. Radiation heat transfer from surface to surface, and radia-
tion absorption and emission by non-transparent gas components in the com-
bustion atmosphere (H 2 0, CO 2 , soot) are modeled with the discrete transfer
model (DTM) as proposed by Lockwood and Shah [2.59]. The DTM method
is based on a direct solution of the radiation transfer equation. For each wall
point the entire solid hemisphere is divided into a discrete number of solid
angles (beams) for which the radiation intensity is assumed to be constant.
The radiation transfer equation is integrated along each beam. All beams
are traced towards opposing wall points whose temperatures are known from
a previous iteration. The incident heat flux on the wall point, which comes
from opposing wall points after crossing the non-transparent gas mixture,
is thus obtained by summation. The heat flux leaving the wall point is the
sum of the reflected incident radiation heat fluxes and the emitted wall ra-
diation. From the surface energy balance (the net radiation flux on the wall
point should balance the heat flux through the wall and the increase in wall
temperature), the wall temperature can be determined. The net radiation
absorption and emission for each control volume appear as source terms in
the energy-balance equation.
The gas-mixture extinction coefficient is calculated from the emission co-
efficients of H 2 0 and CO 2 . The emission coefficients depend on temperature,
partial pressure, and on a characteristic length scale (flame diameter or fur-
nace size). They are taken from the Hottel charts given by Hottel and Egbert
in [2.60]. The total emission of the gas is the sum of the radiation emissions
of H 2 0, CO 2 , and soot.
Combustion Model. In the flame sheet model, the combustion process in a dif-
fusion flame (mixing of fuel and oxidant during combustion) is approximated
by a one-step, irreversible, infinitely fast global reaction (Burke-Schumann
approximation). The reaction occurs at the interfaces or flame front, where
the mixture fraction f has the stoichiometric value. Outside this very thin
flame sheet, no combustion reactions are assumed to take place. The hot
combustion products are mixed with the oxidant or fuel stream. On the
flame front itself only combustion products (plus inert gases) are present;
outside the flame front either fuel and products (plus inert gas components)
or oxidants and products (plus inert gas components) are present. Fuel and
oxidants cannot coexist at the same time and in the same place due to the
infinitely fast reaction assumption. It can be shown that all species mass frac-
tions are piecewise-known functions of temperature and one conserved scalar
variable: the mixture fraction. For this single variable, a convection-diffusion
differential equation must be solved, coupled with the Navier-Stokes, turbu-
lence, and energy equations. Using the previously mentioned piecewise func-
tions/relations, mass fractions of the reactants or inert components (fuel, ox-
idant, product, and inert gases) are deduced from the mixture fraction. Mass
38 2. Melting and Fining

fractions of individual species (such as CH 4 , CO 2 , H 2 0, etc.) are calculated


from the reactant mass fractions.
In the case of oil combustion, instead of natural-gas combustion, a di-
luted two-phase model is used. In this model, the oil-droplet trajectories
through the furnace are calculated in a Lagrangian way. As the droplets
move through the combustion space, momentum, mass, and heat transfer
from the liquid phase to the gaseous phase, and vice versa, are taken into
account. The droplets evaporate according to these laws and, as soon as the
temperature of the particle has reached the boiling temperature, the parti-
cle temperature is assumed to remain constant. Once the oil is in the gas
phase, combustion will take place when enough oxygen is available. The two-
phase oil-combustion model is used in conjunction with the above-mentioned
conserved scalar mixture fraction model.
Dissociation. A drawback of the flame-sheet model is the absence of radical
species in this model. The existence of intermediate radical species (e.g. OH,
0, H) in the reaction zone results in lower flame temperatures. This effect
can be modeled by an empirical correction function, presented by Post [2.61].
This correction function effectively increases the mixture-specific heat and
thus reduces the temperature for a given available enthalpy H. The correc-
tion function becomes effective for T > 1920 K. At the very high temperatures
prevailing in glass-melting furnaces, this correction is essential for obtaining
realistic temperature predictions. The dissociation correction function has
been tuned for maximum flame temperatures with a constrained-equilibrium
model by Peeters [2.62]. This extremely detailed model takes over 150 chem-
ical reactions into account, including the formation and recombination of
many intermediate radicals. Even super-computers are presently unable to
allow this level of detail for 3D or even 2D simulations. Other models pre-
sented in the literature [2.57] have been applied to estimate the formation
and concentrations of NO and N0 2 in the combustion space of glass furnaces.
Soot Model. The presence of soot in a furnace promotes radiation and hence
the efficiency of heat transfer from the flame to the glass melt. In order to es-
timate this effect, the amount of soot and its distribution in the furnace must
be known. Despite the highly complex nature of the chemistry of soot for-
mation and oxidation in flames, some general principles are well understood.
Four, more or less separate, processes can be distinguished .
• Nucleation: fuel is decomposed and the first soot nuclei appear. These
are, amongst others, unsaturated hydrocarbons, particularly acetylene,
and polycyclic aromatic hydrocarbons. Such particles are very small (d <
2 nm), so that even if they occur in large numbers their contribution to the
bulk mass of soot is negligible.
• Coagulation: the roughly spherical soot particles form clusters, thereby de-
creasing the number of soot particles, while the total mass of soot remains
unchanged.
2.1 Heat Transfer and Convection Flows in Glass Melting Tanks 39

• Surface growth: this is the mechanism by which the bulk of the soot mass
is generated. It involves the chemical attachment of gas-phase species to
the surface of the soot particles, by which the soot mass is increased, while
the number of particles remains constant.
• Oxidation: soot is oxidized mainly by O 2 and OR to form products such
as CO and CO 2 plus H 2 0.
The amount of soot being formed depends strongly on the type of fuel.
Methane and ethane give relatively low soot concentrations compared to ethy-
lene and acetylene. Soot formation and decay rates are kinetically limited,
particularly in the first part of the flame. This means that state relationships,
for example soot mass concentration as a function of the fuel-oxidant (air or
oxygen) mixture ratio, cannot be expected to be very accurate because the
concentration of soot will depend on the residence time. Accurate simula-
tion of soot requires solving a transport equation for the soot mass fraction
with appropriate source terms. The instantaneous source term is written as
a formation term minus an oxidation term. It is assumed that, because in
natural-gas flames only small amounts of soot are formed, the other species
concentrations are not influenced by the formation of soot. Moreover, the
reaction enthalpy hardly influences the gas temperature. Hence, the mixture
fraction determination is decoupled from solid soot contents, greatly facilitat-
ing the modeling. The only interaction of soot with the flow field is through
the enhancement of flame emission in the presence of soot.
The soot concentration is coupled to the radiation model via the Planck
mean of the absorption coefficient of soot, which is added to the local ab-
sorption coefficient of the gas. For the description of the soot-formation pro-
cess the Khan and Greeves [2.63] formation model is used. For the soot-
oxidation process an eddy-break-up model as formulated by Magnussen and
Hjertager [2.64] is applied. This model assumes that turbulence decay con-
trols the rate of soot oxidation.
In the combustion model, the field variables (temperature, velocity, fuel-
air mixture, emission coefficient) are computed in an iterative way, using the
results from the previous iteration. If implicit time steps are chosen, several
iterations are executed until certain convergence criteria are satisfied. In a
combustion simulation for a glass-melt furnace, the code solves, on average,
nine transport equations simultaneously. These are:
• the continuity equation,
• the three Navier-Stokes equations for the velocities (momentum transfer
in three directions),
• the transport equation for the turbulent kinetic energy,
• the transport equation for the turbulent energy dissipation,
• the energy equation for the enthalpy/temperature (both in fluid domain
and solid walls),
• the mass transport equation for the mass fraction of a species or the mix-
ture fraction,
• the mass transport equation for the mass fraction of soot.
40 2. Melting and Fining

All gas properties (density, viscosity, thermal conductivity, heat ca pacity,


and extinction coefficient) are functions of both temperature and gas com-
position.
Different groups have developed combustion models [2.9,17,33,57,61] for
glass furnaces. The connection with the glass-tank model is particularly im-
portant for an accurate calculation of the local heat fluxes from the combus-
tion space to the melt, see Fig. 2.5b.

a)

o 1
Absolute velocity Ims- 1

b)

Fig. 2.5. (a) Calculated average gas flow patterns in the combustion space of
a float-glass furnace. (b) Coupled combustion-chamber~glass-melt tank modeling.
Temperature field at the refractory walls and in the melt for a float-glass furnace
2.1 Heat Transfer and Convection Flows in Glass Melting Tanks 41

Electric Boosting

When applying voltage differences between electrodes positioned in the glass


melt, electric currents flow through the melt, for instance from one electrode
to another or several other electrodes. All-electric furnaces are equipped with
sufficient electrodes to provide all the energy required for melting, by the
Joule heat dissipation of the electric current in a melt with an electrical
conductivity C>e. The electric field E (vector), which is generated by the
electrodes, relates to the local electric potential ¢ (scalar) in the melt:

E= -V¢. (2.6)

The electric current J (vector) in this field is

J = C>eE. (2.7)

In general, accumulation of electric charge is negligible and this means that

VJ=O. (2.8)

The dissipation of electrical energy in the form of heat formation Q in the


melt through which the current flows is given by:

(2.9)

The values of Q, C>e, and of the vectors J and E depend on the position in
the tank and on time, temperature, glass composition, electric field, and so
on.
Equation (2.8) is valid for a steady electric field and current. In glass-melt
tanks with several electrodes, alternating electric currents (AC) between the
electrodes are applied. Often, a balanced three-phase system is used, which
is composed of three single-phase voltages with the same amplitude ¢max,
frequency w, and a phase difference of 120 0 between the voltages of the
differently connected electrodes.
The potential field in the melt at location vector x at time t, with an
amplitude ¢max(x) and a certain phase shift B(x) can be written in the
complex number expression:

¢(x, t) = ¢max(x) expi(wtH(x)) (2.10)

Only the phase shift B(x) and the value of ¢max(x) vary with location in
the melt. The electrode potential (potential at the electrode surface) itself
is given in the same way. Assuming a certain temperature field in the melt
and applied electrode potentials (often AC potential field), the electric cur-
rent and electric field (E) in the melt can be calculated, dependent on the
temperature dependency of C>e. Based on (2.7) and (2.9) and averaging of the
heat dissipation over one AC cycle, the local average heat dissipation Q(x)
42 2. Melting and Fining

can be calculated from the imaginary and real amplitudes of the electric field
vector E:

(2.11)

The average value of the heat dissipation is of interest, as the source term
in the energy conservation partial differential equation (2.1). IIErel1 is the
absolute value of the real component of the electric field strength, IIEim II is
the absolute value of the imaginary component at location x. Starting from
a certain potential AC field on the electrode arrangement, the electric field in
the melt can be calculated depending on the distance between the electrodes,
the connections, and the electrical conductivity CJe. At the location n of the
nth electrode at time t, the supplied AC potential <P can be described by

<Pelectrode(n, t) = <Pmax(n) [cos(wt + B(n)) + i sin(wt + B(n))]


(2.12)
= <Pmax(n) exp[i(wt + B(n))] .

Separation of the imaginary and real part of the potential field and obeying
(2.8) leads to the Laplace equations:

\7CJe(X)\7<Pre(X) = 0 , (2.13a)
\7 CJe ( X) \7 <Pim (x) = 0 , (2.13b)
II<pmaxl1 2 = II<pe11 2 + II<pim11 2 .

<Pre(X) and <Pim(X) can be calculated by solving the differential equations


(2.13a) and (2.13b) and using the boundary conditions given by (2.12).
Knowledge of the values of <Pmax(x) and the phase B(x) in each grid cell
(at position vector x) is sufficient to calculate the imaginary and real com-
ponents of the voltages at a certain phase B( x):

<Pim(X) = <Pmax(x) sin(B(x)j360) and <Pre(x) = <Pmax(x) cos(B(x)j360) .


(2.14)

Equation (2.12) can be converted to equations similar to (2.14) for the bound-
ary conditions separating the imaginary and real part of the potential value:
<Pelectrode(n, t). Using the boundary conditions and solving the Laplace equa-
tions, results in the potential field in the molten glass. Using the electrical
conductivity values of the melt, the local average heat dissipation in the grid
cell can be calculated by (2.11). The electric current caused in the melt by
the electric field, together with the magnetic field induced by the current in
the electrode, results in a Lorentz force on the melt. Chaudhary [2.65] cal-
culated this force for a cylindrical electrode vertically immersed in the glass
bath. Only in the tiny area surrounding the electrode does the Lorentz force
cause some extra glass-melt flow, but in most cases the effect is negligible
compared to the other forces and the free-convect ion-driven flows.
2.1 Heat Transfer and Convection Flows in Glass Melting Tanks 43

In modeling the electric heating by electrodes in all-electric-heated tanks,


or in tanks using electric boosting, the maximum potential and the phase at
each electrode are important input parameters. The calculation of the heat-
ing process of the melt by electrodes is an iterative procedure, starting with
a first assumption of the temperature field; then the electrical conductivity
and the electric field can be calculated using (2.8), (2.9), and (2.11). The
electric current can be determined by (2.6) and (2.7), and the heat dissipa-
tion by (2.12). This, however, leads to heat dissipation and the temperatures
and velocities have to be adapted taking into account convection and heat
conduction (using (2.1)). The values of the temperature-dependent electrical
conductivity must be adapted in each grid cell, and the electric field and cur-
rents must be recalculated. This calculation process has to be repeated until
convergence of the temperature values and the electric currents is achieved.
However, this requires an additional iteration process, first assuming an AC
potential on each electrode and then recalculating the potential until the re-
quired power input or required current is realized. It is possible to fix - as the
electrode boundary condition - the value of the potential on the electrodes
and the electric power input or electric currents through the electrodes. The
electric field and heat dissipation can then be calculated.
Figure 2.6 shows the temperature distribution, calculated for an all-
electric melter. The electrodes are also presented in this figure. In the ar-
eas surrounding the horizontal electrodes, elevated temperatures (red-orange
color) can be observed; these local hot spots drive the free-convection flows
in the melting tank. Colder spots (green-yellow color) can be recognized in
the center of the tank.

1260 1280 1300 1320 1340 1360


Temperature IGTM

Fig. 2.6. Example of calculated temperatures in cross sections of an all-electric-


heated melt tank with a cold top of freshly charged batch
44 2. Melting and Fining

Results of the Solution of the Conservation Equations


with Boundary Conditions

In this section, some examples of the results of CFD calculations will be


presented. Accurate solution of the differential equations for mass and energy
transfer yields a large set of calculated data (e.g., temperatures at the nodes
of the grid cells, glass-melt velocities in three directions in the nodes of the
grid cells). Figure 2.7a shows the contours of temperatures determined for
the superstructure walls and the melt in a tank of a container-glass furnace .
Another example for a furnace equipped with separated sections for melting,
fining, and refining and a segmented superstructure is given by Fig. 2.7b.

a)

glass melt

b)

Temperature /"C

1200 1300 1400

Fig. 2.7. (a) Temperature distribution at the surface and cross section of the glass
melt and at the back wall, of an end-port-fired container glass furnace. The tank is
equipped with a weir, electrodes, and a deep refiner. (b) Temperatures calculated
in the glass melt in a segmented tank with segmented superstructure
2.1 Heat Transfer and Convection Flows in Glass Melting Tanks 45

Figure 2.8 shows the calculated temperature field (indicated by color


fields) and flow vectors for a TV-glass melting tank, induding the refiner
and the feeders. The hot-spot zone is the section in the tank with the highest
temperatures. The lower density of the melt at these zones compared to the
colder zones induces free-convection flow. In general, the melt flows upward
(from bottom to surface) in the hot-spot sections. The part of the tank with
the most intense flow from the bottom to the surface is called the spring zone.
The spring-zone region often coincides with a sectional division of the tank
for melting and fining; this is the ideal situation. The first section - between
the doghouse and the spring zone - has a glass-melt circulation flow from the
hot spot along the surface of the melt underneath the batch blanket to the
so-called back wall (wall at the batch-charging side of the furnace) and then
downwards to the bottom and along the bottom to the hot spot or spring
zone. The second section - between the spring zone and the bridge wall or
the neck - forms a second re-circulation loop. A shortcut glass-melt flow over
the bottom underneath a second convection roll in a tank and downstream of
the spring zone, or a shortcut between the two sections, can impair the glass
quality because part of the melt may not pass through the hot-spot zone
at the surface and thus reaches the neck or throat along the bottom within
a very short time. A distinct spring zone can be obtained by bubbling gas
through bubbling pipes mounted in the bottom of the tank in the vicinity of
the hot-spot zone. The gas bubbles, typically of 10-15 cm diameter in size,
bring relatively cold glass melt from the bottom to the surface. This improves
the heat transfer from the combustion chamber into the glass melt and en-
forces the re-circulation flows. The surface temperature of the glass melt is

Spout

Fig. 2.8. Temperatures and flow vectors derived from CFD modeling of a melt
tank, refiner, and feeders of a TV-glass furnace
46 2. Melting and Fining

decreased by mixing the glass melt of the lower layers with the top layers.
This increases the temperature difference between the combustion space and
the glass-melt surface and, consequently, increases the net radiation heat flux
according to (2.5).
On the other hand, intense gas bubbling may result in very high glass-
melt velocities along the bottom and thus create new shortcut flow paths, or
the increased convection may enhance refractory corrosion. Figure 2.9a shows
a cross section of the tank of an end-port-fired furnace with a hot spot close
to the front wall (close to the throat); here, the spring zone is not developed
as in the case of Figs. 2.8 and 2.9b. As Fig. 2.9a indicates, part of the glass
melt flowing from the doghouse to the throat never reaches the surface areas
of the melt in the end-port-fired furnace.

Particle Tracing

After the determination of the temperatures and the velocities of the glass-
melt flow in all three directions in every grid cell or node, the path of a small
volume of material, first traveling through the batch blanket and then passing
the glass melt, can be tracked. There is an infinite number of different possible
paths starting from the charging end to the throat or canal of the furnace. At
each initial position in the doghouse, a path starts. Traveling along the path
takes a certain time period, and thus the material following each path has
a certain residence time in the tank. By calculating the residence times of a

Hot spot
a)

Spring zone
b)

950 1150 1350 1550


Temperature /GTM

Fig. 2.9. (a) Illustration of calculated glass-melt flows in the melt tank of an end-
port-fired furnace in the center plane of the tank in the length-height direction.
(b) Temperatures and flow patterns in a cross-fired furnace, with a distinct spring
zone
2.1 Sand-Grain Dissolution, Behavior of Gas Bubbles, Glass-Quality Index 47

large number of paths (starting from different positions at the entrance of the
doghouse), a residence-time distribution (RTD) can be derived. An accurate
presentation of the RTD requires from 5000 up to 20 000 different paths.
The path with the minimum residence time is often called the critical path
or the shortcut flow trajectory. The calculated RTD can be compared with
measured RTD curves (Fig. 2.10) by using a tracer compound in the batch.
ZnO or SrO are frequently used for this purpose. Other components that are
not present in the regular glass composition and can be easily analyzed at
low concentrations in the glass product by analytical techniques (e.g., X-ray
fluorescence) are also suitable. However, only tracers following the same flows
as the parent glass melt should be chosen.

2.1.3 Sand-Grain Dissolution, Behavior of Gas Bubbles


in Glass Melts, and Glass-Quality Index

Complete melting of the raw materials, particularly the conversion of sand


grains into (glass melt dissolved) Si0 2 , and the removal of gases from the
molten glass are essential in industrial glass-melting processes. The final melt
should be homogeneous and, for most applications, free of such cords (or at
least free of cords leading to unacceptable distortion or mechanical weakness
of the final glass products), free of non-molten material (sand or refractory
grains), and free of seeds and blisters. The glass melt should be well "out-
gassed" to avoid reboil in later stages of the glass-fabrication process.

0.05
- 0 - Feeder (right)
0~ 0.04 - 0 - Feeder (left)
~ - e - Modeling
0
c::
N 0.03
'0
c::
0

~ 0.02
C
Q)
(J
c::
0
() 0.01

0.00
0 10 20 30 40 50
Time since adding the tracer Ih

Fig. 2.10. Residence-time distribution measured (D, 0) with ZnO tracers for glass
exiting the tank through both feeders of an end-port-fired container-glass furnace
and calculated from a CFD model (.) for this furnace [2.31]
48 2. Melting and Fining

Sand Dissolution

In the cristobalite phase, pure Si0 2 only melts at temperatures above


1723°C. In most industrial glass-melt tanks such temperatures are not
reached. Only for melting of some special glasses and for vitreous silica pro-
duction do higher temperature levels have to be achieved. Melting of sand is
the wrong expression for most situations in the glass industry. Sand partly
reacts in the batch blanket at the grain surfaces with other batch components
or with the first-appearing melt phases. Particularly reactive towards sand
are alkali components such as sodium and potassium carbonate: alkali-silicate
melts are formed. These melts further react and mix with other raw mate-
rials components. In the batch blanket, the sand-grain size is decreased by
the reactions with soda, potassium carbonates, or early alkali-rich or alkali-
earth-rich melt phases.
At temperatures exceeding 1100 °C, still undissolved components such as
not completely reacted sand grains, do not melt separately, but dissolve into
the already-existing melt phases by diffusion processes. In conventional glass
types, quartz sand usually is the slowest-dissolving component, unless coarse
aluminum oxide- or zircon-containing ingredients are used. Nearly 90% of the
total time required for complete melting ("batch-free time") is determined by
the dissolution process of quartz sand in the melt phases. Hrma [2.56] quali-
tatively described the kinetics and mechanisms of the batch-melting process.
Beerkens [2.13] presented a mathematical description of the sand-dissolution
or reaction process in the batch blanket and in the glass-melt phases. Ex-
amples [2.13] have been given of the modeling of sand-grain dissolution in
industrial melting tanks by post-process modeling of sand-grain dissolution
after the calculation of the glass-melt temperatures and glass-melt flows by
CFD modeling.

Sand-Grain Dissolution Rate

In this section, we discuss the dissolution rate of quartz sand in the molten
glass or primary melts. As will become clear in the following, the dissolution
rate strongly depends on the temperature and the kind of melt phases present.
At 800°C, initially alkali-rich or alkali-earth-rich aggressive melt phases with
a low Si0 2 content are being formed. Then, the dissolution rate of sand grains
is relatively high. In the final glass melt the Si0 2 content is higher, the driving
force for dissolution of sand in the melt becomes smaller and therefore the
dissolution rate becomes much lower.
Batch components containing high concentrations of mineral "impurities"
or compounds that are very difficult to melt or dissolve (e.g., particles con-
taining chromite, or zircon or aluminum oxide) may considerably increase the
melting time required.
The dissolution rates of single silica grains are diffusion controlled. The
Si0 2 at the grain surface dissolved in the surrounding melt has to be trans-
2.1 Sand-Grain Dissolution, Behavior of Gas Bubbles, Glass-Quality Index 49

ferred into the melt by diffusion. Therefore, the diffusion of Si0 2 within the
melt is very important. The driving force for the diffusion process is the dif-
ference between the equilibrium concentration of Si0 2 in the molten glass
(maximum Si0 2 solubility in the glass melt) at the interface of the silica
grain, Ge , and the Si0 2 concentration in the bulk of melt, Gs (see Fig. 2.11).
The value of Ge for soda-lime silica glass (typical for float glass or
container-glass types) is of the order of 85 wt% at the melting temperature.
For most soda-lime glasses the Si0 2 content, Gs , is 71-74wt%. In this case,
the driving force for Si0 2 dissolution (or sand-grain dissolution) is a concen-
tration difference of about 11-14 wt%. Within the batch blanket, when most
of the Si0 2 has not yet been dissolved, the concentration difference can be
substantially higher (up to about 50%).
For a spherically shaped sand grain, a relation describing the decrease of
its radius can be derived:
dR
PSi0 2 ill = h (Ge - Gs ) Ps , (2.15)

where R is the radius of the sand grain (m), t is time (s), Ge is the concen-
tration of Si0 2 at the interface of the sand grain (wt fraction), Gs is the Si0 2
concentration in the melt phase (wt fraction), Ps and PSi0 2 are the respective
densities of the melt phase and quartz (kg/m 3 ), and h is the mass transfer
coefficient (m/ s).
Outside the batch blanket section, the weight fraction Gs in the glass melt
almost equals the Si0 2 content of the final glass composition. A high Si0 2
concentration also increases viscosity and slows down convection (mass trans-

[Si02] Quartz
wt%

Reaction Si02 with molten glass


Ce = concentration Si0 2
at interface
Si0 2 diffusion

/
Cs = concentration Si0 2 in melt
(about 72 wt% in most
soda-lime silicates)

Dissolution of Si0 2

Ce is maximum Si02 solubility


to (depends on glass composition
and temperature)

Fig. 2.11. Scheme of the dissolution process of an Si02 grain (assumed to be


spherical) in molten soda-lime silica glass
50 2. Melting and Fining

fer coefficient h depends strongly on convection, which determines the con-


centration boundary layer thickness) and the diffusion process. During short
residence times in the batch blanket (20-40 min), sand therefore dissolves
faster in a more aggressive and silica-poor melt (primary melting phases)
than within the bulk of the Si0 2 -rich glass melt, despite the lower tempera-
tures in the batch.
The value of C e also depends on the composition of the melt phase and
on the temperature (it normally increases with T). Ce is determined by ther-
modynamics and can also be measured experimentally. The mass transfer
coefficient h depends on:
• the diffusion coefficient D,
• the grain radius R,
• the time (in view of the creation of a time-dependent Si0 2 concentration
profile in the molten glass surrounding the sand grain),
• the local convection flow of the melt, relative to the sand grain (glass flow-
velocity gradients due to forced convection) in the melt, and
• the free convection within the boundary layer around the quartz grain,
caused by the different densities of sand grain and glass melt and by the
density differences in the vicinity of the sand grain due to the Si0 2 con-
centration profiles [2.66] in the melt.
The Si0 2 mass transfer coefficient h is given by:

(2.16)

where D is the diffusion coefficient of Si0 2 in the molten glass, based on


concentration profiles given in wt fraction (D in m 2 S-l); R is the grain
radius (m); and t is time (s). Sh is the Sherwood number for mass transfer
from spherical grain (see reference [2.13]):
1/3
Sh>':::i 2 + 0.89 [ Re Se + (Gr Se)3/4 ] (2.16a)

= 2 (no convection)
= !(R 2 / 3, D- 1/ 3, (grad v)1/3) (convection flow of the glass melt)
v = flow velocity of the melt relative to the sand grain (m/s)
= !(R3 / 4 , 7]-1/4, D- 1 / 4 ) (free convection of surrounding melt
relative to the sand grain)
7] = viscosity (Pa s) .
Re is the Reynolds number, based On the velocity (vd of the melt relative
to the velocity of the sand grain, depending On the local velocity gradient
(dv/dR) of the melt and buoyancy of the lighter sand grain:
2.1 Sand-Grain Dissolution, Behavior of Gas Bubbles, Glass-Quality Index 51

Re = ±2Rpvt/T/ (2.16aa)
p = density of the molten glass (kg/m3)
Vt = velocity of the melt relative to the sand grain
depending on dv/dR and VS'

For very strong convection:


Vt = Rdv/dR orRJJgradvJJ (JJgradvJJ = absolute value of the

velocity gradient vector in the three directions in the melt)


Vs is the sand-grain velocity relative to melt due to buoyancy,

(2.16ab)

(For a strong forced convection, Vs can be neglected in the Reynolds relation.)


Se is the Schmidt number (= T//(D p)). G, is the Grashof number for free
convection caused by the rising sand grain.

G, = ±8g p~pR3 /T/ 2 (2.16ac)


p = density of molten glass (kg/m3)
~p = density difference between melt and sand grain (kg/m3) .
For small particles dissolving slowly, it can be assumed that R « V7r D t, so
relation (2.16) can be simplified to

(2.17)

Generally, there will be a combination of different types of convection: free


convection (particularly the coarser particles rise within the melt, silica often
being lighter in density than the glass melt) and forced convection, depending
on the position in the glass-melting tank. The value of the Sherwood number
lies between 2 and 20. Then, the dissolution time is almost proportional to
the sand-grain radius to the power 1.25-2 and to the reciprocal value of the
diffusion coefficient of the Si0 2 in the glass melt to the power 0.667-1.

Gas Exchange Between Bubble and Glass Melt

After the fusion or dissolution of most batch components, the glass melt
contains dissolved gases and gas bubbles in sizes varying from 0.03 mm up to
several mm. The concentration of dissolved gases in the fresh melt is relatively
high. The most important gases in the melt or in the gas bubbles are nitrogen,
CO 2 , water vapor, oxygen, S02, argon, and sometimes NO or CO.
The bubbles can be removed by two different processes: bubble ascen-
sion to the glass-melt surface, where the bubbles collapse, or complete re-
absorption of the gases by the melt.
52 2. Melting and Fining

The rising velocity v of a bubble is given by the formula derived from


Stokes' law:

(2.18)

where v is the ascension velocity (relative to the melt) (m/s); R is the bubble
radius (m); p is the density of the glass melt (kg/m 3 ); g is the acceleration of
gravity (9.8m/s2), 71 is the viscosity of the melt (Pas); and c is a factor with
a value between 2/9 and 1/3. If the bubble surface can be considered to be
rigid, then c = 2/9 (Stokes), whereas for a completely mobile bubble surface
c = 1/3 (Hadamard-Rybczynski).
The bubble ascension is proportional to R2 and 1/71. A small seed hardly
ascends in the melt, and the seed neither grows nor shrinks when an equilib-
rium exists between the partial vapor pressures of the gases in the bubble and
the same gases dissolved in the molten glass. Then, there is an equilibrium
for each gas species i:

(2.19)

where Li is the Henry solubility of gas i in the melt at 1 Pa vapor pressure


(mol/ (m 3 Pa)), C gi is the concentration of the dissolved gas i in the melt
(mol m- 3 ), and Pi is the vapor pressure of gas i in equilibrium with the
dissolved gas i (Pa). (For gases dissociating in the melt, Pi = Li/C~, where n
is the number of ions in the melt formed from one gas molecule. For example,
for H 20 dissolving in the melt, 20H ions are formed.)
The gas can dissolve chemically, by reacting with glass-melt components,
or physically, by occupying open sites in the structure of the melt without
chemical bonding. In general, the chemical solubility is orders of magnitude
higher than the physical solubility of a gas. However, the chemical solubil-
ity is much more temperature dependent. Special gases, the so-called fining
gases, dissolve chemically at low temperatures. Examples are S02 gas or
oxygen. A fining agent dissolved in the molten glass releases these gases at
increasing temperatures. Sulfates produce S02 and oxygen gas, and polyva-
lent ions in the most oxidized state are reduced by temperature increments,
releasing oxygen gas. A temperature increase in a melt containing a fining
agent increases the equilibrium pressure of the fining gases: The equilibrium
of the fining gas in the bubble and in the melt is disturbed, and the fining
gas diffuses into the bubbles. As a consequence, these bubbles grow, and the
original gases in the bubble become diluted by the fining gas in the melt.
This process disturbs the gas-phase-melt-phase equilibrium of the other gas
species, which now start diffusing into the growing bubbles as well. Because
bubble ascension depends on the bubble size (v cv R2), bubble rise acceler-
ates during the fining process. Thus, the bubble ascension rate increases with
increasing temperature, and the glass melt is more readily stripped of the
dissolved gases.
The total bubble pressure Pt is given by:
2.1 Sand-Grain Dissolution, Behavior of Gas Bubbles, Glass-Quality Index 53

Pt = 2:Pi = Po + P g H + 20"1 R , (2.20)

where Po is the atmospheric pressure (Pa), p is the density of the melt


(kg m -3), 0" is the surface tension (N m -1), and H is the height of melt
above the bubble (m). The bubble pressure decreases as the bubble grows
(R increases) and ascends in the melt (H decreases). Above a certain tem-
perature level, the total equilibrium pressure of the fining gases may exceed
the value of Pt. Then the bubble grows continuously because an equilibrium
between the gases dissolved in the melt and in the bubble cannot be reached.
At this temperature level, the removal of gases becomes very efficient. The
temperature is called the fining onset temperature Tonset.
H the concentration of a dissolved gas is not in equilibrium with a bubble,
an exchange of the gas (gas transport) takes place between the gas bubble
and the melt. Different authors presented models describing the kinetics of
the gas exchange between gas bubbles and glass melt [2.67-72]. The bubble
may grow or shrink by multi-gas species diffusion. This is described by the
following relation:

dR Sh i Di
-
dt =
I 2 Pt g T(C·SI - C·)
2:·--R II
(2.21 )

where Rg is the gas constant (= 8.31432J/(moIK)), t is time (s), Di is the


diffusion coefficient of gas i in the molten glass (m 2 s-l), C ii is the concen-
tration of gas i in the melt in equilibrium with the vapor pressure of gas i in
the bubble (molm- 3 ).

C ii = Li Pbi ( or C ii = Li JPhl for i = water vapor) , (2.21a)

where Li is the solubility of gas i in the molten glass (moll (m 3 Pa)), Pi is the
partial vapor pressure of gas i within the bubble (Pa), and Pbi is the partial
pressure of gas i in the bubble (Pa).
C si is the concentration of gas i in the glass melt (mol m -3), Sh i is the
Sherwood number. This Sherwood number takes into account the effect of
convection of the molten glass (relative to the ascending gas bubble) on the
mass transfer rate of gas into or from the bubble. The convection is caused
by the flow pattern of the glass melt itself (there are velocity gradients in the
melt), and by bubble ascension due to buoyancy. The difference between the
average flow velocity of the bubble and the surrounding melt determines the
value of Sh i .
H, for most gases C si > C ii is valid, we have bubble growth, whereas with
C si < C ii we have bubble shrinkage (re-absorption). At high temperatures,
the fining-gas solubility in the melt decreases (the value of C si for the fining
gas i increases), and the bubbles start growing. During cooling, the solubility
of the fining gases increases again and the value of C si for the fining gases
54 2. Melting and Fining

may become much smaller than C ii . The slow-cooling stage of the melt to re-
absorb fining gases or other gases, which can dissolve in the melt at decreasing
temperatures, is called the refining process.
The calculation of gas exchange between the bubbles and the melt and
bubble ascension rate by (2.18) and (2.21) requires initial conditions:
• the initial bubble composition and size,
• the position of the bubble in the melt,
• the initial concentration level of dissolved gases, and
• the concentration of the fining agent and the redox state (the redox state
often determines the decomposition equilibrium of the fining reaction).
Very important in modeling bubble behavior in glass melts are the diffusion
coefficient (Dd, the solubility (L i ), and the fining reaction equilibrium con-
stants. These values often depend strongly on the temperature and the glass
composition. Unfortunately, accurate values of these properties are missing
for most glass types.
Figure 2.12 shows the effect of the concentration of sulfate as a fining agent
on the predicted bubble growth and bubble position of a CO 2 -containing
bubble with an original diameter of 0.2 mm initially positioned at a depth of
1 m in a static soda-lime silica melt (containing dissolved N 2 , H 2 0, CO 2 , O 2 ,
and 80 2 ).
In a glass-melt tank, the bubbles follow a path from the position where
these bubbles are formed to the outlet of the tank or the surface of the melt.
This path differs from the glass-melt flow pattern due to the buoyancy forces
acting on the bubble. The concentration of dissolved gases in the glass melt
and the redox state are both location dependent.
The concentration change of a dissolved gas species i in the melt is deter-
mined by the convection and diffusion process descriptions:

arg;;) = _ div(Csiv)+ div(Di grad CsJ + Qi ,


I I I (2.22)
convection diffusion reaction

1.6...---------------,
E14 ~ / 1.4 - - - 80 3 = 0.1 wt%
E . .§ 1.2 - - - - 80 3 = 0.15 wt%
--12
.... . ., ,/ ' / 1l - - - - 80 3 = 0.2 wt%
"*E 1.0 ' ,
,.-,....'" E
.s
1.0
.' I
• I

III 0.8 • , 0.8 .' I


'5 0.6 ."'.-'/' 1:: 0.6 • I

"I
~ 0.4
.c a2
...·F - - - 803 = O.lwt%
- - - - 80 3 = 0.15 wt%
·~0.4
,'/
.' /

~ . - - - - 803 = 0.2 wt%


I 0.2
.,."'"
O+----.--~T_--~ o+-~~._--.--_,--~
a 5000 10 000 15 000 a 5000 1a 000 15 000 20 000
Time/s Time/s
Fig. 2.12. (a) Growth and (b) ascension of a CO 2 -containing bubble during sulfate
fining of a mildly reduced soda-lime silicate melt at 1500°C
2.1 Sand-Grain Dissolution, Behavior of Gas Bubbles, Glass-Quality Index 55

where Gsi is the local concentration of gas i in the melt (mol m -3), Di is the
diffusion coefficient (m 2 S-1), Qi is the formation rate by chemical reaction
(mol/(m 3 s)), and v is the glass-melt velocity vector (vector components in
m/s).
The term Qi is very important for oxygen gas or other fining gases. Gas
is released by redox or fining reactions, and this release increases the concen-
tration of the physically dissolved fining gas (Gsi ). For instance, a temper-
ature increase will shift the fining reaction equilibrium, which leads to gas
(0 2 and/or S02) production. This gas diffuses into the seeds or bubbles and
makes them grow. Redox reactions in the melt, for instance with the ion pairs
Fe2+ /Fe3+, Cr3+ /Cr2+, Cr3+ /Cr6+, Sb3 + /Sb5+ or the ions S6+ /S4+ /S2-,
will lead to gas formation. In the last case, physically dissolved S02 (sulfur
in S4+ state) can be formed in the melt [2.36].
An example is the sulfate fining: SO~- {=? 0 2- + 1/2 O 2 (gas) +S02 (gas).
The equilibrium constant is given by:

(2.23)

GS0 2 , G 02 , and G S0 2- are the concentrations (mol m -3) of S02, physically


4
dissolved O 2, and the sulfate dissolved in the melt. C 0 2- is the oxygen con-
centration, or, more precisely, the oxygen ion activity of the silicate melt.
The equilibrium constants for the other redox reactions can be given in a
similar way. If various polyvalent ions are present in the melt, the fast kinet-
ics of the redox reactions at high temperatures almost always create a local
equilibrium, depending on the temperature between the dissolved oxygen (or
oxygen equilibrium pressure) and all polyvalent species. A change in temper-
ature will shift the redox equilibria, and oxygen or sulfur dioxide is consumed
or produced. The production of gas from such redox reactions can be calcu-
lated in each grid cell if the reaction equilibria as functions of temperature
are known. Generally, these models assume that the oxygen concentration in
the melt, related to the oxygen partial vapor pressure (p02 = G02 /L 02 ), is
in equilibrium with the polyvalent ion species and that the reaction equilib-
rium, such as that given in (2.23), is also valid at each position in the melt.
Calculation of the oxygen or fining-gas concentration in the melt requires
that the temperature-dependent equilibrium constants are known or can be
derived from thermodynamics or via analytical techniques, for instance with
square-wave voltammetry (described by Russel and Freude [2.73]).
The local concentrations of all polyvalent ions in the melt or the oxygen
partial equilibrium pressure are all in chemical equilibrium according to the
reaction equations given by relation (2.23) or by similar relations. From the
reaction equilibria, the local (physically dissolved) oxygen and sulfur oxide
concentrations can be calculated. Solution of (2.22) gives the local concen-
tration of the relevant gas species in the melt including the concentrations
of the gases formed or absorbed by redox reactions. The term Qi in (2.22) is
equal to zero for gases that dissolve only physically. Figure 2.13 shows, as an
56 2. Melting and Fining

a)

o 2 X 10-6 6 x 10-6 8 x 10- 6 1 x 10- 5


O2 (redox) fmol m- 3

b)

o 2 x 10- 2 4 X 10- 2 6 x 10- 2 8 X 10- 2 1 x 10- 1

Fig. 2.13. Distribution of concentrations of dissolved gases in a glass melt. A tank


cross section in the center axis (x-z plane) is shown. (a) physically dissolved gas
O 2 , (b) gas CO 2

example, the concentrations of dissolved oxygen and CO 2 in a cross section


of a container-glass-melt tank with vertical electrodes, deep refiner, and weir.
The stripping of CO 2 from the melt can be clearly observed.
During the travel of the bubble or seed along a path in the tank with
the calculated local gas concentrations, the gas exchange between the bubble
and the surrounding melt can be determined with (2.21) or as described by
Kramer [2.38]' the bubble size and the bubble composition as described by
Beerkens [2.72]' and the buoyancy velocity with (2.18). The trajectory and
changes in contents and size can be determined from the initial position of
the bubble in the tank till the glass-melt surface is reached or, in the case of
poor fining, till the bubble ends up in the glass fed to the forming process.
Figure 2.14 shows the calculated trajectory of a small seed formed at the
glass-melt-batch blanket interface. The figure also shows the change in seed
size and the oxygen concentration in the seed. Note the increase in bubble
size in the fining zone, where extra oxygen (fining gas) is released.
2.1 Sand-Grain Dissolution, Behavior of Gas Bubbles, Glass-Quality Index 57

Bubbling section

a)

Top view

1.8 X 10-4 2.4 x 10-4 3.0 x 10-4 3.6 x 10-4 4.2 x 10-4 4.8 x 10-4
Diameter !Trace

Cross
section

1.8xl0-4 2.4x10-4 3.0xl0-4 3.6xl0-4 4.2x1n-4 4.8x10-4


Diameter !Trace

Bubble reaches surlace (> 80% fining gas: 02J------,


b)

0.00 0.16 0.32 0.48 0.64 0.80


Volume fraction O 2

Fig. 2.14. (a) Trajectory of it seed (originally CO 2 ) in a TV-panel glass-melt tank


with forced bubbling; the size of the bubble during its trajectory is indicated by
the color. (b) Oxygen concentration change in the seed/bubble during its passage
through the glass-melt tank

Figure 2.14a,b shows that the bubble changes in size and composition. In
the top layers, the traced bubble contains mainly oxygen, which is liberated
at high temperatures (> 1350°C) by the antimony oxide fining agent. Fig-
ure 2.14a shows that the seed entrapped in the glass-melt flow from the cold
bottom hardly grows in the bubbling section even close to the glass-melt sur-
face. The surface temperature of the glass melt at this position is relatively
58 2. Melting and Fining

low. The bubble finally escapes from the melt with a diameter of about 5 mm,
containing about 80% oxygen. Bubble size and composition depend strongly
on the redox state and the contents of the fining agent.

Modeling of Single Sand-Grain Dissolution or Bubble Behavior


in Industrial Glass-Melting Tanks

The dissolution of sand grains in the batch-blanket section or in the glass


melt can be calculated, after CFD modeling, by mathematically following
the sand grain during its path in the melt tank. At each position in the tank,
the temperature, velocity direction, and velocity value are known from the
CFD results. The velocity gradients in the melt surrounding the traveling
sand grain can also be derived from the main CFD model. This information
(Ilgrad v II) and data on the initial sand-grain size, Si0 2 diffusion coefficient,
and Si0 2 solubility in the melt allow the calculation of the local sand-grain
dissolution rate by (2.15)-(2.16ac).
Following the path and dissolution process of a large number of sand
grains in the melt yields information about the melting behavior. In each
volume element, the size of the largest sand grain can be determined. In
the grid cells of a glass-melt tank with a maximum sand-grain size equal
to zero, all sand is assumed to be dissolved. The absence of non-dissolved
sand grains (or other batch grains) in the neck or throat at the tank outlet
is very important for the glass quality. However, often the positions with
no sand grains should be situated much more upstream in the tank. The
course of the fining process is only effective if almost all sand grains are
completely dissolved prior to fining. During sand-grain dissolution new very
small seeds are formed and fining can only be effective when the formation
of new seeds can be excluded. Thus, this requirement (at least for a glass
product with a high quality) needs complete melting of all grains in the
section of the tank before reaching the spring zone. Figure 2.15 shows the
calculated maximum sand-grain size for a container-glass furnace in a cross
section of the tank. As can be seen, most sand grains are already dissolved
in the tank section underneath the batch blanket, and in the hot-spot zone
dissolution is complete.
Melting Factor and Melting Index. Different approaches are used to charac-
terize the melting performance of a glass furnace. The melting performance
depends on the trajectories the material is following in the tank. A path in the
melt tank involving high temperature, long residence time, and high glass-
melt velocity gradients yields better melting conditions than a path with short
residence time, almost static conditions, and low temperature such as taken
by shortcut flow from the doghouse along the tank bottom to the throat. For
each path, a melting factor can be defined. A frequently applied approach
is the integration of temperature divided by viscosity over the time along a
path from the doghouse till the throat, for a certain path i:
2.1 Sand-Grain Dissolution, Behavior of Gas Bubbles, Glass-Quality Index 59

Hot spot Throat


.n----~_==

a 200 400 600 800 1000


Maximum sand-grain size /~m

Fig. 2.15. Example of the maximum residual sand-grain size in the glass-melt
tank at different locations. The cross section at the center axis of the tank in the
length-height direction is shown

F1i = j t end T
-dt, (2.24)
t=O T]

where tend is the total residence time of the "particle" in the tank. However,
this expression has hardly any physical meaning and is not dimensionless.
Figure 2.16 shows a typical melting-factor distribution curve. The first peak
at a value of about l.5 million represents a fraction of glass melt with a
relatively low melting factor, meaning that this fraction might be critical for
the melting quality of the glass furnace. On the other hand, the absolute
value of the melting factor is not directly decisive for the absolute melting
performance of the tank.
A better approach for a dimensionless melting index takes into account
the diffusion of Si0 2 from an average-sized sand grain using the expressions
given by (2.16)-(2.16ac) and (2.17). The effect of the diffusion coefficient and
convection of the melt are incorporated in the value of the so-called melting

14.----------------------------------,

12
c
.Q 10
:;
.0

~ 8
i5
~ 6
.Q
t;
~ 4

O +L----.-----.-----.-----.-----.---~
o 5 10 15 20 25 30
Melting factor / 106

Fig. 2.16. Distribution of the melting factor (definition by (2.24» for a melting
tank
60 2. Melting and Fining

index Mh

M Ii = j t=t
t=O
end 0.756.Cs ·o Ilgradv 11 1 / 3 D2(3
1 2

(PSi0 2 I p) R4/3
S,02 dt
'
(2.25)

where 6.CS i 0 2 is the difference between the concentration of Si0 2 in the melt
at the interface of a sand grain and Si0 2 in the bulk of the glass melt (weight
fraction), Ilgradvil is the absolute value of the velocity gradient (s-l or m/s
per s), D Si02 is the diffusion coefficient of Si0 2 in the glass melt (m 2 S-l),
tend is the residence time in the tank (s), R is the radius of the largest sand
grain (s), PSi0 2 is the density of the sand grain (kg/m 3 ), and P is the density
of the molten glass (kg/m 3 ).
If the value of the melting index is » 1 for all possible trajectories, all sand
grains will dissolve in the melt with high probability. The truly dimensionless
melting index given with (2.25) has a value of 1 for sand grains just being
dissolved when reaching the throat, and value k for sand grains apparently
being dissolved k times during their passage in the tank. The integration
takes place along the path a small volume of the melt is following in the
melting tank: from the charging end of the batch in the doghouse (t = 0)
down to the throat or canal of the tank (tend = residence time). Because
there are many different paths in the melting tank, a large number of different
melting-index numbers will be calculated. From these calculations, a melting-
index distribution can be derived. The minimum value of the melting index
is characteristic for the flow pattern yielding worst glass-melt quality. Glass-
melt portions following this critical path can spoil the glass quality. Therefore,
process conditions guaranteeing high minimum melting-index values must be
searched for. Figure 2.16 shows that a small fraction of the glass melt has
a relatively small melting factor (2 x 106 ); this part of the melt will be
responsible for thc final melting quality of the glass.
Fining Index. Bubble growth and bubble shrinkage, bubble trajectories in
the melt tank and changes in the bubble composition can be calculated by
following individual bubbles in the melt. The concentration field of dissolved
gases in the melt and the contents of the fining agent in the melt tank at
each position must be calculated first. For a first indication of the fining
performance a so-called fining index can be defined and an expression for
this fining index can be derived from (2.21) and from the relations for the
mass transfer coefficient h. Similar to the melting index, the fining index and
the fining-index distribution, depending on the trajectory in the melt tank,
can be calculated. The fining index (F Ii) for a path i is then given by:

._jt=t
FI~ -
end (6.CICb )2/3 Di/ 9p5/9 g 5/9
1/3 dt , (2.26)
t=O 57]5/9 H finer

(assuming the Stokes relation describing the ascension of a single bubble)


where Hfiner is the depth in the fining zone of the furnace (m); Dg is the
2.1 Evaporation and Superstructure Refractory Attack by Vapors 61

diffusion coefficient of the most critical gas (or gas determining gas-bubble
growth) in the molten glass (m 2 s- 1 ); C b is the actual (critical) gas con-
centration in the glass melt (mol m -3); 6.C is the concentration difference
of the critical gas component dissolved in the melt and the concentration
of this gas dissolved in the melt in equilibrium with this gas species in the
bubbles (molm- 3 ); g is the acceleration due to gravity (ms- 2).
In fact, the driving force for gas diffusion into the growing bubbles, 6.C,
depends on the production of gases by redox or fining reactions. The fining
index can be calculated after the fining-gas concentrations in all grid cells are
determined. A fining index> 1 means that, in all probability, all the bubbles
have been removed. For values < 1, severe fining problems must be expected.
For such calculations, the flow path of a small glass-melt volume in the
furnace is being followed. Along this path, the integration of the indicated
quantities is carried out over the residence time. These formulae contain
temperature-, composition-, substance-, or raw material-dependent quanti-
ties (R, 6.C, C b ), glass-melt properties (p, 77), and furnace characteristics
(H, Ilgradvll). Unfortunately, bubbles do not follow the same path as a glass-
melt volume element. Therefore, this approach is not very accurate, but it
is still useful for comparing fining-performance trends after changes in pro-
cess conditions or furnace design. In another method, the individual seeds or
bubbles are followed mathematically, starting, for instance, at the interface
batch blanket-glass melt, as described in Sect. 2.1.3. This method has the
drawback that the initial bubble sizes as well as the population are often
unknown. Some information can be obtained from samples taken from glass
furnaces [2.74].

2.1.4 Models for Evaporation and


Superstructure Refractory Attack by Vapors

Depending on the glass composition and on batch contamination, the most


important components evaporating from the glass-melt surface are alkali com-
pounds (e.g., NaOH, KOH, Na, K, NaCI), heavy metals (e.g., lead), boron
species (e.g., HB02, NaB0 2, KB0 2), chlorides, fluorides, and sulfur com-
pounds. S02 is mainly released by the bubbles formed during fining of the
melt. Lead and alkali vapor may react - in particular at relatively low temper-
atures « 1370-1450°C) with silica-containing refractory materials in the
_0

combustion chamber or in the regenerator, forming low-viscous alkali or lead


silicates. The width of the temperature window where this attack takes place
increases for increasing concentrations of these vapors. This has recently been
observed in oxygen-fired glass furnaces [2.75]. The evaporation leads to emis-
sions of metals or dust and fouling or plugging of the regenerators or flue-gas
channels during the cooling of the flue gases, due to the condensation of salts
formed from the evaporated material.
The evaporation rates at the glass-melt surface depend on the gas flow
rates in the combustion chambers, particularly in the vicinity of the glass-
62 2. Melting and Fining

melt surface, on the temperature at the glass surface, the composition of the
glass melt, and the composition of the furnace atmosphere. When reducing-
gas components, such as CO or hydrocarbons, touch the glass-melt surface,
evaporation of alkali or lead is enhanced due to reduction of sodium, potas-
sium, or lead oxides at the glass-melt surface. In static melts, or rather static-
melt volumes, exposed to the furnace atmosphere, diffusion processes bring
the volatile glass components to the surface. At high evaporation rates, the
diffusion process in the melt is too slow to compensate for all evaporation
losses, and depletion of volatile glass components at the glass surface takes
place. This causes a change in surface viscosity, density, and composition.
The evaporation kinetics are determined by mass-transfer processes in
the melt and in the gas-phase boundary layer above the melt as well as by
the thermodynamics of the evaporation process at the glass-melt-furnace
atmosphere interface.

Evaporation Modeling

Assuming that a glass-melt volume element is flowing along the surface ex-
posed to the furnace atmosphere and that convection within this volume
element is limited, the evaporation rate can be estimated from the mass-
transfer rate of the volatile component to the surface and at the glass-melt
surface, expressed as molar flux Qm,j (mol m- 2 S-I). For the diffusing volatile
component j in a static melt this diffusion rate to the surface at x = 0 (x is
coordinate with direction perpendicular to melt surfaced) is given by

(2.27)

The concentration profile (often extending in a top layer of 100-5000 fJ.m) of


component j in the melt can be calculated, for diffusion only, by solving the
differential equation and boundary and initial conditions:

(2.28)

The boundary and initial conditions, assuming an infinite deep melt are:

t=O -oo<x<O - Cj
0J -
at bu1k

at t>O x-t-oo - Cj
0J - bu1k

at t>O x=O OJ = qurface(t) = OJ,x=o(t) . (2.29)

The interdiffusion coefficient Dm,j is assumed to be constant during the time


of the evaporation process and within the diffusion distance. However, this
assumption may not be correct for all cases [2.76-78]. For compositions within
a relatively narrow range, Dm,j probably changes only moderately.
At the surface, the glass-melt component j (for instance Na20 or PbO) is
evaporating or reacting [2.79,80], forming volatile species i (Na, NaOH, Pb
2.1 Evaporation and Superstructure Refractory Attack by Vapors 63

or PbO). crlk is the concentration of component j in the bulk, for instance


given in mol m -3, and Cjurface (t) is the time-dependent surface concentra-
tion of component j. The concentration profile in the melt and the slope
(8Cj j8x)x=o can be determined by the mathematical solution of (2.28) and
the boundary conditions given in (2.29). However, Crrface(t) also depends
on the mass transfer of the evaporated species in the gas phase. For values of
Dm,j in the range of 10- 9 _10- 12 m 2 S-l and static top layers of > 3-5 mm, the
assumption of a semi-infinite layer is mathematically correct. At the glass-
melt surface, component j in the melt can be transformed by evaporation or
reaction with component k into gaseous component i: nj + mk -+ q i. For n
moles of j, q moles of gas i are formed.
The molar flux of the volatile i formed and transported from the surface
of the melt into the main gas stream (R g = gas constant) is given by:

(2.30)

pt is the vapor pressure of component i formed from component j in the melt


in equilibrium with Crrface(t). The average mass transfer coefficient hg,i for
mass transport through the boundary layer at the gas-phase side at position
y (distance) downstream from the leading edge (at which the gas flow starts
contacting the glass-melt surface) can be estimated by Sherwood or Nusselt
relations [2.81-83] and by:

(2.31 )

Shg,i is the Sherwood number for local mass transfer [2.81] of component i
in the gas phase. The subscript g refers to the gas phase.
Under laminar-flow conditions (Reynolds number < 3 x 105 and no ob-
stacles in the gas flow) and Schmidt numbers > 1 (this condition is almost
always met in the case of evaporation of glass-melt components into the com-
bustion chamber), the hydrodynamic boundary layer thickness exceeds the
concentration boundary layer thickness [2.84]. Then, the average Sherwood
number over a length Lg (downstream from the leading edge: leading edge is
the position where gas flow starts at the surface of the melt) can be calcu-
lated:

Shay'
g,l
= 0.662 Re O.5 SCO.
g,L g g,l
333 (2.32)

for laminar gas flow parallel to the melt surface with:

Reg,y = Vg Y Pgj'T}g (Reg,L g = Vg Lg pC;: j'T}g) (Reynolds number), (2.33)

where Y is the distance from the leading edge of the gas flow touching the
melt surface (m), and

SCg,i = 'T}gj(Pg Dg,i) (Schmidt number) (2.34)


64 2. Melting and Fining

Cable and Fernandes [2.85] showed that for parallel flow, at larger gas
velocities and at turbulent flow conditions, the Sherwood numbers are pro-
portional to v~·8, in agreement with Schlichting's analysis [2.82] for turbulent
flows.
For turbulent conditions, at high Reynolds number (> 3 X 10 5 or disturbed
flows):

Shay
g,l
(turbulent) o.8
= A Re g,L g
ScO.
g,l
333 . (2.35)

The value of A (depending on geometry and size) can be derived from [2.81~
83] and is in the range of 0.03~0.04 for turbulent gas flow (Reynolds number
> 3 x 105 or for disturbed flows) parallel to a horizontal surface.
For the flux at the glass-melt~gas-phase boundary, the continuity in mass
transfer and formation of vapor i (for instance NaOH) from glass-melt com-
ponent j (for instance Na 2 0) requires:

q moles of vapor i are formed from n moles of glass component j. If p~ is


proportional to the concentration of component j at the surface in the melt
(p~ = B q,x=o(t)) and pfulk (vapor pressure of evaporating component i in
the main gas stream) is almost zero (pfulk « pj'), the following expressions
hold for the average time-dependent evaporation rate of component i:

Qrn,j. = '!q.!. A v gO. 8 p O


g
.47 '1l~0.47 D O. 667 L ~0.2 R~IT~1 Bcsurface(t)
'tg g,l g g J . (2.37)

Parameter B (relating the equilibrium vapor pressure of i to the concentration


of the volatile component in the melt) depends on the furnace atmosphere
composition and on the chemical activity of the volatile component in the
melt. For NaOH evaporation, the B value depends also on the water vapor
pressure in the furnace atmosphere, B rv pH 2 00. 5 . The values of D g, Pg, and
TJg are derived for the average gas temperature in the boundary layer.
Equation (2.37) shows that the influence of length Lg of the gas flow along
the glass-melt surface is very small because Qrn,j rv L;:0.2. The mass transfer
is only weakly dependent on distance y or Lg along the glass-melt surface.
The concentration profile or depletion of component j can also be cal-
culated. Concentration q in the melt of component j (assuming a very low
concentration of evaporation product from j in the bulk gas flow) at a distance
x from the surface after exposure time T [2.76-78,85, 86] is:

q(x) = erik { erf[x/2(Dm,j T)I/2] + [exp(k d x + k~ Dm,j T)]


(2.38)
. erfc[x/2(Dm,j T)I/2 + kd(Dm,j T)0.5]} ,
2.1 Evaporation and Superstructure Refractory Attack by Vapors 65

a = (n/q) A vo. s pO.47 ",-0.47 O.667 L -0.2 R- 1 T- 1 B


D g,l (2.3Sa)
g g ·tg g g ,

kd = a/Dm,j . (2.3Sb)
Dmj is the interdiffusion coefficient of component j in the glass melt, Di,g is
the diffusion component of the evaporated species i in the gas phase. Vapor
pressures can be derived from thermodynamic modeling [2.S7] or from esti-
mations of the Na20 activity in soda-lime silicate melts as applied by Conradt
and Scholze [2.SS] and by Argent et al. [2.S9]. From the dependency of vapor
pressure on Cj,x=o, the factor B can be estimated.
Figure 2.17a shows the calculated sodium sulfate emissions from oxygen-
gas-fired soda-lime silicate furnaces, assuming that the sodium sulfate is
formed from the NaOH evaporated from the glass-melt surface. Figure 2.17b
shows the dependency of gas flow and temperature on sodium sulfate emis-
sions caused by NaOH evaporation for an end-port air-fired furnace. The
strong influence of glass-melt surface temperature and gas velocity on the
evaporation of sodium and, consequently, on the emission of sodium sulfate
dust becomes quite obvious from Fig. 2.17. The diagram also shows the ef-
fect of turbulent flow conditions on evaporation and emission: turbulent flows
increase the emission levels. The data given in the two diagrams must be
compared with care. Although the evaporation and the sodium sulfate lev-
els at the same gas velocity and gla8s-melt temperature are much higher for
oxygen-fired furnaces (due to higher pH 20 level) than for air-fired furnaces,
the gas flow above a melt can be drastically reduced when changing from air
to oxygen firing. The conditions of relevance (temperature8 and gas veloci-
ties) are shown as in Fig. 2.17. Reducing the local gas velocities by a factor
of 5-S is possible by conversion to oxygen firing; this can lead to considerable
reductions in specific dust emissions.
The calculated specific sodium sulfate emission is 0.160 kg/ton molten
soda-lime glass (Na20 = 13wt%, MgO + CaO = 14wt%) for an end-port
furnace fired with air-natural gas with a gas velocity of 10-15 m/s and an
average flame length of 12 m, assuming 65% free surface and a glass-melt
surface temperature of 1530 ac.This calculated value is close to the val-
ues measured in similar container-glass furnaces, which lie between 0.15 and
0.25 kg/ton molten glass. Measured levels in the glass industry are 0.200-
0.300 kg Na2S04 emis8ion per ton molten glass from air-fired soda-lime sili-
cate furnaces and about 0.12-0.16 kg sodium sulfate emissions per ton glass
from oxygen-fired soda-lime silicate furnaces (vg = 2-5 m/s). These values are
in good agreement with the values predicted from the previously described
model. The model shows that reducing the gas velocities by 30% and decreas-
ing the maximum glas8-surface temperatures by 25 ac,
for example by burner
modifications, will reduce the sodium sulfate dust emissions of container- or
float-glass furnaces by 35-40%.
The evaporation of volatile glass-melt species such as sodium, potassium,
fluorides, lead, and boron may lead to depletion of these components at the
glass-melt surface. The depleted layer may have different properties from
66 2. Melting and Fining

1.4 a)
turbo1500 · C Oxy-gas fired
en
II/) 1.2 tUrbo1550 · C soda-lime glass
O.!!! turbo1600 · C 13wt% Na02
ClIOl
z
.. ..
C
1.0 laminar 1500 · C
EEl .""".
ou laminar 1550 · C

.. ..
-=5 0.8 laminar 1600 · C ,/
C QI
.2 E

ji$
.~~ 0.6 Relevant for /" "

0'yg,,;:, .: .. ~. _.;.:._~.:...- --
E-
.,..-
QlC
0
0<0 04
(/)0 ~ ."_0. . _.~-::.._ _- - - - -
"'a.
ClIClI
Z1» 0.2 ~.-
--- - -- -- -----------------
0
0 5 10 15 20
Gas velocity above free melt 1m s- 1

-.
0.8 b)
• turbo 1500 · C
en Air-gas fired

----
I I/) turbo1550 · C
O.!!!
" " soda-lime glass
eaOl
0.6 " turbo 1600 · C 13wt% Na02
Z C
EEl
e ·2
laminar 1500 · C
- - laminar 1550 · C Relevant for air firing
,
-!::
c QI - - - - laminar 1600 · C
.2 E 0.4
,.... . '" . . .
."."",.

.. ... . . . . . . . .
-- -- -
.~~ °
E-
0

QI C

..,..,
-
0
'V'';

..
OCll
(/)0 0.2
NO.
~ ...........
- --- - ---- ' .- .-.
--
ea ClI
Z> -
",'

---
Q)

~-- ---
o
o 5 10 15 20
Gas velocity above free melt 1m S- 1

Fig. 2.17. (a) Specific sodium sulfate emissions (oxygen firing) caused by evapora-
tion of NaOH from a soda-lime silicate glass melt with 13 wt% Na20, calculated by
mass-transfer models for turbulent and laminar gas flows as a function of temper-
ature and gas flow. (b). Sodium sulfate emission caused by evaporation of NaOH
from a soda-lime silicate glass melt with 13 wt% Na20 calculated by mass-transfer
models as a function of temperature and water-vapor pressure for air-fired condi-
tions, Lg = 12 m (pH 2 0 = 0.18 bar)

the parent melt and this might cause formation of high-viscous cords or a
layer with surface-tension gradients leading to extra convection in the top
layer of the melt. This is especially the case for low diffusion coefficients of
these components in the melt and for a high mass transfer coefficient for the
transport of the evaporated species from the glass-melt surface to the bulk
of the gas phase.
2.1 Evaporation and Superstructure Refractory Attack by Vapors 67

Figure 2.18 shows results of the calculated local NaOH evaporation rates
in a glass furnace. First, the temperatures at the glass-melt surface and the
flows in the combustion space have to be derived from coupled glass-tank and
combustion models. The local evaporation rates can be calculated from the
NaOH vapor pressure, and the mass-transfer relations are given by (2.36) or
(2.37).

Simulation of Refractory Attack by Alkali Vapors

Alkali vapors may react with the silica superstructure refractory. For instance,
sodium hydroxide reacts with Si0 2 or CaO·Si02 (binding phase in most silica
types):
2xNaOH + ySi0 2 ¢} xNa20·y Si0 2 + xH 2 0
2xNaOH + zCaO·Si0 2 ¢} xNa20·zCaO·zSi02 + xH 2 0 .
The sodium silicate or sodium-calcium silicate formation depends on tem-
perature and NaOH vapor pressure. Along the surface of the crown the gas
flow will form a boundary layer. In this gaseous boundary layer, a diffusion
process of NaOH to the surface of the refractory takes place. The driving
force is the difference between the vapor pressure of NaOH in the main gas
flow and the NaOH vapor pressure in equilibrium with Si0 2 and the reaction
products as given in the equations above. At relatively low crown temper-
atures, this vapor pressure in equilibrium with the phases at the refractory
surface is very low and the driving force for mass transfer is high. A high gas
velocity results in the formation of a thin boundary layer and a very steep
NaOH pressure drop over this layer thickness, which results in a high sodium
species diffusion rate.
The transport rate JNaOH (mol/m 2 s) of NaOH to the surface of the re-
fractory is given by

Glass surface

o 80 160 240 320 400


Fig. 2.18. Evaporation of NaOH from a soda-lime glass melt in an industrial glass
furnace, calculated from combustion and glass-tank modeling and using a post-
process evaporation model (concentration of NaOH vapor in vol. ppm)
68 2. Melting and Fining

J (Sh NaoH DNaOH) ( bulk * ) (2.39)


NaOH = Lg Rg T PNaOH - PNaOH ,

where PNaOH is the vapor pressure of NaOH in the bulk or at the refractory
surface (*) (Pa), T is the absolute temperature (K), ShNao H is the Sherwood
number to be derived from Bird, Stewart, and Lightfoot [2.81]' Lg is the
characteristic length (gas flow path from leading edge) (m), Rg is the gas
constant (8.31432 in J/(molK)), and DNaOH is the diffusion coefficient of
NaOH in the gas phase (m 2 S-1).
For each x mol NaOH, y mol Si0 2 or z mol CaO·Si02 reacts, forming
sodium (calcium) silicate. Below the liquidus temperature, this is a solid
reaction product; above the liquidus temperature, the silicate drips or flows
as a liquid phase downward along the refractory crown and the side walls.
The reaction product formation and composition is determined by ther-
modynamics. The thermodynamic model calculations show that liquid phases
are formed already above 1200 °C for> 200 vol. ppm NaOH in the gas phases.
Above 1480°C, no reactions are expected. The crown sections with surface
temperatures below 1480°C are vulnerable to sodium-silicate formation at
these NaOH vapor pressures, as is shown in oxygen-fired furnaces. Above
±161O °C, silica starts to deform. The temperature should be between 1400-
1610 °C. At lower vapor concentration levels in the combustion space, the
critical refractory temperature for NaOH attack shifts to lower levels; this
means that the temperature window for using the refractory without severe
attack by glass-melt vapors becomes wider.
The thermodynamics and mass-transfer relations can be used to calcu-
late the silica loss by these reactions. After calculation of the temperatures
and flows in both glass-melt tank and combustion chamber, the evapora-
tion rates of sodium species can be calculated by using the models described
in the subsection "Evaporation Modeling", p. 62 ff. The evaporated NaOH
molecules distribute over the combustion space by convection. The boundary
layer thickness at the surface of the refractory materials can be calculated
from boundary layer theories as described by Schlichting [2.82]. The mass
transfer of NaOH to the silica crown can be calculated at each position,
assuming thermodynamic equilibrium between NaOH and silica and the re-
action product formed [2.75].
Figure 2.19a shows the results gained by modeling refractory attack by
sodium vapors in an oxygen-fired soda-lime silica glass furnace at the (silica)
side walls; Fig. 2.19b shows the loss of refractory at the inner crown surface.
The attacks are most severe in the exhaust ports and at the crown in the
doghouse areas. At these locations either the gas velocities are very high or the
temperature is relatively low. At low temperatures, NaOH becomes reactive
to silica and CaO·Si02 (one of the phases present in most silica refractory
types). Above a certain temperature, NaOH remains in the vapor phases and
is weakly reactive. Such high temperatures are given in the hot-spot area.
2.1 Evaporation and Superstructure Refractory Attack by Vapors 69

Bulk concentration NaOH: 200 ppmv


Critical temperature: 1465°C
Na20 diffusion coefficient (at 1350 °C): 2 x 10-11 m2/s
Maximum crown corrosion rate: 9.7 cm/year

o 2 3 4 5 6
Refr. corr. rate fcm year- 1

Q)
en
:J <a
o
.r:
C>
e
.r:
o I-
o

o 2 3 4 5 6 7 8 9 10
Refr. corr. rate fcm year - 1

Fig. 2.19. (a) Predicted silica refractory loss (em/year) by NaOH attack in the
combustion chamber of an oxygen-fired soda-lime silica furnace. (b) Predicted silica
loss (em/year) by NaOH attack at the crown surface of an oxygen-fired soda-lime
silica furnace

Other Models

Refractory Attack in Melt Tanks. There are hardly any submodels simulating
the attack of refractory materials by the glass melt. Important work on this
subject has been performed by Briickner [2.90]. Krabel [2.91] measured and
described the corrosion rate of refractory materials exposed to glass melts.
The description enables the modeling of refractory corrosion and contami-
nation by refractory material depending on the glass composition, local flow
conditions, temperature and composition of the refractory material. There
are at least four different, often interacting, mechanisms that determine, to
70 2. Melting and Fining

a certain degree, the rate of refractory attack and contamination of the melt
by refractory components:
• dissolution and diffusion of refractory components in the melt enhanced by
convection flows [2.90];
• diffusion and penetration of glass-melt components in the refractory ma-
terials;
• electrochemical processes involving polyvalent ions in the refractory mate-
rial, leading to gas-bubble formation;
• erosion of the refractory grains by the flowing high-viscous melt.
The modeling of refractory-glass-melt interaction will be a topic for future
glass-melt-tank simulations.
Foam Formation. The formation of foam [2.39] is an important phenomenon
in the mathematical modeling of glass-melting processes. A foam layer either
blocks or reflects the heat radiation from the flames to the glass-melt surface.
Foam influences the local heat transfer to the melt and the temperature gra-
dient in the melt. This often undesired process of foaming may have dramatic
effects on the temperatures and flow patterns of the melt in the tank and on
the energy consumption of the furnace.
Foaming on top of the melt is caused by degassing of the glass melt
during fusion of the batch blanket (primary foam) and/or during the fining
process (secondary foam). Foam formation, with bubbles containing a large
amount of 80 2 and often also O 2 gas, is most likely during sulfate fining
because many bubbles rise to the surface from the hot spot or batch-blanket
tip [2.39]. At the surface, the bubbles are often relatively stable for a while,
due to stabilizing effects caused by surface tension gradients (by very local
differences in the sulfate concentration at the liquid-film surfaces) in the foam-
bubble lamellae. The foam is strongly insulating; it prevents heat penetration
from the combustion chamber into the melt. This effect is unfavorable for
fining processes because particularly during primary fining high glass-melt
temperatures are required. The insulation effect of the foam layer can be
estimated by assuming that:
• the heat conduction of the foam is determined by the dispersion mechanism
of heat radiation at the bubble surfaces;
• each bubble will disperse 33% of the incident radiation (8% by reflection
at the bubble boundaries plus 25% by total reflection due to the bubble
shape);
• the penetration of radiation (l/e = ±33%) equals a foam-layer thickness
of about the height of two bubbles (or a little more);
• the heat conductivity can be described by the Rosseland formula, in which
the reciprocal value of the indicated penetration depth (two bubbles) is
used as absorption coefficient:
16 n 2 uT 3
Afoam =:3 k ' (2.40)
2.1 Evaporation and Superstructure Refractory Attack by Vapors 71

where Afoam is the foam heat conductivity (W /(mK)), n is the refractive


index of the melt in the foam (;::,j 1.50), (J is the Stefan-Boltzmann constant,
T is temperature (K), and k is the absorption coefficient (m -1), whose value
is equivalent to l/(penetration depth) or 1/(2 x bubble diameter).
Example
Bubble diameter Afoam at 1400 DC
mm W/(mK)
1.0 6.54
2.0 13.08
3.0 19.6
4.0 26.2
5.0 32.7

A foam layer of 2 cm thickness, with bubbles having a diameter of 3 mm, has


a thermal resistance of 0.001 K m 2 /W. So a heat flux of 50 kW /m 2 will result
in, or will be caused by, a temperature difference across the foam layer of
50 D C!

2.1.5 Dynamic Modeling

Most model calculations and demonstrations of the simulation of the glass-


melting process are presented for steady-state conditions. Temperatures,
chemical compositions, and flows are only regarded as position dependent and
not as time dependent. However, disturbances or adjustments and changes in
the input parameters (e.g., batch composition, firing rate, power distribution
among electrodes) will lead to a longer period of changes in temperatures,
flows, and chemistry. Examples are pull-rate changes and changeover from
one glass type to another [2.92] or from one color to another. An additional
complexity connected with changing the glass or batch composition is the
change in the glass properties. To calculate the properties of the "mixed"
melt, the composition of the glass or the old-to-new glass ratio in the grid
cells must be known for each position in the melt at a certain time. Such
properties are, for instance, effective conductivity (particularly important for
glass color changes), density, viscosity, and redox state.
Use of simulation models for model-based predictive control requires fast
dynamic models that predict the future performance of the furnace and are
capable of following input-parameter changes. Figure 2.20 shows the results
of a dynamic glass-furnace model, simulating a color change in a float-glass
furnace.
Future developments in dynamic modeling of glass-furnace processes in-
clude the combustion-chamber process and the chemical changes during a
transient period in the tank. During changes in the settings of furnaces, tem-
72 2. Melting and Fining

~ ··- iS1:.'.n
0.0 0 .1 0 .2 0 .3 0.4 0.5 0 .6 0 .7 0 .8 09 1.0
Fraction of "new" glass composition in melt

Fig. 2.20. Color change in a melting tank, calculated by a dynamic CFD glass-tank
simulation model for a float-glass furnace, 40 h after the batch composition change
(average residence time = 70 h)

perature changes as well as redox changes in the product and changes in the
fining or melting index values can be calculated.
Dynamic models help furnace operators in predicting the response of the
furnace to changes in the process settings, and in finding furnace settings en-
abling faster color changes. Within about 1 h, the performance of the melting
process during a prediction horizon of 2-5 days can be simulated [2.93]. The
development of faster simulation models is essential for the further develop-
ment of model-based control, the detailed numerical models in some sections
of the furnace can partly be replaced by analytical models or simplified ap-
proaches [2.94]. For use of CFD simulation models in process-control systems,
their prediction of the process behavior must be about 100-1000 times faster
than the real procedure.

2.1.6 Concluding Remarks

Mathematical simulation models based on computational fluid dynamics


(CFD) for glass furnaces have replaced most of the physical models today.
Physical models, however, are still used for complex flow fields, such as flows
around stirrers or bubblers.
The CFD-based simulation models describe mass, heat, and momentum
transport processes as well as chemical conversions occurring in glass fur-
naces. Over the last decades, they have mainly been used to understand the
mass- and heat-transfer processes and temperatures in the combustion spaces
and melting tanks of existing glass furnaces and to optimize the adjustment
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 73

of process parameters such as batch composition, fuel distribution, bubbling,


and boosting conditions.
Application of such models for designing new furnaces or for process con-
trol requires more validation of the modeling results. The currently available
models are often only validated with respect to some glass-melt temperatures
and residence-time distributions at only one process setting. However, val-
idation of the models, describing the batch-blanket's thermochemistry and
rheology, the local flow rates in melt and combustion space, and the redox
reactions in the melt, requires new experimental methods that have to be
applied in industrial furnaces.
The calculated flow patterns and the temperature field in glass furnaces
depend on each other and on the presence of a foam layer on top of the melt
or the shape and size of the batch blanket. More accurate first-principle-based
models are needed to describe foam formation and foam decay or the behav-
ior of the batch materials floating on the melt. Today, submodels linked to
CFD models exist for the description of gas removal, evaporation, refractory
corrosion, redox changes, and melting kinetics. These models require accu-
rate thermodynamic data, such as vapor pressures, chemical activities, or
solubility values, as well as mass-transfer data such as diffusion coefficients,
for example.
Application of CFD-based glass-furnace simulation models will extend
in the future, but requires the possibility of fast time-dependent modeling,
for instance for process control or simulation of pull or color changes. Use
and availability of more accurate glass-melt property data and validation
will help improving the models' reliability. In short, future efforts will focus
on three major issues: validation, improved glass-property data (particularly
important for submodels describing the chemical processes), and increased
calculation speed for the application of model-based predictive control.

Acknowledgement
The author gratefully acknowledges the support by Dr. Ir. A.M. Lankhorst,
Ir. P. Simons, Dr. Ir. O. Op den Camp and lng. A. Habraken from TNO
TPD, who provided results and examples of CFD glass-furnace modeling.

2.2 Mathematical Modeling of Batch Melting


in Glass Tanks

Wolf S. Kuhn
Introduction
The melting of the raw-material blanket that floats on the glass melt is a
highly complex process. Hydrodynamics, thermodynamics, reaction chem-
istry, and thermal science are involved in the process, from a microscopic
74 2. Melting and Fining

scale in sand grains to a macroscopic scale in the batch blanket of large glass
tanks. As of today, a direct numerical simulation covering the whole scale
range of the melting process is still not possible.
This is why numerous studies have proposed models with different simpli-
fication levels. However, before deciding in favor of one model or another, it
is necessary to define the target of the model, within the scope of a complete
furnace modeling. The correct simulation of the energy flux into the batch
by glass recirculation remains one of the basic objectives of modeling. The
quality of the glass is determined, to a large extent, by this recirculation. It
is therefore essential that batch models properly describe the energy fluxes
on both the upper and lower faces.
In the following, we survey the batch-melting process from a fundamental
and technological point of view. Mathematical modeling is already helpful in
small-scale, laboratory studies on batch melting. However, a complete model
of the microscopic melting on the grain scale remains extremely challenging.
The mathematical modeling of the macroscopic furnace-scale melting is of
considerable technological importance and this is reflected by the numerous
publications on this topic. Before we start discussing some typical batch-
modeling approaches, we highlight the theory of macroscopic modeling. Some
discussions follow on some key values in batch melting together with the
characterization by non-dimensional numbers. On this basis, an assessment
of already-published batch models is carried out and the need for systematic
validation is stressed.

2.2.1 Motivation and Requirements on Batch Modeling

Mathematical modeling of the glass melt and combustion are now well estab-
lished in glass-tank technology. A key point in tank modeling is the reaction
and fusion of the raw materials forming the batch blanket. Batch melting is a
process of considerable complexity regarding reaction chemistry, heat trans-
port, and rheology. Additionally, in a large batch blanket without cullet, the
number of sand grains is of the order of 1010_10 12 . Owing to the size and
number of particles making up the raw materials, direct numerical modeling
of a batch blanket starting with microscopic melting is still not possible.
For these reasons, batch models rely on more or less simplifying assump-
tions that, as a result, limit their validity. However, what are the criteria that
determine the accuracy required for batch modeling? A basic requirement on
batch models can be highlighted with a very simple example:
The Technical Committee 21 (TC 21) of the ICG proposed in its first
round-robin test (RRT 1) a simple model glass tank, as a benchmark test for
tank-simulation codes [2.95]. This model tank comprises a rectangular cavity
for the glass, a throat as a glass outlet, and a batch area of fixed length.
This batch area is defined as an inlet boundary condition, with constant
glass flow and temperature. A slight modification of this boundary condi-
tion enables the allocation of a defined heat flux to this batch area [2.96].
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 75

This heat flux directly determines the intensity of the glass recirculation (see
Fig. 2.21). The higher the batch heat consumption, the stronger the recircu-
lation. Consequently, the minimum residence time decreases, defined by the
shortest time-path trajectory of the glass through the tank (see Fig. 2.22).
An increase of the batch area heat flux from 30 kW to 60 kW reduces the
minimum residence time by about 6%. This effect is accompanied by a de-
crease of the glass bottom temperature. The impact of such effects on the
glass quality is obvious. Hence, it is compulsory that the batch model pro-
vides correct heat-flux distributions if conclusions on glass recirculation and
glass quality are requested as a model output.
The demands on batch modeling may be classified as follows:
• Basic demands
- length extension,
- energy consumption (thermal balance),
- heat-flux distribution between the upper and lower sides.
• Advanced demands
- batch surface extension (length, width, shape) and flow velocities,
- thickness profile depending upon the local melting rates,
- heat-flux density profiles on the upper and lower sides,
- temperature field in the blanket and the surrounding melt.
• Very advanced demands
- size, number, and positions of Si0 2 grains released into the melt,
- size, number, and positions of gas bubbles released into the melt,
- foam surface extension and thickness,

Fig. 2.21. Melt temperatures


and recirculation in a 2D model
of the TC 21 RRT 1 tank (after
[2.96]). The recirculation intensity
is determined by the energy con-
sumption of the batch blanket

6.2 - , - - - - - - - - - - - - - - - - - ,
6.1
6 .0
5.9
-E 5 .8
~ 5 .7
i= 5.6
5.5 Fig. 2.22. Minimum res-
5.4 idence time of the TC 21
5 .3 +----,----.,.-----,r---.----i RRT 1 model tank as a
20 40 60 80 100 120 function of the boundary
Lower-side heat flux /kW heat flux in the batch area
76 2. Melting and Fining

- redox state of the primary melt as a function of the raw-material reac-


tions,
- dusting and evaporation of batch components.
In addition, advanced batch models should cover the widest possible range
of charging modes and should not be technology dependent. However, for
tank modeling, different batch models with different levels of complexity are
necessary, depending on both the required precision and the available data.
The simplest model, which could be used to represent a batch blanket
in a melt simulation, is a boundary condition representing a moving wall
with heat consumption and a mass flux. The correct implementation of such
boundary conditions is already in itself a difficult task [2.95,96]. Furthermore,
the space occupied by the batch above and under the glass level is not given
by this approach.
The next step is to represent the batch using 2D heat-penetration mod-
els or 3D computational fluid dynamics (CFD) simulation. Advanced batch
models are indeed based on a CFD approach and guarantee great flexibility
for any type of charging, furnace geometry, and raw materials. The advan-
tages and limits of CFD batch models will be discussed. But first we try to
develop a complete picture of the fundamental aspects and of the technology
of batch melting.

2.2.2 Survey of Batch Melting

Before entering into the subject of mathematical modeling, let us review


the thermochemistry and technology of the batch-melting process from the
modeler's point of view. For a detailed review on the thermochemistry of the
melting process, the reader is referred to [2.97].

Thermochemical Reactions

For the purpose of clearance, and to limit the size of the following sections,
we will restrict the discussion to a very simple glass-melting system - the
sodium silicate fusion. Sodium silicate, or water glass, is an important inter-
mediate product for the chemical industry. Examples from the wide range
of possible applications are: water treatment, detergents, adhesives, binders,
soil stabilization, and so on. Sodium silicate is mainly produced by a reaction
of soda ash, with sand, in gas- or oil-fired glass tanks.
The stoichiometry of the product sodium silicate glass is determined by
the mixture ratio of the educts. In sodium silicate production, 1 mole Na 2 C0 3
is frequently mixed with about 3 moles Si0 2 , thus resulting in glass with 70-
75wt% Si0 2 .
Furthermore, in the fusion of soda-lime silicate batches, the reaction
pathway can either follow the so-called "silicate route" or the "carbonate
route" [2.97]. The silicate route starts with a reaction between Si0 2 and
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 77

Na2C03. Let us study this very simple reaction of silicon dioxide with sodium
carbonate:

(2.41 )

First of all, we need to look at the reaction enthalpy of this reaction. For
a value of n = 3, f:1Hchem is about 550 kJ /kg. This reaction enthalpy is
positive, and as a result, the reaction is endothermic, like most of the batch
reactions in glass technology. It should be noted that this reaction enthalpy
comprises both the energy for the chemical reaction and the energy for phase
transformations (crystalline solid to "glassy" liquid and gas).
The reaction enthalpy, often referred to as "chemical energy demand", is
an important contribution to the total energy requirement for the fusion of
the batch raw materials. Extensive literature deals with the calculation of
chemical and total energy for the fusion of the numerous batch compositions
in glass making (e.g. [2.98-100]).
For more complex glass compositions and for thermal balances, it is cur-
rently preferred to formulate a global batch reaction:

(1 + a) kg glass batch(Tin) =? 1 kg glass melt(Tout) + a kg gas(Tgas)


(2.42)
(+ energy per 1 kg glass melt) .

If the temperatures Tin, Tout, and Tgas relate to the same reference tempera-
ture, the energy requirement of the glass-melting reaction will correspond to
the reaction enthalpy. If these temperatures refer to the glass-tank process
temperatures, this energy requirement includes the sensible heats of batch,
melt, and gas.
Next, we will look at the chemical equilibrium of the sodium silicate re-
action. The equilibrium constant is, in principle, given by

K = [a(Na 2 0 - nSi0 2 )][a(C0 2 )]


(2.43)
[a(Si02)f'[a(Na2C03)]
The equilibrium constant expresses the equilibrium concentrations of prod-
ucts and educts that are the final concentrations after a sufficiently long reac-
tion time in a closed vessel, at a given temperature. The activity coefficients
a can be replaced by appropriate concentration units. Gas concentrations
are usually represented by partial pressures, and condensed phases by mole
fractions and fugacity coefficients. For pure condensed phases such as Si0 2,
Na2C03, and sodium silicate, a = 1 holds. Hence, the equilibrium constant
reduces to

K =Peo 2 , (2.44)
where Peo 2 is the partial pressure of CO 2 .
The equilibrium constant is related to the free reaction enthalpy (or Gibbs
free energy) of the reaction.
78 2. Melting and Fining

(2.45)

This relationship expresses the temperature dependence of the equilib-


rium constant. As a result, the equilibrium gas pressure and the concentra-
tions of educts and products are temperature dependent.
An analysis of the equilibrium temperatures (K = 1) of sodium silicate
and calcium-sodium silicate formed from sand and carbonates can be found
in [2.98]. At temperatures above 300 ac, the glass formation is clearly favored
for silicate-rich stoichiometry.
The partial pressure of CO 2 or other reaction gases, influences the batch
fusion reaction. Experimental evidence for this statement is, for example,
given in [2.101]. The calcium carbonate dissociation is a nice example of
the CO 2 pressure impact. At 700 ac, the CO 2 partial pressure is as low as
0.02 bar, which indicates that no significant amount of CaO would be created,
even after longer reaction times. At 800 ac, the CO 2 pressure reaches 1 atm.
If no efficient transport process removes the liberated CO 2 from the reaction
area, dissociation stops.
However, for heterogeneous, multicomponent systems such as granular
batch mixtures, the description by a concentration-related equilibrium con-
stant is inappropriate. If equilibrium quantities for solid, liquid, and gaseous
species of a multicomponent system are required, the theoretical treatment
becomes more challenging [2.102,103]. Here, only a brief outline of this prob-
lem will be given.
The solution method is based On the Gibbs energy minimization. The free
energy of a system can generally be expressed as

(2.46)

where J..1i is the chemical potential of each involved substance and qi is the
mole number of the substance i. This represents a sum of Gibbs energies of
all condensed and gaseous species. Next, mass conservation of all involved
species must be considered. From numerical algorithms, the minimum Gibbs
energy has to be found under the constraint of mass conservation. With this
method, equilibrium quantities for condensed (raw materials and melt) and
gaseous (reaction gases) species in a closed system can be determined.
These equilibrium quantities for raw materials, glass, and gas indicate
the thermodynamic direction of the batch reaction for given temperatures
and input masses for both semi-closed and open systems (pot-melting or
continuous tank).
The free reaction enthalpy of a global batch reaction is given by the
reaction enthalpy reduced by the reaction entropy:

l1G = l1H - Tl1S . (2.47)


2.2 Mathematical Modeling of Batch Melting in Glass Tanks 79

However, batch reactions always occur in the same temperature range in


glass technology. For this reason, no distinction is made between b.G and
b.H in most studies on batch energy, and this means that the entropy term
is frequently dropped for batch-modeling purposes because the entropy is
already comprised in the literature data on batch reaction enthalpy. However,
the entropy and the reaction-gas partial pressure should be taken into account
in advanced batch modeling because of their impact on both the reaction
equilibrium and reaction temperature.

Reaction Kinetics on the Molecular Scale

In reaction kinetics, the basic approach consists in determining the reaction


partners, the reaction order, and the mass-conservation equation, which de-
fines the time evolution of all concentrations.
The above sodium-silicate reaction (2.41) then leads, in principle, to the
following differential equation:

(2.48)

where dx / dt describes the speed of the reaction as a function of the concen-


tration of the educts, for each time t. The proportionality constant k is the
rate constant of the reaction. The variable x is related to the concentrations
as

(2.49)

The reaction order is given by n + 1 and defines the number of molecules


necessary to undergo one elementary reaction.
The temperature dependence of the rate constant k is frequently given by
an Arrhenius-type relationship:

(2.50)

Ao and Ea represent the pre-factor and the activation energy of the reaction,
respectively, and R is the gas constant.
However, batch raw materials, such as sand and soda ash, are in general
granular. They do not represent a homogeneous mixture on the molecular
scale! In the contact areas of the grains and in the locally formed first melt,
the transport and diffusion of the reaction partners seems to be the rate-
controlling mechanisms. Unfortunately, very few studies deal with this sub-
ject [2.97, 104-106]. The following discussion reveals the whole complexity of
the reaction kinetics of granular batch materials.
80 2. Melting and Fining

Kinetic Effects on a Grain Scale

The global reaction (2.41) does not proceed homogeneously on the grain scale.
Significant local deviations from stoichiometry appear, which are related to
different solid, liquid, and gas phases that occur during the melting process.
At this point, it is very interesting to look at the binary phase diagram
Na20-Si02 in Figure 2.23. Principally, liquidus lines in phase diagrams in-
dicate the appearance of crystalline phases during cooling of a given melt
composition, or the disappearance of the last crystalline phases in the heat-
ing of a liquid-crystal mixture. Heating or cooling are supposed to be very
slow processes, which maintain the system near thermodynamic equilibrium.
Although the batch fusion reaction is principally a non-equilibrium pro-
cess, interesting conclusions can be drawn from phase diagrams. However, the
starting point in the phase diagram is not the batch-mixture composition on
a macroscopic scale that represents the desired end point of a homogeneous
glass composition. The starting point is given by the stoichiometry of the
product first formed during heating of the mixture.
A comparison of the melting temperatures in Table 2.2 reveals that the
first raw material to melt is the soda ash. Na2C03 is therefore available in
excess at the surfaces of the Si0 2 grains, thus shifting the stoichiometry to the
sodium-rich side. Immediately after reaction and carbonate release, a sodium-
rich Na20-nSi02 mixture prevails in the reaction area. Hence, orthosilicate
(n = 0.5), metasilicate (n = 1), and disilicate (n = 2) are discussed as
intermediate products in the course of the fusion reaction of sodium silicate
with one part Na 2 0 and three parts Si0 2 (see [2.97,104,105, 1l1]). Sodium-
rich sodium silicate has higher melting temperatures, lower melt viscosities,
and greater tendency to crystallize. The final sodium-silicate composition will
have low melting temperatures but high viscosities (Table 2.2). However, for

1200.-------------------------------,

,/,/102 //
",
/////,/10 3
~ 1000 1021
9- /

800
799
40 60 80
Na20 ----------------------~--
Content /mol%

Fig. 2.23. Melting phase diagram of the binary system Na20-Si02 (after [2.97].
The solid line represents the liquidus curve. The dotted lines indicate the melt
viscosities in dPas
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 81

Table 2.2. Temperatures and approximate viscosities of compounds involved in


sodium silicate melting

Compound Tm or Tref Viscosity (dPa s)


Si0 2 1723°C
Na 2 0 1132°C <1
Na2 C0 3 858°C <1
Na2 0·3Si02 850°C ::::0 104

2Na 2 0·3Si02 840°C ::::0 103


Cullet (soda-lime silica) 900 °C > 105

glass-forming melts, the melting temperature is no longer clearly defined.


In conclusion, one can expect a complicated course for the heterogeneous
sand-soda ash reaction, due to thermodynamic constraints. Such interactions
have been studied by Manring and Bauer [2.108] with a hot-stage microscope.
Two sand grains on a bed of soda ash start to react vigorously at about
880°C (see Fig. 2.24). The sand grains are quickly wetted and enveloped by
the molten soda ash. The crunchy morphology indicates either a very inho-
mogeneous reaction at the sand-grain surfaces or the orientation-dependent
growth of small sodium silicate crystals. Manring and Bauer assume the de-
velopment of a "solid crust of metasilicate" , albeit without extending further
evidence or giving information about its supposed composition and structure.
At about 1200 DC, the crust has vanished and a liquid droplet, containing the
yet undissolved silica grains, is left (see Fig. 2.24).
As a result, one can resume the soda-ash sand reactions as follows: at
temperatures below the melting temperature of Na2C03, slow solid-state
reactions occur.

T < Tm = 858°C:

+ nSi0 2 (quartz)
Na2C03 (crystalline solid)
-+ Na20-nSi02 (crystalline solid) + CO 2 .
T> Tm = 858°C:
Na2C03 (melt)+ nSi0 2 (quartz)
-+ Na20-nSi02 (viscous melt) + CO2 forT> Tliquidus ,
-+ Na20-nSi02 (crystalline solid) + CO 2 for T < Tiiquidus .

Above the melting temperature of soda ash, two reaction pathways are
possible. If, in the course of the reaction, the stoichiometry lies on the sodium-
rich side and the temperature remains below the liquidus line in the phase
diagram, the formed sodium silicate is likely to occur as crystalline solid. The
further increase of temperature leads to the melting of the ortho-, meta-, and
disilicate mixture. At the end of the batch reaction, one has a sodium-rich
melt that still contains a high number of undissolved sand grains.
82 2. Melting and Fining

Fig. 2.24. Microscopic images of


the melting of sand grains on soda
ash (after [2.108]). Upper image:
The two sand grains show no sig-
nificant reaction at 800 DC. Mid-
dle image: At 880 DC the grains of
soda ash have disappeared and an
inhomogeneous reaction layer has
formed on the surfaces of the sand
grains. Lower image: At 1200 DC
the inhomogeneous reaction layer
has disappeared but the two sand
grains are still visible. Small bub-
bles of CO 2 can also be seen

This simple example of the sand-soda ash reaction already reveals a con-
siderable complexity on the grain scale. Soda-lime silica and other technical
glasses show even more complex melting reactions. Mathematical modeling of
these processes seems to be a utopian task. A detailed approach to the melt-
ing kinetics on a grain scale must take into account all aspects of chemical
reaction, phase transformation, and mass transport by diffusion and convec-
tion. Nevertheless, certain approaches have been published that try to tackle
reaction kinetics on the grain scale.
The theory of sintering deals with the kinetics of the compaction of gran-
ular materials ([2.109]). The theory of sintering in the presence of a liquid
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 83

phase relates progress parameters, such as density or porosity, to viscosity IL,


surface tension CJ, and particle size d. Conradt [2.98] studied in detail the
liquid-phase sintering for numerous batch compositions. The time T for the
sintering of cullet, as an example, can be estimated by

T ~ 2dl!. . (2.51 )
CJ

Smaller particle size and higher temperatures, and hence lower viscosity, sig-
nificantly accelerate the cullet compaction. However, in reactive sintering of
glass batches, reaction kinetics, gas release, and solid-liquid phase transfor-
mations have to be taken into account. Owing to this complexity, a detailed
theoretical description of batch melting in the framework of the sintering
theory remains very challenging.
Up to now, only phenomenological studies give access to the kinetic be-
havior of glass batches on the grain scale. In a first approach one might try to
interpret the overall batch reaction as a first-order decomposition reaction:

Abatch --+ Brnelt + Cgas . (2.52)

For a first-order reaction, the educt concentration [A] decreases according to

[A](t) = [Aa]e- kIt , (2.53)

where kl is the first-order rate constant. For different reaction orders, different
progress laws result (see standard literature on chemical kinetics).
However, Kroger [2.101] conducted measurements on the rate of CO 2
evolving from sand-soda ash batch mixtures under nearly isothermal condi-
tions. The size of the samples was about 10 cm 3 , which can be considered as
macroscopic, compared to grain size. Two thermocouples in the center and
at the surface of the sample controlled the thermal homogeneity. For fast
reactions at higher temperatures, the temperature homogeneity could not be
guaranteed with this sample size. However, for slow and medium reaction
rates, the temperature-dependent rate constants could be determined. It is
particularly interesting that Kroger states that the Jander relationship best
fits the data

(
1- 3 100%-X(t))2 =kt (2.54)
100% '

where x(t) is the progress of CO 2 escape from the samples. The temperature
dependence of the rate constant k is of the Arrhenius type. Unfortunately,
it is not possible to describe the entire reaction-temperature range by one
Arrhenius relationship. The rate constant also depends on the CO 2 partial
pressure. For further discussions on batch-reaction kinetics, the reader is re-
ferred to [2.97,101,110].
84 2. Melting and Fining

Reactive Grain Dissolution

We have seen in the former discussion that the heterogeneous batch reaction
mainly proceeds with reactions at the surface of the solid silica grains. The sil-
ica grains are successively dissolved to end up with a homogeneous melt. The
dissolution of the silica grains is an alternative way to study batch-melting
kinetics. Furthermore, dissolution of sand grains and other inhomogeneities
is a major subject in glass-tank modeling because these defects considerably
degrade the final glass quality.
Hrma [2.111] developed a reaction-path and grain-dissolution model for
the reaction between sodium carbonate and silica grains. The global reaction
(2.41) is subdivided into four stages: decarbonization, dissolution of orthosil-
icate, dissolution of metasilicate, and final dissolution of silica grains.
The formal description of the first stage relates the grain dissolution to
the evolving flux of CO 2 :

dr = (3 nJ Vs (2.55)
dt C Vc
The decrease of the average sand-grain radius r is given by the grain sur-
face fraction (3c wetted by Na 2 C0 3 • The stoichiometric coefficient n of the
reaction, the CO 2 volume flux J per wetted grain surface, and Vc and Vs
are molar volumes of carbon dioxide and solid silica at the melting tempera-
ture. This approach, in connection with mass-balance considerations, allows
basic insights into the decarbonization kinetics of soda ash-sand mixtures.
Nevertheless, (3c, n, and J have to be determined in experiments [2.104].
For the formal description of the next melting stages, Hrma assumes that
a melt layer exists between the remaining silica grains and orthosilicate and
metasilicate crystals. The decrease of the average silica-grain radius r is now
given by

dr =(3Db. f (2.56)
dt 8 0'
where (38 is the grain-surface fraction wetted by the sodium-silicate melt, Dis
the diffusion coefficient, b.f is the difference of the Si0 2 melt mass fractions
at the silica grain surface and at the ortho- or metasilicate crystal surface.
o is the concentration layer thickness of the separating melt. Again, the ba-
sic parameters (e.g., (38 and 0) must be determined experimentally [2.104].
Moreover, the existence of a thin melt layer between the silica grains and the
metasilicate crystals could not be verified with the hot-stage microscopy by
Manring and Bauer [2.108].
However, Hrma stresses that splitting, agglomeration, and sintering of
silica grains affect the effective wetting values of (3c and (38. Non-uniform
dissolution of silica grains may, furthermore, arise from imperfect mixture,
grain-size variations, buoyant segregation due to low-viscosity first melt and
due to rising of sand grains, enhanced by bubbles attached to the sand grains.
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 85

The diversity of the phenomena of the reactive grain dissolution renders


a detailed theoretical description of the reaction kinetics on the grain scale
extremely difficult. For further discussion on interface reactions and reactive
grain dissolution, the reader is referred to [2.97,105,110,112].

Diffusive and Convective Grain Dissolution

As demonstrated, the dissolution of the silica grains is closely connected with


the complex reaction kinetics on the grain scale. The theoretical description
is considerably simplified if the reactive grain dissolution and gas liberation
is terminated and if the silica grains are completely wetted by homogeneous
melt. This situation is illustrated by the lower image of Fig. 2.24. The re-
maining CO 2 bubbles are quite small and thus do not disturb the wetting
assumption.
The diffusion-controlled dissolution of a spherical particle in the melt can
be considered as quite well studied. The dissolution rate is generally expressed
by the radius decrease

dr h (fSi02 fSi02) (2.57)


dt = - diff interface - bulk

with the Si0 2 melt mass fractions in the volume and at the grain surface.
Depending on material properties and boundary conditions, different, more or
less analytical, solutions exist for the underlying mass-transport differential
equation.
For the quasi-stationary, isothermal situation Weinberg et al. [2.113] de-
termined the mass-transfer coefficient as

hdiff =D (~+
r
~)
V7r Dt
. (2.58)

The dissolution is time dependent with a time root, as frequently found in


diffusion-controlled transient problems. The sand-grain dissolution in a stag-
nant melt is determined by the diffusion coefficient D for Si0 2 in the melt;
this parameter is strongly temperature and concentration dependent [2.112].
In real melting situations, melt convection is an important means to en-
hance the grain dissolution. A few articles deal with the theoretical descrip-
tion and analytical solutions of the convection-enhanced grain dissolution in
glass melts [2.114-118]. However, the convection-enhanced grain dissolution
is even more difficult to study in well-defined laboratory experiments than
the dissolution in a stagnant melt. A different approach to validate simplified
analytical solutions consists in the set-up of a detailed numerical model.
Such a model consists in a transient, 3D-CFD simulation of the dissolu-
tion of a spherical silica grain, the concentration field around the grain, and
the melt convection. Figure 2.25 shows the dissolution of a silica grain in a
disilicate melt at 1200°C, under shear flow [2.119]. The shear flow signifi-
cantly increases the concentration gradients and, as a consequence, reduces
86 2. Melting and Fining

Fig. 2.25. 3D model of a sand grain in a disilicate melt: Si0 2 concentration in a


Na20-2Si02 melt with a shear flow

the dissolution time. For the set-up of this model, the general-purpose CFD
code Fluent was used.
Of course, such a model needs detailed material parameters as a function
of temperature and composition. But considerable insight into the dissolution
process and its limiting factors can be gained.

Final Remarks on Mathematical Models on the Grain Scale

Finally, the question arises as to which of the formerly discussed models is


best suited as a basis or complement for furnace-scale modeling. Phenomeno-
logical models of the batch reaction kinetics on the grain scale are most easily
accessible and can more or less readily be integrated into batch models on the
furnace scale (see, for example, [2.107]). But phenomenological models always
need considerable experimental support and systematic validation. Further-
more, they have difficulties in satisfying advanced batch-model demands on
sand grains and bubbles released into the melt.
Here, detailed mathematical modeling might be a valuable tool in the
study of reactive grain dissolution and the related reactive sintering. How-
ever, such models would be of considerable complexity because they comprise
transient calculations, melting and wetting, convection and diffusion, reac-
tion and solidification, and bubble creation and growth. It is not likely that
numerical modeling of reactive grain dissolution and sintering will emerge in
the near future.
The modeling of diffusive and convective grain dissolution is currently
best tackled by more or less simplified analytical approaches. Such analytical
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 87

approaches have the advantage that they can be coupled by the Lagrangian
approach to mathematical models of the batch and the melt on the furnace
scale.
In conclusion, the mathematical modeling of the batch-melting reactions
is a challenge for advanced batch models. Kinetic reaction models are the
basis for the correct calculation of local melting temperatures, local chemical
kinetics, and local melting rates. Very advanced batch models will predict the
size and number of sand grains and bubbles released into the melt as well as
the concentration of the dissolved gases. The modeling of the grain dissolution
and the bubble behavior in the melt then yields valuable information about
the final glass quality.

Batch Composition, Morphology, and Fusion Above Grain Scale

The standard batch for soda-lime silica glass mainly consists of a mixture of
raw materials, comprising about 70 wt% sand grains (100-500 ~m in diame-
ter) and both sodium and calcium carbonates in the form of powder. Sulfates
and hydroxides are used occasionally. The glass composition is completed by
smaller amounts of K, Mg, AI, Fe, and S oxides. Lead-crystal, borosilicate,
TV-screen, Sr, and Ba glasses have, of course, different batch compositions.
To reduce the energy requirement of the melting process, significant
amounts of recycled glass (cullet) are mixed with the raw materials. The
batch is, in general, made up of a granular mixture of raw materials and
cullet. Pellets and briquettes are occasionally used. Water addition helps to
reduce the dusting of batch powder by the combustion gases. From a theo-
retical point of view, an unfused batch is a granular mixture of solid particles
of different size and composition with air filling the void space between the
particles.
Figure 2.26 illustrates the different process steps during the batch melting,
together with the corresponding batch morphology. The batch is commonly
heated on its lower and upper sides. The granular batch undergoes various
complex changes in the chemical reaction zone. At the end of the reaction,
a high number of undissolved grains and gas bubbles is left. Those grains
eventually start to dissolve but are soon uptaken by the melt convection,
thus leaving the "batch melting zone". The gas bubbles either rise in the
melt that covers the batch, or are uptaken by the melt currents at the lower
side of the batch. The interface between "batch" and "melt" has more the
character of a transition zone. This transition zone is of particular importance
for the heat transfer and the bubble and grain release.
It is obvious that varying chemical batch composition, varying granular
morphology and melting-process complexity represent an extreme challenge
for mathematical modeling.
88 2. Melting and Fining

Glass layer
(liquid-gas)

TIT\>
Unfused grains
Melling of
(solid·gas)
undissolved Undissolved
silica grains silica grains
(liquid· solid-gas) Tm,
Chemical reactions
(Iiquid·solid-gas)
Gas bubbles -t~~;g;;~;ZJ~
Chemical mCR=O
reacUons Melling of
(Iiquid·solid·gas) undissolved
silica grains
Tm3 (Iiquid·solld·gas)
,,,,-__r- -'-Gas Tm2
Unfused grains
(solid·gas) Silica grains

Fig_ 2.26. Morphology and process steps at the upper (a) and lower (b) side of a
batch blanket during melting (after [2.107])

Batch Heating and Morphology on Furnace Scale

The granular mixture is usually charged on the surface of the melt. Owing to
its lower density, the batch mixture floats on the surface until it is transformed
into glass melt. This batch can either be heated on its surface, by combustion,
or on its lower side, by melt convection. Batch-blanket melting can be divided
into different stages [2.120]. However, the process stages at the upper and
lower sides differ (see Fig. 2.27).

Combustion space
Heat transfer by radiation

~ Heating : Reaction ~ Primary melt


:up :start ~formation

~ Cold batch Reaction Core )~


direction core layer t~mperature
---.
-_____~__~--r--,~r~ls~e~--~~~--~~--
t • t t t ...---'----
Recirculating
~Heating ~Reaction Primary melt glass melt
~ up ~start release

Heat transfer by convection and radiation

Fig. 2.27. Sketch of a melting batch blanket with horizontal feed. Different melting
stages can be distinguished
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 89

Upper Batch Side. Raw materials and cullet are first heated at the surface by
radiation and convection of the combustion gas. The radiative heat-flux den-
sity is of the order of 300-500 kW 1m 2 , depending on the source temperature
and source emissivity. Unfortunately, unreacted batch has a high reflectivity.
Furthermore, the batch is a poor heat conductor and develops steep tem-
perature gradients upon heating. High temperature gradients at the start of
heating allow high heat-flux densities. But this advantage is lost when the
batch core rises in temperature. The next step is the reaction onset, which
will, on top of the core heating, consume energy. If the reaction is finished
in the uppermost batch layer, primary melt is formed, which successively
increases in thickness (see Fig. 2.27).
Depending on the surface morphology, down running can remove the
freshly formed melt and limit the surface-melt thickness. The surface melt
has a significantly higher absorptivity than the unreacted batch, which should
enhance the heat-flux density at the batch surfaces. However, the higher sur-
face temperature of the melt and the additional thermal resistance limit the
possible gain. Finally, the foamy primary melt floats on the recirculating glass
melt. Foam presents a considerable thermal resistance for the heat transport
into the melt.
Lower Batch Side. As soon as the unreacted batch comes into contact with
hot glass melt, the batch is heated up and the melt is cooled down. After the
batch has reached the reaction temperature, the reaction layer is formed. The
speed of this process depends on the detailed heat and mass transport in the
batch-charging area. Once the reaction is finished in the lowest batch layer,
primary melt is formed. The release of the primary melt into the recirculating
melt depends on the detailed temperature and flow pattern under the batch
blanket.
Depending on the melting technology, reaction kinetics, melt and batch
properties, deviations from this melting scheme occur. A crucial parameter in
batch melting is the heat-flux density at the lower and upper batch sides. In
the literature, values of about 40-80 kW 1m 2 for the mean heat-flux densities
at the lower and upper batch sides can be found [2.120]. If heat-flux densities
for every single process stage are required, all process stages have to be in-
vestigated thoroughly. Furthermore, the heat-transport processes related to
the batch water and reaction-gas release have to be considered.
The heat- and mass-transport processes in batch-blanket melting in con-
junction with the complex boundary conditions render a closed mathematical
description of the whole melting process impossible. One way to tackle this
problem is to deal with the different melting stages separately. The other way
is the complete numerical simulation of the melting process. Before entering
into the subject of the mathematical description of batch melting, we will
take a glance at the various melting systems in glass technology.
90 2. Melting and Fining

Batch-Melting Technology

The furnace types and the batch-charging modes to be modeled must be


defined before the definition of any batch model. Some typical glass furnace
types and their charging modes are presented in Table 2.3.
Two principal charging modes are used in glass technology.
• Continuous charging
- Screw feeders create batch ribbons or a single blanket.
- Cold-top surface chargers create a closed batch blanket.
- Vibrating chargers with pushing bars or rabbles create ribbons that fuse
to form a blanket of constant thickness.
• Discontinuous charging
- Vibrating chargers with pushing bars or rabbles create separate ribbons
or piles.
- Cold-top surface chargers with tipping arms create a batch with varying
thickness.
In the case of continuous charging, discontinuous heating in regenerator
furnaces often causes fluctuation of the blanket shape, with the frequency of
the heating reversal.
Figure 2.28 shows the shape of a batch blanket in a bottle-glass furnace,
with continuous heating and a capacity of 400 t / day. The raw materials mixed
with cullet are charged at the rear end of the furnace by two chargers with
pushing bars. These chargers first create a closed blanket, but soon afterwards
foamy melt zones begin to appear and break up the closed blanket.
Finally, batch islands appear and move up to the bubbler row. The air
bubblers create an upward convection in the melt, which stops the batch
islands. Foam-free cooler and radiation-absorbing melt create the dark area
around the bubbles. The batch shape in this furnace is a clear illustration for
transient batch melting, due to discontinuous charging. In Fig. 2.28, white
lines mark the shape of the batch lumps. An image analysis software quanti-
fies the extension of the batch lumps. This software is part of an automatic
furnace control, developed by Stein Heurtey. Details on the image analysis
and the control software can be found in [2.153].
All batch models published until now address discontinuous charging by
stationary-melt calculations (see Sect. 2.2.5). Furthermore, batch models are
often set up and more or less validated for one particular charging and furnace
type. Some models take discontinuous batch piles and heating reversal into
account via special algorithms that calculate mean values for temperatures,
heat and mass flux.
Truly versatile batch models should be free from any geometrical restric-
tions and applicable to all principal furnace and glass types, as well as to
different charging and heating modes (see Table 2.3). However, discontinu-
ous charging and heating must be simulated by time integration with special
algorithms, if tedious transient calculations are to be avoided.
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 91

Table 2.3. Principal furnace·-batch combinations

Glass Furnace Typical Typical Charg- Typical Conbust- Typical


prod- type composi- cullet ing charg- ion melting
uct tion content position ing heating flux
mode mode densities
t/m 2 /day
Flat open soda-lime 20-30 front pusher alter- 4-5
float cross- silicate nating
fired

TV cross- barium- 30-50 front screw alter- 3-4


panel fired strontium nating or
with silicate continu-
throat ous
(oxygen)
Bottle side- soda-lime 15-80 front pusher continu- 4
fired silicate ous
unit
melter
Bottle end- soda-lime 15-80 side pusher alter- 5
fired silicate nating

Crystal cross- sodium- 30-80 front or pusher continu- 8-12


glass fired pot as- side ous
with sium lead
throat silicate
Crystal cold-top sodium- 3080 vertically distrib- 2-4
glass electric pot as- utor
sium lead
silicate
Crystal pot fur- sodium- 30 vertically manu- continu-
glass nace potas- ally ous
sium lead inter-
silicate mittent
Sodium end- sodium 0 side pusher alter- 4
silicate fired silicate nating
Sodium cold-top sodium 0 front or distrib- 1-2
silicate electric silicate vertical utor
Waste special fly ash 0 side screw electric
vitrifi- melter arc
cation

Standard Technology Data for Model Set-Up and Validation


Even the simplest batch models require some basic technological data. In
order to set up any specific batch model in a tank simulation, the following
information about batch charging and melting will generally be required:
92 2. Melting and Fining

Fig. 2.28. Image of the combustion chamber of a bottle-glass furnace (after [2.153J,
by courtesy of the Verrerie du Languedoc)

• Material data
- cullet ratio,
- chemical composition of the raw materials and cullet,
- cullet and grain sizes,
- humidity of raw materials and cullet,
- type and quantities of released gases.
• Geometry data
- length and shape of the batch blanket,
- doghouse geometry,
- superstructure geometry,
- glass level,
- batch level after charging,
- feeder type,
- feeder geometry,
- feeder position.
• Thermal boundary conditions
- wall isolation or losses,
- losses by feeder cooling,
- entrance temperature of the batch mixture,
- crown temperatures in the batch area,
- energy balance of the furnace.
• Flow boundary conditions
- raw material feed rate,
- glass pull rate,
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 93

- video of the batch movements,


- velocity at the walls (zero),
- periodical movement of the feeder pusher,
- speed of the feeder screw,
- mean velocity of the blanket after the feeder.
The above information is more or less readily available. However, this is
not the case for more detailed information regarding the batch properties,
batch reactions, melt fluid dynamics, and radiation boundary conditions. As
a result, the modeler has to collect both the general technological data and
the special data. The need for special information will greatly depend upon
the complexity and the performance of the batch model.

2.2.3 Theoretical Basis of Batch Modeling

Principal Mathematical Approaches

In the melting of glass raw materials, the conservation of mass, energy, and
momentum is physical reality, as in many other thermochemical processes.
Any mathematical model of batch melting has to integrate this basic condi-
tion. However, a number of different approaches exist for deriving appropriate
conservation equations for batch melting.
Shell Balance Approach. The batch fusion is considered as a whole and is
studied using shell balances for energy and mass. The heat flux or energy
balance is a well-established tool to evaluate the performance of a glass tank
(see for example [2.121]). Such balances often treat the furnace as a whole,
without considering the details of the heat and mass transport for subsections
in the tank.
However, it is particularly difficult to set up such balances for subsections
such as the batch-melting zone. Figure 2.29 illustrates schematically the dif-
ferent heat fluxes that transport the combustion energy to the upper and
lower batch sides. Part of this energy is directly transferred by radiation to
the batch surface. The remaining energy is first absorbed by the melt and
then transported by convection under the batch.
A simple energy balance can be set up for the melt and batch space.

rnelt
Q surf + Qbatch
surf
= Hbatch
react
+ Qbatch
heat
+ Qrnelt
heat
+ Qwalls
cond . (2.59)

All combustion energy transferred at the batch and melt surface corresponds
to the sum of the reaction enthalpy, the sensible heats of batch raw materials
and glass melt, and the wall losses. The principal difficulty now consists in
..
t h e d etermmatlOn 0 f t he m IVI ua1 contn·b u t·IOns 0 f Qrnelt
. d··d surf an d Qbatch
surf. I n
general, there are no physical separations of the combustion space that allow
both values to be distinguished. There is no simple access to the lower- and
upper-side batch heat fluxes by shell balances.
94 2. Melting and Fining

a batch
Batch surf
with Tin J/J/

Glass-melt convection

Wall losses

Fig. 2.29. Heat fluxes from combustion t o batch and melt space

Continuum Approach. The batch is considered as a fluid in a model based


on the continuum concept. In the continuum approach, individual particles
are ignored and the batch is assumed to consist of continuous matter. At
each point in the batch, one supposes a unique value for velocity, density,
temperature, and other field variables. The continuous matter "batch" is
then required to obey the conservation laws of mass, momentum, and energy.
This leads to a set of differential equations for the field variables, which will
be discussed in the following. The solution of these differential equations
then defines the t emperature and velocity field inside the batch. Moreover,
the shell balances of the model can be established via the integration over
the outer limits of the batch domain. Simplified versions of the continuum
approach consider the batch as solid matter. In such models, the velocity field
is derived from measurements. .
Particle Approach. The batch is considered as a set of discrete particles.
The laws of dynamics and thermics govern the interaction of these particles.
This particle approach is well established in the modeling of granular mate-
rials such as, for example, sand. The macroscopic behavior of the granular
materials is predicted from the interaction of the particles. A particularly in-
teresting feature of the particle approach is that it also yields expressions for
continuum transport coefficients such as thermal conductivity or viscosity.
Quite obviously, the continuum approach is favored in batch modeling.
Nevertheless, care must be taken regarding its validity. The continuum ap-
proach requires that the particle size and distance be very small compared
with the smallest physical length scale of the flow and temperature fields
(such as the thickness of the heat penetration layer or the reaction layer).
Only in this way can meaningful averages over the particles be made and the
particle structure of the batch be ignored. The condition for the continuum
approach to be valid is

_1 « L3 , (2.60)
n
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 95

where n is the number of particles per unit volume and L is the smallest
significant length scale of the flow field. Assuming a thickness of only a few
centimeters for the chemical reaction layer (i.e., L = 1 em) and a batch mix-
ture essentially composed of grains of 0.5 mm diameter, we find
1 .~
4
10 cm-
3 « 1 cm' . (2.61 )

Hence, the continuum approach is valid for batch raw materials composed of
grains < 1 mm.
Next, one assumes a particle diameter of 1 cm, which is a typical size for
cullet and pellets. If the smallest length scale in the batch blanket is still
1 em, one gets
1
--- = 1cm3 . (2.62)
1cm- 3
Now the basic scale condition for the continuum approach is no longer ful-
filled. For batch made up of cullet or pellets, careful analysis of the particle
size and the flow and temperature fields is a prerequisite before any applica-
tion of the continuum approach.

Conservation Equations

If we decide in favor of the continuum approach, we have to identify the


appropriate forms of conservation equations for mass, energy, and momen-
tum. These equations locally describe the change in velocity, temperature,
and density with respect to time and position in the system. Mathematically
speaking, these equations are partial differential equations with terms for
change with time, for convective and diffusive fluxes. Quantity A, at a fixed
position in space changes according to

8A
at = \7)
.A .A
+ \7) diff .
cony (2.63)

Here, 8A/8t refers to the partial time derivative of A with respect to t holding
x, y, z constant. Under stationary conditions, the time derivative becomes
zero. The divergence \7 jA represents the local flux balance of A for a discrete
volume.
Such partial differential equations describe the conservation of mass, heat,
and momentum flux of fluids in a space-fixed framework. The space-fixed
framework, also referred to as a Eulerian coordinate system, is preferred to
solve the majority of fluid-flow problems.
In batch modeling, the formulation of a substantial time derivative fre-
quently proves to be very useful.

(2.64)
96 2. Melting and Fining

Here, we follow the movement of the batch blanket and look at the change of
quantity A. This formulation is particularly favorable if the batch moves in
only one direction at constant velocity, supposing that all other movements
are negligible. As a result, the convection term in (2.63) can be dropped.

DA .A
Dt = \7Jdiff . (2.65)

Such partial differential equations describe the conservation of mass, heat,


and momentum flux of fluids in a moving framework [2.122]. The moving
framework, also referred to as a Lagrangian coordinate system, nevertheless
offers many advantages in batch modeling if the problem can be reduced to
mass and energy conservation.
In the following discussion of the basic conservation equations in batch
modeling, we refer to the Eulerian coordinate system with the balances Over
volume elements fixed in space. But first of all, let us study the relations of
melting modes and mathematical frameworks.

Batch-Melting Modes and Mathematical Frameworks

In Fig. 2.30a-f, the principal batch-melting modes are represented. Pot exper-
iments are the simplest batch-melting systems from the technological point
of view. The unreacted batch and the surface and bottom melt are stationary
except for some internal convection and a vertical movement due to density
variations during melting. However, the progress of the reaction layer renders
the pot experiment transient. Transient heat and mass transfer problems are
already quite challenging.
If the reaction layer is moving, the solution of the partial differential equa-
tion for heat transport involves a moving boundary. The heat transport in
solids in connection with phase transformation is classified as a so-called Ste-
fan problem [2.123]. Modeling a pot experiment can become so complicated
that, finally, numerical methods are needed for the simulation and interpre-
tation of the results.
The caSe of a cold-top electric melter with vertical continuous feed is rep-
resented in Fig. 2.30b. Here, we have of course only one reaction layer. Batch
core and reaction layer are stagnant in the horizontal direction. This fact sup-
ports the model set up in a space-fixed framework. CFD models of the melt
convection of glass tanks are commonly space-fixed. An identical framework
for batch and melt models simplifies the coupling of both models. The mod-
eling of cold-top electric batch melting is of limited complexity and therefore
allows even an approach by analytical methods. However, care has to be taken
with this and all the following examples regarding the bottom melt in contact
with the lower reaction layer. The melt heated by combustion or electrical
boosting reveals quite strong convection. This convection frequently shows
complicated flow patterns that render impossible any simplified theoretical
description.
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 97

a) Pot experiment: transient operation


Model framework: space fixed
Result: transient heat penetration with moving reaction layer
Glass surface
Surface melt layer stagnant except density variation

~!~!J:;J::l;J::·J::"J:!J:: Reaction layer moving position


vx~ 0
vy=o Batch core stagnant except density variation

J:iJ:iJ:iJ:itiJ:iJ:iJ:iJ:: Reaction layer moving position

Bottom melt stagnant except internal convection

b) Cold-top electric: stationary operation


Model framework: space fixed
Result: stationary heat penetration with fixed reaction layer

Continuous feed by vertical charging

Batch surface

Batch core stagnant except feed velocity

Reaction layer fixed position

Bottom melt moving with strong convection

c) Closed blanket: stationary operation


Model framework: moving
Result: transient heat penetration with moving reaction layer

Glass surface
Surface melt layer stagnant except density variation
::'!''::'o::':'!:'l';.~\~~-:}:~
Reaction layer moving position
vx ~ 0
Vy =0 Batch core stagnant except density variation
NN;;'-;!J."":J.":J.-::'-: Reaction layer moving position

Bottom melt moving with strong convection

Fig. 2.30. Operation modes and frameworks of different melting systems. (a) Pot
experiments refer to laboratory-scale crucibles as well as industrial melting pots.
(b) Cold-top electric melters with continuous vertical charging develop closed blan-
kets with negligible lateral movements. (c,d) Closed blankets refer to continuous
feeding and heating in conventional glass tanks. (e,f) These tanks develop dispersed
blankets or batch piles for discontinuous feeding or heating
98 2. Melting and Fining

d) Closed blanket: stationary operation


Model framework: space fixed
Result: stationary heat penetration with fixed reaction layer
Glass surface

Surface melt layer moving

Reaction layer fixed position


Batch core moving
Reaction layer fixed position

Bottom melt moving with strong convection

e) Dispersed blanket: transient operation


Model framework: moving
Result: transient heat penetration with moving reaction layer
Glass surface
vx~ 0 Surface melt layer stagnant except density variation
vy~ 0 and down running
•o:;:-:!-"];t1'~":-::~·
Reaction layer moving position
vx~ 0
vy=o Batch core stagnant except density variation

Reaction layer moving position

Bottom melt moving with strong convection

f) Dispersed blanket: transient operation


Model framework: space fixed
Result: transient heat penetration with moving reaction layer

Glass surface
Vx;t 0
Vy ;t 0 Surface melt layer moving
':~;?;;.~r·;?:···· Reaction layer moving position
Vx;t 0
Vy ;t 0 Batch core moving

Reaction layer moving position

Bottom melt moving with strong convection

Fig. 2.30. continued

Closed-blanket melting with continuous feed and heating is already more


challenging than the cold-top case. Figure 2.30c represents a framework mov-
ing with the feed velocity of the blanket. The batch core and the surface melt
are stagnant with reference to the model framework. As for the pot experi-
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 99

ment, the reaction layers are moving towards the batch core. However, the
recirculation of the melt in glass tanks induces strong convections under the
batch blanket. This marks a significant difference between pot experiments
and melting of blankets in glass tanks. Nevertheless, blanket models with
moving reference frames are attractive because the pot results can be easily
transferred to furnace-scale melting.
Closed-blanket melting can also be modeled in a space-fixed framework
(see Fig. 2.30d). Now, surface melt and batch core are moving with the feed
speed, and the reaction layer is fixed in position. The coupling of such a
model with the melt model is easily feasible. The rheological behavior of the
batch blanket can, in principle, be taken into account.
A dispersed blanket represents the most complicated melting mode. Here,
discontinuous horizontal feeding leads to closed blankets but with varying
thickness. In the course of the melting, the thicker sections have a longer
lifetime and end up as separated island8. Varying thickne88, island creation,
and island melting are extremely challenging for mathematical modeling.
In Fig. 2.30e, we consider the dispersed blanket in a moving framework.
Surface melt and batch core remain stagnant, except for vertical movement8
due to den8ity change8. Furthermore, melt flow on tilted surfaces has to be
taken into account for 8maller structures with strong thickness variations. In
the zone of the more or les8 cl08ed blanket, one could a88ume a constant, mean
blanket velocity and direction for the definition of the moving framework.
This approach is no longer valid for isolated islands, wh08e movement8 often
seem random in batch videos, but in fact are largely determined by melt
convection.
Figure 2.30f shows island melting in a space-fixed framework. In such a
space-fixed model for di8persed blankets, surface melt, batch core, and reac-
tion layer are moving in space and time. As a consequence, the whole coupled
model has to be transient, which requires additional computer resources for
the already "greedy" stationary glass-tank models. Direct coupling of the
flow and temperature fields of the batch and melt model is compulsory if
quantitative information about island melting is required.
In conclusion, moving frameworks are particularly interesting if the heat
transfer in the reaction layer and batch core is principally addressed. Further-
more, pot experiments can easily be used for model set-up and validation.
Fixed-framework mode18 allow direct coupling of heat, mass, and momentum
transfer between the batch and the melt. However, in the moving as well
a8 in the space-fixed frameworks, quite complicated systems of heat- and
mass-conservation equations arise. These equations will be discussed in the
following.

Mass Conservation

Mass conservation seems to be the most basic and simple condition in batch
modeling.
100 2. Melting and Fining

up
ut = -\lpv . (2.66)

For stationary conditions and constant density, the mass conservation reduces
to

p\lv =0 . (2.67)

However, constant density is a very rough approximation in batch modeling,


for two reasons. First, raw materials and cullet undergo compaction to about
half of their granular density during fusion, thus drawing out the air in the
void spaces. Second, the raw material loses about 20% of its initial mass by
the release of reaction gases, mainly CO 2 •
Consequently, a more general mass-conservation equation has to be taken
into account in batch modeling:

at "
UPi = -VPi V + d'IV [D Pi] +ri·
P ij\lp (2.68)

The density change of the component i at a fixed point in space is given


by three factors: convective transport, diffusion in a concentration gradient
(p;j P is the mass fraction of i), and release or consumption of species i in a
chemical reaction.
As already discussed in Sect. 2.2.2 (p. 83 ff.), the assumption of three
species is a first approach to the batch mass conservation. In this first ap-
proach, one assumes the mixture of raw materials and cullet to be the input
species A, the newly formed glass melt to be specimen B, and the released
gas to be C.
It is generally assumed that unreacted batch and the newly formed glass
melt do not mix by penetration or diffusion. Hence, the diffusion terms for
batch and glass melt can be dropped.
The gas transport is very important to batch fusion for several reasons.
The gas, which penetrates through the open pores of the batch, contributes
to the energy transport. The local pressure of the released gas has an impact
on the progress of the batch reaction. Finally, part of the reaction gases is
captured in gas bubbles during fusion. This effect is of considerable impor-
tance for the final glass quality, and its theoretical description is extremely
difficult!

Melting Reactions and Kinetics

As outlined previously, a detailed numerical simulation of the melting kinet-


ics on the molecular and on the grain scale in the framework of a full-size
batch model is actually not feasible. For this reason, one has to rely on phe-
nomenological models of the batch-melting kinetics.
The simplest approach to represent the melting reactions is to use a first-
order global chemical reaction of type
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 101

Abatch --+ Bmelt + Creleased gases tlHreaction. (2.69)

The determination of the enthalpy for such a global fusion reaction presents
no major difficulties and is extensively dealt with in the literature (see
Sect. 2.2.2, p. 77). However, care must be taken in the mathematical def-
inition of the reference mass for the reaction enthalpy (batch or melt). The
reaction enthalpy then enters as a drain term into the energy conservation
equation.
The progress of the reaction is determined by the kinetics on the grain
scale. The simplest approach is a first-order reaction kinetics, according to

d[Abatch] - k [A ] (2.70)
dt - - 1 batch·

Of course, this implies the determination of the rate constant k 1 . Here, the
simplest approach is the assumption of an Arrhenius law for the rate constant.

(2.71)

Owing to this exponential temperature dependence, the temperature field


that results from the energy-conservation equation determines the reaction
rate. In this way, reaction kinetics and mass and energy conservation are cou-
pled in a nonlinear manner! Additionally, reaction kinetics for batch melting
follows more complicated relationships than the simple first-order kinetics
(see Sect. 2.2.2). Advanced batch models should at least account for reaction
kinetics based on phenomenological studies.

Energy Conservation

What is the appropriate form of the energy-conservation equation for a batch-


melting process? First of all, energy conservation in batch melting refers to
the internal energy and comprises both the thermal and chemical energy of
the batch constituents. Heat transport by convection and conduction, as well
as a volumic drain or source, have to be taken into account in a batch-melting
model. The change of the internal energy for a fixed position is given by the
partial differential equation

8H
at = \7 pvcpT + \7(->.gradT) + S . (2.72)

During batch melting, heat conductivity>., density p, and heat capacity cp


vary significantly as a function of temperature and reaction progress. This
renders simplified solutions of this partial differential equation quite diffi-
cult. The non-stationary term can only be dropped if the batch blanket is
stationary.
102 2. Melting and Fining

The source or drain term S is of particular importance in batch model-


ing. This term can be composed of different contributions. The endothermic
reaction of the raw materials is represented by a drain term. The heat for the
water evaporation can be integrated by a further drain term. Other contri-
butions to the source term are given by additional energy transport by gases
in the porous batch, dissipation of electric boosting currents in the primary
formed melt, and radiative energy transport in the case of clear glass melts
and cullet.

Radiation

The radiative energy transport in semi-transparent media such as glass melts


is, generally, quite a challenging problem. The physics of the radiation trans-
port is described by the equation of transfer. The change in radiation intensity
i)., over the distance S is given by

(2.73)
Wi

The right-hand terms describe loss by absorption, gain by emission, loss by


scattering, and, finally, gain by scattering. All coefficients of this equation are
wavelength dependent, a fact not to be neglected for glass. First, the solution
of this integro-differential equation generally needs advanced numerical algo-
rithms. Second, the radiation characteristics of the glass melt (e.g., spectral
absorption and emission coefficients) and the spectral scattering coefficients
related to the number and size of scattering centers (e.g., bubbles and sand
grains) must be known. Third, the boundary conditions for reflection, re-
fraction, and absorption at the three interfaces: 1. primary melt to reacting
batch, 2. primary melt to melt in the tank volume, and 3. primary melt to
combustion space, have to be known and defined.
Because of the complexity of radiation simulations, the Rosseland ap-
proach is frequently used for glass-melt modeling. In this approximation, an
effective thermal conductivity coefficient is defined by [2.124]

Aeff = Aconduction + Aradiation (2.74)

with

(2.75)

In this way, the apparent Fourier heat conduction can be calculated from opti-
cal material constants. n is the refractive index, O"SB is the Stefan-Boltzmann
constant, and aRoss is a gray absorption coefficient calculated from the spec-
tral absorption of the glass melt.
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 103

Careful dimensional analysis is necessary to verify the validity of the


Rosseland approximation. Mean penetration distances can be defined for the
processes of absorption and scattering.
1
lms = - . (2.76)
, (YSA

Via the optical thickness or opacity"" the behavior of a semi-transparent


material can be characterized. The thickness h of the primary melt layer
being in the range of cm, we can compare

• clear glass: '" = h/lm,a'" « 1,


• opaque glass: '" = h/lm,a'" ~ 1,
• primary melt: '" = h/lm,s"" > l.
For a clear glass melt, most of the incoming combustion radiation would
penetrate a thin surface layer (in particular if one considers the spectral
characteristics of high-temperature radiation and the glass-melt absorption).
Opaque glass melts already absorb a significant part of the radiation even at
distances in the cm range.
Primary melt with a high number of gas bubbles and grains represents
an opaque media due to significant scattering of the radiation. The particle
diameter D in relation to the wavelength determines the scattering mech-
anisms. For A ~ D, Mie scattering predominates. For A « D, Fraunhofer
and Fresnel diffraction plus refraction must be taken into account. The typi-
cal radiation wavelength range in glass melting is 0.55 11m. With sand-grain
diameters smaller than 500 11m, gas bubbles ranging from a few 11m to mm
(larger bubbles have only a very short lifetime in glass melts), and cullet
pieces in the range of mm to cm, essentially scattering by diffraction and re-
fraction must be taken into account. Because of the difference in the refractive
index, gas bubbles have the most significant impact on radiation scattering
in the primary melt. A study on the extinction coefficient as a function of
bubble diameter and concentration can be found in [2.125].
In conclusion, particle scattering in the primary melt significantly simpli-
fies radiation modeling. However, the effective thermal conductivity and the
surface absorption and emission of the primary melt become dependent on
the number and size of the gas bubbles. For a high number of bubbles one
enters the physics of heat transfer in foam.
A further problem arises for more or less clear glass melts. For a decreas-
ing number of gas bubbles, the Rosseland approximation is no longer valid.
A radiation simulation including melt absorption and particle scattering is
necessary.
In this article, we will not further discuss radiation modeling in glass and
the reader is referred to [2.126] and SectS. 4.1-4.3.
104 2. Melting and Fining

Momentum Conservation

What is the appropriate form of the momentum conservation equation for a


batch-melting process?

ata pv + [V pvv] = -vp - VT + pg . (2.77)

In the case of stationary batch melting, the time derivative of the momen-
tum change can be omitted. The term of the momentum gain by convection
is negligible for the low velocities in standard batch-melting processes. How-
ever, the pressure gradient, the viscous friction, and the gravitational force
are important terms for the batch motion. The momentum change by viscous
transfer presents a particular problem in batch modeling. For glass melts at
high temperature and with low shear rates, Newtonian flow behavior is a
convenient assumption.

T = JLVV . (2.78)

For a batch blanket, this assumption is no longer valid. Batch is a granular


material comprising sand, various mineral powders, and cullet. Granular ma-
terials show many rather astonishing phenomena. When shaken, sheared, or
poured, they show segregation, convection, and spontaneous fluctuations in
density and stress. A wealth of theories and modeling approaches has been
developed to describe the behavior of granular materials; see [2.127,128].
During melting, the batch undergoes complex solid-liquid phase transi-
tions, which make rheological description even more complex. The primary
melt, which emerges from the fusion, is still highly inhomogeneous. A high
number of undissolved sand grains and numerous gas bubbles are present in
the first glass melt. At high grain and bubble concentrations, the primary
melt could be described as a slurry-type fluid. However, no studies have been
published on this subject so far.

Batch Properties for the Continuum Approach

The type and complexity of the chosen model largely determine which ma-
terial properties of the batch and which parameters and data of the melting
process must be known.
Thermal-balance models require specific heats and reaction enthalpies.
Heat-flux models additionally require thermal conductivity, diffusivity, den-
sity, and appropriate parameters for the chosen boundary conditions (heat
transfer coefficients, emissivities, etc.). If reaction kinetics is to be modeled,
kinetic parameters are needed. Models predicting the flow behavior of the
batch require rheological data. The TC 21 selected for the round-robin test
four (RRT 4) typical data for soda-lime silica batch melting; the chosen values
for batch and melt are compared in Table 2.4.
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 105

Table 2.4. Typical properties of batch and melt as defined by the TC 21 in the
RRT4
Property Batch Melt
Cp 1000J/kgK 1300J/kgK
P 1400kg/m 3 2300 kg/m 3 (1400°C)
A 0.25W/mK « 500°C) 30W/mK
H chem 750kJ/kg

The apparent density of powdered raw materials for soda-lime silica glass
is of the order of 1400kgjm3 . Cullet and glass composition and grain prop-
erties influence the apparent density. The final glass melt has a density of
about 2300 kgjm 3 . During the fusion reaction, the density undergoes signifi-
cant variations, in particular if strong foaming occurs.
But the most significant differences are observed for the thermal conduc-
tivity. Starting with a value < 1 W jm K for the unreacted batch, the thermal
conductivity increases by about one order of magnitude for the final melt.
The high values for the effective conductivity coefficient are due to the ra-
diative contribution in glass melts. The RRT 4 will be further discussed in
Sect. 2.2.5 on p. 124.
Valuable information about density, heat conductivity, and heat capacity
relating to different, unreacted batch materials is widely available in the liter-
ature. However, it is extremely difficult to identify any studies on the rheolog-
ical behavior of granular batch mixtures, and data regarding the properties
of the reacting layer and primary melt are scarce as well.

Final Remarks on the Continuum Approach

The continuum approach requires a very systematic model set-up with the
following steps:
1. The validity of the continuum approach has to be checked.
2. Depending on the batch-melting mode and the chosen framework, the
governing conservation equations must be determined.
3. The material properties, duly representative of the continuum approach
of the batch and melt, have to be determined.
4. Appropriate boundary conditions must be chosen.
5. Finally, the conservation equations have to be solved by analytical or
numerical methods.
Owing to the property changes occurring during batch melting (density,
heat capacity, thermal conductivity, and viscosity), simplified formulations
of the conservation equations are no longer valid. It goes without saying
that assumptions such as the Boussinesque approximation, which assumes
constant density in all terms, except the gravity term, also cannot be used.
106 2. Melting and Fining

For these reasons, numerical solutions are compulsory for advanced batch
modeling.

2.2.4 Key Values and Non-Dimensional Numbers

As shown above, the continuum approach leads to conservation equations


for mass, energy, and momentum. If the corresponding differential equations
are rendered dimensionless, they yield dimensionless numbers. These num-
bers prove to be quite useful for the physical interpretation of characteristic
situations. But before entering into a detailed discussion, let us quickly look
at the typical mass- and heat-flux densities encountered in batch melting.

Mass- and Heat-Flux Density

One key parameter in batch melting is the melting flux density, i.e., the
melting rate per blanket surface, generally expressed in tonnes per square
meter per day. Typical values range from 4 to 7t/(m 2 day) for bottle, TV, or
float glass (see Table 2.3). This specific blanket-melting rate must be clearly
distinguished from the specific melting rate per tank surface. Furthermore,
the specific blanket-melting rate represents a mean value of the mass flow or
melting-flux densities at the lower plus the upper batch sides.
With the energy requirement for melting, typical heat-flux densities can
be determined. As already mentioned, typical values are in the range of 40-
80 kW 1m 2 for the lower or upper blanket side. The different melting stages
of a batch blanket govern the local variations of the heat-flux density (see
Fig. 2.27).

Damkohler Number

The Damkohler number is frequently used in fluid dynamics to compare trans-


port processes to the kinetics of chemical reactions. The diffusion of batch
components compared to the transport being slow, only the Damkohler num-
ber for convective transport is of interest.

Da conv = -.:.::.::.."----
tflow = k~ . (2.79)
treaction v
The characteristic time for reaction treaction is given by the reciprocal value
of the reaction rate constant k. The time for the flow is given by the flow
velocity v and the characteristic length L for the problem. For simplicity, we
will only consider planar ID melting, as in a cold-top melter. (The following
considerations can be applied to most batch-melting systems, if the reaction
layer is taken as a reference position). The Damkohler number is, in principle,
only defined for a homogeneous temperature field over the considered char-
acteristic length. However, if one assumes a negligible temperature variation
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 107

in the reaction layer over a thickness of 1 cm, a Damkohler number can be


defined.
For a melting flux density of 2 tj(m 2 day), a batch flow time of 10 min over
a 1 cm layer thickness results. Figure 2.31 represents this local Damkohler
number as a function of temperature. For Da « 1, the reaction time is
long against the flow time, and the batch stays almost unreacted over the
considered flow length. For Da » 1, the reaction is fast, and the batch leaves
the considered layer almost completely reacted.
Two different reaction rate constants are considered in Fig. 2.31. The data
for the stiff reaction are taken from Kroger's measurements of the reaction
kinetics on sand-soda ash (n = 3) [2.101]. If the temperature increases from
one layer to the next from 850°C (Da = 0.3) to 900°C (Da > 2), the reaction
will be almost completed over a distance of 1 cm.
The data for the soft reaction are taken from a numerical batch simulation
by Ungan and Viskanta [2.107]. If one assumes the same temperature gradient
as in the former case (50°Cjcm), the Damkohler number increases slowly
layer by layer (0.2, 0.3, 0.7, 1.5, ... ). Consequently, the speed of the reaction
increases only moderately, layer by layer, and the reaction will extend over a
number of discrete layers with the chosen characteristic length of 1 cm.
In addition, the Damkohler number allows the temperature of the reaction
layer to be estimated as a function of the reaction rate constant. Here, the
reaction-layer temperature turns out to depend not only on the reaction
kinetics but also on the flow velocity, respectively on the local melting flux
density. Higher melting values shift the reaction layer to higher temperatures.
The quite simple Damkohler number reveals, in a nice manner, the cou-
pling of flow field and reaction kinetics. Moreover, the Damkohler number
relates the reaction kinetics to the temperature field of the batch. But how
can we gain insight into the temperature field? This is the problem to be
tackled next.

en
en 2.0
.!!?
E
~ 1.5
Q;
.c
E
:::J
C 1.0
Q;
:c
'0
~
E 0.5
Cll
0
(ij
<..l
0
0
-' 750 800 850 900 950 1000 1050
Mean layer temperature JOG

Fig. 2.31. Damkohler number as a function of mean temperatures of discrete layers


(ll.x = 1 cm) parallel to the reaction layer. The reaction kinetics has a considerable
impact on the reaction temperature and reaction-layer width
108 2. Melting and Fining

Peclet N mnber

The Peclet number expresses the ratio of time for diffusive heat transport to
convective heat transport.

Pe = tdiff = vL . (2.80)
tflow a
For Pe < 1, heat conduction dominates, for Pe > 10, heat convection domi-
nates.
For simplicity, we consider again a planar 1D melting situation. Here, one
can assume that the melting-front velocity is equivalent to the flow velocity
of cold batch materials through the heating zone towards the reaction layer.
This situation corresponds exactly to the mixed convection-diffusion heat
transport that can be characterized by the Peclet number [2.96].
It is again useful to consider a simple example. The characteristic length
is given by the half value of the batch core thickness (here L = 10 cm). A
typical value for the thermal diffusivity of the batch is 3 x 10- 3 cm 2 Is. For
a local melting-flux density of 2 t/(m 2 day), a Peclet number of 5.5 results.
In Fig. 2.32, the temperature profile for this case is shown. The hot-side
temperature is given by the reaction temperature (here 900°C). The batch
core is assumed to be still cold (i.e., RT). For Pe = 5.5, the influence of the
convective heat transport is already dominant over the considered domain.
The heat-flux density at the hot side of the profile is about 25 kW 1m 2 . The
heat-flux density of 0.1 kW 1m 2 at the cold side is negligible.
For higher melting-flux density such as 4 t/(m 2 day), the hot-side heat-flux
density doubles to 50 kW1m 2 as would be expected from a simple thermal
balance and Pe = 11. At the same time, the thickness of the heat penetra-
tion layer decreases to about 5 cm. For a new characteristic length of 5 cm,

1000

800
0
e....
~ 600
::J

Q) 400
c.
E
- - 4 t (m2 d) Pe = 11
~ 200 eee 2 t (m 2 d) Pe = 5.5

0
0 0.02 0.04 0.06 0.08 0.1
Distance from reaction layer 1m

Fig. 2.32. Peclet number and temperature profiles in the batch core before the
reaction layer. The local melting rate has a considerable impact on the temperature
profile
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 109

Fe = 5.5 results. The intermediate value of Fe = 5 thus allows one to esti-


mate rapidly the thickness of the heat penetration under the reaction layer.
However, it gives no information as to whether the high heat-flux densities of
50 kW 1m 2 can be attained in the lowest layer of the complete heat-transfer
chain of the batch-melting process.
The characterization of the heat penetration by the Peclet number is
particularly valid for melting with advancing reaction layers in the zone of
the still cold batch (see Fig. 2.27). Similar approaches for the temperature
profiles and heat fluxes of the other melting stages can be developed, but will
not be further discussed here.

Biot Number

The Biot number expresses the ratio of the heat transfer coefficient to the
internal heat conduction of the batch.

Bi = hL
k . (2.81 )

For the combined heat transfer by radiation and conduction of the combus-
tion gases, a quite high surface heat transfer coefficient results. With typical
batch dimensions and the low thermal conductivity of the batch core, Bi » 1
results. This indicates that the batch core temperature will stay inhomoge-
neous for a long time.

Fourier Number

The Fourier number is a measure of the rate of heat conduction in comparison


with the rate of heat storage in a given volume element.
cd
Fo = L2 . (2.82)

As already outlined, the batch is a poor thermal conductor. Furthermore,


it is characteristic for batch-core heating that a significant amount of the
conducted heat is at once stored as sensible heat. A heating time of the order
of hours is required to heat the batch core for a typical blanket geometry
and to increase Fo to values greater than 1. However, due to the melting
process, the characteristic batch thickness L decreases, thus reducing the
heat-up time.

Heat Transfer Coefficient

The heat transfer coefficient h is, strictly speaking, not a dimensionless num-
ber arising from the conservation equations. h is commonly defined by
110 2. Melting and Fining

h = -:-=--q=----_=_----,- (2.83)
(Tinterface - T bu1k ) ,

where q is the heat-flux density at the interface between two domains, for
example between the batch core plus reaction layer and the underlying melt.
The interface temperature Tinterface is, of course, identical to both domains for
systems near thermodynamic equilibrium. The temperature Tbulk is given by
the volume temperature of the domain that is not simulated but represented
by a boundary condition. This temperature is, for example, difficult to define
for the melt convection under the batch. Figure 2.21 illustrates that the melt
temperature is strongly position dependent.
One should also keep in mind that the heat transfer coefficient is not a
constant characteristic of a fluid medium. On the contrary, the heat transfer
coefficient depends in a complicated manner on many variables, including the
fluid properties, the system geometry, the flow velocity, and the characteristic
temperature difference (see literature on fluid dynamics, e.g. [2.122]). Local
and mean heat transfer coefficients have to be clearly distinguished. Radiative
contributions to the heat transfer coefficient must be handled very cautiously
because of their power-4 dependence on temperature.
In general, heat transfer coefficients are used in simple batch models for
the calculation of interface heat fluxes, thus avoiding tedious calculations of
the heat transport with conservation equations. Another application consists
in the system analysis by dimensionless numbers, such as the Biot number
or the N usselt number.
Normally, heat transfer coefficients must be determined experimentally.
But the difficulty in measuring the heat-flux density and the temperatures
at the lower and upper side of a batch blanket in a melting tank renders the
determination of local heat transfer coefficients nearly impossible.

2.2.5 Batch Models

Examples

A considerable number of papers deal with batch modeling on the furnace


scale. In the following we try to give a representative overview of the numerous
different batch-modeling approaches. It is not the deliberate intention of the
author if an important batch model contribution has been overlooked.
Pugh 1968
In 1968, Pugh [2.129] already published a model for batch melting that com-
prises some of the essential features of many later models. Based on the
thermal conductivity measurement of Koger and Eliehausen [2.130]' Pugh
assumes that unreacted batch behaves as a nearly perfect insulator. The
batch reaction occurs at a fixed temperature. The heat 6. Hfusion, which com-
prises the processes of batch heating and reaction, enters into the batch and
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 111

is absorbed in a very thin layer. The thickness L of the melt layer formed on
the batch surface is related to the melting time according to

(2.84)

where >"(T) is the thermal conductivity and p is the density of the melt layer.
The time needed to form a layer of melt increases, therefore, as the square
of the layer thickness. Thus, this moving-framework model, in principle, sat-
isfies the basic model demands "length extension" and "energy consumption"
(see introduction). The down running of the surface melt on batch ridges as
formed by screw feeders is also considered by Pugh. Some basic conclusions
are drawn concerning the impact of batch composition and reaction enthalpy
on melting times. General experimental observations are cited to corroborate
the theoretical approach. The major drawback of this simple approach is the
negligence of the lower batch face melting and of the different stages of the
surface melting (see Fig. 2.27).
Trier and Voss 1968
The glass tank and batch model developed by Trier and Voss is summarized
in [2.120]. The model is based on energy sub-balances for melt, combustion,
and batch zone. The total quantity of heat for batch melting Qbatch is trans-
ferred to the batch by combustion according to

(2.85)

The factor 0 < 'P < 1 takes into account that the heat transferred at the
upper batch side is diminished by the heat-flux contribution of the melt
recirculation to the lower batch side (see Fig. 2.29). The authors admit that
there are some difficulties in determining the heat transmission coefficient k,
and, by consequence, the factor 'P. It is clear that 'P cannot be determined
in an independent manner without a melt-convection model. Nevertheless,
the global tank model of Trier and Voss gives general insights into fuel and
specific energy consumption, specific melting rate, glass depth, regenerator
efficiency, and crown temperature.
Fuhrmann 1973
Fuhrmann [2.123] studied in detail the melting of batch blankets and islands
in a moving framework, which leads to the transient form of the energy
conservation equation. The 1D heat transport in the batch core is solved for
temperature-dependent properties and including the internal heat transport
by reaction gases. Furthermore, endothermic reactions by batch additions are
taken into account by a volumic source term.
However, the reaction layer is again assumed to be of negligible thickness
and of constant, fixed temperature. The heat flux from the melt and com-
bustion to the batch model is addressed by heat transfer coefficients. The
112 2. Melting and Fining

formed melt is currently removed from the reaction layer. Despite these sim-
plifications, the underlying differential equations remain complex and require
already a numerical solution.
With this model, principal conclusions can be drawn on the core heat
penetration and reaction front for various shapes of batch blankets and islands
and for varying batch properties. The principal effect of surface-melt-layer
formation and of its foaming is discussed. Unfortunately, a comparison with
the few published measurements on batch melting proved to be difficult.
Austin and Bourne 1973
Austin and Bourne [2.131] set up a complete model for a cold-top electric
melter comprising a 2D CFD melt model and a 1D batch-melting model.
The batch model assumes again that the batch reaction takes place at a
fixed temperature in a layer of negligible thickness. But now, the heat flux
into the batch core is addressed by an analytical solution of the 1D energy
conservation equation:

T(x) - Tsurf 1 - exp(/3 x)


(2.86)
Treact - Tsurf 1- exp(!3l) ,

where x is the vertical distance from the lower batch face, l is the batch
thickness, and /3 = Tn cp /).. (with Tn as the mass-flux density related to the
melting rate). The internal batch core heating by the released reaction gas is
taken into account by a corrected value for the specific heat cpo The thermal
conductivity ).. and the specific heat of the unreacted batch were assumed
constant.
The resulting, nonlinear temperature profile is characteristic of mixed
convection-conduction heat-transport problems (see also the profiles in
Fig. 2.32). The assumed reaction and melting temperature of 1277 °C seems
somewhat too high for the modeled melting of flint, amber, and borosilicate
glasses. However, the authors mention an experimental verification of the
vertical temperature gradients, but unfortunately do not present any data.
Leyens 1974
Leyens [2.132] developed a 2D CFD glass-melt model of a continuous-melting
tank. The batch melting is represented as a boundary condition for the melt
model. Local thermal balances for the batch are established on the basis of a
predefined melting rate. The heat transfer from the combustion to the batch
is calculated by a heat transfer coefficient.
The local heat-flux density from the melt to the batch has to complete
the batch heat demand according to the boundary condition
lowerside total u pperside
qmelt = qbatch = qbatch - qbatch . (2.87)

The CFD melt model then results in a boundary temperature of the melt to
satisfy the heat-flux boundary condition. The boundary temperature differs
from the melting or reaction temperature. This is one of the typical problems
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 113

of simplified batch models, which do not account for the complex coupling of
mass flow and heat fluxes of melting batch blankets.
Mase and Oda 1980
Mase and Oda developed a 2D CFD melt model coupled with a combustion
model based on Hottel's zonal method. In [2.133], this tank model was com-
pleted by a batch model. In the combined model, the batch is considered as
a succession of more or less separated rifts. The melt between the rifts in
covered by foam. The rift length and distribution has to be determined by
observations. The batch, or better, the rift velocity is given by the feed rate
and the initial rift thickness.
The batch model assumes a fixed reaction temperature, an infinite reac-
tion speed, and, by consequence, a zero thickness reaction layer. The batch
core heating is represented by a 1D heat transfer in batch depth. The equa-
tion given for the energy balance can be transformed with the batch velocity
Us = ax/at to

a oaTs
{)tPscsTs = ay ks ay + H chem (2.88)

Ps, cs, ks, and H chem represent the batch core density, specific heat, heat
conductivity, and reaction heat demand.
This model hence represents the batch in a moving framework with tran-
sient 1D heat penetration into more or less separated islands. The boundary
conditions at the lower and upper batch sides are given by the equality

qext aTsl
= ks - - (2.89)
ay interface

The external heat-flux density qext is available from the combustion radia-
tion or melt-convection model. The authors admit that the resulting batch
interface temperatures TB are virtual (Ts does not match Tmeld. Moreover,
the initial model condition that states a fixed reaction temperature does not
coincide with the distributed source term H chem in the energy balance equa-
tion.
Nevertheless, the whole 2D tank model was validated with temperature
and flow measurements on a small float-glass tank. For the batch model
validation, only the batch-length extension was measured.
Hrma 1982
Hrma presents in [2.152] a thermodynamic analysis of 1D batch-melting sit-
uations. On the basis of the continuum approach, local balances of mass and
energy are formulated. The study focuses in particular on the reaction layer.
As an important constitutive assumption, Hrma states that all condensed
material properties are determined by the temperature and the degree of
conversion ~

cp , p, >.. = f(~, T) (2.90)


114 2. Melting and Fining

The rate of conversion is subject to a kinetic equation

a~
at = a(~, T) . (2.91 )

This kinetic equation might be constructed by multiple successive reactions of


the Arrhenius type. However, for a fixed temperature history, the conversion
can be expressed by a pure function of temperature

~ = g(T) . (2.92)

This function can be experimentally determined for a prescribed temperature


history T = f(t). If this temperature history is similar to the one experienced
by a batch particle during melting, the constitutive equation for the material
properties becomes a sole function of temperature

Cp , p, .x = f(T) . (2.93)

Experimental data are, however, useless for temperature histories different


from those used for the measurement of the degree of conversion.
Hrma then considers a stationary melting of a batch blanket from below
(corresponding to the cold-top case). The velocity, temperature, and degree
of conversion fields are analytically solved for this 1D system. With some
simplifying assumptions (~ is a linear function of temperature, all properties
are constant in the reaction layer), quite simple analytical expressions are
found for temperature T, heat flux q, and thickness h of the reaction layer.
The reaction layer thickness is given by

h = .x Tend - Tstart In ( qend) . (2.94)


qend - qstart qstart

The index "start" labels the cold, upper side of the reaction layer and the
index "end" the lower, hot side. The heat flux qstart is given by the preheating
of the batch and the batch surface losses. qend is given by the sum of qstart +
reaction heat + sensible heat for the reaction layer.
For typical melting parameters, reaction layer thicknesses of 5-10 cm are
estimated. These values are not negligible compared to typical batch-blanket
thicknesses, a fact that offends zero thickness assumptions in other batch
models.
Hrma extends his analysis to different, 1D heating situations (i.e., melting
from both sides) and discusses the relations between the rate of conversion,
heat fluxes from below and above, and the rate of pull. However, no vali-
dation data are presented in the paper. The author admits that the model
is too restricted to represent most of the real melting situations. Describing
conversion by a specific function appears to be too crude an approximation.
Thus, a kinetic equation for the reaction seems necessary to simulate the real
melting behavior of the batch.
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 115

Hilbig 1984
Hilbig [2.134] studied the batch melting in a cold-top electric melter. The
solution of the 1D energy conservation equation leads to the same result as
found by Austin and Bourne [2.131].
Discontinuous batch feed causes fluctuations of the temperature field in
the batch core.
T(x, t) = Tstat(x) + 19(x, t) . (2.95)
Hilbig studied the analytical solution of the 1D transient energy conservation
equation to determine 19(x, t). This type of problem arises if the mass units of
the discretely fed batch come into the range of the total mass of the unmolten
batch floating on the melt.
Hilbig and KirmjJe 1986
Hilbig and KirmjJe approach in [2.135] the thickness profile of a batch blan-
ket fed horizontally by the geometry of a circular section. This gives rise to
a batch-core heat conduction equation in cylindrical coordinates. The batch
mass transport is solely included by boundary conditions, which seems some-
what problematic because the convective term in the energy-conservation
equation presents a significant contributions to the batch core heat trans-
port.
Wu and Viskanta 1986
Wu and Viskanta [2.126] set up a 1D model for the transient heating and
melting at the surface of the batch. The paper presents the first study on the
radiative heat transfer in the semi-transparent surface glass layer on melting
batch.
In this study, it is assumed that the opaque batch starts reacting at a
fixed temperature Tm and becomes a semi-transparent melt. Scattering of
radiation by undissolved batch and gas bubbles is assumed to be negligible
for this melt. The batch reacts either completely at T m or over a temperature
interval f}.T. The reaction enthalpy is hence either completely consumed at
the batch-melt interface at T m or distributed over the melt layer thickness.
The thermal model is based on two energy-conservation equations. The
unreacted batch is considered with the Fourier equation comprising a con-
duction and a transient term. For the energy conservation in the glass-melt
layer, the Fourier equation is extended with a term accounting for the density
increase during reaction, a chemical heat sink and a 1D radiative flux term.
The total, local radiative flux in the melt layer accounts for the external,
incident radiation, the emission by the unreacted batch, and the emission
and partial absorption by the molten glass. The spectral glass absorption
is considered with three spectral bands. Furthermore appropriate boundary
conditions were chosen that take combustion radiation and convection into
account.
This system of two energy differential equations coupled with radiation
integrals and moving boundaries has, of course, to be solved numerically. It is
116 2. Melting and Fining

particularly interesting that the numerical nodes were automatically adjusted


to follow the positions of the moving boundaries.
With this batch surface melting model, different cases were studied for a
typical set of batch and melt properties. First, the semi-transparent model
was compared to a model based on the diffusion approximation with an effec-
tive (radiative) conductivity for the melt layer. Significant differences appear
for thicker melt layers where the semi-transparent model yields 20% shorter
melting times for an initial thickness of 4 cm. However, the chosen absorption
coefficients are representative for white container glass. For colored glass, the
model differences should be minor.
Significant differences in temperatures and melting times were also stated
between the localized chemical heat sink at T m and the distributed heat sink.
The authors stress that this result clearly demonstrates the importance of
the kinetic melting model on the conversion of batch to glass.
A validation of this model under well-defined experimental conditions
would be very interesting but is not reported in this paper. An extension of
this study that considers scattering by bubbles and undissolved batch in the
newly formed melt would provide further insight into the heat transfer for
batch melting.
Ungan and Viskanta 1986
Ungan and Viskanta [2.107] presented a detailed batch model, which can
best be described as a 2D CFD simulation of the three phases, solid, liquid,
and gas. The unfused batch core consists of granular solid particles and per-
colating CO 2 gas. The reaction occurs in a transition layer, which consists
of two phases, gas and solid-liquid mixture. The final melt formed above a
transition temperature of 1200 °C is represented by a liquid phase virtually
free of grains or gas bubbles. For all three phases, the conservation equa-
tions for mass and energy are solved in a more or less detailed manner. The
momentum problem is reduced to a unidirectional forward movement of the
batch. The percolation of the CO 2 gas is addressed by Darcy's law. A species-
conservation equation for a simple carbonate + Si0 2 reaction is included in
the model. The reaction rate is calculated using an Arrhenius-type equation
with empirical constants, which have to be determined experimentally. A
spectral, two-flux model represents the penetration of combustion radiation
in the surface melt layer.
For the numerical solution of this system of conservation equations, a co-
ordinate system, which follows the streamlines, is chosen. In this way, the
mesh fits the variable computational domain during the course of the calcu-
lation.
A parametric study of the batch behavior is presented for idealized bound-
ary conditions for the combustion and the melt. Quite steep vertical gradients
for temperature and for the chemical species (carbonate concentration) are
obtained [2.136]. The numerical resolution needed a I-mm grid and matching
step resolution. In the presented case study, about 200000 cells are needed for
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 117

the 2D simulation of a batch with 125-mm entrance thickness and a length


of nearly 2 m.
The model represents a quite complete description of the principal phys-
ical processes for mass and heat transport in batch melting. In [2.136]' this
batch model was coupled with a 3D tank model. The numerical results seem
quite reasonable, but unfortunately no measurements were available to verify
the model.
Schill 1986
Schill presents in [2.137] a batch model for the cold-top configuration. The
reaction-layer problem is considered by an approach slightly different from
Hrma's. First, the batch-melt conversion is represented by a conversion func-
tion ex = f(T) that is deduced from differential thermal analysis (DTA) and
thermogravimetry (TG) measurements. The batch model domain is defined
by the boundary of complete conversion (ex = 1) at a temperature of about
1000°C. The temperature range of the vigorous part of batch reactions typi-
cally extends over 100°C. The measurements have, of course, to be conducted
with the typical heating rates of the batch blanket studied.
The stationary 1D energy conservation equation is considered. Tempera-
ture-dependent properties are taken into account. The source term comprises
the energy of all batch reactions that are identified in the DTA/TG measure-
ments. The energy transport by the released gases is added to the convection
term. A closed analytical solution is no longer possible for the resulting 1D
energy and mass conservation equations. Hence, Schill developed a numerical
and a mixed analytical-numerical solution for this cold-top batch melting.
An electric melter with a capacity of 3 t/m 2 /day for the production
of white container glass is presented as an example. The model predicts
58 kW /m 2 heat-flux density at the batch-melt interface, and a batch thick-
ness of 9.6 cm. Unfortunately, no experimental results could be presented.
De Waal1990
In the framework of a 3D simulation of a glass tank, de Waal [2.114] presented
a quite pragmatic approach to batch-melting kinetics and the rheology prob-
lem. The batch is part of an integral CFD melting space model. The blanket
rheology is assumed Newtonian with an apparent viscosity. The viscosity-
temperature relation is found by fitting to the observed batch shape on the
furnace scale. The melting kinetics is taken from batch-free time experiments.
A set of simplified batch-blanket parameters is defined in such a way that
they fit into the flow model. The best model parameters are then found by
comparing the calculated batch-blanket geometry with the actual one in the
furnace.
This model presents the first approach to the batch problem in the TNO
glass-tank model. It is, furthermore, interesting that the batch model has
already been combined with a reactive sand-grain dissolution model.
118 2. Melting and Fining

Faber, Beerkens, and de Waal1992


Faber et al. [2.154] studied the melting of various glass and batch composi-
tions in a small pot furnace. The heating of the batch layer is followed by
thermocouples at different height positions. The thermal diffusivity was de-
termined by discretization of the 1D Fourier equation. In this approach, the
distance between the neighboring thermocouples is important for the pre-
cision of the calculated thermal diffusivities. However, the values found in
this study agree well with other published values of thermal diffusivities or
conductivities that were determined by different methods.
The thermal diffusivities are then used in a semi-empirical model of the
batch heat penetration. In this model, the batch is considered as an infinite
slab of constant thickness heated from two sides. Below and above a certain
batch transition temperature of typically 800-900 DC, two heating stages are
distinguished. For the first stage, the minimum (inner) temperature 79center
of the batch is calculated according to

(2.96)

The surface temperature is defined by 19surf = (79upper + 791ower) /2. The ther-
mal diffusivity a1 is assumed to be constant, which is a fairly justified assump-
tion for temperatures well below the reaction temperature of about 900 DC.
When, after time ts, the center temperature reaches a certain "batch
transition" temperature, a second heating stage is defined, and the inner
temperature is calculated with an extension of the a t term.

(2.97)

The coefficient a2 takes the increase of the heating rate of the batch into
account that occurs when the batch core exceeds the reaction temperature.
It should be noted that this heating rate increase is due to the increase of the
thermal diffusivity at the reaction temperature as well as to the heat con-
sumption for chemical reaction. The coefficient a2 represents a fit parameter,
which should not be interpreted as a thermal diffusivity.
With this two-stage batch-heating model, the authors fitted the core-
heating curves for a number of typical batch mixtures and heating conditions
in pot experiments. The implementation of this semi-empirical batch heating
model in the complete TNO glass-tank model is not further discussed in the
paper.
Carvalho and Nogueira 1993
Carvalho and Nogueira [2.138] developed a batch model that addresses two
types of melting modes: closed-blanket melting and melting of dispersed is-
lands.
The zone of the closed blanket is modeled by a 2D approach with a solid
batch core and a fluid molten-glass layer that forms on the top of the batch
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 119

blanket. The batch reaction is represented by a phase change at 850°C. An


endothermic single-step reaction is assumed to occur over a finite but small
temperature range. The authors refer to [2.139] concerning the details of the
model.
The zone of the dispersed islands is prescribed by a density function that
represents the distribution of the batch over the melt:
local area occupied by batch unmelted parts
1/J - -----:--:---7----=---:----:-------:-''---- (2.98)
- total local area over the glass melt .
The shape of the distribution function is taken from video observations
and imposed on the model. In order to conserve the overall batch mass bal-
ance, the calculated local melting rates are normalized by the batch inlet
mass-flow rate. This approach to the dispersed, transient batch melting en-
ables us to stay with stationary calculations of batch and melt. The 2D batch
model is coupled with a 3D CFD simulation of the melt.
Schill 1995
In [2.140] Schill extends his conversion-degree-based batch model to three
dimensions. Mass and heat transport are now addressed by 3D stationary
conservation equations. The batch properties (thermal conductivity, viscosity,
density, electrical conductivity) are defined as functions of temperature and
degree of conversion. Additionally, the parameters of the conversion model
have to be defined.
Evidently, numerical methods are required for the solution of this type of
model. The degree of conversion of each numerical node is calculated from
the state of several adjacent modes and an overall balance is calculated in
the batch area.
Modeling examples for front and side charging as well as a cold-top melter
are briefly discussed. Apart from the conversion-degree measurements, the
paper gives no details on the batch and melt properties used. Measurements
on the batch-blanket melting are not presented either. This batch model is
part of the Glass Service CFD tank model.
Lixin and Chengxu 1995
Lixin and Chengxu [2.141] presented a moving-framework batch model based
on thermal and mass balances of blocks moving forward at constant speed.
The progress of batch melting is divided into various steps: temperature ris-
ing, melting beginning, and melting ending. The model assumes a homo-
geneous temperature within each block. Heat input from the melt and the
combustion space is taken into account. The reaction energy is integrated in
the thermal balance of the blocks. Fresh melt from the block surface virtually
penetrates the batch to enter the melt.
The batch model is coupled with a 3D CFD simulation of the melt in a
float-glass tank. The heat transfer from the combustion space to the batch is
described by an experimentally determined radiation heat transfer coefficient.
Unfortunately, the authors present no validation data.
120 2. Melting and Fining

Yamamura et al. 1997


Yamamura et al. [2.142] presented a batch model, which takes into account
both heating from melt and combustion. The thermal problem is addressed
by 3D heat conduction of the batch core and heat transfer coefficients at
the top and the bottom of the core. The reaction energy is not mentioned.
The momentum problem is reduced to a unidirectional forward movement at
constant speed. The removal of molten glass from batch heaps is given by the
heating, during the forward movement of the batch.
Different batch geometries in a container-glass furnace are studied by
linking the batch model with CFD melt calculations. Constant heat transfer
coefficients are assumed for the batch top and bottom. The authors admit
that this heat transfer coefficient is, in reality, variable and a function of the
state of the batch surface. The authors also admit that their model is not
validated because experimental data are lacking.
Beerkens 1998
In a review on the possibilities and performances of the TNO glass tank
model, Beerkens reveals in [2.143] a few details on the implemented batch
model.
The batch blanket is assumed to be flat and is represented by a single
layer. One could presume that the thickness ofthis layer is constant and given
by the batch-entrance thickness in the doghouse. This assumption allows one
to apply the semi-empirical model for the slab heat penetration used for the
previously described pot experiments. For the batch model, the melting rate
is calculated with an effective melt mass flux

S::::::;(l-c)xAexp (T-To)
-c- , (2.99)

were A, C, and To represent coefficients, which are determined in crucible


experiments for the various batch compositions. The virtual porosity c repre-
sents the volume fraction (1- c) of the remaining, unmelted batch. With this
relationship, a temperature-related melting flux S is introduced. However,
the definition or calculation of temperature T is not further specified in the
paper. One might presume that it corresponds to the minimum (inner) tem-
perature of the semi-empirical model for the pot experiments (see [2.154]).
For the batch energy balance, a 3D transient energy conservation equation
is given in the paper. However, it is not further specified whether this equation
is used for shell balances or for the continuum approach.
The global batch blanket or island rheology is represented by two param-
eters. The force parameter expresses the mechanical force of the charging
system and the mutual forces between batch islands. The apparent viscosity
parameter expresses the lateral expansion of the batch blanket. It should be
stressed that both parameters have global character and must be determined
experimentally for every batch and furnace type.
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 121

A few validations of the TNO glass tank model have been published (see,
e.g., [2.144]), but no specific measurement on the batch melting have been re-
ported. Recently, TNO announced the set-up of a full 3D batch model [2.145].
Kawachi and Kawase 1998
The study of Kawachi and Kawase [2.146] illustrates the difficulty in getting
correct starting conditions for the 3D simulation of melt redox and bubble
evolution in a melting tank. Because a detailed batch model is lacking, they
have to rely on small-crucible experiments. The samples were melted under
conditions that should be representative of the batch-blanket melting history.
On these samples, the amount of solved gases (C0 2 , N 2 , O 2 ), the redox state,
and the number, composition, and size distribution of the gas bubbles were
determined.
These values were then used as a constant batch area inlet boundary
condition. This approach, of course, does not account for differences in the
melting conditions, (e.g., between the lower and upper blanket side), and
sampling of the required parameters in the furnace itself is a utopian task.
Advanced batch models seem to be the only way to solve the problem of
starting conditions for bubble and grain simulations in the melt.
Wang, Brewster, Webb, McQuay, and Bhatia 1999
Wang et al. presented in [2.147] a batch model coupled with a 3D CFD simu-
lation of combustion and melt in a float-glass tank. The model considers the
batch blanket as a series of thermally unconnected 2D lanes extending from
the batch feeders in the direction of the working end. The temperature field,
the melting process, the shape of the batch blanket, and the batch coverage
area and local thickness are predicted by the batch model. No further details
on the model are given in the paper and the reader is referred to [2.148].
The batch model is thermally coupled to the melt and combustion model
by temperature and heat-flux boundary conditions. The whole tank model is
validated with temperature and heat-flux measurements in a float-glass fur-
nace. However, no specific measurement on the batch melting are presented.
Prasad, Mukhopadhyay, and Dutta 1999
The 3D batch model presented by Prasad et al. [2.149] comprises a full set
of conservation equations (i.e., mass, energy, and momentum). It is an in-
trinsic part of a 3D CFD glass-melt simulation. Batch temperature as well
as flow fields are model results and are not initially fixed. This gives great
flexibility for modeling various batch shapes and melting modes. The model
then requires, of course, detailed material parameters concerning thermal and
rheological properties in particular.
The batch-melting reaction can be addressed in two ways: one approach
is based on the degree of conversion similar to the model proposed by
Hrma [2.152] and Schill [2.140]. In this approach, the temperature interval
for the batch reaction is fixed and determined by DTA/TG measurements.
122 2. Melting and Fining

The second approach addresses batch melting by reaction kinetics for a


first-order reaction and an Arrhenius-type temperature dependence. The re-
action enthalpy is considered in a source term of the energy conservation
equation. A model tank proposed by Ungan [2.136] was re-simulated by
the authors and comparable results for batch and melt were obtained (see
Fig. 2.33).
For the model set-up, the CFD code Fluent was used. Apparently, a
general-purpose CFD code is suited for the numerical simulation of a spe-
cific problem such as the batch-melting process. The validation of the batch
modeling with real tank data is presented in [2.150].

Classification

The wealth of published batch models can be quite confusing. Here, we will
try to classify these models according to the type of mathematical description
of the physical reality and in terms of predicted outputs (see the classification
of demands in Sect. 2.2.1). Only models coupled to 2D or 3D melt simulations
will be cited .
• 1st level:
- batch is represented by a boundary condition;
- velocity inlet with a fixed temperature;
- wall with heat flux and additional mass inlet;
- examples: [2.95,96] .
• 2nd level:
- analytical or low-order numerical batch models;
- mass and thermal coupling at interfaces;
- discrete thermal balances or calculation of the heat penetration;
- examples: Austin and Bourne 1973, Leyens 1974, Mase and Oda 1980,
Song and Sun 1995, Yamamura et al. 1997, Beerkens 1998.

Fig. 2.33. 3D CFD model of batch melting in a glass tank. The decreasing concen-
tration of the carbonates at the batch surface indicates the progress of the batch
reaction (after [2.149])
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 123

• 3rd level:
- basic CFD batch model directly coupled to the melt model;
- multiphase nature of the batch represented by effective properties;
- these properties depend on a batch-to-melt conversion factor, which is a
function of a temperature history;
- reaction energy is related to the conversion degree;
- integrating the reaction energy into cp = f(T) is equivalent;
- examples: de Waal 1990, Schill 1995, [2.151].
• 4th level:
- advanced CFD batch models with chemical reaction kinetics;
- properties depend on the temperature and conversion degree;
- examples: Ungan and Viskanta 1986, Prasad et al. 1999, [2.150].
• 5th level (for the future):
- free surface for correct batch thickness and surface melt flow;
- detailed non-Newtonian flow behavior;
- spectral radiation and scattering at bubbles and un melted batch;
- transient calculation of discontinuous charging and flame reversal;
- CO 2 + H 2 0 transport by gas percolation in the batch core;
- melting kinetics for the number and size of released bubbles and grains;
- melting kinetics in conjunction with gas transport for melt redox state
predictions;
- batch surface evaporation kinetics and dust-particle transport.
It is clear that level-5 batch modeling is extremely challenging. Future
batch models will have to focus on a selection of the aspects mentioned under
level 5. Nevertheless, some of the points listed there have already been treated
partially in the batch models we discussed. The advanced demands defined in
the introduction can be satisfied only with level-5 batch modeling. Currently,
level-4 batch modeling has to be settled and the lack of property data and
validation measurements to be tackled.

Validation

Validation is, of course, a key element in the development of any batch model.
A correct validation for a batch model consists in the comparison of a com-
plete simulation of the furnace with measurements, focusing on the batch
area. This procedure must be carried out for all principal glass-furnace types
and batch compositions. In Table 2.3, these principal furnace types are listed
together with their usual batch compositions and charging modes. Certain
furnace types as used, for example, in float-glass and TV-panel production
have comparable batch blankets, which simplifies validation. However, the
interior of glass furnaces is hardly accessible and the only information cur-
rently available about batch blankets is their extension on the glass melt.
Temperature measurements by thermocouples introduced vertically through
124 2. Melting and Fining

the furnace crown are not practicable in the batch area, because of the hori-
zontal force of the batch on the thermocouple rod.
There is an urgent need for new validation methods to better access the
basic physical parameters that determine batch melting in industrial furnaces.
These needs consist of the following measurements:
• vertical temperature profiles in batch and melt under the batch,
• radiative heat-flux distribution at the batch and melt surface,
• lateral extension of the batch using video recording,
• temperature profiles at the batch surface.
A very important parameter in batch melting is the heat-flux distribution
between the lower and upper sides of the batch blanket. The primary recir-
culation of the melt is essentially determined by the lower-side batch-heat
consumption (see Sect. 2.2.1). Hence, there is a need to get more detailed
information about the heat-flux distribution between the upper and lower
sides.
As of today, the most important published validation case is the TC 21
RRT 3 Ford float-glass furnace. Regrettably, the only information available
about the batch blanket of this float furnace is its lateral extension.
Laboratory-scale furnaces may be used to investigate the physical pro-
cesses of the batch melting and to develop batch models. A few melting stud-
ies in pot furnaces have been published (see Table 2.5). However, in these pot
furnaces, only transient melting can be studied. Unfortunately, the results of
these melting studies cannot be directly transferred to the melting of blan-
kets in full-scale furnaces (see Sect. 2.2.3). However, careful interpretation of
the pot results allows one to define basic parameters such as the apparent
thermal diffusivity of the batch raw material [2.154].
A completely different approach to the batch model validation was chosen
by the TC 21 in the RRT 4. Here, the idealized glass tank of the RRT 1
was extended by a batch-melting process and a combustion space with heat
exchange solely by radiation. The aim of this benchmark test is a comparison
of the different batch-modeling approaches. Only a few key properties for the
batch are fixed, which gives much scope for different theoretical approaches to
the batch-melting physics. However, the RRT 4 is still running and the final
results have to obtained before drawing any conclusions on this validation
method.

Conclusions

Hrma stated 20 years ago [2.152]: "It is unlikely that the melting conditions
of a glass furnace could be simulated in the laboratory. The time and effort
should be devoted rather to measurements of properties relevant for batch
behavior during melting. Glass technology can benefit only from close co-
operation between both experimental and modeling approaches accompanied
2.2 Mathematical Modeling of Batch Melting in Glass Tanks 125

Table 2.5. Laboratory studies on batch melting in pot furnaces

Heating modes Batch composi- Validation data Model Refer-


tions ence
Surface by com- Ba-Sr batch, T -profile as a func- virtual [2.154]
bustion; melt cullet, moisture, tion of time (6 ther- thermal
in electrically sand, pellets mocouples TC) diffusivi-
heated pot ties
Surface by com- soda-lime clear T-profile as a func- heat [2.55]
bustion; melt by + opaque, pel- tion oftime (12 TC) transfer
MO electrodes lets coeffi-
cients
Surface by com- soda-lime sili- T -profile as a func- diffusivity [2.97]
bustion; melt in cate tion of time (5 TC) coeffi-
heated pot cients
Surface by com- soda-lime sili- T-profile as a func- effective [2.155]
bustion; melt by cate, pellets tion of time (11 TC) thermal
M 0 electrodes conductiv-
ity of
pellets
Surface by com- sheet-glass T -profile as a func- thermal [2.156]
bustion batch and cullet tion of time (4 TC) diffusivi-
fiber-glass batch ties below
700°C

by direct observations of melting in full-size furnaces." This statement is still


valid and needs nO further comments.

Acknowledgements

I would like to thank my company, Stein Heurtey, for entrusting me with this
interesting topic and authorizing me to publish part of the results. Further-
more, I would like to acknowledge Prof. Conradt and my work colleagues, in
particular Bertrand Strock, Frederic Marmonier, and Andre Bilde, for many
constructive discussions on the batch-melting process. Thanks are also due
to our customers for interesting discussions on the theory and practice of
batch melting in our furnaces. Finally, I would like to express my particular
gratitude to my wonderful wife for her patience with my extended working
hours and for her help in editing this manuscript.
126 2. Melting and Fining

2.3 High-Frequency Melting of Glass in Crucibles

Frank- Thomas Lentes


Introduction

Indirect inductive heating of glass melts with frequencies of about 10 kHz


is a well-established technology in glass production. A platinum crucible of
good electrical conductivity is heated via a coil by eddy currents induced
in the metal. The heat of the crucible is then transferred to the glass melt.
Naturally, the temperature of the glass melt can never be higher than that
of the platinum. The upper limit of the maximum glass temperature is the
highest possible temperature of the platinum (~ 1650 QC).
A basically different technology is the direct heating of glass melts with
frequencies in the megahertz range. Already examined by Eden [2.157] in
1950, this heating technology has regained interest in the last few years be-
cause of the very high temperatures to be reached with it. At a certain tem-
perature, the electrical conductivity of many glasses is high enough to couple
the high-frequency magnetic field to the glass melt, thus directly inducing
eddy currents there. These lead to a Joule effect in the glass, i.e., there is no
detour via other transport mechanisms.
Starting with the basics of electrodynamics, a model for high-frequency
heating of glass melts is developed, and computation results for the result-
ing temperature and flow fields are presented. With respect to applications,
this section is restricted to the heating of axisymmetric crucibles; for a gen-
eral introduction into conductive and inductive electrical heating of various
materials see [2.161].

2.3.1 Basics of Electrodynamics

Maxwell Equations

Classic electrodynamics is the theoretical basis for the solution of eddy cur-
rent problems. In general, electromagnetic fields are described by the Maxwell
equations [2.162]:

VD=p VE=-~~ ,
(2.100)
VB=O VH =j + 8ft .
Here, E is the electric field strength (V1m), D the electric flux density (V1m),
H the magnetic field strength (Aim), B the magnetic flux density (Aim), j
the total current density (A/m 2 ), and p the electric-charge density (C/m 3 ).
The field equations are completed by the material equations for the de-
scription of the interaction between electromagnetic fields and matter.
2.3 High-Frequency Melting of Glass in Crucibles 127

D = eoerE
B = J.LoJ.LrH (2.101)
j = o-E,

where eo is the electric permeability of the vacuum (in 81 units 8.8542 x 10- 12
A s V- 1m -1), er the relative electric permeability of the material, J.Lo the mag-
netic permeability of vacuum (in 81 units 47f x 10- 7 ), J.Lr the relative magnetic
permeability of the material, and 0- the electrical conductivity (O-l m -1).
The total current density j consists of two components, namely the cur-
rent density jq of the electrical contacts and the current density jw of the
induced eddy currents. Thus, we get

'M'H . . oD
v = Jq + Jw + ot . (2.102)

According to the requirements of the mathematical modeling of inductive


heating, we restrict ourselves in the following to time-dependent fields with
not too high frequencies, so that the displacement current density (oD)j(ot)
can be neglected compared to the conduction current density j (quasi-
stationary approximation). This is possible if, together with the circular fre-
quency w, the following constraint is fulfilled [2.158]:

0-» Weoer . (2.103)

For the frequencies used in inductive heating of glass (w ~ 27f X 1 MHz), the
displacement current density has no importance. (For a frequency of 1 MHz,
Weoer is less than 10- 3 0-l m -1, whereas the electrical conductivity of hot
glass melts is typically greater than 1 0-l m -1.) Hence, the Maxwell equations
can be simplified and we get
oB
VD=p VE=--
ot ' (2.104)
VB=O VH=j.
The two equations on the right-hand side can be written as one equation if
we formulate the 1st derivative of V H with respect to time, substitute j by
j = 0- E, and multiply this equation by J.Lol),r. This yields
oB 00- oE
V Tt = J.LoJ.Lr ot E + J.LoJ.Lro-Tt . (2.105)

(0 B) j (ot) - V E leads to the so-called telegraph equation (according to


the constraints, without the displacement current):

(2.106)

During a period of high-frequency application, the local electrical conductiv-


ity is constant with respect to time, therefore (oo-)j(ot) = O. Now, we can
further simplify:
128 2. Melting and Fining

8E
V (V E) + MOM r O"7ft = 0 . (2.107)

With the vector analytical identity V (V E) = V (V E) ~ 6.E, this equation


reduces to

(2.108)

because there are no free charges, thus V E = O. (The existence of electro-


static fields requires a conductivity of the insulation material that is almost
zero [2.159]. Because in reality insulation materials have finite conductivities,
for potentials constant with respect to time there is electric flow. The electric
field constant with respect to time is always a flow field.)
For both the electric and the magnetic field, we assume harmonic oscilla-
tions with circular frequency w:

E(x,t)=E(x)e iwt ,
(2.109)
B(x,t) =B(x)e iwt .

Fields E and B as well as the current density j are then complex parameters.
This finally yields

(2.110)

Solution of the Maxwell Equations

In the integration area considered, we are seeking a solution that fulfils the
Maxwell equations (2.104) according to given boundary conditions. A popular
approach for the solution of the field equations is based on the scalar electric
potential V and the magnetic vector potential A. Because the magnetic flux
density B is source-free, it can always be written as the rotation of a vector
field A: B = V A.
There are no constraints for the selection of V and A. By means of an
adequate "gauging transformation", the Maxwell equations can always be
fulfilled. For reasons of uniqueness, in the following we will use the so-called
Coulomb gauging with V A = O. Fields E and B can then be formulated
with the two potentials V and A.

B=VA,
8A (2.111)
E = ~7ft ~ VV.

The following boundary condition for A must always be fulfilled:

A(x-+oo)=O. (2.112)

The total current density is given by


2.3 High-Frequency Melting of Glass in Crucibles 129

j=aE=-a~~-avv. (2.113)

It consists of the induced eddy current density jw = -a[(8A)/(Ot)] and the


current density jq = -a VV caused by the voltage applied. This solution ap-
proach is used in many commercial software packages (e.g., ANSYs, EMAS).
Another possibility for modeling the heating by eddy currents is the direct
(numerical) solution of (2.110).

Computation of the Joule Effect

In order to receive the power input by the Joule effect in a volume element,
the local power density P (W m- 3 ) is required.

(2.114)

The factor 1/2 is due to time averaging over a harmonic oscillation.

2.3.2 Mathematical Formulation of the Simulation Model

Solution of the Axisymmetric Electrodynamic Problem

Semi-Analytical Solution. In the following, we will examine the inductive


heating of an axisymmetric melting device as shown in Fig. 2.34.
A heated glass volume of height H and radius R is surrounded by a
cylindrical coil with n windings per unit of length, through which current I
is flowing. The relative permeability of the glass is JLr = l.
As an approximation, we formally consider the cylindrical coil in the di-
rection of z as infinitely expanded. Thus, the magnetic field caused by the

Glass

Fig. 2.34. Scheme of the coil and the glass


130 2. Melting and Fining

current in the coil is homogeneous in the interior of the coil and almost van-
ishing at its surface. (This is not valid for the total magnetic field including
the influence of the eddy currents!) The electrical conductivity of the glass
strongly depends on the local temperature, hence u = u (T (r, z)).
Owing to the homogeneous magnetic field and the rotation symmetry, in
cylindrical coordinates (2.110) can be written in the following way:

E)2 Eep loEep 1 0 2Eep .


or2 + -:;:or - r2 Eep - OZ2 - lW/LoU Eep = 0 . (2.115)

Because the magnetic field is assumed to be homogeneous, there is no explicit


dependence on z. However, the electrical conductivity being a function of the
local temperature, the dependence on z is implicit. Therefore, we make a
further approximation and write

d 2 Eep
-d
2
r
+ -r1 dEep
-d -
r
1
2Eep
r
.
-lW/LoU
(())
T r,z Eep = o. (2.116)

We solve this ordinary differential equation in r for the azimuth component


of E, but take the actual radial distribution of U (T(r, z)) for every height z.
The solution of the boundary-value problem of (2.116) requires two
boundary conditions. For mathematical and physical reasons, one of these
at point r = 0 must be

Eep (r = 0) = 0 . (2.117)

In order to formulate the second boundary condition at point r = R, we


need a formulation for Eep, respectively (oEep)/(or). A connection between
the electric field and the magnetic field is given by VE = - [(oB)/(ot)]. The
magnetic field has only one z component, therefore the z component of the
rotation of E is given by
oEep Eep oBz
-+-=--. (2.118)
or r ot
For the computation of the magnetic field we need a connection with the ex-
iting current of coil 1. We consider the integration path described in Fig. 2.35.
In the interior of the coil, according to the constraints, there is a homogeneous
magnetic field oriented in the direction of z. At its surface, the magnetic field
is almost vanishing, so that it can be neglected.
With Stokes' theorem, we can write:

f B ds = J
V x B dA = /Lo J j dA . (2.119)

The right-hand side is simplified in the following way:

f B ds = /Lo nI L . (2.120)
2.3 High-Frequency Melting of Glass in Crucibles 131

L Glass
••
A

Fig. 2.35. Integration path for the deter-


mination of the boundary condition

The ring integral yielding only BzL, it finally follows that

Bz = J.1onI . (2.121)

By formulating the temporal derivative of this equation (oBzlot = iWJ.1onI)


and inserting the result into the right-hand side of (2.118), we get for the
second boundary condition

dEep
- - + -E<p =
.
-lWJ.10 r=R .
I (2.122)
dr r
We conclude the solution of the axisymmetric problem by summing up the
results as follows:
d 2 E<p 1 dEep 1 .
dr2 + ~dr - r2 E<p - lWJ.10CJ (T (r, z)) E<p = 0 ,
with the boundary conditions (2.123)
dEep . I
E<p (r = 0) = 0 and -d
r
+ -E<p
r
= -lwJ.1onI r=R .

As an abbreviation, parameter b = V2(J.10J.1rwCJ) is introduced and called the


penetration depth (for reasons soon to become clear).
Discussion of the Solution. Although in reality the electrical conductivity
strongly depends on temperature, it is very useful to discuss the properties
of the solution first for the case of constant conductivity. Equation (2.123)
then has an analytical solution.

E ( ) _ - ( 1 + i) WJ.10 b nI J 1 [( 1 + i) rib]
(2.124)
<p r - 2 J o [(l+i) Rib] .

Here, Jo and J 1 describe the Bessel functions of zero and first order. For large
penetration depths (b » R), we get the approximation
132 2. Melting and Fining

-iwfto nI
E i':::5 2 r. (2.125)

For small penetration depths (6 « R), in contrast, we get


(1 - i) wft6 nI
E i':::5 2 exp [- (1 - i) (R - r) / 6] . (2.126)

Here, the importance of 6 as the penetration depth becomes clear: at depth


6, the field strength has decreased to l/e.
For the power density created by the Joule effect P = 1/20" IEI2, we have
the following relations:

1 2 2 beri (V2r / 6) + beii ( V2r / 6)


(2.127)
P = "2 Wft n I ber6 (V2R/ 6) + bei6 (V2R/ 6)
"ber", respectively "bei", stand for the real and the imaginary part of the
Kelvin functions. (The Kelvin functions are defined in the following way:
ber(x) = 1- [(X/2)4 /2!2] + [(x/2)8 /4!2]_ ... ; bei (x) = (x/2)2 - [(x/2)6 /3!2] +
[(x/2)10 /5!2]_ ... ). For large penetration depths (6)> R), we can write

(2.128)

As approximation for small penetration depths (6 « R), the radial distribu-


tion of the power density is given by

(2.129)

Figures 2.36 and 2.37 show the normalized field strength and the normal-
ized power density depending on radius r (normalized to radius R) for very
large and very small penetration depths [2.160].
Conclusion: In the case of an axisymmetric device, the power input at
point r = 0 is always zero, i.e., in the center of the cylinder there is no
heating. The maximum power input happens at or near the surface. Here we
can also observe the maximum heat loss.

0.5.---------------------------~

0.4
li/R» 1

-I'w E. . .
- ::t
0.3
80.2

0.1 li/R =0.2


0.0 -JL----,.---.....,====;.------,-----I Fig. 2.36. Normalized field
0.0 0.2 0.4 0.6 0.8 1.0 strength depending on the
r IR normalized radius
2.3 High-Frequency Melting of Glass in Crucibles 133

0.5...,-----------------,

0.4

"t: 0.3
/
>'"
a. -::l.
80.2

0.1 I)/R = 1.12


0.0 +--~=r===---,_--_r===--,__----1 Fig. 2.37 . Normalized power
0.0 0.2 0.4 0.6 0.8 1.0 density depending on the
r /R normalized radius

Solution of the Coupled Electrodynamic


and Fluid Dynamic Problem

Formulation of the Physical-Mathematical Problem. In order to model in-


ductive heating, we couple the following physical processes:
• creation of the Joule effect by eddy currents in the glass;
• heat transport by radiation, heat conduction, and convection from the glass
melt to the surroundings;
• convective flow field in the glass melt due to the temperature-dependent
density of the glass.
For a mathematical solution of the "inductive heating with convection"
problem, the following equations must be solved simultaneously:
1. local conservation of energy in the Boussinesque approximation (this
approximation does not account for the temperature-dependent density in
the energy equation)

PoCp (~~ +UVT) = ~ +Pel (2.130)


heat conduction

2. local impulse conservation (Navier-Stokes equation)

Po
au)
(8t + =-v u Vu p + ~ - Fog (!h ~T - To) ),
2
(2.131 )
friction bouyancy

3. equation for the complex-weighted electric field (for the real and imaginary
part)

(2.132)

4. power density of the Joule effect


1 2
Pel = "20" lEI (2.133)
134 2. Melting and Fining

Description of the parameters: T is the temperature, u the vector of the flow


velocity, p the local pressure in the fluid, and g the gravitational acceleration.
Material properties: >"(T) is the heat conductivity, Po the density at the ref-
erence temperature To, cp the specific heat, 1](T) the viscosity, (3 the linear
thermal volume expansion coefficient, and a(T) the electrical conductivity.
We are searching for the following parameters: spatial distribution of the
electric field strength E(x), spatial distribution of the current density j(x),
spatial distribution of the power-density input (local power density through
high-frequency heating) Pel (x), total power input in the glass Ptot , temper-
ature field T(x), and flow field u(x).
Material Parameters. For the computation of the simulations, reliable ma-
terial parameters are necessary. These include electrical conductivity a(T),
density p(T), effective heat conductivity >"(T) , and viscosity 1](T), which,
evidently, are all temperature dependent. For viscosity and electrical con-
ductivity, this dependence is actually very strong; for example, the electrical
conductivity varies by more than six orders of magnitude for a temperature
range between 20 DC and 2000 DC for typical glasses.
Realistic simulations can only be achieved when these material param-
eters are known. Unfortunately, they cannot be measured directly for very
high temperatures; the approximate limit is at 1700-1800 DC. For higher tem-
peratures, the respective material parameters must be extrapolated.
We use an Arrhenius approach to represent the electrical conductivity in
an analytical function, and the Vogel-Fulcher-Tamman equation to represent
the viscosity.

2.3.3 Simulation Results

Example 1

The simulated crucible is a vessel with a volume of : : : ; 20 L and an interior


radius of 12.5 cm, a height level of 35 cm, and a wall thickness of 3 cm. Initially,
the glass has a homogeneous temperature of 1400 DC. The temperature of the
cylindrical surface area of the quarzal crucible (quarzal is a special refractory
material) is 50 DC at the exterior side. The bottom temperature of the crucible
is 600°C. The lid of the crucible is cooled by radiation (c: = 0.7) and natural
convection (he = 10 W / (m 2 K)). For the glass, no-slip at the crucible is
assumed everywhere. The power input happens via a coil with 16 windings
with a current strength I of 57 A at a frequency v of l.5 MHz.
Figure 2.38 shows the temperature field and Fig. 2.39 the flow field of the
stationary computation (because ofaxisymmetry, only the left half is shown).
The temperature distribution with a maximum value of 2067 DC is relatively
homogeneous. The domain with the highest temperatures is displaced up-
wards from the center of the crucible due to convection. The flow shows one
single convection eddy. The maximum flow velocity is 2.11 mms-I.
2.3 High-Frequency Melting of Glass in Crucibles 135

2330 K (2057 · C)
2310 K (2037 · C)
2290 K (2017 · C)
2270 K (1997 · C)
2250 K (1977 · C)
2230 K (1957 . C)
2210 K (1937 · C)
2190 K (1917 · C)

Minimum: 323 K (50 . C)


Maximum: 2340 K (2067 · C)

Rotational axis Fig. 2.38. Temperature field in


a 20-L crucible with a quarzal
wall thickness of 3 cm

Velocity vector plot

Rotational axis Fig. 2.39. Flow field in a 20-L


crucible with a quarzal wall
thickness of 3 cm

Example 2

The simulated crucible again is a vessel of 20 L, but the wall thickness is only
1 em, which leads to considerably higher loss of heat via the wall. The current
strength is 65 A. The rest of the configuration is identical with example 1.
The results are shown in Figs. 2.40 and 2.41.
The heat flow penetrating the cylindrical quarzal wall is nearly four times
higher than that in example 1. Therefore, the temperatures are lower by
::::; 300 K, despite a higher current density. The increased viscosity leads to
136 2. Melting and Fining

1980 K (1707 ec)


1940 K (1667 ec)
1910 K (1637 eC)
1870 K ( 1597 0c)
1840 K (1567 eC)
1800 K (1527 ec)
1770 K (1497 eC)
1730 K (1457 ec)

Rotational axis Fig. 2.40. Temperature field in a


20-L crucible with a quarzal wall
thickness of 1 cm

Velocity vector plot

·:1
:I Color code veloc ity

.. 0 .59
052
0.44
0 .37
0.30
0.22
0.15
0.07
mmfs

Rotational axis Fig. 2.41. Flow field in a 20-L


crucible with a quarzal wall
thickness of 1 cm

lower flow velocities; at 0.6 mm s-l, their maximum value is considerably


lower than in the case of example 1.

2.3.4 Conclusion and Outlook

From the above results, we can draw the following conclusions:


• From theoretical considerations it follows that the maximum heat-power
input by eddy eurrents happens near the wall of the crucible, whereas at
the axis of the cylinder (center of the crucible) the power input is zero.
2.4 Model-Based Glass Melter Control 137

• The simulations show that radiation and convective mass transport are
responsible for the heat transport in the interior of the hot glass.
• Furthermore, the simulations show that the profile of the power input
(strongly increasing from the center towards the wall of the crucible) is
completely different from the resulting temperature field. Therefore, it is
in general impossible to draw any conclusions from the temperature field
with respect to the distribution of the power input!
• The material parameters for electrical conductivity, effective heat conduc-
tivity (as an approximated parameter for the radiation transport), and
viscosity should be known up to the maximum temperature used; if not,
these parameters must be extrapolated for very high temperatures. The
latter two parameters control whether the heat transport in the melt is
primarily due to radiation or convection.
The so-called skull crucible method can be applied to make glass melt "in
its own juice". Here, the wall of the crucible is actively cooled, whereas the
temperature of the glass-melt interior is kept constant. To avoid cooling of the
melt interior, power input and heat loss are balanced over an average time.
For a 50-L crucible, an electric power of ~ 70 kW (without electric power
losses) is necessary, which is instantaneously transferred to the cooling agent
and to the surroundings as a power loss. An additional amount of energy
is needed to heat up the glass from the starting temperature to a certain
maximum temperature.

2.4 Model-Based Glass Melter Control

Ton Backx
Introduction

Glass industries, like most other process industries, have been confronted with
a major change in the market during the past decades. Competition has dras-
tically increased and environmental legislation has been tightened severely.
The strong growth in production capacity in general, and in container-glass
manufacturing in particular, has exceeded the growth in market demand. This
has resulted in a market that is largely customer-controlled and saturated.
In addition, the complexity and costs of production equipment have signif-
icantly increased owing to the tightening legislation on ecosphere load, and
operation of the processes within the ever-tightening constraints has become
increasingly more complex as well.
One of the major reasons for the changes is globalization of the market.
Globalization is one of the results of the recent developments in the fields
of telecommunication, transportation, and advanced automation, which have
emerged from the rapid developments in electronics and computer and in-
formation technology. As a consequence, the process industry is nowadays
138 2. Melting and Fining

confronted with a strongly competitive market. The market has developed


from a supplier-driven market to a demand-driven market. These changes
have far-reaching consequences for producers. In this market, margins on
products are eroding rapidly. Good margins can only be obtained for prod-
ucts that are scarce and in demand. Customer-dictated markets are capri-
cious. Opportunity windows for good margin product sales are tightening.
This requires producers to respond quickly and reliably to product demands.
Products have to be delivered at short notice in strictly defined time windows
in the right quality and the requested volume.
These market changes enforce industries to flexibly produce small series
of a large variety of product types, preferably with existing production instal-
lations. A further consequence of these changes in the market is a continuous
shortening of the life cycle of a significant part of supplied product types.
The innovative power to bring new products to the market quickly, in a pre-
dictable and controlled way, is becoming a necessity for industries to improve
or even maintain their market position.
In order to prepare for these drastic changes, tight control of the pro-
duction processes over a broad operating range is needed. Process operation
has to enable a completely predictable and reproducible operation at and
changeover between different operating points that correspond with the pro-
duction of various product types under different economic objectives (min-
imize costs, maximize production rate, minimize stock, benefit from fluctu-
ating prices, etc.). The strategy yielding the most profitable conditions has
to be selected from a variety of potential operating scenarios to produce the
desired product type. This decision is based on a thorough understanding of
both process behavior and process operation. The freedom of choice offered
in process operation must be used to predictably produce precisely what
is required in terms of quality, volume, and time, with the best achievable
business result.
This section explains how process models and model-based control sys-
tems can be used to support process operation in the most flexible way, in
accordance with market requirements and driving towards conditions that
maximize margins. The use of model-predictive control technology to push
processes closer to their physical limits in order to obtain a better economic
result is discussed. Because the performance of model-based process control
systems relates one-to-one to the accuracy of the models applied, we start
with a short introduction into modeling and model concepts.

2.4.1 Model Concepts

Detailed knowledge of process behavior and extensive use of this knowledge


are the key for obtaining the intended improvements in process operation.
Mathematical models are the vehicles for making knowledge on process be-
havior accessible for automated process operation.
2.4 Model-Based Glass Melter Control 139

Glass-manufacturing processes consist of a sequence of manufacturing


steps. In each of these production steps specific processing conditions need
to be realized to guarantee ultimate product quality. The main steps are:

• preparation of batch and batch transport,


• charging of batch material in the melter,
• melting and fining of glass,
• conditioning of the glass for further processing,
• manufacturing of the products,
• conditioning of the products,
• post-processing of the products.

Each of these processing steps needs to satisfy particular specifications.


The available operating envelopes of the processing units enable realization
of the specified processing conditions for high-performance manufacturing,
if they are appropriately designed. In each process step, a number of vari-
ables determine the course of the process and consequently the characteris-
tics of the resulting products (e.g., component separation, residence time and
residence-time distribution, temperature profile in the melter or forehearth,
hot-spot temperature, hot-spot location, boosting, bubbling, concentration,
homogeneity and purity of batch components, concentration of undesired
components, glass level, furnace pressure, exhaust gas oxygen excess, NO x ,
temperature distribution in forehearth cross section, etc.). A selection of crit-
ical processing variables and a number of product properties of the semi-
manufactured and final products have to be kept within specified tolerance
limits or have to be brought within these limits during a process changeover
to guarantee good ultimate product quality and to ensure high lifetime of the
manufacturing equipment. These process variables are the so-called process
outputs or CVs (controlled variables).
In order to keep the CVs in their predefined region, a set of process vari-
ables are available for manipulation of the process behavior. These variables
have a predefined operating region within which they can be manipulated
by the operator or the control system. These variables, the so-called process
inputs or MVs (manipulated variables) are used to compensate for external
disturbances and changes in the observed process behavior. They have to
drive the process to the desired operating conditions along preferred paths.
The third category of process variables that affect the process behavior are
the so-called process disturbances or DVs (disturbance variables). Examples
of these variables are impurities of the batch components, composition of the
batch, humidity of the batch and the combustion air, ambient temperature,
furnace wear, reversal of firing, Wobbe index, and so on. These variables
determine the process behavior in a manner similar to the MVs, but un-
like them, DVs cannot be manipulated. Consequently, we have to accept the
presence of these disturbances and the resulting effects on the processing.
In the best case, the disturbances affecting processing are measurable. Their
140 2. Melting and Fining

ultimate effect on product properties or on the process may be predictable


over a certain time horizon. In model-based control terminology these mea-
sured disturbances are often referred to as DVs. Unmeasured disturbances
are then considered to be part of the process output noise. Figure 2.42 gives
a graphical representation of a process and the variables defined above.
As an example, a melting tank can be used. The most relevant variables
for control of a melting tank are:
• controlled variables (CVs)
- glass level - crown temperature profile
- batch composition - residence time
- batch position - residence-time distribution
- hot-spot temperature - exhaust oxygen excess
- glass temperature profile -NO x
• manipulated variables (MVs)
- total fuel flow - batch charging
- fuel-flow distribution - reversal time
- air/fuel or oxygen(! air) /fuel - boosting
ratio for each burner - bubbling
- cooling air - anti-foam agent flow
• disturbance variables (DVs)
- batch composition (unmeasured disturbance)
- batch humidity (measured or unmeasured disturbance)
- ambient temperature (measured or unmeasured disturbance)
draught (unmeasured disturbance)
- furnace wear (unmeasured disturbance)
- foaming (unmeasured or measured disturbance)
- Wobbe index or fuel composition (measured or unmeasured disturbance)

Unmeasured disturbances

------------------------------------- -

Process
Measured i
disturbance i
CI=~:=========> Disturbance transfer
variables i

Manipulated i Controlled
Process transfer +
variables variables

Fig. 2.42. General representation of a process and the defined variables


2.4 Model-Based Glass Melter Control 141

In general, the process installations as well as the processes running in


the processing equipment exhibit inertia. When a variable is adjusted, for
example a gas flow, the process starts changing for a while. After the so-
called response time, it arrives at a new steady state that corresponds to
a new operating point. This dynamic behavior, where the process changes
over a characteristic time interval in response to a manipulation of a process
variable or a change in a disturbance variable, is called the dynamic behavior
of the process. A dynamic process model can describe the relevant dynamic
process behavior for the complete transition time interval. The step response
is a well-known example of such a dynamic model. The step response is
the response of process variables and product parameters on a unit step
adjustment of a manipulated variable.
Figure 2.4:1 gives an example. The step response in this figure shows how
the underlying process changes due to a step change at the input; it char-
acterizes the changeover from one operating point as a function of time to
another operating point. Detailed analysis of the response in Fig. 2.43 shows
that the process output, after a short delay time of approximately 2 min, ini-
tially moves in the wrong direction for about 15 min, after which the output
reaches its final value in around 80 min.
The step response model is a specific model representation of the process
dynamics. Other model types that represent dynamic behavior are impulse re-
sponses, transfer functions, differential-algebraic equations (DAEs), and state
space models. Each model type has its specific mathematical representation.
Process models can, within certain limits, be used for simulation and pre-
diction of the expected process responses on arbitrary input signals applied
to manipulated variables and/or disturbance variables of the process. Con-
sequently these models enable the prediction of the process outputs in the
near future on the basis of known adjustments on the manipulated variables
and known behavior of the measured disturbance variables in the recent past.
The process models can also be used to determine which manipulated variable
adjustment8 are to be applied to the process in order to bring it efficiently

Step response
0.05,--'--'---------------,
o
-0.05
-0.10
-0.15
-0.20
-0.25
-0.30
-0.35 +----.------r--,=~=;====r====1
o 20 40 60 80 100 120 Fig. 2.43. Step-response repre-
Time/min sentation of a process transfer
142 2. Melting and Fining

to a desired state, i.e., in accordance with the business goals. The bottom
line is: Models make process behavior more predictable, controllable, and op-
timizable. Model-based control systems explicitly use the knowledge of the
dynamic behavior of the process, as described by the models, to determine
the best possible control strategy under given market and production cir-
cumstances. In the design of classical PID control systems the model is only
implicitly applied for determination of the controller P, I, and D parameters.

2.4.2 Model-Predictive Control

A model-predictive control (MPC) system is an ideal tool for control of mul-


tivariable processes. Multivariable processes are processes whose inputs in-
fluence more than just one process output simultaneously. Characteristic for
MPC is that the control strategy can be adjusted for each calculation of a
following control action. As a result, MPC is very flexible for changing condi-
tions such as, for example, changing requirements, switching-off or failure of
sensors and actuators. Moreover, MPC can deal with constraint-type require-
ments, i.e., it can keep both manipulated as well as, to some extent, controlled
variables in certain predefined ranges. MPC has been developed within the
industry, emerging from the need to operate processes tighter within oper-
ational and physical constraints of the process and applied equipment, and
closer to the operating constraints that maximize margins. From its initial
development [2.163,164], MPC has grown to a widely proven technology, es-
pecially in oil refining. The dominant use of MPC in oil-refining applications
implies robustly pushing the controlled process to operating conditions that
maximize margins and minimize process variability. For most refinery appli-
cations, this results in maximization of the throughput of a certain product
mix. In glass manufacturing, the benefits mostly stem from tight control of
product quality, increase of average furnace load, increase of efficiency, tight
control of emissions and minimization of energy consumption.
The success of MPC within industry is to a large extent due to the fact
that MPC meets industrial requirements. These requirements can be roughly
categorized into three groups.
• Operational requirements: processes have to be operated within a prede-
fined region (safety, emissions, wear, etc.).
• Product-quality requirements: products have to be produced at specifica-
tions (Cpk values, 6-sigma ranges, etc.).
• Economic requirements: products must be produced in such a way that
margins are maximized, without violating operating constraints.
Figure 2.44 shows a block diagram of an MPC control system. Initially,
MPC did not explicitly take constraints into consideration. Refinements of
the technology developed at the end of the 1980s allow constraints on both
input and output variables to be considered in the formulation of the control
2.4 Model-Based Glass Melter Control 143

Disturbance

Manipulated 1---,--- Controlled


variables variables

........... .
............
.............. " ..
Measured
disturbances .............. -
................. , " .
. . . . . . ....
. ... .................
. r'............-'-"-''-'i-'--'-''-....... ................ ' ... ' ," ,. ,
t·........
... - ....... , , .
............ ,"

.............. ,

Setpoints
and
specifications

Optimization and
"constrainr handling .. . . . . . . . . . . .
................. , ................ .
. - . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . .

Fig. 2.44. Schematic representation of a model-based process control system

strategy. A paper by Qin and Badgwell [2.165] gives a good overview of the
MPC technology that is currently applied in industry.

MPC Without Constraints

The basic principle of MPC can best be illustrated on the situation without
constraints. The finite impulse response (FIR) model, describing the dynamic
behavior of a process with m inputs and p outputs, can be used to demon-
strate how input manipulations u(t) applied to the process at discrete time
instances in the past t = k-i, influence the process output y(t) at the current
discrete time instance t = k:
N
y(k) = 'LMiU(k-i) , (2.134)
i=l

where the p x m matrix elements Mi are the so-called finite impulse response
(FIR) parameters or Markov parameters.
Figure 2.45 shows the way the FIR model of the process is applied for
constructing the prediction of the process outputs. The input signal u(t) is
decomposed into a sequence of time-shifted "impulses" that compose the
original process input after summation. The bars with length ai represent
the impulses with amplitude ai that enter the process. They are the input
signal samples resulting from sampling of the continuous process input sig-
nals. The process output signals Yi result from summation of all elements at
row i of the impulse response elements scaled by the sample amplitude of the
corresponding input signal sample.
144 2. Melting and Fining

UT I · a, . a2 •a 3 . a. · as •a 6 . a7 •as . a9 · a,o

a, " Time
y(1 )= =M"a,
y(2)= ~M~l+M"a,

y(3)= • =M C3 J +M ,a2 +Mza ,

y(4)= • a4+M,aJ +M~2 +MJ.a ,

y(5)= • =Moas+ Mta,+M:za3+Ml ai +M",a,

y(6)= • a7
(::: Sum

y(7)= • • • • • (:: Sum

[. .
y(8)=
y(9)= •
• •

--• •


--• ...•


as"
t=:

c= Sum
Sum

y(10)= • • • • • • <:- Sum

y(11 )= • • • • • • • • Sum
• ...
(

-
y(12)= • • • • • • • • < Sum

y( 13)= • • • • • , .. (...= Sum


Time~ :

Fig. 2.45. Construction of the prediction of process output signals with the FIR
model

Hence, the FIR model can be used to describe the process output y(t) at
discrete time instances t = k - i in the past. More interestingly, when the
input manipulations u(t) at the discrete time instances in the future t = k + i
are known, the model can also be used to describe the evolution of the process
output y(k + i) at discrete time instances in the future:

00

L Miu (k - i)
i=O
00

Y (k) L Miu (k +1- i) .. M2 Ml Mo 0 0 0


u (k - 2)
Y (k + 1) i=O
00 .. M3 M2 Ml Mo 0
u (k - 1)
y (k + 2) L Miu (k +2- i) ·. M4 M3 M2 Ml Mo 0
u (k)
y (k + 3) i=O
·. M5 M4 M3 M2 Ml Mo 0 .
00
u (k + 1)
Y (k + 4) L Miu (k +3- i)
i=O u (k + 2)
u (k + 3)
00

L Miu (k +4 - i)
i=O

(2.135)

The future behavior of the process outputs is therefore determined by both


the input manipulations applied to the process in the past (u(k - i) Ii =
1,2, ... ) and the future input manipulations (u(k + j)lj = 0, 1,2, ... ).
By defining }fp(t, NfNp ) as the influence that the past input manipula-
tions over the horizon [t - N p , t -1] have on the future outputs over the time
horizon [t, t+Nf] at time instant t and by defining in addition (Yff(t, Nf, N c ) as
2.4 Model-Based Glass Melter Control 145

the influence that future input manipulations over the time horizon [t, t + Ncl
have on the future outputs over the time horizon [t, t + Nfl, the predicted fu-
ture behavior at the process outputs at time instant t over the time horizon
[t, t + N f ], say Yr(t, N f ), is determined by
Yr (t, N f ) = Yrp (t, Nf, N p ) + Yff (t, Nf, N c) (2.136)
= H (Nf' N p) Up (t, N p) + T (Nf' N c) Uf (t, N c) ,

where H (Nf' N p ) is the so-called Hankel matrix:

(2.137)

This Hankel matrix is the tool that enables prediction of future process output
responses On the basis of known past process input signals.
T (Nf' N c) is the so-called Toeplitz matrix:

Mo 0 0
Ml Mo 0
M2 Ml Mo
0
T(Nf,Nc) = M Nc - 1 M Nc - 2 M Nc - 3
(2.138)
Mo
MNc M Nc - 1 M Nc - 2 Ml

M Nf - 1 M Nf - 2 M Nf - 3 MNf-Nc-l

This Toeplitz matrix is the tool that enables prediction of future process
output responses to future process input manipulations.
Three vectors, the vector containing the predicted future process output
responses, Yr (t, Nt) E R(NfP)xl, the vector with past process input manipu-
lations, Up (t, N p ) E R(Npm)xl, and the vector with the future process input
manipulations, Ur (t, N c ) E R(Ncm)xl, are defined as

y (t)
y (t + 1) U(t~NP)1
Yr (t, Nr) = U (t, N p ) = and
y (t + Nr - 2) U (t - 2)
y (t + Nf - 1) u (t - 1)

u (t)
u (t + 1)
(2.139)
u (t + Nc - 2)
u (t + Nc - 1)
146 2. Melting and Fining

In MPC terminology, the horizon t to t + Nf - 1 is called the prediction


horizon. The control horizon equals the time horizon t to t+Ne-l. The above
distinction between the influences of past and future input manipulations on
the predicted future behavior of the process outputs is visualized in Fig. 2.46.
The distinction between the influence of the past and future input ma-
nipulations on the future outputs, respectively Yfp (t) and Yff ( t), is relevant
for MPC because
• past input manipulations have already been applied to the system and are
therefore fixed;
• future input manipulations have not yet been applied to the process and
are therefore still free to be chosen.
In MPC, these future input manipulations are chosen such that the future
behavior at both the process outputs and process inputs is close to the desired
behavior of these process variables. Hence the future input manipulations
are the degrees of freedom that can be used to optimize the future process
behavior. MPC uses a quadratic criterion function for the minimization of
deviations of the desired process output responses:

min { IIWsp (Yref (t, N f ) - Yf (t, Nr))II~ + Ilp~U (t, Ne)ll; } (2.140)
U(t,Nc)

with

u (t) u (t - 1)
u (t + 1) u (t)
~U (t, N e ) = u (t + 2) u (t + 1) (2.141)

u(t+Nc -1)

Past Future

---+----
..... -----~~/-\ ~/
Hao"" maIO, 1'\ TO~l maIO'
Outputs

........ ................................................. )......J ...... / "' '- / '- ..... --'. /'
..................•..•.. I -
I
---====)
k Time t

Fig. 2.46. Relation between the past and future process inputs and the future
process outputs
2.4 Model-Based Glass Melter Control 147

The above optimization problem is solved for each controller interval because
new information, i.e., new measurements from the process, becomes available
to refine the solution. This is called the receding horizon principle of the
controller. The input manipulations are determined over the complete control
horizon. However, only the first sample of the calculated control solution
vector u(t) is actually sent to the process. The matrix Wsp is an output
weight that enables the control system designer to define the distribution of
the error between the desired output behavior Yref(t, N f ) and the actually
predicted future process output behavior Yt( t, Nt) over the different outputs.
In MPC, the matrix Wsp generally is a diagonal matrix with a constant value
per output. This value is frequently specified by its inverse: the so-called
equal concern factor. The move weight p is also a diagonal matrix and is
specified by one parameter per input. This parameter is frequently called the
move suppression factor. The move suppression factor is used to trade-off fast
changes of the corresponding input against the other inputs and against the
outputs.
Observe the dominant role that the process model plays in the above
formulation of the control problem. It is clear that the attainable performance
of the controller is closely related to the quality of the applied process models:
Accurate models enable high-performance control.
In the prediction of the future output behavior it is easily possible to in-
clude disturbance models, i.e., models describing the relation between mea-
sured disturbances and the process outputs. Including the effect of these
disturbances on the future output behavior in the optimization criterion en-
ables the optimization to account for these effects during the calculation of
the future input moves. In fact, this is a feed-forward control action, i.e., the
controller already starts compensating for the disturbance before it actually
becomes visible at the process output. This resembles the behavior of a person
who retracts when someone tries to hit him/her. Retracting minimizes the
pain of the offense. The incorporation of disturbance models in the controller
may drastically improve the controller performance: instead of waiting for
the negative effects of the disturbance to become fully visible at the process
outputs, they are anticipated already on the basis of predictions of process
outputs. But note that, because the model predicts the effect of the measured
disturbance at the output, the actual improvement is completely determined
by the quality of the predictions of the disturbance model.

MPC with Constraints

An essential extension of MPC with respect to the MPC described in the


previous section, is the optimization with constraints. The inclusion of con-
straints gives the MPC the characteristics and flexibility desired by industry.
Constraints can be defined on process inputs, process outputs and additional
variables whose relation with the process inputs can be described by some
linear function:
148 2. Melting and Fining

min {
Uf(t)
Ilwsp (Yref (t) - Yr (t)) II; + Ilp~Uf (t) II; } (2.142)

subject to
adi) S; u (t + i) S; av (i) for i = 1,2,3, ... ,
rdi) S; ~u (t + i) S; rV (i) for i = 1,2,3, ... ,
S; y (t + i) S; /3v (i)
!h (i) for i = 1,2,3, ... .
Expressions aL, /3L, rL, and av, /3v, rV represent the respective lower
and upper limits defined on input variables, output variables, and the rate
of change in the input variables. The constraints not only give the control
system its desired flexibility, but also enable the implementation of complex
control strategies with control hierarchies, as discussed in the introduction to
this section. Constraints are, for example, frequently used to define the oper-
ational requirements, i.e., the operational envelope within which the process
may be operated. Note that constraints in general will limit the attainable
performance of the process as soon as they become active. This is due to the
fact that each active constraint implies loss of a degree of freedom in pro-
cess operation. Constraints usually originate from safety limits and operating
limits related to equipment constraints, which take priority over fulfilling the
criterion function. The criterion function usually represents process perfor-
mance considerations. The process performance directly relates to process
economics.

2.4.3 Extensions of the MPC Technology

The generation of MPC systems widely applied to oil-refining processes has


a number of limitations that restrict broader industrial applicability. On the
one hand, the restrictions are caused by the way the criterion function is
minimized. On the other hand, the models applied in most of these MPC
systems have severe limitations.
The first restriction is related to the fact that the solution of the criterion
function, subject to constraints over the complete future horizon at each sub-
sequent sampling instant, still requires significant computational power. The
actual optimization problem is therefore in general approximated by a sim-
plified problem requiring less computer power. A generally applied approach
is to split the original formulation into two sub-problems: a steady-state prob-
lem and a dynamic problem, which is successively solved. The steady-state
problem rigorously defines an optimal solution that fulfils all constraints and
minimizes the criterion function at steady-state conditions. The solution for
the input and output variables obtained from the steady-state optimization
is then used as a target for the dynamic optimization. The dynamic opti-
mization defines the path that brings the process variables from their current
values to these steady-state targets. In particular the rigorous implementa-
tion of the optimization of the dynamic control problem is computationally
demanding. A number of simplifications are applied especially in this step.
2.4 Model-Based Glass Melter Control 149

These simplifications may significantly deteriorate the dynamic performance


of the controller.
Another limit to the performance stems from the models applied in these
MPC systems and the identification techniques used to determine these mod-
els. Nowadays, the most frequently applied types are
• finite step response (F8R) models,
• finite impulse response (FIR) models,
• low-order transfer function (TF) models, and
• low-order state space (88) models.
These models are obtained from dedicated identification tests applied to
the process. In general they describe only the part of the process dynamics
that is relevant for control. The low-frequency behavior, i.e., the slow process
responses and the steady-state process behavior, is well described by these
models. The restricted validity of the dynamic model is directly determined
by the identification techniques used. The fact that the models do not accu-
rately describe the faster process dynamics relevant for control can have a
direct impact on the performance of the MPC. The restricted validity of the
dynamic model limits the MPC operation to a reduction of the variance of the
slow variations of the process outputs only. The controller cannot compensate
for the faster variations of the process outputs.
The quality improvement of critical product properties to be obtained
with the current MPC generation is therefore restricted. This is important
for problems where quality control, i.e., control of the so-called C pk value
of specified product and process parameters is an important objective (see
Fig. 2.47).
The application of process-identification techniques for determination of
the models for the MPC system generally requires on-line tests, which can
cause a temporary loss of production. The associated high costs severely re-

Probability density function


1.4 .--~~~--.------:--~~----,,-----~---,
CPk = minlllol+,IOli - ml/ 3s
1.2 ,

1.0

£0.8 .:tol+
:0
~ 0.6
e
c.. 0.4
0.2

00 2 3 4 5- 16 7 8 9 10
m Property value

Fig. 2.47. Optimization of the "capability" (Cpk ) of important process variables


and product parameters using model-predictive control
150 2. Melting and Fining

strict the application field of MPC technology. Current process-identification


techniques almost always result in linear dynamic models. Sometimes simple
static, nonlinear functions at the inputs and outputs are applied to approx-
imately describe nonlinear process behavior. This type of MPC systems is
therefore restricted in its ability to control fast changeover between different
operating points of the process and batch processes.
Hybrid models, i.e., models obtained from the integration of first-principle-
based process models (e.g., CFD-based simulation models of melters, refiners,
and forehearths) and models obtained with process identification techniques,
are applied in the latest MPC systems. Hybrid models can not only increase
the accuracy of predictions, they may also drastically reduce the costs asso-
ciated with the modeling phase.
The latest generation of MPC systems copes with the above-discussed
problems. These systems enable operation of processes closer to their physico-
chemical operating limits. In this way, the problems posed in the introduction
regarding the requirements on flexibility, predictability, and complete repro-
ducibility of process operations in conformity with defined specifications be-
come solvable.

2.4.4 Application of MPC in the Glass Industry

A typical application of MPC in the glass industry is the control of crown,


glass, and bottom temperatures in melters, refiners, and forehearths.
Melters have particularly slow dynamics, typically with response times of
several hours up to one day. This is where model-predictive control performs
very well. It consistently updates and keeps track of all applied changes in
heating/cooling adjustments, and the way they work out On all individual
glass temperatures taking into consideration the full history of process ma-
nipulations over several shifts. Moreover, the process of glass melting is a
highly interactive system with both spatial and temporal flow patterns that
connect glass temperatures and the related glass-processing conditions in a
dynamic way. Every change in heating/cooling simultaneously affects almost
all glass temperatures and therefore the processing conditions relevant for
glass quality. The desired temperature profiles are adjusted in such a way
that the average residence time and the residence-time distribution together
with the time-temperature history of each small volume of glass meet speci-
fications that link to product quality.
Finding an optimum for the operation of such a process is not a straight-
forward task. In general, there are three optimization criteria that should be
satisfied with decreasing priority:
1. safety - constraint demands to protect the construction and the equip-
ment from damage;
2. quality - control to meet product specifications and imposed environmen-
tal constraints;
2.4 Model-Based Glass Melter Control 151

3. economic optimization of operation - maximize efficiency and minimize


energy consumption.
To protect the furnace from unacceptable control solutions (e.g., chang-
ing the heating/cooling too fast, damaging the construction), constraints on
heating/cooling levels and crown-temperature profiles and ranges are applied.
This means that the MPC will never violate these safety constraints in order
to satisfy a control objective of a lower priority: "Safety first!"
Most of the time, the process is controlled in a safe operating region,
with room to move the MVs for the purpose of keeping quality variables on
target with minimum variability - despite ever-present disturbances, such as
changing batch compositions and temperature disturbances.
A final optimization objective is minimization of the operating costs. In
the glass industry, this mostly means saving energy, maximizing throughput
at a given quality level, and maximizing efficiency. For each particular control
interval, the "cheapest" solution satisfying all constraints and quality require-
ments is determined. The combined adjustments on all heating and cooling
flows is additionally chosen to minimize costs. In particular for melting fur-
naces, which typically consume a lot of fuel, the potential for cost reduction
is considerable in general.
Normally, a refiner connects to a number of forehearths for the distri-
bution of the glass melt to the forming equipment (e.g., a press for TV
panels/funnels, containers and equipment for drawing tubes). Production
problems or product changeover on one forehearth can severely degrade the
operation of the other forehearths in the form of (inlet) temperature distur-
bances. Applying MPC on the refiner can anticipate problems and minimize
the disturbing effects. Furthermore, the individual MPC of each forehearth
can compensate for the remaining disturbances, long before the effect is felt at
the forehearth exit, where the forming process takes place. Because normal
forehearths use both heating and cooling, conflicting simultaneous adjust-
ments of heating and cooling flows can be avoided, thus saving some energy,
without degrading quality control.
Figure 2.48 shows the dynamic interaction matrix of a typical forehearth,
exhibiting the step responses from each MV to each CV and the corresponding
gains. As can be seen, almost all CVs are simultaneously influenced by almost
all MVs. This is called the "multivariable" character of the process.
A control objective for MPC control on a forehearth in general is to
drive glass-temperature distribution on a vertical cross section near the bowl
or gobber to a specified profile. The aim is to improve the temperature-
distribution conditions of the glass to an optimum profile for further process-
ing.
Figure 2.49 shows a typical operator interface to an MPC-controlled fore-
hearth (ProfileExpert@). Notice the graph, showing converging glass temper-
atures, after the MPC was switched on.
152 2. Melting and Fining

--
[-J [-J [-J [-J [ -J [-J [- J [- J
MV1 MV2 MV3 MV4 MV5 MV6 MV7 MV8
r-- I r -I----
CV 1.97 1.37 0.897 0 .722 0.503 0.187 0 0
.J 1
1, -
CV
r-- I r - r-- ! - - -
.J 2 1.19 1.84 1.46 1.23 0.533 0.188 0.347 0
,.---I r - ,,---
CV

----
.j 3 0.431 0.522 1.02 1.54 1.12 0.623 0.238 0.209

v----.. .

--
CV 0.395 0.279 0.362 0.699 1.2 0.893 0.44 0.239
·j4

CV
V-- -
.j 5 0.24 0.134 0 .228 0.24 0 .453 0.968 1.12 0.511

CV
- I~
.J 6 0.127 0 0.168 0.145 0.233 0.292 0 .537 1.49
o [hours] 10 [hours] 10 [hours] 10 [hours] 10 [hours] 10 [hours] 10 [hours] 10 [hours] 10

, date == 1-Feb-2000, 16:27:13.05

Fig. 2.48. The multivariable character of a typical forehearth model applied in a


model-predictive control system

~~~~----~----~----~----~~--~ ~­
-"'
r ...
~~~~~~~--~~--~~--~~~~~ ~
~

---
~~:~ "",'.

, !llb3ItOI">«IDI1.t ' 21DJ M041 ' · Il\ M1)Ij2 ' 1t01

Fig. 2.49. Model-predictive controller applied to control glass-temperature homo-


geneity
2.4 Model-Based Glass Melter Control 153

.,.,.
D ..........c: .. "U'
U .c,. ...
1:1-""_",,
......
tt• •
lI"th..CI_Ml.
ft..,._~

l'. I'OP'_Cl_ ... .


MI_
ftJ"t.c ...... .
ft_" ..._"'.
~.( _.

... . ..-
. '
. ..
Fig. 2.50. Converging glass temperatures near the bowl (objective: maximum ho-
mogeneity)

Figure 2.50 shows these converging temperatures in more detaiL Fig-


ure 2.51 shows a comparison of the behavior of model prediction versus the
actual behavior of one of the controlled temperatures. The actual tempera-
ture changes match the prediction very closely.
An example of an industrial application of the model-predictive control
system is the control of a forehearth of a TV-panel production line intended
to stabilize temperature profiles and to minimize gob-weight variations.

~~ II
tJ
1\ : 11"6101 .. 1
"" .~ .
...::- ---
1 _ ....- -
~

.. ---....... -........ .... . _.....-...... -··· ......... ....--__.. ................ -··· ...... .....•. ....
"'~
~.,

...
"Ol "Ol ,- aID ,.Ol
w
..,
"'~ .

. -... . --.. ... .. . ..


..... ,"v

.. ...... ··· .. .... ··· . .•


~

. " .... ..oo .. oo


n• •
" .

';... Ju. . . . . IHl • ...................... a,

Fig. 2.51. Actual past and model-predicted future temperature during transition
control
154 2. Melting and Fining

Figure 2.52 shows a comparison between the performance of a traditionally


controlled process and the results obtained with the model-predictive control
system under similar conditions.

••

Fig. 2.52. Performance of a traditionally controlled process compared with the


results gained with the model-predictive control system ProfileExpert ®
References 155

Conclusion and Outlook

An overview has been given of model-based control systems, which are more
and more applied in process industry. The discussed MPC technology is
widely applied in oil-processing industries today. An extension of this proven
technology that is optimized for control of glass-manufacturing processes is
an emerging new technology in glass manufacturing. The bottom line driver
for applying this technology is its widely demonstrated capability to improve
business performance. The break-even point of investments in applications of
this technology is in general reached well within one year.
Dedicated product development based on the MPC technology is ongoing
to extend its applicability to a larger range of processes. The latest develop-
ments of the MPC technology in this respect are:
• robust high-performance control of melters, refiners and forehearths. These
control systems stabilize temperature profiles at conditions that result in a
significant reduction of the variance of critical product parameters and pro-
cess variables. This enables production at desired Cpk values for specified
product quality parameters thus maximizing the margins on the products.
• control of changeovers from one operating point to another along a trajec-
tory in a completely predictable and reproducible way (maximum flexibility
with regard to color, pull, or product-type changes);
• realization of control systems that provide a good balance between devel-
opment and maintenance costs on the one hand, and profitability on the
other hand.
The power of the latest MPC technology has been illustrated by a de-
scription of typical MPC applications in the glass industry. MPC can cope
with safety, quality, and economic demands in the proper context. It is ideally
suited for application to typical glass processes with their dense interaction
matrices and extremely slow dynamics. MPC technology is currently rapidly
developing. Dedicated MPC-based applications for a broad range of glass-
manufacturing processes are just entering the market.

References
2.1 D.A. Nolet, R.A. Murmane: "Development of modelling techniques for glass
furnaces", in 1st Int. Conf. on the Fusion and Processing of Glass (Am.
Ceram. Soc., Alfred, NY 1988) pp. 13.1-13.21
2.2 E. Buckingham: "On physically similar systems: illustrations of the use of
dimensional equations", Phys. Rev. 4(4), 345-376 (1914)
2.3 F.C. Flint, A.K. Lyle: "The flow of glass in tanks", J. Am. Ceram. Soc. 15,
410-418 (1932)
2.4 G. Leyens: "Contribution to the calculation of two-dimensional convection
flows in continuously operated glass-melting tank furnaces", Part 1, "Mathe-
matical model", Glastechn. Ber. 47, 251-259 (1974), Part 2, "Results of the
calculations", ibid., 261-270
156 2. Melting and Fining

2.5 T. Chen, R.E. Goodson: "Computation of three-dimensional temperature


and convective flow profiles for an electric glass furnace", Glass Techno!. 13,
161-167 (1972)
2.6 H. Mase, K. Oda: "Mathematical model of glass tank furnace with batch
melting process", J. Non-Cryst. Solids 38/39, 807-812 (1980)
2.7 A. Moult: "Two and three dimensional mathematical models for glass tank
furnaces", Glass Techno!. 23, 106-112 (1982)
2.8 F. Simonis, H. de Waal, R. Beerkens: "Influence of furnace design and oper-
ation parameters on the residence time distribution of glass tanks predicted
by computer simulation", in Collected Papers XlV. Int. Congress on Glass,
Vo!' 3 (Indian Ceram. Soc., Calcutta 1986) pp. 118-127
2.9 M.G. Carvalho, M. Nogueira, J. Wang: "Mathematical modelling of the glass
melting industrial process", in Proc. XVII. Int. Congress on Glass, Vo!' 6
(Chinese Ceram. Soc., Beijing 1995) pp. 69-74
2.10 H. de Waal: "Mathematical modeling of the glass melting process", Glas-
techno BeL 63K, 1-18 (1990)
2.11 S. Kawachi, Y. Kawase: "Evaluation of bubble removing performance in a TV
glass furnace", Part 1, "Mathematical formulation", Glastechn. BeL Glass
Sci. Techno!. 71, 83-91 (1998)
2.12 J. Chmelar, A. Franek, P. Schill: "Computer aided tracing of bubbles - glass
quality control", in Proc. XVI. Int. Congress on Glass, Vo!' 6 (Deutsche
Glastechnische Gesellschaft, Frankfurt/Main 1992) pp. 45-50
2.13 RG.C. Beerkens, H.P.H. Muijsenberg, T. van der Heijden: "Modelling of
sand grain dissolution in industrial glass melting tanks", Glastechn. BeL
Glass Sci. Techno!. 67, 179-188 (1994)
2.14 R Beerkens, T. van der Heijden, H. Muijsenberg: "A model study on de-
gassing of sulfate refined glass melts", in Proc. XVI. Int. Congress on Glass,
Vo!' 6 (Boletin de la Sociedad Espanola de Ceramica y Vidrio 31-C, Madrid
1992) pp. 15-21
2.15 M.K. Choudhary: "Mathematical modeling of transport phenomena in glass
furnaces: an overview of status and needs", in Proc. ICG Annual Meeting
on Glass in the New Millennium - Challenges and Break Through Technolo-
gies, Amsterdam 15-17 May 2000, session T1.1 (CD Rom, ed. by National
Committee Netherlands Glass Industry, Eindhoven 2000)
2.16 M.K. Choudhary, N.T. Huff: "Mathematical modeling in the glass industry:
an overview of status and need", Glastechn. BeL Glass Sci. Techno!. 70,
363-370 (1997)
2.17 J. Wang: Three-Dimensional Mathematical Modeling of Thermal Phenomena
Occurring in Industrial Glass Melting Tanks, PhD Thesis (Instituto Superior
Tecnico, Universidade Tecnica Lisboa 1998)
2.18 M.G. Carvalho, N. Speranskaia, J. Wang, M. Nogueira: "Modeling of glass
melting furnaces: applications to control, design and operation optimization" ,
Ceramic Transactions 82, 109-135 (1997-1998) (Proc. of the Advances in
Fusion and Processing of Glass II, Toronto 1997)
2.19 M.G. Carvalho, J. Wang, M. Nogueira: "Investigation of glass melting and
fining processes by means of comprehensive mathematical model", Ceramic
Transactions 82, 143-152 (1997-1998) (Proc. of the Advances in Fusion and
Processing of Glass II, Toronto 1997)
2.20 M.G. Carvalho, M. Nogueira, J. Wang: "Mathematical modelling of the glass
melting industrial process", in Proc. XVII. Int. Congress on Glass, Vo!' 6
(Chinese Ceram. Soc., Beijing 1995) pp. 69-74
References 157

2.21 R. Beerkens: "Heutige Moglichkeiten zur Modellierung von Glasschmelzofen


- Voraussagen zur Qualitat des Glasschmelzprozesses", Glastechn. Ber. Glass
Sci. Technol. 71(4), N35-N47 (1998)
2.22 R.G.C. Beerkens, T. van der Heijden, H.P.H. Muijsenberg: "Possibilities of
glass tank modeling for the prediction of the quality of melting processes" ,
Ceram. Eng. Sci. Proc. 14(3-4), 139-160 (1993)
2.23 R.S.O. Prasad, A. Mukhopadhyay, A. Dutta: "Implementation of a glass
batch-melting model in the general purpose three-dimensional CFD code
FLUENT", Proc. V. Int. Seminar on Mathematical Simulation in Glass Melt-
ing, June 17-18, 1999 (Horni-Becva, Czech Republic 1999) pp. 1-9
2.24 A. Ungan, R. Viskanta: "Melting behaviour of continuously charged loose
batch/blankets in glass melting furnaces", Glastechn. Ber. 59(10), 279-291
(1986)
2.25 M.-R. Safaieh: "Glasstromungsuntersuchungen in einer Modellwanne",
Part 2, "Untersuchung der rein thermisch bedingten Konvektionsstro..
mung und ihre Beeinflussung durch Lufteinblasen (Bubbling)", Glastechn.
Ber. 41(4), 129-137 (1968)
2.26 J. Kucera, J. Auerbeck: "Hydrodynamics of the glass-melt flow through a
feeder head", Glastechn. Ber. Special Edition LVI K, Vol. 1 (Glastechnische
Gesellschaft, Frankfurt/Main 1983) pp. 136-14
2.27 V.J. Astanin, K.V. Klimova, L.P. Lebedeva, V.J. Shutkinov: "Stirring the
glass mass in an oval-shaped working tank", Steklo Keram. 38(11/12), 557
(1981) (physical models for stirrers)
2.28 V.O. Aume, B. Hong: "Application of bubbling in a CRT panel furnace",
Ceramic Transactions 82, 251-256 (1997/1998)
2.29 O.A. Prokhorenko, O.V. Mazurin: "High-temperature radiative conductivity
of commercial iron-containing glasses", Ceramic Transactions 82, 323-333
(1997/1998)
2.30 G. Leyens, J. Smrcek, J. Thyn: "Use of isotope measurement technology and
a mathematical model to determine the residence time distribution in a glass
tank", Glastechn. Ber. 53(5), 124-129 (1980)
2.31 J. Bauer, U. Roger, P. Simons: "Validation of a mathematical glass tank
model", Glastechn. Ber. Glass Sci. Technol. 72(6), 171-181 (1999)
2.32 H.P.H. Muijsenberg, F. Simonis: "Verification of 3D mathematical simulation
with measured temperature profiles during furnace operation", Glastechn.
Ber. Glass Sci. Technol. 68C 2, 55-62 (1995)
2.33 J. Wang, B.S. Brewster, M.Q. McQuay, Q. Mardson, B.W. Web: "Valida-
tion of an improved batch model in a coupled combustion space/melt/batch
melting glass furnace simulation", Glastechn. Ber. Glass Sci. Technol. 73(10),
299-308 (2000)
2.34 A.M. Lankhorst, H.P.H. Muijsenberg, M.P.J. Sanders: "Coupled combus-
tion modelling and glass tank modelling in oxy- and air-fired glass-melting
furnaces", in Proc. Int. Symp. on Glass Problems, ICG Annual Meeting,
Sept. 4-6, 1996, Istanbul, Turkey, Vol. 1 (~i§ecam, Istambul, Turkey 1996)
pp.378-384
2.35 A. Backx, J. Ludlage, A. Koenraads: "Application of model predictive control
for the quality of glass melting processes" , International Workshop on Energy
Efficiency in Glass Production, May 18-19, 2000, Amsterdam (TNO-TPD,
Einhoven 2000)
2.36 F. Simonis: "Estimation of the local redox distribution in the melt by numeral
flow modelling", Glastech. Ber. Special Edition LXIIIK, 29-33 (1990)
2.37 J. Klouzek, J. Matyas, L. Nemec, M. Trochta, J. Ullrich: "Mathematical
model of glass quality - bubble distribution in a glass melting furnace",
158 2. Melting and Fining

in Pmc. 5th ESG Conference: Glass Science and Technology for the 21st
Century, ed. by A. Helebrant et a!. (Czech Glass Society, Prague, Czech
Republic 1999) pp. A2.25-A2.34
2.38 F. Kriimer: "Simulation of fining in specialty glass", Ceramic Transac-
tions 82, 7-17 (1997/1998)
2.39 P.R. Laimbock: Foaming of Glass Melts, PhD Thesis (Eindhoven University
of Technology 1998)
2.40 H.A. Lenhart, H.A. Schaeffer: "Elektrochemische Messung der Sauerstoffak-
tivitiit in Glasschmelzen", Glastechn. Ber. 58(6), 139-147 (1985)
2.41 R.G.C. Beerkens: "Redox and sulphur reactions in glass melting processes",
in Pmc. 5th ESG Conference: Glass Science and Technology for the 21st
Century, ed. by A. Helebrant et a!. (Czech Glass Society, Prague, Czech
Republic 1999) p. A1.1
2.42 L. Ortmann: Zum Einfiufl von Sauerstoff und Schwefel auf polyvalente Ele-
mente in Alkali-Erdalkali Silikatglas (AES)-Gemenge und AES-Scherben-
schmelzen, PhD Thesis (TU Bergakademie Freiberg, Germany 1999)
2.43 H.J. Barklage-Hilgefort, K.-W. Mergler, H.-J. Voss: "Stromungsmessungen
im Liiuterbereich einer rekuperativ beheizten Glasschmelzwanne", Glas-
techno Ber. 53(1), 27-36 (1980)
2.44 W. Geffcken: "Vorgiinge der Homogenisierung in der Schmelze, Ausziehen
der Schlieren, Diffusion", Glastechn. Ber. 30, 143-145 (1957)
2.45 H. Becker: "Abbau von Si0 2 -Schlieren in Gliisern durch Diffusion und Stro-
mung", Glastechn. Ber. 35, 387-392 (1962)
2.46 U. Fotheringham, F.-T. Lentes: "Active thermal conductivity of hot glass",
Glastechn. Ber. Glass Sci. Techno!. 67(12), 335-342 (1994)
2.47 F.T. Lentes, N. Siedow: "Three-dimensional radiative heat transfer in glass
cooling processes", Glastechn. Ber. Glass Sci. Techno!. 72(6), 188-19 (1999)
2.48 S.V. Patankar: Numerical Heat Transfer and Fluid Flows (McGraw-Hill, New
York 1980)
2.49 R. Bergman, D.P. Tselepidakis: "Simulation of a TV glass refiner using cou-
pled combustion and glass flow models", Ceramic Transactions 82, 239-244
(1997/1998)
2.50 O.S. Verheyen, R.G.C. Beerkens: "Description of batch-to-melt conversion
for floatglass", in Pmc. of Glass in the New Millennium - Challenges and
Breakthmugh Technologies, Amsterdam, May 15-17, 2000, session T3-T3.8
(CD Rom, ed. by National Committee Netherlands Glass Industry, Eind-
hoven 2000)
2.51 C. Kroger, H. Eligehausen: "Uber das Wiirmeleitvermogen des einschmelzen-
den Glasgemenges", Glastechn. Ber. 32(9), 362-373 (1959)
2.52 R. Conradt, P. Suwannathada, P. Pimkhaokham: "Local temperature dis-
tribution and primary melt formation in a melting batch heap", Glastechn.
Ber. 67(5), 103-113 (1994)
2.53 A.J. Faber, R.G.C. Beerkens, H. de Waal: "Thermal behaviour of a glass
batch on batch heating", Glastechn. Ber. 65(7), 177-185 (1992)
2.54 H. Fuhrmann: "Beitrag zur niiherungsweisen Berechnung des Abschmelzens
von Glasgemengeschichten", Glastechn. Ber. 46, 201-218 (1973)
2.55 M. Daniels: "Einschmelzverhalten von Glasgemengen", Glastechn. Ber. 46,
40-45 (1973)
2.56 P. Hrma: "Complexities of batch melting", in Pmc. 1st Int. Conf. on Ad-
vances in Fusion and Pmcessing of Glass, Alfred NY (Am. Ceram. Soc.,
Westerville, OH 1988) pp. 10.1-10.18
References 159

2.57 F. May, H. Kremer: "Mathematical modeling of glass melting furnace design


with regard to NO x formation", Glastechn. Ber. Glass Sci. Techno!. 72(1),
1-6 (1999)
2.58 B.E. Launder, D.B. Spalding: Mathematical Modelling of Turbulence (Aca-
demic Press, London 1972)
2.59 F.C. Lockwood, N.G. Shah: "A New Radiation Solution for Incorporation
in General Combustion Prediction Procedures", in Proc. 18th Int. Symp. on
Combustion, Waterloo, Canada (The Combustion Institute, Pittsburgh, PA
1981) pp. 1405-1414
2.60 W.M. Rohsenow, J.P. Hartnett (Eds): Handbook of Heat Transfer, contribu-
tions by Hottel and Egbert in S.S. Penner, "Radiation" (McGraw-Hill, New
York 1973) pp. 15.76-15.77
2.61 L. Post: Modelling of Flow and Combustion in a Glass Melting Furnace, PhD
Thesis (Delft University of Technology 1988)
2.62 T.W.J. Peeters: Numerical Modeling of Turbulent Natural-Gas Diffusion
Flames, PhD Thesis (Delft University of Technology 1995)
2.63 LM. Khan, G. Greeves: "A method for calculating the formation and com-
bustion of soot in diesel engines", in Heat Transfer in Flames, 1st edn., ed.
by N.H. Afgan, J.M. Beer (Wiley, New York 1974) pp. 391-404
2.64 B.F. Magnussen, B.H. Hjertager: "On mathematical modeling of turbulent
combustion with special emphasis on soot formation and combustion", in
16th Symp. (Int.) on Combustion (The Combustion Institute, Pittsburgh
PA 1976) pp. 719-729
2.65 M.K. Choudhary: "A modeling study of flow and heat transfer in the vicinity
of an electrode", in Proc. XVII. Int. Congress on Glass, Vo!' 6 (Chinese
Ceram. Soc., Beijing 1995) pp. 100-107
2.66 M.K. Choudhary: "Free convection effects on the dissolution of a spherical
particle", in 1st Int. Conf. on the Fusion and Processing of Glass (Am. Ce-
ram. Soc., Alfred NY 1988) pp. 11.1-11.23
2.67 F. Kramer: "Mathematisches Modell der Veranderung von Glasblasen in
Glasschmelzen", Glastechn. Ber. 52(2), 43-50 (1979)
2.68 L. Nemec: "The behaviour of bubbles in glass melts. Part 2. Bubble size
controlled by diffusion", Glass Techno!. 21(3), 139-144 (1980)
2.69 L. Nemec: "The behaviour of bubbles in glass melts. Part 1. Bubble size
controlled by diffusion and chemical reaction", Glass Techno!. 21(3), 134-
138 (1980)
2.70 L. Nemec, M. Muhlbauer: "Verhalten von Glasblasen in der Schmelze bei
konstanter Temperatur", Glastech. Ber. 54(4),99-108 (1981)
2.71 M.C. Weinberg, P.LK. Onorato, D.R. Uhlmann: "Behavior of bubbles in glass
melts. Part 2. Dissolution of a stationary bubble containing a diffusing and
a non-diffusing gas", J. Am. Ceram. Soc. 63(7/8), 435-438 (1980)
2.72 R.G.C. Beerkens: "Chemical equilibrium reactions as driving forces for
growth of gas bubbles during refining", Glastechn. Ber. 63K, 222-242 (1990)
2.73 C. Russel, E. Freude: "Voltammetric studies ofthe redox behaviour of various
multivalent ions in soda-lime-silica glass melts", Phys. Chern. Glasses 30(2),
62-68 (1989)
2.74 H.O. Mulfinger: "Gasanalytische Verfolgung des Lautervorganges im Tiegel
und in der Schmelzwanne", Glastechn. Ber. 49, 232-245 (1976)
2.75 A.J. Faber, R.G.C. Beerkens: "Reduction of refractory corrosion in oxy-fuel
glass furnaces", Proc. 18th Int. Congress on Glass, July 5-10, 1998, San
Francisco (Am. Ceram. Soc., Westerville, OH 1998)
2.76 M. Cable, C. Apak, M.A. Chaudhry: "The kinetics of volatilization from lead
glasses", Glastechn. Ber. 48(1), 1-11 (1975)
160 2. Melting and Fining

2.77 J. Matousek, J. Hlavac: "A study of the volatilisation of lead glass", Glass
Technol. 12(4), 103-106 (1971)
2.78 M. Cable, M.A. Chaudhry: "Volatilization from soda-lime-silica melts at one
atmosphere and reduced pressures", Glass Technol. 16(6), 125-134 (1975)
2.79 D.M. Sanders, H.A. Schaeffer: "Reactive vaporization of soda-lime-silica glass
melts", J. Am. Ceram. Soc. 59(3/4),96-101 (1976)
2.80 D.M. Sanders, W.K. Haller: "Effect of water vapor on sodium vaporization
from two silica-based glasses", J. Am. Ceram. Soc. 60(3/4), 138-141 (1977)
2.81 R.B. Bird, W.E. Stewart, E.N. Lightfoot: Transport Phenomena (Wiley, New
York 1960) p. 513
2.82 H. Schlichting: Boundary Layer Theory (Pergamon, New York 1955) pp. 438-
439
2.83 V. Gnielinski: "Warmeubertragung bei erzwungener einphasiger Stromung",
in VDI Wiirmeatlas (VDI-Verlag, Dusseldorf 1991)
2.84 J.M. Coulson, J.F. Richardson: Chemical Engineering, Vol. 1, 4th edn.
(Pergamon, Oxford 1990) chapter 10, pp. 499-500
2.85 M. Cable, M.H. Fernandes: "Volatilization of molten sodium metaborate with
convection of the furnace atmosphere", Phys. Chern. Glasses 39(4), 228-235
(1998)
2.86 J. Crank: The Mathematics of Diffusion (Clarendon, Oxford 1956) p. 34
2.87 C.W. Bale, A.D. Pelton: FACT- Win-User Manual (CRCT, Ecole Poly tech-
nique de Montreal, Quebec 1999)
2.88 R. Conradt, H. Scholze: "Zur Verdampfung aus Glasschmelzen", Glastechn.
Ber. 59(2), 34-52 (1986)
2.89 B.B. Argent, K. Jones, B.J. Kirkbride: "Vapors in equilibrium with glass
melts", Chern. Soc. Spec. Publ. 34, 379-390 (1980)
2.90 R. Bruckner: "Wechselwirkungen zwischen Glasschmelze und Feuerfestmate-
rial", Glastechn. Ber. 53(4), 77-88 (1980)
2.91 B. Krabel: Korrosionsverhalten moderner feuerfester Baustoffe in Glass-
chmelzen, PhD Thesis (Technical University, Berlin 1996)
2.92 H.P.H. Muijsenberg, F. Simonis, E.G.J. Peters: "Studying time transient
behaviour of glass melting tanks by mathematical simulation models", in
Proc. XVII. Int. Congr. on Glass, Vol. 6 (Chinese Ceram. Soc., Beijing 1995)
pp. 139-144
2.93 R. Beerkens: "Development of a method for the control of glass melt quality
in industrial glass furnaces - 'REGLA' ", project proposal Economy-Ecology-
Technology program (June 1999)
2.94 W.S. Kuhn: "Approximate analytical solutions for the heat transfer in glass
melting furnaces", Glastechn. Ber. Glass Sci. Technol. 92(2), 27-41 (1999)
2.95 W. Muschick, E. Muysenberg: "Round robin for glass tank models",
Glastechn. Ber. Glass Sci. Technol. 71(6), 153-156 (1998)
2.96 W.S. Kuhn: "Approximate analytical solutions for the heat transfer in glass
melting furnaces", Glastech. Ber. Sci. Technol. 72(2), 27-41 (1999)
2.97 R. Conradt, P. Suwannathda, P. Pimkhaokham: "Local temperature distri-
bution and primary melt formation in a melting batch heap" , Glastechn. Ber.
Glass Sci. Technol. 67(5), 103-113 (1994)
2.98 R. Conradt: Einschmelzen von Gemenge und Glasscherben. Grundlagen des
industriellen Glasschmelzprozesses (HVG, Frankfurt/Main 1999)
2.99 C. Kroger: "Theoretischer Warmebedarf der Glasschmelzprozesse", Glas-
techno Ber. 26(7), 202-214 (1953)
2.100 C. Madivate: "Thermochemistry of the glass melting process - energy re-
quirement in melting soda-lime-silica glasses from cullet-containing batches",
Glastechn. Ber. Glass Sci. Technol. 69(6), 167-178 (1996);
References 161

C. Madivate: "Calculation of the theoretical energy requirement for melting


technical silicate glasses", J. Am. Ceram. Soc. 81(12), 3300-3306 (1998)
2.101 C. Kroger: "Gemengereaktionen und Glasschmelze", Glastechn. Ber. 25(10),
307-324 (1952)
2.102 D. Angermeier, W.S. Kuhn, R. Druihle, D. Ballutaud, R. Triboulet: "Initial
growth processes in the epitaxy of Ge with GeH4 on oxidized Si substrates" ,
J. Electrochem. Soc. 143(2), 694-697 (1996)
2.103 G. Eriksson: "Thermodynamic studies of high temperature equilibriums",
Acta Chern. Scand. 25,2651-2658 (1971)
2.104 L. Bodalbhai, P. Hrma: "The dissolution of silica grains in isothermally
heated batches of sodium carbonate and silica sand", Glass Technol. 27(2),
72-78 (1986)
2.105 W. Buss: "Erhitzungsmikroskopische Untersuchungen von Vorgangen in
Glasschmelzen mit und ohne Sulfat", Glastechn. Ber. 35(4), 167-176 (1962)
2.106 L. Stoch, S. Kraishan: "Interface phenomena accompanying the early stages
of glass batch reactions - a model study", Glastechn. Ber. Glass Sci. Tech-
nol. 70(10), 298-305 (1997)
2.107 A. Ungan, R. Viskanta: "Melting behavior of continuously charged loose
batch blankets in glass melting furnaces", Glastechn. Ber. 59(10), 279-291
(1986)
2.108 W.H. Manring, W.C. Bauer: "Influence of batch preparation process on melt-
ing and fining of glass", The Glass Industry, 354-357 (1964)
2.109 H. Scholze, H. Salmang: Keramik, Teil 1, Allgemeine Grundlagen und wich-
tige Eigenschaften (Springer, Berlin, Heidelberg 1982)
2.110 P. Hrma: "Batch melting reactions", in Chemistry of Glass, 2nd edn., ed. by
A. Paul (Chapman & Hall, London 1990) pp. 157-176
2.111 P. Hrma: "Reaction between sodium carbonate and silica sand at 874°C
< T < 1022 °C", J. Am. Ceram. Soc. 68(6), 337-341 (1985)
2.112 M. Cable, D. Martlew: "Effective binary diffusivities for the dissolution of
silica in melts of the sodium carbonate-silica system", Glass Technol. 26(5),
212-217 (1985)
2.113 M.C. Weinberg, P.I.K. Onorato, D.R. Uhlmann: "Behaviour of bubbles in
glass melts. Pt. I, Dissolution of a stationary bubble containing a single
gas", J. Am. Ceram. Soc. 63(3-4), 175-180 (1980)
2.114 H. de Waal: "Mathematical modelling of the glass melting process",
Glastechn. Ber. 63K, 1-18 (1990)
2.115 R.G.C. Beerkens, H.P.H. Muijsenberg, T. van der Heijden: "Modelling of
sand grain dissolution in industrial glass melting tanks", Glastechn. Ber.
Glass Sci. Technol. 67(7), 179-188 (1994)
2.116 M.K. Choudhary: "The effect of free convection on the dissolution of a spher-
ical particle in a viscous melt", Glass Technol. 29(3), 100-102 (1988)
2.117 M.K. Choudhary: "Dissolution of polydisperse silica grains in glass melt-
analysis", J. Am. Ceram. Soc. 73(10), 3053-3058 (1990)
2.118 A.R. Cooper, Jr., W.D. Kingery: "Dissolution in ceramic systems: I. Molec-
ular diffusion, natural convection and forced convection studies of sapphire
dissolution in calcium aluminium silicate", J. Am. Ceram. Soc. 47, 37-43
(1964)
2.119 Stein-Heurtey, France: unpublished document
2.120 W. Trier: Glass Furnace Design Construction and Operation (Society of Glass
Technology, Sheffield 1987)
2.121 R. Conradt: "Some fundamental aspects of the relation between pull rate
and energy consumption of glass furnaces", Int. Glass J. 108, 22-26 (2000)
162 2. Melting and Fining

2.122 RB. Bird, W.E. Stewart, E.N. Lightfoot: Transport Phenomena (Wiley, New
York 1960)
2.123 H. Fuhrmann: "Beitrag zur naherungsweisen Berechnung des Abschmelzens
von Glasgemengeschichten, Teil I. Theoretische Abteilungen, Teil II. Nume-
rische Ergebnisse", Glastechn. Ber. 46, 201-208, 209-218 (1973)
2.124 M. Czerny, L. Genzel: "Energiefluss und Temperaturverlauf im Glasbad von
Schmelzwannen als Folge von Warmeleitung und Warmestrahlung", Glas-
techno Ber. 25, 387-392 (1952)
2.125 R Viskanta: "Heat transfer in glass", AIChE Symp. Ser. 81(245), 63-69
(1985)
2.126 X. Wu, R Viskanta: "Modelling of heat transfer in the melting of a glass
batch", J. Non-Cryst. Solids 80, 613-622 (1986)
2.127 H.J. Hermann: "Simulating granular media on the computer", in 3rd Granada
Lectures in Computational Physics, ed. by P.L. Garrido, J. Marro (Springer,
Berlin, Heidelberg 1995) pp. 67-114
2.128 A. Hansen, D. Bideau (Eds): Disorder and Granular Media (North-Holland,
Amsterdam 1992)
2.129 A.C.P. Pugh: "A method of calculating the melting rate of glass batch and
its use to predict effects of changes in the batch", Glasteknisk Tidskrift 23,
95-104 (1968)
2.130 C. Koger, H. Eliehausen: "Heat conductance of melting glass batches",
Glastechn. Ber. 32, 362-373 (1959)
2.131 M.J. Austin, D.E. Bourne: "Mathematical model of an electric glass furnace",
Glass Techno!. 14(3), 78-84 (1973)
2.132 G. Leyens: "Beitrag zur Berechnung zweidimensionaler Konvektionsstromun-
gen in kontinuierlich betriebenen Glasschmelzwannen, Glastechn. Ber. 47,
251-259 (1974)
2.133 H. Mase, K. Oda: "Mathematical model of glass tank funace with batch
melting process", J. Non-Cryst. Solids 38/39, 807-812 (1980)
2.134 G. Hilbig: "Die raumlich gemittelten Temperaturfelder von Schmelze und Ge-
menge in Elektroglasschmelzwannen, Silikattechnik 35(11), 331-333 (1984)
2.135 V.G. Hilbig, H. KirmBe: "Das Temperaturfeld im Gemengekeil brennstoffbe-
heizter Glasschmelzwannen", Glastechn. Ber. 59(6), 169-173 (1986)
2.136 A. Ungan: Three-Dimensional Numerical Modeling of Glass Melting Process
PhD Thesis (Purdue University, West Lafayette, Indiana, USA 1985)
2.137 P. Schill: "Mathematisches Modell des Abschmelzens der Gemengeschicht in
vollelektrischen ()fen", Glastechn. Ber. 56K(I), 179-184 (1983)
2.138 M.G.M.S. Carvalho, M. Nogueira: "Glass quality evaluation via three-
dimensional mathematical modelling of a glass melting furnace", 1st Int.
Conf. on the Fundamentals of the Glass Manufacturing Process, Sheffield,
UK, Sept. 1991;
M.G. Carvalho, M. Nogueira: "Modelling of glass melting industrial pro-
cess", J. Phys. IV, Colloque C7, supplement au Journal de Physique III,
Volume III, 1357-1366 (1993)
2.139 M.G. Carvalho: Computer Simulations of a Glass Furnace, PhD Thesis (Im-
perial College London, London 1983)
2.140 P. Schill: "Batch melting in mathematical simulation of glass furnaces", in
Proc. 3rd Int. Seminar on Mathematical Simulation in Glass Melting (Glass
Service Ltd., Vsetin 1995) pp. 97-101
2.141 L. Song, Ch. Sun: "Mathematical model of float glass tank furnace", Proc. of
the XVIIth Int. Congr. on glass, Beijing 1995, J. Chinese Ceram. Soc. 23(5),
75-80 (1995)
References 163

2.142 S. Yamamura, H. Itoh, H. Todoriki, M. Iga: "Evaluation of the effect of batch


shape on glass flow by mathematical model", in Pmc. 4th Int. Seminar on
Mathematical Simulation in Glass Melting (Glass Service Ltd. Vsetin, Czech
Republic 1997) pp. 96-104
2.143 R. Beerkens: "Heutige Moglichkeiten zur Modellierung von Glasschmelzofen:
Voraussagen zur Qualitiit des Glasschmelzprozesses", in FachausschujJbe-
richt Nr. 75, Glasschmelzofenbau - Stand der Technik - FachausschujJ II,
(Deutsche Glastechnische Gesellschaft, Frankfurt/Main 1998)
2.144 J. Bauer, U. Roger, P. Simons: "Validation of a mathematical glass tank
model", Glastechn. Ber. Glass Sci. Techno!. 72, 171-181 (1999)
2.145 R.A. Bauer, O. Op de Camp, P. Simons, O. Verheijen, M. Noot,
A.M. Lankhorst: "Advanced possibilities of the TNO glass tank model and fu-
ture developements", in Selection of Activities TNO Glass Technology (TPD
Materials Research and Technology, Eindhoven 1998)
2.146 S. Kawachi, Y. Kawase: "Evaluation of bubble removing performance in a TV
glass furnace, Part 1. Mathematical formulation", Glastechn. Ber. Glass Sci.
Techno!. 71(4),83-91 (1998); "Evaluation of bubble removing performance in
a TV glass furnace, Part 2. Verification using real furnace data", Glastechn.
Ber. Glass Sci. Techno!. 71 (5), 111--119 (1998)
2.147 J. Wang, S. Brewster, B.W. Webb, M.Q. McQuay, K. Bhatia: "A coupled
combustion space/batch/melt tank model for an industrial float glass fur-
nace", in Pmc. 5th Int. Seminar on Mathematical Simulation in Glass Melt-
ing (Glass Service Ltd., Vsetin, Czech Republic 1999) pp. 84-93
2.148 J. Wang: Three-Dimensional Mathematical Model of Thermal Phenomena
Occurring in Industrial Glass Melting Tanks, PhD Thesis (1ST, Lisbon, Por-
tugal 1998)
2.149 R.O.S. Prasad, A. Mukhopadhyay, A. Dutta: "Implementation of a glass
batch melting model in the general purpose three-dimensional CFD code
Fluent", in Pmc. 5th Int. Seminar on Mathematical Simulation in Glass
Melting (Glass Service Ltd., Vsetin, Czech Republic 1999) pp. 43-51
2.150 W.S. Kuhn, F. Marmonier, D. Bessette, A. Duttar Muralidharan: "First
principles batch modelling and validation on a large cold top sodium silicate
me Iter" , Contribution at the XIX. Int. Congr. on Glass, Edinburgh, July
2001
2.151 Stein Heurtey: unpublished document (1998)
2.152 P. Hrma: "Thermodynamics of batch melting", Glastechn. Ber. 55, 138-150
(1982)
2.153 E. Obser, S. Lepert, S. Lelandais: "Image processing for glass industry", in
QCA V '98, Int. Conf. on Quality Control by Artificial Vision (Takamatsu,
Kagawa, Japan 1998)
S. Lepert, C. Barret: "Apport de la logique foue pour l'industrie verriere",
Journees francophones sur la logique floue et ses applications, LFA (Valenci-
ennes, France 1999) proceedings published by Cepadues
2.154 A.J. Faber, R.G.C. Beerkens, H. de Waal: "Thermal behaviour of glass batch
on batch heating", Glastechn. Ber. 65(7), 177-185 (1992)
2.155 P. Costa: "Untersuchung des Einschmelzverhaltens von pelletiertem Gemen-
ge zur Glasherstellung", Glastechn. Ber. 50(1), 10-18 (1977)
2.156 H.R.S. Jack, J.A.T. Jacquest: "Heat transfer in glass-batch materials", Re-
search Laboratories, Pilkington Brothers Ltd., St. Helens, Lancashire (Eng-
land)
2.157 C. Eden: Hochfrequenzerwiirmung von Glas, Diploma Thesis (Jena 1950)
2.158 A.J. Schwab: Begriffswelt der Feldtheorie, 4th edn. (Springer Verlag, Berlin,
Heidelberg 1993)
164 2. Melting and Fining

2.159 K. Kupfmiiller: Einfiihrung in die theoretische Elektrotechnik, 13th edn.


(Springer, Berlin, Heidelberg 1990)
2.160 RF. Sekerka, RA. Hartzell, B.J. Farr: "Instability phenomena during the
RF heating and melting of ceramics", J. Crystal Growth 50, 783-800 (1980)
2.161 E.J. Davies: Conduction and Induction Heating (Peregrinus, London 1990)
2.162 J. Jackson: Classical Electrodynamics, 2nd edn. (Wiley, New York 1975)
2.163 J. Richalet, A. Rault, J.L. Testud, J. Papon: "Model predictive heuristic
control: applications to industrial processes", Automatica 14,413-428 (1978)
2.164 C.R Cutler, B.L. Ramaker: "Dynamic matrix control - a computer con-
trol algorithm", Proc. Joint Automatic Control Conf. (IEEE, Piscataway,
NJ 1980) paper WP5-B, pp. 1-6
2.165 S.J. Qin, T.A. Badgwell: "An overview of industrial model predictive control
technology", AIChE Symp. Ser. 316(93), 232-256 (1996)
3. Homogenizing and Conditioning

3.1 The Intensity of Mixing Processes

Henry Eisermann, Ulrich Lange, Horst Loch, Gunter Weidmann


Introduction

Tracer particles in high-viscosity fluids, such as polymer or glass melts, ex-


hibit complex kinematics during mixing processes. At first glance, this is sur-
prising because high viscosities are usually associated with "simple" flows.
Nevertheless, a more thorough analysis shows that deterministic chaos is at
work and that this is synonymous with good mixing.
The purpose of this work is to present feasible mathematical methods for
a realistic assessment and improvement of the mixing effect of stirrers in glass
melts. This is important because platinum stirrer systems are very expensive,
and if they do not guarantee the desired homogeneity, the economic conse-
quences for the production are serious. Of course, the same methods can also
be applied for the analysis of the mixing effect of melting and refining tanks.
Predicting the homogeneity of glass melts quantitatively after the mix-
ing process is difficult if all parameters of influence, such as diffusion and
chemical reactions, are taken into account. However, statements about the
absolute mixing quality are not necessarily required in practical process de-
velopment. It is usually more important to improve and standardize already
existing stirrers, and this can be achieved by comparing the mixing intensity
of different types of stirrers. In this chapter, we only discuss the most funda-
mental mixing mechanism, namely the mechanically induced increase of the
interfacial area between striae and basic glass, which is called mixing in the
narrow sense.

3.1.1 Description and Quantification of Mixing Processes

The description and evaluation of the homogenization of glass melts is a


complex problem. In this introductory section, we will precisely define the
task and give an outline of a feasible working strategy.
The following considerations introduce the matter very briefly. They are
far from being complete. The purpose is simply to make the state-of-the-art

H. Loch et al. (eds.), Mathematical Simulation in Glass Technology


© Springer-Verlag Berlin Heidelberg 2002
166 3. Homogenizing and Conditioning

accessible to the reader in a very intuitive and condensed manner. Readers


who are interested in details may refer to the bibliography at the end of this
chapter.
First, let us clarify the meaning of two concepts: "mixing" denotes the
purely mechanical process in which a substance (solid, fluid, or gas) is mixed
with another one. If a fluid 1 is soluble in a fluid 2 and if the two fluids are
mixed and dissolved in each other, this process is called "homogenization".
We will not deal with the large number of other possibilities (emulsifying,
suspending, etc.). The two concepts defined above are sufficient for our ap-
proach.
We start with a thought experiment: let a vessel contain a homogeneously
colored fluid where a small blob of a passive tracer with another color mark-
ing has been introduced locally. Now we stir the fluid in the vessel in an
unspecified way for a finite time interval and then assess the homogeneity.
We define perfect mixing as a state where the tracer color is spread out in the
fluid so uniformly that colored structures are no longer visible. Instead, only
a homogeneous, mixed color remains. Obviously, two questions arise: how can
the mixing state be assessed at any time, and which flows lead to a good mix-
ing state? As already mentioned, we will only deal with the second question,
because this leads to the most effective progress in the situations typically
occurring in practice. We are interested in improving existing systems (that
is, in relative statements) rather than in theoretical absolute evaluations.
The local inhomogeneities in glass melts are called "striae", i.e., all inho-
mogeneities that can in principle be dissolved without any trace on technical
time scales. In the following, we will deal exclusively with the elimination of
such striae.
Owing to the high viscosities that are usual in glass melts, virtually no
turbulent flow can be generated. Hence, microscopic homogenization can be
brought about only via diffusion. Diffusion in fluids is a very slow process
that is not effective for technically relevant time and length scales. Hence, a
premixing is necessary in order to enlarge as much as possible the interface
where diffusion takes place. Let us state this first important conclusion: A
strong growth of the interface between the basic glass and the striae is a
prerequisite for good homogenization in glass melts.
The behavior of the interface in real glass melts is, of course, not only
influenced by the flow, but also by many other mechanisms:
• temperature distribution,
• density and viscosity of the striae relative to the basic glass,
• interfacial forces between the striae and the basic glass,
• diffusion processes,
• chemical reactions, and so on.
We will concentrate on the simplest and most important case, which will
help us to find a feasible way for assessing and improving glass stirrers in
practice: the mixing of passive tracers with a homogeneous basic glass.
3.1 The Intensity of Mixing Processes 167

By definition, a passive color tracer follows the fluid flow without any
reaction (no resistance to flow, no diffusion, no chemical reactions, etc.),
that is, the mixing is a purely kinematic process in the fluid. In the context
of a simulation, following a tracer thus means mere post-processing for the
computation of the flow.
How can we follow the tracer numerically? There are two possibilities:
solving an additional diffusion equation with an extremely small diffusion
constant modeling the propagation of the tracer, or following the paths of
many particles. More accurate investigations have shown that the method
of solving an additional diffusion equation in our case leads to considerable
numerical problems. Therefore, we chose the second possibility. The basis for
our visualization and evaluation of the mixing process is provided by following
many particle paths and streaklines.
Returning to the mixing process: how can a strong growth of the interfaces
between the species be achieved in a purely kinematic way? Consider the
following prototype of a simple mixing strategy: a fluid volume with the
shape of a parallelepiped of edge lengths a and b and thickness d is stretched
in the direction of b to double its original length (i.e., half of its original
thickness), and cut in the middle of length b into two pieces with half the
thickness of the original parallelepiped and the same dimensions a and b.
Next, the two pieces are laid on top of one another so that a parallelepiped of
the original dimensions is obtained. This process obviously generates layers.
n transformations result in 2 n layers. If the fluid volume consists of different
layers of species lying on top of each other in the beginning, the number
of interfaces between the species grows exponentially during this process.
The efficiency of this transformation can be demonstrated with a simple
experiment with layers of clay or plasticine of different colors (rolling out
and folding back ten times already leads to 1024 layers!). Thus, we have a
hot trail to follow in order to generate a strong growth of interfacial area
systematically: A strong growth of interfacial area is achieved by exponential
stretching and folding in the entire fluid volume.
Up to now, we have considered the above-mentioned transformation only
with respect to the formation of layers (that is, orthogonally to the layer
planes). It is interesting, however, that this transformation is also the clas-
sic prototype of a kinematic process that leads to deterministic chaos in the
stretching direction (that is, tangentially to the layer planes). If we mark a
point in the beginning and follow its path, the point jumps back and forth
"randomly" each time the volume is stretched and folded back although its
path is generated according to a very simple (however, nonlinear!) law. Ac-
tually, this law is the classic prototype of a deterministic random number
generator, which is a standard part of every mathematics software library:
dividing with a remainder. Chance is involved in so far as, due to exponential
growth, the rounding errors in the last decimal digit make any predictions
168 3. Homogenizing and Conditioning

impossible after a few iterations. Those readers who are interested in details
should refer to Davies [3.5].
If we apply these principles to mixed flows in high-viscosity fluids, we find
that in "simple" Eulerian flow fields (compared to really turbulent flows),
"Lagrangian chaos" can also occur, that is, chaotic kinematics of particle
paths. Exponential stretching and folding is equivalent to deterministic chaos.
On the one hand, such a flow is exactly what we want for a good mixing
effect. On the other hand, it causes a serious numerical problem because in
flows with complex kinematics, the computation of particle paths is, of course,
extremely sensitive to any kind of numerical or model-dependent error and
to initial and boundary conditions. Particles that are close together at the
inlet of the stirrer are far away from each other at the outlet. However,
because particle paths are to be the basis of all our assessments, it is obvious
that, apart from a great deal of care in following the paths, only a statistical
evaluation of many paths can be of any help.
What has been said up to now can be summarized in the following me-
thodical approach for a comparative evaluation of the efficiency of glass-
mixing flows: The assessment of the mixing efficiency in glass melts is made
by statistical evaluation of the mechanically induced interfacial growth along
many particle trajectories.
We would like to emphasize that the graphical representation of many
particle paths and spot diagrams alone often provides interesting qualitative
insight into the behavior of a stirrer. Therefore, such a representation should
always be the starting point of an analysis. The more thorough methods
described in this work complete this rather intuitive procedure. They offer
insight into the local action of the "mixing motor" and allow a quantitative
assessment of stirrers.
Here are a few references for those readers who would like to look into
the original literature: Hiby [3.3] gives a brief introduction into the definition
and measurement of the mixing quality (the subject we omitted). In Pahl et
al. [3.4] you find a survey of the state-of-the-art in 1998 from the point of
view of chemical process engineering including, in particular, typical stirrers.
Ottino's book [3.1] is a standard work on the kinematics of mixing. Hinch [3.2]
gives a very short survey that covers similar material. Cable [3.6] in 1990
presented the topic from the point of view of glass production. The thesis
by Riehl [3.7], an early publication, already contained much of the material
discussed here, particularly tailored to melting tanks. Finally, Aref et al.
[3.8] gives a survey of the application of other mathematical methods from
the field of deterministic chaos, which in part are far more sophisticated, to
the problem of mixing. This last publication demonstrates in particular that
we still have no complete, comprehensive theory of mixing. Instead, there is
only a variety of approaches with different focuses.
3.1 The Intensity of Mixing Processes 169

3.1.2 Flows and Particle Paths in Stirrers

Let us start with some technical preliminary remarks on the computation of


flow fields and on the particle paths derived. We concentrate on the case of
rotating stirrers, that is, rigid stirrer blades moving in the flow (see also movie
on CD). This case, with the boundaries of the flow field varying in time, is the
most difficult one and, of course, it implicitly contains the simpler situations
where all boundaries are constant (as, for example, in a tank, a channel, or
a static stirrer).
We adopt the following conventions:
• The set of all points in the system is called D C R3 .
• D is defined in such a way that at any time any particle can be assigned
a position in the system even if the particle has already left the limited
space of the stirrer we are interested in (this is achieved, for example, by
extending the stirrer space by an infinitely long pipe at the outlet). This
condition, which at first may sound a little cryptic, will save us tedious
case distinctions in the following discussion.
• All mappings that occur are sufficiently smooth.

Velocity Fields

From the Navier-Stokes equations and the specific boundary conditions of


the system, the following velocities are derived:

v:DxR----+
(3.1)
(q, t) r--+

For moving stirrer blades, the computation of the velocities is in itself already
a challenging numerical problem. In the commercial finite-element and finite-
volume packages that we apply for this purpose, several methods have proved
their feasibility in the last few years. Up to now, we have used three of them.
The first method takes advantage of the fact that a stirrer with a rota-
tionally symmetric outer wall (crucible) and corresponding axial inflow and
outflow can be computed in a coordinate system rotating with the stirrer
blades, where the outer wall has a constant velocity (namely the rotational
speed). The stirrer is not necessarily rotationally symmetric, i.e., we still have
a 3D computation. The problem of moving stirrer blades is thus reduced to
a simple coordinate transformation. The contours of the blades are modeled
exactly in a fixed mesh. Under the above-mentioned conditions, this method
is by far the most efficient one, and it does not contain any additional ap-
proximations. However, it is not applicable if the outside contours are not
rotationally symmetric (unfortunately, this is the most common case).
170 3. Homogenizing and Conditioning

With the "moving body method", moving parts are not meshed and
moved explicitly. Instead, a stationary mesh is introduced, and during the
computation the actual position of the stirrer blades is determined for each
time interval. Then, a velocity constraint is imposed on all nodes that are
inside the stirrer blades at that time, so that the movement of the fluid at
that point is equal to that of the stirrer. The moving parts behave, so to
speak, like "ghosts" that do not appear physically, but influence the fluid ap-
proximately like a real moving solid body at that point. The accuracy of this
method obviously depends in particular on the fineness of the mesh, its great
advantage being the fact that it requires relatively little additional numerical
effort. Moreover, the stirrer blades do not have to be meshed explicitly. It is
sufficient to describe their interior by equations. This approach is useful if
one wishes to get a basic survey quickly. For flows with large local velocity
gradients (good mixing flows are exactly of this kind), however, it introduces
a systematic error, which finally reduces the accuracy of the computation of
particle paths. Strictly speaking, one must examine in every individual case
whether the method can be applied. Here, a comparison with the method
of rotating coordinate systems suggests itself if rotationally symmetric outer
walls can be assumed.
The "sliding mesh method" meshes the domains in space with rigid and
moving geometry separately, and the flow equations are solved separately for
each of these domains. For the rotating parts (i.e., the stirrers), the method of
rotating coordinate systems discussed above suggests itself. At the interfaces
between the domains, an interpolation is made that takes into account the
actual relative positions of the different meshes for each time interval and
then establishes an appropriate connection between the conditions of flow on
both sides of the interface. This method requires the most computing power,
but it also provides the most accurate results in general situations if the
meshes are sufficiently fine.
From now on, we consider the velocities as given. The system can be
stationary or non-stationary. The latter case can be further divided into pe-
riodic and non-periodic systems. The system is periodic with the time period
TO E R if

v(q, t) = v(q, t + TO) V(q, t) E R x fl . (3.2)


The glass flows considered here are highly viscous, therefore the stirrer
flows can be regarded as periodic (the actual velocity state of the melt virtu-
ally depends only on the actual position and speed of the stirrer blades, and
hardly ever on the history).

Particle Paths and Flow Functions

A parameterized set of ordinary differential equations can be defined through


the velocity field: for an arbitrary (q, s) E fl x R, let cq,s be the solution to
the initial value problem
3.1 The Intensity of Mixing Processes 171

Cq ,s(s) = q . (3.3)
The existence and uniqueness of cq,s follows from the theory of ordinary
differential equations. cq,s is called a particle path in D with the initial point
q and the initial time s.
This yields the flow functions <Ps of the vector field v with the property
a
at l(q,t)<P s = v( <Ps (q, t), t), <P s (... , s) = id.n (= identity on D). (3.4)
Thus, <Ps is a parameterized set of solutions to initial value problems. In the
following, we use the abbreviation

<P := <Po . (3.5)


It follows that

<P(q, t+s) = <Po(q, t+s) = <Pt(<P(q, t), s) = <Ps(<P(q, s), t) 'Vs, tEn 'Vq ED.
(3.6)
In order to avoid extensive indexing, the flow function <P is investigated in
the following instead of the whole parameterized set of all flow functions.
Exceptions will only be made if they seem necessary. Figure 3.1 illustrates
the working of <P.
Let to E n
be an arbitrary fixed time. It has a vector field
v( ... , to) :D ----+ n3
V1(q,t O) ) (3.7)
q f---t ( V2 ( q, to)
V3(q , to)

Time 0

- <1>(00. ,1)

r1
- V
<1>(00 .,1)

<I>(q,l) Time 1

Fig. 3.1. Working of <P ( ... , t)


172 3. Homogenizing and Conditioning

The mapping
d : n -----+ Q
(3.8)
s f------t d( s)
is called the streamline of the system at the time to through the point q E Q
if

d(so) = q (3.9)

Streamlines are curves that are tangential to all velocity vectors whose base
point they pass at a fixed instant.
Let I = [a, b] ~ n be a (time) interval and q E Q a point of passage. Let
to E n be an arbitrary instant. Then S = tJ>( q, to + 1) is called a streakline at
time to. In a stationary flow field, particle paths and streaklines are identical.
An algorithm for the computation of several thousand particle paths
through a stirrer has to meet great numerical challenges. We will only point
out three of them. First, the path integration must be reliable, because the
velocity gradients in stirrers are very large and, therefore, the paths are very
sensitive to small perturbations. Additionally, there is the danger of a par-
ticle colliding with the wall if numerical imprecision occurs near the wall.
We use a high-order Runge-Kutta method for the path integration, which
leads, of course, to longer CPU times in comparison with simpler, but less
precise methods. The second problem concerns the organization of the data,
namely the search algorithm that determines the element where the parti-
cle is located at each time interval. From the velocities at the nodes of this
element, the actual particle velocity is derived by interpolation. Typically,
this procedure must be carried out several million times (thousands of time
intervals On thousands of particle paths). Because we do not impose any
restrictions On the geometric meshing structure of our stirrers (e.g., axial
parallel and equidistant meshes), this search is very complex, in principle.
We use a pigeonhole-type algorithm where, at the beginning of the computa-
tion, appropriate sets of neighboring elements are joined in subsets, so that
each subset can be identified by a simple check of the coordinates. When
we search for the actual element of interest during the process of following
the particle paths, we first make a rough estimate in which of these subsets
the particle might be, and then only the elements of these few subsets are
searched thoroughly. An efficient formulation of the algorithm is necessary
for enabling the program to be run On a workstation. The third problem has
to do with the computation of stretching (to be discussed in the following
sections) along the particle paths, which is also very complex. Here, in a pre-
processing run we once determine the flow function for one rotation of the
stirrer for all grid points. The determination of the stretching at each point
of a particle path is then reduced to an appropriate interpolation.
Despite all these efforts, we have a typical computation time of several
days for real 3D situations On modern workstations.
3.1 The Intensity of Mixing Processes 173

3.1.3 Statistics of Residence Time and Dispersion

Procedure

As mentioned in the introduction, we want to observe many particle paths in


the system and then set up statistics of the deformations occurring along the
paths in order to assess a stirrer. This section deals with two standards of
assessment, which are conceptionally simple and can be determined without
much computing power. We assume small spots in the inlet cross section,
each consisting of many particles. We follow the particles of these spots on
their way through the stirrer. This is correlated with the tracer experiments
common in laboratory experiments with plexiglass models where a colorant
is injected in concentrated form at a specific point, and its behavior is then
observed.
The simplest way of assessing consists in observing the residence times
of the particles. Long residence times increase homogenization because they
support diffusion and enhance the probability of a strong interfacial growth.
Of course, this approach gives only a very rough impression of the mixing
behavior. We are looking for better methods.
A more thorough analysis starts from the assumption that an ideal stirrer
would disperse the particles of each small entry spot equally over the outlet.
We can thus evaluate a stirrer according to its dispersion potential. We as-
sume that the stirrer outlet is contained in a circle of diameter D and proceed
in the following way:

1. We start with an equally and randomly distributed set of points at the


outlet cross section. Let n be the number of points and (Xi, Yi)iE{1, ... ,n}
their positions.
2. We compute the distances between these points. The number of compar-
isons is [n(n - 1)]/2, and

di •j := II(Xi,Yi) - (xj,Yj)11 wherei,j E {l, ... ,n} andj < i. (3.10)


3. When the distances are determined, a probability distribution is com-
puted:

Fref(8) := #{ di,j < 81 i, j E {I, ... , n} j <i }


n n(n-l)
(3.11)
--2-

4. For large n, F~ef converges to a differentiable function Fref. The function


Fref is differentiated, which yields the probability density

ref = :s F ref . (3.12)

The probability density is always positive, and zero at 0 and D. Between


these zeros, there may be other zeros (or domains of zeros).
174 3. Homogenizing and Conditioning

After the computation of F ref and 1'ef, a corresponding analysis is per-


formed for the distribution of the penetration points of real particle paths.
Assuming the starting position p of a spot, functions Fspot and j"pot are
calculated analogously. The mixing quality is determined by the deviation of
the functions 1'ef and j"pot:

(3.13)

This is the so-called dispersion, which, of course, depends on the reference


point p. Due to the factor 0.5, (J has values between 0 and 1.

Typical Plots

For the purpose of illustration we present the dispersion analysis of a pro-


duction stirrer without explaining its geometry in detail. The stirrer has a
circular inlet and a circular outlet. Figure 3.2 shows the residence-time spec-
trum.
Let us now turn to the dispersion analysis. First of all, a uniform prob-
ability distribution over the outlet cross section was generated with 2000
particles (see Fig. 3.3). This distribution will serve as a standard (the "ideal
stirrer") for comparison.
Figure 3.4 shows the spots in the inlet cross section that have been fol-
lowed through the stirrer (each spot consisting of a few hundred particles).
Figure 3.5 shows the statistic assessment of the spot diagrams and the COIIl-
parison with the ideal stirrer.
According to (3.13), the following dispersion indices are obtained:

0.06 ,-----,----,-----,---,----,---..,--,---.,----,,-------,

0.05

~ 0.04
l!..
~
~ 0.03
::J
0-

~ 0.02

0.01

OL-~~~WW~WW~wwwwwwww~ww~~~~

o 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Time /s

Fig. 3.2. Typical residence-time spectrum


3.1 The Intensity of Mixing Processes 175

a)

b)
0.8
~ 0.6
:c<tl
..c
e
a.
0.4
0.2

0
0 r 2r
Distance of two particles

Fig. 3.3. Random distribution with 2000 particles (a) and resulting probability
function (b)

Reference position Dispersion 0-

eO (center) 0.3214007
e1 (3 o'clock, outside) 0.3969149
e4 0.3927481
e5 0.3030272
e6 0.6149949
e7 (6 o'clock, outside) 0.3728627
e8 0.3060501
d1 (3 o'clock, half of radius) 0.4128288
d4 0.41181
d5 (9 o'clock, half of radius) 0.3020983
d6 0.2837924
d7 (6 o'clock, half of radius) 0.2899852
d8 0.3095072

The dispersion method is a relatively suitable tool for comparing stirrers.


However, information about the fluid deformations occurring in the stirrer
is given only in an encrypted form, which is by no means complete. This
method can be varied, for example, by starting particles not only at the
time 0, but also at later times, and by then studying the dispersion of all
these particles from different time levels. However, our practical experience
with this method has not been really satisfying in the end. Several phenomena
cannot be understood on the basis of this method. Therefore, we will now
turn to another method that deals with the problem of assessing deformations
much more directly (and, unfortunately, is also much more complex).
176 3. Homogenizing and Conditioning

0.50 b)

0.45

• ••
•• •
c: Contour
.2 0.40
U of entrance
~ 'Start_d1 '
'6 'Stared2'
N 0.35
'Start_d3'
'Start d4'
'Start_d5'
0.3 'Start d6
'Stared?'
'Start d8'
-0.1 -0.05 0 00
.5 0.1
Y direction


0.50 a)

• •
~

0.45
Contour

• • •
c: of entrance
.2 0.40
U 'Start_eO'
~ 'Start e1 '
'6 'Start_e2'

•••
N 0.35
'Start e3'
'StarC e4'
Start e5'
0.30
'Start_e6' Fig. 3.4. Starting posi-
'Start_e?' tions in the inlet (a) on
'Start_e8' the halved radius of the
-0.10 -0.05 0 0.05 0.10 inlet cross section, (b) in
y direction the center and outside

3.1.4 Deformation of Infinitesimal Test Bodies


Along Particle Paths

We continue investigating how well a stirrer stretches the interface between


inhomogeneities (striae) and the basic glass. In the following, the approach for
a quantitative assessment consists in computing the deformation of infinites-
imal test bodies along many particle paths and then performing a statistical
evaluation. The initial geometry of the test bodies can be chosen arbitrarily,
and we will give a relatively detailed description of three different approaches
(cube, cylinder, sphere) because this sheds light on the problem from different
points of view and provides interesting insights. The discussion will lean on
the method described in Ottina [3.1], but will in much more detail extend on
obvious questions that have arisen in the context of the dispersion of striae.
As we are considering incompressible fluids only, excluding any kind of
diffusion and reaction, the volume of the test body is constant in time. The
3.1 The Intensity of Mixing Processes 177

14.---~----~----~----~----~
a)
... Randomly distributed set 01 points
~ 12 'g1 _zieLdat'
'g2 zleLdat'
~ 10 'g7 _zieLdat'
Ll
~ B 'dB zlel.dat'

'0 6
~
'i:!! 4
Q)
Cl

O.4r O.Br 1.2r 1.6r 2r


Distance of two penetration points

14 .----~=---~--.,..-.",....--:--~:---:--...., b)
... Randomly distributed set 01 points
_Z ' 12 '10 ziel.dat'
'f1_zieLdat'
~ 10
Ll
'12 zieLdat'
eO. B '16_zieLdat'
'17 ziel.dat'
'0 6 'IB_zieLdat'
~
'u;
c:
Q)
Cl
Fig. 3.5. Dispersion statistics,
start position (a) half of ra-
O.4r O.Br 1.2r 1.6r 2r dius, (b) center and 80% of ra-
Distance of two penetration points dius

mapping ifJ is volume preserving due to the continuity equation. Of course,


the geometry can be deformed arbitrarily, and this is exactly what leads to
considerable numerical problems.
All the following considerations are valid on an infinitesimal scale. Even
if huge stretchings occur, everything is regarded as happening locally. We
do not claim that the resulting parameters give a correct description, for
example, of the deformation of a sphere with real dimensions in a flow field,
because they presuppose identical conditions at every point of the test body
(which is not the case for a real body of finite dimensions in a non-trivial
flow field).

Deformation of Cubes Under the Influence of "\lcP

Two particles occupying the positions q and q + WI at time 0 are transported


to the positions ifJ( q, t) and ifJ( q + WI, t) by the flow ifJ in time t. During this
transport the distance between the two points changes. The ratio of the new
distance and the distance between the original points is the length stretch A,
which is
178 3. Homogenizing and Conditioning

'( ) ._ 1. 114J(q + 8WI, t) - 4J(q, t)11


/\ q, t, WI .- 8~ 11 8w III
Ill ·Im8-+0 <P(q+8wl,t)-<P(q,t) II
8

IlwI11

IIVI(II~I~WIII = IIV,(q,t)4J CI:~II) II (3.14)

in an infinitesimal formulation. In order to relate this definition of stretch


to the usual concepts of "true stretch" and "technical stretch", which are
common in classic mechanics of solids, we consider a rod of initial length l(O)
between the points q and q + w, which is stretched (compressed) to the final
length l (to) = l (0) + f1l. As is well known, the real stretch Ereal is obtained
from the technical stretch Etech as follows:

Ereal:= l /(t O )

/(0)
dl f1l
-l = In(l(to))-ln(l(O)) = In(1+-(-)) =: In(1+Etech). (3.15)
l to

With the notation l(t) := IIVI(q,t)4Jwll, that is, in particular l(O) = IIVI(q,o)4Jwll
= Ilwll, we get the following relation:

= 1 (l(t )) -1 (l(O)) = 1 ( IIVI(q,t)wll ) = 1 (1IVI(q,t)4Jwll)


Ereal non n IIVI(q,o)4Jwll n Ilwll (3.16)
=In(.\(q,t,w)) .

If two vectors q + W2 and q + W3 are considered, the parallelogram spanned


by them in q has the area IIw2 x w311. In analogy to the length stretch .\, the
area stretch 7] is defined:
11(4J(q + 8W2, t) - 4J(q, t)) x (4J(q + JW3, t) - 4J(q, t))11
7](q, t, W2, W3) :=liIl!
8-+0,8-+0 118w2 x JW311
·
Ill Im8-+0 <P(q+8w2.t)-<P(q,t)) x 1· _
8 Im 8 -+ 0
<P(q+ 8W 3,t)-<P(q,t) II
8
IIw2 x w311
IIVI(q,t)4Jw2 x VI(q,t)4JW311
(3.17)
IIw2 x w311
Because invertible matrices A and vectors WI, w2 satisfy
AW2 x AW3 = det(A)(A-Ifw2 x W3 , (3.18)
7] can also be defined differently:

Finally, the volume stretch J of a parallelepiped spanned by the vectors


WI, W2 and W3 can be defined in the same sense:
3.1 The Intensity of Mixing Processes 179

( ) ._ det(V'(q,t)<PWl,
J q,t,Wl,W2,W3 .- d (
V'(q,t)<PW2' V'(q,t)<PW3)
)
et WI, W2, W3
= det(V'(q,t)<P) . (3.20)
In the case of an incompressible flow, it can be expected that J is time
independent. We will shortly prove the following relation:

(3.21 )

For any direction w, we have


a p':p.
at V'(q,t)<PW

lim :
Ll.t-+O u t
(lim ~(<P(q + s w, t + ~t) -
s-+O S
<p(q + w, t + ~t))
- lim
s-+O S
~(<P(q + s w, t) - <p(q + w, t)))

lim
s-+O S
~(lim : (<p(q+s
Ll.t-+O u t
w,t+~t) -<p(q+s w,t))
- lim :
Ll.t-+O u t
(<P(q+w,t+~t)-<p(q+w,t)))
(3.4). 1
= hm
s-+O S
-(v(<p(q + s w, t), t) - v(<p(q, t), t))

= V'(<P(p,t),t)V V'(p,t)<p W . (3.22)


Therefore,

(3.23)

applies generally. This equation is not only influenced by the permutation


rule for the differentiation of continuously differentiable functions, it also
essentially depends on the main equation (3.4). The derivative of J with
respect to time is now "easy" to calculate:
a
a/(q, t, WI, W2, W3) = D(q,t) (det oV'<p)
= Dv<p det(V'(<p(p,t),t)V V'(p,t)<p)
= det(V'(<p(p,t),t)V V'(p,t)<p el, V'<P e2, V'<P e3) (3.24)
+det(V'<P el, V'(<P(p,t),t)V V'(p,t)<p e2, V'<p e3)
+det(V'<p el, V'<p e2, V'(<P(p,t),t)V V'(p,t)<p e3) .
If det(V'(p,t)<p) = 0, the right-hand side of (3.24) vanishes and the statement
is proven. However, V' (p,t)<P is a bijective mapping, and therefore the matrix
A := V'(p,t)<p is regular. With the standard normal basis ei, we have
180 3. Homogenizing and Conditioning

det(V(p(p,t),t)V A el,A e2,A e3)


+det(A el, V(p(p,t),t)V A e2,A e3)
+det(A el,A e2, V(p(p,t),t)V A e3)
Ainvertible
= det A
(A- l V(p(p,t),t)V A el, A e2, A e3 )

+ det(A el, AA-lV(p(p,t),t)V A e2, A e3)


+det(A el,A e2,AA- l V(p(p,t),t)V A e3)
det(A)(det(A-lV(p(p,t),t)V A el, e2, Ae3)
+ det(el' A-lV(p(p,t),t)V A e2, e3)
+ det(el' e2, A-lV(p(p,t),t)V A e3))
det(A)trace(A -1 V (p(p,t),t)v A)
det(A)trace(V (p(p,t),t)v)
det(V(p,t)p) div(v(p(p, t), t) . (3.25)
In the penultimate step, we made use of the fact that the traces of similar
matrices are equal. Because of the continuity equation, the statement is thus
completely proven.
It follows that the measures of deformation defined in this way not only
depend on time and space coordinates (like p), but also on the given orien-
tations of the test bodies. More precisely,
>":JlxRxS 2 -----+ R
(3.26)
(q, t, w) f----t >..(q, t, w) = IIVp(w)11
and
Ti : [l X R X S2 -----+ R
(3.27)
(q, t, w) f----t Ti(q, t, w) = II det(VP)((Vp)-l )T(w)11 .
Instead of this additional dependence on directions in the individual tangen-
tial spaces, it would be desirable to generate space directions in a canonical
way and to replace WI, W2, and W3 with mappings that depend on time and
space coordinates only.
This is possible due to a relation between >.., Ti, and p( ... , t). If t = i Oi,
we have
p( ... , t) = p( ... , i !:l.t) = P(i-l) .6.t(..·, t)OP(i-2) .6.t(. .. , (i - 1) !:l.t)o ...
o P.6.t(. .. , 2 !:l.t) 0 po( ... , !:l.t) , (3.28)
and thus
VI(. .. ,tJ P = VI(. .. ,i .6.t)P(i-l) .6.t 0 VI (i-I) .6.t P (i-2) .6.t 0···

OVI<. .. ,2 .6.t)P.6.t Oi) 0 VI(. .. ,.6.t)PO

=: II VI<. .. ,j .6.t)P(j-l) .6.t . (3.29)


j=l
3.1 The Intensity of Mixing Processes 181

Deformations can thus be computed by iteration.


Because of (3.18), Vlc ... ,t)cJ> must be computed for each individual particle
path in the algorithm by means of (3.29). At the end, the results can be
stored. With Vlc ... ,t)cJ>, the information about length and area stretch is then
available and can be processed further. According to the definitions of (3.26)
and (3.27), the values of the unit sphere are used in the computation:
W
A(q, t, w) = IIVlc ... ,t)cJ> Ilwllll (3.30)

'T/( q, t, w) = II det(VI C... ,t)cJ» (VI c... ,t)cJ» -1 f wll . (3.31 )


Because the unit sphere is compact, the two mappings have a minimum and
a maximum. The range of stretch is given by

Amin = min{>..(q, t, w)lllwll = I} and (3.32)

Amax = max{>..(q,t,w)lllwll = I}. (3.33)


Moreover, from this range the corresponding directions can also be derived.
The problem of space directions is thus eliminated. A canonically provides
two values and two directions that are characteristic for each particle path.
To each particle starting point q and the time t = i~t at which the particle
leaves the system, correspond the two non-directional numbers mentioned
above. The minimum is the worst case for a stria, whereas the maximum is
the best case.
A unit cube (spanned by the vectors WI, W2, and W3 in the tangential
space) attached to a particle path with a vertex has transformed into an
oblique parallelepiped at the end of the path. The edge lengths as well as the
lateral surfaces may change. According to the continuity equation, only the
volume remains constant. The stretch of the edge lengths has already been
computed. The stretch of the lateral surfaces results from
,(WI, W2, W3) := IIVlc ... ,t)cJ>Wl x VIC ... ,t)cJ>W211
+IIVlc ... ,t)cJ>Wl x VIC ... ,t)cJ>W311
+IIVlc ... ,t)cJ>W2 x V!c. .. ,t)cJ>w311
II det(Vlc ... ,t)cJ»(Vlc ... ,t)cJ>- l f WI x w211
~

+11 det(Vlc ... ,t)cJ»Vlc ... ,t)cJ>- l f W2 x w311 . (3.34)


~
WI

With the abbreviation

(3.35)
182 3. Homogenizing and Conditioning

the problem
= (1IBwIiI + IIBw211 + IIBw311) -+ min
I(Wl, W2, W3)
< Wl,Wl > = 1
< W2,W2 > = 1 (3.36)
< Wl,W2 > = 0
W3 = WI X W2
must be solved. This is an optimization problem with constraints. It is pos-
sible to describe the problem in a parameterized form (without constraints).
However, it is useful to rewrite the function IIBwIiI + IIBw211 + IIBw311 first:
IIBwIiI + IIBw211 + II Bw311 =
J< BWl,Bwl > + J'<-B=-W-2-,-=B-w-2->-+ J< BW3,Bw3 > =
vwnBTB)Wl + Vwr(BTB)W2 + vwf(BTB)W3 . (3.37)
The matrix BT B is symmetric and thus diagonalizable. All eigenvalues are
positive (because the standard scalar product is positive definite). Therefore,
we can assume without loss of generality

BT B = (6o ~2 ~
1

0 13
) (3.38)

Now we determine that


COS(() cos(a))
WI = ( cos(() sin(a) ,
sin(()

W2 = cos(iJ) (~~~(i))) + sin(iJ)


o
(= ::~~g~~~[~J) ,
cos(()
(3.39)

W3 = WI X W2 = sin(;3) (~~~(~))
o
(= ::~~g ~~~~~J)
+ cos(;3)
cos(()
With the free parameters a, iJ, and ( we obtain for T
I(WI, W2, W3)

= V,I cos 2(() cos 2(a) + 12 cos 2(() sin 2(a) + 13 sin 2(()

+vcOS 2(iJ)('I sin 2(a) + 12 cos 2(a)) + sin 2(iJ)(rI sin 2(() cos 2(a) ...

... + 12 sin 2(() cos 2(a) + 13 cos 2(())

+ vcoS 2(iJ)(r1 sin 2(() cos 2(a) + 12 sin 2(() sin 2(a) + 13 cos 2(()) ...

... + sin 2(iJ)('I sin 2(a) + 12 cos 2(a)) . (3.40)


3.1 The Intensity of Mixing Processes 183

We abbreviate

(3.41)

which yields
')'1 sin 2(a) + ')'2 cos 2(a) = ')'1 + ')'2 - ~ . (3.42)
Therefore,

')'(Wl' W2, W3) = V~ cos 2(() + ')'3 sin2(()


+vcos 2 (j3)bl + ')'2 -~) + sin2(j3)(sin2(()~ + ')'3 cos 2(())

+Vcos2(j3)(~sin2(() + ')'3 cos 2(()) + sin 2(j3)bl + ')'2 - ~)


?: V ~ cos 2(() + ')'3 sin 2(() + J ')'1 + ')'2 - ~ + V sin 2(()~ + ')'3 cos 2(()
?: ~ + v03 + J ')'1 + ')'2 - ~
(3.43)
This minimum is indeed assumed for the eigenvectors of BT B.
Now we are able to define a measure for the surface stretch of the cube:
1
')'cube := 3(yfh + Vf2 + v03) . (3.44)

The advantages and disadvantages of directly computing the deformation


matrix are:
• No information about the deformation along a particle path is lost. Effects
of deformation can be evaluated later with respect to arbitrary directions.
In particular, the resulting linear mapping can be examined.
• It is thus possible to compute the above-mentioned,), measure for surface
deformations.
• If the sections of the flow function are computed one after the other (this
corresponds to multiplying the matrices), numerical errors increase. There-
fore, at least at this point we should search for methods that account for the
accuracy of the computation. Apparently, the problem does not arise from
the A measure, but from the 'I] measure. The matrix obviously degenerates
during the iteration and becomes "singular".

Deformation of Cylinders Under the Influence of ViP

From a geometric point of view, string-shaped striae are, of course, to be


compared rather with cylinders than with cubes. Which effect does a linear
mapping have on a cylinder? The answer is as simple as the details involved
are tedious: linear mappings turn ellipses into ellipses.
184 3. Homogenizing and Conditioning

Remark 1: For every endomorphism of a finite-dimensional vector space


with a scalar product there is an orthonormal basis that is mapped on mu-
tually orthogonal vectors.

cp : (V, < ., . » --+ (V,< .,. ». (3.45)

The adjoint cpa of cp is characterized by

VV, wE V < cp(V), w >=< v, cpa(w) > (3.46)

cpa 0 cp is self-adjoint. It follows that for this composite mapping there is an


orthonormal basis {viii E {I, ... dim(V)}} of eigenvectors associated with
the real non-negative eigenvalues (1, ... (dim(V). This basis satisfies

Vi,j E {I, ... dim(V)}

The remark is thus proven. Obviously,

Vi E {I, ... dim(V)} (3.48)

Remark 2: Remark 1 is also true if cp is a mapping between two vector


spaces (V, < .,. >v) and (W, < .,. >w) of equal dimensions. As V and W
have orthonormal bases, an angle-preserving mapping

1/J: (W, <.,. >w) --+ (V, <.,. >v) (3.49)

can be defined via these bases. Then, 1/J 0 cp fulfils the conditions of remark 1,
and, with the same notation, we have

Vi, j E {I, ... dim(V)} (jOi,j =< 1/J 0 cp(Vi)' 1/J 0 cp( Vj) >v
(3.50)
=< cp(Vi), cp(Vj) >w .
Remark 3: If cp : (V, < ., . >v) --+ (W, < ., . >w) is a bijective mapping,
then

cp(v) = cp (
dim(V)
~ < v,v, > v,
1
Fourier st~etch of v ~n the ON~ Vl"",Vn

(3.51 )

Therefore,
3.1 The Intensity of Mixing Processes 185

(3.52)

Unit circles (with respect to < .,. >v) are thus mapped onto ellipses (with
respect to < ., . >w). Every ellipse is a unit circle (if an appropriate parallel
shift is applied and then an appropriate scalar product is chosen).
Cylinders Z consist of a space direction r (axis) and a circle K, which is
drawn into the space along this axis, that is, Z = {s r + qls E [0, l],p E K}.
K lies in the plane U corresponding to r. After the application of a linear
mapping ¢ (restriction of the linear mapping: ¢Iu : U ----+ ¢(U)) into another
plane ¢(U), the circle will at least remain an ellipsis ¢(K) according to the
considerations above. The original axis of the cylinder is also rotated (in
general, it is no longer orthogonal to the plane of the ellipse). The resulting
structure is ¢(Z) = {s ¢(r) +pls E [0, l],p E ¢(K)}, that is, an ellipse drawn
obliquely into space.
The surface of the "cylinder" in Fig. 3.6 is

0: = 2 F(\74>(rcos(a)vl
\.
+ rsin(a)v2)) #
V

area of the base under 'V <P


+ U(\74>(cos(a)vl + sin(a)v2))
, v .f
\74>(V3) (3.53)
perimeter of the base under 'V <P

= 2 F( ... ) + U( ... ).,(p, t, V3) .


In this formula, the deformation of the cylinder base and the stretch of the

°
cylinder axis are nicely discernible. The stretch of the surface relative to the
original cylinder at t = is
o
'Y'=
. -41f' (3.54)

There is an initial orientation of the longitudinal axis of the initial cylinder


for which a minimal surface stretch under the influence of \74> is assumed.

Fig. 3.6. Effect of a linear mapping on a


cylinder
186 3. Homogenizing and Conditioning

Deformation of Spheres Under the Influence of ViP

The last case we want to consider is that of spheres. If at the time 0 a sphere
(instead of a cylinder or a cube) is started at a point in the flow area, it is not
necessary to account for different initial orientations in the consideration of
surface stretches, because the description of a sphere in space does not require
specific directions. After the application of the linear mapping V' q,tiP, only
the square roots of the (positive) eigenvalues of the mapping V' q,tiPa 0 V' q,tiP
are relevant. Let them be denoted by 11, 12, and 13 (we assume, without loss
of generality, 11 < 12 < 13). According to the derivation above, 11, 12, and
13 are the lengths of the axes of the resulting ellipsoid, whose surface can be
computed as follows. With

k '= J
13 I~ - Ii a := arcsin (
V12 _12)
~3 1 (3.55)
. 12 J,~ -Ii'

and the elliptic integrals

(3.56)

the surface 0 is given by

(3.57)

Thus, the surface stretch relative to the initial sphere at t = 0 can be defined
in the same way as in the cylinder case:
o
,:=- . (3.58)
47r

Cube, Cylinder, or Sphere?

The cube, the sphere, and the cylinder are three candidates for the assessment
of surface stretch. Which one is the most appropriate for what?
If we consider length stretch, where we are dealing with a space direction
and its deformation, the cube is appropriate because it is spanned by three
vectors. In addition to the fact that a cube is a slightly strange geometric
model for a small perturbation in the glass, its start position in space has an
influence on the deformation parameter I (as has been described above, the
dependence is three-dimensional). If the cube orientation with the smallest
surface stretch is to be found for each particle path, a minimization problem
must be solved afterwards.
3.1 The Intensity of Mixing Processes 187

At first glance, the cylinder seems to be a more suitable model for a stria.
However, as with the cube, there is a preferred orientation (in this case, a
two-dimensional problem). The). measure is a very comprehensible measure
of the longitudinal stretch of the cylinder.
The sphere has no preferred orientations; these are defined only by the
flow itself. The resulting ellipsoid contains the entire information about the
effects of stretching and compressing. The model of a tiny sphere that is
deformed in a flow with its surface expanding seems natural.
We have decided to use the ). measure as a comprehensible means to
describe the longitudinal stretch of cylinders and the r measure as a compre-
hensible means to describe the surface stretch of spheres.

3.1.5 Deformation Statistics

After the previous section, which is very technical, we can now define new
appropriate assessment measures that represent the best current state-of-the-
art with respect to the evaluation of mechanically induced surface stretch.

Procedure

Let a sufficiently large number of particles 8 = {Pi liE {I, ... , n} }, which are
randomly distributed over the whole inlet domain, be started (it is assumed
that the particle Pi reaches the exit or control point at time t i ) and then
pass a control point (8 only contains start positions that actually reach the
control point; the maximum particle running time t max must be chosen so
large that a sufficient portion (70-80%) of the particles started actually passes
the control point). Then, ). and r can be assigned to each particle. For)' the
initial velocity usually defines the direction. In order to determine the spectra,
the probability distributions

() #{).(pi,ti,V(p,O)) < xli E {l, ... ,n}}


~x= , (3.59)
n
() #{r(Pi, t i ) < xli E {I, ... ,n}}
P,,! X = n (3.60)

are considered. Both curves are monotonically increasing and run between 0
and 1. For a comparative assessment, the decisive statement is: The curve
belonging to the better stirrer system lies below that of the worse system.
When the stretch measures are computed, we often reach numerical limits,
namely if a test body is stretched extremely so that volume preservation
can no longer be guaranteed. This problem can be overcome with a simple
but effective trick: it is not the course of the two curves at extreme stretch
values that we are interested in, because at a certain order of magnitude the
stretching is so good that it may be considered as ideal (special thanks to
David Gelder, who gave us this tip in a seminar in Cambridge). Below this
bound, differences between two curves correspond to relevant differences in
188 3. Homogenizing and Conditioning

quality between the respective stirrers. Above this bound, spheres (or cubes or
cylinders) are deformed so strongly that diffusion mechanisms can act ideally.
On condition that the ratio of the set S and the set of started particles is
sufficiently large, those particles having not yet passed the control point after
the time t max are of no interest, because they tend to be subject to a stronger
deformation and, at the same time, to a longer period of diffusion processes.

Typical Plots

Again, for the purpose of illustration we present results for a real production
stirrer without going into detail with respect to its geometry. Figure 3.7 shows
the distribution of the length stretch ,\ (above) and the surface stretch (below)
along the particle paths of a large number of particles randomly distributed
at the inlet.
We can see that the stretches are rather large (,\ and I on the abscissa
are given in a decadic-logarithmic scale). These curves contain the complete
statistical information about stretches. The boldface vertical lines drawn as
examples through the abscissa at a value of 10 each show that approximately
70% of all (in this case 3735) particles for which the computation has been

1.0 ,------,------,---r---",:o===,.----,
a)

0.8
,,
~
.0
0.6 ,
. _ .. _ .. -.- ---.. -- --- ----··f·
,
/
,,
al
.0
£. OA ! - ~t-
,
, ,,
0.2 - - -,·r·'t-··_---
,,
, /"
OL..---.::...L.-_--'-_-----'-_ _L - - _ . . . L . . - _ - - '
o 5 10 20 25 30

0.50 ,------,------,---r----r-----,------,
b)
OA5
OAO
>- 0.35
§ 0.30
.0
~ 0.25
£. 0.20
0.15
0.10
0.05
o ~-~~~~-~-~--~-~
Fig. 3.7. Typical stretch
o 5 10 15 20 25 30 spectrum: (a) length stretch,
I091QY (b) surface stretch
3.1 The Intensity of Mixing Processes 189

performed have undergone a length stretch .A ::; 10 10 , and that approximately


3% have undergone a surface stretch 'Y ::; 1010. The curves also show, for
example, that approximately 22% of all these particles have undergone a
length stretch 1010 ::; .A ::; 10 15 and that approximately 14% of them have
undergone a surface stretch 1010 ::; 'Y ::; 10 15 .

3.1.6 Example: a Simple Paddle Stirrer

In the following, we demonstrate the application of the measures for the as-
sessment of stretching in the process analysis of a (strongly idealized) stirrer.
The example shows the quantitative influence of typical process parameters
on the mixing result, whose basic effect can be immediately understood in-
tuitively. The stirrer consists of two paddles that are attached to an axis and
rotate in a cylinder where there is an axial flow (see Fig. 3.8). The diameter
of the cylinder is 100 mm, the throughput is 1.5 kg/min, and the viscosity is
250 Pa s. In this case, the flow can be computed without approximation ac-
cording to the method of rotating coordinate systems. We have investigated
the influence of the wall distance and the speed of rotation on the stretching.
Figures 3.9 and 3.10 show the results of the stretching analyses. The mean-
ing of the boldface lines and the point clouds is the following: the starting
points of those particle paths in the inlet cross section are represented whose
stretching is less than , or equal to , the stretching marked by the boldface
line. This supplementary figure gives a quick survey of those domains in the
inlet cross section that are mixed the least (the boldface line can, of course,
be set to an arbitrary value). In practice, this is a very important information
as it is often known a priori where certain typical striae arrive at the inlet
(e.g., striae at bottom or top).
Finally, Fig. 3.11 shows a snapshot of some streaklines. We suggest to
watch the time behavior of these streaklines as a video with the appropriate
software, in order to get a feeling for the dynamics.

Fig. 3.8. Simple paddle stirrer


190 3. Homogenizing and Conditioning

1.0 a)

0.8
~
:E 0.6
'"
£ 0.4
.0

0
0 20 40 60 80 100
10 910'"
1.0 b)
0 .8
~
:E 0.6

£'" 0.4
.0

0.2

0
0 20 40 60 80 100 120 140
109 10Y
1.0 c)

0.8
~
:E 0 .6

£'" 0 .4
.0

0
0 5 10 15 20 25 30 35 40 45 50
10910'"

1.0 d)

0.8
~
:E 0.6
'"
.0
004
It .

00 10 20 30 40 50 60 70 80
10910 Y

Fig. 3.9. (a, b) Case 1: 3mm wall distance, 20 revolutions per minute,
(b, c) case 2: 3 mm wall distance, 5 revolutions per minute
3.1 The Intensity of Mixing Processes 191

1.0 i -.--------:========j a)

0.8
?;
:B 0.6
co
.0
a:004.
0.2

O WL--~--~----~---L--~
o 20 40 60 80 100
10910).

1.0 i -r-----r-===:=::::!:=::==, b)

0.8
?;
:B 0.6
co
.0
a:004.
0.2

00 20 40 60 80
10910 Y

c)

5 10 15 20 25 30 35 40 45 50
10910A

d)

10 20 30 40 50 60 70 80
109l0 Y

Fig. 3.10. (a, b) Case 3: 6mm wall distance, 20 revolutions per minute,
(c, d) case 4: 6 mm wall distance, 5 revolutions per minute
192 3. Homogenizing and Conditioning

Fig. 3.11. Some streaklines (see also movie on CD)

3.1.7 Outlook

In the previous sections, we have presented ways of assessing mechanically


induced interface stretching in flows within stirrers. The basis for the mea-
surement of the mixing intensity was the deformation of (otherwise passive)
infinitesimal test bodies. As mentioned in the beginning, this is only the most
elementary mechanism of several others that are at work. One of the next
steps will be the introduction of diffusion effects, that is, following the test
bodies we will not only account for the deformation, but also for the diffusion
that happens simultaneously. As a consequence, long residence times without
strong mechanical deformation will become as important for the evaluation
of dissolving striae as short residence times with strong deformation. One
problem will be to find representative parameters for the diffusion constants,
in particular if the temperature dependence is taken into account. Another
additional step will be to account for chemical reactions. In this case, ade-
quate reaction rates must also be determined first. In both cases, not only are
the processes going on in the stirrer relevant, but also the diffusion and reac-
tion processes that take place until certain lower temperatures are reached.
In the melting and refining sections of the melting device, such processes are,
of course, particularly important.
In our above treatment, we have omitted another problem: real striae
have, in general, different viscosities and densities from the basic glass. There-
fore, real paths of striae will only approximately be flow streaklines (the stria
no longer follows the basic glass "without resistance"). The relevance of this
effect for the assessment of stirrers is not yet clear. Mathematically, a two-
phase flow must be computed in order to include this behavior correctly.
However, one of these phases is spatially concentrated, which makes classic
numerics very complex (extremely high spatial mesh resolution). Probably.
this problem can only be solved satisfactorily with new computation meth-
ods where the computation of the flow itself consists in following the particles
(Lattice-Boltzmann method).
3.2 Instabilities and Stabilization of Glass Pipe Flows 193

3.2 Instabilities and Stabilization of Glass Pipe Flows

Ulrich Lange, Horst Loch


Introduction

During the production of high-quality special glass, platinum pipes are often
used for transport and conditioning. The glass is transported through these
pipes from the tank to the hot-forming unit, and it is conditioned there to the
appropriate temperature. The main force that drives the glass flow is gravity,
the temperature profile in the glass is determined by the loss of heat through
insulation and the Joule effect through electric heating. For the design and
operation of such a pipe system, the following two questions must be answered
above all:
• How to dimension the system ("hydrostatic" height, pipe cross section, in-
sulation material and wall thicknesses, electric pipe heating, ... ) to achieve
the desired flow rates at the desired temperatures?
• How to control the system in order to guarantee safe and stable operation?
At first sight, it might seem strange that a technical scientific publication
today deals with such questions, the subject being high-viscosity flows in
circular pipes. The first thing that students of fluid mechanics learn is that
this is the most harmless of all imaginable situations. Of course, the relevant
material parameters are temperature dependent, but there is a range of com-
putational fluid dynamics software (CFD) able to simulate such situations
quickly. So, where is the problem? This section will show that the exponential
dependence of the glass viscosity on the temperature causes unexpected sta-
bilization problems. In order to solve these problems systematically, a deeper
global insight into the dynamic behavior of such systems is necessary. This
is the reason for our approach with a combination of closed approximation
equations and CFD simulations. As a useful by-product, design equations are
developed that are easy to apply.

3.2.1 Stationary Temperature and Pressure Profiles


in the Pipe

We start our considerations with a pipe of length Land (a much smaller) ra-
dius R, which is actively heated by an electrical heating circuit (see Fig. 3.12).
The first question is which initial pipe temperature and pressure gradient
arise if the volume flow 11 and the temperature of the pipe inlet TA are
given. We make the following assumptions:
• The flow field and the temperature field are axisymmetric.
• Due to the very small radius R compared to the length L, heat diffusion is
only important in the direction of the radius; in the direction of the axis
the convective heat transfer dominates.
194 3. Homogenizing and Conditioning

q(z)
, ~-------------------------.
PE
R PA z
, r---= ........ E:>
v
TA
L
Fig. 3.12. Pipc with one heating
Heating circuit circuit

• The heat flux escaping via a small segment of the pipe wall is proportional
to the wall temperature of this segment (with a heat transfer coefficient) .
• The velocity field is one-dimensional by approximation, i.e., only the influ-
ence of the axial component of the velocity is accounted for in the momen-
tum and energy balance.
The analytical equations derived on the basis of these assumptions will later
be verified by CFD.
In the following, the axial component of the velocity is called w, the
pressure p, and the temperature T. The coordinates are the direction of the
radius r and the direction of the axis z. The heat transfer coefficient at the
pipe wall is h, the effective heat conductivity of the glass is k. The balance
equations for momentum and energy are (in cylindrical coordinates):

1 8 ( 8W) 8p (3.61 )
;: 8r r TJ(T) 8r = 8z '

8T (8 2T 8 ( 8T))
pCp W 8z = k 8z2 + ;:1 8r r 8r (3.62)

The (elliptic) momentum (3.61) fulfils the symmetry condition 8w/8r(0) = 0


and the no-slip condition at the pipe wall w(R) = O. The (parabolic) energy
equation (3.62) has the following initial and boundary conditions:

T(r, 0) = TA constant temperature of the pipe inlet,


~~ (0, z) =0 symmetry condition, (3.63)
-k~~ (R, z) = hT(R, z) loss of heat via pipe wall.
For the following asymptotic considerations, the intensity of the heat transfer
in the direction of the radius is very important. The Nusselt number is a
dimensionless number for the heat transfer at the pipe wall:

hR
NU=T' (3.64)

In the situations we are interested in, Nu is a (very) small number: typical


values are h = 5 W/Km2, R = 0.030m, and k = 3W/mK, thus Nu = 0.05.
3.2 Instabilities and Stabilization of Glass Pipe Flows 195

A Coarse Model

In order to develop a coarse model, we assume the ideal case N u = O. This


means that temperature differences (i.e., also differences of viscosity) in the
direction of the radius are compensated at once. Therefore, we can expect a
parabolic velocity distribution for the glass flow. Thus, we can determine the
profile of the mixing temperature 1'(z) quantitatively from a global energy
balance (see appendix):

- 27fRhz
T(z) = TA exp( - - - . - ) . (3.65)
pCp V

For z = L, the equation yields in particular the mixing temperature TE at


the end of the pipe.
From the momentum balance for the pipe, we can determine the pressure
gradient (see appendix):

8LV
p(z) = PA - - 1
4 -L
7fR
l0
z -
7](T(z))dz. (3.66)

This equation corresponds to the Hagen-Poiseuille law for the pipe flow with
the viscosity averaged over the pipe length, which can be calculated from the
profile of the mixing temperature determined above.

The Refined Model

With the rough information about the loss of pressure and the temperature
profile given by the above equations, in a second step we can now derive the
corrected equations for the realistic case of a small (however, not vanishing)
Nusselt number. By a local energy balance we can determine, as a first ap-
proximation, the temperature distribution in the direction of the axis, as well
as in the direction of the radius (see appendix):

5 r2 r4)) (3.67)
T(r,z)=1'(z) ( l+Nu ( 12-R2+4R4

Thus, the smaller the Nusselt number, the smaller becomes the difference
between the temperature in the center T(O, z) and the temperature of the
wall T(R, z). Here, the expectation is quantified that a smaller pipe radius
R, respectively a more gentle cooling (small h), leads to a more homogeneous
temperature profile.
We can now compute the local viscosities on the basis of this detailed
knowledge about the temperature distribution. If we use this information for
the momentum balance, we get the following refined equation for the velocity
w in the direction of the axis (see appendix):
196 3. Homogenizing and Conditioning

2V
w(r, z) = 7rR2

(3.68)

and a more exact equation for the pressure gradient:

PA -PE = 8L~ ~
7rR
r
Lio
L
1J(T(z)) (1 _~Nu
8
In (1J(!E)) T(z) ) dz.
1J(TA) TA-T E
(3.69)

In order to check the validity of our assumptions with respect to the heat
diffusion, the predictions of the above equations are compared with results
from numerical simulations with a concrete glass G 1 for two geometrically
different situations. As a test example, we assume that the glass in the pipe
with a flow rate of 1 kg/min is cooled from 1300°C to a wall temperature
of ~ llOO °C (i.e., the respective heat transfer coefficient h depends on the
pipe length). The pipes have a uniform radius of 32.5 mm, the pipe length is
assumed to be 2.5 m in one case and 5.0 m for the second case. This yields
a Nusselt number of Nu = 0.057 for the short pipe and Nu = 0.029 for
the longer one. Figure 3.13 shows the predictions for the temperature pro-
files at the end of the pipe system. The numerical results corroborate our
approximate analytical model: the predicted mixing temperatures differ only
by a few degrees. A parallel displacement of the analytical predictions by
this difference shows clearly that the radial temperature profiles agree even
better. The predictions for the velocity profiles shown in Fig. 3.14 differ more
strongly (up to 7%), which might be expected because of the strong tempera-
ture dependence of the viscosity. Nevertheless, the relatively good agreement
confirms that our analytical model includes the essential mechanisms of the

1150
---
1140

~ 1130
~

e 1120
~

Ql
a.
E
~ 1110 Numerical (2.5 m) _
Analytical (2.5 m)
Parallel-displaced (2.5 m)
Numerical (5.0 m) 0
1100 Analytical (5.0 m)
Parallel-displaced (5.0 m)
1090 0 0.005 0.01 0.015 0.02 0.025 0.03
Distance from the pipe axis 1m

Fig. 3.13. Comparison of the temperature profiles (at the end of the pipe system)
3.2 Instabilities and Stabilization of Glass Pipe Flows 197

0.005

0.004
I"Ul
.§ 0.003
~
·u
0
CD .~ ..
> 0.002 Numerical (2.5 m) •
(ij
·x Analytical (2.5 m)
<t Numerisch (5.0 m) 0
0.001 Analytical (5.0 m)
Parabolic profile

o~----~----~----~----~----~----~~
o 0.005 0.01 0.015 0.02 0.025 0.03
Distance from the pipe axis 1m

Fig. 3.14. Comparison of the velocity profiles (at the end of the pipe system)

flow through the pipe and that it can even be used for solving very detailed
problems of design. For example, one can recognize very well how the velocity
profiles become increasingly steeper due to the radial temperature gradient,
compared to a simple parabolic profile. Thus, the different residence times in
the pipe resulting from different cooling processes, for example, can also be
estimated.

3.2.2 A Stability Phenomenon

Now, we want to discuss a basic and, at first sight, paradoxical result of


our models, which leads to decisive practical consequences for the control of
the pipe heating. As an example of this phenomenon, we again consider a
pipe of length 2.5 m and radius 0.0325 m. The glass flowing through this pipe
has an initial temperature of 1300 ac, the loss of heat via the pipe wall is
controlled by one single heating circuit. Figure 3.15 shows the dependence of
the flow rate on the pressure difference in the pipe, predicted according to
the momentum balance. The mean temperatures for the respective flow rates
at the end of the pipes, resulting from the energy balance, are also given at
some points of both curves. Note that both curves have a turning point for
a final temperature of approximately 1200 ac.
Below this turning point, astonishing things happen. If we follow the
curve, we can see that an increase of the pressure difference (hence, of the
main driving force) leads to a decrease of the flow rate. Moreover, a further
mathematical analysis shows that the operation points lying on this lower
branch of the curve are unstable, i.e., although they are formal solutions of the
stationary balance equations, the system reacts to the slightest fluctuations
in these points in such a sensitive way that these (stationary!) states could
never be physically realized in reality.
198 3. Homogenizing and Conditioning

5
"Weak cooling"
I"
c:: 4 (h = 5.3) -.".-
·E
Cl
~ "Strong cooling"
(]) 3 (h =7.5)--

~
u::: 2
0

OL---~--~--~----~--~--~--~--~
40000 60000 80000 100000 120 000
Pressure difference PA - PE /Pa

Fig. 3.15. Flow rates for TA = 1300°C depending on the pressure difference (one
heating circuit)

The solution to this riddle, which will be substantiated in the following


two sections, is: if there is only one single heating circuit, indeed only the
points on the upper branch of the curve correspond to real operation states.
In reality, however, there are usually several heating circuits connected in
series like cascades. In this case, the typical operation states are situated on
the stable upper branch of the (displaced!) curve belonging to these other
operation states.
The following detailed analysis of this stability phenomenon particularly
leads to important consequences with respect to the limits of single-stage
heating circuits, and to design criteria for heating cascades.

Single-Stage Heating Circuit

In reality, apart from the two branches of the curve there is also a third
solution of the balance equations, which our asymptotic model cannot predict
because it neglects the heat conduction in the direction of the axis. This
third solution corresponds to the "frozen" glass. Under this condition, the
heat transport in the direction of the axis practically takes place exclusively
by heat conduction, therefore the glass temperature is very low and the flow
rate resulting from the pressure difference is almost zero. We could define
this third curve as an (almost) horizontal straight line, according to a flow
rate of Okgjmin in Fig. 3.15. From the mathematical theory of dynamic
systems (bifurcation theory), we can derive that the middle one of these
three simultaneously existing solutions of the stationary balance equations
(i.e., the branch of the curve below the turning point in Fig. 3.15) must be
unstable.
3.2 Instabilities and Stabilization of Glass Pipe Flows 199

This abstract result can easily be explained by a dynamic interpretation


of the curves in Fig. 3.15: for example, if we consider point (1) on the lower
branch of the curve for h = 5.3 and proceed to a stronger cooling h = 7.5 at a
constant pressure difference, we cannot expect that we reach point (2), which
refers to a higher temperature and a higher flow rate. Hence, state (2) cannot
be reached and is therefore unstable, as well as state (1), which is proved by
inversion of the above consideration. If, however, we consider point (3) on
the upper branch of the curve, it is absolutely plausible that state (4) is
reached if we proceed to a stronger cooling. Points (3) and (4) are stable. It
is also interesting what happens if we start with state (5) and proceed to a
stronger cooling: the turning point of the curve for h = 7.5 is situated above
the pressure difference referring to state (5), therefore the glass will freeze
at once. Another example for sudden freezing results from a decrease of the
pressure difference in the pipe at constant cooling to such a degree that the
turning point is crossed.
A further remarkable aspect of this stability limit is that it is determined
by a similarity law. A comparison of the equation for the mixing temperature
at an arbitrary point z and the equation of the mixing temperature at the
point z = L yields

(3.70)

This means that, independent of the concrete pipe geometry and the cooling
and flow rate, the profile of the mixing temperature only depends on the
initial temperature and the desired final temperature. Hence, for the case
N u = 0, this is also valid for the mean viscosity, which can be seen particu-
larly well by a transformation of the integral in the momentum balance (3.66)
with (3.70):

PA - PE = 8L17 1 rTA
T/(T) dT . (3.71)
7f R4 In (~:) lYE T

According to the energy balance, the volume flow is eliminated from this
relation, which finally yields the similarity law:

pCp R3(pA - PE)


(3.72)
16T/(TA)h£2

The essential statement of the similarity law is that the right-hand side
depends only on TA, T E , and the VFT parameters of the respective glass.
A detailed explanation of these relations is given in the subsection "Relation
Between Stability and Viscosity Profile" (page 203 ff.). Figure 3.16 shows this
right-hand side for the concrete glass Gland different initial temperatures
200 3. Homogenizing and Conditioning

1.00
tUJ
0.99 Initial temperature:
TA = 1400 °C
I- 0 .98 T A '" 1300 · C - -
l!?
::I
0.97 TA '" 1200 °C - -
Ii!
Q)
TA = 1100 · C - -
a. 0.96
E
2 0.95
Iii
;§ 0 .94
"
Q)
.~
0 .93 Stability limit
Iii 0.92
E
<s
z 0 .91
0 .90
20 25 30 35 40 45 50 55 60
Normalized pressure difference f (T E ,T A,"G1 ")

Fig. 3.16. Stability diagram for glass Gl (one heating circuit)

compared to the temperature relation 'FE/TA. According to the similarity


law, the turning points of these curves exactly correspond to minima of the
(normalized) pressure differences and thus to the stability limit. Obviously,
the position of the turning points only weakly depends on the initial temper-
ature. In conclusion, we can say:
The flow of glass through a cooled pipe has a critical final temperature,
below which the glass freezes at once. This critical final temperature cannot
be influenced by modifications of the pipe geometry. It only depends on the
kind of glass and on the respective initial temperatures. For the glass Gland
adequate typical values of T A , for example, this minimum final temperature
lies between 91% and 94% of the initial temperature.
Figure 3.17 shows the stability diagram for a second, shorter glass G2.
The viscosity of this glass is extremely temperature dependent; therefore the
critical final temperatures are even higher than for G 1.
Of course, the potential instability of the glass flow in single-stage heat-
ing circuits has been known for a long time now in practice (at least, this
is the tiresome experience from numerous experiments). An obvious solution
are multi-stage heating cascades. This approach naturally also results from
an objective motivated by completely different aspects: we would like to in-
fluence the temperature profile of the glass along the pipe by means of a
finer resolution. Therefore, several segments with different heat transfer coef-
ficients are necessary. The following considerations are helpful for the design
of such cascades.

Multi-Stage Heating Circuit

We will now consider the practically relevant case of the cooling of the pipe
system controlled by several heating circuits. Thus, each single heating circuit
3.2 Instabilities and Stabilization of Glass Pipe Flows 201

1.00 r---~--~----r---'---~----r---~---.
<{
Initial temperature:
t. 0 .99
TA = 1400 °C
w
I-
Q)
0 .98 TA = 1300 °C -
~ 0.97 T A = 1200 °C -
<tI TA = 1100 °C -
~ 0.96
E
.2! 0.95
tii
.§ 0.94
-c
Q) 0.93
.':::'
~ 0.92
~ 0.91
0.9~0 25 30 35 40 45 50 55 60
Normalized pressure difference I (T E.T A."G2")

Fig. 3.17. Stability diagram for glass G2 (one heating circuit)

is only responsible for the control of a smaller temperature difference that


lies above the stability limit. This can be seen in Fig. 3.18: if we assume
only one single heating circuit for the entire pipe system, we get the curve
known from Fig. 3.15 with the operation point (lkgmin-l) situated on the
unstable branch. So, if we had only one single heating circuit, the unit could
not be operated. However, if we divide the pipe into two control circuits
of identical length and assume that the first control circuit is responsible
for maintaining a constant mixing temperature of 1208 °C in the center of

3.0 ,-------.---/-r-"--.,--------,-------.-----,
One heating circuit
2.5 Two heating circuits
Three heating circuits
~t: 2 .0
'E
g Here h = 5 .3 lor the complete pipe!
Q) 1.5
~
g 1.0
u::

0.5

O ~----~----~----~----~----~--~
40000 45000 50000 55000 60000 65000 70000
Pressure difference p A - PE IPa

Fig. 3.18. Flow rates for different heating cascades. The last heating circuit of each
cascade is fixed at h = 5.3, the control response time of all other heating circuits is
"infinite"
202 3. Homogenizing and Conditioning

the pipe system, the operation point is situated on the upper branch of the
momentum balance curve of the second control circuit. Hence, in principle
the unit could be operated with two heating circuits, although the operation
point is dangerously near to the stabilization limit. If a small deviation of
the operation point (which might be a momentary decrease of the pressure
difference in Fig. 3.18) is not corrected quickly enough by the control circuit,
the stability limit may be crossed. Now, if we divide the pipe into three control
circuits of identical length and assume that the first two are responsible for
maintaining a constant mixing temperature of 1178°C after two thirds of
the pipe length, the operation point is situated on the stable branch of the
momentum balance curve of the third heating circuit, at a safe distance from
the stability limit. Figure 3.19 proves that the assumption made for the two
heating circuits can also be fulfilled: it shows the momentum balance curves
for all three heating circuits. For each one of them it is assumed that the
other two heating circuits are able to maintain the constant temperature
differences for which they are responsible. The operation point is obviously
stable for all three heating circuits, although the last one is the most critical
one.
In summary, our stabilization considerations lead to the following practi-
cal consequences:
• The glass flow through a pipe segment controlled by only one heating circuit
becomes unstable if the temperature decreases along the pipe segment until
a certain critical value is reached. The consequence is either freezing or
uncontrolled flow.
• By a cascade of heating circuits, a correspondingly larger temperature dif-
ference can be reached.

1.4 First heating circuit: h = 5.3


Second heating circuit: h=5.3 -
1.2 Third heating circuit: h = 5.3

I"c:
1.0
·E
Ol
~ 0.8
Q)

"@
:;: 0.6
0

C
u:::
0.4

0.2

Pressure difference PA -PE IPa

Fig. 3.19. Sensitivity of the heating cascade; heating circuits 1~3 are fixed at
h = 5.3, the two other heating circuits react "infinitely fast"
3.2 Instabilities and Stabilization of Glass Pipe Flows 203

• The stability limit does not depend on the pipe geometry.


• A decrease of the temperature difference per heating circuit leads to a more
robust control, because stronger fluctuations can be allowed.

Relation Between Stability and Viscosity Profile

In the following, we want to illustrate that the above-discussed stability phe-


nomenon can only occur in case of a viscosity profile "similar to that of glass" .
The discussion of (3.72), see page 199, has shown that the stability limit is
given by a minimum of the function

(3.73)

Now, the question is which properties the viscosity curve ry(T) must have in
order to minimize this function? If we introduce the normalized final temper-
ature XE = 'IE/TA, j also reads in the following way:

(3.74)

If the viscosity curve is given and the initial temperature TA is constant, a


necessary constraint for a minimum is:

of (3.75)
aXE =0 .

If we explicitly differentiate j, after several transformations we have the


equivalent constraint

1 1

XE
ry('IE) - 2ry(TA y) dy = 0 .
Y
(3.76)

A trivial conclusion from this constraint is that the viscosity curve must at
least be so steep that ry('IE) > 2ry(TA), otherwise the integrand is always
negative. So, would the stability phenomenon also occur for a sufficiently
steep, linear viscosity profile? If we insert

(3.77)

into constraint (3.76) for the stability limit, we get an explicit constraint for
the critical slope a of the viscosity curve:

1 InxE
a - - - - --c;------- (3.78)
- 2TE .XEl - 1 + In XE
204 3. Homogenizing and Conditioning

However, this value for slope a is so large that ",(TA ) would be negative, which
is unphysical, of course. Hence, the stability phenomenon cannot occur for a
linear
viscosity profile. Rather, the viscosity curve must be very steep on the
one hand, and on the other hand it must become more so even for higher
temperatures, so that constraint (3.76) can be fulfilled. The VFT curves
",(T) = lO-A+B/(T-To) , (3.79)

which can describe the viscosity of glasses, show this quality. Which are the
concrete constraints for the VFT parameters A, B , and To following from
constraint (3.76)? After several transformations, the substitution of (3.79)
yields the constraint

(3.80)

Here, (3 = B/TA and Xo = To/TA are the normalized VFT parameters (in
order to avoid unphysical cases, in addition to constraint (3.80), XE > Xo
must also be fulfilled). Thus, the VFT parameter A is not important for the
stability phenomenon. Figures 3.20 and 3.21 show the numerically computed
zeroes of FVFT in the ((3, xo) plane for different values of XE. From a purely
mathematical point of view, for each parameter combination ((3, xo) there is
obviously a root. For an increasingly smaller parameter (3, the stability limit
XE comes closer to Xo. For the practically relevant cases, however, (3 > 1: for
a shorter glass, i.e., for a larger parameter (3, the stability limit XE is closer
to 1. The critical final temperature 'IE for a stable pipe flow is increasingly
closer to the initial temperature TA for short glasses. Figure (3.20) shows

1.0 r---'-~~----,,--'---~----r---'----,

FyFT (/J.xo.0.50) = 0
~ FVFT(tJ. xo.0 .60) '" 0
-
-
II 0.8 FvFT(/J.Xo.0.70) =0 -
~

v
Qi
FyFT (/J.xo.0.80) == 0 -
E 0 .6
~
<II
a.
ti:: 0.4
>
"0
.~
til 0.2
E
o
z
0.1 0.2 0 .3 0.4 0.5 0 .6 0 .7 0 .8
Normalized VFT parameter Xo == To f T A

Fig. 3.20. Stability constraint depending on the VFT parameters


3.2 Instabilities and Stabilization of Glass Pipe Flows 205

14 r-~--~---r--~--.---.---r-~--~---,

~ 12 FvFT (fl.Xo.0.80) 0 - =
a:l FVFTVI. xo. 0.85) = 0 -
II
FVFT(8,Xo,0.90) = 0 -
<:Q..
Qj 10
~ ............
··· ............ =
FVFT(li. xo. 0.95) 0 -
Qi . · .. .. ..
E · .., ... ... ... ... .. . ~

·····
•••• : •• • • 0 •• : ••••••• : ._ ••••• : •••• • •• : • • •• • •• : ••• 0 • • • • •• • •

~
. . . . . . .. ...
.. ·........ . .
ttl ,
0.

Ii: ····
~
. .. . ...
;.. .... ...:,. ....... ;.. ... -... ;.. .... .

> .. ..
"0
Q)
.!:!
-.~.- ......~ .......( ..····f·······:· .. .
(ij .. o

.. •
.. •

.,
,

E 2
(;
z
00 0.1 0 .2 0 .3 0.4 0.5 0.6 0 .7 0 .8 0 .9 1 .0
Normalized VFT parameter Xo =TofT A
Fig. 3.21. Stability constraint in the technically relevant range of VFT parameters

that the critical final temperature for nearly all technical glasses lies between
80% and 95% of the initial temperature.

3.2.3 Appendix: Derivation of Several Equations

The following problem-adapted normalization is obvious:

z--+L(, --+RC
2V (3.81 )
w --+ 7rR2 W ,

Then, the two normalized balance equations are:


~~ ( TJ(TA 8)
~ &~ ~ TJ(TA )
ow) _&Il
&~ - &( , (3.82)

2pcp V &8 R2 &28 1 & ( &8)


7rkL w &( = L2 &(2 + Z&~ ~ &~ (3.83)

The normalized initial and boundary conditions for the energy equation read:

8(~,0) = 1,
&8
&~ (0, () = 0, (3.84)
&8
- &~(1,() = NuT(l,() .

In addition to the Nusselt number, the following numbers are defined:


206 3. Homogenizing and Conditioning

Reynolds number ,

Prandtl number , (3.85)

aspect ratio .

Owing to the given conditions, the aspect ratio e is very small. Therefore,
the term of order e 2 is neglected in the energy equation (3.83) (i.e., the heat
conduction in the direction of the axis), which simplifies the energy balance:

1 a ( ae)
ae = Z
2Re Pr e W ae a~ ~ a~ . (3.86)

Coarse Model: Averaging Over the Cross Section

If the normalized boundary conditions are used and (3.86) is averaged over
the pipe cross section, it follows that

2Re Pre Jo
r ae
1ae ~ d~ = -Nu e(l, () .
W (3.87)

For the coarse model, we are now searching for a solution of the system of
equations given by (3.82) and (3.87). In order to make up for the information
"lost" by averaging, we make the following closure ansatz:

e;:::;(j(() . (3.88)

By approximation, in the coarse model the temperature is substituted by the


temperature averaged over the pipe cross section. Naturally, this temperature
only depends on the axial coordinate (, therefore the viscosity also does not
depend on ~, and, by a double integration of (3.82) over ~, we get

(3.89)

Because the volume flow is independent of e, by averaging (3.89) over the


pipe cross section we can see that

(3.90)

In the coarse model, the normalized velocity profile W is just a parabola with
the maximum value 1. If we insert this wand the closure ansatz for e into the
averaged energy equation (3.87), we can integrate the left-hand side, which
yields the ordinary differential equation
3.2 Instabilities and Stabilization of Glass Pipe Flows 207

1 de -
-Re Pr c;- = -NuB. (3.91)
2 d(

If the initial conditions are accounted for, the solution of this equation is

-B(() = exp (2NU)


- Re Pr c; ( . (3.92)

If the normalization is reversed in (3.92), the result is (3.65) for the mixing
temperature. The normalized version of the equation for the pressure gradient
over the pipe segment in the coarse model results from the integration of
(3.90) over (:

Jl(O) - Jl(I) = - (4- )


T] TA
11
0
T](TAe(())d( . (3.93)

Refined Model: Asymptotic Expansion According to the


N usselt Number

In order to solve the temperature profile in the direction of the radius, B is now
defined as an asymptotic expansion in the Nusselt number in the following
way:

B(~, () = e(()(1 + cp(~)Nu + O(Nu 2 )) . (3.94)

Because the zeroth-order approximation of B is given bye, the zeroth-order


approximation of w is the simple parabola from the coarse model. If we insert
this result into the normalized energy balance (3.86) and neglect terms of
order N u 2 , we get an ordinary differential equation for the function cp:

-4(1 - ~2 ) = -1 -a ( ~-
acp) (3.95)
~ a~ a~

A double integration yields

cp(~) = cp(O) - e + 4e
1
. (3.96)

Of course, the cross-sectional mean value of B must remain equal to there- e,


fore cp(O) can now also be determined by the integration of (3.94) over the pipe
cross section. This finally yields the normalized version of the temperature
distribution in the direction of the radius given in Sect. 3.2.2:

(3.97)

By means of (3.97), we are now able to write a more exact equation for the
axial velocity. First, the integration of (3.82) yields
208 3. Homogenizing and Conditioning

(3.98)

Now, the viscosity profile is approximated by the following exponential ap-


proach, which gives a good approximation of the VFT curve in the domain
between TA and T E :
7](TA )) (TA -T)/(TA -TE)
7](T) = 7](TA) ( ---=--
7](T E) (3.99)
= 7](TA) exp (b 1
1- BE
-.!!) ,
where b = In(7](TA)/7](TE)) and e E = TE/TA. If we insert this viscosity curve
together with (3.97) for the temperature into (3.98), this yields

w(~,()
loll 7](TA)
= -----_-
2 o( 7](TAB) ~
11 ( -
~exp
'P ) de
NubB-_-
1 - BE
(3.100)

Because we assume that Nu « 1, the exponential term of the integrand of


the right-hand side of (3.100) can be linearized:

exp (NU b e~) Rj 1 + Nu be ~ . (3.101)


1 - BE 1- BE
Thus, the integrand is a polynomial in ~, and the integration yields:

Averaging of (3.102) over the pipe cross section results in

(3.103)

If this equation is solved for the normalized pressure gradient, we get


- --1
_all = 47](TAB) [l-~b~NU] . (3.104)
o( 7](TA) 8 1 - BE
The substitution of (3.104) into (3.102) now yields the normalized version
of the refined equation for the axial velocity. The normalized version of the
refined equation for the pressure gradient is given by the integration of (3.104)
over (. Here, the following approximation for small Nusselt numbers is also
applied:

(3.105)
3.3 Shape Optimization of Flanges 209

3.3 Shape Optimization of Flanges

Norbert Siedow, Horst Loch, Sandra Manservisi


Introduction

The electric heating of platinum tubes, where the cooling of the glass (and
thus, e.g., the throughput) can be influenced by generating variable amounts
of Joule heat in the pipe wall, is state-of-the-art. The electric current is fed
into the pipe via so-called flanges, which, as a first approximation, can be
described as thin annular disks around the pipe. Apart from their main task,
namely the homogeneous conduction of a given current into the pipe in the
circumferential direction, flanges should have no local thermal influence on
the glass at the electrical contact area, in order to avoid the development
of crystals or striae. In high-quality production, such inhomogeneities can
negatively influence even the forming process.
The current is fed into the pipe by one or more electrical contacts and
then distributes in three dimensions according to the local electrical resistance
(depending on the materials, temperatures, and geometries) until it reaches
the electrical contact region with the pipe. Hence, from a mathematical point
of view, a continuum problem for several coupled variables (potential, current,
temperature) must be solved. Therefore, the design of optimum flanges is a
highly difficult inverse problem where economic factors (e.g., reduction of the
amount of platinum needed) are also very important, apart from the technical
aspects mentioned above.
In the following, we will show how the design of flanges can be supported
by mathematical methods. First, we will apply general shape-optimization
methods where the geometry (i.e., the exterior contour and the thickness
distribution) of the flange can, in principle, vary arbitrarily. We decided for
this first approach despite the fact that the results (e.g., continuously varying
sheet thicknesses) cannot be realized in technical constructions. Our objective
was to learn something about typical structures of optimized flanges and to
use that knowledge in order to design simplified geometric forms as a basis
for further realistic optimizations with only a few degrees of freedom. In
Sect. 3.3.2, we will describe such a reduced approach, i.e., a 3-ring/spoke
structure, and we will show that it leads to good results. This structure has
the advantage that it can be optimally adapted to each special situation.
This can be done, for example, with the commercial finite-element-method
software ANSYS and the finite-dimensional optimization algorithms included
there.
210 3. Homogenizing and Conditioning

3.3.1 General Shape Optimization: Continuously Varying


Thicknesses and Contours

The 3D Problem

Let us consider the three-dimensional platinum flange fl with a platinum


pipe T. The upper and lower flange facing is called r f , the electrical contact
area of the flange ra, the electrical contact area of the pipe rro and the
remaining side areas of the flange are called ro. The entire surface of the
flange is thus defined by

(3.106)

The entire surface of the pipe is denoted by E, which includes the interior
surface E g , the end of the pipe E e , and the electrical contact area with the
flange E r . The remaining areas are called Eo, so that the entire surface is
determined by

E = Eg U Ee U Er U Eo (3.107)

(see Fig. 3.22).


The space-dependent potential distribution ¢( x) in such a flange-pipe
system is described by an elliptic differential equation:
1
-b.¢(x) = 0 , xEfluT. (3.108)
Pe
Pe is the electric resistance. In order to complete this equation, several bound-
ary conditions are necessary. At the electrical contact boundary ra, a constant
current density C a is given:

.! o¢ _ C
Pe on - a,
(3.109)

and at the end of the pipe Ee the electric potential is zeroed:

¢= 0, (3.110)

x.

~-z Fig. 3.22. Front and side view of the 3D


flange
3.3 Shape Optimization of Flanges 211

At the remaining boundaries, we assume that there is no electric current.

~a¢ -0 (r u E) \ (ra u Ee) .


Pe an - , x E (3.111)

The electric current produces Joule heat within the area (n u T),

J(x) = (V'¢(x)? , (3.112)


Pe
which leads to a heating of the flange and the pipe. Hence, the temperature
distribution can be described by the heat conduction equation

(V' ¢)2
ktD..T(x) = - - - = -J(x), x E nuT. (3.113)
Pe
Here, T(x) denotes the temperature and k t the heat conduction coefficient
in the platinum. The temperature at the electrical contact boundary ra is
described as Ta (x ) :

T(x) = Ta(x), (3.114)

A temperature exchange can take place between the pipe and the hot glass
melt E g , which is described by a boundary condition of mixed type:

x E Eg . (3.115)

a g is the heat transfer coefficient between the platinum pipe and the glass
melt, and Tg the glass temperature (which is assumed to be constant). All
the other boundaries are adiabatic:

x E (r U E) \ (ra u Eg) . (3.116)

As already described in the introduction, the temperature of the glass flowing


through the pipe should not be influenced by the flange. Hence, the temper-
ature of the flange in the immediate proximity of the contact area should be
equal to the temperature of the glass T g . Thus, we have to solve the following
shape-optimization problem:
Find a surface shape of the flange n that minimizes the integral
~ J (T(x) - Tg)2dx ----+ min (3.117)
s
within a small section Sen at the pipe boundary rn and a temper-
ature distribution T (x) that fulfils (3.113) with the boundary condi-
tions (3.114)-(3.116).
212 3. Homogenizing and Conditioning

Such a three-dimensional shape optimization problem is mathematically very


complex. A general theory for the solution of such problems can be found,
for example, in [3.9]. In our case, however, the spatial dimensions of the
flanges have a special feature: they are very thin. So the space coordinates
change on different scales: the flange thickness z = h(x, y) is very small
compared to the length x and the width y. This motivates the introduction
of an asymptotic expansion, in order to reduce the three-dimensional shape
optimization problem to a simpler two-dimensional problem.

Asymptotics: Small Thickness

Let us consider only the flange st and a heat transfer equation with a gener-
alized source term and with boundary conditions of mixed type at the flange
facings rf = rf(X, y):

kt~u(x) = F(x) , x Est,


au
kt an = a(u - u) , x E rf. (3.118)

n = (nx, ny, nz)T is the vector of the exterior normal. At the side areas
ra u rr u ro, additional boundary conditions are necessary, which will be
discussed a little later. We also assume that the flange is symmetric in the
thickness direction. The plane z = 0 is drawn through the middle of the
flange. We have

au = 0 for z =0 . (3.119)
az '
With the shape transformation (x, y, z) ---+ (x, y, z) with x = x, y = y, z =
cz, c « 1, the flange is rescaled, c being a small parameter. h(x, y) denotes the
thickness of the flange. After this transformation, from (3.118) there follows
a2U a 2u) a 2u
c 2 k t ( ax 2 + ay2 + kt az 2 = c 2 F(x) , x Est, (3.120)

x E rf. (3.121)

If we account for the relations nx/nz = -hx(x, y) and ny/nz = -hy(x, y),
for the boundary condition (3.121) it follows that

aU h -au) + -au a
ck t ( -h x -
ax
-
Yay az
= c-(u
n
-
z
u) , x E ra. (3.122)

The solution of the differential equations (3.120) and (3.122) is described as


a series expansion with respect to the small parameter c:

u(x, y, z) = uo(x, y, z) + c2Ul(X, y, z) + .... (3.123)


3.3 Shape Optimization of Flanges 213

A comparison of the coefficients with respect to E yields


o. a2uo_
E • kt az 2 - 0 ,
auo
kt az = 0, z =0, z=h; (3.124)

z = h;

(3.125)

From (3.124) it follows immediately that Uo cannot depend on z: Uo =


uo(x, y). If we presuppose that the right-hand side of (3.118) is independent
of z, F = F(x, y), the first equation (3.125) yields the relation

(3.126)

Therefore, also the second derivative of U1 with respect to z can only be a


function of (x, y).
Integration of (3.126) results in a linear function in z:

(3.127)

From the symmetry condition kt(auI/az) = 0 for z = 0 it follows that


C 1 (x, y) = O. Thus, the boundary condition for z = h in (3.125) yields the
function C 2 (x, y):

(3.128)

A second integration results in a term for U1(X,y,Z). We conclude that U1


can only be a function quadratic in z:

C 2 (x, y) 2
U1(X, y, z) = 2 z + Co(x, y) . (3.129)

C 2(x, y) can be determined by (3.128), and Co(x, y) is a still unknown func-


tion, which, however, is unimportant in the following.
However, what is important is that, from (3.126) and (3.128), we have
found a differential equation for the function uo(x, y):
214 3. Homogenizing and Conditioning

a ( auo) a ( auo) a _ h(x,y)


ax hex, y) ax + ay hex, y) ay - nzkt (uo - u) = ~F(x, y) .
(3.130)

As a first step, u(x, y, z) can be approximated by the function uo(x, y),


which is the solution of the two-dimensional differential equation (3.130).
The boundary condition at the flange facings n
has thus become part of this
differential equation.
Based on the discussion of the general problem (3.118), (3.121), for the
thermoelectric problem (3.108), (3.113) we get a two-dimensional substitute
instead of the three-dimensional equations.
¢(x, y, z) = ¢o(x, y) + E¢l(X, y, z) + ...
a ( h(x,y) a¢o)
ax ax a
+ ay ( h(x,y) a¢o)
ay = 0, (x,y) E D (3.131)

and
T(x, y, z) = To(x, y) + ET1 (x, y, z) + ... (3.132)
~
!::>
(h( )aTo)
x, y!::> +~!::>
(h( )aTo) = _ h(x,y) (V'¢O)2
x, Y !::> k ' (x,y)ED.
uX uX uy uy t Pe
In order to receive complete two-dimensional substitute problems, we will
now derive the boundary conditions for (3.131) and (3.132).
Let us first consider the electrical contact boundary Ta. Here, the exterior
normal of the area is given by the vector n = (-1,0, O)T. This immediately
yields

(x,y) E Ta , (3.133)

and

To(x, y) = Ta(x, y) , (x,y) E Ta , (3.134)

where Ta is supposed to depend exclusively on (x, y). Except for the pipe
boundary Tn we have homogeneous Neumann conditions.

(3.135)

and
aTo
kt an = 0, (3.136)

Through the asymptotic expansion, the flange could be approximated as


a two-dimensional object. However, if we consider the combined system of
flange and pipe, the asymptotic expansion can no longer be applied, because
3.3 Shape Optimization of Flanges 215

the pipe has a significant length in the direction of z. Therefore, we will ne-
glect the pipe. Nevertheless, in this case we must find new rr conditions for
the electrical contact boundary that did not exist in the original problem.
The new boundary conditions should approximate the influence of the plat-
inum pipe on the flange as realistically as possible. For the approximation of
the boundary condition for the potential, we reduce the system flange-pipe
to a one-dimensional object:

pipe, flange,

where lr denotes the pipe length. Due to the homogeneous potential equation
and the boundary condition <Po(O) = 0, the behavior of the potential along
the pipe is linear, <Po (z) = C z. If we use the boundary condition h( x, y) ~<p;: =
etr<po for z = lr, the heat transfer coefficient is
h(x,y)
et r = - - - . (3.137)
lr
Therefore, for the two-dimensional substitute problem we use

h( X, Y) a<po=h(x,y)A-,
an lr 'Po,
f
or rr . (3.138)

As for the two-dimensional potential equation, the condition at the flange-


pipe boundary must also be modified for the temperature equation (3.132) in
order to be able to account for the pipe. An energy balancing approximately
results in

(x,y) Err. (3.139)

Now, the two-dimensional substitute problem is uniquely described.

Thickness Optimization as a 2D Problem

Equations (3.131), (3.133), (3.135), (3.138) and (3.132), (3.134), (3.136),


(3.139) completely describe the direct problem connected with the temper-
ature computation of the two-dimensional flange. Now, we will turn to the
inverse optimization problem:
Find a function h( x, y) (thickness distribution of the three-dimension-
al flange) which, in a small section S c fl at the pipe boundary rr)
minimizes the integral

under the following constraints:


216 3. Homogenizing and Conditioning

• To(x, y) fulfils the temperature equation,


• c/Jo (x, y) fulfils the potential equation,
• the function hex, y) is bounded from above and below:
hmin ::; hex, y) ::; hmax .

Instead of a three-dimensional shape-optimization problem, we must now


solve a two-dimensional parameter-identification problem, which is a much
easier task. This inverse problem is ill-posed. Therefore, a regularization is
necessary. As a new optimization criterion, we use

~ J
s
(To(x) - Tg)2dx +~ J
n
h2(X, y)dx --+ min (3.140)

The second term - the regularization term - means that the amount of mate-
rial necessary for the flange should also be minimized. This is obviously im-
portant for economical reasons. The parameter 'T weights both requirements:
uniform temperature profile and minimal use of material. For the solution of
this optimization problem, we use the Lagrange multiplier method:

.c(h, T,p, c/J, v, s, fL) =~ J


s
(To(x, y) - Tg)2dx +~ J
n
h2(X, y)dx

+ J( h(x,y)
"'\l(h(x, y)"'\lTo(x, y) + -k-t
-
Cev c/Jo(X,y))2)
Pe p(x, y)dx (3.141 )
n
+ J
n
("'\l(h(x, y)"'\lc/Jo») vex, y)dx

.c(h,To,p,c/Jo,V,S,fL) is the Lagrange objective function, where the first two


terms form the optimization term, and the other three terms describe one
constraint each. p(x, y), vex, y), and fLeX, y) are the respective continuous La-
grange multipliers, where p(x, y) stands for the temperature in the constraint,
vex, y) for the potential in the constraint, and fLeX, y) for the thickness in a
usefully formulated constraint. We demand that

We apply the Gauss theorem to the third and fourth integral of the right-hand
side, which results in the weak formulations of the potential and temperature
equations:
3.3 Shape Optimization of Flanges 217

.c(h,To,p,¢o,V,S,p,) ="2 11 (To(x,y) -Tg) 2 dX+"2 "11 h 2 (x,y)dx


s n
+1( n
-h(x, y)\lTo(x, y)\lp(x, y)
h(x,y) (\l¢o(X,y»2
+ -k-t - Pe
)
p(x, y) dx

+1 r,
lrar
k;(To(x, y) - Tg)p(x, y)ds

+1 (-h(x, y)\l¢o(x, y)\lV(X, y» dx + 1 h(x, y)PeCaV(X, y)ds


n ra
h(x,y)
+ 1 -lr-¢O(X' y)v(x, y)ds

r r)
r,

+ 1(
n
(h(X, y) - h max ; h min + S2 - (hmax ~ h min p,(x, y)dx .

We form the variation of this objective function

+ 1(
n
-h\lTo \l8p + h(~~:)2 8P) dx + 1
G
(lr:r (To - Tg)8P) ds

+ 1(k~~e
n
\l¢o\l8¢oP) dx + 1(-h\l8¢o\lV)dX +
n
1~8¢oVdS
r,
+ 1(-h\l¢o\l8V)dX +
n
1
ra
hpeCa8vds + 1
r,
0¢o8vdS

+ 1
n
2s8sp,dx + 1(
n
(h - h max ; hmin ) 2+ s2 _ (h max ~ hmin ) 2) 8p,dx

+ 1
n
"I Mhdx + 1
n
(-8h\lTo\lP + 8h~~~O)2 p) dx

+ j(-8h\l¢o\lV)dX+ j 8hpeCavds+ j~~¢oVdS


n ra r,
+ j 2 (h _ h max ; hmin ) 8hp,dx 1

n
where 8To1 8p, 8¢o, 8v, 8s, 8p" and 8h are linearly independent. If we choose
in succession
218 3. Homogenizing and Conditioning

r5v = w E HI, r5To = r5p = r5c/JO = 158 = r5/-L = r5h = 0 ,


r5p = W E Hfj, r5TO = r5c/JO = r5v = 158 = r5/-L = r5h = 0 ,
r5TO = W E Hfj, r5p = r5c/JO = r5v = 158 = r5/-L = r5h = 0 ,
r5c/JO = W E HI, r5TO = r5p = r5v = 158 = r5/-L = r5h = 0 ,
this results in
• the direct potential problem

(x,y)ED,

(3.142)

an = !!.-A,
h Oc/JO lr '1-'0, (x, Y) E rr ,

ha;;:=O, (x,y) Er\(rrUra),


• the direct temperature problem

~ )OTo) + ~
;:) (h( x, y;:) ;:) (h( x, Y )OTo)
;:)
= _ h(x,y)
k
eV'c/Jo)2 ' (x,y) ED,
uX uX uy uy t Pe

oTo lrar ( )
h (x, y ) an =T To - Tg , (x,y) E rr, (3.143)

T(x, y) = Ta(x, y) , (x, y) E ra ,


• the adjoint temperature problem

a ( h(x, y) ax
ax op) a
+ oy ( h(x, y) oy
op) = -(T - Topdr5(S) , (x,y)ED,

op lrar
h(x,y) an = k;P, (x,y) Err, (3.144)

op
h(x, y) an = 0, (x, y) E r \ (rr u ra) ,

p(x,y) = 0, (x,y) Era


• and the adjoint potential problem
3.3 Shape Optimization of Flanges 219

:x (h(X, y) ~~) + :y (h(X, y) ~~) = k t2Pe \1p(h\1¢) , (x,Y)ED,

av 2
han = kt hpCa, (x,y) Era, (3.145)

a -
h av _ 2hp '1-',
A.
o ( ) 1"'
X, Y E L r ,
n ktPelr

h~~=O, (x,y)Er\(raurr ).

If 58 = W E HI, 5To = 5p = 5¢o = 5v = 5JL = 5h = 0, this yields the

J
integral equation
2 8JLwdx = 0 .
n
It follows immediately that 8JL = O. Because JL is an arbitrary test function,
we have 8 = O. From 5JL = W E HI, 5To = 5p = 5¢o = 5v = 58 = 5h = 0,
it follows that

J( (h - h max ~ hmin) 2_ (h max ~ hmin ) 2) 5JLdx = 0,


n
which leads to a quadratic equation

The solutions of this equation are h = h max and h = h min . Hence, it is


unnecessary to formulate an equation for the constraint hmin :s: h :s: h max in
the solution algorithm. When a new function h(x, y) is found, it is easy to
check whether h lies within acceptable bounds. If not, h(x, y) is set back to
the value of the respective bound.
For the new thickness distribution h(x, y), we finally get the equation

p(\1¢O)2)
hnew = h+ (Y ( '"jh - \1To\1p - \1¢o\1v + ktPe . (3.146)

Here, (Y is a damping parameter. In conclusion, let us summarize the solution


algorithm:
• initialize k = 0
(i) determine the initial thickness of the flange h = h(x, y), choose the
damping parameter (Y and the regularization parameter '"j, select the
maximum number of iterations k max and a termination bound E;
(ii) solve the potential equation (3.142) ::::} ¢o, \1¢o I I,
220 3. Homogenizing and Conditioning

(iii) solve the temperature equation (3.143) '* ITo, V'To I,


(iv) compute the objective function

E(h) = ~j (To (x, y) - Tg)2dx +~j h2(X, y)dx ,


s n
• iteration: k = k + 1
if k > kmax goto (xii)
(v) solve the adjoint temperature equation (3.144) '* Ip, V'p I '
(vi) solve the adjoint potential equation (3.145) '* I v, V'v I'
(vii) determine a new flange thickness h

(viii) solve the potential equation (3.142)'* I¢~, V'¢~ I '


(ix) solve the temperature equation (3.143) '* ITt, V'Tt I '
(x) compute the objective function

Ek(h) = ~ j(T!f(x,y) - Tg)2dx + ~ j(hk(x,y))2dX ,


s n
(xi) if Ek < E k - 1 , then
- (J = 1.2(J
- goto (v)
otherwise
- (J = O.25(J
- if (J < c goto (xii)
- goto (vii)
(xii) stop.

Numerical Tests I: Thickness Optimization

The numerical tests are done for a circular flange with a one-sided electri-
cal contact, such as shown in Fig. 3.23. (A denotes the electrical contact
area whose thickness is supposed to be constant. D corresponds to the opti-
mization area S.) We nOw consider a quarter of the flange that ought to be
optimized. AU the dimensions mentioned refer to this quarter of the flange.
A Dirichlet Problem. We start the numerical tests with a problem whose
boundary conditions have been changed. Nevertheless, this academic problem
appears interesting because the basic optimization operations become very
clear here.
3.3 Shape Optimization of Flanges 221

L2 Fig. 3.23. Flange geometry

We consider the two-dimensional potential and temperature equations


(3.131), (3.132) with the following boundary conditions:

cPo (x, y) = 0.225, (x, Y) Era,


cPo(x,Y) = 0, (x,y) Err

for the electric potential and

To(x, y) = 573.15, (x, y) Era,


oTo
h(x, y) on (x, y) = 0, (x, y) Err

for the temperature. Thus, at the electrical contact region of the flange, po-
tential and temperature are constant. At the electrical contact area of the
pipe, the potential must be zero, and there must not be any heat exchange.
Of course, this condition is very academic. The thickness of the flange can
vary between hmin = 5 X 10- 6 m and h max = 5 X 10- 2 m. The glass tempera-
ture is Tg = 1430 K. The thickness of the electrical contact region is 0.004 m
and cannot be changed. We start the computation with "y = 0 and a constant
flange thickness of 0.004 m.
The optimization results in the thickness profile shown in Fig. 3.24. The
constant thickness of the electrical contact region can clearly be seen. Then,
the flange thickness is set to the minimum value before it returns to the
initial thickness. The optimization algorithm has created an indentation in
the flange, trying to "separate" the optimization region - the circular flange -
from the electrical contact region. Thus, the mathematical problem becomes
singular, and the algorithm has every degree of freedom in order to reach the
optimization objective. Figure 3.25 shows the temperature profile along line
L1 before and after the optimization. After the optimization, we can clearly
recognize a temperature plateau around 1430 K.

0.500E·05
_ 0.449E·03
0.893E·03
_ 0.0 01337
_ 0.001781
0.002224
c::::J 0.002668
c::::TI 0.003112
_ 0.0 03556
0.004

Fig. 3.24. Optimized thickness profile


222 3. Homogenizing and Conditioning

1496.438-r-----------==~-, a)
1404.109 L1
1311.780
1219.451
1127.122
1034.794
942.465
850.136
757.807
665.478
573.150 +----.-----.-----r--.--.-----.---'
o 0.576 1.153 1.730 2.307 2.884 3.461
Dis! x 10-1
1431.542 .---------====----"""] b)
1345.703 L1
1259.864
1174.024
1088.185
1002.346
916.507
830.667
744.828
658.989
573.150 .jL--...----.----.----.-----.-----r----'
o 0.576 1.153 1.7302.307 2.884 3.461
Dis! x 10-1
Fig. 3.25. Temperature (a) before and (b) after the optimization

More Realistic Boundary Conditions. The next example is a flange with more
realistic boundary conditions
• for the potential:

a¢o
h(x, y) an = h(x, Y)PeCa , (x,y) Era,

h( )a¢o = h(x,y) A,
(x,y) Err
x, Y an lr '1'0,

• and for the temperature:

To(x, y) = 573.15, (x, y) Era,


aTo l;at
h(x,y) an = --,z;(To -Tg), (x,y) Err.

Here, Pe = 5.5 X 10- 7 n m, lr = 0.5 m, C a = 2.08333 x 106 A/m 2 , at =


40 W /K m 2 , and Tg = 1450 K. The flange thickness can vary between
5 x 10- 6 m and 5 x 10- 2 m. The optimization starts with a thickness of
0.02 m within the optimization region. We get the thickness profile depicted
3.3 Shape Optimization of Flanges 223

in Fig. 3.26. Figure 3.27 again shows an indentation at the fixed electrical
contact boundary.
Figure 3.28 shows the temperature profile along the line L1 before and af-
ter the optimization. The entire temperature profile is described by Fig. 3.29.
The optimization algorithm increases the entire temperature level and
creates an evident plateau at 1450 K. With a result of 0.03 m, the optimized
flange is relatively thick and expensive. Therefore, in the last example we
want to minimize the complete objective function and thus the amount of
material necessary for the flange.
Additional Parameters. We change the data of the previous example at two
points. The thickness of the flange may only vary between 5 x 10- 4 m and
5 x 10- 1 m. In addition, the regularization parameter must fulfil "Y i- O. The
initial thickness of the flange must decrease linearly, starting at the electrical
contact region (see Fig. 3.30a). Figure 3.30b shows the optimized thickness
profile. Here, too, the indentation is characteristic.
Figure 3.31a,b shows the temperature profiles in a three-dimensional
flange. After the optimization, the temperature distribution in the flange is
essentially more homogeneous. Figure 3.32a,b shows the heat flow before and
after the optimization. The heat flows in the case of the optimized thickness
are obviously smaller.

_ 0.5 00E-05
_ 0.0 0556
0.011115
_ 0.01667
_ 0.0 22225
0.02778
c:::J 0.033335
0.03889
_ 0.0 44445 Fig. 3.26. Optimized
0.05 thickness profile

x 10-2
5.000,----------------,
4.515 L1
4.030
3.546
3.061
2.577
2.092
1.608
1.123
0.638
0.154 +==:;::::,...L.--_,....-_,....-_.---,~
o 0.576 1.153 1.730 2.307 2.884 3.461 Fig. 3.27. Optimized thickness
Dist x 10-1 along line L1
224 3. Homogenizing and Conditioning

1425.186 a)
1339.982 L1
1254.778
1169.575
1084.371
999.168
913.964
828.760
743.557
658.353
573.150
0 0.576 1.153 1.730 2.307 2.884 3.461
Dist x 10-1

1450.866 b)
1363.095 L1
1275.323
1187.551
1099.780
1012.008
924.236
836.465
748.693
660.921 Fig. 3.28. Temperature
573.150 profile along line L1 (a) be-
0 0.576 1.153 1.730 2.307 2.884 3.461 fore and (b) after the opti-
Dist x 10-1 mization

573.15 a)
_ 670.365
767.579
864.794
_ 962.009

.L _ . . ~
c:::J 1059
c:::::J 1156
c::::J 1254
_ 1351
1448

573.15 b)
- 670.745
- 768.34
865.935
_ 963.53
c:::::J 1061 Fig. 3.29. Temper-
c:::::J 1159
1256 ature (a) before and
1354 (b) after the opti-
1452 mization

Contour Optimization as a 2D Problem

Up to this point, we have developed a shape-optimization procedure resulting


in the optimal thickness profile of the flange. The exterior contour - the
3.3 Shape Optimization of Flanges 225

_ 0.001 a) l)
_ 0.0 01333
0.001667
0.002
- 0.002333
c:::::J 0.002667
0.003
c:::::J 0.003333
_ 0003667
0.004

_ 0.500E-03 b) ,)
_ 0.0 02072
0.003644
_ 0.0 05216
0.006788
c:::::J 0.008361
c:::J 0.009933 Fig. 3.30. (a) Ini-
c;::::::J 0.011505 tial and (b) opti-
0.0130n mized thickness pro-
0.014649 file
350 a)
570
790
1010
1230
1450
1670
1890
2150

350 b)
570
790
1010
1230
1450
1670
1890
2150

Fig. 3.31. Tempera-


ture profile in a three-
dimensional flange (a)
before and (b) after the
optimization

straight electrical contact boundary and the circular flange boundary - was
left out. Now, we want to examine a shape optimization where the contours
of these boundaries can also vary.
We consider a smooth section of the boundary between the points
(xo, y( xo)) and (Xl, Y( Xl))' The boundary curve between these points is given
226 3. Homogenizing and Conditioning

MIN = 5.833 MAX = 1 'uu,~_ _ _ 5.833 a)


_ 1500
3000
5000
_ 7000
9000
c=J 11000
_ 13500
16500

M" . >'529 MAX . "D _


_

_
1500
3000
5000
6500
1.529 b)

8500
c=J 11000
13000
16500

Fig. 3.32. Thermal gradient (a) before and (b) after the optimization

explicitly as a function ,6(x) and is changed in such a way that within the
section 5, the temperature of the flange differs only minimally from the glass
temperature. An additional condition is that the boundary that ought to
be optimized is smooth. The basic method is the same as for the thickness
optimization described above. The new objective function is

J J
X,

E(,6) = ~ (To(x, y) - Tg)2dx + ~ q2(x)dx -----+ min (3.147)


S Xo

q( x) fulfils the ordinary differential equation

d 2 ,6
dx 2 (x) = q(x), x E (xo, xd ,
,6(xo) = Yo, ,6(xd = YI , (3.148)

and is a measure for the smoothness of the curve ,6(x). Equation (3.148) is
also called the shape equation. Hence, the optimization problem is:
Find a function ,6(x), Xo :::; x :::; Xl describing the boundary, which
minimizes the objective function E(,6) fmm (3.147) within a small
3.3 Shape Optimization of Flanges 227

section SeD at the pipe boundary rr. The following constraints


must be accounted for:
• To(x, y) fulfils the temperature equation,
• cPo(x, y) fulfils the potential equation,
• (3(x) fulfils the shape equation (3.148).
With the Lagrange multiplier method, we can derive the necessary equations
in the same way as for the thickness optimization. Here, we only want to
present the final algorithm:
• initialize k = 0
(i) determine the initial shape of the flange y = (3(x) and set q = (3"(x),
select the damping parameter a- and the regularization parameter "
determine the maximum number of iterations kmax and a termination
bound 10;
(ii) solve the potential equation (3.142) =? I cPo, \lcPo I '
(iii) solve the temperature equation (3.143) =? I To, \lTo I '
(iv) compute the objective function (3.147),
• iteration: k = k + 1
if k > k max goto (xiii)
(v) solve the adjoint temperature equation (3.144) =? p, \lp I I,
(vi) solve the adjoint shape equation

f.l"(x) = h(\lT \lp)ny\!l + ((3'(x))2, x E (xo,xt)


f.l(xo) = f.lo, f.l(X1) = f.l1 (3.149)
=?~,

(vii) determine a new q:

qk = qk-1 _ 5(f.l + ,qk-l) ,

(viii) solve the shape equation (3.148) =? ~ ,

(ix) solve the potential equation (3.142) =? I cP~, \lcP~ I '


(x) solve the temperature equation (3.143) =? IT~, \lT~ I '
(xi) compute the objective function (3.147) ,
(xii) if Ek < E k - 1 , then
- a- = 1.2a-
- goto (v) ,
otherwise
- a- = O.25a-
- if a- < 10 goto (xiii)
- goto (vii)
(xiii) stop.
228 3. Homogenizing and Conditioning

However, practical application of the algorithm is connected with a consid-


erable difficulty: after the solution of the shape equation in step (viii), a new
meshing of the area is necessary. If the flange thickness is constant at each
point, no problems occur. However, if h(x, y) is a space-dependent function,
it must be mapped appropriately onto the new flange geometry. We now
present several test examples.

Numerical Tests II: Combined Contour and Thickness Optimiza-


tion

Variation of the Flange Contour. We consider a flange of constant thickness


with a temperature of 1450 K at the pipe boundary. The optimization refers
to the exterior shape of the initially circular flange. The algorithm yields
the shape depicted in Fig. 3.33. As in the previous examples, the algorithm
again tries to "separate" the contact boundary from the optimization region.
Here, this becomes clear through the attempt to enlarge the flange as far as
possible. Because in the direction of the x axis the expansion of the flange is
fixed, the results are a kind of "menhir" flanges.
Variation of the Electrical Contact Boundary. In the thickness optimization
examples, the algorithm always produced an indentation at the electrical con-
tact boundary. Therefore, in the next example we want to optimize the shape
of this contact region. We start with a straight line (see Fig. 3.34a). After 14
iteration steps, the algorithm leads to the result depicted in Fig. 3.34b.
There is a waist at the electrical-contact boundary similar to the inden-
tation. Thus, our former conclusion is confirmed:
The optimization algorithm tries to "separate" the flange area from
the electrical contact. The mathematical problem becomes singular
and there are more possibilities available for the algorithm in order
to reach the optimization objective.
Combined Thickness and Contour Optimization. Finally, we want to combine
the optimization of thickness and contour. We start with a flange of constant
thickness and with a straight electrical contact (see Fig. 3.34a).
First, the electrical contact is optimized. We again assume a glass tem-
perature of 1450 K. The result of this optimization is shown by Fig. 3.35.

573.15
_ 745.911
918.673
_ 1091
_ 1264
1437
c::::::J 161 0
1782
_ 1955
2128 Fig. 3.33. Temperature after optimizing the
flange contour
3.3 Shape Optimization of Flanges 229

573.15
- 670.365 a)
- 767.579
_ 864.794
_ 962.009
1059
c::J 1156
1254
_ 1351
1448

_ 573.15 b)
_ 672.31
771.47
_ 870.63 Fig. 3.34. (a) Tem-
_ 969.791
perature in the initial
c::::J 1069
c::::J 1168 flange and (b) temper-
1267 ature after the opti-
_ 1366 mization of the electri-
1466 cal contact
573.15
- 670.254
767.358
_ 864.462
_ 961 .566
1059
c::::::J 1156
1253 Fig. 3.35. Temperature
_ 1350 after the shape optimiza-
1447 tion

Figure 3.36a,b shows the temperature profile along the line Ll before, re-
spectively after, the shape optimization.
Based on this new shape of the contour, we now search for the corre-
sponding optimal thickness profile. The results yielded by the algorithm are
depicted in Fig. 3.37.
Figure 3.38 shows very clearly that the attempt to keep the temperature
distribution in the flange at a nearly constant level of 1450 K has been suc-
cessful. The temperature difference now is only 2.61 K, compared to 175.88 K
for the initial flange. The temperature difference at the pipe boundary has
decreased from originally 22.89 K to 0.08 K.
Finally, Fig. 3.39a,b shows the thickness distribution in the optimum
flange. This optimization does not account for the limits with respect to
the maximum thickness. We can see the indentation as well as the waist at
the electrical contact.

3.3.2 Finite-Dimensional Shape Optimization:


the 3-RingjSpoke Flange

We have applied the general shape-optimization methods to typical situations


in very different ways. Based on repeated observations of the same phenom-
ena, we developed a simple flange concept that has the advantage of being
230 3. Homogenizing and Conditioning

1425.186.---------------="""' a)
1339.982 L1
1254.778
1169.575
1084.371
999.168
913.964
828.760
743.557
658.353
573.150 +----r------,----,--,..---,----,----'
o 0.576 1.153 1.730 2.307 2.884 3.461
Dist x 10-1
1433.5441------:=========~-1 b)
1347.505 L1
1261.465
1175.426
1089.386
1003.347
917.307
831.268
745.228
659.189 Fig. 3.36. Temperature
573.150 + - - - . - - - - - . - - - . - - , - - - . . . - - - - r - - - - ' along L1 (a) before and (b)
o 0.576 1.153 1.730 2.307 2.884 3.461 after the shape optimiza-
Dist X 10-1 tion
573.15
- 670.605
768.059
865.514
_ 962.969
c::J 1060
c::::J 1158
1255
_ 1353
1450

Fig. 3.37. Temperature after the thickness optimization

applicable to a wide range of cases. In our following description we assume


that, for reasons of process technology, the entire current feed into the flange
is known and fixed.
The flange consists of three concentric rings around the heated pipe, each
with a specific function:
• exterior ring = "electrical contact ring": electrical contact and distributor,
• middle ring = "barrier ring": thermal barrier,
• interior ring = "distributor ring": pipe contact area and thermal distribu-
tor.
3.3 Shape Optimization of Flanges 231

1450.094,-----------------,
1362.400
1274.705 L1
1187.011
1099.316
1011.622
923.927
836.233
748.538
660.844
573.150 + - - . . , - - - - , - - - - - , - - , - - - - , - - - - . . , - - ' Fig. 3.38. Temperature along
o 0.576 1.153 1.7302.307 2.884 3.461 L1 after the thickness opti-
Dist x 10-1 mization
_ 0.500 E-03 a)
_ 0.0 2832
0.05614
_ 0.0 8396
0.11 1 78
- 0.1396
0.16742
c::::::J 0.19524
0.22306
- 0.25088

X 10- 1
2.508 . , - - - - - - - - - - - - - - - , b)
2.260 L1
2.011
1.762
1.514
1.265
1.017
0.768
0.519 / Fig. 3.39. (a) Thick-
0.271 / ness distribution in
0.022 +--...,...--"""'--.---.-----.---,.---.--' the optimum flange
o 0.576 1.153 1.730 2.307 2.884 3.461 and (b) thickness dis-
Dis! x 10 - 1 tribution along L1

The exterior ring feeds the current from one or more electrical contacts into
the flange and distributes it. The middle ring creates a Joule effect to such an
extent that the temperature in the ring becomes equal to the glass tempera-
ture. The interior ring provides the contact area with the pipe, and homog-
enizes current and temperature near to the pipe. The exterior and interior
ring are mostly "thick", because in this case the electrical and thermal resis-
tances are small, which has a positive effect on the distribution. The middle
ring is mostly "thin", because a smaller cross section generates a larger Joule
effect. For stability reasons, the sheet thickness of the middle ring cannot
be made arbitrarily small. Therefore, the entire cross section is reduced by
the insertion of slots, which force the current to flow through the remaining
232 3. Homogenizing and Conditioning

spokes. Figure 3.40 shows a typical situation for a flange with altogether four
spokes of the same size: electrical contact ring = blue, barrier ring = green,
distributor ring = red. In general, it is not necessary for the spokes to have
the same size or be distributed regularly.
The function of the exterior ring does not necessarily require high tem-
perature. By appropriate cooling, it is possible to avoid the use of platinum.
The consequence is an increased heat loss from the hot middle ring towards
the exterior ring, which must be compensated by a reduction of the cross
section. This leads to an increase of the entire electrical resistance. Here, one
must weigh up for each single case the costs for the platinum and for the
necessary electrical transformers.
The concept presented here obviously offers the following finite number
of degrees of freedom:
• the sheet thicknesses in the three rings (assumed to be homogeneous from
now on),
• the radii of the three rings,
• number, position, and size of the spokes in the barrier ring,
• number, position, and size of the electrical contacts at the electrical contact
ring.
These degrees of freedom can be adapted to the respective boundary
conditions and constraints in the best possible way, for example, with the
FEM software ANSYS and its included optimization algorithms. We have
first considered two-dimensional models where the pipe is only implicitly
accounted for by a thermal boundary condition. The electrical contacts are
also only implicitly taken into account by an appropriate boundary condition
in the model. At the respective points of the electrical contact ring, the given
entire current is fed into the ring homogeneous distribution. The objective
function F that ought to be minimized consists of two parts:

where and

Maximum and minimum must be computed over the interior boundary (Le.,
the contact line between flange and pipe), i denotes the current density. Fl
keeps the heat flow between flange and glass as low as possible, F2 makes
sure that the current is conducted into the pipe as uniformly as possible.
Actually, the second question we are interested in is whether the temperature

Fig. 3.40. 3-ring/spoke flange (2D vari-


ant 1: four equal spokes of 45 degrees)
3.3 Shape Optimization of Flanges 233

in the pipe is uniform. Because this information cannot be directly gained


from 2D models (where the pipe is not explicitly accounted for), a similar
substitute condition is formulated for the homogeneity of the current density.
The parameter T weights those parts of the entire objective function F that
are accounted for by the independent partial objectives defined by FI and F 2 .
In a 3D model, FI is evaluated at the contact area between flange and
pipe. In F 2 , i is substituted by T, and minimum and maximum are computed
over the entire interior side of the pipe. We have applied the optimization
method to the initial situation presented in Fig. 3.40, where only the two
sheet thicknesses of the barrier and distributor ring and the width of the
distributor ring are to be optimized (three degrees of freedom). Because this
is a concrete industrial problem, we will only present general results and
no details here (dimensions, current intensity, ... ). Figures 3.41 and 3.42
show the effect of the parameter T if T = 1, only the current homogeneity
is optimized, if T = 0, only the temperature is optimized. The figures show
relative oscillations, Isoll is the (known) current density in the case of optimum
homogeneity, Tsoll is the glass temperature (if the temperature of a narrow
concentric ring around the interior boundary equals the glass temperature,
there is no heat flow). The percentage deviations for current and temperature
are inevitably of different orders of magnitude, because a relative deviation
of 1% for the current has a different influence on the quality than the same
deviation for the temperature. The reference case is a flange without ring-
and-spoke structure (i.e., a simple circular ring) of a constant thickness and

°
with the same exterior cross section as the ring/spoke flange. In both figures,
a horizontal straight line of the value was the ideal case. The results show
that FI and F2 indeed follow opposite objectives.

0.005 .,---,;-----:--:---------,..,..----...._-----,

-0.005
~

(5
(1)1-0.D1 0

i
-0.015

-0.020

-0.10 -0.05 o 0.05 0.10 0.15


Coordinate along the tube boundary

Fig. 3.41. Oscillations of flange temperature at the interior boundary (2D)


234 3. Homogenizing and Conditioning

0.4
/.
0.2 I "
. ~ .
\

. ."'-../0'
·0.2 : i'~ O • 0. .

...\...../.. O< y < l


~ =1
'

·0.4 Reference case D

-0.15 -0.10 -0.05 o 0.05 0.10 015


Coordinate along the tube boundary

Fig. 3.42. Oscillations of flange current at the interior boundary (2D)

In the second step, in addition to the mentioned three degrees of freedom,


the size of the four spokes (i.e., the aperture angle) is also to be optimized
(altogether 3 + 2 = 5 degrees of freedom), and I = o. Figure 3.43 shows
the resulting flange, the spokes have become smaller and the distributor ring
wider. Figures 3.44 and 3.45 show the respective relative oscillations of tem-
perature and current. Here, the geometry resulting from the first step with
I = 0 is called flange variant 1. The temperature has become considerably
more homogeneous and, due to the larger distributor ring, there are also less
current oscillations than in variant 1 (although, due to I = 0, there was no
current optimization).
In the third step, I was varied with the same five degrees of freedom as in
step 2. Figures 3.46 and 3.47 show the result, the green curves corresponding
to the blue curves (flange variant 2) in Figs. 3.44 and 3.45. The variant
with I = 0.2 yields a rather good compromise between the two objectives of
achieving good temperature and good current distribution. The geometry of
this variant (i.e., above all the position and size of the spokes) lies between

-
flange variants 1 and 2.

293.15
451.142
609.135
767. 127
925.12
1083
1241
1399 Fig. 3.43. Flange vari-
1557 ant 2 (2D, with tempera-
1715 ture distribution)
3.3 Shape Optimization of Flanges 235

0.006 , . . - - - -- - - - - - - - - - - - - - - - ,

0.004
.-,, ,~
, ' .•
0.002

o
'" o
S l
-0.002
Flange variant 1 '
-0.004 Flange variant 2 x
"

-0.006 +-......f._,---_ _-,-_---..::¥-_ _.---_ _,..-.....L-I


-0.15 -0 1 -0.05 o 0.05 0.1 0 .15
Coordinate along the tube boundary

Fig, 3.44, Oscillations of flange temperature at the interior boundary (2D)

0.4
0.3
0.2 I .
"
~
/'.
\
0.1
~

0 0
'"I
S -0.1
-0.2
.
-0.3
Flange variant t • .,.,'-./"
-0.4 Flange vanant 2 x

-0.5
-0.15 -0.1 -0.05 o 0.05 0.1 0.15
Coordinate along the tube boundary

Fig. 3.45. Oscillations of flange current at the interior boundary (2D)

In the fourth step, the reference case (flange without rings and spokes,
homogeneous thickness) and the optimized flange variant 2 were modeled
three-dimensionally. Figures 3.48 and 3.49 show the resulting temperature
fields. For reasons of simplification, we assumed here a perfectly adiabatic
pipe isolation, i.e., the input of heat energy into the pipe exclusively occurs
through the glass and the Joule effect, and a heat loss of the pipe only re-
sults from the heat conduction through the flange. Of course, this is not a
realistic model yet, but it served our purpose of assessing the quality of the
optimization. The three-dimensional feeding of the current into the flange
236 3. Homogenizing and Conditioning

0.005

-0.005
~
-0.D1 0

.(\..
(5
<Il
I -0.015
~ i ~.
-0.020
:i \ ,.
-0.025

-0.030

-0.035
-0.15 -0.10 -0.05 o 0.05 0.1 0 0.1 5
Coordinate along the tube boundary

Fig. 3.46_ Oscillations of flange temperature at the interior boundary (2D)

0.20
"-"-" ,= 0
0.15 y= 0.05 .
i' = 0.1
0.10 i'=0.2 a
=0.5 •
~
0.05
(5
<Il
I 0
S
-0.05

-0.10

-0.15
-0.15 -0.10 -0.05 o 0.05 0.10 0.1 5
Coordinate along the tube boundary

Fig. 3.47. Oscillations of flange current at the interior boundary (2D)

is not shown. The pipe temperatures in the optimized ring/spoke flange are
obviously more homogeneous than in the reference flange.
rt has probably become clear that there is more than one useful defini-
tion of the objective function to be minimized. Depending on the situation,
we can apply variants of the definition described above. A third term, for
example, can minimize the necessary amount of platinum. Another idea is
to introduce a further objective function in order to keep the temperature
in the flange from exceeding a critical value, which is obviously important
for the operating safety of the flange. Of course, the weighted application of
more than one objective function always implies that the different objectives
-
References 237

316.946
1680
1690

-
1700
1710
1720
c::J 1730
c::J 1740
1750

--
Fig. 3.48. Tube/flange temperature in the reference case (3D)

291 .099
1680
1690
1700
1710
1720
c ::J 1730
c::J 1740
1750

Fig. 3.49. Thbe/flange temperature for flange variant 2 (3D)

usually compete with each other. Thus, the result is a compromise between
the intended single objectives.

References
3.1 J.M. Ottino: The Kinematics of Mixing: Stretching, Chaos, and Transport
(Cambridge University Press, Cambridge 1989)
3.2 E.J. Hinch: Mixing: Turbulence and Chaos - an Introduction (Kluwer Aca-
demic, Plenum, New York 1999)
3.3 J.W. Hiby: "Definition und Messung der Mischgiite in fliissigen Gemischen",
Chem.-Ing.-Tech. 51(7) , 704- 709 (1979)
3.4 M.H. Pahl (Ed.): Mischen und Riihren. Grundlagen und moderne Verfahren
fiir die Praxis, VDI- GVC conference Baden- Baden , 23.-24.11.1998 (VDI ,
Paderborn 1998)
3.5 P. Davies: Prinzip Chaos, 4th edn. (Goldmann, Miinchen 1993)
3.6 M. Cable: "The principles of homogenizing of glass melts" , Glastechn.
Ber. 63K , 308- 322 (1990)
3.7 F .F. Riehl: Beitrag zur Berechnung der Homogenisierungs- und Liiutervor-
giinge in Glasschmelzwannen, PhD Thesis (RWTH, Aachen 1976)
3.8 H . Aref, M.S. El Naschie (Eds): Chaos Applied to Fluid Mixing (Pergamon,
Oxford 1995)
3.9 O. Pironneau: Optimal Shape Design for Elliptic Systems (Springer, Berlin,
Heidelberg 1983)
4. Shaping at Low Viscosities

4.1 Heat Transfer Between Glass and Mold


During Hot Forming

Matthias Brinkmann, Norbert Siedow


Introduction

The technological control of all subprocesses in glass making (melting, hot


forming, annealing) requires a correct description of the heat transport within
the glass and the heat transfer between glass and vessels (crucibles, tanks),
molds, or cooling equipment (lehrs, vans). Especially in hot forming, the heat
transfer from the glass into the molds is of decisive importance. The heat
balance of the glass-mold system has considerable influence on the process
control and, finally, on the product quality.
A precise numerical simulation of the heat transfer during hot forming
requires - as input parameters - the thermal conductivity of the glass and
the mold material, and the coefficient of heat transfer between glass and
mold. The present contribution reports on a theoretical and experimental
investigation on the heat transfer between glass and mold during hot forming.

Phononic

Opaque

Semi-transparent

Transparent

Gas film Cover

Fig. 4.1. Heat transport in hot glass and heat transfer between the glass and the
mold; dashed arrows = phononic heat conduction, continuous arrows = radiative
heat conduction

H. Loch et al. (eds.), Mathematical Simulation in Glass Technology


© Springer-Verlag Berlin Heidelberg 2002
240 4. Shaping at Low Viscosities

Figure 4.1 illustrates the complex heat transport behavior in the glass-
mold system. Hot glass gets into contact with a much colder mold. In the
mold as well as in the glass, heat can be transported by phononic heat conduc-
tion (dashed arrows). Within the glass, heat can also propagate by radiation
(at wavelengths in the semi-transparent and transparent range) and interact
with the surface of the mold (continuous arrows). The mold may be covered
with an oxide layer, a coating, or any kind of lubricant, which the heat must
"penetrate" to reach the mold [4.1]. Additionally, a gas layer (e.g., air) can
develop between the glass surface and the cover, in particular when the glass
cools down and "breaks loose" from the mold because of the different coeffi-
cients of thermal expansion [4.2-4]. In addition, the glass and mold surfaces
are not ideally smooth, they are rough. Therefore, the thicknesses of the inter-
face layers (cover and gas film) may also vary laterally. If, due to the surface
properties and the temperature of the mold, the glass does not "break loose" ,
it will "stick" to the mold at the end of the forming process. This adherence
(above the so-called sticking temperature), which must absolutely be avoided
in hot forming, and is extensively discussed in the recent literature [4.5,6].
Of course, all these effects have a tremendous influence on the heat transfer
between the glass and the mold and its time dependence during hot forming.
In contrast to opaque materials, heat can be transported in semi-transpar-
ent glass by (phononic) vibrational heat conduction as well as via radiation
(particularly at high temperatures) [4.7-10]. Apart from pure phononic heat
conduction (dashed arrows in Fig. 4.1), a heat radiation field exists inside
the glass that contributes to the heat transfer between the glass and the
mold [4.10,11]. Figure 4.2 shows the internal optical transmittance of a typ-
ical Schott glass. This example allows for the division of the radiative field
into the opaque, semi-transparent, and transparent wavelength range. Re-

~ 100
Thickness
/\\1
Q)
0
c = 1 mm
<II 80
::::
'E / !
Cfl
c
§
60 Planck (1200 'c) I
"iii 40
0
E. i I
0
"iii 20 ! .................................
c
CD
C 0
U i
./ Transparent Semi-transparent Opaque
0 2 5 6 7

Fig. 4.2. Internal optical transmittance of a I-mm-thick glass pane in the wave-
length ranges: UV (opaque), transparent, semi-transparent, and opaque. The dashed
curve displays Planck's function at 1200°C
4.1 Heat Transfer Between Glass and Mold During Hot Forming 241

garding heat transport by thermal radiation, the UV range may be neglected


at temperatures below 1200°C (Planck's function in Fig. 4.2).
In the wavelength range .\ ;::: 5 !-Lm, the glass is opaque, that is, no heat
can be transported by radiation within the glass. Only the surfaces of the
glass and the mold can exchange heat via radiation (Fig. 4.1).
In the semi-transparent wavelength range (2.5-5 !-Lm), heat can be trans-
ported by radiation within the glass (even over large distances). If this radia-
tion escapes from the glass and reaches the mold surface, it is either absorbed
or reflected. By absorbing this radiation, the mold surface removes heat from
the glass which is transferred into the mold. Radiation emitted from the mold
surface can also penetrate into the glass and be absorbed there.
In the transparent wavelength range (0.5-2.5 !-Lm), the radiative field inter-
acts only weakly with the glass. Therefore, hardly any heat transfer between
the glass and the mold takes place. Radiation is mainly exchanged between
opposite surfaces of the mold. This process is called radiative heat transition
through glass.
Aside from phononic heat conduction and heat transport by radiation,
heat is also transferred by convection in molten glass. At typical processing
temperatures, this contribution is already very large [4.8]. Nevertheless, we
will not consider heat transfer by convection here, because this process de-
pends very much on the details of the forming process and must be treated in
numerical-simulation experiments for each individual case. However, three-
dimensional computations require - as input parameters - the thermal con-
ductivity and the coefficient of heat transfer, which should be experimen-
tally determined undisturbed by convective influence. Without convection,
the mathematics used in the following is considerably simplified. However,
when the coefficient of heat transfer is determined experimentally, it must
be ensured that convective flows of the glass do not distort the results (see
Sect. 4.1.5).
Because we intend to describe the basics of the heat transfer between the
glass and the mold, the following discussion is restricted to "one-dimensional"
geometries (for example, the process of pressing flat glass products) where
the heat transfer basically takes place perpendicularly to the glass-mold in-
terface. In this case, the necessary numerical computations of heat transport
can easily be carried out (see Sect. 4.1.2).

4.1.1 Heat Transfer Coefficient Between Glass and Mold

The two different, but common, definitions of the parameter "heat transfer
coefficient" 0: for the contact between a fluid and the surface of a solid border
are illustrated in Fig. 4.3. In both cases, a cooled metal or molding plate at
"low" temperature comes into contact with a (semi-infinite) fluid bath at
"higher temperature". A heat flow q between the fluid and the mold arises
instantaneously. The red lines in Fig. 4.3 show the steady-state temperature
profiles for the contact between a fluid of high temperature (which is infinitely
242 4. Shaping at Low Viscosities

Fluid Glass
a) b)

Boundary layer Contact layer

Fig. 4.3. Two different definitions of the heat transfer coefficient: (a) definition in
heat engineering, (b) definition for hot forming in the glass industry

extended in one direction) and a cooled solid plate of low temperature and
small thickness D. The outermost layer of the fluid is called the boundary
layer (Fig. 4.3a). The outermost surface layer of the solid plate (cover and/or
gas layer) is the contact layer (Fig. 4.3b).
In heat engineering and chemical-plant technology, the definition in
Fig. 4.3a is used (e.g., to design heat exchangers):

(4.1)

where 0: is given in W /K m 2 . In this case, the difference between the in-


ner temperature Tfl of the hot fluid and the average temperature T S of the
metal sheet is used. This definition is based on the so-called boundary layer
model [4.12]. In this model, the complete temperature difference (TfI-T") de-
creases within a thin fluid boundary layer (Prandtl boundary layer). Because
of its contact to the metal surface this layer is free of any convection [4.13].
Apart from the properties of the contact layer from Fig. 4.1 (not drawn in
Fig. 4.3a), the heat transfer coefficient strongly depends on the fluid dynam-
ics and the geometry of the heat exchanger, which have a huge impact on the
fluid boundary layer [4.14].
For hot forming in the glass industry, however, it makes sense to use the
definition according to Fig. 4.3b ([4.15-17]):

(4.2)

Here, the temperature difference between the glass and the mold surface
temperature is chosen (instead of the "inner temperatures") for the definition
of the heat transfer coefficient. In contrast to the boundary layer in Fig. 4.3a,
the contact layer does not consist of a liquid phase (in this case, glass),
but of the combination of the air/gas layer and the surface cover described
in Fig. 4.1. This second definition of the heat transfer coefficient for glass
making according to (4.2) can be justified by the following considerations:
4.1 Heat Transfer Between Glass and Mold During Hot Forming 243

• Compared to heat engineering, the temperature differences between the


glass and the mold are very large in hot forming (Tg=1000-1300 °C,
Till=300-600°C, i.e., Tg - Till=400-1000K).
• Due to the large temperature differences Tg - Till, the thermodynamic and
thermo mechanical properties of the glass change greatly in the boundary
layer. In particular, a very strong increase of the glass viscosity occurs
near the mold. This change suppresses convection in the glass and leads
to considerably thicker boundary layers (>1 mm), so that the definition of
the heat transfer coefficient according to (4.1) and Fig. 4.3a is no longer
valid.
• Typically, the final glass thickness is small (1-50 mm) in hot forming. If the
boundary layer happens to be very thick, the concept of the heat transfer
coefficient definition according to (4.1) and Fig. 4.3a breaks down. In an
extreme case (final glass thickness 1 mm), the entire glass pane only consists
of boundary layers.
Thus, in hot forming we do not have the situation illustrated in Fig. 4.3a,
with an extremely sharp temperature drop in the Prandtl boundary layer
(order of magnitude 100 K per 0.1 mm), but deal with the situation illustrated
in Fig. 4.3b, where the temperature gradient at the glass surface is smaller
(order of magnitude 100 K per 1 mm), followed by a further temperature drop
in the contact layer, which, for simplicity, comprises the cover and the air or
gas layer (Fig. 4.1). If a finite thickness Dcontact and a thermal conductivity
kcontact are assigned to the contact layer, (4.2) can be regarded as Fourier's
law:
TgO _ T illO
q = kcontact D (4.3)
contact

where the thickness and the thermal conductivity of the contact layer are
combined into the heat transfer coefficient
kcontact
a= (4.4)
Dcontact

The heat transfer coefficient a defined by (4.1) and Fig. 4.3a thus describes
the thermophysical processes in the boundary layer (i.e., the outermost fluid
layer), whereas the definition of (4.2) and Fig. 4.3b is appropriate to describe
the thermal behavior of the contact layer (i.e., the cover and the air/gas
layer).
In this context, the special case of an ideal heat contact between glass and
mold is reviewed. Here, the thermal resistance of the contact layer vanishes
aideal --+ 00. If, additionally, the radiative contribution to the heat transport
and to the heat transfer between the glass and the mold is neglected, we can
determine the time-dependent behavior of the temperature profile in the glass
Tg(x, t) analytically. For the case of a half-infinitely extended (x < 0) one-
dimensional glass layer of constant temperature, which at t = 0 comes into
contact with a half-infinitely extended (x> 0) molding plate, we get [4.18]
244 4. Shaping at Low Viscosities

Tg ( x, t) = Tcontact + (Tg O - Tcontact) erf ( ( 4.5)

(erf: error function). For the mold (x> 0), the temperature profile Tm(x, t)
IS

Tm(x, t} ~ T mO + (1'on',,' - 7""0) eefe ( (4.6)

(erfc: complementary error function). Tcontact denotes the time-independent


contact temperature of the glass~mold interface (x = 0):

(4.7)

The thermophysical material parameters (thermal conductivity kg,m, heat


capacity cg,m, and density Pg,m), which are assumed to be not temperature
dependent here, will be introduced more precisely in the next section. Fig-
ure 4.4 shows an example (Tg O = 1000°C, T mO = 500°C) for the time-
dependent behavior of the temperature profiles Tg,m(x, t) in the glass and in
the mold in the case of an ideal thermal contact.
In the following section, the physics and mathematics of the heat transfer
between the glass and the mold will be presented, additionally accounting for

1000
T lOG
800

0.10

Fig. 4.4. Time-dependent behavior of the temperature profiles in the glass Tg(x <
0, t) and in the mold Tm(x > 0, t) for the special case of ideal thermal contact
O!ideal ---+ 00 without (radiative contribution to heat transport and heat transfer
neglected)
4.1 Heat Transfer Between Glass and Mold During Hot Forming 245

the heat transfer by radiation within the glass and the non-ideal heat transfer
between the glass and the mold. The glass-mold heat transfer coefficient
depends on many intrinsic properties of the contact layer, for example: the
thickness and thermal conductivity of the cover and the air/gas layer (see
(4.4)), the surface roughness, the temperature of the mold and of the glass,
and the pressing pressure.
Besides, a decreases during the process of hot forming from large values
(~20000W/Km2, almost "ideal" heat transfer) by more than one order of
magnitude (~ lOOOW/Km2) [4.15]. This drastic change can occur within
a few seconds. A sound physical model, which can explain this behavior
quantitatively, does not yet exist. In the following, the heat transfer coeffi-
cient will therefore simply be assumed to be a function of time, a(t). This
phenomenological approach is already a preparation for the solution of the
inverse problem described in Sect. 4.l.5 (computation of the time-dependent
behavior of a using experimental data).
Although only the definition according to Fig. 4.3b and (4.2) will be used
in the following for the glass-mold system, the first definition according to
Fig. 4.3a and (4.1) also becomes important when heat transfer by radiation
is considered (discussed in Sect. 4.l.4).

4.1.2 Physics and Mathematics of the Heat Transfer

In this section, the exact one-dimensional transport problem of the heat


transfer between the glass and the mold will be described and solved nu-
merically [4.19].
The model discussed in the following is illustrated in Fig. 4.5. An (in-
finitely extended) glass pane of thickness l is in thermal contact with a mold-
ing plate of thickness L -l. Let this pile of glass and mold be placed between
two "domains" UI and U2 with given constant temperatures T a1 ,2' Heat may
also be exchanged between these domains and the pile.
According to Fig. 4.1, the heat transport inside the glass is a combined
effect of phononic and radiative heat transport. The heat transfer glass-mold

o Glass Mold L
x
I I I
U1 U2
TD(x) T"l(x)
CgPg kg CmPm

~)"a km

~gr a"" a mr Fig. 4.5. Illustration of the one-dimensional


domain geometry for the heat transfer be-
n. "a tween the glass and the mold. The distances
Ta2
between the zones UI-glass and glass-mold
represent the air/gas layers, which are addi-
tional optical obstacles to the radiative heat
transfer
246 4. Shaping at Low Viscosities

and glass- VI can also be divided into a pure "contact" contribution, which is
described by "contact" heat transfer coefficients agr and a gm , and a radiative
portion due to interaction between the heat radiation from the interior of the
glass and the mold surface. For this combined transport problem, the time-
dependent behavior of the one-dimensional temperature profiles Tg,m(x) in
the glass and in the mold will be determined numerically.
Let the following thermophysical material parameters for the two mate-
rials (glass and mold) be known:
• the temperature dependent heat capacities cg,m(T),
• the temperature-dependent densities Pg,m(T),
• the temperature-dependent phononic heat conductivities kg,m(T).
Convection inside the glass and thermal expansion of the two materials will
not be taken into account.
At the interfaces VI- glass, glass- mold, and mold-V2, the time-dependent
"contact" heat transfer coefficients agr, a gm , and a mr will be used. The glass,
the mold surface, and the domain VI have the following optical properties:
• wavelength-dependent absorption coefficient ,,;(,\) of the glass,
• wavelength-independent (simplification!) refractive index ng of the glass,
• wavelength-dependent emissivities f2 ,1 (,\) and specular reflectivities r2,1 (,\)
of the surfaces of the mold and VI.
Figure 4.6 shows the definition of the optical surface properties applied
here. When radiation of wavelength ,\ and intensity I impinges on the surface
with an incident angle e,
a portion E('>") is absorbed. The remainder, (1 -
E('>")) , is reflected. The portion (1 - E('>"))r('>") undergoes specular reflection,
and (1 - f('>"))(1 - r('>")) is isotropically diffuse-reflected [4.20]. Anisotropic
diffuse reflection will not be considered. An ideal black-body radiator is thus
characterized by E = 1 (r arbitrary), a perfectly reflecting mirror by E = 0,
r = 1, and an ideally diffuse mirror by E = 0, r = 0.
In Fig. 4.7, some of the important radiative transfer processes are illus-
trated. In this context, it is important to account for the additional air/gas

10(A,£I) T

Fig. 4.6. Illustration of the optical surface prop-


erties. The incident radiation 1(>", B) can either
( 1-~(A))·r(A) be absorbed (dashed arrow) or undergo specular
(arrow below) or isotropic diffuse (semicircle) re-
flection. The corresponding factors represent the
relative contributions
4.1 Heat Transfer Between Glass and Mold During Hot Forming 247

Glass Mold

Glass Mold

Fig. 4.7. Illustration of the heat transfer by radiation between glass and mold.
Several processes are shown as examples. I-l and {t are the directional cosines in the
glass and in the air/gas layer, respectively (right drawing)

layer (Fig. 4.1 and Fig. 4.7) between the surfaces of the glass and the mold
with optical refractive index na. Of course, the presence of a (non-absorbing)
air layer of finite thickness (larger than the "interesting" wavelengths >. of
the heat radiation) is a further barrier for the radiative heat transfer. In this
case, Fresnel reflection occurs at the glass-air boundary layer obeying

(4.8)

with

ng sin( ¢) = na sin( 8) (4.9)

(law of refraction). Thus, up to 40% of the heat radiation from the interior
of the glass may be reflected back. In the following, the system of differential
equations for the heat transport problem from Fig. 4.5 will be presented.
The temperature profile in the interior of the glass or metal is described
by the heat transfer equation:

in the glass CgPg


aTg = ax
at a (aTg )
kg ax - qr ,0::::: x : : : l , (4.10)

in the metal CmPm


aT
8t
m
=
a ( k aTffi)
ax ax ,l : : : x : : : L ,
m (4.11)

where qr is the radiative flux, which is obtained from the spectral radiation
intensity 1(>', x, /-1) according to
248 4. Shaping at Low Viscosities

J
+1
qr(X) = 21f L Ik(x, J1)J1dJ1 . (4.12)
k -1

Here, the division of the intensity spectrum into bands of wavelengths


Pk, >'k+1}:

J
)..k+l

Ik(x, J1) = 1(>', x, J1)d>. and


)..k

has already been made. J1 = cos( ¢) describes the polar angle of the radiation
direction n inside the glass (Fig. 4.7). Besides, note that all wavelength-
dependent optical properties of the material are given as functions of the
wavelength in air 5.. The radiative transfer equation in glass (0 < x < l)
reads

(4.13)

where B is Planck's function


2hc 2 1
B(>.,T) = ~ (~) -1 (4.14)
exp )"kBT

(Planck constant: h = 6.626 X 10- 34 J s, vacuum light speed: c = 2.998 x


108 m/s, Boltzmann constant: kB = 1.381 x 1O- 23 J/K). However, the differ-
ence between Planck's function Bg in glass and Planck's function Ba in the
air layer must be taken into account, namely
n~Bg(>', T) = n~Ba(\, T) ,
na>' = n g >. ,

J Bg(>', T)d>. = n~ J Ba(5., T)d5.


CX) CX)

n~ ,
o 0

n~B;(T) = n~B~(T) (4.15)

(>.: wavelength in the glass, 5.: wavelength in the air/gas layer ~ vacuum
wavelength).
The boundary conditions of the heat transfer equation describe the heat
transfer between the surface of the glass and the domain VI and between the
surface of the mold and the domain V2:
boundary of the glass towards VI (x = 0):

opaque
4.1 Heat Transfer Between Glass and Mold During Hot Forming 249

boundary of the mold towards U2 (x = L):


uTm
-kID ux = CYmr(T m - Ta2 ) . (4.17)

The factor l' is a measure for the "emissivity current" of the glass surface in
the opaque wavelength range:

l' = 27f J
iL>O
(1 - P (f-L(Jt))) JtdJt , (4.18)

where Jt = cos(8) is (analogously to f-L) the directional cosine of 8 in the air


layer (Fig. 4.7). We have
n~ (1 - Jt2) = n~ (1 - f-L2)
n~JtdJt = n~f-Ldf-L . (4.19)
For the radiative heat transport in the glass, we use the boundary conditions
described in Fig. 4.6, which take into account emission as well as specular
and diffuse reflection:
boundary of the glass towards U1 (x = 0):
Jk(O, f-L) = p(f-L)Jk(O, -f-L) + r~(1- Ef)(l - p(f-L))2 Jk(O, -f-L)
+(1- rf)(l- E~)(l- p(f-L)) J (1- p(f-L(Jt)))Jk(O,f-L(Jt))dJt (4.20)
iL<O
+E~(l - p(f-L))B:(Tar) ;
boundary of the glass towards mold (x = l):
Jk(l, -f-L) = p( -f-L)Jk(l, f-L) + r~(1 - E~)(l - p( _f-L))2 Jk(l, f-L)
+(1- r~)(1- E~)(1 - p(-f-L)) J (1- p(f-L(Jt)))Jk(I,f-L(Jt))dJt (4.21)
iL>o
+E~(1- p( -f-L))B:(TID(l, t)) .
The right-hand sides give the contributions to the radiation intensity (at
the boundaries of the glass) by reflection at the glass surface (first term),
by specular reflection at the mold (U1) surface (second term), by diffuse
reflection at the mold (U1) surface (third term), and by emission from the
mold (U1) surface (fourth term).

°
Aside from the boundary conditions of the heat conduction equation for
x = and L, so-called interface conditions must also be imposed at the glass-
mold contact surface x = l. For a complete mathematical description of the
problem, two conditions are necessary:
for the glass (x = l-):
-kg afa; = CYgm(Tg(l, t) - Trn(l, t))
+1' J E2(5.) (Ba(5., Tg(l, t)) - Ba(5., TID(l, t))) d5. , (4.22)
opaque
250 4. Shaping at Low Viscosities

for the mold (x = l+):


~km 81x = = Q;gm(Tg(l, t) ~ Tm(l, t))
+i J E2(~) (Ba(~, Tg(l, t)) ~ Ba(~, Tm(l, t))) d~ (4.23)
opaque

+27r J (1 ~ P(fL)) "£ E~Ik(l, fL))fLdfL


{t>O k

J
>-(tr + semi-tr)

With the last two terms in (4.23), the energy exchange between the glass
and the mold via radiation in the transparent and semi-transparent range
(= limits of integration) is taken into account. Here, the mold surface interacts
with the interior of the glass via radiation, without any heat being absorbed
or emitted by the glass surface. A portion of the radiation

J
{t>O
(1 ~ P(fL)) LE~Ik(l'fL))fLdfL27r
k

from the transparent and semi-transparent wavelength ranges leaves the in-
terior of the glass and is absorbed by the mold. On the other hand, the part

J
>-(tr + semi-tr)

is emitted by the mold and absorbed inside the glass. The difference between
these two parts of the radiative field is the net flow that is absorbed by
the mold via radiation in the transparent and semi-transparent ranges. This
radiative contribution to the heat transfer occurs only for semi-transparent
materials, such as glass.
In the boundary and interface conditions (4.16)-(4.23), the contribution
of multiple reflection in the air/gas layer to the radiative field has been ne-
glected. The reason for this is that the formulae describing the boundary
conditions would "explode", if diffuse radiation and multiple reflection were
taken into account. In consequence, a "radiation leakage" occurs, through
which multiply reflected radiation can escape from the interfaces Ul-glass
and glass-mold. This is unphysical. However, the corresponding losses are
small. In the case ng = na (no air layer), there is no reflection at the surface
of the glass, and therefore no multiple reflection and no radiative leak.
In order to solve the one-dimensional transport problem (4.10)-(4.23),
a numerical algorithm has been developed that is able to compute the
time-dependent behavior of the temperature profiles in the glass and in the
mold. Therefore, a semi-implicit finite volume method is used. The intervals
0= Xo < x < Xn = l, l = Xn < X < XN = L are divided into elementary cells

(xo, Xl/2), (Xl/2' X3/2), ... , (Xn-l/2, x n ), (xn' Xn+l/2), ... , (XN-l/2, XN)
4.1 Heat Transfer Between Glass and Mold During Hot Forming 251

with the nodes

Analogously, only discrete times to = 0, t 1 , t 2 , ... are considered. At time tj,


the corresponding heat equation (4.10) for the glass or (4.11) for the mold is
integrated over the ith elementary cell, and the first derivatives are approxi-
mated by differences. The partial differential equations are thus transformed
into linear algebraic equations. The radiative flux of (4.12) is computed from
the results of the previous time interval tj -1. For the radiative transport
equation, a formal analytic solution exists [4.10]:

+; J x

Bk(T(s, tj_de-I<k(x-s)/f'd/k, for /k > 0,


o
(4.24)
Ik-(x,/k) = Ik-(l,/k)e-I<dx-l)/f'

-; J I

Bk(T(s, tj_d)e-I<k(x-s)/f'd/k, for /k <0.


x

The spatial integrals are evaluated numerically if the temperature profile


T(Xi' tj-d is known. Because of the specular and the diffusive boundary
conditions, (4.20) and (4.21), the two terms in (4.24) are coupled for all
directions /k at the boundary. Ik+(O, /k) and Ik-(l, /k) in (4.24) are determined
by setting up a system of linear equations and solving it, for example, by
means of the Gaussian algorithm. After determining the radiation intensity
and the radiative flux, a tridiagonal system of equations is obtained that
allows an efficient computation of the temperature at the nodes Xi for each
time tj by a shortened Gaussian algorithm.
In the following, two special cases of this model are presented that allow
the computation of real one-dimensional problems in hot forming:
• For £1 = 0, 1'1 = I, and a gr = 0, we have the situation of casting hot glass
between two symmetric molding plates, because this selection of the heat
transfer parameters generates adiabatic boundary conditions with respect
to the heat transfer towards U1. In consequence, the domain (0 < X < L)
is virtually mirrored at the plane x = O. The glass thickness D = l is thus
doubled. In the next section, this situation is demonstrated by a few model
computations .
• If we select a mr = 0, the system is physically mirrored at the plane x = L.
Now, a molding plate of thickness 2(L - l) is described that is immersed
in the middle of a glass melt of thickness 2(L -l) and thus heated. This is
the situation of the immersion experiment described in Sect. 4.1.5.
Concluding this section, we want to emphasize once more that this section
differs from the previously published literature in so far as the contributions
252 4. Shaping at Low Viscosities

to the heat transfer between the glass and the mold through material contact
as well as through radiation are taken into account.

4.1.3 Sample Computations

The numerical algorithm presented in the previous section will now be used
to study the influence of the contact and radiative contributions to the
heat transfer between the glass and the mold with an example. Let a 5-
cm-thick glass plate (Fig. 4.2) of temperature 1500 ac (homogeneous) come
(at time t = 0) into contact with a molding plate of the same thickness and
temperature 500 ac. The two plates are assumed to be insulated thermally
(O:gr,mr = 0) and optically (El = 0, Tl = 0) from the domains U1 and U2.
Physically, this case corresponds to the symmetric casting of a glass plate
of double thickness (2l = 10 cm) between two 5-cm-thick molds that are
insulated towards the outside. We take advantage of the symmetry of this
problem with respect to the mirror plane in the middle of the glass. The
heat balance between the glass and the molding plate is simulated over a
time interval of 30 s. Figures 4.8 and 4.9 show the results of three separate
simulation computations, which differ only with regard to the kind of heat
transfer between the glass and the mold. In the red case, an (almost) ideal
thermal contact (O:gm(t) = 106 W/Km 2 ) was used in conjunction with an
ideally reflecting mold surface (E2(A) = 0, T2(A) = 1). In contrast, in the
blue and the green cases a completely insulating contact heat transfer was
selected (O:gm (t) = 0), combined with ideal absorption at the mold surface
(E2(A) = 1). The blue and the green cases differ in so far as in the blue case
the air layer between the glass and the mold is missing (n a = ng).
Figure 4.8a shows the temperature profiles in the glass and the mold
after 30 s. Whereas in the red case the profiles are continuous at the glass-
mold interface (ideal thermal contact), temperature differences of 200K and
500 K, respectively, are obtained in the blue and the green cases. Here, contact
heat transfer is suppressed, and heat exchange occurs only via radiation. If,
however, the temperatures in the interior of the glass are compared, the
temperature in the blue case is clearly below that of the red case (100 K).
This is possible only because in the blue case heat from the interior of the glass
can leave the glass directly via radiation (in the semi-transparent wavelength
range, Fig. 4.2) and is then absorbed by the ideal black mold surface. In the
red case, this heat radiation is reflected at the perfectly reflecting surface and
returned into the glass. Therefore, heat transfer cannot take place through
this transport channel. In the green case, heat radiation from the glass can
also be absorbed by the mold surface. However, a part (~ 40%) is reflected
back into the glass at the glass-air/gas interface. This influence is clearly
discernible (in comparison with the blue curve).
Figure 4.8b shows the heat-flux profiles in the glass and the mold, also
after 30 s. The phononic contribution to the heat flux in the glass is rep-
resented by solid lines, whereas dashed lines correspond to radiative flow.
4.1 Heat Transfer Between Glass and Mold During Hot Forming 253

1500 a)

1300
()
!l..
Q)
::; 1100

Q)
a. 900
E
~
700

500
0 0.02 0.04 0.06 0.08 0.10
x/m

5 x 105 b)

4 x 105

'"E
I 3 x 105
Total flow
~ 2 x 105
Radiative flow
0- Phononic heat flux

1 x 105

0.02 0.04 0.06 0.08 0.10


x/m

Fig. 4.8. (a) Temperature profile and (b) heat flux profile of the example under
consideration at t = 30 s for the three simulation computations with the following
contact conditions: red: agm(t) = 10 6 W/Km2, E2(A) = 0, r2(A) = 1; blue: agrn(t) =
0, E2(A) = 1, r2(A) = 1, na = ng = 1.5; green: agm(t) = 0, E2(A) = 1, r2(A) = 1,
na = 1.0

The total heat flow in the glass and the mold is represented by the dotted
lines. It must be continuous at the interface air-mold. In the red case, the
radiative flow in the glass drops to zero (dashed red curve) at the interface
(glass-mold) because of the ideally reflecting mold surface. In the blue and
the green cases the radiative contribution represents the total heat flux at
the glass-air interface.
Figure 4.9a displays the time-dependent behavior of the temperature on
the left (x = 0) and the right (x = 0.0499) glass surface, and on the left
(x = 0.0501) and the right (x = 0.1) mold boundary. In the red case, the
surface temperatures of the glass and the mold coincide after a very short
time period (ideal thermal contact), whereas in the blue and the green case
(pure radiative heat exchange) the surface temperatures approach each other
very slowly.
Figure 4.9b describes the time dependence of the heat content inside the
glass and the molding plate (reference temperature 500°C). The total heat
254 4. Shaping at Low Viscosities

1600
x - O a)
1400 1\
.....................
~ 1200 .. ..... - .... - ...
Q)
x = 0.
0499
5
"§ 1000
Q)

....--:::=:
a.
E 800 x - 0.501

c:=:-
~
600 x =0.1
'--
400
0 5 10 15 20 25 30
tis

Total b)
~_~"io.""''''''''''_'"C'IO

10 8 Glass
'"l'
E
~
E
Q)

~ 107
u
iii
Q)
I

106
0 5 10 15 20 25 30
tis
Fig. 4.9. (a) Time dependence of the temperature for the two glass surfaces (x = 0
(towards VI) and x = 0.0499 (towards mold)) and the two mold surfaces (x =
0.0501 (towards glass) and x = 0.1 (towards V2)). (b) Time dependence of the heat
content (logarithmic graph) of the glass, the mold, and the sum of both (black)

content of the system, which must be constant because of the thermal and
optical insulation from the domains Ul and U2, is also shown. A comparison
between the different heat contents inside the glass plate shows that in the
blue case, where heat can reach the mold via radiation in the best way, we
have the strongest decrease.
From this example, we can therefore conclude that the heat transfer be-
tween the glass and the mold via radiation may not be neglected in com-
parison with the pure contact heat transfer. From the development of the
surface temperatures (Fig. 4.9a) it becomes obvious that, in the beginning
(t = 0 ... 2 s), heat is transferred into the mold via thermal contact Cl: gm . How-
ever, when a temperature gradient is established at the glass surface, purely
phononic contact heat transfer is reduced because of the slow heat trans-
port from the interior of the glass towards the surface. Now, the radiative
transport channel (in the semi-transparent wavelength range) becomes more
important, because it enables a very efficient heat transfer from the interior of
4.1 Heat Transfer Between Glass and Mold During Hot Forming 255

the glass to the mold surface (blue case). The two transport channels are once
more illustrated in Fig. 4.10. In the first few seconds after the contact has
been established, purely phononic transfer is dominating (Fig. 4.10 above).
Then radiation from the interior of the glass reaches the mold surface, where
it is absorbed and conducted by phonons into the mold (Fig. 4.10 below).
Note that this kind of combined phononic and radiative heat transfer can
occur only in semi-transparent materials such as glass.

4.1.4 Radiative Contributions to the Heat Transfer

In Sects. 4.1.2 and 4.1.3, the exact heat transport problem (contact and ra-
diative contribution) has been treated in one dimension for the glass-mold
system (Fig. 4.5). This section deals with an approximation that allows the
computation of the "static" radiative contribution to the coefficient of heat
transfer arad. Let us once more consider the situation of Fig. 4.7. Heat radi-
ation is exchanged between the surfaces of the glass and the mold. Formally,
heat transfer via radiation between two parallel surfaces (1 and 2) with
wavelength-, temperature-, and angle-dependent emissivities £1,2(>', T 1 ,2, e)
can be described by the following coefficient of radiative heat transfer:

J
(X)

arad = arad(>')d>' (4.25)


o

(4.26)
x fl(A,Tl,8)f2(A,T2,8) de.
£1 (>', T 1 , 8) + £2(>', T 2) e) - £1 (A, T 1 , e) E2(>', T 2, e)

arad(>') is the wavelength-dependent coefficient of radiative heat transfer.


Therefore, the corresponding heat transfer coefficient between hot glass and
a mold can be obtained from (4.26) by replacing the emissivity El (>', T 1 , 8)
with the corresponding emissivity of a hot glass plate. In contrast to opaque

... ......
o Glass Mold L x
I I I ~

U1 U2
PhalIOIIic Phononic
.-: : ~
t = 0... 2 s

RadIIIIMt
..
rI>2s
. ..
Phononic
Fig. 4.10. Illustration of the competing heat-
transfer processes between the glass and the
mold
256 4. Shaping at Low Viscosities

materials, radiation from the entire glass bulk interacts with the mold surface
in the glass-mold system. However, one can also treat the glass formally as
an opaque solid with a special emissivity, which is generated by radiative
contributions from the glass surface and the interior.
The emissivity of a glass plate of thickness D with an arbitrary temper-
ature distribution Tg(x) (-D ~ x ~ 0) is, according to [4.21],

Eg().., Tg(x), B) = K:()..) 1 - p(¢(B))


B().., TgO) 1 _ 2("'(B)) ex (-2Dr;;()..))
p '+' P cos(¢(O))

J
1/ cos(¢(O))
x B()", Tg((l- s) cos(¢(B))) exp((l- s)K:()..))

x (1 +°p(¢(B)) exp (-2 (l/ (cos(¢(B))) - (l- s)) K:()..))) ds


(4.27)

(D: thickness of the glass plate). Equations (4.25)-(4.27) show a very general
way of computing arad. However, we accept with the following limitations:
• The geometry of the problem is one-dimensional.
• Limitations of the optical properties of the glass:
- The surface of the glass is ideally smooth, i.e., there is no diffuse scat-
tering or diffuse reflection.
- The glass has no dispersion, i.e., n()..) = const.
- The internal optical transmittance does not depend on the temperature,
i.e., K:(T) = const.
• Limitation of the optical properties of the mold surface: no diffuse scatter-
ing (r = 1).
However, the most important disadvantage of this method for determining
the radiative coefficient of heat transfer between the glass and the mold arises
when inserting arad in (4.2). This would imply the assumption that the radia-
tive heat exchanged between the mold and the glass "gets stuck" in the glass
surfaces before leaving or entering the interior of the glass. For the opaque
mold and its surface, this is correct. But for glass in its semi-transparent
wavelength range, the relevant transport mechanism is quite different. Here,
the interior of the glass exchanges heat with the mold without involving the
glass surface. Therefore, an exact calculation of the heat transfer between
the glass and the mold requires the heat equation and the radiative trans-
port equation in the glass to be solved simultaneously (Sect. 4.1.2). However,
the above approximation can be further refined by replacing (4.2) with

qrad = aopaque
rad
(Tg O _ T mO ) + atrans
rad
(Tg1 _ T mO ) (4.28)

Tg1 is the interior temperature of the glass, and


4.1 Heat Transfer Between Glass and Mold During Hot Forming 257

a~~;que = J
opaque
arad(oX)doX (4.29)

and
atrans
rad J
trans
arad(oX)doX. (4.30)

Here, the internal optical transmittance of the glass is (rigorously) divided


into two wavelength ranges (opaque and (semi- )transparent). In the opaque
wavelength range (aopaque), the radiative heat is exchanged between the sur-
faces of the glass and the mold (TgO - T illO ). In the (semi- )transparent wave-
length range (atrans), the radiation interacts between the interior of the glass
and the mold surface (Tg 1 - TITlO). Thus, the two definitions of the heat
transfer coefficient in Sect. 4.1.1 are mixed up in (4.28). This improved ap-
proximation is motivated by the combined radiative and phononic transport
phenomenon, which cannot be described by a simple boundary condition of
the heat transport problem.
In a certain way, a~~;que and a~~~ts can be compared with the radiative
contributions to the active k:~dive and the apparent k;~d thermal conductiv-
ity [4.9]. a~~;que and k~~dive are the radiative contributions to the free cool-
ing of the glass surface and the glass bulk, respectively, whereas a~~~tS and
k;~d describe the effect of the radiative heat transition (without interaction)
through the glass surface and the glass bulk, respectively.
Equations (4.26)-(4.30) become very important when heat transfer via
radiation is implemented in 3D simulation tools. At Schott, thermal models
such as the concept of active heat conductivity [4.9] are used in order to
account for heat transfer via radiation in three-dimensional glass geometries.
The determination of the heat transfer via radiation according to (4.26)-
(4.30) is an approximation that is relatively easy to realize. However, it can
only be applied to flat glass products, because otherwise the one-dimensional
approximation of the glass may no longer be adequate.
Figures 4.11 and 4.12 show some results for arad(oX) and a~~;que,trans for
the glass type of Fig. 4.2 that comes into contact with a 500°C hot black
(E2(oX)=1) mold surface. In Fig. 4.11, the glass thickness D has been varied
at a constant and homogeneous glass temperature (1500 DC). At D = 0.5 m,
the glass is completely opaque, and the coefficient of heat transfer takes its
maximum value of arad = 246 W /K m 2 . Note that in these computations,
an air layer between the glass and the mold was assumed (Fresnel reflection
at the surface of the glass). Without this air layer, all coefficients of heat
transfer rise by approximatively 40%.
If the glass thickness decreases, arad(oX) also decreases in the transparent
wavelength range (oX ::::; 4.5 ~m), and so does the value of a~~~ts because for
thin plates less radiation is generated in the interior of the glass.
In Fig. 4.12, the homogeneous glass temperature Tg has been varied for
a constant glass thickness (5 cm). If the glass temperature increases, the
258 4. Shaping at Low Viscosities

150 ,----------------.------------------,
0.5

o 2 3 4 5 6 7
?'/lJm

Fig. 4.11. Wavelength-dependent coefficients of heat transfer via radiation arad(>')


between 1500 ° C hot glass (of thicknesses 0.001 , 0.005 , 0.01 , 0.025 , 0.05 , 0.1, and
0.5 m) , and an ideally absorbing mold surface of 500 °C. In the inset, the coefficients
of heat transfer a~:;que , trans between glass and mold are plotted as a function of
the glass thickness

250 s---~~--------~------------------.

200
E
::::I.

E 150
~
Opaque

~ 100
U
~ 1000 1250 1500 1750
ts
50

2 5 6 7

Fig. 4.12. Wavelength-dependent coefficients of heat transfer via radiation arad(>')


between an ideally absorbing mold surface of 500°C and a 5-cm-thick glass plate at
constant temperatures 750, 1000, 1200, 1500, 1750 DC, and for the red temperature
profile in Fig. 4.8. In the inset, the coefficients of heat transfer a~:;que, trans between
the glass and the mold are plotted as a function of the glass temperature .

temperature difference between the glass and the mold (mold temperature:
5"00 DC) increases, too. This leads to an increasing coefficient of heat transfer
via radiation.
4.1 Heat Transfer Between Glass and Mold During Hot Forming 259

Finally, let us discuss the absolute values of OOrad once more. If these values
(of the order of magnitude 10-100 W jK m 2 ) are compared with typical coeffi-
cients of contact heat transfer in the range 500-10000 W jK m 2 , the question
arises as to why radiation actually has such a strong influence as computed
in Sect. 4.1.3. In order to answer this question, the corresponding radiative
contribution to the heat transfer coefficient has been computed for the red
temperature profile from the model computation in Fig. 4.8a and displayed in
Fig. 4.12. The result is clearly larger in the transparent range, amounting to
oo~~~tS = 400 W jK m 2 . This large value is achieved after normalizing the heat
flux in (4.2) to the surface temperature difference, which, however, is zero
in the red case (ideal thermal contact). In order to avoid this singularity, we
used a mold surface temperature of 500°C for the computation of OOrad()..).
The previous consideration has again demonstrated two important results
regarding heat transfer via radiation:

• The radiative contribution to the heat transfer between the glass and the
mold in the semi-transparent wavelength range is associated with an in-
teraction of the heat radiation from the interior of the glass and the mold
surface, so that the definition according to (4.3) is no longer valid .
• When the surface temperatures of the glass and the mold have approached,
after the first few seconds of the heat exchange, the phononic heat transfer
is then determined by the "slow" process of heat conduction from the
interior to the surface of the glass (Fig. 4.10). Now, the radiative heat flux
from the interior of the glass takes a considerable quantity of heat to the
mold surface and dominates the transport process. This can be described
by a heat transfer coefficient according to (4.1) and Fig. 4.3a.

4.1.5 Laboratory Experiments

In this section, the results of two laboratory experiments with one-dimensional


sample geometry will be presented. In both cases, the contact contribution
to the heat transfer OOgm (t) has been determined by means of the numerical
solver of Sect. 4.1.2. In contrast to the set-ups described in the literature up to
now [4.17,22-24]' the experiments discussed here are strictly one-dimensional
and have well-defined, controllable boundary conditions towards the domains
Ul and U2. The experimental set-ups are illustrated in Fig. 4.13. In the im-
mersion experiment illustrated in Fig. 4.13a, a molding plate (dimensions:
40 x 40 x 2 mm 3 ) which has been preheated to 400°C is "shot" within a sec-
ond into a 1400 °C hot glass melt at time t = O. The temperature of the glass
is chosen in such a way that the convective contribution to the heat transfer
between the glass and the mold is negligible. Within 10-20 s, the molding
plate heats up to temperatures of about 1300°C, and then starts to melt.
During this time interval, the temperature profile is detected by three ther-
mocouples in the center of the plate T m , at the surface Tl, and at a distance
of 20 mm from the plate Tg.
260 4. Shaping at Low Viscosities

b)

c:
o
~
"S
'"c: Fig. 4.13. Illustration of the im-
mersion experiment (a) and the
casting experiment (b)

In the casting experiment (Fig. 4.13b), hot glass (1500 0c) is cast at time
t = 0 into a homogeneously preheated, well-insulated steel mold. The cooling
of the glass and the heating of the mold are also measured by three thermo-
couples.
In order to determine the time-dependent coefficient of heat transfer
agm(t) from the measured temperature data T m , Tg, and T i , an inverse
mathematical problem must be solved. This calculation is based on our model
of heat transfer described by (4.10)-(4.23). The numerical solution of these
equations for a given coefficient of heat transfer yields the values T[':, Ti~'
Ti
and 3 • Now, for each time tj, the coefficient a gm must be determined from
the measured data by minimizing the sum of the mean square deviations of
the temperature differences:

(4.31 )

Equation (4.31) is a minimization problem with constraints (4.10)-(4.23),


which can be solved by means of the Lagrange method described in detail
in Sect. 3.3. The coefficient of heat transfer a gm is determined iteratively
by a gradient method where the direct problem (4.10)-(4.23) and an adjoint
problem must be solved in each iteration step.
Figure 4.14 compares typical results from both experiments described
above with experimental data from the literature [4.15]. Due to experimental
circumstances of the immersion experiment, a gm can only be determined
for t > 0.5-1 s. In the casting experiment, evaluation is possible only after
t = 4 s, because of the finite casting time. For both experiments, a gm has been
determined in two different ways. One evaluation of a gm has been carried out
accounting for the radiative heat transfer (properties of the mold surface were
measured optically, curves in the diagram denoted "without radiation"), and
a second one neglecting the radiative contribution (mold surface considered
to be ideally reflecting in the simulation, curves in the diagram denoted "with
radiation"). We did this in order to enable a comparison of our data with the
results from the literature, where the radiative contributions have not been
taken into account. As can be seen, the results from the literature (red curve)
4.1 Heat 1l-ansfer Between Glass and Mold During Hot Forming 261

:1
10000 Immersion (with radla!lon)
l

15000
~

~ McGraw (pressing)
E
~'"
1000

500

0.05 0.1 0.5 5 10 50


tis
Fig. 4.14. Comparison of the heat transfer coefficient measured in our laboratory
experiments (blue: casting experiment , green: immersion experiment) with results
from the literature by McGraw (red)

are nevertheless in good agreement with the results of our casting experiment
(blue data, "with radiation"). The corresponding curve "without radiation"
(i.e. , pure contact contribution to the heat transfer) is clearly below the
results from the literature, because this coefficient of heat transfer is free of
radiative contribution.
In the immersion experiment, however, cygm(t) decreases similarly in the
beginning, but then a steep rise occurs. This behavior is due to the fact that
the thermal conditions in the immersion experiment are clearly different from
those in the casting experiment. In the first 3-4 s, the cold plate cools the glass
surface. At this stage, the glass breaks loose from the mold surface, and a layer
of gas forms between the surfaces, as described in Sect. 4.1.1. The decrease of
CY gm is comparable with the results from the literature (pressing) and those

obtained in the casting experiment . At the same time, the temperature of the
molding plate rises at a high rate in the immersion experiment. The "cold"
glass surface of high viscosity warms up and softens again after 3- 7 s. The air
layer disappears and the thermal contact improves very quickly. Reheating of
glass surfaces (e.g., after the removal of the molds) has also been investigated
in the literature [4.25]. In our experiment , however, it causes the melting of
the molding plate after about 15 s at t emperatures above 1300 °C .

Summary and Conclusions

In this contribution, the basics of the heat transfer between glass and mold
have been introduced. Aside from pure contact heat transfer , a considerable
amount of heat is transferred from the interior of the glass into the mold by
radiation. The physics and mathematics of this combined transport problem
262 4. Shaping at Low Viscosities

are very complex and a numerically exact solution is only given in one dimen-
sion. Using a 1D solver for the associated system of differential equations, the
characteristics of the heat transfer between the glass and the mold have been
demonstrated. As a first step towards the implementation in 3D tools, an
approximative method of calculating the radiative contribution to the heat
transfer coefficient has been introduced. In this approximation, the interac-
tion of the radiation from the interior of the glass with the mold surface is
separated from the contribution of the glass surface in the opaque wavelength
range.
By applying a numerical 1D algorithm to measured temperature data
from laboratory experiments, coefficients of pure contact heat transfer can
be determined and successfully interpreted in the framework of the air/gas
layer model.

4.2 Remote Spectral Temperature Profile Sensing

Matthias Brinkmann, Norbert Siedow, Thomas Korb


Introduction

A useful counterpart to numerical modeling of the heat transport in hot glass


is the experimental measurement of the time-dependent temperature profile,
for example during hot forming. Here, a sensing method is required that works
in the harsh environment of glass making. This sensing technique should
determine temperature profiles inside permanently distorting and moving
glass without disturbing the manufacturing process [4.26]. (Hence, methods
in which, for example, a sensor must continuously be in touch with the glass
or even penetrate it are, of course, unsuitable.)
Figure 4.15 illustrates the principles of this type of remote optical sens-
ing method. The optics of the instrument is directed towards the glass gob
and records optical signals from the glass (generated by the glass itself or
induced by an external source). From these signals, the temperature distri-
bution inside the glass body may be computed by a numerical algorithm.
Of course, the basic requirement for any remote optical method is appropri-
ate and undisturbed optical access to the glass, which mayor may not be
attainable during the entire hot-forming process (e.g., during pressing).
Four possible remote optical sensing methods have been examined at
Schott to date; three of them, namely Raman thermometry [4.27]' absorption
thermometry [4.20]' and index thermometry [4.28]' have only been applied
in laboratory set-ups. The fourth method, an advanced modification of pyro-
metric temperature sensing [4.29-31] known as spectral pyrometry [4.32-42]'
can also be applied in production and is presented here. For reasons of sim-
plification, mainly flat glass products will be considered (lD approximation).
Examining 3D glass bodies is also possible, but at such a small instrumental
4.2 Remote Spectral Temperature Profile Sensing 263

3D :

Instrument Reconstruction a)

10:
Reconstruction b)

z
Fig. 4.15. Remote sensing of the temperature profile in hot glass. (a) General
case of an arbitrarily shaped glass body, (b) special case of a slab (aligned along
the observation axis). The measuring volume observed by the "lD instrument"
typically has lateral expansions h ~ 1 cm

field of view (typical diameter,::::: 1 cm) that the lateral temperature distri-
bution in this part of the volume can be safely assumed to be homogeneous
(Fig. 4.15b). In this 1D approximation, the temperature profile is determined
along a "virtual bore hole" through the glass body of diameter h. However,
in Sect. 4.2.5 three-dimensional glass shapes will be completely examined by
means of a spectral thermocamera.
During glass making, the spatial and the temporal development of the
temperature profile are of critical importance. Both high annealing rates
(e.g., 100 K/min) at the glass surfaces and fast movements of glass gobs (e.g.,
1 m/s) in production lines must be accounted for. In Sect. 4.2.5, an on-line
measurement during hot forming will demonstrate that the spectral pyrome-
ter used at Schott, possessing a time resolution far below 1 s, is sufficient for
almost all temperature-sensing problems during glass making.

4.2.1 Thermal Radiation in Hot Glass

Spectral pyrometry is an integral sensing method that registers the entire


thermal-intensity spectrum radiated from the hot glass. The application of
an inversion algorithm then calculates the temperature profile inside the
glass [4.32-37]. This is illustrated in Fig. 4.16 with a 2S-cm-thick quartz
glass block. While the radiation intensity at wavelength>. = 4.5!-Lm origi-
nates from the surface, at 2.5-2.7!-Lm thermal radiation is recorded from the
interior of the glass. At 2.0 !-Lm, the glass is transparent, i.e., no thermal ra-
diation is generated. Here, only the radiation from the backside of the glass
(e.g., from a hot mold) can reach the surface.
The wavelength- and temperature-dependent absorption coefficient K,(>.,T)
determines the depth from which the radiation reaches the front glass surface.
For the quartz glass considered here, K,(>., T) was measured in the wavelength
range 1.5-4.5 !-Lm and for the temperatures of 20, 500, 600, 700, SOO, 900, and
264 4. Shaping at Low Viscosities

2.0j.Jm

2.Sj.Jm

2.7j.Jm Glass
Spectrometer
4.Sj.Jm - z
o D = 28 cm
Fig. 4.16. Thermal radiation of a quartz glass block, thickness D = 28 cm

1000 ac. The results are presented in Fig. 4.17. The absorption coefficient of
glass results from the transmission T of a slab of thickness Dtrans, according
to

T('\, T) = (1 - 2p('\, T))e-",()..,T)Dtrans , (4.32)

where p('\, T) is the Fresnel reflectivity. The inverse absorption coefficient


l('\, T) = ",(i,T) is called the radiative penetration depth. In Fig. 4.17, this
parameter is also shown as functions of wavelength and temperature. For a
quartz glass block of thickness D = 28cm (Fig. 4.16), we get temperature
information about the whole 1D temperature profile from the analysis of the
intensity spectrum in the wavelength range 1.5-4.5 Il-m.
In addition, surface reflectivity p('\, T) is also important for the thermal
radiation of a glass gob. In the case of vertical incidence (which is exclu-
sively considered here), the reflectivity is connected with the refractive index

1000 10000
SOO SOOO

100 1000
SO 00
~~ 10 E
:!::: S , ,
---"'--~',
0 --
E

'" 1 !i~\\~~~l~~%i%~:'\, ____


O.S , ' 10
,/ '''~----'''''' S
0.1
O.OS
--:~\~~i~~
O.S
1.S 2.0 2.S 3.0 3.S 4.0 4.S
,qJm

Fig. 4.17. Wavelength dependence of the absorption coefficient '" (continuous


curves) for quartz glass at the temperatures 20, 500, 600, 700, 800, 900, and
1000°C (from bottom to top). The dashed lines show the radiative penetration
depth l = 1/", as function of wavelength and temperature (now, 1000°C at the
bottom)
4.2 Remote Spectral Temperature Profile Sensing 265

n('x, T), according to [4.29]:

(,X T) = (n(,X,T)
p, n(,X,T) +1
_1)2 (4.33)

Figure 4.18 shows the wavelength dependence for quartz glass. Until now,
there is no commercial device available for measuring the temperature de-
pendence of this material parameter. Because the temperature dependence
of the glass reflectivity is usually very small, the data for room temperature
will be used here.
The basis for the computation of the temperature profile T(z) (z = 0 at
the glass surface, z = D at the opposite side of the slab) from the measured
intensity spectrum Jill(,X) of the thermal radiation is the following analytical
relation (see Sect. 4.2.2):

J~(T,
D

Jill(,X) = 1 _ ~~~2KD 'x)B(T, ,X)[e- Ks + pe-2KD+Ks]ds


o (4.34)

J~(T(s'),
s

~s = 'x)ds' ,
o
where B(,X, T) is Planck's function (see previous section). If the material
parameters ~ and p are known and the temperature profile T(z) is given, this
equation yields the intensity J ('x, T (z)). Equation (4.34) describes the so-
called direct problem of spectral pyrometry, because from the "cause", namely
the temperature profile inside the glass, we can calculate the "effect", the
thermal radiation spectrum.
However, for spectral pyrometry we would like to use the reverse relation,
because the thermal radiation 1(>..) (Le., the effect) is measured and the tem-
perature distribution T( z) (the cause) is to be determined. Equation (4.34)

1.5 0.04

0.03
1.45

c: 0.02 "'-

1.4
0.01

1.35 0
1.5 2 2.5 3 3.5 4 4.5
A/lJm

Fig. 4.18. Wavelength dependence of the reflectivity p(>.., T) and the refractive
index n(>.., T) for quartz glass at room temperature
266 4. Shaping at Low Viscosities

cannot be solved analytically for T(z). Although I = 1(>', T (z)) is unique,


this is not the case for the relation T = T (z, 1(>')). For identical material
parameters, the spectrum 1(>') can be generated by numerous (mostly un-
physical) temperature profiles T(z).
The mathematical problem of calculating the temperature profile T(z)
inside the glass from the experimentally measured spectrum I (>.) is called
the inverse problem of spectral pyrometry. Except for the first publications in
the years 1973-1975 [4.32-37]' this problem has not been examined. During
the last two years, the Institute for Industrial Mathematics (ITWM, Kaisers-
lautern), in cooperation with Schott, and the Eindhoven University of Tech-
nology [4.42] have independently developed numerical algorithms for solving
the inverse problem. The theoretical basis of the ITWM method will be pre-
sented in the next section. Results for numerically generated spectra (i.e.,
computed by means of (4.34) are discussed in Sect. 4.2.3. In Sect. 4.2.4, the
performance of the algorithm is demonstrated with measured spectra.

4.2.2 The Inverse Problem of Spectral Temperature Sensing

The relation between the radiation intensity 1(z, >.) of a hot glass gob and the
temperature profile T(z) is described in (4.34). For a one-dimensional domain
(e.g., a slab of thickness D), the radiative transfer equation (see Sect. 4.1)
reads
OJ
jL oz (z, jL, >.) + K,(T, >')1(z, jL, >.) = K,(T, >')B [T(z), >.], 0::; z ::; D, (4.35)
where jL is the directional cosine vertical to the surface normal. The glass
boundary, z = Zg, is assumed to be specularly reflecting:

1(zg, jL, >.) = p(jL)1(zg, -jL, >.) + [1 - p(jL)] n~B(Ta, >.) , (4.36)
where Ta is the temperature of the surrounding medium and ng is the refrac-
tive index of the glass. A detailed derivation of this equation can be found
in [4.7,8,10]. Here, we tentatively assume that the external medium is a
black-body radiator.
Based on the formal solution of this equation for the boundary point
Zg = 0, the radiation intensity of the glass can be calculated (generalization
of (4.34)):

1ill (>.) = 1 _ ~~~21<D {(I - p)e-I<Dn~[B(Ta2' >.) + pe-I<D B(Tal' >.)]

J
D

+ K,(T, >')B(T, >')[e-I<s + pe- 21<D+I<S]ds} ,


(4.37)
o

J
s

K,s = fi:(T(s'), >')ds' .


o
4.2 Remote Spectral Temperature Profile Sensing 267

Tal and Ta2 are the exterior temperatures at the left (z = 0) and the right
(z = D) boundary (black-body radiators). It is assumed that the spectral
pyrometer is installed at the left boundary (z = 0). Therefore, Tal is ap-
proximately 300 K « T(z), so that the respective term B(Tal')..) can be
neglected in (4.37). From the radiation intensity Im()..) recorded by the spec-
tral pyrometer and (4.37), the temperature profile T(z) can be determined
via the inverse approach.

Radiating Background

Now, we extend the intensity equation (4.37) for radiating backgrounds that
are not perfect black-body radiators. Let us again consider the radiative
transfer equation (4.35). At the left boundary, z = 0, the radiation is reflected
specularly:

Here, PI (p,) is the reflection coefficient between the glass and the left exterior
medium with the refractive index nl, Cl ()..) is the emissivity of the left exterior
medium, and B(Tal')..) is Planck's function for the temperature Tal.
At the right boundary, we consider a radiating background with the emis-
sivity C2()..) , the temperature T a2 , and an interface medium between the glass
and the background with refractive index n2.

I(D, p,,)..) = p2(p,)I(D, -p,,)..) + (1 - p2(p,))2r()..)(1 - c2()"))I(D, -p,,)..)

+ (1 - P2(p,))(1 - r()..))(l - C2()..)) J (1 - P2(p,'))I(D, /-L',)..)


1"'>0

(4.39)

The first term of the right-hand side of (4.39) describes the specular reflection
at the glass surface, the second term describes the specular reflection at the
background, the third represents the diffuse reflection at the background, and
the fourth defines the emission of the background. The parameter 0 :s; r()..) :s;
1 describes that part of the reflected radiation that is specularly reflected at
the background. However, the following discussion is based on the condition
that the background medium is not diffusely reflecting, i.e., that r()..) = 1. In
this case, the system of equations (4.35), (4.38), and (4.39) has an analytical
solution if the temperature profile T(z) is known. For the boundary point
z = 0 and the direction p, = -1 (observation direction of the pyrometer),
this yields
268 4. Shaping at Low Viscosities

fmp..) = 1- PI 2 { e-I<D [(1 - P2)c2B(Ta2' A) + Te-I<D (1 - pd


1- PITe- I<D
D

cIB(TaI , A)] +jK,(T(S) , A)B(T(s), A)e-I<s [1 + Te- 2(I<D-I<S)] dS} ,


o
S

T = P2 + (1 - P2)2(1 - c2(A)) , K,s = j K,(T(s'), A)ds' .


o
If we take into account that the left exterior medium is air at room temper-
ature, and 1 - PI Te- 2I<D ~ 1, this results in the mathematical description of
the radiation fm(A) measured by the spectrometer:

fm(A) = (1 - PI){ e-I<D (1 - P2)c2B(Ta2, A)


D
(4.40)
+j K,(T(s),)..)B(T(s),)..)e-l<s [1 + Te- 2(I<D-I<S)] ds}.
o

Simple Approximation Method for the Solution of the


Inverse Problem

The Rosseland approximation [4.7,8,10] is a simple asymptotic method for


determining the radiation intensity, i.e., the direct problem. There is an analo-
gous method for the inverse problem - the Eddington-Barbier approximation
[4.43]. Like the Rosseland approximation, this method has also been devel-
oped in astrophysics and is based on the same ideas:
• The considered medium is assumed to have infinite extent, i.e., D ---+ 00.
Thus, we consider an optically thick medium. In the linear case, from (4.37)
it follows that
00

fm()..) = (1 - p)K, j B(T(s), )..)e-I<Sds . (4.41)


o
• The Planck function is approximated locally at a point z* (which will be
determined more precisely below) by a linear function.
dB dB*
B(s) ~ B(z*) + (s - z*) dz (z*) := B* + (s - z*) dz '
(4.42)
fm()..) ~ (1 - p) { B* + d~* (~ - z*)} .

The position z* is selected in such a way that the derivative of the black-body
radiation in 4.42 is zero, hence
4.2 Remote Spectral Temperature Profile Sensing 269

1
z* = -,." . (4.43)

This results in the Eddington~Barbier approximation for the radiation inten-


sity:
1
I m(>,) ~ (1 - p)B(T(z*), A), z* = - . (4.44)
,."

Due to the constraints mentioned above, this method can be applied for de-
termining the temperature profile close to the glass surface (in the opaque
wavelength range). In the semi-transparent wavelength range of the glass,
the conditions of the Eddington~Barbier approximation are not fulfilled com-
pletely. Hence, for spectral pyrometry the method will only yield mediocre
results (see discussion in Sects. 4.2.3 and 4.2.4). However, similar to the
improved diffusion approximation [4.10]' we can also try to weaken the con-
ditions of the Eddington~Barbier approximation here. We neglect that the
medium is assumed to have infinite extent. In the linear case, we again sub-
stitute Planck's function in 4.37 by a linear approximation at a point z*.
Integration leads to

Im(A) ~ 1 _ ~~~2"D {(I - p) e~"Dn~ [B(Ta2' A) + pe~"D B(Tal' A)]


1 dB*
+ B*(l - e~"D)(l + pe~"D) + - - [(1 - e~"D)(l - pe~"D)
,." dz
- ,."D(l- p) e~"D - z*,.,,(l- e~"D)(l + pe~"D)]} . (4.45)

As in the case of the Eddington~Barbier approximation, we select z* in such


a way that the derivative of the Planck function is zero:
(1 - e~"D)(l - pe~"D) - ,."D(1- p)e~"D
z* = ,.,,(1- e~"D)(l + pe~"D)
(4.46)

This yields a modified Eddington~Barbier approximation:

Im(A) ~ 1 _ ~~~2"D {(I - p)e~"Dn~ [B(Ta2' A) + pe~"D B(Tal' A)]


(4.47)
+ B(T(z*), A)(l - e~"D)(l + pe~"D)} .

Again, the term B (Tal, A) can be neglected. In contrast to the original


Eddington~Barbier approximation, point z* now lies within the considered
medium: 0 :S z* :S D /2.

Regularization Method

Equation (4.37) is a nonlinear integral equation. If we assume the surrounding


air to be at room temperature and further take into account that the term
p e~2"D+"s is small, (4.37) can be simplified:
270 4. Shaping at Low Viscosities

J~(T, J~(T(s'),
D s

Imp..) = (1- p) )..)B(T, )..)e-I<sds, ~s = )..)ds' . (4.48)


a a
Such a nonlinear integral equation can only be solved numerically. First, the
interval [0, D] is divided into

°= Zl < ... < Zi-l < Zi < Zi+l < ... < ZN = D , (4.49)

and the integral in (4.48) is approximated by a sum:


N-l
Im()..) ~ (1 - p) L B(T(Zi),)..) (e- ai - e- ai + 1 ) ,

i=l

J~(T(s),
Zi (4.50)
ai = )")ds .
a
Instead of the nonlinear integral equation, (4.48), we now have a nonlinear al-
gebraic equation. If the intensity for different (e.g., M) wavelengths is known,
and if we want to determine the temperature at N positions in the glass, we
have a system of M nonlinear equations for N unknown temperatures. This
system can also be written in the following way:

(4.51 )

Here, Ti is the temperature to be determined at the ith grid point Zi, and If
the measured intensity for the kth wavelength )..k. Using Newton's method,
this nonlinear system of equations can be linearized, which yields the follow-
ing iteration method:

G'(Tz)(Tz+l - Tl) = Iill - G(Tz), l = 0, 1,2, .... (4.52)

Because the number of the considered grid points N (number of columns)


must not be larger than the number of measurements M (number of wave-
lengths), the system is, in general, (N < M) overdetermined. By multiplying
with the transposed matrix G'*(Tz), (4.52) is transformed to a linear system
with the same number of equations and unknowns:

G'*(Tz)G'(Tz)(Tz+l - Tz) = G'*(Tz)(I ill - G(Tz)), l = 0, 1,2, .... (4.53)


However, one can show that (4.53) is ill-conditioned (the relation between the
largest and the smallest eigenvalue is very large). This means that a small
error of the right-hand side 1m - G (Tz) (e.g., experimental or rounding errors)
results in a large error of the temperature reconstruction. Therefore, the sys-
tem of (4.52) must be regularized. A detailed discussion of the regularization
techniques known from literature can be found in [4.44]. Here, we will only
concentrate on several essential aspects.
4.2 Remote Spectral Temperature Profile Sensing 271

Instead of the system of equations (4.52) with G' (Tz) : X -+ Y (X defines


the space of the temperature functions and y the image space), we consider
the following minimization problem:

inf IIG'(Tz)(TI +1 - Tl) - (1m - G(Tz))11 (4.54)


TEV

with the constraint

(4.55)

By means of a Lagrange multiplier, this minimization problem can be trans-


formed into a minimization problem without constraints:

For the solution of this minimization problem, the following theorem [4.44]
is useful:
Theorem: The minimizing element Tz+1,,,, of J", from (4.56) solves the reg-
ularized normal equation

Ts is an appropriate approximation of the solution and a is the so-called


regularization parameter. If the regularization parameter is very small, the
solution of (4.57) converges towards the solution of (4.53). For large a,
the constraint is fulfilled. In practice, a series of regularization parameters
aj, j = 0, 1, ... , will be selected that monotonously converges towards zero.
Thus, it is guaranteed that a regular problem in the vicinity of the real (ill-
posed) problem is solved. The details of selecting the regularization parameter
will be described later.
In the literature, the constraint, (4.55), is often called classical Tikhonov
regularization. For concrete problems, for example the inverse problem of
spectral pyrometry, it is now important to find a regularization (constraint)
tailored to the problem.
If we assume that the temperature in the glass results from the combined
heat transport through conduction and radiation, T(z) fulfils the stationary
equation:

a ( khaz
az aT -q(T(z)) ) =0. (4.58)

q(T(z)) represents the temperature-dependent radiation flux vector. As we


have already mentioned above (see, e.g., [4.10]), the influence of radiation
can be calculated by the Rosseland approximation. Instead of (4.55), as a
constraint we choose
272 4. Shaping at Low Viscosities

(4.59)

A very common method for the determination of the regularization parameter


a is the Morozov discrepancy criterion. If only the right-hand side of the
problem to be regularized is disturbed by errors (which we will assume first),
the idea of this method simply is to select the regularization parameter in
such a way that the norm of the residual (limited by an upper, predefined
value 6) corresponds to the error norm of the right-hand side E. Let us assume
that there are errors with respect to the measured intensity lID:

(4.60)

Now, the regularization parameter is selected in such a way that


1

IIIID - G(TI+dll = (t,(Ik' - G(Tz+1, Ak))2 d Ak ) "2 <6= T E (4.61 )

holds. IIIID - G(Tz+d II is called the norm of the residual, and T is an arbitrary
real number greater than 1. If a parameter a is found that fulfils the con-
straint, (4.61), the temperature computed with this parameter is accepted as
the result.
The method described here has been implemented into a software package
that automatically reconstructs the temperature profile. The following input
parameters are necessary:
• the material parameters /'£(A, T) and ng of the glass,
• the glass geometry: thickness D and number of spatial nodes N where the
temperature ought to be computed,
• the measured intensity spectrum Ik'(Ak) for M 2: N wavelengths,
• several control parameters of the algorithm: the maximum error of the
measured data Ik'(Ak), the maximum number of iterations, ....
The program computes the temperature profile T(z) and the intensity spec-
trum I(A, T(z)) calculated according to (4.34) (solution of the direct prob-
lem), which ideally corresponds to the measured spectrum IID(A).

4.2.3 Sample Computations

The algorithm for the solution of the inverse problem of spectral pyrometry,
developed in Sect. 4.2.2, will now be examined using a numerical example.
4.2 Remote Spectral Temperature Profile Sensing 273

Let us again consider the quartz glass slab of thickness D = 28 cm from


Figs. 4.16 and 4.17. After establishing an initial homogeneous temperature
profile of 1200 o e, free annealing in air for 50 s was modeled. The result is
the symmetric temperature profile displayed in Fig. 4.19 (black curve, input).
For this profile, the vertically (towards the left-hand side) radiating intensity
spectrum I()") was computed according to (4.34) (black curve in Fig. 4.20,
input). On the basis of these 120 spectral data for I()..) , the reconstruction
algorithm is supposed to compute the temperature profile T(z). Figure 4.19
shows the result.
First, we will only consider the red curve (0%). In this case, the "exactly"
computed intensity spectrum I()") was used as input file for the algorithm.
The result corresponds very well to the input profile from z = 0 up to the
center of the slab (z = 0.14m). Then, deviations occur. In particular, the
temperature gradient at the backside does not correspond well to the input
data.
This behavior demonstrates that the inverse problem of spectral pyrom-
etry is ill-posed. The slightest inexactness during computation (e.g., due to
numerical rounding) leads to large deviations of the temperature profile at
the backside of the slab. The physical reason is that the relatively cold back-
side produces less thermal radiation than the hot interior. In addition, this
radiation is also absorbed On the way through the glass slab. Although the
temperature in the back half of the slab (14 cm ::; z ::; 20 cm) is changed by
the reconstruction algorithm, this results in almost exactly the same inten-
sity spectrum I()..) (red curve in Fig. 4.20) for the reconstructed temperature
profile, compared to the input spectrum (black curve in Fig. 4.20). This

1200

~ 1100
~
~ 1000
Qj
E 900
~
800

700 +-~-r----.---'----.----.-~
o 0.05 0.10 0.15 0.20 0.25
z /m

Fig. 4.19. Temperature profile along a quartz glass slab of thickness 28 cm.
Black: given input profile, red: reconstructed profile by use of the "exact" inten-
sity spectrum for the input profile, blue: reconstructed profile by use of the input
spectrum with a 1% error, green: reconstructed profile by use of the input spec-
trum with a 5% error, cyan: reconstructed profile if one takes advance of profile
symmetry, brown: reconstruction according to Improved Eddington~Barbier
274 4. Shaping at Low Viscosities

1.0 - . - - - - - - - - - - - - - - - - ,

vo.s Hot background

o
~ 0.6

6:
CO 0.4

~ 0.2

o +---,,----.-~,-~-._~_1
2 .0 2.5 3.0 3 .5 4 .0 4 .5
}./lJm

Fig. 4.20. Intensity spectra of the 28-cm-thick quartz glass slab. Black : directly
computed from the input profile, red : computed from the reconstructed red pro-
file (previous figure), blue : like black, but additionally accounting for a radiating
background (with a temperature of 1200 °C)

demonstrates that strong temperature gradients at the back boundary can


hardly be determined by spectral pyrometry.
The brown curve in Fig. 4.19 (Impr. Eddi) shows the result of the re-
construction according to Improved Eddington- Barbier ((4.46) and (4.47)).
Again, the initial slope (0-5 mm) is correctly represented. However, inside
the glass there is a strong deviation of the temperature profile from the in-
put profile. The temperatures in the interior are always too low because this
approximation does not account for the absorption of the radiation on the
way from the interior towards the surface.
When we apply the reconstruction algorithm to experimental results (next
section), we almost never get completely accurate results because the mea-
surements are always subjected to errors. In order to examine the influence
of such errors, the exact values of the input intensity spectrum (black curve
in Fig. 4.20) were numerically disturbed by a statistical error of 1 % and 5%
here. The spectra manipulated in such a way were then used as input files for
the reconstruction algorithm to compute the temperature profile. The results
are shown in Fig. 4.19 (blue and green curves). Near the front (0 :::; z:::; 2cm)
surface, both profiles are still close to the input data. However, in the interior
one finds deviations of the order of magnitude ±10 K (±20 K) for the input
data error of 1% (5%). The temperature gradient at the backside is also not
represented correctly.
The blue curve in Fig. 4.20 again shows the intensity spectrum computed
from the input temperature profile, though additionally accounting for a black
body radiator of the t emperature 1200 °C behind the slab.
Compared to the black curve (without background), a difference occurs
at wavelengths close to 2.4 f.l-m because the penetration depth of the radia-
tion here is approximat ely 50-100 cm. For these wavelengths, radiation from
4.2 Remote Spectral Temperature Profile Sensing 275

the background penetrates the quartz glass slab and reaches the spectral
pyrometer.
If this spectrum I(A) and the radiation properties of the background
(black-body radiator at 1200°C) are used as input parameters for the re-
construction algorithm, this results again in the red temperature profile (an
additional statistical error has not been introduced here), i.e., the radiating
background is "correctly" taken into account by the algorithm.
Finally, we want to present a method to enable the correct reconstruction
of the temperature gradient at the backside. If we know that the temperature
distribution is symmetric with respect to the center plane of the slab, the
reconstruction algorithm can be applied to a substitute problem. Though
only half of the thickness is used as input parameter, an ideal mirror is used
as the background (see sketch in Fig. 4.19). A hot glass slab of thickness
D /2 = 14 cm in front of a mirror radiates in the same way as a glass slab of
thickness D with a symmetric temperature profile.
The result of the reconstruction (with D /2 and mirror) is one half of the
temperature profile that can then be mirrored at the center plane. The cyan
curve in Fig. 4.19 shows these data, which now almost perfectly correspond
to the input temperature profile.
In summary, using the inversion algorithm, temperature profiles can be
reconstructed even in the case of noisy input data. Decreasing temperature
profiles at the backside of the glass gob are problematic. However, such tem-
perature profiles can at least be computed correctly up to the maximum
temperature.

4.2.4 Laboratory Experiment

In Sect. 4.2.3, we explored the reconstruction algorithm by using a numerical


example. Now, the method is applied to the measured spectra of a laboratory
experiment. A 28-cm-Iong quartz glass rod (diameter 25 mm) is placed into a
horizontal, electrically heated tube furnace. The material parameters K, and
p of the glass are represented in Figs. 4.17 and 4.18. The entire discussions
in Sects. 4.2.1 and 4.2.3 are relevant to this experiment. The only difference
is that here the temperature profile along the glass rod does not result from
free cooling but from external heating instead.
Figure 4.21 shows the experimental set-up. A quartz glass rod is placed
in a commercial tube furnace (open at both sides) and heated to different
temperatures in successive steps. The tube furnace has an interior diameter
of 28 mm, so that the quartz glass rod fills the furnace tube almost com-
pletely. The heating wire is symmetrically wound around the furnace tube
with respect to the center. The temperature profile of the furnace (decreas-
ing towards the exterior) is transferred to the quartz rod during the heating
process. The corresponding temperature profiles T(z) inside the glass were
determined by means of spectral pyrometry. In six longitudinal grooves in
276 4. Shaping at Low Viscosities

Cross section

<I)
Q)
Ci
o=>
(.)
o
E
Spectrometer ~
.c
f-
Fig. 4.21. Experimental set-
up for the det ermination of
the temperature profile in hot
glass

the quartz rod, eleven thermocouples were placed, whose sensor tips were
equidistantly distributed along the furnace axis.
To check the temperature profile T( z ) by some kind of reference contact
measurement , six longitudinal grooves were drilled into the glass rod (see
cross section in Fig. 4.21). Into these grooves, 11 thermocouples were glued
by a special adhesive. The sensor tips of the thermocouples are equidistantly
distributed along the furnace axis. Figure 4.22 shows the temperature profiles
measured by the thermocouples for different mean furnace temperatures. In
this measurement, errors occur due to the temperature gradients between the
thermocouples and the glass surface and due to the radial temperature gradi-
ents which have not been measured. These inexactnesses can be estimated as
a total error of approximately ±30 K, which represents the difference between
the glass temperature along the axis of the glass rod (= optical axis of the
spectral pyrometer) and the temperatures measured by the thermocouples.

1000 .---------------------------------~

800
U
.!!...
~ 600
=>
Cii 500 C
Qj 700 · C
~ 400 900 C
~

200

O +-~---r----_.------._~--_.----_,r-~

o 0.05 0.10 0.15 0 .20 0 .25


z /m

Fig. 4.22. Temperature profiles measured by the thermocouples (along the quartz
glass rod) for the mean furnace temperatures 500, 700, and 900°C
4.2 Remote Spectral Temperature Profile Sensing 277

The temperature profiles (Fig. 4.22) are symmetric with respect to the
center of the rod, possess very steep gradients (:=:::; 100 K/cm) at both ends,
and display a fiat plateau in the interior. From Fig. 4.21, one can see that
for the three temperature profiles T(z), the intensity spectra I()..,) were mea-
sured from the left side by the spectral pyrometer. Figure 4.23 displays the
results. These data have been used to reconstruct the temperature profiles.
Figure 4.24 compares the results for the furnace temperature of 500°C to
the thermocouple data. The reconstructed temperature profile as well as the
respective spectrum are drawn in red in Fig. 4.24 and Fig. 4.23. Up to the
center of the quartz glass rod, the temperature profile measured by the ther-
mocouples corresponds well to the reconstructed profile. Deviations occur in
the range of 10-30 K. The measured and the reconstructed spectra (Fig. 4.23)
correspond very well. Nevertheless, the temperature gradient in the back part
of the quartz glass rod can only roughly be determined by the reconstruction,
as has already been demonstrated in Sect. 4.2.2. Besides, the result of the
Improved Eddington-Barbier approximation shows temperatures that are far
too low in the center (Sect. 4.2.3).
In order to examine whether the remaining temperature difference be-
tween the reconstructed and the thermocouple temperature profile is caused
by the inaccuracy of the experiment or by the reconstruction algorithm,
we chose the following method. From the thermocouple temperature profile
(mean furnace temperature 500°C), the respective spectrum was computed
according to (4.34) (blue curve in Fig. 4.23), which should be identical with
the measured spectrum in the case of a complete correspondence between the

1.5 - , - - - - - - - - - - - - - - - - - - - - - ,

6
o
1.0
o
o
U'l

2.0 2.5 3.5 4.0 4.5

Fig. 4.23. Radiative intensity spectrum measured at the mean furnace temperature
of 500 0 C (black squares), compared to the reconstructed spectrum (red points) and
the spectrum computed according to the thermocouple profile (blue line). For short
wavelengths, several wavelengths (black squares) have been omitted that were not
used for the evaluation
278 4. Shaping at Low Viscosities

600

u 500
~
Q)
:5
~ 400
Q)
c. Rec. from experiment
E
Rec. from computed spectrum
~ 300 Thermocouples
Impr. Eddington- Barbier approximation

200

o 0.05 0.10 0.15 0.20 0.25


z /m
Fig. 4.24. Temperature profile measured by the thermocouples ( black squares, con-
trol temperature 500 °C) , compared to the reconstructed t emperature profile (red
line) , the result of the Improved Eddington-Barbier approximation (green points) ,
and the reconstruction from the computed spectrum (blue line, see text)

spectral pyrometry and the thermocouple measurement. However, there is an


evident deviation. Now, this computed spectrum was used as input data for
reconstructing the temperature profile (blue curve in Fig. 4.24) . Compared
to the reconstructed temperature curve from the measured spectrum, this
profile follows the thermocouple data more closely up to the center of the
glass rod. However, an additional edge-smoothing effect (compared to the
thermocouple profile) occurs.
Hence, the evaluation presented here shows that the differences between
the temperature profile reconstructed from the measured spectrum and the
thermocouple profile are not due to the reconstruction algorithm. Instead,
they are probably caused by the temperature differences between the thermo-
couples and the glass temperature at the center axis of the rod, as discussed
above.
In summary, the reconstruction algorithm presented here can be used for
the computation of temperature profiles from measured thermal radiation
spectra. Compared to the Improved Eddington- Barbier approximation, there
is an evident improvement with respect to the interior of the glass.
In order to complete the analysis of the laboratory experiment , Fig. 4.25
shows the analogous evaluation of the measurements at the mean furnace
temperatures 700 and 900 DC. Again, the thermocouple profiles correspond
very well with the reconstructed profiles on the left side of the glass rod,
whereas the decrease following the maximum temperature can only be rep-
resented qualitatively. The result according to Improved Eddington-Barbier
shows clear deviations even near the left surface.
4.2 Remote Spectral Temperature Profile Sensing 279

1000 I------=::::=====----~~
900 0 e
900

u 800
l!...

.2
Q)
700
~
~ 600
E Rec. from experiment
~ Rec. from computed spectrum
500 Thermocouples
Impr. Eddington- Barbier approximation
400
700 0 e
300+-~---.--~-,--~--.-~---.--~-,,-~

o 0.05 0.10 0.15 0.20 0.25


z /m
Fig. 4.25. Temperature profiles measured by the thermocouples (black squares,
mean furnace temperatures 700 and 900 DC), compared to the reconstructed (from
the measured spectrum) temperature profile (red line), the result of the Improved
Eddington-Barbier approximation (green points), and the reconstruction from the
computed spectrum (blue line)

The differences between the reconstructed temperature profile and the


thermocouple curves on the left-hand side of the quartz glass rod are not
due to the reconstruction algorithm (as has already been shown for the mean
furnace temperature of 500°C), but to the experimental set-up, which is
not perfect from a thermal point of view. If we use the spectrum computed
from the thermocouple profile for the reconstruction instead of the measured
spectrum, this yields a very good correspondence between the reconstruction
results (blue curves) and the thermocouple curves, at least for the front part
of the glass.

4.2.5 Spectral Imaging of Hot Glass

In the previous section, spectral pyrometry has been tested in a laboratory


experiment. Now we want to find out how far this method can also be applied
directly during glass production, particularly in hot forming. As we have
already mentioned in the introduction, the following new challenges must be
met:
• harsh production environment (e.g., high room temperatures, vibrations,
etc. ),
• limited optical access to the hot glass (obstructed by molds, machines,
chilling equipment, etc.),
• usually complex, not one-dimensional geometry of the glass gob,
• continuously distorting shape of the glass gob,
• fast movement of the glass gob.
280 4. Shaping at Low Viscosities

The basis for a successful application of this technique is a spectroradiome-


ter, which is able to operate outside the laboratory environment and records
a complete spectrum within a short time. In the following, results from ap-
plication of spectral pyrometry to glass production will demonstrate this
capability of the method. The inset of Fig. 4.26a displays a thermo camera
image showing a glass drop falling into a mold. The thermal radiation of the
glass gob was registered at a wavelength of 5 Jl.m. Here, the glass is opaque,
i.e., the surface temperature distribution was determined.
This thermo-image already shows the fundamental problems for three-
dimensional glass bodies. Although the cylindrical geometry of the drop is
relatively simple, the temperature profile of the glass gob already shows errors
at the left and right boundary. Here, the glass appears to be cooler (red area)
than in the center of the image (yellow area), although in both cases the
glass surface was measured. The temperature distribution in the drop being
almost rotationally symmetric, the temperatures ought to be identical. The
difference is caused by the varying angle between the observation direction of
the camera and the normal vector of the glass surface. The emissivity of the
glass surface decreases with increasing tilt angle [4.21,45,46]. The evaluation
of the thermo-images should account for this fact. However, commercially
available thermocameras in general are not able to do so.
The analysis of three-dimensional images of a "spectral thermo camera"
also ought to account for this with respect to radiation from the interior of
the glass. Such a combination of image analysis and complex reconstruction
algorithm is not yet available. Therefore, we will apply the 1D variant of spec-
tral pyrometry, which has been described before, to this glass drop. Although
the glass geometry is not one-dimensional here, the thermal radiation can,
nevertheless, be approximated as locally one-dimensional (for a vertical ob-
servation direction), because the diameter of the instrumental field of view h
(Fig. 4.15) is smaller than the local drop diameter D (here: hiD « 1).
Figure 4.26a shows the temperature distribution at a certain point (green)
and time, determined by the spectral pyrometer and the reconstruction algo-
rithm from Sect. 4.2.2. Again, we took advantage of the fact that the temper-
ature profile in the glass drop is symmetric. By using the maximum scan rate
of the spectral pyrometer (less than 1 s), enough spectra were successively
measured for obtaining the time-dependent development of the temperature
profiles of two drops (optics fixed to a fixed position, green point in Fig. 4.26).
Using the reconstruction algorithm, time-dependent temperature profiles at
different glass depths (at the position of the green point) could be computed
from these data. Figure 4.26b shows the result.
Interpreting the data is relatively easy. As soon as the drop appears in
the field of view of the optics, there is already a temperature gradient of
approximately 100 K at the surface caused by free cooling in air. Because
at this stage the glass flows from the open feeder at low velocity (driven by
the hydrostatic glass pressure), one can also detect further cooling of the
4.2 Remote Spectral Temperature Profile Sensing 281

1000 , . . - - - - - - -- - - - - - - - - - ---, a)

Spectral pyrometer

Thermal image
(.J. = 51..1m)
800 ~--------------------------------~
z
b)
z = 50 mm

z = 10 mm
z = 5mm

z=Omm

0>
Ol
C
:::l
1. Gob a:: 2. Gob

Time

Fig. 4.26. Temperature distribution in a glass drop during production. (a) Tem-
perature profile along the diameter vertical to the axis of the cylinder, measured at
a certain time t = 5 s. The inset shows the picture of a thermo camera, containing
the field of view of the spectral pyrometer (green point). (b) Time-dependent devel-
opment of the glass temperature at different depths of the growing drop, determined
at a fixed observation position (green point)

glass surface (temperature decrease in the curves z = 0, 5, and lOmm). In


the interior of the drop (z = 50 mm), however, there is almost no change
in temperature. At the end of the gob-feeding process, hot glass that has
remained in the feeder is pressed into the drop by a plunger. By this enforced
glass flow, hot glass from the interior of the drop reaches the surface, which
leads to an increase of the surface temperatures (z = 0, 5, and lOmm). Then,
the drop is cut and leaves the field of view of the spectral pyrometer (green
point). After a certain time, the next drop appears in the observation field
of view and is registered.
This example proves that the above-described ID variant of spectral py-
rometry can be successfully applied to the measurement of temperature pro-
282 4. Shaping at Low Viscosities

files in hot glass during hot forming. Moreover, Schott has already made an
initial attempt to determine a complete three-dimensional temperature dis-
tribution in hot glass with a prototype "spectral thermo camera" [4.47]. Here,
simple glass shapes were examined (again in laboratory experiments).
The sketch in Fig. 4.27 shows the experimental set-up for a glass paral-
lelepiped (50 x 50 x 90 mm 3 ), whose front side is observed vertically (to the
plane normal).
In this experiment, the glass sample was first heated in a laboratory fur-
nace up to approximately 780°C (homogeneously), then removed from the
furnace and placed on a cold base (marked black in the thermo-images).
During the cooling phase, the temperature distribution in the parallelepiped
was recorded by a spectral thermo camera. Figure 4.27 shows the tempera-
ture distribution after approximately 1 min. The camera registers (nearly)
simultaneously at seven wavelengths in the range 2~5 Il-m. From the radia-
tion intensity images, temperature profiles along cross sections at different z
positions have been computed according to the Improved Eddington~Barbier
method, maximally up to the penetration depth D/2 (see above). The out-
line of the parallelepiped and the base are manually drawn into the thermo-
images. In the image z = 39 mm, a hot spot is marked. This is a black-body
radiator behind the parallelepiped, which radiated through the glass at the
measured wavelength.

z =0.5 mm z = 1.0 mm z = 1.5 mm

iii! •

• .

o !It=JJ'"~rg
I 90mm I

Fig. 4.27. Thermographic images of a glass parallelepiped during free cooling in


air
4.2 Remote Spectral Temperature Profile Sensing 283

A comparison of the images shows very well the increasing temperatures


with increasing depth z. In addition, cooling can also be observed with respect
to the lateral temperature distributions. The corners and edges of the paral-
lelepiped have the lowest temperatures. The larger heat transfer towards the
base can also be visualized (compared to the heat transfer to the surrounding
air).
The images z = 36 mm and 39 mm display thermal radiation outside the
glass body. However, this combined scattering and reflection process (e.g.,
caused by the background) will not be discussed in detail here.
In the center of the image z = 39 mm, a hot spot appears (white arrow).
This is the radiation from a black-body radiator behind the glass sample
which radiates through the glass block for K,(>..)D :::::: 1 at this wavelength.
Although the glass geometry for this experiment is not one-dimensional,
the Improved Eddington- Barbier approximation can nevertheless be applied
locally. The temperature maps shown in Fig. 4.27 are based on this method.
In addition, Fig. 4.28 presents one-dimensional temperature profiles along
three different channels through the glass parallelepiped (parallel to the z
direction). A comparison between the profile at the center (A), the boundary
(B), and the edge (C) illustrates again the influence of the glass geometry
on the cooling process. In particular, the strong temperature decrease at the
corner (curve C for z :::; 3 mm) indicates fast cooling of this area, which in
general induces high thermal stresses and may lead to edge chipping.
Figure 4.29 shows the temperature distribution along a vertical scanning
line in the center of the parallelepiped (vertical to the z axis). The profiles at
the surface and at a depth of 36 mm are compared. Again, the nonsymmetry

--
800 ~----------------------------------~

A
._00- - - - 0 -0

700 .. 0
o
o
o
::; 600
QJ ____.o.-.-....--...-----.-O..- ..(]
B ....-_.__...-
'@

l~
QJ 0 -0

E 500
o
0

~
A
400 0

10 20 30 40
z /mm

Fig. 4.28. Temperature profiles along three different channels through the glass
block in the z direction. A: center, B: boundary, C: edge
284 4. Shaping at Low Viscosities

800
z =36 mm

700
0
l!...
~
:J
tii 600
Q;
C. <D
E ,E
~ c
500 ~
CJ)

I
400
0 10 20 30 40 50
y /mm

Fig. 4.29. Temperature profiles in the glass parallelepiped along a lateral scanning
line for z = 0.5 mm and 36 mm

of the profiles indicates the increased heat transfer from the glass towards
the base.
If the glass geometry is more complex than that of a parallelepiped, the
temperature profiles cannot be computed any longer from the intensity im-
ages according to Improved Eddington-Barbier. In the case of a 3D body,
e.g., the projection of the glass thickness observed by the optics is, in gen-
eral, locally different. A future 3D variant of the method must account for
this fact.
To stress this need, another problem is presented in Fig. 4.30. The cooling
of a glass cylinder standing upright was observed vertically to the cylinder
axis (here, only the surface temperatures are shown). A vertical and a hori-
zontal section through the thermo-image indicate the cooling behavior along
the cylinder axis and the influence of the angle-dependent emissivity of the
glass surface in the x direction. Because of the rotational symmetry of the
transport problem (the cooling by air is identical on all sides), one would
not expect any temperature variation along the x direction. However, the
lateral section through the thermo-image in the x direction (continuous blue
curve) shows a strong temperature decrease at the boundaries Ixl > 12mm.
The reason for this phenomenon is the increasing angle between the plane
normal of the glass surface and the observation direction [4.21]. If the angu-
lar dependence of the emissivity of hot glass is accounted for, this "virtual"
temperature decrease can be computed (dashed curve in Fig. 4.30). Apart
from a slight nonsymmetry of the experimental curve, the apparent temper-
ature profile along the x direction can now be explained. A similar effect has
already been observed in the example with the hot glass drop in Fig. 4.26.
However, the measured temperature profile along the cylinder axis (y direc-
4.2 Remote Spectral Temperature Profile Sensing 285

450
Theory (radial)

.......
0
400
~
:>
li1
Q)
a.
E
~ 350

300
-15 -10 -5 0 5 10 15
x, y/mm

Fig. 4.30. Temperature distribution at the surface of a free-cooling glass cylinder


(diameter 36 mm, height 35 mm) standing upright on a cold base

tion, red curve) is not influenced by this angle-dependent effect. Here, the
"real" temperature profiles along the y direction can be observed.
A further difficulty of spectral temperature sensing for three-dimensional
glass bodies are additional reflections at diagonal glass surfaces. Figure 4.31
shows a drastic example. During fast cooling in air, cracks caused by stresses
occurred in the interior of a glass slab with a hexagonal cross section. While
at the surface z = 0.5 mm these cracks cannot be observed yet, they can
be realized very clearly at the corresponding wavelength for z = 12 mm. The
surfaces of the fractures reflect thermal radiation from almost all directions in
the glass. These reflections have considerable influence on the thermographic
image. A correct determination of the temperature profile is now impossible.
A special feature of the thermo-image z = 0.5 mm is the registration of a
diagonal groove (1 mm deep and wide) at the surface. The temperature of the
glass in the groove exceeds that of the neighboring surface by approximately

z" 0.5 mm z " 12 mm

Fig. 4.31. Temperature


distribution in a 29-mm-
thick glass slab with a hexa-
gonal cross section (diame-
480 'c 330 'c
ter 55mm)
286 4. Shaping at Low Viscosities

5 K. This is because in the groove, the glass surface is cooled less by the
surrounding air.
The results of the previous studies of spectral temperature sensing clearly
show that a full 3D method has high potential and that the evaluation al-
gorithm for the computation of the temperature field from intensity images
will meet considerable challenges.

Summary and Conclusions

We have presented the method of spectral pyrometry for the remote sensing
of temperature profiles inside hot glass during hot forming. The thermal
radiation from the interior of the glass is used for the determination of the
temperature distribution. Hence, the inverse problem of spectral pyrometry
must be solved. The ID reconstruction algorithm presented here is able to
determine the respective temperature distributions from both numerically
generated and measured intensity spectra. Difficulties only occur in the case
of temperature gradients at the backside of the glass, due to the nature
of the ill-posed inverse problem. Preliminary numerical and experimental
approaches for the development of a 3D method (spectral thermocamera)
have been discussed.

4.3 Heat Transfer During Casting Experiments

Matthias Brinkmann, Thomas Korb


Introduction

In the previous sections, two different methods have been described for the
determination of the temperature distribution in glass during hot forming.
Mathematical modeling uses the system of differential and integro-differential
equations describing the heat transport inside hot glass in order to compute
time-dependent temperature profiles from known initial and boundary con-
ditions. Numerous material data are needed as further input parameters.
Although this system of equations can be solved nearly exactly with today's
numerical methods, deviations can occur with respect to the real process be-
cause of incorrect material parameters and/or wrong initial and boundary
conditions. Because no analytical solution exists for the heat transport in
hot glass, the corresponding numerical deviations can only be estimated by
sample computations.
The second method for determining the temperature distribution in glass
is a remote sensing technique using spectral pyrometry, which has been pre-
sented in the previous section. This method can also be subject to errors
that are due to inaccurate optical material data, as well as to the poor reso-
lution of the instrumental equipment in combination with the reconstruction
4.3 Heat Transfer During Casting Experiments 287

algorithm. Here, too, a classical error estimation is only possible in special


cases.
Because the two methods for determining temperature profiles in hot glass
are subject to errors, we would like to compare them in a special laboratory
casting experiment. Apart from the comparison, this experiment also illus-
trates again several complex properties of the heat transport process in hot
glass.

4.3.1 Experimental Set-Up

The laboratory experiment presented here deals with the casting of a crucible
melt (with an almost homogeneous initial temperature T) into a casting mold
of depth D = 13.5 cm. The inset in Fig. 4.32 sketches this process. For better
comparability of the simulated and measured results, the same experiment is
performed with two glasses having almost identical material parameters (not
presented here) and very similar thermophysical material parameters, except
for their optical transmission (transparent glass: K, ~ 5 m -1, semi-transparent
glass: K, ~ 15 m- 1 ). After the casting, the glass slabs cool down, transferring
heat into the mold as well as into the air above. Because the glass slabs
are evidently wider than they are thick, the heat transport problem can be
considered as one-dimensional in a first approximation.
During the cooling phase, the glass slabs are observed with the spectral
pyrometer, which is almost vertically aligned from above. Thus, the time-
dependent development of the temperature profiles is recorded. As can be
seen in Fig. 4.32, the absorption spectra of the two glasses differ significantly
for wavelengths less than 2.5 ~m.

1000
500
Casting Spectrometer
1>0s
II
I" 100
~
'< 50

10 = 10 cm Transparent
5

1.5 2.0 2.5 3.0 3.5 4.0 4.5


A/lJm

Fig. 4.32. Wavelength dependence of the absorption coefficient of two glasses


(transparent: K, ~ 5m-1, semi-transparent: K, ~ 15 m- I ) during casting of a cru-
cible melt
288 4. Shaping at Low Viscosities

The optical absorption of a glass determines the influence of radiation on


the heat transport. For the modeling, however, the phononic part of thermal
conductivity k phon is also necessary.
In [4.9] and [4.28]' a comparative heat flow method for determining the
apparent thermal conductivity is described. From the experimental results
of this equipment the phononic contribution to the thermal conductivity can
be extracted. For the transparent glass, this measurement was carried out
between 100 and 900 DC. The phononic thermal conductivity was computed
using the apparent data for the thicknesses 25 mm and 40 mm. Figure 4.33
shows the result. The difference between the apparent (red curve) and the
phononic conductivity (blue curves) represents the radiative part of the ap-
parent thermal conductivity. Only this contribution depends on the glass
thickness.
A linear increase from 100-500 DC can be realized for the phononic ther-
mal conductivity. At higher temperatures, this material parameter remains
temperature-invariant at a value of approximately 1. 7 W /K m, which is used
for the numerical simulation at temperatures higher than 900 DC.
A further critical parameter is the experimental casting time of approxi-
mately 10-15 s, which can vary slightly for each experiment (casting by hand).
It is therefore difficult to define an initial time t = 0 where the casting process
is finished and convection inside the melt has stopped. This problem will be
discussed in detail in the following section.

4.5

4.0

3.5
I"
E 3.0
:.::

--
!i: 2.5
-""

2.0

1.5

1.0
0 200 400 600 800
Temperature I·G

Fig. 4.33. Apparent and phononic thermal conductivity of the transparent glass.
Squares: additional results from another instrument that is only applicable up to
glass transition temperature Tg
4.3 Heat Transfer During Casting Experiments 289

4.3.2 Comparison Between "Exact" Modeling and Measurement

Figure 4.34 shows the result of the simulation for a 13.5-cm-thick glass plate
with an initial temperature of 1500 ac. At this slab thickness, the interior
of the semi-transparent glass is cooling faster than the transparent glass,
whereas the glass surfaces have more or less the same "cooling speed". This
behavior shows a special property of the heat transport in hot glass: due to
the higher absorption coefficient K, = 15m- 1 = (6.7cm)-1 :::::: (D/2)-1, more
thermal radiation is produced in the interior of the semi-transparent glass
for A ::; 2.5!-Lm than in the transparent glass with K, = 5 m -1 = 20 cm -1.
Because of the greater amount of thermal radiation, more heat is extracted
from the semi-transparent than from the transparent glass.
The profiles shown in Fig. 4.34 are symmetric with respect to the center
plane, although the bottom surface of the glass slab (z = 0.135 m) touches
the mold. Hence, there ought to be a larger heat transfer from the bottom
of the slab into the mold than from the surface to the air above. However,
this potentially stronger heat transport is intentionally suppressed by a small
mold thickness and a good insulation between mold and mold bearing.
The heat transition through the bottom of the mold cannot be modeled
here. However, one can estimate that, after the initial heating of the mold
(:::::: 20 s), more or less the same heat escapes from the bottom and from the
top of the glass. Therefore, a symmetric transport problem was computed in
order to simplify the situation. Besides, the comparison between simulation
and measurement only refers to the area from the top glass surface (z = 0)
to the center (z = 0.06 m), so that the details of the cooling at the bottom
have no significant influence on the result.

Os
1400
10 s
25 s
u 1200
!!.... 50s
~
:J 100 S
1000
~
Q)
a. 250 s
E Transparent
800
~
Semi-transparent 500 s
600
1000 s

400 2000 s
0 0.02 0.04 0.06 0.08 0.10 0.12
z /m

Fig. 4.34. Modeling of the temperature distribution in the casting experiment -


comparison between both glass types
290 4. Shaping at Low Viscosities

Figure 4.35 shows the comparison between numerical simulation and mea-
surement at t = 250 s. The evaluation of the measured spectra according to
Improved Eddington-Barbier (squares, see previous section) yields a temper-
ature profile from the top glass surface to the center. Although the "exact"
reconstruction (dashed lines) yields complete temperature profiles, these do
not represent the temperature gradient at the backside of the glass slab cor-
rectly (see previous section).
Figure 4.35 clearly proves the correspondence of the temperature profiles
from 0 up to 0.06 cm for both glass types, i.e., the temperature-profile sensing
during the experiment uniquely confirms the prediction of the modeling. For
almost identical surface temperatures, the center of the semi-transparent glass
slab is cooling faster than the interior of the transparent glass.
Figure 4.36 compares the modeled and measured time-dependent devel-
opment of the temperature profiles. Again, simulation and experiment corre-
spond well with respect to the time dependence of surface and center temper-
atures. However, a detailed examination of the data shows some deviations.
Here one finds the limits of exactness for both methods. Notwithstanding
this fact, in summary we can state that the results of this experiment have
proved the efficiency of both methods in describing thermal processes during
hot forming.

4.3.3 Alternative Modeling Using the


Active Thermal Conductivity
In [4.9]' the concept of "active thermal conductivity" for free cooling has been
presented. Accordingly, two types of temperature and thickness-dependent

1150 t = 250s
.. -------.
"...... --·- Measurement

1100
"
,," a
,' Measurement
C!!, __
- •••
.. -.
I. ,"' ~
o
e... 1050 .. ..
:J
"§Q) 1000
a.
E
~ 950

900

850
o 0.02 0.04 0.06 0.08 0.10 0.12
z/m

Fig. 4.35. Comparison of the simulation results (continuous curves) and the ex-
perimentally determined temperature distributions at the time t = 250s (squares:
Improved Eddington-Barbier, dashed lines: reconstruction). Red curves: semi-
transparent glass, blue curves: transparent glass
4.3 Heat Transfer During Casting Experiments 291

1400 300
'I
E ~
(,) 1200 E Ql
e.. ~ 200 ~
Q)
E 't:

~ Q) :J

Q)
1000 i5 30
e! !
Q)
0.. OJ
E Q) 100 ~
~ 800 o 20 o
oCalculation ~ i="
- Experimental :J <I
C/)
600 10
o
Semi-transparent Semi-transparent
400 +....--~.........--~.,....,...---i-I
5 x 10 1 5 x 102 5 x 103 5 X 102 5 x 103
Time Is Time Is

Fig_ 4_36_ Comparison between modeling and measurement of the time-dependent


surface and center temperature, surface gradients and temperature differences
( center-surface)

thermal conductivities can be defined for one-dimensional problems: the "ap-


parent thermal conductivity" kapp(T, D) describes the heat transport for heat
transition through hot glass (which can be applied, e.g., at melting devices).
In this case, thermal radiation from the hot boundary walls can penetrate
the glass without interacting with the material. The "active thermal conduc-
tivity" kact(T, D) represents the free cooling of glass in a cold environment
(e.g., the casting experiment described above). In this case, the transport
process only includes that part of the thermal radiation that is generated in
the semi-transparent wavelength range.
A differentiation rule describes the relation between both conductivities:

(4.62)

kact(T, D) is always smaller than kapp(T, D). For the active thermal con-
ductivity, kact(T, D) represents the combined heat transport (phononic and
radiation) in the heat transport equation without accounting for an explicit
radiative heat flow.
This approximation is now applied to the casting experiment. Figure 4.37
shows the thickness and temperature dependence of kact for both glass types.
At typical hot-forming temperatures, there is an interesting cross-over at
D = 0.2 m between both glasses: for D < 0.2 m, the active thermal con-
ductivity of the semi-transparent glass is larger. The casting experiment for
D = 0.135 m (green arrow in Fig. 4.37, indicated by A) confirms this result
experimentally as well as the "exact" modeling. For D > 0.2 m = 1/ Ii =
(5 m- 1)-1, the active thermal conductivity of the transparent glass is larger,
292 4. Shaping at Low Viscosities

C 102
.Q
lli 5
'C
[!!
>-
c:
.£. 10'
'"j 5
E
::.:::
~ 100
t5 5
-'" '"
10- 1
10-2 2 4 6 8 10-2 2 4 6 810-{J 2
D Im

Fig. 4.37. Thickness and temperature dependence of the active thermal conduc-
tivity for both glass types. The typical (T,D) process parameter ranges for melting,
hot forming and cooling are also indicated. The green arrows indicate the cooling
process of the casting experiments considered here

because the transparent glass is constantly becoming less transparent for


A :::; 0.25 Il-m with increasing glass thickness. The amount of thermal radia-
tion generated inside the transparent glass increases, as does the respective
part of the heat transport. Additionally, the ratio between the radiative mean
free path (6.7 em) and the thickness becomes more and more inconvenient for
D > 0.2 m in the semi-transparent glass. The radiation from the interior is
increasingly absorbed on the way towards the surface. Figure 4.38 shows the
modeled time dependence of the temperature distribution in a glass slab with
thickness D = 25 em (initial temperature 1000 °C, green arrow B in Fig. 4.37)
for both glass types. As a matter of fact, the center of the transparent glass
slab cools faster (in contrast to D = 13.5 em, Fig. 4.34), i.e., the heat trans-
port through radiation in the semi-transparent glass is more effective now.
The surface temperature of the transparent glass slab is higher than that of
the semi-transparent glass, because the amount of heat transported from the
interior to the surface is considerably larger in the transparent case.
Although in this case the active thermal conductivity describes the "real
behavior" of the transport process correctly, it is in general not convenient
to compute the temperature profiles by directly inserting kact(T, D) into the
heat transport equation. Particularly, the sharp temperature gradients at
the glass surfaces (Fig. 4.34 and Fig. 4.38) are not described correctly [4.10].
However, the active thermal conductivity is a very useful parameter for cal-
culating global heat balances.
4.4 Thin Layer Flows of Glass 293

1 s
5s
10 s
20 s
30 s
U
!!...
Q)
:J Transparent
'§ 950
Q) Semi-transparent
a.
E
~

900 ~~----.-------'------'r------'-U~
o 5 10 15 20
z /cm

Fig_ 4.38. Modeling of the temperature distribution in 25-cm-thick glass slabs of


the transparent and semi-transparent type, initial temperature 1000°C

Summary and Conclusions

Two methods for the determination of the thermal behavior of hot glass dur-
ing hot forming have been successfully tested in a laboratory casting experi-
ment. The results of the simulation and the experimental data from spectral
pyrometry are in good agreement. Hence, both methods provide new possi-
bilities for the thermal analysis and optimization of hot-forming processes in
industrial glass making.

4.4 Thin Layer Flows of Glass

Ulrich Lange
Introduction

The fluid mechanics in the hot-forming zone of many important glass manu-
facturing methods can be characterized as a highly viscous "thin-layer flow":
The thickness of the glass is typically much less than the other geometrical
dimensions. Typical examples are the float process and other methods for the
manufacture of plate glass, such as the overflow-fusion or the slot downdraw
process (see, e.g., Hynd [4.51] for a description of these processes). The spin-
ning of optical fibers, blow molding of hollow glass as well as the final stage
of pressing processes also fall into the category of thin layer flows.
Because the shape of the hot-forming zone of those processes is very sen-
sitive to changes of the operating conditions (e.g., the cooling rate), numer-
ical solvers for the full system of Navier-Stokes equations for those three-
294 4. Shaping at Low Viscosities

dimensional free surface flows often exhibit a rather shaky convergence behav-
ior. Moreover, these numerical methods tend to be time-consuming because
of the high aspect ratio between thickness and length of the flow domain:
A large number of nodal points are necessary for the discretization of such
domains in order to avoid finite elements with high aspect ratios and their
well-known numerical weaknesses (see, e.g., Braess [4.48] on "h:-regularity").
The drawbacks of the direct numerical methods can be avoided by us-
ing model equations, which are tailored for thin layer flows, rather than the
full Navier-Stokes equations. A surprisingly large number of fluid mechani-
cal problems relevant to applications in the glass industry can be efficiently
treated with such simplified models, without significant losses of accuracy. In
the present section, we will discuss the basic concepts of the thin layer flow
models and present some simple examples of applications.
There are two different ways of deriving the simplified equations: The
first one exploits the heuristic expectation that variations of the flow field
in the thickness direction should not be relevant because of the high aspect
ratio. Hence, one may consider mass, momentum, and energy balances that
are averaged over the thickness of the glass sheet. (In Sect. 3.2, even pipe
flows are treated as thin layer flows by this averaging method.) The sec-
ond method is based on an asymptotic expansion with the inverse aspect
ratio as a small parameter. This approach is less intuitive, but more rigor-
ous from the mathematical point of view. Moreover, extensions of the basic
model (e.g., appropriate inclusion of inertia and surface tension effects) are
more straightforward in the framework of asymptotic methods. A detailed
discussion of many aspects of this approach can be found in Howell [4.50].
In order to illustrate the main ideas of the asymptotic methods, we will
derive simplified model equations for the flow in the hot-forming zone of the
slot downdraw process in Sects. 4.4.1 and 4.4.2. Some examples of useful
design formulas for the process are derived from these model equations. The
asymptotic methods provide a valuable tool for the systematic derivation of
more comprehensive simplified models. As an example, we derive a formula
for the difference between core and surface temperature of the glass layer by
evaluating higher-order terms of the asymptotic expansion.
In Sect. 4.4.3, the predictions of the simplified model are compared to
numerical solutions of the full Navier-Stokes equations. Extensions of the
simplified model to fiber- and tube-drawing applications are presented in
Sect. 4.4.4. Finally, we present the prediction of the thickness distribution in
a plate glass sheet manufactured by the slot downdraw process as an example
of the numerical solution of the thin layer equations in a three-dimensional
situation.

4.4.1 Example of a Thin-Layer Model

In order to avoid tedious algebra as much as possible, we restrict ourselves


to the special case of a two-dimensional thin-layer flow, which is symmetric
4.4 Thin Layer Flows of Glass 295

with respect to the y axis (see Fig. 4.39). Moreover, we neglect surface ten-
sion effects. An example of such a flow is the hot-forming zone of the slot
downdraw process, also shown in Fig. 4.39, which is applied in the production
of ultra-thin plate glass. The starting point of our derivation are the general
mass , momentum, and energy balances for the flow:
\7u=O , (4.63)
Du
P Dt = \7 T + P9 , (4.64)

DT
pCp Dt = \7 ()"(T)\7T) , (4.65)

T = -pI + 'r/(T)(\7u + \7u T ) . (4.66)


Here, u = (v, w) T denotes the velocity vector and T is the stress tensor. In
the material law (4.66), the unit tensor is denoted by I, p is the pressure,
and 'r/(T) is the viscosity (strongly dependent on the temperature T in the
case of glass). Further material properties occurring in the equations are
the density p, the conductivity )..(T) , and the specific heat Cpo At the free
surfaces y = ±D /2, the kinematic condition must hold and the normal and
tangential stresses must vanish. Furthermore, we assume that heat transfer
from the surface to the surroundings is dominated by radiation:
laD waD
27ft + 2& = TV (4.67)

tT Tn =0 (4.68)

(4.69)
O"E 4 4
n \7T = T )..(T) (T - Tcx,) . (4.70)

The ambient temperature is denoted here by Too. The emissivity of the free
surfaces is c, a is the Stefan-Boltzmann constant, and t , n denote the tan-
gential and normal unit vectors to the free surfaces.

Hot-forming z
zone
L

Annea linir·· ..
zone
Drawing
"' r
rolls
Fig. 4.39. Cross-sectional view of the
slot downdraw process
296 4. Shaping at Low Viscosities

One may non-dimensionalize the above equations based on the length L


of the layer in the z direction, a characteristic value U for the flow velocity
(e.g., the drawing speed) and a characteristic temperature e (e.g., the die
temperature). Because we are considering thin-layer flow here, i.e., we assume
that

(4.71)

where DAis a typical length scale of the layer thickness (e.g., the die width),
the appropriate scaling for the dimensional quantities is given by

z = Lz' , y=8Ly', D=8LD', t = b.t'


U '
_ UrMl ,
w=Uw' , v = 8U v' , P- -L-P , T=eT' , (4.72)
Too = eT!:xo , 71 = ry(e) r/ , A = A(ep' .

If we apply this scaling to the equations, we get


Vy + W z = 0, (4.73)
Re 84 (vt + VVy + wVz ) = -Py + 2(ryvy)y + (rywy)z + 82 (ryvz)z , (4.74)
Re 84 (wt + VWy + wWz ) = -8 2 pz + 28 2 (ryw z )z + (rywy)y + (4.75)
+ 8 (ryvz )y + 8 St ,
2 2

RePr 8(Tt + vTy + wTz ) = (ATy)y + 82 (ATz )z . (4.76)


(In order to keep the notation simple, we omitted the primes denoting the
dimensionless quantities and abbreviated partial derivatives using subscripts
(fy := al jay).) The three dimensionless numbers occurring here are defined
by

pUD L 3
Re = ~ 8- (modified Reynolds number)

P = cpry(e) 83 (modified Prandtl number) (4.77)


r A(e)
pgL2
St = ry(e)U (Stokes number)

We have multiplied the usual Reynolds and Prandtl numbers by factors of


8- 3 and P, respectively. (This is motivated by the fact that the order of
magnitude of the modified numbers is 1 for typical glass applications with
8 ~ 0.01. Thus the order of magnitude of all the terms in the scaled equations
is exclusively given by their respective powers of 8.) The scaled boundary
conditions at y = ±Dj2 are
4.4 Thin Layer Flows of Glass 297

1 W
"2Dt + 2Dz = =fV , (4.78)

-P + 2T]vy ± ~Dz (8 2 vz + w y) = 0 , (4.79)


82
T](8 2 vz +wy) =f 2: Dz(-p + 2T]w z ) = 0, (4.80)
82 82
)..(Ty =f 2:Tz) ± B 8(1 ± 4D;)1/2(T4 - T!) = 0 . (4.81)

There is an additional dimensionless number occurring in the radiation con-


dition:

(4.82)

We will consider only the pure mechanical problem (4.73)-(4.75) with bound-
ary conditions (4.78)-(4.80) for the moment. Note that there are only even
powers of the small parameter 8 in these equations. Hence, we may substitute
asymptotic series of the type

(4.83)

for the unknowns v, W,p, and D into the mechanical problem and determine
a hierarchy of equations for the coefficients Vi, Wi,Pi, Di (i = 0,2,4, ... ) of
the series by equating like powers of 82 . The zeroth order of this hierarchy is
VOy + wOz = 0, (4.84)
0= POy + 2(T]voy)y + (T]WOy)z , (4.85)
0= (T]Woy)y . (4.86)
The zeroth-order boundary conditions are found to be
1 Wo
"2Dot + 2Doz = =fvo , (4.87)
1
-Po + 2T]voy ± "2T]woyD oz = 0 , (4.88)
T]WOy = 0 . (4.89)
By a Taylor expansion of the original boundary conditions (4.78)-(4.80) one
finds that it is consistent to pose these conditions at y = ±Do/2. From (4.86)
and (4.89), we conclude that Wo must be independent of y. Hence, from (4.84)
and (4.87), we get the following equations for Vo and Do:
Vo = -WOz Y , (4.90)
Dot + (woDo)z = 0 . (4.91)
Finally, by substitution of (4.90) into (4.85) and (4.88), the zeroth order of
the pressure is found to be

Po = - 2T]woz . (4.92)
298 4. Shaping at Low Viscosities

Thus, if we knew wo, we could easily determine Vo,Po, and Do from (4.90)-
(4.92). But how do we find wo? We used up all zeroth-order equations and
only found that it must be independent of y. The crucial step in the determi-
nation of Wo is the observation that a series solution can only exist if certain
solvability conditions are satisfied: The second order of the z-momentum
equation is
(ryW2y)y = -(4rywoy)z - (ryvoz)y - St , (4.93)
and the second-order boundary conditions at y ± D o/2 are
ryw2y = ±2rywozDoz - ryvoz . (4.94)
Equation (4.93) with boundary conditions (4.94) is only solvable, if
(4TjwozDo)z = -St Do . (4.95)
This compatibility condition, which can be found by integrating (4.93) once
with respect to y, corresponds - in mathematical terms - to the so-called
Fredholm alternative. It is the missing equation for the determination of woo
Note that Tj denotes the cross-sectional average of the viscosity, because the
viscosity may depend on the y coordinate via the temperature field.
If we return to the dimensional variables, (4.92) and (4.95) become
Dt+(wD)z=O, (4.96)
(4Tj wzD)z = -pgD . (4.97)
This equation system, the so-called Trouton model, can now be used to cal-
culate the local vertical velocities wo(z) and the sheet thickness Do(z) in
the hot-forming zone of the slot downdraw process. Various useful design
formulas can be derived from the Trouton equation by further simplifying
them.
For example, note that in the framework of the Trouton model the vertical
tension in the glass layer is given by T(z) = 4Tj(z) wz(z). From (4.97), we
derive

T(z) = T(O) - pg 1 ~~;~


z
d( . (4.98)

U sing an extension of the asymptotic analysis presented here, it can be shown


that a viscous sheet under compression will buckle (see [4.49]). Thus, (4.98)
provides a stability criterion for the drawing process,

T(O) - pg
r
Jo
D(()
D(z) d( > 0 , (4.99)

which can be easily evaluated from the solution of the model equations (4.96)
and (4.97). If a desired process state (i.e., drawing speed, die width, flow rate,
and cooling rate) is prescribed, the inequality (4.99) can be interpreted as a
condition for the maximum feasible length of the sheet (from the die to the
drawing rolls).
4.4 Thin Layer Flows of Glass 299

4.4.2 Simplified Energy Balance

Using asymptotic analysis, we can also derive a simplified model for the
temperature field in the glass sheet, which we had assumed to be given so
far. Now we assume a given layer thickness and a given velocity field (v,w).
The vertical velocity component w is assumed to be independent of y. We
substitute a series of the type

(4.100)

into the energy balance (4.76) and its boundary conditions (4.81). Equating
like powers of Ij yields the following zeroth- and first-order equations:
o=
(AoToy)y , (4.101)
RePr(Tot + vTOy + wToz) = (AoTly)y + (AlToy)y . (4.102)
Here, we used the abbreviations AO := A(To) and Al := A'(To)Tl. The bound-
ary conditions for To and Tl at y = ±D/2 become
AoToy = 0, (4.103)
AoTly + AlToy = ~B(T6 - T!,) . (4.104)
From the zeroth-order equation (4.101) with boundary condition (4.103) we
conclude that To must be independent of y. Then, integrating the first-order
equation (4.102) once with respect to y yields the solvability condition

RePr(DTot + wD· Toz) = -2B(T6 - T~J , (4.105)

which determines To. By coupling (4.105) with the Trouton model we obtain
a simple model for the description of non-isothermal thin-layer flow. A draw-
back of the simple model is that it does not predict temperature variations
across the layer thickness, which may be important if the surface of the layer
is heavily cooled. The model can be extended by considering the higher or-
ders of the equation hierarchy. From the first-order equation (4.102), we get
the following formula for Tl by two integrations with respect to y:

(4.106)

We have used the fact that To must be independent of y here. The integration
constant c = c(t, z) can be found via the second-order solvability condition,
but also by the following heuristic shortcut: It is plausible to assume that the
y-independent zeroth-order solution To corresponds to the temperature aver-
aged across the layer thickness. Then the average of the first-order solution
Tl must vanish:
300 4. Shaping at Low Viscosities

(4.107)

Using this condition, the constant c can be easily determined and we get

T1 = B (T.04 _ T4 )(D02 _ 12y2) . (4.108)


12.\oDo oc

Thus, in each cross section of the layer, the temperature profile is approxi-
mately a parabola and the difference tlT between core and surface temper-
ature can be estimated by

tlT = (j ~(T.4
4.\0 0
- T4)D .
CXl 0
(4.109)

If we return to dimensional variables, (4.105) and (4.109) become

pCp(DTt + wDT z ) = -2uc(T4 - T!,) (4.110)


tlT = uc~ (T4 _ T4 ) . (4.111)
4'\(T) 00

(Here and in the following, we have denoted the zeroth order of the temper-
ature by T in order to emphasize that it is the cross-sectional average of the
temperature.) These equations extend the Trouton model for the hot-forming
zone of the slot downdraw process. The cooling by surface radiation can now
be taken into account appropriately. Again, simplified design rules can be
derived from these model equations:
For example, if we assume a stationary process (T t = 0), we conclude
from (4.110) that the average temperature depends on the flow field only
via the product wD. Hence, if we change the drawing speed, but keep the
flowrate constant, the average temperature field in the hot-forming zone does
not change! (At least not significantly. There are, of course, the higher-order
corrections to our asymptotic model.) The estimate for the difference between
core and surface temperature of the glass layer given by (4.111) yields another
important design rule. Of course, the reliability of this estimate depends
strongly on how well the emissivity parameter c of the glass surface and the
(effective) conductivity .\ model the combined heat transport by conduction
and radiation inside the semi-transparent glass.

4.4.3 Validation of the Model

As a test case for the Trouton model, we consider a two-dimensional glass


sheet that is drawn at a speed of WE = 20 mm s-l from a slot die with a
width of D A = 10 mm. The final sheet thickness should be DE = 1 mm. A
constant ambient temperature of Too = 750°C is prescribed. In Figs. 4.40 and
4.41 the predictions of the Trouton model for this situation are compared to
the solution of the full Navier-Stokes equations obtained by the commercial
4.4 Thin Layer Flows of Glass 301

10 ~------.-----------~~~~~--~ 20

8 16 ~
E b
.§ - - - Polyflow, layer thickness E
(J)
(J) 6 - - - Trouton model. layer thickness 12 ..§
Ql
C - -- Polyflow, vertical velocity -~
o
.><: o
0 - - - Trouton model , vertical velocity a;
:£ 4 8 >
CiJ Cij
>. o
co -,;::;
-.J CiJ
2 4>

o 0
o 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08
Vertical distance from die exit 1m

Fig. 4.40. Comparison of POLYFLOW and Trouton model predictions: shape of the
hot-forming zone and vertical velocity

1000 80
-- Polyflow, core temperature
975 - Polyflow, surface temperature 70
-- Trouton model, mean temperature
-- Polyflow, l\T
950 Trouton model \ T 60
()
~ 925 50 ~
~ ~
:l :l
(jj 900 40 (jj
CiJ CiJ
n. n.
E 875 30 E
~ ~
850 20

825 10

800 0
0 0.01 0.02 0 .03 0.04 0.05
0.06 0 .07 0 .08
Vertical distance from die exit 1m

Fig. 4.41. Comparison of POLYFLOW and Trouton model predictions: mean tem-
perature and difference between core and surface temperature

solver POLYFLOW [4.52]. Except for the first few millimeters after the die, the
two approaches are in very good agreement in their predictions of the hot-
forming zone as well as the velocity and the mean temperature field. Also,
the first-order correction (4.111) to the Trouton model, which predicts the
temperature difference 6.T between core and surface of the sheet, agrees well
with the POLYFLOW results. However, the agreement becomes worse as one
gets closer to the die. This is due to the fact that the model based on the
asymptotic limit (j --t 00 is not appropriate in the vicinity of the die, where
the flow field must reorganize from the Poiseuille-type flow inside the die
302 4. Shaping at Low Viscosities

to the "plug flow" described by the Trouton model. In mathematical terms,


the Trouton model is the only outer expansion of a singular perturbation
problem. A rigorous construction of a uniformly valid model would require
an additional inner expansion (see e.g. [4.54] for the method of matched
expansions). For many practical purposes, however, the Trouton model is
sufficiently accurate and it is not necessary to know details of the transition
region downstream of the die.

4.4.4 Fiber- and Tube-Drawing Models

Applying the methods outlined in Sect. 4.4.1 to the Navier-Stokes equations


in cylinder coordinates, we obtain a simple model for the drawing of (axisym-
metric) fibers:
At + (wA)z = 0 , (4.112)
(317Awz)z = 0 . (4.113)
Here, A = 7r R2 denotes the cross-sectional area of the fiber, w its velocity
in the drawing (z-) direction and 17 is the cross-sectional average of the
local viscosity. Comparing (4.113) and (4.97), one notes the "Trouton ratio"
is 3 instead of 4 in the tension term: Stretching an axisymmetric fiber is a
little easier than stretching a two-dimensional sheet. Because the fiber can
shrink in two directions, its radius is decreased less than the thickness of
a two-dimensional sheet stretched by the same amount. Hence, less viscous
resistance to deformation must be overcome.
As a simple application example for the fiber-drawing model, we study a
fiber re-drawing process, in which the cross-sectional area ofthe preform is 10 4
times larger than that of the final fiber. Figure 4.42a shows the devolution of
the (non-dimensionalized) cross-sectional area for three different temperature
profiles in the hot-forming zone. The prescribed (non-dimensonalized) tem-
perature profiles are shown in Fig. 4.42b. One notes that shape changes more
rapidly and the necessary pulling force (non-dimensionalized values given in
the legend) increases as the heating zone becomes narrower. Because of the
high aspect ratio, the numerical solution of the full Stokes equations for this
problem would be a very challenging task.
The model equations (4.112) and (4.113) can also be used to study the
basic mechanisms of the drawing of fibers with non-circular cross sections A
(see, e.g., [4.53]).
The extension of the fiber-drawing model to a model for (axisymmetric)
tube-drawing processes is straightforward, although the additional algebra is
quite cumbersome: In the tube case, there is a second (inner) free surface,
and one has to take into account that the pressure at this inner surface may
be different from the outer pressure, because this is a common means to
achieve different combinations of the tube diameters and wall thicknesses.
The resulting model equations are
4.4 Thin Layer Flows of Glass 303

1.0000 '---~=
, .. - ;;=_ ""' .. """
.;.;;:- . '. -
.;.:-< ;;;-
;;;""".~,.=
,..=.;.=';.='. --~==~-.., .. a)
....•. _.•. ,' ': .• - Case l ' Pulling force = 8 "
Case 2: Pulling force = , ,
- Case 3: Pulling fprce = 16
<U
Q)
0.1000
iii
<ii
c:
.2
13 0.0100
Q)
fJ)

'"
fJ)
0
0 0.0010

0.0001
-0.3 -0.2 -0.1 0 0.1 0.2 0.3
Drawing direction

1.0
Case 1 - - b)
Case2--
0.9 Case 3 - -

~ 0.8
~
:>
~ 0.7
<I.l
a.
E
~ 0.6

0.5

0.4
-0.4 - 0.2 0 0.2 0.4
Drawing direction
------ c)
-------
E
.E
~
0..

l'!®
Q)
®

I
®
® ®
iii
.0
u::

Fig. 4.42. Drawing of a fiber from a pre-


form
304 4. Shaping at Low Viscosities

(4.114)

(4.115)

(4.116)

In these equations, R is the mean radius and D is the wall thickness of the
tube (see Fig. 4.43). The pressure inside the tube, which may vary in the
drawing direction, is denoted by P. As usual, w is the velocity in the drawing
direction, 'ij is the cross-sectional average of the local viscosity, and p is the
glass density. The pgRD term is, of course, only relevant in the case of a
vertically drawn tube (Vello process) . If this term is cancelled, the model can
also be used to obtain many useful insights into the mechanisms of horizontal
tube drawing (Danner process), even though the axis of the tube in the hot-
forming zone of the Danner process is, of course, not really straight.
It is interesting to analyze these equations for a stationary process. In this
case, (4.114) yields that the flow rate wRD is constant. Hence, if one divides
(4.116) by wRD, one obtains

(4.117)

From this equation, it is obvious that the ratio of the tube radius to the tube
diameter remains unchanged in the hot-forming zone if no inner pressure is
applied. As an application example, we consider a tube-drawing process with
a prescribed flow rate from a mandrel with an end diameter of 200 mm. At
the end of the mandrel, the glass layer is 100 mm thick. Suppose we want
to draw a tube with a mean radius of 30 mm and a wall thickness of 5 mm.
Given the flow rate, one can easily determine the appropriate drawing speed
from (4.114). However, if we do not apply an inner pressure, the ratio RID
remains constant throughout the hot-forming zone. Thus, we end up with a
smaller tube with thicker walls than desired, as is illustrated by Fig. 4.44.
U sing the model, the appropriate inner pressure can be calculated, and the

I
I

I I
I I
I I
I I
I I
I I

,
I I
\
I I
;";" ". ....
\ \ ..t..-,_-
I' ....,/
,/
/
\ I
Fig. 4.43. Drawing of an axi-
, .Jt ' _ ....
symmetric tube
4.4 Thin Layer Flows of Glass 305

200 ~----------------------------------~

180
• No Inner pressure - -
160 Applied inner pressure: 104 Pa - -
140

~ 120
~ 100
'i5
/:. 80 . T
60
40 Case 2: RF"onal= 30 mm, DF"InBI= 5 mm

20 ase mm. mm
O+---------r--------.r--------.--------~
o 0.5 1.0 1.5 2,0
Distance from mandrel 1m

Fig. 4.44. Thbe drawing: calculated shapes of the hot-forming zone with and
without applied inner pressure

resulting shape of the hot-forming zone is also shown in Fig. 4.44. Of course,
we would get different values for the appropriate inner pressure if we changed
the cooling of the tube in the hot-forming zone.

4.4.5 More Comprehensive Thin-Layer Flow Models

The extension of the above considerations to the three-dimensional situation


is straightforward. The mass and momentum balances result in the following
system of equations:
D t + (uD)x + (wD)z = 0 , (4.118)
(7] D(4u x + 2w z ))x + (7]D(u z + wX))z = rhs 1 , (4.119)
(7] D(uz + wX))x + (7] D(2ux + 4wz))z = rhs 2 . (4.120)
In these equations, the two in-plane coordinates are denoted by (x, z), and
w = (u, w)T is the in-plane velocity vector. For the right-hand side vector
rhs, we have to consider two different cases: For applications like the slot
downdraw or overflow-fusion process, the gravity vector 9 is in the plane of
the sheet, and we have

rhs = pgD. (4.121)

For the float process, the gravity vector is perpendicular to the plane of the
sheet and we have to take into account buoyancy forces due to the tin bath.
Then the right-hand side results in

rhs = pg (1 - L) D VD ,
Ptm
(4.122)
306 4. Shaping at Low Viscosities

where Ptin is the tin density. It should be mentioned that the centerline S of
the glass sheet can no longer be assumed to be a straight line for the float
process, which results in some additional technicalities in the derivation. The
local position S = S(x, z) of the centerline is found to be dominated by
the balance of buoyancy of the tin and the weight of the glass and can be
described by the following formula, if the tin-air interface is located at y = O.

S= (~-
2
L) D.
Ptin
(4.123)

A solver for (4.118)-(4.120) is implemented in the commercial package


POLYFLOW [4.52]. With this tool, three-dimensional effects in the hot-forming
zone can be studied at the expense of solving a two-dimensional problem. As
an example, we examine the influence of the temperature boundary conditions
on the sheet thickness distribution for a slot downdraw process. We consider
a sheet of glass that is drawn at a given speed from a die shaped like a flat
rectangle. We assume a uniform flow-rate distribution across the die width.
We distinguish three different cases for the temperature boundary conditions,
which are illustrated on the left half of Fig. 4.45: In the simplest case, we
assume that the glass exits the die with a uniform temperature distribution.
It is then cooled by radiation against a uniform ambient temperature Too. In
the second case, we assume that the inlet temperature decreases near the ends
of the die. In practice, this could be achieved approximately by cooling the
small sides of the glass tank above the die. In the third case, we assume that -
in addition to the non-uniform inlet temperature - the ambient temperature
Too for a small spot close to the die is 200 DC lower than for the rest of the
sheet. In practice, this could be achieved by placing a small cooled object
very close to the sheet surface. In the upper right quarter of Fig. 4.45, the
shape of the sheet calculated for case 1 is shown. The contour lines show the
calculated variation of the thickness distribution. One notices that the sheet
is thinning very rapidly after the die except at the edges, where bulb edges
are formed. The final thickness distribution in the sheet, which is reached
already at a very short distance from the die, is shown in the lower right
quarter of Fig. 4.45: For case 1, the sheet thickness is not very uniform. The
inlet temperature profile prescribed in case 2 leads to cooler edges. Because
the stiffer edges span the sheet, this results in a better thickness distribu-
tion. The cold spot close to the die in case 3 causes a localized hump in the
thickness distribution. In practice, this effect is exploited by devices for fine
tuning the thickness distribution.
It is also possible to derive simplified flow equations for arbitrarily curved
thin layers that are suitable for the investigation of hollow glass manufactur-
ing processes. We refer to the work of Howell [4.50] for a discussion of these
"viscous shell" equations.
4.5 Pressing of Drinking-Glass Stems 307

See small diagram below for


inlet temperature distribution
[1

/
Case 3: lower Too
in this spot Red. high thickness
Blue: low thickness

Uniform Too
in cases 1 and 2

Temperature boundary Calculated sheet-thickness


conditions distribution (case 1)

., .. 1 . 10 '--~-C
-a~s-
e-'-_
---' 1.1 0 '--..--""-~-~~~~---;"----r-'
~ ~ 1.05 Case 2
.,
Qj ' .08
i!l2 Case3 -
~ ~ 1.00 'lii,'.06
1
EO.
(; E 0.95
z~
0 .9~, -0.5 0 0.5
\ ,
~ :; ,.04
~.,,,,,*, .02
~ ~ 1.00
Normalized distance
Irom sheet centerline [-I
§ - 0.98 -Case 1 -
o
Z 0 9. 6 Case 2
Case 3 -
0.94 '---'---'---'-"---'---'---'--'--'---'
-1 .0 -0.8-0.6-0.4 -0.2 0.0 0 .2 0.4 0.6 0 .8 1.0
Normalized distance from sheet centerline [-I

Fig. 4.45. Effect of temperature boundary conditions on sheet-thickness distribu-


tion in the slot downdraw process

4.5 Pressing of Drinking-Glass Stems

Gunter Weidmann, Kurt Holtberg, Henry Eisermann


Introduction

The pressing of stems is an important part of drinking-glass production.


Schott Zwiesel, a subsidiary of Schott Glas, produces high-quality wine
glasses. The design of such products strongly depends on the spirit of the
times, therefore the market permanently requires new models equivalent to
the pretensions of fashion. A few years ago, designs such as "Sevilla" were
en vogue, whereas now, glasses such as "Carisma" are more popular (see
Fig. 4.46).
There are two simultaneous production steps. On one line, cups are pro-
duced by a press-blow method. The stern and its foot are pressed syn-
chronously on a second line. At the end of this pressing phase, the stem
comes into contact with the cup, and both are merged. In this section, we
will focus on the stem production.
308 4. Shaping at Low Viscosities

a) b)

Fig. 4.46. Wine glasses: (a) Sevilla and (b) Carisma

For the production of attractive wine glasses, it is not sufficient to design


interesting contours. These contours must also be manufacturable. The range
of possible shapes of such drinking glasses is particularly limited by the stem
pressing, which is due to the method itself. A hot gob of glass of defined
temperature and volume falls on the entrance of a two-part steel mold (see
Fig. 4.47). The inner contour of the mold corresponds to the glass design, its
volume to the volume of the gob. A plunger begins to press the glass in the
interior of the mold.
The procedure is finished when the gob fills the mold completely. Dur-
ing the pressing, the glass is in contact with the steel of the mold, and is
therefore cooled down. The decreasing temperature causes an increasing vis-
cosity requiring a higher pressure on the plunger. The permitted tolerance of
pressure is ~ 100 bar and must not be overstepped. A lower diameter of the
mold induces an increasing pressing time and can cause an abort because of
freezing glass: the design cannot be produced, the stem is wasted glass. Up
to now, a test of a new design prior to manufacture has not been possible,
and costs of failed attempts were painful. Therefore, the development of a
tool that was capable of predicting the feasibility of a design before the first
"real" mold is machined appeared to be useful.

4.5.1 Model 1: Finite-Element Modeling

The first attempt for a mathematical model of the problem was made with the
commercial fluid dynamics program FIDAP. The calculation was performed
by means of the volumes-of-fluid method (VOF). This approach enables us
4.5 Pressing of Drinking-Glass Stems 309

1. Feeding 2. Pressing 3. Pressing 4. Deforming 5. Cheers II!


(begin) (end)

Fig. 4.47. Schematic of the most important process steps

to consider all major aspects of the process: the annealing of the glass with
a computation of the heat loss at the interface to the mold, and a precise
description of the temperature-dependent viscosity field. The idea of the VOF
method is a Eulerian arrangement, i.e., a fixed grid. The moving glass front
is realized by switching a layer of volume elements from air (or empty) to
glass, including all parameters of the material. This is an alternative method
to the Lagrangian arrangement, where the glass front is bound to the grid.
This
.- method requires a computation of the new grid at every time step .
Extreme deformations of the glass (which is the case here) result in extreme
deformations of the grid and will rapidly make the grid unfit for further
computations.
In the present model, the most important condition is the pressure at
the entrance. The real plunger does not appear in the model. The glass is
considered to be a Newtonian fluid, viscoelastic effects are neglected. The
viscosity is treated as a function of temperature (VFT), further material
parameters are density, heat conductivity, and heat capacity. Radiation effects
are not considered. Figure 4.48 shows the velocity field of the Sevilla glass at
an intermediate time step.
The computations deliver the filling-front position of the glass in the mold
with time. These functions, which can be plotted as simple curves, describe
the feasibility of the desired design. If the modeled mold is not filled in a
given time (the pressing time), the real process will not work. In many cases,
simple modifications of the model (i.e., the shape of the stem) are sufficient in
order to receive a pressable design. The problem with this approach was the
handling of the simulation, which could only be done by the department for
310 4. Shaping at Low Viscosities

2.70e+OO

2.43e+OO

2.16e+OO

1.8ge+OO

1.62e+OO

1.35e+OO

1.08e+OO

I 8.lle-Ol

5.41 e·Ol

2.70e·Ol
Fig. 4.48. Velocity field during the
O.OOe+OO pressing of the Sevilla glass

simulation. The reasons were the complexity of the preprocessing (generation


of grid and input file) and the requirements of the FEM calculations with
respect to the hardware. Therefore, the development of a simpler model that
was based on an analytical approach appeared to be useful.

4.5.2 Model 2: Analytical Modeling


Here, the basic equations for the analytical model are derived. The meaning
of the symbols is explained in Fig. 4.49 .
• Volume rate and averaged velocity:
Glass is considered as an incompressible Newtonian fluid. The volume rate
through a cross section with radius r(x) is

J
r(x)
V = 7rr 2 v m = 27r v Y dy = const. , (4.124)
y=o
and defines the average velocity of the fluid Vm (x) .
• Equation of momentum in x direction:
The inertia force and body force are neglected. Thus, the equation of mo-
mentum is reduced to the equilibrium of pressure and friction force. Es-
sentially, the pressure depends on x. We obtain

(4.125)

with viscosity 7].


4.5 Pressing of Drinking-Glass Stems 311

pp

HHtHHHH x
y
axial coordinate
radial coordinate
x' front of glass at time t
I length of drain
y r(x) radius of drain
pp pressure
pu surrounding pressure
Tr temperature at r

To

Pu mold

TrL---+-________--r

Boundary layer
of temperature

Fig. 4.49. Sketch of the model set-up with mathematical abbreviations appearing
in the equations

• Equation of heat transfer:


In this equation, axial heat conduction and dissipation are neglected, which
yields a simplified equation consisting of the thermal energy flux in the axial
direction and heat transfer in the radial direction.

vaT yay (yOT)


ax = a~~ oy (4.126)

with (constant) temperature conductivity a and temperature T.


• Viscosity:
The viscosity of glass melts is extremely dependent on temperature. It can
be written as

log (~) = T ~C (4.127)

with the specific constants A, B, and C. Instead of (4.127), we apply the


approximation

log (1,) = B' + C'T (4.128)

with the constants A', B', C ' .


312 4. Shaping at Low Viscosities

• Dimensionless basic equations:


If we apply the dimensionless numbers
pressure: P = P-Pu radial coordinate: Y = ~ ,
po-Pu '
velocity: V=...!'...
Vm '
radius of drain: R=L ro '
temperature: drag number: l/f = 7r(PO-pu )rt
11)0 V '
viscosity: C = .!L Peclet number:
1)0 '

viscosity at radius r: Cr = !k. ,


1)0
glass number:
axial coordinate: X = T'
(4.124)-(4.128) are transformed to

J J
1 1

1 = 2 VY dY = V dy2 , (4.129)
y=o y=o

l/fR4 dP
dX
=!~
Yay
(Y av) Cay , (4.130)

q5Va8 =!~ (ya8) (4.131)


ax yay ay ,

(1+r)8 ]
C = exp [ (8 + r) In Cr , (4.132)

c=exp(8Incr) . (4.133)
The integration of equation (4.130) yields

dV _ ~l/f 4 X dP Y
(4.134)
dY - 2 R ( ) dX C (X, Y)

(Consider: ~~ Iy=o = 0 (symmetry condition).) Repeating the integration


(consider: V IY=l = 0 (no-slip condition)), we obtain

V (X, Y) = _~l/fR4 (X)


4
dP
dX
J 1
2(d( .
c(X,()
(4.135)
(=Y

Now, we multiply (4.134) by y2. Regarding (4.129) and the no-slip condi-
tion V IY=l = 0, the integration over the radius of drain R leads to

(4.136)

The average viscosity is defined by


4.5 Pressing of Drinking-Glass Stems 313

_1_ _
Em (X) -
J
1
4(3d(
E(X,() (4.137)
(=0

Equation (4.136) transforms to

1 4 dP 1
-1 = -l}/R (X) - - - . (4.138)
8 dX Em (X)
Together with equation (4.135), this yields a new equation for the distri-
bution of velocity:

J
1
2(d(
v (X, Y) = 2Em (X) c(X,()
(4.139)
(=y

After the integration, (4.138) leads to

8
P (X) = 1 - (fr
Jx
Em (~)
R4 (0 d~ . (4.140)
1;=0

At the front of glass X=X', the condition P (X') = 0 is valid, which yields
the drag coefficient
x'

l}/ (X') =8
1;=0
J ~~ ~~; d~ . (4.141)

• Approximation of the boundary layer:


We define the normalized thickness b. (of the temperature boundary layer)
and the normal coordinate of the wall Y
. o(x)
thlckness: b. = r (x) ,
(4.142)
. - r(x)-y 1-Y
normal coordmate of wall: Y = 0 (x) = b. (X) .

We assume that 0 is much smaller than the radius r. Furthermore, let the
normalized temperature 8 be a function of Y. Then, (4.126) is transformed
to

_<pyd~ Vb. db. = d 2_8 . (4.143)


dY dX dy2
Equation (4.139) is applied in order to derive the gradient of velocity close
to the wall of the drain:
314 4. Shaping at Low Viscosities

dV 1
- - Iy-+o = 4Em (X) ~ (X) ( _) (4.144)
dY E X,Y

Integration yields the velocity field in the boundary layer:


y

V ( X, Y) = 4Em (X) ~ (X) J (i, () .


(=0
E (4.145)

We assume further a linear temperature profile in the boundary layer. It


is a function of Y only:

8 = { 1 - Y for 0::; ~ ::; 1 , (4.146)


o for 1::; Y .

We obtain the viscosity as a function of Y:

E={exP [(l-Y)ln(Er)] for O::;~::;l (4.147)


1 for 1::; Y .

Alternatively, we can substitute the temperature distribution in the bound-


ary layer by a polynomial of third degree:
3
e = Laiyti , (4.148)
i=O

where

thickness of boundary layer : ~' = ~g] , (4.149)


normal coordinate of wall .. Y' -- Ji(X)
r(x)~y -- l~Y
t,.'(X)

The polynomial may fulfil the following conditions:

This yields a more realistic temperature distribution in the boundary layer,

-
8= { 1 - ~2 Y' + 1.2 y,3 for 0 <
- Y'_ <- 1 , (4.151)
o for 1 ::; Y' ,

which leads to the "true" thickness of the layer,

Li -- ~~
2 . (4.152)

The insertion of (4.147) into (4.145) yields the velocity distribution in the
boundary layer:
4.5 Pressing of Drinking-Glass Stems 315

exp (Ylncr) -1
v (X,y) =4cm(X)~(X) l '
Cr nCr
(4.153)

We insert this into (4.143) (equation of energy) and obtain

- 4 -(jj- Y- [ exp (Y1ncr


- de Cm (X)
) -1 ] -_ ~2 (X) -d~ = -_-
d 2 e . (4.154)
Cr In Cr dY dX dy2

Regarding the linear temperature distribution (see (4.146)), we can inte-


grate over the whole boundary layer:

(4.155)

After further transformations, the averaged viscosity is derived as


In E.,.

exp ((1-~) IllEr) Jll.


-
Cm
1
= (1 - ~)4 + 4
(IllEr)
4 L':. e exp (-~)d~ . (4.156)
~ (l-L':.) In';r

With the abbreviations


a- L':.
- (l-L':.) III Er '
(4.157)
b=~
InE r '

we obtain

-
1
= (1 - ~)4 + 4 [(1_~)3~ (1 + 3a + 6a 2 + 6a 3 )
Cm lncr (4.158)
- !!..- (1 + 3b + 6b 2 + 6b 3 )] .
Cr
The thickness of the boundary layer is small, ~ -+ 0, the viscosity at the
wall of the drain is huge, Cr -+ 00. Therefore, a -+ 0 and b -+ 0 hold, and

~= [1 - ~ (X)]4 + -41 [1 - ~ (X)]3 ~ (X) (4.159)


Cm nCr
becomes a reasonable approximation of (4.158).
An equation for the thickness of the boundary layer ~ can be obtained
by the application of (4.159) and (4.155):
316 4. Shaping at Low Viscosities

With some transformations, we can derive an implicit equation for Do:

(4.161)

• Pressing:
The position of the glass front in the drain is determined by the volume
flow

. 2 I dx'
V = 1fT (x) cit . (4.162)

With the dimensionless time

(4.163)

the (normalized) volume flow can be obtained:

cp = R2 (X') d~' . (4.164)

The dimensionless pressure is defined by

(4.165)

The drag coefficient determines the (normalized) volume flow:

(4.166)

Now, we have all the relevant equations at hand in order to describe the
process. With these equations, an algorithm was programmed and tested.

4.5.3 Comparison of Model 1 and Model 2

Both models yield estimations about the design's practical feasibility, i.e.,
whether it is suitable for pressing or not. Now, the results of the FEM com-
putations are compared to those of the analytical model. The models gener-
ate very similar results (see Fig. 4.50). Obviously, the computations of the
analytical model are not as accurate as the Fidap computations. However,
the analytical model can be applied without much effort on slow pes, the
geometries can easily be programmed and - last but not least - it works
excellently.
4.6 The Use of Remeshing Methods in Pressing Simulations 317

0.10 a)
0.09
0.08
0.07
.€
c: 0.06
·'(ii2 0.05
1 !
0
~ 0.04 + • ~ 1: ~

«
<1l
'x 0.03
I .
.
!
i

0.02 , Analyllcal model


I
I
. Fipap ----
0.01 I

0.00
0 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40
Time Is

0.09
0.08
b)
0.07

0.06 .,,
.€ i
c:
0
;;::;
0.05 ...
'(ii
&. 0.04
"iii
'x 0.03
«
0.02
Analytical model
0.01 Fidap ----
0.00
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09
Time Is
Fig. 4.50. Computed curves of the filling-front position performed by (a) model
Carisma and (b) model Sevilla

4.6 The Use of Remeshing Methods III


Pressing Simulations

Ulrich Lange
Introduction

As was exemplified in the previous section, the so-called VOF method is


a possible simulation method for glass-pressing processes. The idea of this
method is to solve the N avier-Stokes equations on a fixed (Eulerian) mesh
with material parameters depending on a marker concentration. This marker
318 4. Shaping at Low Viscosities

concentration is determined by an additional transport equation. Its value for


a given cell of the grid indicates whether this cell is currently "filled" with
glass or not. In order to accurately resolve the evolution of the free surface
and possible steep temperature gradients, a fine numerical grid - and hence
a high computational effort - is required for the VOF method.
An alternative to the VOF method are (Lagrangian) remeshing methods.
In these methods, only the actual glass volume is meshed, and the grid is
modified in each time step according to the motion of the glass. Thus, one may
expect to maintain good accuracy with less computational effort. However,
one has to cope with the following crucial characteristics of glass-pressing
processes:
• During the calculations, one has to detect when and where the glass surface
contacts moving or non-moving tools and handle the fact that the boundary
conditions at the glass surface change at contact points.
• The shape of the computational domain often undergoes extreme changes
during the process (e.g., from a half-spherical gob to a thin disk).
• The contact between the hot glass and a cold mold or plunger results in
steep temperature gradients, which must be resolved.
Contact detection algorithms are well established in several commercial sim-
ulation packages: In each time step of the calculation, these methods test
geometrically if any node of the glass surface comes into contact with a tool.
Then, large forces are imposed at those nodes of the glass surface that are in
contact with a tool, thus preventing further relative movement. The second
challenge in the above list is more difficult: The built-in remeshing algorithms
of commercial packages (e.g., method of spines, elliptical or pseudo-elastic
mappings, Lagrangian remeshing) result in severely distorted meshes, if the
shape of the computational domain changes too much. The following proce-
dure, proposed by the package POLYFLOW [4.55] for example, is a solution
to this problem:
1. Calculate only a few time steps of the evolution of glass shape and field
variables using a built-in remeshing algorithm.
2. Import the current glass geometry into a mesh generator and generate a
completely new, undistorted mesh.
3. Interpolate the field variables from the old grid to the new grid and
restart the calculation.
POLYFLOW, together with its pre-processor GAMBIT [4.56]' provides all the
ingredients (i.e., automatic mesh generation, mesh-to-mesh interpolation,
etc.) to perform this procedure automatically. Moreover, GAMBIT'S "mesh
boundary layer" feature is a convenient tool to handle the third challenge
mentioned in the above list, i.e., the steep temperature gradients at the glass
surface.
In this paper, we report our experience with the application of the
above procedure using POLYFLOW /GAMBIT to glass-pressing processes. In
4.6 The Use of Remeshing Methods in Pressing Simulations 319

Sect. 4.6.1, we consider some technical aspects of the method - in particular


the choice of suitable values for the parameters controlling contact detection
and constraint forces. As application examples, we consider the pressing of
a tumbler (Sect. 4.6.2) and the pressing of a TV screen (Sect. 4.6.3). These
problems were formulated as benchmark problems for pressing by TNO and
adapted by Technical Committee 25 of the International Commission on Glass
in June 2001. However, we idealized the TV-screen problem to an axisym-
metric geometry, in order to decrease the computational effort. Rubin et
al. [4.57] treated the original 3D benchmark in the isothermal case using the
same POLYFLOW /GAMBIT method.

4.6.1 Some Technical Aspects of the Method

The iterative algorithm sketched above is readily implemented as a UNIX


shell script, for example. A sensible enhancement of the basic algorithm is
an automatic transition from relatively coarse meshes, which are suitable for
the initial - usually bulky - glass geometry, to finer meshes at the end of the
process, when the glass geometry is thin. A suitable element size at a given
stage of the process can be calculated from the distance between plunger and
mold in that stage, for example. In many cases, such an automatic mesh re-
finement allows for considerable reduction of the computation times without
any manual interventions into the algorithm.
Another critical factor concerning computational effort is the contact de-
tection. It is controlled by two parameters lOp ("penetration accuracy") and
Ed ("element dilatation"): During the time step tlt, in which the contact with
a tool occurs for some node of the glass surface, this node can still move a
distance of l = Ilvlltlt. If this distance is larger than lOp the time step is re-
peated with a decreased step size 6.t. In addition, POLYFLOW uses a heuristic
trick to anticipate the occurrence of contact: The contact elements are simply
enlarged by the length Ed.
Obviously the contact surface is found more accurately, if smaller values
for lOp and Ed are chosen. On the other hand, smaller values of lOp re~mlt
in smaller time steps, and hence more computational effort. A reasonable
compromise is to set lOp and Ed to about 1% of the final glass thickness.
Once contact with a tool has been detected for a node, normal and tan-
gential constraint forces are imposed. The normal force in is given by:

in = -k( Vglass - Vtool)n . (4.167)

Here, Vglass denotes the glass velocity in the node, Vtool is the velocity of the
tool, and n is the unit normal of the tool surface. The penalty coefficient k has
to be chosen sufficiently large in order to prevent further relative movement
between glass and tool. A suitable value for k, which should yield a relative
displacement of the order of magnitude of the penetration accuracy lOp, can
be estimated by
320 4. Shaping at Low Viscosities

k ~ P char tchar .
(4.168)
cp

Here, P ch a r is a typical pressure occurring during pressing and tchar is the


process time. The tangential force Is is given by:

(4.169)

where s is a unit tangent to the glass surface. In order to impose a no-slip


condition, one should choose a slipping coefficient Fslip of the same order of
magnitude as the penalty coefficient k. It is also possible to impose slippage
laws, like Navier's law, via the tangential force Is.

4.6.2 Example: Pressing of a Tumbler

A sketch of the initial and the final stage of the pressing process for a tumbler
is shown in Fig. 4.51. The plunger is assumed to move downwards with a

32.5 mm
31 mm

u-
!!l
ti
<II
c:
0
C
0
~ ~
v
t() '<t E
en en E
co co co
C1>

BS7 (No functi


- - l
~
iii BS3 (Free surtace)
E
E I
>- I 23mm
!:e.
~jlr-
en
In BS2 (No-slip) (')
---l

BS6 (No function)


--t~
. ~~
27mm

Fig. 4.51. Benchmark problem: an axisymmetric tumbler


4.6 The Use of Remeshing Methods in Pressing Simulations 321

velocity v(t) decreasing in time according to

v(t) = _(0.480e-18.96t - 0.01906) m/s . (4.170)

Thus the whole process lasts only 0.0778 s. The initial glass temperature is
1000°C. The mold and the plunger are assumed to remain at a constant
temperature of 500°C throughout the process. The heat transfer coefficient
between mold/plunger and glass is 5000 W /m K. Apart from this heat trans-
fer, plunger and mold appear only as contact surfaces in the model. The
material data of the glass are:

viscosity: 7](T) = 1O- 2 .8+47oo/(T-220) Pas (T in °C),


conductivity: )"'(T) = 2 W /m K,
specific heat: cp = 1400 J /kg K,
density: p = 2500kg/m3 .
The penalty and the slip coefficient were set to 10 13 Pas/m. The penetra-
tion accuracy and the element dilatation were set to 10- 5 m. The initial pave
size of the grid was 0.9 mm. The pave size was refined down to 0.3 mm during
the run. Thus the initial grid had about 1000 elements and the final grid
had about 5000 elements. A mesh boundary layer consisting of two rows of
thinner elements was used at the free surface of the glass, in order to account
for the steep temperature gradient. It had to be switched off, however, at
t = 0.04, when the automatic remeshing with boundary layers failed because
of the large curvature of the glass front. The computation needed 49 hours
on an HP J5000 (however, since there were several tasks running simultane-
ously on that machine, only about 50% of performance of one processor was
available) .
Figure 4.52 shows that the numerical meshes do indeed look nice and
undistorted throughout the computation. Figure 4.53 illustrates the steep
gradients in glass speed and temperature that develop in the narrow gap
between plunger and mold during the short process time.

4.6.3 Example: Pressing of an "Axisymmetric TV Screen"

As a second example, we consider an idealized "axisymmetric TV screen"


pressing process from a spherical gob, as sketched in Fig. 4.54. The plunger
is assumed to move downwards with a velocity v(t) decreasing with time
according to

v(t) = -(0.3675 e-1. 535t - 0.03675) m/s . (4.171)

The initial glass temperature is 1000°C. Mold and plunger remain at a


constant temperature of 500°C and the heat transfer coefficient between
mold/plunger and glass is 5000 W /m K. The material data of the glass are
I
322 4. Shaping at Low Viscosities

- I
I

I
t " 0.20 s
1- 0.00,
I

II

I
~]
I1 t '" 0.50 s I

I
t" 0.30 s I

I
I
I

Fig. 4.52. Axisymmetric tumbler: meshes at different times


4.6 The Use of Remeshing Methods in Pressing Simulations 323

t=06
.0 s 1= 0.70, I~ \=0.778s I

I
I

Fig. 4.52. continued

Contour levels t = 0.02 s t = 0.04 s Contour levels


speed 1m s- l temperature 1°C
1.60 1000

1.44 975

1.28 950

1.12 925

0.96 900

0.80 t = 0.06 s t = 0.0778 s 875

0.64 850

0.48 825

0.32 800

0.16 775

0.00 750

Fig. 4.53. Axisymmetric tumbler: glass speed and temperature at different times
(see also movie on CD)
324 4. Shaping at Low Viscosities

lv
SD3 BS6 (No function)
BS5 (Contact)

BS2 (Free sur1ace)

E
rJ) E
co E 0
r--
BS4 (Contact) E
o ·
SD2 BS3 (No function) 235mm

Fig. 4.54. Benchmark problem: an "axisymmetric TV screen"

the same as in the "tumbler" example except for the thermal conductivity .\,
which now depends on temperature:
.\(T) = 0.7725+(S.3xlO- 4 T)+(2xlO- 8 T 3 )W/mK(Tin Kelvin). (4.172)
Moreover, we impose a radiation heat transfer condition at that part of the
free glass surface, which is not in contact with a tool: The emissivity of the
glass is O.S and the ambient temperature is 500 DC.
The penalty and the slip coefficient were set to 10 15 Pa s/m. The penetra-
tion accuracy and the element dilatation were set to 10- 4 m. The calculation
was run twice on different grids in order to check grid dependence. A rule
for the automatic grid refinement was that the element size at a given time
should be at most 5 times (first run) or 10 times (second run) smaller than
the current thickness of the glass. Moreover, a mesh boundary layer was used
at the free surface. Thus, the first run started with a grid of 300 elements
and ended with a grid of 1550 elements. The computation time was about
2.5h on a HP J5000 (at about 50% of the performance of one processor).
The second run started with a grid of 550 elements and ended with a grid of
3350 elements. The computation time was about 11 h.
The meshes at different time steps ofthe second run are shown in Fig. 4.55.
The evolution of glass speed and temperature is shown in Fig. 4.56. The flow
and temperature fields predicted on the different grids are in good agree-
ment. However, the glass front moves slightly slower in the run with coarser
grids. This variation may be due to some volume loss caused by insufficient
resolution of the highly curved glass during the remeshing. Another possible
reason is a slightly higher penetration of the glass into the mold in the case
of coarser meshes.

Conclusion
The adaptive remeshing technique based on POLYFLOW /GAMBIT is a tool
for the simulation of glass-pressing processes, which can provide useful in-
4.6 The Use of Remeshing Methods in Pressing Simulations 325

t " 0.00 s
J t = 0.20 s

t = 0.40 s
t " 0.60 s

t =0.80 s t " 1.00 s

t" 1.20 s t" 1.25 s

Fig. 4.55. "Axisymmetric TV screen": meshes at different times

sights connected to many practical questions, such as necessary pressures


and thermal interaction between glass and tools.
Certainly the most unrealistic simplification in the benchmark examples
discussed here is the neglect of the warming of the tools. This idealization is
not necessary, though. In fact, the prediction of heat conduction in the tools
can be easily included in the modeL
From the numerical point of view, the method is very robust and easy
to handle, but the computations can be quite time-consuming. This compu-
326 4. Shaping at Low Viscosities

Temperatu J

J
Contour levels Contour levels
speed Ims- 1 temperature loe

0.350 t =0.00 s 1000

0.315 980

0.280 960

0.245 940
t = 0.40 s 920
0.210

0.175 900

0.140 880
t =0.80 S
0.105 860

0.070 840

0.035 820
t = 1.20 s
0.000 800

t = 1.25 s
Fig. 4.56. "Axisymmetric TV screen": evolution of glass speed and temperature
(see also movie on CD)

tational effort still prevents practically relevant 3D applications, at least on


smaller workstations.

4.7 Chill Ripples in Pressing and Casting Processes

Ulrich Lange
Introduction

"Chill ripples" , which are also known as "press ripples" or "flow ripples" , may
be produced on the glass surface under certain process conditions in many
hot-forming - in particular pressing and casting - processes. Such ripples are
frequently perceived in low-quality consumer glassware. The feet of cheap
wine glasses, for example, often show concentric waves on their surfaces.
4.7 Chill Ripples in Pressing and Casting Processes 327

The name "chill ripples" reflects the fact that the waves usually appear
if the tool temperature is too low. The obvious counter-measure of allowing
higher tool temperatures is viable only to a limited degree, because glass
tends to stick to forming tools having temperatures that are too high. The
window between the critical temp~ature for the onset of ripples and the
sticking temperature can be quite small. Hence, for high-quality products
(e.g., the optical fiber preforms shown in Fig. 4.57) either the process condi-
tions must be controlled very carefully, or an expensive post-processing must
be accepted.
Different opinions are given in the glass industry regarding the cause of the
ripples. In this section, we will investigate the mechanism of the evolution
of chill ripples using a model based on the adaptive POLYFLOW /GAMBIT
remeshing technique [4.55], which was described in detail in the previous
section.
In a first case study, we consider a simple casting process. In fact, the
model predicts the development of ripples on the glass surface, if the mold
temperature is too low. Close inspection of the temperature and flow field
corroborates the conjecture that chill ripples are a consequence of the strong
temperature dependence of glass viscosity. In particular, it is obvious from
the results that thermal expansion effects are not necessary for the emergence
of chill ripples.
In a second case study, we compare our model predictions for a press-
ing process with experimental results published by Kluge (see [4.58,59]).
Although the model fails to reproduce Kluge's results for low initial glass

Fig. 4.57. Optical fiber preforms with


(right) and without (left) chill ripples
328 4. Shaping at Low Viscosities

temperatures, the agreement for higher initial glass temperatures is good.


Moreover, the model reveals interesting details of the evolution of chill ripples
in pressing processes, which would be difficult to observe in an experiment.

4.7.1 A Simple Casting Process

As a first example for the development of chill ripples, we consider the casting
process sketched in Fig. 4.58: An optical glass (SF 10) exits a feeder with a
prescribed flow rate of 170 glmin and flows into a cylindrical mold. The glass
temperature in the feeder is 1200 DC. The mold is initially at 350 DC. Contrary
to the benchmark examples in the last section, heat conduction in the mold
is fully modeled. At the exterior surfaces of the mold, we assume a heat
transfer boundary condition with coefficient am = 100 W 1m2 K and ambient
temperature T = = 350 DC. The interior mold surfaces are insulated except
when in contact with the glass. At the free surface of the glass, we assume
a radiation boundary condition (T= = 350 DC). The feeder wall is insulated.
In the case of contact, ideal contact heat transfer and no-slip conditions are
assumed.
The predicted evolution of glass surface and temperature field during the
casting is shown in Fig. 4.59. A closer look shows that small waves have devel-
oped on the surface where the glass touches the side wall and the outer part
of the bottom of the mold. Thus, the model is indeed capable of reproduc-
ing the infamous chill ripples. In agreement with experience, the predicted
glass surface is smooth everywhere, if the initial mold temperatures is chosen
higher than in the case shown in Fig. 4.59.
In Fig. 4.60, the evolution of a ripple is shown in detail, thus illustrating
the underlying mechanism: When a point on the glass surface comes into
contact with the mold, its environment is cooled rapidly. Due to the resulting

Inflow (mass flow prescribed)

E
E
<0 E
E

Mold E
E
II)

Fig. 4.58. Model of a simple axi-


30mm symmetric casting process
r
4.7 Chill Ripples in Pressing and Casting Processes 329

II t = 0.00 s t = 0.12s t = O.24,

;....J

Contour levels
temperature 1°C

1200
1115
1030
945
860
775
690
605
520
435
350

Fig. 4.59. Evolution of glass shape and temperature field in the casting process

strong increase in viscosity, the vicinity of the contact point behaves almost
like a solid body (see the solid red zones on the right-hand side in Fig. 4.60,
which correspond to glass velocities smaller than 10 Il-m/s).
In particular, a part of the glass surface close to the contact point virtually
stops moving towards the mold. Thus, a wave crest is formed. The less-viscous
glass from above flows around the "frozen" contact region and eventually
touches the mold at a new contact point, some distance away from the first
one. The "frozen" region is larger if the glass temperature level is already
low. This explains why the ripple amplitudes increase in the radial direction
and are usually not visible in the center.
330 4. Shaping at Low Viscosities

Velocity vectors
Isotherms (solid red area: IvJ < 1 e - 5 ms- ')

t = 16 s

Contour levels
temperature I'C

900 t = 17 s
845
790
735
680
625
570
515
t = 18 s
460
405
350

t = 19 s

Fig. 4.60. Evolution of a ripple during casting

4.7.2 A Model for Kluge's Experimental Set-Up

As another example for chill ripples, we consider a model for Kluge's experi-
mental set-up for their development during pressing (see [4.58,59]): A sketch
of the initial and final stage of the pressing process indicating the dimensions
is shown in Fig. 4.61. Note that the model is axisymmetric, and that pressing
tools are idealized in the model by two cylinders. The plunger moves down-
wards with a constant velocity Wp. The calculations were stopped when the
distance between plunger and mold reached 2.5 mm.
It is assumed that glass and mold are initially at a common constant
temperature TGA and the plunger is initially at the constant temperature TpA.
4.7 Chill Ripples in Pressing and Casting Processes 331

30mm

E
E
~ C\I
Q; , Plunger ~
E ~~------~~ Q;
E ~ 2.5 mm Plunger
>-
UJ
'0 :::'>-1, -= > __. . .
G:.o.;la""s;.::.
~.
s.:...-
UJ
E Mold o Mold
·x UJ
<: E ·x
<:

11 .5 mm - 18 mm

Fig. 4.61. Sketch of the model used to simulate Kluge's experiments

All surfaces of the tools are assumed to be insulated, except when in contact
with the glass. At the free surface of the glass a radiation boundary condition
with TpA as ambient temperature is assumed. In the case of contact, we
assume an ideal contact heat transfer and no-slip conditions.
The material data for the tool material (X22CrNi17) was adapted from
Kluge's thesis. We considered either BK 7 or SF 10 glass. The glass mate-
rial data was also adapted from Kluge, except for conductivity, which was
increased to .A = 4 W 1m K (for BK 7) or .A = 3 W 1m K (for SF 10) in order
to consider (very roughly, of course) the increased energy transport due to
radiation.
In Fig. 4.62, the predicted evolution of the glass surface during the press-
ing process is shown for the case of glass SF 10, which is initially at a temper-
ature of TCA = 750 ac.
The plunger is at TpA = 400 ac in the beginning and
moves with Wp = 0.3 mm/s. In agreement with Kluge's observations, ripples
appear at the glass surface facing the plunger. According to Kluge, the rip-
ples should vanish if the plunger temperature is increased to T pA = 475 ac.
The model predictions for this case are shown in Fig. 4.63. Although ripples
develop initially, they indeed vanish again towards the end of the process.
The curious phenomenon of vanishing ripples, which would be hard to ob-
serve in experiments, can be understood by looking at the total force needed
to squeeze the bulk of the pressing further apart. As the plunger moves down-
ward, it conveys this force to the ripple crests. Because the pressing force
increases very strongly as the gap between plunger and mold narrows, it may
become sufficiently large to level ripples that have not cooled too much.
As a measure for the ripple amplitude, we define the deviation:

()" = [~21
R
Is(r) - 81 2
r dr
]1/2
(4.173)

Here, s(r) specifies the upper part of the final glass surface contour up to a
radius of R = 0.014m, and 8 is the mean of s(r) on the disk with radius R.
332 4. Shaping at Low Viscosities

Contour levels
temperature rc
750

715

680

645

610

575

540

505
t =15.00 s .'

------- ....
ft. 470

435

400

Fig, 4.62. Evolution of glass surface for SF 10 (TGA = 750 o e, T pA = 400 o e,


Wp = 0.3mm/s). Vertical direction stretched by a factor of 20 in the zoomed
pictures on the right-hand side

The curved front of the pressing is well outside of this disk. For both glasses
and all considered values of TCA, the deviation (J is found to be monotoni-
cally decreasing with the plunger temperature TpA. A comparison between
Kluge's experimental data and our model predictions is shown in Fig. 4.64.
The boundary curve for the onset of ripples according to Kluge's experi-
4.7 Chill Ripples in Pressing and Casting Processes 333

Contour levels
temperature 1°C
750

715

680

645

610

575

540

505

470

435

400

Fig. 4.63. Evolution of glass surface for SF 10 (TGA = 750°C, T pA = 475°C,


Wp = 0.3 mm/s). Vertical direction stretched by a factor of 20 in the zoomed
pictures on the right-hand side

ments corresponds to measured amplitudes of less than O.5!-lm. Because such


small amplitudes cannot be predicted reliably due to the limited accuracy
of the contact detection algorithm, we compare the experimental curve with
predicted a isolines for 5 !-lm and 10 !-lm.
334 4. Shaping at Low Viscosities

IJl 6.0 , -- - - - - - - - - - - - - - - - - - - - - ,
<lI BK7 Exp. (Kluge). 0 .5 Ilm SF10: Exp , (Kluge). 0,5 Ilm
0.. Model -a-
5 Ilm Model. 5 Ilm -e--
- 5.5 Model, 10 Ilm _ Model. 10 Ilm---+-
1
I- 5.0
-;:;
o
Ol 4 ,5
.2
~ 4 .0
'u;
o
.s:~ 3.5
IJl
~ 3.0
Ol
~ 2,5
'co
400 450 500 550
Initial plunger temperature T PA rc

Fig. 4.64. Comparison of predicted ripple amplitudes and experimental data

For low values of the initial glass viscosity TJ(TCA ), the predicted IJ isolines
are in good agreement with the experimental boundary. However, the steep
decrease of the critical plunger temperature for higher initial glass viscosities
could not be reproduced by the model. Unfortunately, we could not inves-
tigate whether the model would predict this steep decrease at still higher
initial viscosities, because the time steps of the numerical algorithm became
too small in these cases.

Conclusion

The model calculations presented here corroborate the conjecture that chill
ripples are caused by an "arching" flow pattern in regions of contact between
glass and tools: The "arching flow" evolves due to immobilized regions around
cold contact points which are overflown by the following, less-viscous, glass.
The results show that - contrary to, for example, the conjecture of Zschim-
mer et al. [4.60]- heat expansion effects are not necessary for the development
of chill ripples, as they were not even included in the model.
Many aspects of Kluge's experiments are not reproduced by the calcu-
lations. Still, the extent of agreement is surprisingly large, considering the
simplifications of the model (ideal contact heat transfer, approximation of
radiative energy transport by increased thermal conductivity). In order to
get a more conclusive assessment of the model predictions, one should prob-
ably try to reproduce experimental data of a casting process yielding larger
ripple amplitudes.
References 335

References
4.1 M. Coenen, H. Sendlbeck: "EinfluB der Formenschmiermittel auf den War-
meubergang Glasschmelze/Form", Keram. Z. 35(12), 640-644 (1983)
4.2 W. Trier: "Temperaturverteilung und WarmefluB in Glas in Vorformen von
Hohlglasmaschinen", Glastechn. Ber. 33, 449-456 (1960)
4.3 W. Trier: "Temperature distribution and heat flow in glass in blank molds
of container machines", J. Am. Ceram. Soc. 44, 339-345 (1961)
4.4 D. Steingrube, W. Trier: "Warmekontakt zwischen Glasschmelze und offen-
porigem Sintermetall", Glastechn. Ber. 51, 104-110 (1978)
4.5 P. Manns, W. Doll, G. Kleer: "Glass in contact with mould material for
container production", Glastechn. Ber. 68, 389-399 (1995)
4.6 M. Falipou, C. Donnet: "Sticking temperature investigations of glass/metal
contacts - determination of influencing parameters", Glastechn. Ber. Glass
Sci. Techno!. 70, 137-145 (1997)
4.7 M.F. Modest: Radiative Heat Transfer (McGraw-Hill, New York 1993)
4.8 R. Viskanta, E.E. Anderson: "Heat transfer in semi-transparent solids", Adv.
Heat Transfer 11, 317-441 (1975)
4.9 U. Fotheringham, F.-T. Lentes: "Active thermal conductivity of hot glass",
Glastechn. Ber. Glass Sci. Techno!. 67, 335-342 (1994)
4.10 F.T. Lentes, N. Siedow: "Three-dimensional radiative heat transfer in glass
cooling processes", Glastechn. Ber. Glass Sci. Techno!. 72, 188-196 (1999)
4.11 I.H. Farag, M.J. Beliveau: "Heat transfer during glass forming", Chern. Eng.
Comm. 52,21-32 (1987)
4.12 D. R. Pitts, L. E. Sissom: Theory and Problems of Heat Transfer, Schaum's
Outline Series (McGraw-Hill, New York 1998)
4.13 W. Vauck, H. Muller: Grundoperationen chemischer Verfahrenstechnik,
(Deutscher Verlag fUr Grundstoffindustrie, Leipzig 1994)
4.14 VDI-Wiirmeatlas: (Springer, Berlin, Heidelberg 1997)
4.15 D.A. McGraw: "Transfer of heat in glass during forming", J. Am. Ceram.
Soc. 44, 353-363 (1961)
4.16 R. Schuhmacher: "Thermische und mechanische Vorgange beim Verpressen
einer ziihflussigen Glasschmelze", Glastechn. Ber. 47, 239-250 (1974)
4.17 C.J. Fellows, F. Shaw: "A laboratory investigation of glass to mould heat
transfer during pressing", Glass Techno!. 19, 4-9 (1978)
4.18 H.D. Baehr, K. Stephan: Wiirme- und Stoffiibertragung (Springer, Berlin,
Heidelberg 1996)
4.19 J. Manthuruthil, T.R. Sikri, G.A. Simmons: "Simplified mathematical model
simulating heat transfer in glass-forming molds", J. Am. Ceram. Soc. 57,
345-350 (1974)
4.20 N.D. Eryou, L.R. Glicksman: "An experimental and analytical study of re-
diative and conductive heat transfer in molten glass", J. Heat Transfer 94,
224-230 (1972)
4.21 R. Gardon: "The emissivity of transparent materials", J. Am. Ceram.
Soc. 39, 278-287 (1956)
4.22 D.A. McGraw: "Heat transfer in the glass-mold interface", Glastechn.
Ber. 46, 89-91 (1973)
4.23 T.K.G. Howse, R. Kent, H. Rawson: "The determination of glass-mould heat
fluxes from mould temperature measurements", Glass Techno!. 12, 84-93
(1971 )
4.24 M. Patka, M. Klauz, J. Nisou: "Warmeubergang zwischen Glas und Form
beim Pressen kleiner Glasgegenstande", Glastechn. Ber. 44, 241-244 (1971)
336 4. Shaping at Low Viscosities

4.25 H. Rawson: "Radiative heat transfer in glass manufacture - one- and two-
dimensional problems", Glastechn. Ber. 66, 77-84 (1993)
4.26 Hiittentechnische Vereinigung der Glasindustrie: Mefltechnik in Glashiitten,
HVG vocational training course (Frankfurt/Main 1971)
4.27 O. Sohr: Ramanspektmskopie an technischen Gliisern und Glasschmelzen bei
hohen Temperaturen, PhD Thesis (Free University of Berlin, Berlin 1999)
4.28 T. Kunc, M. Lallemand, J.B. Saulnier: "Some new developments on coupled
radiative-conductive heat transfer in glasses - experiments and modelling" ,
Int. J. Heat Mass Transfer 27, 2307-2319 (1984)
4.29 D.P. DeWitt, G.D. Nutter: Theory and Practice of Radiation Thermometry
(Wiley, Chichester 1988)
4.30 W. Pepperhoff: Temperaturstrahlung (Steinkopff, Darmstadt 1956)
4.31 U. Gliickert: Erfassung und Messung von Wiirmestrahlung (Franzis, Miinchen
1972)
4.32 R.E. Chupp: Development and Experimental Evaluation of a Technique to
Determine the Temperature Distribution in Semitransparent Solids fmm Re-
motely Sensed Spectral Emission Data, PhD Thesis (Purdue University, West
Lafayette, IN 1973)
4.33 P.J. Hommert: Application of a Thermal Remote Sensing Method to Mea-
surement of Temperature Distribution in Glass, Master Thesis (Purdue Uni-
versity, West Lafayette, IN 1973)
4.34 R.E. Chupp, R. Viskanta: "Development and evaluation of a remote sensing
technique for determining the temperature distribution in semitransparent
solids", J. Heat Transfer, 391-397 (August 1974)
4.35 R. Viskanta, J.P. Hommert, G.L. Groninger: "Spectral remote sensing of
temperature distribution in semitransparent solids heated by an external
radiation source", Appl. Opt. 14, 428-437 (1975)
4.36 R. Viskanta, R.E. Chupp, J.S. Toor, P. Hommert: "Spectral remote sensing
of temperature distribution in glass", IEEE Trans. lA-II, 438-446 (1975)
4.37 R. Viskanta: "Infrared radiation techniques for glass surface and temperature
distribution measurements", IEEE Trans. lA-II, 494-505 (1975)
4.38 LH. Farag, R.L. Curran: "Pyrometric measurement of glass plates with non-
uniform temperature distribution", Chern. Eng. Comm. 52, 291-300 (1987)
4.39 C. Ades, R. Tsiava, J.-P. Traverse: "Spectral remote sensing of temperature
distribution in molten glass for optimizing the glass manufacturing process" ,
Glastechn. Ber. 62, 261-265 (1989)
4.40 R. Tsiava, C. Ades, J.-P. Traverse: "Measurement of temperature in molten
glass by multi-wavelength pyrometry", High Temp.-High Press. 24, 285-294
(1992)
4.41 R.E. Fields, R. Viskanta: "Measurement and prediction of dynamic temper-
atures in unsymmetrically cooled glass windows ", J. Thermophysics Heat
Transfer 7, 616-623 (1993)
4.42 T.M. Nagtegaal, C.C.M. Rindt, A.A. VanSteenhoven: "Temperature mea-
surements in hot glass melts", in Pmc. ICG Annual Meeting 2000: Chal-
lenges and Break-thmugh Technologies, ed. by H. de Waal, R.G.C. Beerkens,
J. Peelen, J.H.A. van der Wouden (Amsterdam 2000) paper T3-7
4.43 H. Scheffler, H. Elsasser: Physik der Sterne und der Sonne, 2nd edn. (Spek-
trum, Heidelberg 1990)
4.44 A.K. Louis: Inverse und schlecht gestellte Pmbleme (Teubner, Stuttgart
1989)
4.45 J.O. Israd: "The emission of thermal radiation from hot glass. Part 1. Emis-
sivity of sheets, spheres, cylinders, and tubes", Glass Technol. 27, 24-31
(1986)
References 337

4.46 V. Le Dez, J.C. Hudin, M. Lallemand: "Directional emissivities of semi-


transparent media - cases of dihedron, infinite cylinder and cone" , Glastechn.
Ber. 66, 127-137 (1993)
4.47 Special thanks to Thermosensorik GmbH, Erlangen, Germany, for their
friendly support.
4.48 D. Braess: Finite Elemente: Theor·ie, schnelle Loser und Anwendungen in
der Elastizitiitstheorie, 2nd edn. (Springer, Berlin, Heidelberg 1997)
4.49 J.D. Buckmaster, A. Nachman, L. Ting: "The buckling and stretching of a
viscid a" , J. Fluid Mech. 69, part 1, 1-20 (1975)
4.50 P.D. Howell: Extensional Thin Layer Flows, PhD Thesis (Oxford 1995)
4.51 W.C. Hynd: "Flat glass manufacturing processes", in Glass: Science and
Technology, Vol. 2, ed. by D.R. Ullmann, N.J. Kreidl (Academic Press, Or-
lando 1984)
4.52 Polyflow S.A.: POLYFLOW User Manual, Version 3.6.0 (May 1998)
4.53 A. Milutinovic-Nikolic, R. Jancic, R. Aleksic: "Mathematical modelling and
simulation of drawing thin glass sheet from a rectangular preform", Glass
Technol. 39(5), 166-172 (1998)
4.54 M. van Dyke: Perturbation Methods in Fluid Mechanics (Parabolic, Stanford
1975)
4.55 Fluent Inc.: POLYFLOW User Manual, Version 3.8.0 (August 2000) Chapter 5
4.56 Fluent Inc.: GAMBIT User Manual, Version 1.3 (April 2000)
4.57 Y. Rubin, L. Fondin, T. Avalosse, E. Pichelin: "Non-isothermal simulation of
3D glass pressing involving an adaptive remeshing technique", in Pmc. Int.
Congr. Glass, Vol. 2, Edinburgh, Scotland, 1-6 July 2001, pp. 962-963
4.58 W.D. Kluge: Ein Beitrag zum Kontaktverhalten von Glas mit Formgebungs-
werkzeugen, PhD Thesis (Freiberg 1988)
4.59 W.D. Kluge: "Kleben und Wellenbildung beim Pressen von Glas", Silikat-
technik 38(11), 367-368 (1987)
4.60 E. Zschimmer, K. Hesse, K. Meures: "Uber die Ursachen der Gefrierrisse und
Runzeln (Wellen) des PreBglases", Glastechn.Ber. 3, 342-343 (1924/25)
4.61 W. Grundmann: "Einfliisse auf die Bildung von 'FlieBwellen' bei PreBglii-
sem", Glastechn. Ber. 39(8), 370-376 (1966)
5. Reshaping at High Viscosities

5.1 Temperature-Dependent Elasticity III


Reshaping Simulations

Thomlf Johansson, Ulrich Fotheringham


Introduction

In the standard model of Tool~Narayanaswamy for the viscoelastic material


behavior of glass, the shear modulus of elasticity is assumed to be temperature
independent. The motivation is the exponential temperature dependence of
the viscosity, which implies an exponential temperature dependence of the
characteristic shear stress relaxation times. The usual assumption is that the
decrease of the modulus of elasticity slows down significantly with increasing
temperature.
In contrast to this assumption, measurements of the modulus of elasticity
at Schott revealed a drastic drop at temperatures ~ 100°C above the trans-
formation temperature. We will show that this has decisive consequences for
reshaping simulations.

5.1.1 Model

Viscoelastic Material Model

The basic equation for the description of viscoelastic behavior in hot glass in
the transformation region reads:

J J
t t

O"ij (t) = 2G (t - t') eijdt' + 6ij K (t - t') .pdt' , (5.1)


o 0

where O"ij(t) is the stress tensor, G is the time-dependent shear modulus of


elasticity, K is the compression modulus of elasticity, eij is the deviatoric
strain tensor, and P is the volumetric strain.
The shear modulus of elasticity G, the compressive modulus of elastic-
ity K, the modulus of elasticity E, and the Poisson ratio v are connected
according to

H. Loch et al. (eds.), Mathematical Simulation in Glass Technology


© Springer-Verlag Berlin Heidelberg 2002
340 5. Reshaping at High Viscosities

E= 9KG and G= E (5.2)


3K+G 2(1 + v)
In the following, we will assume K to be constant, so that the second part
of the viscoelastic equation's right-hand side is constant. We concentrate on
the shear stress relaxation:

J
t

Tij (t) = 2G (t - t') eijdt' . (5.3)


o
We define

() _G(t) (5.4)
9 t - Go '

where Go is the instantaneous shear modulus of elasticity, which implies

J
t

Tij (t) = 2G o 29 (t - t') eijdt' . (5.5)


o
For a temperature-dependent shear modulus of elasticity Go(T), we get:

J
t

Tij (t) = 2Go (T) 29 (t - t') eijdt' . (5.6)


o
A degradation of the instantaneous shear modulus of elasticity implies an
analogous variation of the stiffness matrix: the material is more easy to de-
form.

Model of an Example Reshaping Process

We will demonstrate the effect of a temperature-dependent shear modulus


of elasticity by a representative reshaping process. An initially flat sheet is
positioned on a mold. By heating in the edge region and subsequent vacuum
suction, the sheet is made to sag and its geometry eventually adapts to the
mold geometry. In Fig. 5.1, the sheet was heated to a well-defined shaping
temperature where the (initially purely gravity driven) sagging just started.
Here, all the geometries are rotationally symmetric. The following simulations
were made with the commercial finite-element program ABAQUS.

Fig. 5.1. Glass sheet on mold, heated to shaping temperature (red: high temper-
ature, blue: low temperature); see also movies on CD
5.1 Temperature-Dependent Elasticity in Reshaping Simulations 341

Material Parameters. In this example, we use the following material param-


eters:

• heat conductivity: 1.2W/mK


• density p: 2230kg/m 3
• heat capacity cp : lkJ/kgK
• modulus of elasticity
(room temperature): 63GPa
• Poisson ratio: 0.2
• coefficient of thermal
expansion (10- 6 /K): 3.3

In the first simulation, we used the standard viscoelastic model with the
temperature-independent shear modulus of elasticity. Measurements at the
Schott labs revealed that this modulus drops very strongly at temperatures
only 100 DC above the glass transition temperature Tg • The temperature
where the shaping takes place in our example is significantly higher. In a
second simulation, we therefore used the temperature-dependent modulus
from Fig. 5.2.

5.1.2 Simulation Results

Temperature-Independent Shear Modulus of Elasticity

In a first step, the transient temperature distribution in the sheet was com-
puted. In this simulation, the sheet was heated until a typical temperature
was reached at a specified location known from the real process. This tem-
perature distribution was the basis for the following mechanical simulation,
i.e., we did a sequentially coupled thermomechanical computation.
If a temperature-independent shear modulus of elasticity is used, the
gravity-induced sagging of the sheet is rather small (see Fig. 5.3a). Even
with additionally applied vacuum suction, the final geometry is not reached

70.-----------------------------,
60
50
~ 40
~
W 30

20
10 Fig. 5.2. Temperature-de-
O+-,_~_.~_.~_.~--._~._~,_~ pendent modulus of elasticity:
500 550 600 650 700 750 800 E drops to ::::; 1/10 of its value
TI"C at room temperature
342 5. Reshaping at High Viscosities

(see Fig. 5.3b,c). This behavior is in contradiction to the observations in the


real process. Because the temperature distribution is known to be correct,
the mechanical material behavior must be the reason for this discrepancy.

Temperature-Dependent Shear Modulus of Elasticity

In the second part of the simulation, we used the same temperature regime,
except for the measured temperature-dependent modulus of elasticity. The
sheet now deforms much faster at higher temperatures because its stiffness
has been significantly reduced. When the reshaping temperature is reached,
the sheet has already deformed significantly (see Fig. 5.4a), which corresponds
to the real observations. Additional vacuum suction induces a pronounced
sagging of the sheet into the mold, even at the upper edges (see Fig. 5.4b,c).
The simulation now corresponds very well with the real process.

a)

high

T low
b)

~
high

T low
c)

~
high

Fig. 5.3. Temperature distribution and deformation (a) after reaching the reshap-
ing temperature (initial sagging by gravity), (b) 1 s after starting additional vacuum
suction, (c) 2 s after starting additional vacuum suction
5.2 Sagging and Pressing of Glass Sheets 343

a)

high

T low
b)

~ ./
high

T low
c)

high

Fig. 5.4. Temperature distribution and deformation of the sheet after (a) reaching
the deformation temperature (sagging by gravity), (b) 1 s of additional vacuum
suction, (c) 2 s of additional vacuum suction

Summary

The comparison between the simulations of a reshaping process and the ob-
servations revealed very dearly that the temperature dependence of the shear
modulus of elasticity has a strong influence on the material behavior at tem-
peratures ~ 100°C beyond the transformation temperature T g . Reliable sim-
ulations must account for this by an appropriate modification of the standard
viscoelastic material model.

5.2 Sagging and Pressing of Glass Sheets

Thoral! Johansson
Introduction

Curved glass sheets are produced from plane sheets by thermomechanical


shaping. The process analyzed in this report consists of the following steps:
344 5. Reshaping at High Viscosities

1. Positioning of the glass sheet on a bottom mold - the sheet is supported


by the boundary of the mold (see Fig. 5.5).
2. Preheating - the sheet is preheated by the warmer bottom mold until it
reaches the heating zone in the furnace.
3. Heating and sagging - by the upper heating, the sheet is heated to a
temperature higher than the transformation t emperature T g . The sheet
sags due to its own weight (see also movie on CD).
4. Pressing of the final shape - a plunger presses the sheet to its final shape,
simultaneously cooling it.
5. Lifting and cooling of the transformed sheet.
The final shape of the workpieces must fulfil different specifications. The
best shaping results are reached if the deformation due to sagging already
corresponds as closely as possible to the final shape. The obj ective of the
present process is a spherical shape. The simulations were intended to opti-
mize temperature control.

5.2.1 Model and Boundary Conditions

The shaping process takes place on a frame. Round plane sheets are posi-
tioned on the boundary of a round mold. The sheet is transformed by sagging
due to its own weight and by subsequent pressing with a plunger. The sagging
of the sheet is influenced by the geometry of and the temperature distribution
on the bottom mold. If the temperature differences in the workpiece are too
large, the transformation is inhomogeneous, so that large deviations from
the desired shape occur. When the pressure is relieved, there is an elastic
spring-back of the sheet , which the plunger geometry must account for.
Process parameters influencing the shaping process are the geometry of
the bottom mold, the temperature distribution on the bottom mold, and the
upper heating temperature. In order to examine the bottom mold tempera-
ture, a parabolic temperature distribution of the bottom mold was assumed
(according to existing thermographic measurements):

(5.7)

Radiator

Original glass sheet

Form Fig. 5.5. Position of the sheet in the furnace


5.2 Sagging and Pressing of Glass Sheets 345

The boundary of the sheet is supported by the bottom mold, therefore a


linear heat transfer was modeled there:

(5.8)

For the rest of the sheet, heating of the glass through thermal radiation of
the bottom mold was implemented:

(5.9)

In these equations, j represents the heat flow into the sheet, TUF the bottom
mold temperature, TK the sheet temperature, and TM the temperature in the
center of the sheet.
On heating the sheet, the upper radiation heater has a decisive influ-
ence on the temperature distribution in the sheet, which considerably deter-
mines the shaping process. Therefore, as a further process parameter in the
model computations, the upper heating temperature was varied. The radia-
tion heater is assumed to be parallel to the sheet.
The heat flux towards the upper surface of the sheet is inhomogeneous
and results from the view factor F12 of two parallel circular sheets:

(5.10)

with

F12 ="21 { l+(h/rd 2 +(rdrd 2 - [1 + (h/rd 2+ (r2/rd 2] 2- 4 (rdrd 2} .


(5.11)

1>12 is the heat flux reaching the sheet from the upper heating, and 1>1 is the
heat flux radiating from the upper heating. One can see that the input of heat
energy is larger in the center of the sheet than at its boundary. The model
computations were carried out in a coupled thermomechanical analysis with
the commercial finite-element software ABAQUS.

5.2.2 Results of the Model Computations

The sagging of the sheet is due to gravitational effects on the glass, which
"softens" at high temperatures. The objective of this process phase is a trans-
formation of the sheet that corresponds to the desired final shape as closely
as possible. In the present case, the final shape of the sheet is intended to be
spherical with a sagging depth of 22.5 mm in the center.

Influence of the Upper Heating Temperature

First, three different upper heating temperatures were tested with a plunger
temperature Tpll remaining constant. Figure 5.6 shows the resulting sagging
346 5. Reshaping at High Viscosities

22.5

_____ After pressing +


elastic spring-back
After sagging
Goal

T,

Fig. 5.6. Dependence of the sagging depth in the center of the sheet on the upper
heating temperature (plunger temperature T pll )

depths in the center of the sheet. At the temperature T 2 , there is the smallest
deviation from the desired shape. For T 3 , the temperature is so high that the
deformation is too strong. In this case, the plunger no longer touches the
sheet.
The spring-back of the sheet after the relieving of the pressure, which is
caused by the elastic forces occurring during the deformation, also strongly
depends on the upper heating temperature. If the radiation temperatures
increase, the elastic part of the deformations is evidently reduced.
Figures 5.7-5.9 show the temperature profiles and the sagging depths for
the radiation temperature T2 in the different phases of the process.

Influence of the Plunger Temperature

The next step was the variation of the plunger temperature at the constant
upper heating temperature T 2 . Table 5.1 illustrates the sagging depth in the
center for two different plunger temperatures. The results show a further
reduction of the elastic spring-back. The displacement of the center of the

T low U3 VALUE
a) b)
:rg~d
.. .,....
Z displacement
·1'3£.f)2
·112E.(I2
·17IE-G2
·16QE-(t2
-, ,9(-02
·131E-02
·121E-G2
·1.17£-02
., OIEo02
-9s5E:-oa
.,.7E-03
-7.39E.03
131E-Ol
-5.,.-oJ
... 15£-oa
Time: 126 s -3.07E-03
-'19£-00 Time: 126 s 2
high ·910£.(M 1
... , HE-04

Fig. 5.7. (a) Temperature profile after sagging and (b) sagging depth for the upper
heating temperature T2
5.2 Sagging and Pressing of Glass Sheets 347

T low a)
Temperature
z displacement b)

3
Time: 132 s 2 3
1 2
high Time: 132 s
1

Fig. 5.S. (a) Temperature profile and (b) depth after the pressing phase for the
upper heating temperature T2

T low
b)

Time: 155 s 2 2
1 Time: 155 s
high 1

Fig. 5.9. (a) Temperature profile and (b) depth after relieving for the upper heating
temperature T2

sheet is only 22.3 mm in the final shape, i.e., the deviation from the desired
shape is reduced to only 0.2 mm in this simulation.

Deviations from the Desired Shape

The concluding evaluations compare the shapes resulting from the simulation
with the desired final shape. Figures 5.10 and 5.11 illustrate deviations of the
sagging depth of the sheet from the desired shape. Obviously, the configu-

Table 5.1. Influence of the plunger temperature on the displacement of


the center of the sheet

Tplunger Max. displacement of the center of the sheet


after heating and sagging after pressing and relieving
°C mm mm
18.9 21.2
18.9 22.3
348 5. Reshaping at High Viscosities

o ,---------------------------~
- - Sagging a) b)
- - Loading
-5 - - Unloading
e · • •_ e Optimum
E -10

~
<l -15

-20

- 25 +-~-.~--.-~_r----._-~~
o 50 100 150 200 250
r /mm

Fig_ 5.10. (a) z displacements after the single process phases for a radiation heater
temperature T 2 /plunger temperature Tpll and (b) corresponding deviation from the
desired shape

o .---------------------------~ a) b)
- - Sagging
- - Loading
-5 - - Unloading
--.-- - Optimum
l10
~

<l -15

-20

-25 +-_--.-_ _,.-_-,.-_--,r--_-.---l


o 50 100 150 200 250
r /m m

Fig. 5.11. (a) z displacements after the single process phases for a radiation heater
temperature T 2 /plunger temperature Tp12 and (b) corresponding deviation from the
desired shape

ration with the plunger temperature Tpl2 yields the best results also with
respect to the quality of the entire shape of the sheet.

Summary

Model computations with the FEM software ABAQUS have shown which pro-
cess parameters have the decisive influence on the quality of the shaping.
Particularly, the temperature of the upper radiation heater and the plunger
temperature decide whether the quality is sufficient. For sagging and pressing,
optimization possibilities are indicated (temperatures of the upper heater,
bottom mold, and plunger).
5.3 Calibration of Glass Tubes 349

5.3 Calibration of Glass Tubes

Thoralf Johansson
Situation and Objective

Glass tubes are calibrated by shrinking them onto metal mandrels by exterior
pressure at adequately high temperatures. The precise inside diameter of the
tubes results from the defined outside diameter of the metal mandrels. The
simulations were intended to show optimization potential for existing pro-
duction processes (particularly with respect to a reduction of process time).

5.3.1 Model Description

The glass tube is heated by a vertically moving circular radiation furnace.


The heat-flux density j through the outer wall of the glass is determined by
thermal radiation. Here, for the heat flow between the furnace and the outer
tube wall, the Stefan-Boltzmann law of radiation holds:

(5.12)

(Tfurnace, T tube : furnace and exterior tube temperature; emission factor c =


0.95). The heat losses of the tube due to convective cooling in the atmosphere
are sufficiently accounted for by an effective furnace temperature. As a refer-
ence for the quality of the furnace model, temperature curves were used that
had been measured experimentally by a thermocouple in the glass during the
calibration process.
The thermal coupling between the inner tube wall and the metal mandrel
outside of the contact area resulted from radiation with an assumed emission
factor of c = 0.9, analogously to (5.12). During the contact phase of the inner
tube wall and the metal mandrel, the heat transfer was defined to be linear:

j = -A (Ttube - Tmandrel) . (5.13)

(Tmandrel, T tube : outside temperature of the mandrel and inner tube temper-
ature). A constant heat transfer coefficient was assumed.
During the simulation, the furnace first remained at the bottom of the
tube until the tube was heated to a sufficiently high temperature. Then,
the furnace moved upwards at constant velocity, generating a quasi-steady
state in the heated zone (stationary in a moving coordinate system). In the
following, the results for this zone will be discussed. The mechanical loads
of gravitation and air pressure at the outer wall of the glass were taken into
account.
In the simulation models, the glass tube and the metal mandrel were
fixed at the bottom in the direction of the axis, the top of the glass tube
was moving freely, and the furnace was moving upwards (see Fig. 5.12). As a
350 5. Reshaping at High Viscosities

Vacuum

Glass

i i
I I Oven

....... Metal
Fig. 5.12. Sketch of the process

consequence, the glass tube can be "squeezed upwards" in the direction of the
axis during the shrinking on the metal mandrel. The models were computed
with the commercial FEM software ABAQUS.

5.3.2 Results of the Model Computations

Based on a standard tube and tool geometry, the processes listed in Table 5.2
were simulated. Figure 5.13 shows a section of the model. We first analyze
process I, which corresponds to the actual state. Then, the other process
variants are presented.
Process 1: Figures 5.14 and 5.15 show the results for the established stan-
dard process. Figure 5.14a illustrates the temperature profile in the glass
wall and in the metal mandrel in the transition region between separation
and contact. The maximum temperature in the tube is reached shortly before
the contact.
During the contact, a large amount of heat is transferred from the glass
tube (Fig. 5.14b) to the metal mandrel due to the temperature difference
between the inner wall of the glass and the surface of the metal mandrel,
according to (5.13). This also involves the transition from a free inner tube
wall to a pressurization (Fig. 5.14c) through the mandrel due to contact.

Table 5.2. Process parameters (THI > Ti and VHl > Vi)
Process Max. temperature Feed velocity
radiation furnace
°C
1 = standard
2
3
4
5
6
5.3 Calibration of Glass Tubes 351

f - - - - - Oven middle

Metal Glass Fig. 5.13. Section of the model with


rod positions of the measuring points

When the temperature decreases, the inner wall of the glass is again detached
from the mandrel due to the stronger thermal contraction of the metal.
Figure 5.15a shows the temperature profile at three points of the glass
tube (see Fig. 5.15b) with identical vertical positions, and at one point of
the mandrel. After continuous temperature increase at the glass tube points
up to a maximum, the temperature decreases again by t:.Tk due to contact.
In all the figures with temperature profiles, the maximum changes of the
temperature at the outer and inner glass-tube wall are clearly evident. They
are a measure for the thermal shock the glass is subjected to.
Further input of heat leads to a new temperature increase until the ob-
served part of the tube leaves the furnace. Then, the temperature of the
glass tube decreases due to the losses caused by thermal radiation towards
the surroundings.

a) b) c)
Heat flux Stress
T ol w compr.

high + tension

Fig. 5.14. Process 1: (a) temperature profile, (b) radial heat-flux density, (c) radial
stresses in the contact region
352 5. Reshaping at High Viscosities

....---------;=:::-----,-- - - - - - , a) Outer surface of


- Glo Steel rod glass tube (GLo)
- Grm
GI- i
..... MS

b)

Time
Fig. 5.15. Process 1: (a) time-dependent temperature at the outer surface of the
glass tube (GLo), in the center of the glass wall (GLm), at the inner glass-tube
wall (GLi) , and at the metal surface (MS), (b) positions of the respective glass-
tube points

The surface of the metal mandrel receives the thermal radiation of the
inner tube wall, which is the reason for the delayed temperature increase
of the mandrel with respect to the glass. The contact with the inner tube
wall leads to another small temperature increase of the metal surface, before
the temperature of the mandrel also starts to decrease. Owing to the better
thermal conductivity of the metal mandrel, there is an overlapping of the
temperature profiles during cooling.
The radial displacements

Ur = tlr (t) = r (t) - r (0) (5.14)

(r(O) = ret = 0): initial radius) in the glass tube according to Fig. 5.16 repre-
sent the temperature profiles and the properties of the glass at the different
temperatures. Due to the increased temperature, the glass expands in the
beginning. The air pressure does not cause any large deformations because
the glass is still sufficiently stiff. In the vicinity of Tg (= glass transition
temperature), the glass becomes softer and the air pressure causes a strong
deformation of the glass, which is limited by the mandrel. At the time of the

............. ~ ..
t3
Ur c'o" Fig. 5.16. Process 1: time-dependent
radial displacements U r at the outer
U
surface of the glass (G Lo), in the cen-
ter of the glass wall (GLm), at the
inner glass-tube wall (GLi), and at
Ie - the metal surface (MS); the positions
of the respective glass-tube points are
Time given in Fig. 5.15b
5.3 Calibration of Glass TUbes 353

contact, the thermal shock effect is also evident in the displacement curve.
After the first contact, the glass is again detached from the mandrel for a
short time before it fits completely close. The displacements in the glass are
almost identical, except for the time of the contact between glass and metal
mandrel.
In order to understand the calibration process, t he process parameters
were varied according to Table 5.l.
During process 2 (see Fig. 5.17), the temperature of the outer glass-tube
surface increases too much because either the furnace temperature is too
high or the feed velocity too low. Here, already small disturbances (e.g. ,
small temperature fluctuations) can lead to strong deformations of the tube.
The discrete time steps of the FEM computations cause small temperature
differences between neighboring points, which are, nevertheless, sufficient for
a non-uniform deformation of the outer tube surface. Figure 5.18 shows the
waviness of the outer tube surface after the tube left the furnace (depending
on z; the trend is due to the cooling of the tube). Although this waviness is a
numerical artefact due to time discretization, such "numerical instabilities"
usually also indicate real process instabilities.
Process 3 (see Fig. 5.19) leads to a lower maximum temperature of the
glass surface than process 2. Thus, a non-uniform surface of the outer tube
wall is avoided. The higher feed velocity results in a lower temperature of the

a)

- Glo
T - GI- m
GC)
..... MS

Time

b)

u
~o
u

Fig. 5.17. Process 2: (a) time-


dependent temperature and (b)
time-dependent radial displace-
Time ments U r
354 5. Reshaping at High Viscosities

Fig. 5.18. Process 2: waviness of the


Vertical position outer tube surface

metal mandrel in the pre-contact phase. The resulting larger temperature


difference between the inner tube wall and the mandrel leads to an evidently
stronger temperature decrease than in the standard process. The displace-
ment curves (Fig. 5.19b) clearly show an increased wall thickness of the glass
tube.
During process 4, the selected temperature of the radiation furnace is
too low. The glass cannot be deformed sufficiently, and there is no contact
between tube and metal mandrel. In the temperature curve (Fig. 5.20a) and
in the displacement curve (Fig. 5.20b), the sharp bending accompanying the
contact is missing.

a)

Time

b)
~-' \l
U
1\1 - GI_o
ur 'E - GL m
0
u GI i
- MS
Fig. 5.19. Process 3: (a) time-
1_ dependent temperature and (b)
\" ~
time-dependent radial displace-
Time ments U r
5.3 Calibration of Glass Tubes 355

a)

No contact

- GLo
T .... - GL m
GI i
........... ..... MS

Time

b)
~} ............................................. .
No contact

- GLo
Ur - GI_m
GU
.... MS
Fig. 5.20. Process 4: (a) time-
dependent temperature and (b)
time-dependent radial displace-
Time ments U r

Processes 5 and 6 differ with respect to the feed velocity. The resulting
temperature profiles show a basically different behavior, starting with the
first contact. The gradient of the glass temperature in process 5 (Fig. 5.21a)
is evidently larger than in process 6, because due to the lower feed velocity
the interaction during contact takes longer. The glass temperature decreases
until it almost equals the temperature of the metal mandrel.
During process 6, the glass tube is detached from the metal mandrel
after a very short contact time. Therefore, the glass temperature (Fig. 5.22a)
decreases more slowly, so that the temperature increase of the metal mandrel
is also distinctly smaller. Because of the very short contact, the inner diameter
of the glass tube is not exactly calibrated. Obviously, the feed velocity of
process 6 is too high.

Summary

The model computations of processes 1, 3, and 5 (see Table 5.1) led to the as-
sumption that a higher furnace temperature can increase the process velocity.
As the simulations showed, the feed velocity can be increased by about 30%
without decreasing the precision of the calibration. However, owing to the
higher feed velocity, the temperature of the inner glass-tube wall decreases
more strongly during contact. Thus, the possibility of increasing the feed
velocity is limited by the thermal shock resistance of the glass tube.
356 5. Reshaping at High Viscosities

a)

- GI_o
- GI_m
GU
..... MS
".

Time

b)
~.J···'·II ···"""·"·"""""""""""""·"

'0
CII - Glo
C - GCm
o
U GI i
... MS
Fig. 5.21. Process 5: (a) time-
dependent temperature and (b)
time-dependent radial displace-
Time ments U r

a)

- GL o
- GI_m
T GU
.. MS

Time

............ .............. b)

- GI_o
- GI_m
Ur GU
..... MS

Fig. 5.22. Process 6: (a) time-


I--. dependent temperature and (b)
-- time-dependent radial displace-
Time ments U r
5.3 Calibration of Glass Tubes 357

The simulations clearly show which processes will not result in successful
calibration:
• During process 2, either the furnace temperature is too high or the feed
velocity is too low, and therefore the temperature of the outer tube surface
becomes too high. This indicates instabilities during the forming process.
The FEM simulation shows an axial waviness of the outer glass tube sur-
face.
• During process 4, either the furnace temperature is too low or the feed
velocity is too high, so that the necessary contact between glass tube and
metal mandrel does not take place.
• Although the feed velocity is also too high during process 6, nevertheless
a contact between glass tube and metal is established. However, the con-
tact time is too short, so that the metal mandrel cannot define the inner
diameter precisely enough.
In simulation, successful calibration of a glass tube is most clearly indi-
cated by a sharp bend in the temperature profile of the glass, which proves
that appropriate contact took place.
6. Thermal Treatment

6.1 Verification of Relaxation Models

Peter Thomas, Clemens Kunisch


Introduction
The thermomechanical material behavior of glass during the cooling process
from the liquid to the glassy state is very complex, because it is influenced
by viscoelasticity as well as by structural relaxation. The widely used model
by Tool-Narayanaswamy ("TN model") is able to model these processes very
well under certain conditions. Of course, a fundamental requirement is the
knowledge of the thermal conditions. However, in real production processes
it is exactly the thermal simulation that causes large uncertainties, because
the initial and boundary conditions are often not completely known and mea-
surable. If the discrepancies between the computed and measured mechanical
stresses are unacceptably large, one must always answer the difficult question
whether the entire model is unsuitable for this case or whether it is "only" the
temperatures that are not sufficiently modeled. Exactly such a situation was
the reason for the present study. The concrete process partly implied high
cooling rates (molding and subsequent transportation by a cooling belt), and
it was not evident whether applicability of the TN model might be restricted
by uncertainties of the material parameters resulting from laboratory exper-
iments.
In order to examine the efficiency of the TN model, laboratory exper-
iments in a rapid lehr with different cooling conditions were carried out.
Unlike in the mold, the history of the temperature in the lehr is well known,
i.e., it can be neglected as a source of errors. Thus, the computed stresses
only depend on the material behavior. The temperatures measured during
the experiments were taken as the direct basis for the parallel simulations of
the experiments. Our study indicates and explains the difficulties typically
connected with such experiments.

6.1.1 Mathematical Models


We begin with a very condensed presentation of the mathematical models
applied for the relaxation behavior of glass in the transition range, and the
relation between stresses and birefringence.

H. Loch et al. (eds.), Mathematical Simulation in Glass Technology


© Springer-Verlag Berlin Heidelberg 2002
360 6. Thermal Treatment

Relaxation

In order to start with well-defined conditions, all prehistory effects in the


glass must be avoided. Therefore, we choose a fixed reference temperature
that is high enough to establish internal equilibrium, i.e., at the beginning
of the computation the glass is totally relaxed. By using this fixed reference
temperature Tref, a reduced time

~(t) = j exp [D.H


t (1 x 1- x)]
R T ref - T(t) - Tf(t) (6.1)
o
is defined. R is the gas constant, x and D.H are the material parameters. The
internal state of the glass is given by its fictive temperature. It is calculated
from the implicit differential equation system
n
Tf(t) =L WkTfk(t) (6.2)
k=1
Tfk ( t) = _ Tfk (t) - T( t) ~(t) (6.3)
Tk,ref

with the weighting factors Wk and the relaxation times for the glass structure
Tk,ref in n parallel-arranged Maxwell models. For the weighting factors, it
holds that
n
LwZ=1. (6.4)
Z=1

In an unconstrained volume element of glass, temperature changes lead to


the thermal strain

(6.5)

with the thermal expansion coefficients Ct g and Ctz of the glassy and liquid
state, respectively, depending on temperature, not on time. For the calcula-
tion of stresses from strain, the shear modulus and bulk modulus are used
instead of Young's modulus and Poisson's ratio. The shear modulus

G (~(t)) = Go Lm

Z=1
[~(t)]
Vz exp - -
TZ,ref
(6.6)

(Go = instantaneous shear modulus) is no longer constant. Instead, it is time


dependent. For the weighting factors,
m

(6.7)
6.1 Verification of Relaxation Models 361

holds again. The shear stress no longer follows from Hooke's law (iT = G E),
but from the Maxwell relation

J (~(t)
t

iT(t) = G - ~(t')) E(t')dt' (6.8)


o
with the strain history E(t). Contrary to the shear modulus, we assume that
the bulk modulus does not depend on time.

Stress Birefringence

Stresses change the refractive index n of isotropic materials such as glass by

(6.9)

with

(6.10)

(6.11)

(6.12)

(6.13)

ql and C -.L are the stress optical coefficients for light polarized parallel
and perpendicular to the stress orientation, respectively. For real materials,
t1n « n holds.
For pure hydrostatic pressure p = -iT this simplifies to

(6.14)

For uniaxial stress, the path difference t1¢ between two waves of light prop-
agating in the x direction and being polarized perpendicularly to each other
can be calculated from

(6.15)

where D is the sample thickness and >'0 is the wavelength of light in vacuum.
More complicated stress distributions usually can be dealt with by the
linearization of equations, if the path of light is not significantly affected by
the stress field, which means that the stress differences must not be too high.
362 6. Thermal Treatment

6.1.2 Experiments in the Lehr

The experiments were carried out in a rapid lehr, which allows cooling rates
up to 80 K/min, depending on the size of the glass samples. The surface tem-
perature of the glass and the air temperature were measured by thermocou-
ples in the center of the samples. In order to measure the surface temperature
of the glass, a thermocouple was glued to a groove in the glass sample. Thus,
the thermocouple measures only the surface temperature of the glass instead
of a mixture of glass and air temperature.
The samples consisted of the glass used in the real production process. In
order to measure stress and compaction, different sample sizes were used:
• stress samples: (100 x 100 x 14.6) mm 3 ,
• compaction samples: 5 mm, 100 mm long.
In the lehr, the samples were positioned on a rack, so that their bottom side
was cooled as closely as possible in the same way as their top side. Figure 6.1
shows the position of the samples in the oven, seen from above.
Two different types of experiments were carried out:
• continuous cooling at a constant cooling rate,
• continuous cooling, interrupted by a soak temperature.

Continuous Cooling at a Constant Cooling Rate

In four experiments, the glass samples were cooled at constant cooling rates
of 2, 5, 12, and 50 K/ min. The advantage of this type of experiment is that
it does not depend on an exact t emperature measurement, because due to
the constant cooling rate a stationary temperature profile develops over the
glass thickness, which is responsible for the stress. Thus, the stress essentially
depends on the cooling rate; the absolute temperature is of secondary impor-
tance. The starting t emperature for the cooling was chosen so high that in the
transformation domain there was already a stationary temperature gradient.
The cooling at this constant rate was continued up to ::::::: 100 °C below T g •

1"~'~"'"'~'M'M"" _ _-'-'-'-----'-'-'-'-'--'-'-'-'--'- "j


j i Oven wall
\i Stress sample
i;V
i !

Thermocouple groove

Fig. 6.1. Position of the sam-


Compaction samples ples in the oven (top view)
6.1 Verification of Relaxation Models 363

Cooling with Soak Temperature

In two further experiments, the samples were cooled at 5 K/min. Cooling


was interrupted for 15 min at 500°C, respectively 490 °C (soak temperature).
Here, the absolute temperature is essential. The inhomogeneities of the oven
became evident, and it was already difficult to measure the temperature
distribution for the entire sample consistently and to transfer the results to
the simulation.

6.1.3 Simulation

Each experiment in the oven was simulated by means of a finite-element


model. The measured temperature profile of the air temperatures with re-
spect to time was directly used as an input parameter. The computed glass
temperatures were again compared with the measured glass temperatures.
For the simulation, the temperature conditions in the surroundings of the
sample were assumed to be homogeneous, i.e., the air temperature measured
in the center of the sample was assumed for the entire sample. The heat trans-
fer by forced convection (fan in the oven) over the sample was also assumed
to be constant. The reason for these simplifications is not the simulation it-
self; they are necessary because of difficulties in gaining measurement data
with adequately high spatial resolution.

6.1.4 Measuring Stress and Compaction

Measuring the entire stress field is technically impossible; therefore measure-


ments are carried out only at selected local points and with different methods.
Because the samples with low cooling rates have small residual stresses, only
the stress-optical path difference could be measured with sufficient exact-
ness for comparison with the simulation results. Only for the sample with a
cooling rate of 50 Klmin, could the surface stress be measured directly. For a
further characterization of the cooled glass, the compaction behavior was also
examined with a standard temperature program relevant for this product.

6.1.5 Results

Temperature

For the experiments with a constant cooling rate, the simulation of the mea-
sured temperatures was successful. Figure 6.2 shows a very good correspon-
dence between measurement and simulation for the cooling experiment at
5 K/min. The air temperatures we measured and used as input parameters
for the simulation (green curve in Fig. 6.2) yielded a computed glass sur-
face temperature (cyan curve) that is identical to the measured glass surface
temperature (red curve).
364 6. Thermal Treatment

570c-----------------------------------~

520
470
~ 420
Q) 370
:i
li!
Q)
320
~ 270
~ 220
170 Glass temperatures simulated
Glass temperatures measured
120 Air temperatures in simulation and measurement
70 +-----.-----r---~----~----_.----~~
o 600 1200 1800 2400 3000 3600
Time Is

Fig. 6.2. Comparison of the measured and simulated temperatures at a cooling


rate of 5 K/min

Figure 6.3 shows a cooling run for which the agreement is less perfect.
In this example with a soak plateau, a temperature of 475°C was reached,
which is distinctly below the nominal soak temperature of 490 DC. In this
experiment, the relation between air and glass temperature was not always
conclusive, so that additional assumptions became necessary for the simu-
lation. In order to receive consistent results, the temperature control in the
batch oven and the temperature measurement must be further improved.

550,---------------------------------- .
535
520
u 505
~
~ 490
::>
li!
Q)
475
~ 460
~ 445
430 Glass temperatures sImulated
Glass temperatures measured
415 Air temperatures In Simulation and measurement
400 +-----.----.-----,----.-----r----.~
o 400 800 1200 1600 2000 2400
Time Is

Fig. 6.3. Comparison of the measured and simulated temperatures at a cooling


rate of 5 K/min with I5-min residence time at 490°C
6.1 Verification of Relaxation Models 365

Stress

For the sample cooled at a rate of 50 K/min, the stresses were relatively high
and could be measured in our lab, so that direct comparison with the sim-
ulation became possible. For reasons of symmetry, it is sufficient to model
only 1/16 of the quadratic sample for the simulation. Figure 6.4a shows the
position of the simulation model (red) within the entire glass sample (blue)
and the position of the measurement path for the stress measurement. Fig-
ure 6.4b shows the stress distribution for this part of the glass sample, as
well as the point and the result of the stress measurement with the GASP
device. The groove for the thermocouple prevents measurements being taken
exactly in the center of the sample. The measured stress is 31 MPa, the com-
puted stress in the measuring area is 28-32 MPa. Thus, measurement and
simulation correspond relatively well.
For the samples with lower cooling rates, the residual cooling stresses in
the glass could only be measured with sufficient exactness via the stress bire-
fringence. Along a measurement path from the center of the sample towards
the boundary (Fig. 6.4a), the stress-optical path difference was measured over
the sample thickness.
Figure 6.5 shows the measured and the computed path difference for the
experiments at constant cooling rates. The relative changes from run to run
in the computations agree very well with the experiments, but a systematic
absolute difference between measurement and computation can be observed.

a)

SX b)
- :~~
- ·32
Measured: 31 MPa - ·28
- ·24
- :~~

- -4
-4
- . 12
- ·8
- 0
8
12
_ 16
_ 20
- ~:
Fig. 6.4. (a) Position of the simulation model within the glass sample and po-
sition of the measurement path for the stress measurement, (b) computed stress
distribution and a measured result (cooling rate 50 K/min)
366 6. Thermal Treatment

2500,_----------------------------~--,

Measurement
Sample 2: 2Klmin
Sample 1: 5Klmin
Sample 3: 12 Klmin
Sample 4: 50 Klmin

-500 +-~--.--~___r-~---,;--~-,_~-_._----1

o 10 20 30 40 50
Distance from sample center Imm

Fig_ 6.5. Comparison of the path difference between measurement and simulation
at constant cooling rates

This difference can, for example, be due to the fact that the conditions within
the lehr (in contrast to the simulation) are not completely symmetrical (e.g.,
temperature differences between front and back, top and bottom).
In Fig. 6.6, the results from the experiments with a soak temperature are
presented, and the result from the 5 Kjmin constant-cooling experiment is
shown for comparison. Here, the relative changes due to the soak also fit the
experiments. However, the small stress differences measured between the two
experimental runs with a soak do not show up adequately in the simulation:

350 ,_--------------------------------,
Measurement Simulation
E 300 Sample 1: 5 Klmin
c Sample 5 : 5Klmin, 490 · C
::; 250 Sample 6 : 5Klmin, 500 · C
Ql

a;
<J
200
Cii
~ 150
, ,
.c ,, , ,
~ 100 ,, ,,
, +
~
<J 50 ,,'..... "
E. ".;,,-,'
o
° t---~~~~~~~~-~
~-~-~------1
-50 +-----___r-~-,_---___r-~-..._--_r'

o 10 20 30 40 50
Distance from sample center Imm

Fig. 6.6. Comparison of the path difference between measurement and simulation
for the cooling with a soak temperature
6.2 Stresses and Crack Growth in Continuously Formed Slabs :367

400 ~-------------------------------------,

350 Measurement
_ Simulation
300
E
a.
.e- 250
c:
o
'5 200
'"
a.
E 150
5 Klmin
490 ' C
o
()
100

50

o
Experiments in the lab oven

Fig. 6.7. Compaction: comparison between measurement and simulation

for both experiments, practically the same path difference is computed (the
blue line corresponds exactly to the green line in Fig. 6.6!). This inaccuracy
is mainly due to difficulties in temperature measurement.

Compaction

Figure 6.7 compares the experimental and computed results for compaction.
The relative changes in compaction agree very well, but again the absolute
values show a constant difference between measurement and computation.
The simulation corresponds to approximately 90% of the measured results.
For the experiment at 50 K/min, it corresponds to 80%.

Conclusion

The data for the TN material model were determined with sufficient ex-
actness. The model is also appropriate for processes implying high cooling
rates, such as the molding process we examined. Our study again proved that
knowledge of the time-dependent temperature distributions is indispensable
for an exact computation of the cooling states.

6.2 Stresses and Crack Growth III


Continuously Formed Slabs

Clemens K unisch
In glass production, the way the glass is cooled down to room temperature
after the forming process governs important quality aspects of the glass and
368 6. Thermal Treatment

the final product. Glass may even be destroyed if cooling stresses make ex-
isting flaws (originating, e.g., from contact with the belt) grow to a critical
size.
The stresses can be classified in the following way:
• temporary stresses, which (mainly) depend on the actual temperature dis-
tribution at each instant in time, and
• residual stresses, which remain forever in the glass if it is equilibrated at
room temperature.
To a large extent, the latter ones depend on the temperature distribution in
the glass during the cooling process through the glass-transition range.
In this section, we analyze the temporary stresses in a continuously formed
strip during cooling in a lehr, and the resulting residual stresses after the strip
was cut into slabs. We use a finite-element model with the commercial soft-
ware ANSYS and model only the glass (Fig. 6.8). The influences of oven walls,
air, belt, etc., on the glass are formulated as boundary conditions. Unfortu-
nately, perfectly homogeneous lehrs are difficult to find in real production.
However, we assume such a perfect lehr in order to clarify the principal effects.
If we look at region I of the crack velocity versus stress intensity factor
relation, the increase of the crack length can be written as

da = v(T) (Kr(u,a))n (6.16)


dt K rc

(see The Properties of Optical Glass, Schott Series on Glass and Glass Ce-
ramics, 2nd edn., Springer, Berlin, Heidelberg 1998, p. 191, for explicit ex-
planation; a = crack length, t = time, K rc = room temperature fracture
toughness, u = stress, n = stress-corrosion constant) with the temperature
dependence restricted to

v (T) = Va exp (Ti2) (6.17)

(Va, Ta = constants), and the stress-intensity factor

Kr(u, a) = Y uVa (6.18)

Fig. 6.8. Finite-element model of a glass


strip section and half the slab, respectively
6.2 Stresses and Crack Growth in Continuously Formed Slabs 369

(Y = factor accounting for the geometry of the crack) assuming pure mode I
loading on the crack. Integration from step i to i + 1 yields

which can be calculated by using the results from the finite-element analysis
run.

6.2.1 Cooling a Continuous Strip

When the glass strip enters the lehr after forming, its surface temperature is
lower than the temperature of the bulk (Fig. 6.9). At the soak temperature,
the temperature differences in the strip decrease (compare Fig. 6.11a). After
entering the first cooling rate CR 1 , they increase to an almost constant value
that depends on the cooling rate (Fig. 6.9).
For geometrically simple products such as strips, the soak temperature
should be chosen high enough in order to enable the relaxation of stresses
induced by forming (Fig. 6.10).
After the soak period, the glass strip is cooled below the glass transition
range. Cooling usually starts with small cooling rates (C Rl in Fig. 6.9) to
keep the temperature differences within the strip low and thus to prevent
the development of residual stresses. Slow cooling allows the stresses to relax
almost immediately at temperatures well above the transition range, and the
strip is almost stress-free there (Fig. 6.10).
Temperature inhomogeneities still existing during freezing mean that the
various parts of a glass piece freeze in at different times. Figure 6.10 shows

500
..:. . . - ....'::'
Soak .
~ ~
Center
400 ~ ~
Edge
. ~

~~ Lehr
~ ~
,:-.
300 ~ .. :---
() CRl
ll.
I-
200

100

o+----.----,----.----,----.,----,--~
0.0 0 .5 1 .0 1.5 2.0 2.5 3.0 3.5
tlh

Fig. 6.9. Lehr temperature and temperatures in the glass strip versus time
370 6. Thermal Treatment

10 ~---------------------------------.

,
/ \
5 \

I \
\
'\
'\

Center Edge
- 10

- 15
• •
0 0
Equilibrated at RT
After cutting

0.0 0 .5 1.0 1.5 20 2.5 3.0 3.5


tlh
Fig. 6.10. Stress a zz at center and edge of the glass strip versus time (compare
with Fig. 6.9)

how stresses are induced by such a time lag. Owing to its lower temperature,
the glass at the edge freezes in before the glass in the center, forming a kind
of hollow rod inside which the still liquid glass portion starts contracting as
soon as it has cooled to the freezing temperature. At temperatures below the
glass transition range we therefore observe compressive stresses close to the
strip surface and tensile stresses in the strip center (Figs. 6.10 and 6.11 b).
When the cross-sectional temperature of the glass piece is low enough,
the cooling rate is usually increased from CR l to CR 2 . Because the glass
is already frozen in, the residual stresses (compression at the surface, ten-
sion in the center) are no longer influenced by this change. The temporary
temperature differences, however, increase with the cooling rate, and so do
the temporary stresses (tension on the surface, compression in the center,
Fig. 6.11c). Depending on how strongly the cooling rate increases, the first
principal stress at the strip surface changes from compression at CR l to ten-
sion at CR 2 (Fig. 6.10). Tension stresses at a glass surface are potentially
dangerous, because they can result in breakage if the induced crack growth
is too large.

6.2.2 Crack Growth

The lower the temperature, the lower is the thermal conductance and the
heat transfer coefficient. Therefore, a constant cooling rate of the oven results
in larger temperature differences in the strip, which lead to higher stresses.
These stresses drive the growth of flaws (Fig. 6.12). In the case shown here,
the temperature dependence of the subcritical crack growth speed is of minor
importance. In other cases, the decrease in temperature can even stop the
crack growth, although the stresses remain at a high enough level.
6.2 Stresses and Crack Growth in Continuously Formed Slabs 371

----
471 .475 a)
471 .946
472.416

---
472.887
473 .357
473.828
474.298
474.769
475.239
475.71

---
· 11
-9
-7

---
·5
·3
·1
1
3
5

--
7
·11

-
·9
·7
·5
3
5
7

---- 265
266.498
267.996
b)

---
269.494
270.992
~72.49
73 .988
275.486
276 .984
278.482

---
·11
-9
-7
·5
·3

-
·1
1
3

--
5
7
·11 Fig. 6.11. Temperature

--
·9
·7
-5
(OC) and stresses (MPa) in
3
5
the glass strip at the end of
7 (a) soak, (b) CR 1

Very small flaws do not grow at all (Fig. 6.13, black curve). Flaws with
sizes of about 0.25-0.31 mm grow by a few percent, whereas flaws larger than
~ 0.31 mm cause fracture and therefore must be avoided. If this is impossible,
the temperature profile in the lehr must be adjusted in such a way that stress
development in the final cooling phase in the lehr is reduced. This can be
achieved, for example, by increased cooling rates at the beginning and reduced
rates at the end of the lehr run. Suggestions for adjusted temperature profiles
can be checked in the same way until the calculated growth of presumed flaws
is reduced to an acceptable level.

6.2.3 Modified Temperature Program in Order to A void Cracking

If flaws with sizes> 0.31 mm are unavoidable, we can try to limit the crack
growth by modifying the temperature program in the lehr. The easiest way
is to change the temperature settings while maintaining the belt speed. To
reduce the maximum tensile stress at the end of the lehr run, we increase the
initial cooling rate and reduce the final cooling rate (e.g., by about 4% each),
372 6. Thermal Treatment

--
106.567 c)
111 .293
116.018
120.744
125.469

--
130.195
5
1 4.921
1 9 6. 46
144.372
149.097

--
· 11
·9
·7
·5
·3

--
·1
1
3
5

-
7
·11
·9

-
·7
·5
3
5
7

---- ·11
·9
·7
·5
·3
d)

--
·1
1
3
5
7

--
· 11
·9
·7
·5

-
·3
·1
1
3

-
5
7
·11
6.11. continued: (c) CR 2 ,

-
·9
-7
3
5 and (d) after equilibration
7 at room temperature

which means that we must lower the temperature setting in the middle part
of the lehr by :::::: 9 K. Generally speaking, these changes seem to have little
effect on the stresses in the glass (Fig. 6.14). Viewed in terms of our specific
purpose, namely to avoid breakage, they make a decisive difference. Owing
to the strong nonlinear dependence of crack growth on stress, the acceptable
flaw size increases by about 20% (compare Fig. 6.13, gray curve).
Of course, in the real world there is nothing for free. After equilibration at
room temperature, the glass coarsely annealed with the modified temperature
program will have slightly higher stresses than with the original temperature
program. Depending on the further use of the slabs, however, this should be
acceptable if cracks are avoided.

6.2.4 Cutting the Strip into Slabs

After equilibration at room temperature, there are residual stresses in the


strip (Fig. 6.11d) that change when the strip is cut into slabs. In the cutting
cross section, all stresses O"zz, T xz , and Tyz become zero. The stresses red is-
6.2 Stresses and Crack Growth in Continuously Formed Slabs 373

0.38,..---,---.----,---,--------,

0.36 aj/mm
- 0.29.
E - 0.30L .

-
0.34
'--6- 0.31.

co 0.32 .~.... 0.314

0.30

0.28 -t--...--...--_+_--+---+---r-----1
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5
tlh
Fig. 6.12. Crack size a(t) at the edge of the glass strip versus time for four different
initial crack sizes ai (compare with Fig. 6.10)

1.20 -,---,..----~--~--------,
___ Edge
___ Edge_mod.
1.15 I· I
I
<if
~ 1.10
.,

.. ; :
~
I

Ii I
~ III
1.05 ,. I
,. !II

~. .."
~·-·--~~·-~
1.00+-----,----.·1.. •.-~1II--·~----~
0.20 0.24 0.28 0.32 0.36 0.40
aj/mm

Fig. 6.13. Relative increase in flaw size a f / ai versus initial flaw size ai (compare
with Fig. 6.12)

10~----------__----------------,

5~···································'··········'····· . . . . • )'\, ................ .

tI!. 0
~
~-5~······IJ ....................~--~~-- .............................. ~ . . . . . . I

-10+·········· • "' .....................•.......

-15 +-----.-----r-----T-----;-------,----r-~
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5
tlh
Fig. 6.14. Stress (]" zz at the edge of the glass strip versus time for the original and
the modified temperature program (compare with Fig. 6.10)
374 6. Thermal Treatment

--
a)

-11

---
-9
-7
-5
-3
-1
1
3
5
7

b)

_ -11
_ -9
_ -7
_ -5
_ -3
-1
1
3
5
7

Fig. 6.15. Stresses in the strip before cutting (a) and in the slab after cutting (b).
For a close look inside the glass, only 1/8 of the slab is shown with the center of
the slab being closest to the reader

tribute in the slabs in order to achieve a new minimum level of stored elastic
energy (Fig. 6.15).

6.3 Thermal Tempering of Drinking Glasses

Volker Seibert, Andreas Bensberg


Introduction

Hollow glassware (e.g., drinking glasses) is rather sensitive to external me-


chanical influences such as shocks, particularly at the edges. Any chips or
cracks are, of course, potentially dangerous to the user of drinking glasses.
The risk of injury is even higher in the case of small flaws, which are often
not easily visible. If such pre-damaged glasses are exposed to temperature
differences, they can eventually break.
6.3 Thermal Tempering of Drinking Glasses 375

In the following we describe methods of thermal tempering of thin-walled


drinking glasses that have two objectives:
• The product strength is increased for normal usage by inducing a compres-
sive stress in the bulb edge, where the mouth comes into contact with the
glass .
• The product will break completely if critical damages emerge in the upper
edge region of the glass. This is achieved by inducing a tension stress just
below the bulb edge that magnifies initially small and potentially dangerous
flaws. This is advantageous for quality control (visibility of damages) and
for the safety of the user.

6.3.1 Principles of Thermal Tempering

Thermal tempering is by far the most important method for inducing sur-
face compressive stress. Tempered glass has an increased strength against
mechanical loads and temperature changes, and upon breakage it will disin-
tegrate into many small and obtuse pieces that are much less dangerous than
sharp-edged glass parts.
The first step in thermal tempering is to heat the glass to a homogeneous
temperature that is approximately 100°C above the transformation temper-
ature. Because of the viscoelasticity of the glass, all existing stresses will relax
within a short time. In the second step, the glass surface is cooled as fast as
possible by induced surface convection. The interior temperature decreases
much more slowly because of the relatively low heat conductivity. Thus, the
surface layer shrinks more strongly than the kernel, so that tension stresses
are induced in the surface and compressive stresses in the interior. These
stresses can partly relax at temperatures well above the transformation re-
gion. If the surface temperature approaches the transformation temperature,
these stresses increase. When the surface temperature is below the transfor-
mation temperature, the kernel (which is still hotter) continues shrinking for
a while, whereas the surface is already "frozen in" . The stress distribution in
the glass is reversed: now, the surface is under compression, and the interior
is under tension.
The intensity of the remaining permanent surface compressive stress ob-
viously depends on the transient temperature gradients between surface and
interior, which on the other hand depend on the wall thickness and the
heat convection at the surface. The maximum heat transfer coefficient to
be reached, for example, by air convection, is limited. Additional effects are
needed in order to achieve sufficiently high surface compression stresses in
thin-walled glassware. We will demonstrate for drinking glasses that inten-
tionally spatially inhomogeneous quenching processes can be used for this
purpose.
376 6. Thermal Treatment

6.3.2 Results for Spatially Inhomogeneous Quenching

The following finite-element models show how to induce permanent surface


stresses in the edge region by an appropriate inhomogeneous quenching pro-
cess. Axially symmetric models were used. The homogeneous starting tem-
perature of the glass was 610 °e, which was 100 °e above the transformation
temperature Tg = 511 °e for the chosen crystal glass. In the simulations
presented here, the viscoelastic behavior was described by a simple I-term
Maxwell model. The crystal glass had the following material parameters:
modulus of elasticity 70GPa
Poisson number 0.223
density 2537kg/m 3
specific heat capacity 1320J/kgK
heat conductivity 0.80 at ooe 1.17 at 400 0 e
(W/mK) 1.00 at 100 0 e 1.19 at 500 0 e
1.10 at 200 0 e 1.21 at 600 0 e
1.15 at 300 0 e
viscosity 1og -.!L ~ -1 52 + 4350.26
dPas ~. o~ -215.84 .

In the model, the quenching is represented by a linear heat transfer be-


tween the glass and a cooling medium. For a heat transfer coefficient up to
300 W /m 2 K, this can be realized by convective cooling with compressed air
(Vair ~ 20 ... 30 m/ s), which takes about 6 s. During this period, the bulb
edge is cooled to about 100 0 e below T g . Then, the glass is further cooled
to room temperature in a lehr. The blue area in Fig. 6.16 is the analyzed
region, the lower part of the drinking glass was unimportant for this simula-
tion. Figures 6.17 and 6.18 give the results of two representative simulations
showing all the relevant mechanisms. Figure 6.18 illustrates the temperature
distribution after 5, respectively 10 s.
Figure 6.19a shows the permanent axial and tangential stresses after
quenching at the outer surface. At the bulb edge, a tangential tension stress
is visible. Even small damages would induce splintering here. However, the
tension stress ceases just a few millimeters from the edge, and turns into
compressive stress in the lower part. Splintering will therefore occur only in
the small upper zone, and the glass probably will not disintegrate completely.
The images given in Fig. 6.19b illustrate an opposite behavior. Quenching
takes place only at the bulb edge, which results in a tangential compressive
stress at the inner and outer surface of the bulb edge. About 3 mm below,
the compressive stress changes into tension stress. In this case, damage in the
bulb edge region will probably induce complete disintegration of the drinking
glass.
The simulations were done with an axially symmetric model, i.e., the
heat transfer coefficient was supposed to be circumferentially constant. It is
6.3 Thermal Tempering of Drinking Glasses 377

a) b)

a =300 W/m2K

Fig. 6.16. Drinking glass Fig. 6.17. Boundary convection: quenching


geometry (a) at the outer surface, (b) at the bulb edge

--
a)
T lOG T lOG
415 _ 330

---
t=5s 426 t = 105
438 - ~;
449
461 - 380

-
- 397
472 - 414
484 - 430
495 - 447
507 - 464
518 - 480
530 - 497
541 = 514
553 _ 530
564 _ 547
576 _ 564
587
599 - ~~

---- ---
b)
T lOG T lOG
t =5 s 347
363 t = 10 S 283

----
302

-
378 322

---
394 342
410 361
426 381
441 401
457 420
473 440
489 460 Fig. 6.18. Tempera-
504 479
520 499
ture distribution in the
536 ~
518 cup: (a) quenching at
552 538
567 558 the outer surface, (b)
583 577
599 597 quenching at the bulb
edge

important to realize this in a concrete process because asymmetries could


produce transient stresses resulting in breakage.
378 6. Thermal Treatment

a) b)

Uax IMPa uax IMPa


-24 ·73 ·60 -112
-21 .9 ·67 ·53 -102.5
· 19.8 · 61 ·46 -93
-17.7 ·55 ·39 -83.5
-15.6 -49 ·32 -74
· 13.5 ·43 -25 -64.5
E · 11.4 ·37 -18 ·55
E -9 .3 ·31 · 11 -45.5
-7 .2 -25 ·4 -36
<Xl -26 .5
-5.1 -19 3
-3 -13 10 -17
·.9 ·7 17 -7 .5
1.2 -1 24 2
3.3 5 31 11 .5
5.4 11 38 21
7.5 17 45 30.5
9.6 23 52 40

Y
Fig. 6.19. Permanent stresses in the vicinity of the bulb edge: (a) quenching at
the outer surface, (b) quenching at the bulb edge

6.3.3 Realization of a Quenching Process

The next step in our work was the development of a concrete device in order
to quench the glass in the aforementioned way. Figure 6.20 shows a sketch of
such a device for drinking glasses.
Compressed air is blown into the interior volume via a cylindrical nozzle.
A small slit with a height of approximately 1 mm is visible at the bulb edge
where the air will escape to the exterior at high velocity. Sixteen bolts are
arranged circularly around the glass, each containing two nozzles that can
blow at the bottom of the drinking glass.
The air flow simulations were done with the commercial software FLUENT,
their result being the distributions of the heat transfer coefficients (depending
on time and location); see Figs. 6.21 and 6.22. These were the inputs for the
following temperature and stress computations with the commercial software
ANSYS.
The results in Fig. 6.23 confirm that this quenching process is able to
produce stress distributions of the desired type .

.....-Cylindrical nozzle

....-Glass

Fig. 6.20. Sketch of a quenching device


6.3 Thermal Tempering of Drinking Glasses 379

..!.. v
a) b)

---
·C ms- 1
Sll~ :2 55e+01 a I W (m 2 K)-1

---
" o
5081 •• 02 2_33.,01

5 4lt1+02 21'b..ol 26.25


Cylindrical nozzle 52.5
:S23e.G2 '11 • ..01 78.75
105
504. . 02 110..0 1

-
131.25
4 &.02 , 461+01 157.5
183.75
_.-02 121..01 210

---
-4 ... &..02 , oe,..01 236.25
262.5
.f.2?.Oi

.. oa.t02
1.461..00
.,....., = 288.75
315
341.25
367.5
3 .... 02 42 ....+00
393.75
36S40.02 :2 1:z...00 420
3so..02 , 310-0'

Fig. 6.21. (a) Air flow inside the cup (t 6 s) and (b) resulting heat transfer
coefficient distribution

---- 344
360
376
U ax IMPa Ulan IMPa

-
-73
392
408
- :~~.5 -67 .375

-
- -23 -61 .75
424 - -20.5 -56.125
440 - -18 -50.5
456 - -15.5 -44 .875
- -13 -39.25

---
472 - -10.5 -33.625
488 - -8 -28
504 - -5.5 -22.375
520 - -3 -16.75
536

-
- -.5 -11 .125
552 24 .5 -5.5
_ .125
568 _ 7
5.75
584 11.375
600 ~~ 17

Fig. 6.22. Temperature in Fig. 6.23. Permanent stresses near the


cup and foot (t = 6 s) bulb edge
7. Post-Processing by Laser Cutting
Kurt N attermann, Christoph Hermanns

Introduction

There are two methods for brittle cutting of glass sheets, namely the full-body
and the micro crack cutting. We will describe the relevant processes, how to
model them, and how to find advantageous process data via simulation.
In full-body cutting (see Fig. 7.1), the glass sheet is heated along the de-
fined cutting path with a laser beam to a temperature below T g • The result-
ing mechanical stresses cause crack propagation throughout the sheet. This
method is suitable for thin glass sheets (glass thickness d < 1 mm, depending
on the glass properties) and has the following advantages: with well-adjusted
process parameters the cutting direction can be precisely determined, non-
splintering smooth cuts result, and the glass temperature remains below T g .
For thick sheets, microcrack cutting (see Fig. 7.2) is the method of choice.
First, a micro crack is produced at the surface and then the actual separa-
tion of the glass sheet is done by subsequent breaking. Basically, microcrack
cutting means "laser scribe" and break.

7.1 Rough Estimation of Process Parameters

7.1.1 Stress Levels

Both full-body cutting and micro crack cutting cause mechanical stresses as
a result of inhomogeneous heating of the glass. The stress levels attainable
in this way can be estimated as follows:

Cutting path

Fig. 7.1. Full-body cut Fig. 7.2. Microcrack cut

H. Loch et al. (eds.), Mathematical Simulation in Glass Technology


© Springer-Verlag Berlin Heidelberg 2002
382 7. Post-Processing by Laser Cutting

a r:::; ~aE D.T , (7.1)

where E is Young's modulus, a is the coefficient of thermal expansion, and


D.T is the temperature difference caused by laser heating. Based on (7.1) we
can state a simple fact: glasses with extremely small coefficients of thermal
expansion such as, for example, "fused silica" , or "zero-expansion materials"
such as Ceran® cannot be separated by brittle laser cutting! Table 7.1 shows
the stress levels for the most important technical glasses.
Stresses as given in Table 7.1 are sometimes insufficient for the cutting
of certain glass sheets. For example, glass surfaces are often subject to com-
pressive stress, which must then be overcome by the laser-induced (tensile)
stresses. Therefore, laser cutting is only possible for sheets whose initial com-
pressive stress is sufficiently low because of their small thickness, or has been
reduced by annealing (see Table 7.2)

7.1.2 Laser-Beam Profiling

Crack propagation is induced by tensile stress in the glass or at the sur-


face. But compressive stresses are expected to occur around the heating zone
when glass is heated. Let us consider the radial stress a r and the tangential
stress at of a stationary glass sheet after heating with an axial-symmetrical
(Gaussian-shaped) laser beam. Figure 7.3a sketches the development of the
stress distribution. As can be seen, the radial stress a r around the heated spot

Table 7.1. Stress levels for brittle laser cutting

Glass type
Soda-lime glass (windows, container glass) 50 ... 100MPa
Borosilicate glasses (fire-protection windows, laboratory 25 ... 50MPa
glass articles, e.g. Pyrex®, Duran ®, Suprax®

Table 7.2. Toughening levels for brittle laser cutting

Application Glass type Typical toughening level Laser


cutting
possible
MPa
Thin glass sheets borosilicate glass <2 yes
(d < 1mm)
Glass windows, soda-lime glass 10-30 yes
car windshields (technically annealed)
Fire-protection borosilicate glass 40-100 no
windows (if toughened)
Safety glass soda-lime glass 80-150 no
(extremely toughened)
7.1 Rough Estimation of Process Parameters 383

~--------~~
~e~
n~s~i
o~n-------------'a) ~--~----~------~

at
o
::i
.!!! - 1
b Compression

-2

-3+------,------.-----,-----~
o 2 4 6 8
rlro

Fig. 7.3. (a) Stresses at a circularly heated spot in a glass sheet, (b) probable
crack propagation (top view on sheet)

is everywhere compressive (convention: negative stresses mean compression,


positive stresses mean tension). For the shear stress at, in contrast, we find
tensile stresses to be active at sufficient distance from the heating spot. As
Fig. 7.3b indicates, the crack will propagate radially towards the center of
the laser spot.
This means that an axial-symmetric spot already influences the direction
of the crack path. The same is true for moving spots. Finite-element analysis
and experimental observations show that this directivity is very distinct when
"elongated spots" oriented cross-directionally to the motional direction are
used.
Elongated heating spots naturally develop during use of higher transversal
laser modes. TEMOl modes are examined in this report. The corresponding
intensity profile of the laser radiation is, for example,

2PL (X) 2 (X2 +


h(X'Y):::::;7TW 2 ~ exp - w 2
y2) (7.2)

(so-called Gauss~Hermite mode). PL is the laser power, w is the mode radius


of the beam, and x and yare (Cartesian) coordinates relative to a plane
normal to the laser beam. Figure 7.4 shows the intensity distribution for the
TEMOl mode.
The mode radius is adjusted via the focusing optics of the laser. As can
be seen, the intensity profile has two maxima relative to the x direction.
According to (7.2), the distance of the maxima is not freely variable in this
case, but results from mode w. The "track precision" of the crack is about
w /2 for the TEMol mode, i.e., for precise cuts it is generally favorable to use
strongly focused laser beams. Using a TEMol mode is the obvious means for
improving the directivity. We will only discuss heating by this beam profile.
384 7. Post-Processing by Laser Cutting

1/10

·2

Fig. 7.4. Intensity pro-


Propagation direction file of a TEMol mode

The intensity distribution in (7.2) is standardized as follows:

JJ
+00

dxdyh (x , y) ~h . (7.3)
-00

Temporal modulation of the laser power is possible in principle. Here, only


the simplest case with PL = const. will be analyzed.
The motional direction of the laser b eam on the glass surface is also given
in Fig. 7.4. The beam is moved in the y direction. Assuming the laser beam
is moved along the surface of the glass sheet with constant velocity v, the
following time-dependent intensity distribution of the beam strength on the
surface can be derived from (7.2):

h (x , y,t) ~
2PL
7fW2
(x
( ;X)2 exp - 2 +(y-Vt)2)
w2 (7.4)

Possible generalizations of (7.2) for curvilinear beam paths are quite obvious.
Here, we only deal with results for laser beams moved with constant velocity
in the y direction, i.e., the y direction is always the cutting direction.

7.1.3 Selection of Laser

Emitters for laser cutting must have the following qualities:


• the spectral range of the laser must be favorable, i.e. , the radiation must
be absorbed by the glass;
• continuous (cw) or quasi-continuous operation of the laser must be possible
(e.g., pulsed operation with repetition rate in the kHz range);
7.2 Numerical Analysis of Cutting Processes 385

• beam configuration (or, e.g., a stable TEMol mode) must be possible;


• they must be suitable for industrial application (reliable and easy to main-
tain).
At present, only CO 2 lasers fulfil these conditions. The small penetration
depth of the radiation in silica glasses is a disadvantage. Because of the
strong coupling to the polar vibrations of the Si-O bonds in silica glasses, we
find low values for the absorption coefficient in the range of (3 ;:;::; 10 ... 150 x
103 m -1. The absorption coefficient varies in dependence on the exact spectral
range of the laser. (C0 2 lasers can radiate in the range of 9.3 ... 10.9 J..lm - at
10.6 J..lm wavelength, which is favored in the technique, the laser has highest
efficiency.) Our following numerical calculations are restricted to a value of
(3 = 25 X 103 m- 1 .

7.2 Numerical Analysis of Cutting Processes

Numerical analysis of cutting processes can be carried out with the finite-
element method (FEM) or with the boundary-element method (BEM). In
BEM programs, "perfect brittle-failure modules" and "oo-elements" can be
realized. In this way, the processes at the cutting line can be well monitored
and the CPU time can be considerably reduced because far fewer elements
are necessary. On the other hand, commercial BEM programs are still far
less sophisticated than FEM programs regarding their operability (in partic-
ular during pre-processing). This is particularly disadvantageous in the early
simulation stage of laser cutting of glass, where the search for favorable pro-
cess parameters necessitates many model variations. We therefore chose FEM
simulations, which we carried out with the explicit FEM program ANSYs.
The simulation calculations are divided into the following steps:
• heating of the sheet by laser radiation, with subsequent cooling if necessary;
• stress development in the sheet as a result of thermal heating;
• fracture-mechanical evaluation of the stresses (calculation of stress inten-
sity).
In principle, the first two steps can be condensed into one thermoelastic
calculation. This would simplify the simulation processes (pre- and post-
processing, data transfer between the models). But because the thermal and
the mechanical calculations require different meshings (mesh density, model
size) and time discretizations, problems will occur regarding the model size
and the CPU time. On the other hand, the sheet is only insignificantly de-
formed during cutting, i.e., the laser beam is not influenced by the deforma-
tion of the sheet. The thermal and the mechanical calculations can therefore
be carried out separately. We decided on this approach.
386 7. Post-Processing by Laser Cutting

7.2.1 Calculation of Temperature Distributions

The absorbed laser radiation heats the glass. Numerical calculations account
for this by introducing the heating density q (q = heat generation per unit
volume). The strong absorption of the laser radiation in the glass results in
a strong depth dependence of the heating density:

q (x, t) ~ f3 exp (-f3z) (1 - r) Isurf (x, y, t) . (7.5)

Here, z denotes the distance from the exposed surface (i.e., the "depth").
Isurf(x, y, t) is the place- and time-dependent beam intensity at the surface,
r is the reflection coefficient for the radiation through the glass surface, and
x and yare the coordinates for the surface of the sheet. (In this section we
only investigate in model calculations the radiation emitted by a CO 2 laser
normal to a smooth glass surface. For this purpose, we can assume r ~ 0,
see Table 7.3. Caution: CO 2 laser radiation is strongly absorbed by water.
For "moist glass surfaces", for example when the glass sheet is additionally
cooled by moist air, strong surface reflection may occur. Moreover, completely
different thermal boundary conditions must then be assumed.)
Figure 7.5 shows the depth dependence of the heat generation for a CO 2
laser and an IR burner (.\ ~ 4 !-Lm). As can be seen, the heat generation in-
duced by a CO 2 laser beam only reaches depths of a few tens of !-Lm. There-
fore, a complete heating of the glass volume can only be achieved by thermal
conduction.
If a laser beam in a TEMOl mode hits the glass surface at normal inci-
dence, the intensity distribution according to (7.4) can be used to calculate

Table 7.3. Material and process parameters for FE calculations (all material pa-
rameters are assumed to be independent of temperature)

Symbol Value Unit


Young's ratio E 70 GPa
Poisson's ratio v 0.22
CTE a 5 10- 6 K- I
Glass density p 2500 kgm- 3
Specific heat cp 1000 Jkg- I
Heat conductivity K, 1.0 W/(m 20 C)
Glass transition temperature Tg 550 °C
Wavelength A 10.6 f!m
Reflection coefficient r 0
Absorption coefficient f3 50000 m- I
Laser power PL 10 W
Mode radius w 0.5 mm
Scan speed v 36, 100 mms- I
Thickness of glass sheet d 0.4-1.1 mm
Size of glass sheet 500 x 500 mm
7.2 Numerical Analysis of Cutting Processes 387

50
Depth profile of
heat generation
40 C02-laser
j3 = 50 000 11m
1:E
-.....
N
30

~
x 20 IR burner
CD
'<l..
j3 = 10000 11m

10
o L ______________________ _

o 25 50 75 100 125
depth z/lJm

Fig. 7.5. Depth profile of heat-generation rate

Isurf(x, y, t) in (7.5). Under the initial condition T (x, t = 0) = 0, temperature


T(x, t) in the sheet is then calculated with the thermal conduction equation

(7.6)

where p is the density, cp is the specific heat, and r;, is the thermal conductiv-
ity of the glass. (The material data used in the model calculations are listed
in Table 7.3; for our present discussion we assume material data to be inde-
pendent of temperature.) Cooling of the sheet surface by convection or by a
cooling agent is neglected, i.e., we have no heat losses. (As the calculations
show, the heating of the glass can easily result in temperatures well above
1000 cC. Then strong thermal radiation by the glass would occur, i.e., heat
losses by re-emission and a strong contribution to the thermal conduction by
thermal radiation would have to be considered. When the process parameters
for "brittle cutting" of glass are optimally adjusted, the glass temperature
amounts to less than 600°C; the re-emission of the glass is then negligible.)
Now, the thermal conduction equation with a volume heating term has to be
solved. The actual calculation of the temperature distributions is done with
FEM in a transient calculation.
Heating is restricted to a thin surface layer. This fact must be accounted
for by a sufficiently fine meshing of the FEM model at the surface. We found
that the volume heating is quantitatively modeled correctly (see note below)
only if the thickness of the first layer of the FEM mesh at the surface is
< 1/(413). Because 13 = 50mm~1 (see Table 7.3), the first layer must then
have a thickness of 5 ~m. As Fig. 7.5 shows, the heat generation can already be
neglected for z > d1 = 4/13. With the adaptive mesh refining implemented in
ANSYs, convergent results are achieved if the entire glass thickness is meshed
388 7. Post-Processing by Laser Cutting

with at least 20 layers and the first layer has thickness 1/(4,6). To save CPU
time, we chose this method.
Whether the meshing is sufficiently fine can be tested by carrying out an
energy calculation. With the entire energy contributed by the laser remaining
in the sheet,

2: Wi = (1 -
2
r) it
0
dt' PL (t') (7.7)

must be valid. Wi denotes the thermal energy stored by the ith element of
the FEM model, and summation is carried out for all elements; planes of
symmetry must be considered. (In the purely thermal model, the fact that
part of the energy induced by heating is transformed into elastic energy or
into fracture energy (enlargement of the body surface) can be neglected.)
The mesh density is assumed sufficient if at least 95% of the radiation energy
input is registered via the heat stored by the model and if the temperature
distribution is "sufficiently smooth" .
Note: Alternatively to modeling the heat source by volume heating for ex-
tremely high absorption coefficients (extremely high ,6 d), a heat flow through
the surface can be predefined instead of the volume absorption. The condi-
tions in the top layers are not precisely represented, but this can be neglected
if the values ,6 d » 1 are sufficiently high. The system of equations to be
solved then reads

-"'/1zT (x, t) = jsurf (x, t) = - (1- r) Isurf (x, y, t) '}


(7.8)
(pcpgt - '" \7 2 ) T (x, t) = 0.

Because of the high temperature gradient at the surface, the mesh must also
be refined in this case. As a result, CPU time savings remain small.
Figure 7.6 shows the result of a temperature calculation (numerical solu-
tion of (7.6) with FEM). In Fig. 7.6a, the heat-generation rate is represented
as a color contour plot, while Fig. 7.6b shows the resultant temperature dis-
tribution. A sheet positioned symmetrically to the path of the laser beam

qfNmm- 3
v _0
i
iii
~

- 500

z
Fig. 7.6. (a) Heat generation rate and (b) temperature distribution (laser illumi-
nation from the top, view on top face; results for t = 30 ms, i.e., the laser spot has
moved 3mm from the edge); see movies on CD
7.2 Numerical Analysis of Cutting Processes 389

was taken as a basis. Table 7.3 gives the material and model data used. One
half of a symmetrical model is shown; the represented part has the dimen-
sions 0.7 x 10 x 4 mm 3 and consists of 60 000 elements! The model was cut to
make the in-depth and lateral temperature development easily visible. (The
"cut-out angle" is available in the FEM model; we simply left it out for
convenience. )
Glass sheets to be cut typically have lateral dimensions of about 0.5 m.
Modeling sheets with such large dimensions is, of course, unnecessary because
a relevant temperature increase occurs exclusively in a narrow strip along the
cutting path. For temperature calculation, only this strip must be modeled.
To describe the temperature development in a sheet with sufficient precision
we must therefore determine the necessary dimensions of the thermal model
(strip) and find a suitable discretization.
In a first step, we determine what width Bth (concerning the x direction)
the strip must have by making the following estimate: due to laser heating,
(approximately) homogeneous temperature conditions exist in the strip at a
large distance from the laser spot. The temperature increase there amounts
to

(7.9)

If we allow maximum temperature increase for this strip, we can derive its
minimum width:
11 IPL 1
B th ?,: - - - - - - - (7.10)
2 pcp d v 6.T

For 6.T ::::0 10 °C (only approximately 2% of the entire temperature increase


according to Fig. 7.6) follows B ?': 2.9 mm. In this report, results for 4 ... 6 mm
width are shown.
The strip need not to be uniformly meshed over its whole width. As
Fig. 7.4 shows, the region with high heat generation is Ixl < 2w (w being
the mode radius). For the model data according to Table 7.3 this region is
Ixl < 1 mm. There, the strip is uniformly meshed with 6.:r ::::0 w/4. Outside
this range, far coarser meshing will suffice because the adaptive mesh refining
implemented in ANSYS is employed there.
Estimating the length L (concerning the y direction in Fig. 7.6) over which
the model must extend is more complex. Our approach to a solution is via the
one-dimensional thermal conduction equation. As the typical time constant
for which the temperature is balanced over the sheet thickness (z direction)
we get
~ pCp 2
Tdepth ~ --d . (7.11)
4K:
With the data from Table 7.3 we get Tdepth ::::0 0.3 s. For t < Tdepth, the tem-
perature distribution is strongly structured over the sheet thickness, as in
390 7. Post-Processing by Laser Cutting

Fig. 7.6. For t > Tdepth, on the other hand, the temperature inhomogeneities
are quickly equilibrated over the sheet thickness. If we now assume (as con-
firmed by numerical calculations), that the temperature distribution over the
sheet thickness significantly influences the stress distribution and therefore
the cutting performance, we can estimate the minimum length of the strip
as follows:
pcp 2
L? V Tdepth ~ --d v . (7.12)
4K
Using the model data in Table 7.3 we get L > 3 cm. The strips our present
thermal calculations are based on have a length of at least 50 mm (Fig. 7.6
shows only a section).
Concerning the y direction, the strip is also meshed with a uniform ele-
ment size t::J.y ~ w/2 (here, t::J.y = 250 !-lm). The meshing in the y direction
can be coarser than that in the x direction because the distribution of light
of the laser beam is also coarser in the y direction than in the x direction
(see Fig. 7.4). In total, the y discretization comprises 200 layers.
On the CD enclosed in this book you will find several movies showing, for
example,
• the movement of the heating spot in the strip;
• the temperature development in the strip;
• the development of the x component of the mechanical stress in the strip
(see Sect. 7.2.2);

With a surface area of 4 x 50 mm 2 , the FEM model for thermal calculations
has considerably smaller dimensions than the glass sheets to be cut. This
poses the question of how to transfer the temperature distributions of the
small thermal model to the larger mechanical one. Assuming a straight cut,
a quasi-stationary temperature distribution in the sheet can be found via
transformation of coordinates.

T (x, y, z, t) = T (x, y - v . t, z) = T (x, 7], z) , (7.13)

where 7] = y - v t or 7] = y - Y (t) for a laser spot moved at inconstant rate


(distance: Y (t)). The temperature data of small thermal models are then
transferred to larger, mechanical ones with

T mechModel(X,7],Z) = Tt hermModel(X,7],Z) forlxl:::; B, and 7]:::; L


= 0 (or initial temperature) otherwise,
(7.14)

or with suitable interpolation formulae.


Surprisingly, a temperature of ~ 500°C is observed at the glass surface
in Fig. 7.6b already for the very low laser power of P L = lOW. This phe-
nomenon is due to the high absorption coefficient (3 and the small thermal
7.2 Numerical Analysis of Cutting Processes 391

conductivity of the glass: the laser energy is absorbed in a very thin surface
layer and penetrates into the glass volume relatively slowly.
In any case, the temperature is already close to the Tg of the glass. The
process parameters must therefore be adjusted very carefully in order not
to exceed T g . If thermal conduction in the glass is neglected in a rough ap-
proximation, the following scaling formula for the maximum "temperature
increase" can be derived with (7.5):

(3
6.Tmax ex
(7.15)
material process
paramo paramo

6.Tmax is the temperature difference between the middle of the laser spot (for
the TEMOl mode the temperature in the middle of an intensity maximum)
and the temperature of the surrounding cold glass. Equation (7.15) does not
yield quantitatively correct values (with the data from Table 7.3, 6.Tmax R::
4000°C follows; the discrepancy with the FEM results with 6.T R:: 500 DC
arises from the disregard of conduction), but within the framework of a linear
theory useful estimates are gained concerning the modification of process
parameters necessary in order to receive the same temperature increase. In
cutting a glass type, for example,

(7.16)

is valid; i.e., if laser power P L is increased by a factor, the mode radius or


the driving speed can be increased by about the same factor.

7.2.2 Calculation of Stress Distributions

Temperature calculations form the basis of stress calculations. For this pur-
pose, the temperature loads are imposed on the mechanical model with (7.14).
The sheet is not subject to any further mechanical loads; its own weight, in
particular, is neglected.
The mechanical model represents a sheet with the dimensions 500 x
500mm 2 (or 250 x 500mm2 for "symmetrical cutting"). Although meshing
the mechanical model is far less critical than meshing the thermal model, still
small elements are required in the surrounding of the cutting line in order to
detect the details of the stress distributions (see, e.g., Figs. 7.18 and 7.22).
Therefore, a strong mesh refining along the cutting line is required. Modeling
in a strip along the cutting line (about Ixl <:::: Bth' B th : width of the thermal
model) is done with "bricks" (e.g., ANSYS element type solid95). Outside of
the strip, shell elements (e.g., ANSYS element type shell93) can be used to
save CPU time.
392 7. Post-Processing by Laser Cutting

The sheets are fixed


• at y = 500 mm (rear edge) concerning the y and z movement, and
• at x = 250 mm (right edge) concerning the x and z movement.
All rigid-body translational and rotational degrees of freedom of the sheet
are fixed in this case. (With a plane of symmetry at x = 0, the system would
be mechanically overdetermined as a rigid body - but not as an elastic body.)
Via these constraints, the friction of the sheet at the support is basically
accounted for. The constraints are so far removed from the cutting line that
an effect on the stresses in the cutting zone is no longer observed. In the
model version "sucked sheet", all nodes on the bottom of the sheet are fixed
concerning the z movement.
Figure 7.7 shows the result of stress calculations for an undamaged sheet
(without initial crack in the plane of symmetry). The x component of stress
is represented because this component is relevant for a cutting path along the
symmetry line. The represented section of the sheet corresponds to that in
Fig. 7.6. In addition, the entire bottom of the sheet shown in Fig. 7.7 is fixed

a)
Ox fMPa i -120

v =
=
=
-
=
= b)

Symmetry 0
plane
= 25

N/x
c)
0 y fMPa -- -120
~Y x
=
§=
=
=
0

= 25

Fig. 7.7. Stress distribution caused by thermal loads (t = 30ms). (a) x component
of stress; note the compressive stress on the top face. (b) x component of stress, view
on the bottom face ; note the extended region with tensional stress; (c) y component
of stress
7.2 Numerical Analysis of Cutting Processes 393

concerning the z movement. This corresponds to the case of a sheet sucked


by vacuum to the support (sucked sheet).
The starting edge of the sheet in Fig. 7.7 is cut exactly vertical. No round-
ing (C or D edge) or faceting exists. This corresponds to the most frequent
case of application in laser cutting of glass sheets.
In Fig. 7.7, compressive stress occurs at the top face of the sheet, whereas
tensile stresses seemingly arise only at the bottom. As can be gathered from
Fig. 7.8, this is only true in the initial phase of the cutting process. Figure 7.8
shows the time development of the x components of the stresses at the top
face (solid curves) and at the bottom face (dashed curves) of a thin sheet
with thickness d = 0.4 mm and, in terms of laser cutting, a relatively thick
sheet with d = 1.1 mm (at y = 0 and z = 0 or z = d, i.e., directly at the sheet
edge). At first, tensile stresses are indeed present only at the bottom (0" > 0),
and compressive stresses at the top face. Later (for t > 0.25 s), tensile stresses
also occur at the top face, and as a rule they even exceed the tensile stress
at the bottom.
The tensile stress maximum occurs at the top face after Tdelay ~ 0.44 .. .48 s.
How does this time relate to the in-depth volume heating time of the sheet
according to (7.11)7 For the O.4-mm-thick sheet we determine Tdepth ~ 0.08s
and for the 1.1-mm-thick sheet Tdepth ~ 0.6s. Although the order of magni-
tude of Tdelay still roughly matches Tdepth, we do not find the dependence on
sheet thickness t (Tdelay ex d 2 ) expected from (7.11).
A laser rate of 36 mmls having been assumed in the model these calcu-
lations are based on, the laser beam is now about 16 ... 17 mm away from
the edge. The maximum tensile stresses at the surface of the sheet are only
30 ... 62 MPa in Fig. 7.8. In agreement with theoretical considerations, we
find the following dependence for 0" max:

60 Bottom face

40
t =0.4 mm
<1l
a..
~ 20
~
0
t = 1.1 mm
- 20
Sucked sheets

0 0.5 1 1.5 2
tis

Fig. 7.8. Time development of x stress at the starting edge (sucked sheets with
thicknesses of 0.4 mm and 1.1 mm; solid line: top face, dashed line: bottom face)
394 7. Post-Processing by Laser Cutting

1
a max ex d (7.17)

As will be discussed in detail below, tensile stresses of 30 ... 62 MPa are usu-
ally insufficient for inducing instantaneous fracture in a glass sheet being not
heavily predamaged.
Suction of the sheet to the support strongly influences the stress distribu-
tion. Figure 7.9 compares the stresses in a sucked sheet with those effective
in a free sheet. The mechanical stresses, in particular the tensile stresses at
the bottom, are decidedly smaller in the free sheet. These additional limiting
conditions (sucked +-+ free sheet) must be correctly considered in modeling
the laser cutting of thin glass sheets!
Figure 7.10 shows the time development of the x stress component along
the edge of the free sheet (as for Fig. 7.8). A comparison with Fig. 7.8 yields

a) b)
:: -120
(Jx IMPa
55

o
25

y~x ~x
y

Fig. 7.9. Comparison of stress distributions (a) in a sucked and (b) in a free sheet
(t = 30 ms); see movie on CD

60

40
t = 0.4 mm
(1!
a..
--~ 20
~

- 20
Free sheets
0 0.5 1.5 2

Fig. 7.10. Time development of x stress at the starting edge (free sheets with
0.4 mm and 1.1 mm thickness; solid line: top face, dashed line: bottom face)
7.2 Numerical Analysis of Cutting Processes 395

no obvious qualitative differences; only the stress levels are a little reduced
in Fig. 7.10.
Figures 7.8 and 7.10 show that tensile stresses in the range ~ 30 ... 60 MPa
occur at the upper face of the starting edge. To decide whether these stresses
are high enough to initiate a cut at the edge we must know the edge strength
of the glass sheet. Figure 7.11 shows measurement results gained by 4-point
bending of approximately I-mm-thick glass sheets. We calculated the cumu-
lative fracture probability versus the tensile stress. (The resistance of glass
to instantaneous fracture can only be calculated via fracture probabilities.)
The dots represent measured values and the solid straight lines approxima-
tions by Wei bull distributions. Two series of measurements for samples with
differently processed edges are shown. For laser cutting of glass edges, the
"polished-edge" samples are representative.
The fracture probability of a glass edge, i.e., the probability of a cut
starting at a glass edge loaded over length L with tensile stress rY is

F (rY, L 1oad ) = L ref (rY)


1 - exp [ - Lload rYe
m]
' (7.18)

where L ref = 75 mm, rYe = 160 MPa, and m = 7 are the data of the strength
measurement. If the starting point of the cut can be determined with an
accuracy of better than 1 mm (L 1oad = 1 mm), we find a probability of ~ 10- 5
for rY ~ 60 MPa.
A cut starting at an undamaged edge is therefore improbable. The stresses
can only be expected to cause propagation of an already present crack, pre-
viously produced at the glass edge with a suitable tool. This effect will be
analyzed in the following model calculations. We begin with investigating
which stresses result from a crack in the sheet. In Fig. 7.12 we see the results

0.99
Edge strength
0.9 (Lref = 75 mm)
Wei bull distribution

0.5

I
F I
0.1 • I
Polished I
edge I
0.05
I
I Polished
and toughened
0.01
50 100 200 300 400 500 600
(J I MPa

Fig. 7.11. Fracture probability of glass sheet versus edge stress


396 7. Post-Processing by Laser Cutting

a) b)

T I"C

--
t1x IMPa
~ = -50

V -
Crack
tip
5 0

- -
<=>

= 50
Fig. 7.12. (a) Temperature and (b) stress distribution in sucked sheet with a 5-
mm-Iong initial crack at the symmetry plane, top-side view (t = 130ms, i.e., the
laser spot has moved 13 mm)

of an FEM calculation. Figure 7.12a shows the temperature, gained with


(7.14) from the preceding thermal calculation, and Fig. 7.12b gives the stress
component O"x. As can be seen, the stress level is still low; at the crack tip
the tensile stresses amount to 5 ... 10 MPa only. Nevertheless, a crack open-
ing is already distinctly visible. In the context of fracture mechanics (see
Sect. 7.4.1) this means that the crack is subject to a mode-I load! The crack
growth is induced by this crack load.

7.2.3 Condition for Cut Elongation

Thus we arrive at the central question, how to assess stresses in glass and how
to include the assessment in a finite-element algorithm? We start by calculat-
ing and assessing the so-called stress intensity factors K r , where i = I, II, III
refers to the fracture modes. (For a concise fundamental introduction to the
fracture mechanics of brittle solids please refer to Sect. 7.4.) To keep discus-
sion within reasonable limits, we will only describe the symmetric cutting of
glass sheets.
The laser beam and the fixing of the sheet are symmetrical to a sym-
metry line that coincides with a straight cutting path. Thus, a pure mode-I
load occurs (Kn = KIll = 0, see Sect. 7.4.1). The scheme for the numerical
calculation of the stress intensity factors is given in Sect. 7.2.4.
We consider the stress intensity factor Kr for crack mode I. A crack will
increase if Kr exceeds the so-called fracture toughness K rc of the glass (see
(7.23) and Table 7.4):
7.2 Numerical Analysis of Cutting Processes 397

Kr > K rc . (7.19)
The proper scheme for the analysis of (7.19) is to calculate the stress inten-
sity factor Kr = Kr(a) in dependence on the cut length a. Useful process
parameters for laser cutting are reached, if for all t > 0 a cut length a with
o < a < Llas(t) = VLt can be found that matches (7.19) (Lias = laser path
length). However, applying this condition in FEA is very difficult.
Cut length a cannot be treated as a continuity in FEA. The adequate FEA
modeling is based on an incremental increase of the cut length by equidistant
steps (e.g., l:1y = 0.5mm). For (7.19) to "hold continuously", a sufficiently
large range (amin(t), amax(t)) for any Llas(t) must be found where (7.19) is
valid.
Consider a laser path length Lias = 50 mm. A cut length increment of
l:1y = 0.5mm requires N = 100 Kr calculations (with coarse modeling and
submodeling, see Sect. 7.2.4). Because this must be done along the whole
laser path (0 < Lias < 50mm), the total amount of calculations required is
:::::: 1/2 N 2 = 5000 K r . Running these calculations, even with completely au-
tomatized FEA codes, would take up too much computing time. We therefore
restrict ourselves (1) to a few representative laser path lengths and (2), for a
fixed path length, to a few representative cut lengths.
In reality, the crack tip has a (non-stationary) sickle-shaped contour,
which presently cannot be determined by FEA. We therefore use a straight
vertical tip contour such as sketched in Fig. 7.13b. However, this simplifica-
tion presupposes that the stress intensity factor Kr does not vary strongly
along the crack tip.
Figures 7.14 and 7.15 show the time dependence of the stress intensity
factor Kr for thin (d = O.4mm) and thick (d = 1.1mm) sheets at the top
faces (solid lines) and bottom faces (dotted lines) of sucked and free sheets,
respectively. The process parameters are equal to those for Figs. 7.8 and
7.10 (e.g., cutting speed 36mm/s). The dotted horizontal lines indicate the
fracture toughness K rc of glass (see Table 7.4). The cut will elongate in the

Cutting}
direction contour

Fig. 7.13. Crack tip contours: (a) curved crack tip, (b) straight crack tip
398 7. Post-Processing by Laser Cutting

1.4 Sucked sheets


a= 0.5 mm
1.2 ;'
I

,
1.0
~
~E
I
III 0.8
a.. -------------r-./-
:2
'::.. 0.6

L'" ....
---
~

l
0.4
t = 1.1 mm
0.2
I
0.0
0 0.2 0.4 0.6 0.8 1.0 1.2
tis
Fig. 7.14. Time development of Kr for a O.5-mm-long crack in sucked sheets (thick-
nesses: O.4mm and 1.1mm; solid lines: top face, dashed lines: bottom face)

1.4 Free sheets


a=0.5 mm Top face
1.2

~
1.0
E
III 0.8
a..
:2
'::.. 0.6
~

0.4

0.2

0.0
0 0.2 0.4 0.6 0.8 1.0 1.2
tis
Fig. 7.15. Time development of Kr for a O.5-mm-long crack in free sheets (thick-
nesses: O.4mm and 1.1 mm; solid lines: top face, dashed lines: bottom face)

regions where the curves exceed Krc. Judging from Figs. 7.14 and 7.15, no
problems are to be expected for thin sheets, whether free or sucked, because
the Kr values clearly exceed the fracture toughness.
Figure 7.15 (free sheets) shows only small differences for the stress inten-
sities at the top and bottom face. Both curves reach their maximum at the
same time, and the maximum values are approximately the same. We find
only negligible Kr variations along the crack contour. This vindicates the
assumption of a full-body cut with an approximately straight vertical crack
contour.
7.2 Numerical Analysis of Cutting Processes 399

L Cut face

I (microcrack)

Crack tip Fig. 7.16. Crack tip contour for a micro-


contour crack

For the thin sucked sheet in Fig. 7.14, in contrast, the Kr maximum
occurs definitely earlier at the bottom than at the top face. Again we expect
a full-body cut, because Kr exceeds KIc at both sides, but the crack contour
will presumably be sickle-shaped (see Fig. 7.13a). For a full-body cut of thin
sheets, the free-sheet arrangement is therefore preferable!
Figures 7.14 and 7.15 reveal that for thick sheets, sucked or free, the
stress intensities never clearly exceed the fracture toughness Krc, meaning
that cutting thick sheets is definitely more critical.
The Kr value of the thick free sheet (Fig. 7.15) exceeds K rc only at the top
face. (A sheet with thickness d = 1.1 mm has already high bending strength
and cannot be brought into sucked condition by (vacuum) suction alone. The
sheet must be mounted by mechanical devices (pressing-on of the sheet).) In
this case, laser cutting is expected to result only in a micro crack at the top
face. For real separation, the sheet must be broken after cutting.
The Kr value of the thick sucked sheet (Fig. 7.14) exceeds KIc only at the
bottom face. The differences between the top and bottom face are significant
here. With this arrangement, scribing the bottom is advisable.
A comparison between Fig. 7.14 and Fig. 7.17 indicates how conditions
change when the cracks in the sucked sheets reach length a = 2.5 mm. The
thin sheet shows no qualitative changes, only the stress intensity factor is
slightly smaller than in Fig. 7.14. Extended calculations show the Kr values
for longer cracks to be stationary, i.e., they do not decrease with the further
development of the cutting path.
The Kr value of the thick sheet in Fig. 7.17 does not exceed K rc at all,
meaning that the crack gets stuck in the range a = 0.5 ... 2.5 mm. We see
that the process parameters are unfavorable for cutting 1.1-mm-thick glass
sheets; the cutting speed of 36 mmls is too high and the laser power too low.

7.2.4 Calculation of Stress Intensities for Laser Cutting

Because very small crack tips (partly « 1 mm) must be modeled in large
sheets (500 mm), either extremely fine meshing or the so-called submodeling
technique must be applied. We decided on the last method, i.e., we first de-
400 7. Post-Processing by Laser Cutting

1.4 Sucked sheets


a=2.5 mm
Top face
1.2

~ 1.0
~E

8!. 0.8 --
;z- 0.6
:2:
-,
I
0.4 /
/
0.2 / t = 1.1 mm
I
0.0 +-L.JI!'"--,-.....L--,--L-_--,_ _, -_ _,-_---.---'
o 0.2 0.4 0.6 0.8 1.0 1.2
tis
Fig. 7.17. Time development of Kr for a 0.5-mm-Iong crack in sucked sheets (thick-
nesses: 0.4 mm and 1.1 mm; solid line: top face, dashed line: bottom face)

vel oped a coarse model that already correctly considered the crack as well as
the heating and the boundary conditions (fixing or clamping) of the sheet.
In this way, the "coarse distribution" of the mechanical stress and the defor-
mations in the sheet is computed. Subsequently, another model is built with
a considerably finer mesh appropriate for the calculation of stress intensity
factors (see Sect. 7.4.2). Then the thermal loads (in volumes), the stresses on
the edge of the submodel, and the deformations of the edge of the submodel
are transferred from the coarse model to the submodel. The deformations
and stresses in the immediate surrounding of the crack tip are then com-
puted with the sub model. With a suitable meshing of the elements at the
crack tip, the stress intensity factors at the crack tip can be computed. Via
these factors we can then predict whether the crack will run on, and if so, in
which direction (the directional aspect will be ignored here because we only
consider symmetrical cutting).
An appropriate meshing for the coarse model was found by trial and
error. Figure 7.18 shows the meshing of the model in the sheet plane. In
the strip Ixl :::; Bth' the temperatures from the preceding thermal calculation
are transferred. In the present model, Bth is 4.5 mm. A regular meshing is
employed in this strip.
To keep the CPU time within reasonable limits, the y discretization is
not carried out over the entire sheet length (500mm). Stationary conditions
are reached very fast. The finely meshed region can usually be limited to
y :::; Ll ~ 50 mm. Three to four element layers are used for the meshing in
the depth direction, the ratio between the topmost and the lowest layer being
0.1 (for sheets with thickness d < 0.7mm), respectively 0.05 (for t > 0.7mm).
Beyond this strip, "faster" shell elements are employed because they must
only represent the "plate stiffness" of the sheet and its clamping or fixing.
7.2 Numerical Analysis of Cutting Processes 401

Crac
tip

B
Solid elements Shell elements
>-----.x Fig. 7.18. Meshing of a coarse model

In Fig. 7.18 a crack tip is indicated at y = 0.5 mm. In the subsequent


submodeling, a special crack-tip meshing is carried out in the surrounding
area.
Figure 7.19 shows the results gained by coarse modeling a O.4-mm-thick
sheet with a crack of length a = 0.5 mm in the plane of symmetry after 0.5 s.
The chosen feed rate of the laser was 36 mmls (meaning the laser spot in
Fig. 7.19 is at y = 18mm). Figure 7.19a shows the stress distribution in a
free sheet, Fig. 7.19b that in a sucked sheet. In Fig. 7.19a, the support is
indicated by a dashed line to make the upward curvature of the free sheet
more easily visible.
For simplicity, the front of the 0.5-mm-Iong crack is vertical (i.e., in the
case of y = const., 0 :::::: z :::::: t). For modeling a crack, the nodes in the
generated surface with 0 :::::: y :::::: a are free to shift in the x direction. The
maximum stress occurs at the crack tip. A comparison between the free and
the sucked sheet shows that the type of mounting influences the stress values
only slightly. Yet the fracture-mechanical evaluations (calculation of stress
intensity factors) indicate that the loads differ vastly.

b)

,
,
,,
,
,,
,,
,
y z ', Symmetry
~X '\" plane

Crack tip
a = O.5 mm

Fig. 7.19. Stress distribution in a free (a) and in a sucked (b) sheet with a 0.5-
mm-long crack at the symmetry plane (t = 500 ms, displacement magnified x 200) ;
see movie on CD
402 7. Post-Processing by Laser Cutting

In Fig. 7.20, finally, a O.4-mm-thick and a 0.7-mm-thick sucked sheet,


both with a 2.5-mm-long crack in the symmetry plane, are compared (other
model data as above). The stresses at the crack tips are smaller than those
in Fig. 7.19 despite the crack being five times longer.
The crack elongation to a length of 2.5 mm can be realized in FEM models,
for example, by using doubled nodes instead of single nodes in the generated
surface. One node is always placed at the symmetry plane and is fixed con-
cerning the x movement. The associated nodes are placed in the generated
surface of the sheet. In the unloaded state the nodes coincide. To model the
cutting in the ANSYS solving mode, the associated nodes can be successively
decoupled: to carry out the calculation for Fig. 7.19, we simply decoupled the
second y-node row, for Fig. 7.20 the sixth node row (f1y = 0.5 mm; the first
y-node row corresponds to y = 0).
This method is not unproblematic: a special problem in modeling sharp
cracks (i.e., cracks where the crack edges touch in the unloaded state) is that
the crack edges of the FEM model may penetrate under load. In reality, this is
of course impossible because the crack edges then will collide and penetration
is prevented by the corresponding compressive stresses.
This penetration of elements also occurs in the FEM modeling of laser
cutting if the modeling is not done carefully enough. Figure 7.21a shows how
the elements penetrate (penetration of the generated surface: 0.4 !-lm; the
deformation is magnified by a factor of 1000 in Fig. 7.21). The FEM meshing
of the coarse model is indicated by thin black lines.
Penetration of the elements can be prevented in the FEM models by ap-
plying contact elements in the symmetry plane (e.g., ANSYS type contac52).
Figure 7.21b shows the results and also indicates that the use of contact el-
ements essentially changes the stress distribution in the crack tip: although
the penetration of elements is insignificant, this effect must be considered!

a) Oxx IMPa b)
:: -60

40

Crack tip
a = 2.5 mm

Fig. 7.20. Stress distribution in a sucked sheet with a 2.5-mm-Iong crack at the
symmetry plane for thicknesses (a) d = O.4mm and (b) d = 0.7mm (t = 500ms,
displacement magnified x200); see movie on CD
7.2 Numerical Analysis of Cutting Processes 403

oxx /MPa a)
=-60
oxxJMPa b)
=60.

Fig. 7.21. Top-side view of a 0.7-mm-thick sucked sheet with a 2.5-mm-long crack
at the symmetry plane. (a) Penetration, (b) penetration prohibited by applying
contact elements (t = 200 ms; laser spot has moved 3.6 mm; displacement magnified
x 1000; pink triangles: contact elements); see movies on CD

The use of contact elements makes the model nonlinear, and this signifi-
cantly increases the CPU time. In addition, the insertion of contact elements
constitutes an essential modification of the model that cannot be carried
out in the solving mode but must be done in pre-processing. The modes of
operation of the FEM program must be switched accordingly.
Numeric investigations on laser cutting show that penetration of the ele-
ments actually occurs only for a few moments. One could therefore attempt
calculation without taking into account the contact elements by simply reject-
ing the time steps in which element penetration occurs. However, contacting
may already have occurred in the history of a cutting process before the
crack reached the contact-free state. It is then unclear whether the state of
contact-free crack opening will be reached at all.
The submodel employs a crack-adapted meshing at the crack tip for the
calculation of stress intensity factors (see Sect. 7.4.2). The data computed
above with the coarse model were transferred to the submodel for this pur-
pose. Figure 7.22 shows examples of submodels. The meshing and the stress
distribution in the respective submodel are represented. For better orienta-
tion the position of the submodel in the unloaded sheet is also given as a

r
Oxx IMPa a)

50

:: 150
b)
Fig. 7.22. Examples of submodels, top

r
0xx IMPa
view, t = 500ms, displacement magni-
fied xlOOO. (a) O.4-mm-thick sheet with
50
a 0.5-mm-long crack (left: free sheet,
right: sucked sheet), (b): 0.7-mm-thick
=150 sucked sheet with a 2.5-mm-long crack;
see movies on CD
404 7. Post-Processing by Laser Cutting

white outline. The submodel consists of five layers. Such meshings are used
to calculate the stress intensity factors in Sect. 7.2.4.

7.3 Practical Realization

Realizing this technology in production machines is a highly demanding task


for mechanical and control engineering. An assembly also suitable for clean-
room conditions has been developed (see Fig. 7.23).
The torsionally stiff mechanics consists of high-quality materials such as
stainless steel or anodized aluminum in order to allow particle-free operation.
The flow-optimized concept guarantees a particle-free environment in the
machine. Special attention has been given to the process-gas feeding, for
example the use of pneumatically switched diaphragm valves and ultrahigh-
purity gas filters.
All axes are carried out as linear motors or torque motors and offer the
following advantages:

Fig. 7.23. Laser cutting system DLC 600


7.3 Practical Realization 405

• high accuracy,
• low friction and therefore few particles,
• reduced use of critical materials in clean rooms (e.g. grease),
• clearance-free operation.
The standard equipment also includes a camera system, which enables highly
accurate cutting of coated glass sheets. Such sheets are first aligned and
subsequently measured in the x-y direction by means of reference crosses
with a CCD camera. Detected deviations can be conveniently corrected as
translational displacement and angular deviation by the Siemens Sinumerik
840D control via zero shift.
The DLC 600 machine is made highly productive by application of sev-
eral cutting heads. Several cuts can be simultaneously carried out with a
maximum of three laser heads. Preparatory to cutting, these are positioned
independently and then cutting along the x axis is carried out. The sheets
are turned with a directly driven rotary table, time-parallel during cut prepa-
ration.
The standard dimensions of DLC 600 are 1350 x 1950 x 2000 mm 3 . Pro-
cessing glass raw formats of up to 1500 x 1500 mm 2 is possible with this
generation of cutting systems. Custom-built special solutions are available
for larger glass formats.
Typical examples of laser applications are illustrated in Fig. 7.24. Coming
from various branches, they demonstrate the vast application range: optical
filters, hard discs, solar cells, fiat-panel displays, and mirrors. As different as
these applications are, they have one thing in common: production with con-
ventional cutting techniques requires extensive post-processing and/or pro-
duces an immense amount of waste.

Fig. 7.24. Typical examples of laser cutting


406 7. Post-Processing by Laser Cutting

Another example of the development potential of laser cutting is the in-


crease in cutting velocity achievable with the microscribe method. Figure 7.25
shows the results gained with a testing machine by Schott Advanced Process-
ing (Mainz) for commercial soda-lime glass without coating. On the x axis
the glass thickness is indicated in mm, on the y axis the scribing velocity in
m/ min. The three graphs illustrate the maximum velocity achieved in the
years 1998- 2000. Whereas the velocities reached in 1998 were very low, they
are now in the range of conventional scribing methods using diamond or
carbide cutting wheels.
The most important progress was made for thin glass (glass thickness
< 1.1 mm). Improved comprehension of the process has led to a significantly
increased energy input. By means of several tests and FEM simulations, a
beam profile was found that allows high cutting velocities as well as flexible
adjustment to different glass types and thicknesses.
Figure 7.26 shows the top view on a glass plane that is to be separated
along the blue line. A laser beam profile symmetrical along the cutting line
heats the glass, which is moved in the direction of the arrow. The U-profile is
open in the feed direction. The beam can be shaped by means of a scanner,
for instance.
Figure 7.26 shows three sections each through the intensity and tempera-
ture fields. It is clearly visible that the graph marked by section C- C' has two
peaks, with a certain distance between them, which are considerably higher
than the two peaks of section B-B'. In the vertex, marked by section A-A',
the intensity peak is again much lower. Because preheating has taken place
near the separation line through the two peaks, the temperature profile in
the range of section B-B' only shows a low local maximum. This maximum
will be compensated by a relatively low laser intensity in the vertex so that

35

30
'Ic
·E 25 Soda-lime glass
.€ 20
u
Q)
Q)
Co 2000
t/)
15
Ol
c
:0
.§ 10
en
5

0
1.1 2 3 4 6 8 10
Thickness Imm

Fig. 7.25. Increase in cutting speed


7.3 Practical Realization 407

z-
'Cij
cQ)
E

• y

Tg ...,. ........ ..
A- A'
~
:J
~
Q)
a.
;-- B-B'
E
~ C-C'

Fig. 7.26. Patented Schott laser-cutting technology

there will be a temperature gradient in the vertex with a temperature peak


in the range of the separation line.
The mechanical features of laser-cut edges are far superior to those of con-
ventionally cut edges. How significant the difference is becomes obvious when
the cutting edges produced with the different cutting methods are compared
under the microscope (see Fig. 7.27).

Scribed D 263 a) b)
Scribed D 263

0.4 mm
~eed
~eed
O.4mm

Laser spot

Fig. 7.27. (a) Conventional scribe-and-break edge versus (b) laser-cut edge
408 7. Post-Processing by Laser Cutting

7.4 Appendix: Fundamentals of Fracture Mechanics

7.4.1 Fracture Mechanics for Brittle Solids

The limit for the breaking strength of glass is O"limit > 1 CPa. Such high
values are only reached for glass fibers; "normal glass items" already break
at 0" < 100 MPa. The strength is reduced by flaws in the glass volume and
in the glass surface. Flaws enhance the stress in their neighborhood to such
a degree that even small loads may cause breakage. The simple rule that "a
body will break if the stresses exceed a critical value" does not hold for glass.
Fracture mechanics for glass must consider possible (initial) damages - which
are usually neglected in stress calculations - as well as initial stresses. These
so-called nominal stresses can have several causes. Internal stresses may be
induced, for example, by toughening, external loads by bending, impact, or
inhomogeneous temperature profiles.
Moreover, the geometry of the workpiece, the damage site, the stress sta-
tus (direction, homogeneity, etc.) and the orientation of the crack in the stress
field must be considered. In this case, we speak of crack modes. Figure 7.28
shows the fundamental crack modes for edge cracks:
• mode I - tensile forces perpendicular to the crack plane, crack propagation
perpendicular to the forces, with crack opening;
• mode II - forces parallel to the crack plane, crack propagation parallel to
the forces, crack edges slide without opening;
• mode III - forces parallel to the crack plane, crack propagation perpendic-
ular to the forces, crack edges slide without opening.
An explanatory model for glass fracture must consider the stress increase
in the surrounding area of cracks at least qualitatively, for example with the
following rule of thumb: "The crack increases the stresses in the body until
they exceed the limiting fracture strength - then the body breaks". This model
serves for qualitative discussions - but it fails in quantitative calculations!
We find an explanation for this failure when we calculate the stress increase
caused by the crack. For a crack stressed in the ith mode (i = I, II, III) one
discovers in ideal-brittle materials (materials showing brittle behavior until
fracture, i.e., showing only elastic deformation that completely disappears
when the force is no longer active, but no viscous or plastic deformation, nor

v[) II ,.... ". III


Fig. 7.28. Fundamen-
tal crack modes:
1. opening,
2. longitudinal shearing,
3. transversal shearing
7.4 Appendix: Fundamentals of Fracture Mechanics 409

any structural changes) a characteristic dependence on the distance r from


the crack tip for local stresses in the region around the crack:
Ki
local stresses: O"loc(r) ex Vi i = I, II, III (7.20)

local deformations: 7Lloc(r) ex KiVi . (7.21 )


For r --+ 0 we find O"loc --+ 00, which meanS that the local stresses diverge at
the crack tip! The stress intensity factors Ki are key quantities of fracture
mechanics. Note that they are proportional to the nominal stress 0"0. For
brittle materials, a fracture criterion based On the stress intensity factors can
be employed:
Ki 2:: K ic with i = I, II, III . (7.22)
If one of these conditions is fulfilled, fracture occurs according to the respec-
tive mode. Usually only the most important mode I
(7.23)
is considered. K rc is the so-called fracture toughness (see Table 7.4). The
fracture loads and critical crack sizes can be calculated with (7.23). For edge
cracks, for example, Kr ::::0 20"0Va (0"0: tensile stress; a: crack depth) is valid.
Thus, a body with a crack of depth a > (Krcl20"0)2 in the surface will fracture.
For 0"0 = 50 MPa and Duran® with K rc ::::0 0.85 MPaVrn we find a c ::::0 70 11m,
which means that even cracks with small depth are relevant for the fracture
toughness of glass!
Note that for crack mode I (i.e., Kr > 0) only tensile forces (i.e., 0"0 > 0)
open the crack! Compressive forces, in contrast, would close the crack without
its growing. For crack modes II and III, crack opening will occur for both
0"0 > 0 and 0"0 < o.
Note also that (7.22) and (7.23) are only valid for sharp crack tips, i.e., for
defects having a radius of curvature p --+ 0 at the tip. At glass defects with
p > 0, which are called notches, the stresses at the vertex remain finite (see
Fig. 7.29). The ratio of the local stresses O"loc occurring at the notches to nom-
inal stresses 0"0 active in the glass sheet is described by stress concentration
factors, for example
O"loc = Kt , (7.24)
0"0
where K t is the finite stress cOncentration factor.

Table 7.4. Fracture toughness of some glasses

Glass KJc (MPay'ffi)


BK7 1.08
FK50 0.51
Duran® 0.85
410 7. Post-Processing by Laser Cutting

b)

Fig. 7.29. (a) Sharp crack tip and


(b) blunted crack tip

7.4.2 FEA Calculation of Stress-Intensity Factors

Figure 7.12 shows a crack opening due to laser-induced heating, which cor-
responds to the deformation of the crack for mode 1. According to (7.20),
we would therefore expect diverging stresses at the crack tip. Against ex-
pectation, Fig. 7.12 shows a smooth stress distribution in the region around
the crack tip. The reason is that the meshing of the FEM model is not fine
enough for calculating the stresses at the crack tip (there is no sufficiently
fine meshing at crack tips!). Trying to judge the risk of fracture originating
from a crack via stress calculations is not at all sensible: we must expect
stress rates at the crack tip that depend on the meshing (with infinitely fine
meshing, the stresses would even diverge, see (7.20)), but do not reflect the
actual fracture stress!
For laser cutting, the appropriate consideration of fracture stress is based
on the evaluation of the stress-intensity factors. With FEM, these factors can
be calculated with (7.21) or

(7.25)

where Wi are linear-independent, stress-induced fundamental deformations


in the crack tip region. They are known from analytical investigations. For
example, the pure mode-I load of a straight, sharp crack (with the crack edges
touching under unloaded condition) is

Ux}
~~
K
= 2~V
rr{ cos(cp/2) (K, - cos(cp))
i; ~i~(CP/2)(K, - cos(cp)) (7.26)

G is the shear modulus of elasticity, K, is a combination of material parameters


that also depends on the specific stress condition. The calculation scheme
for the stress intensity factors is designed to calculate the deformations in
the crack tip region and split them according to Wi. (Note: Deformations
in the crack-tip region show a vrdependence. To achieve a highly accurate
representation in FEM models, elements with central nodes should be used
at the crack tip and the edge of the first element should be divided as 1:2,
with the short section at the crack tip.)
Figure 7.30 gives an example. The stress intensity factor KI is calculated
for a small quarter-penny-shaped crack (crack depth a = 0.1 mm) at the
corner of a glass edge. The glass edge is under uniform tensile stress of 40 MPa.
7.4 Appendix: Fundamentals of Fracture Mechanics 411

a) b)

--
min

--
--=
max

Fig. 7.30. FEA of a quarter-penny-shaped crack at a glass edge. (a) FEA mesh,
(b) zoomed detail displaying the first-principle stress

From FEM calculation


• a value of K J ~ 0.60 MPavm can be expected at the surface (the value
quoted in the literature is 0.57 MPavm) , and
• a value of K J ~ 0.55 MPavm for the crack middle, i.e., below 45° to the
surface (value in the literature: 0.57MPavm).
As can be clearly seen in Fig. 7.30b, the crack opens over its entire length.
However, this is not the rule. Under more complicated loads, the crack edges
often touch in some places (also with dominant mode-I loads!) and exert
compressive stresses on each other in these contact regions. Therefore, special
caution is necessary in FEM modeling. It is best to mesh the crack edges with
contact elements in such a way that penetration of the volume elements is
impossible. At least, the computations must be carefully checked for possible
penetration of elements.
Figure 7.30 also shows that the stress intensity factors vary along the crack
tip, so that considerable computing power and analysis efforts are required
for calculation. In practice, we must make simplifications.

7.4.3 Prediction of the Crack Path

We know that under the condition given in (7.22) and (7.23), immediate
crack elongation results. We do not know, however, which direction it will
take. Because laser cutting implies precise control of the cutting path via the
laser beam, we need criteria to predict the direction of the crack elongation.
The "principle of virtual crack" , introduced by Rice and Cot erell, supplies
such a criterion. An existing crack is "virtually" elongated (see Fig. 7.31).
In this way, elastic energy is released because the stress distribution in the
sheet is modified by the crack elongation. The probable crack direction is
then the direction for which the most energy is released by the virtual crack
elongation.
412 7. Post-Processing by Laser Cutting

Crack

Virt. crack

Fig. 7.31. Virtual elongation of a


a
, crack by an angle a with respect to
the primary crack

For ideally brittle-elastic solids, the principle of Rice and Coterell is equiv-
alent to assigning an effective stress intensity factor that depends on the
elongation direction to the resultant crack.

(7.27)

where Kr(O:) and Kn(o:) denote the stress intensity factors for the pure crack
modes I and II. There is no contribution from the crack mode III; crack elon-
gation only occurs in the "1/11 plane". The crack then runs in the direction
for which Keff( 0:) is maximal.
Application of this principle with FEM is basically possible. It would
mean, however, that the "virtual" crack must be elongated in many directions
- a disadvantage that makes modeling quite tedious. According to Nuismer,
the Kr(o:) and Kn(o:) values for ideally brittle-elastic solids may also be
derived from the stress intensity factors for the primary crack:

Kd 0:) = ~ cos ( ~) (( 1 + cos (0:)) Kr - 3 sin (0:) Kn) ,


Kn(o:) = ~cos(~) (sin (0:) Kr - (1-3cos (0:)) Kn) . } (7.28)

Based on (7.27) and (7.28), the direction of the crack elongation can be mod-
eled rather easily. Figure 7.32 shows a scheme of the prediction principle. The
stress intensity factors are represented as coordinates in a coordinate system.
For simplicity, the stress intensities Kr and Kn have been standardized to
Keff = 1 (vectors (Kr,Kn) form a semicircle). The arrows indicate the prob-
able directions of the crack elongation. Roughly speaking, the crack will run
in that direction for which the mode-II stress intensity factor decreases. The
crack favors elongation in the pure mode 1.

Fig. 7.32. Prediction of the direction ofthe crack


elongation based on the stress intensities KJ and
KII < 0 KII > 0 Kn
8. Glass Products Under Mechanical and
Thermal Loads

8.1 Strength Optimization of Airbag Igniters

Kurt Nattermann, Harald Krummel, Ludwig Frank


Introduction

Airbag igniters are small assemblies consisting of a sealing component with


specified electrical properties and an explosive. Robust igniters must guaran-
tee the reliability of the electrical properties and of the explosive for many
years. Thus, the components must not exhibit any form of corrosion and
the assembly must be air-tight. Otherwise, the explosive could degrade by
moisture, for example.
The necessary sealing properties of the igniters can be achieved with glass-
to-metal seals. These vacuum-tight assemblies of glasses with metals are used
to feed electrical conductors through the walls of hermetically sealed pack-
ages. A typical glass-to-metal seal consists of an external metal part (in the
following called the "ring"), in which a glass element is sealed [8.1, 2]. The
glass element, in turn, encloses one or more leads (in the following called the
"pins"), which are sealed into it. Glass-to-metal seals are successfully used
in electric engineering and cover a wide range of applications in which the
sealing glass serves as an excellent insulator.
A special feature of airbag igniters is that their "normal operation" is
an "explosion" in close proximity to a person. Therefore, the strength of
the igniters must be sufficiently high in order to avoid personal injury by
components splitting off.
We have investigated the functionality and reliability of air bag igniters
with two-step finite-element analysis (FEA), using the general-purpose FEA
program ANSYS.

8.1.1 FEA for Axial-Symmetric Models

We first consider a simplified axial-symmetric FEA model. The material pa-


rameters and the size (Table 8.1) of this model are close to those of the final
device.

H. Loch et al. (eds.), Mathematical Simulation in Glass Technology


© Springer-Verlag Berlin Heidelberg 2002
414 8. Glass Products Under Mcchanical and Thermal Loads

Table 8.1. Material parameters and geometrical data of axial-symmetric models

Unit Pin Sealing glass Ring


M ateTial parameters
CTE 10- 6 K- 1 9 lOA 18
Young's modulus GPa 150 70 200
Poisson's ratio 0.28 0.22 0.3
Yield strength MPa 300 00 300
Geometrical data
Radius 0.5mm
Inner radius l.355mm
Outer radius 3.52 mm
Height after grinding 2.22mm
Height before grinding 5.2 mm
Hcight after grinding 4.99 mm

Figure S.la shows the ring, sealing glass, and pin of an igniter. The sealing
glass does not completely fill the ring's bore. The dotted horizontal line indi-
cates the grinding face, and the dashed line the rotational axis. The ring-glass
and pin-glass interfaces are modeled by "doubled nodes", i.e., we assign at
an interface separate nodes to each material, all exactly in the same position
as in the adjacent material. Normally the nodes are coupled, but the cou-

a) b)

l.
Grinding
Ro face
axi ~~
~ ~

---
ovM/MPa
0

---
110
~ §§~ 220
Glass ..
330
E
440

--
-
1=
l- 550
Cl
I- =
Cl 660
l- 770

Pin 880

Ring

\
Fig. 8.1. (a) FEM mesh, (b) von-Mises stress for elastic model
8.1 Strength Optimization of Airbag Igniters 415

pIing can be freed (or replaced by contact elements) for fracture-mechanical


considerations (e.g. de-adhesion).
First we treat the device as fully elastic. The stresses are caused by dif-
ferent thermal shrinkage of the components upon cooling the device from the
sealing temperature of 467°C down to 20 0. The colors in Fig. 8.1 b indicate
the von-Mises stress avM in the device in MPa. avM corresponds to plasticity.
The metal ring and pin flow plastically if avM exceeds the yield strength. In
Fig. 8.1 b, avM exceeds 880 MPa in the ring, and thus is higher than the yield
strength (see Table 8.1): The ring will exhibit plasticity.
Figure 8.2a shows the von-Mises stress avM and Fig. 8.2b the correspond-
ing plastic strain EvM if plastic flow is taken into account. The arrow in
Fig. 8.2a indicates the region of maximum avM, which is located in the glass
not exhibiting plasticity. The extended region with plastic flow in the grind-
ing zone of the ring, indicated by an arrow in Fig. 8.2b, will be removed by
grinding.
The ring tapers in the overhanging part (displacement in the figures mag-
nified x30) result in a pivoting at the marked point in Fig. 8.2a and can cause
tension at the ring-glass interface. Because tension at an interface can cause
de-adhesion, the stress resulting from thermal shrinkage and pivoting must
be compressive. Figure 8.3a depicts the radial stress component an i.e., the
stress normal to the interface. It reveals that the entire ring-glass interface
is under compression (by convention, positive stress means tension, negative
stress means compression).
The encircled region in Fig. 8.3a is enlarged in Fig. 8.3b. The enlargement
reveals a discontinuity of the radial stress at the sharp notch of the glass-
ring junction. The discontinuity cannot be removed by choosing a finer FEA
mesh. Mathematically, stresses diverge (i.e., a or T -+ 00) for a notch with a
tip radius R -+ O. FEA is not appropriate for handling sharp junctions!

a) b)

---
0vM IMPa

---
t vM

--=
0 0
84

-=
.002
168
252 .004
336
.006
420
= =
- -
=
504 .008
588
.010
672

Fig. 8.2. (a) Von-Mises stress, (b) plastic strain


416 8. Glass Products Under Mechanical and Thermal Loads

a) b) c)

ur /MPa "I fMPa


-530 -240
-455 -180
-380 -1 20
-305 -60
-230 o

--
60
=
=
-155

= -80
-5
120
180
Ring
70 240

Fig. 8.3. (a) Radial stress, (b) enlarged detail, (c) first principal stress

Such discontinuities can be treated in FEA either by modeling a small


curvature (R > 0) for the junction (because such a curvature is arbitrary, the
FEA results would then depend on arbitrary assumptions), or by ignoring
the small region (.6.z ~ 0.3 mm) where incorrect or less-precise FEA results
are gained. We took this approach.
Figure 8.3c shows the first principal stress O"r in the device. Figure 8.4b
shows O"r and Fig. 8.4a the radial stress O"r in the glass (zoomed sealing glass
and fitted color code)_ The top face of the glass is exposed to compression
(negative values) and should not fracture. The lack of flaws within the glass
(pores, grain boundaries, etc_) means that any failure mode of the glass would

a) b)

---
"r /MPa
--
"1/MPa

-
-550 -210

-- --
-472
-141
-355
-72
-238

= =
=
-3

- --
=
=
-121
=
= 66
-4
135
74

Fig. 8.4. (a) Radial stress, (b) first principal stress


8.1 Strength Optimization of Airbag Igniters 417

require nucleation and propagation of cracks from one surface through the
entire depth of the glass seal (e.g., microcracks will not propagate into the
compression zones of the glass seal). The compressive glass-to-metal seal pre-
cludes the propagation of cracks from the glass surface through the glass vol-
ume. However, at the top of the pin-glass interface we see a small tension
zone that may cause de-adhesion. We neglect this tension zone because it will
be removed during grinding.

Grinding of Top Faces

The grinding process can be modeled in FEA by


• severely reducing the Young's modulus of elements subjected to grinding,
• applying the "doubled-node method", for example by successive decoupling
of nodes.
We applied the first method. Starting at the top, the elastic modulus is
"switched-off" layer-by-Iayer until the final contour is reached. The removed
material is not displayed in the following plots. We discuss the final results.
Figure 8.5a shows the von-Mises stress O"vM and Fig. 8.5b the plastic
strain cpJ after grinding. Figure 8.6a,b gives the radial and the first principal
stress, and Fig. 8. 7a, b the stresses in the sealing glass.
Because the FEA elements are rectangular-shaped, the deformation of
the device causes a curved top face, as can be seen in Fig. 8.5a, for example.
However, the deformation in Fig. 8.5a is actually very small (please note that
the displacements are magnified by x 30), so that the height ("non-flatness")
of b.z ;:::0 6 ~m can be neglected.
From Fig. 8.6b we see that O"I is compressive in the sealing glass. Thus
the sealing glass will not break. The stress discontinuity at the glass-ring
junction (arrow in Fig. 8.6a) can be neglected.

---- --==
a) b)
C1vM IMPa t vM
0 _ 0
77

-
.002

-
154
231 .004
308
- .006
= 385
=

-
=
=
c:::::l
462 §S .008
539
.010
616

Fig. 8.5. (a) Von-Mises stress (~z :::::: 6.3 ).lm, displacement magnified x30),
(b) plastic strain in the ground device
418 8. Glass Products Under Mechanical and Thermal Loads

a)

---
(]r fMPa
-500
-436
b)

---
0 1fMPa
-205
-150

--- -
-372 ·95
-308 -40
-244 15
=

-
-180
= 70

--
= ·116
-52
=
= 125
180
12 235

Fig. 8.6. (a) Radial stress, (b) first principal stress in the ground device

Figure 8.7a shows that the entire ring~glass and pin~glass interface is
exposed to compression. This is important for the tightening of the device.
The compression at the top (arrow in Fig. 8.7a) of ~ -10 MPa prevents glass
fracture during grinding.
The first principal stress in the glass amounts to +40 MPa (Fig. 8. 7b).
In general, such high tensions are relevant for glass. However, this particular
stress appears in the interior of the pin~glass interface and is not exposed
to moisture (no subcritical crack growth by stress corrosion). Thus, rupture
caused by this stress component is unlikely.

Rupture of Glass-to-Metal Interfaces

Sticking of materials can b e caused by adhesion or friction. Adhesion is pre-


ferred because it results in outstanding tightening. Sections 8.1.3 and 8.1.4
show experimental results for pins and glasses exposed to high loads. The

- 10 MPa - 10 MPa
a)
(]r IMPa
b)
I

-
_ -500 UI fMPa
40 MPa
= -435
- -370
=

-
- 190

- 142
-305
-240 - 94
~ - 175 =
=
--
-4S
~ -110
=
-
-45 2
20 40 MPa 50

Fig. 8.7. (a) Radial stress in the sealing glass, (b) first principal stress
8.1 Strength Optimization of Airbag Igniters 419

tests reveal that the adhesion strength at t he pin- glass and ring- glass inter-
face is very high and de-adhesion is unlikely. However, reliable devices must
tighten without adhesion. Therefore, the ripped pin-glass or ring- glass inter-
face must be exposed to compressive stress. In axial-symmetric models, the
radial stress uris just the pressure (despite the sign, negative stress values
mean pressure).
The ripping is modeled by freeing "doubled nodes" at interfaces (see
the model description on p. 414). We consider the failure of a sealed device
(Fig. 8.8). Two assumptions complete the model description:
• We apply Coulomb friction. Thus, the materials stick if the shear stress at
the interface fulfils

171 < fLlunl for Un < 0 (compression) (8.1)

or slide with 7 = ±fLlul, where 7 is the shear stress at the interfaces and fL
is the friction coefficient. We apply fL = 0.3 for a worst-case estimation .
• The ripping is modeled by successively freeing bonds element-for-element
along the paths (see Fig. 8.8). The results depend on the direction (i.e.,
bottom-up or top-down), however, the difference is too small to be clearly
resolved in Fig. 8.9.
Figure 8.9 shows the radial stress in the device for a freed pin- glass and
ring-glass interface, respectively (compare with Fig. 8.3a). Figure 8.9a (freed
pin-glass interface) reveals that the pin moves to the top. This is found for the
top-down as well as for the bottom-up rupture. In Fig. 8.9b (freed ring-glass
interface), the pin-glass core moves to the top.
Figure 8.10 shows the stress along the pin-glass interface for the sticking
(dashed bold curve) and slipping interface (thin solid curves; the curves for
bottom-up and top-bottom freeing coincide). The shaded area is the region
with incorrect FEA results due to the sharp ring- glass junction. Figure 8.10
reveals that the stress distributions do not change essentially if de-adhesion
at the interface occurs: An extended region with high compression remains

Grinding
fac~~
' - -'
Pin-
glass
Ring-
glass
interface interface

I---

Fig. 8.8. Coordinates of slipping inter-


- faces
420 8. Glass Products Under Mechanical and Thermal Loads

=
a) or /MPa b) or /MPa
_ -367

=
_ - 503

=
= -439 - 316

=
= -374 - 266
= - 310 - 215
= -246 - 165

=
= - 182 - - 114
_ -11 7 ~

- 14
-53
- 11 36

Fig. 8.9. Normal stress for bottom-up ripping, displacement magnified x30,
(a) pin- glass, (b) ring-glass interface freed

along the pin-glass interface. Thus we conclude that the device tightens even
in the case of de-adhesion of the pin-glass interface. Note that the shear stress
is very low.
Figure 8.11 shows the stress along the ring-glass interface for adhering
(dashed bold curve) and freed interface (thin solid curves; curves for bottom-
up and the top-bottom freeing coincide). Here, the stress changes drastically
by de-adhesion. Because in all cases a large region with very high compressive
stress along the pin- glass interface exists, we conclude that the device is tight
in the case of de-adhesion of the ring- glass interface. Conclusion: The airbag
igniter tightens even for ripped pin- glass or ring-glass interfaces.

100
Pin-glass interface

o ----- ....,..--.,...~_ _/~


- .r..-I

ro -1 00 rn
(L

E
:5 - 200 Top-down
bottom-up

- 300
- - Sticking
- - Friction (u = 0.3)
-4OO +-------,--------r-------.-------.--~
o 0.5 1.0 1 .5 2.0
z Imm
Fig. 8.10. Normal and shear stress for slipping at the pin- glass interface
8.1 Strength Optimization of Airbag Igniters 421

200 ~---------------------=~--------~
Ring-glass interface

- - Sticking
100 - - Friction (p. =0.3)

ra
a.
E....
ti' -100

Or'""\. .... _
-200
, , - .----~
-300 +---~~~~;:::=----r---____r-~
o 0.5 1.0 1.5 2.0
z Imm
Fig. 8.11. Normal and shear stress for slipping at the ring- glass interface

8.1.2 FEA of 3D Models

Figure 8.12 shows the mesh of a three-dimensional FEA model of the opti-
mized air bag igniter. The model is cut at the plane of symmetry. The igniter
has no rotational symmetry for electrical purposes (being not the subject of
the present optimization).
Figure 8.13a shows the von-Mises stress (TvM and Fig. 8.13b the corre-
sponding plastic strain in the ground 3D device. The displacement in the
figures is again given in enlarged scale 1:30. Note the bending of the pin due
to the asymmetry of the device.
The maximum von-Mises stress of 672 MPa appears in the glass, in the
metals (TvM ::;; 300MPa holds. Figure 8.13b shows an extended zone of plastic
strain at the grinding face (mark a). However, the maximum strain of l.6%
appears at mark b; see the discussion of Fig. 8.2b for an explanation. The

Fig. 8.12. FEA mesh for the airbag


igniter
422 8. Glass Products Under Mechanical and Thermal Loads

=
O"vM/MPa £vM I MPa
o

=
_ 0

=
0.002

=
0.004

=
0.006
0.008

=
0.01
0.012
0.014
- 0 .016

Fig. 8.13. (a) Von-Mises stress, (b) plastic strain

radial stress (normal to the interface) and the shear stresses Trz (stress com-
ponent at the glass-pin interface acting in axial direction) and Trip (stress
component on the glass-pin interface acting in tangential direction) of the
igniter are represented in Fig. 8.14. The excess stress marked at the edge of
Fig. 8.14b is caused by the sharp glass-ring junction. Figure 8.15 shows the
stresses in the ring, and Fig. 8.16 the first principal stress CTr in the sealing
glass. Obviously, compressive stress dominates in the glass part.
Some igniters have pins consisting of softer metals. The yield strength of
such pins is only 170 MPa. All other material parameters are as specified in
Table 8.1. Figure 8.17 shows the von-Mises stress in a device modeled with

a,/MPa

=
'rz IMPa

=
= -490 _ - 105

=
-45
-420

=
- 350 15
-280 75
-210 135

-
- 140 195
255
315

=
' r." fMPa 375
_ -82

- 70
= - 58
= -46

=-34
- -22
- 10
Fig. 8.14. (a) Radial stress, (b) shear
2 stress Trz, (c) shear stress Tr'P - pin removed
- 14 for clarity
8.1 Strength Optimization of Airbag Igniters 423

Or / MPa trz /MPa


_ - 104
- 72
--40
-8
24
56

-
88
120
152

----
Tr<p IMPa

- -26

---
- - 13

-
Fig. 8.15. Stresses in the ring: (a) radial
stress (normal to the interface), (b) shear
component Trz, (c) shear component Tr'P -
- 13 pin and glass removed for clarity

these data. A comparison with Fig. 8.13a reveals that the stress level (in the
glass) is not affected by the softer pin material.
For this device, Fig. 8.18 shows the shear stresses at the pin-glass and
ring-glass interface, respectively. The pin and the glass have been succes-
sively removed in the plots for clarity. The stresses are almost the same as in
Figs. 8.14 and 8.15. We therefore conclude that the yield strength of the pin
has a minor effect on the stress distribution in the device.

ol/MPa
_ - 200
--
ol /MPa

--
- 160

- -80

---
o c::::J
0
40
80 80
120

Fig. 8.16. First principal stress in the glass, pin and ring removed. (a) View into
bore, (b) view on circumference from the bottom
424 8. Glass Products Under Mechanical and Thermal Loads

O"vM IMPa
o
84
168
252
336
420
504
588
672 Fig. 8.17. Von-Mises stress in a device with
a reduced yield strength of 170 MPa of the
pin

Trz IMPa Trz IMPa


_ -105 _ -104
-45 - 72
15 -40
75 -8
135 24
195 56
255 88
315 120
375 152
Fig. 8.18. Shear stresses for a device with a yield strength of 170 MPa of the pin,
(a) at the pin-glass interface and (b) at the ring-glass interface

8.1.3 Pull-Out Tests

During the handling or installation of igniters in air bag systems some force is
applied to the igniter pins. In order to guarantee that no damage occurs, the
robustness of igniters to such loads is assessed in pull-out tests. The testing
apparatus consists of a support unit for mounting the ring of the igniter and
of a cramp for applying a tensile force on the pin (see Fig. 8.19).

a) b)

Fig. 8.19. Photographs of the testing apparatus, (a) top view on igniter and (b)
side view on pin and cramp
8.1 Strength Optimization of Airbag Igniters 425

An Instron testing machine has been used with the jig moving to the
top. The igniter's pin is fixed by the cramp at the bottom of the apparatus.
The igniter is supported by the mounting, which is pulled to the top until
the fracture/rupture occurs. The testing machine operates in force-controlled
mode with a load rate of 10 N Is. Two types of specimens have been tested:
type-I specimens with a ceramic solder (predecessor) and type-II specimens
with a glass solder (optimized igniter).
Table 8.2 shows the statistical data for the failure force. All data points
are taken into account, irrespective of the different failure reasons. We usu-
ally apply the normal and the Weibull distribution on fracture data. The
statistical procedures for data evaluation are briefly described in Sect. 8.1.6.
Note the X 2 values in Table 8.2. For a high X2 value, the statistical dis-
tribution (or statistical hypothesis) should be rejected. For a confidence level
of 95% an upper limit of 5.99 for the X2 value is recommended. Obviously,
the statistical distribution of the failure data for the type-I specimens can
be neither described by a normal nor by a Wei bull distribution. The reason
can be gathered from Fig. 8.20, where the Wei bull distribution of the data is
depicted: There is one "outlier" with an excess strength value. If we neglect
the outlier (Table 8.2), the data fit the normal and the Weibull distribution.
Because in fracture mechanics the Wei bull distribution is preferred, we
focus our further discussion on the Wei bull distribution:

(8.2)

F denotes the cumulative fracture probability, I the applied force, Ie the


characteristic value of the distribution and m the Wei bull modulus (see Ta-
ble 8.2). Figure 8.20 shows a so-called Weibull plot of the data (note the
nonlinear scale). Dots depict data points and closed lines represent Wei bull
approximations (straight lines in a Weibull plot) to the data for the type-I
(outlier neglected) and type-II specimens. Obviously the approximations fit

Table 8.2. Statistical data for the failure tensile stresses


Type I Type I Type II
outlier
neglected
sample size N 50 49 50
minimum 205.6 205.6 446.1 N
maxImum 398.4 263.6 462 N
Weibull characteristic value fe 241 234 457 N
distribution Wei bull modulus m 13.3 22 191
X 2 value (5 classes) 23 5.6 2.1
Normal mean value <f> 232 229 456 N
distribution standard deviation b 27 13 3 N
X2 value (5 classes) 27 4.6 3.2
426 8. Glass Products Under Mechanical and Thermal Loads

0.99 .----------,-.----:--------..,~__,r__l

0.9
Outlier
lor type I
0.5

u..

0.1

,Type II

0 . 01 +-------L,-------r-----r--~r_--r_~r_~

150 200 250 300 350 400 450 500


fiN
Fig. 8.20. Weibull plot of the failure forces (f) for the pull-out test

the data points quite well. The slopes of the lines were proportional to the
Weibull moduli. The dashed line represents a Weibull approximation to the
data points with the outlier taken into account. In this case the line does not
fit the data points.
Figure 8.21 shows an extrapolation of the Weibull distribution to small
fracture probabilities based on (8.2) and the data in Table 8.2. The axis for
the fracture probability has a logarithmic scale. The dashed lines correspond
to the 95% confidence range of the Wei bull approximations, i.e., the true
fracture probability lies with a probability of ;:: 95% within the corresponding
range.

0.99

0.10 Type II

0.01
II
u.. 10- 3
If

10-4
x 2.5 "' 1
II
/I
10- 5 1/

1~ +--.-~r--.--.--.--.--r~-r-~
50 100 150 200 250 300 350 400
"450 500
fiN
Fig. 8.21. Extrapolation to small fracture probabilities
8.1 Strength Optimization of Airbag Igniters 427

If we suppose that the pin of an airbag igniter is not pulled by more than
150 N, the fracture probability is approximately 10- 4 for type-I specimens,
and far below 10- 6 for type-II specimens. (An extrapolation on the basis of
the data in Table 8.2 yields F < 10-90, but an extrapolation to such low
probabilities makes no sense for a sample size of N = 50).
It is informative to see what happens if we take the outlier for the type-I
specimens into account for extrapolation. With (8.2) we then get F::::; 0.2%,
i.e., a fracture probability about 20 times higher than calculated without the
outlier [8.3-5]. At first sight this is surprising because the outlier has twice
the strength of the other specimens of this sample. In fact, the median failure
force increases slightly if we take the excess strength into account. However,
the outlier has a broader fracture distribution, and this results in lower values
for the "safe strength" .

Fractographic Failure Analysis

A visual inspection of the broken type-I (ceramic solder) igniters shows that
the pin was pulled out of the ceramic solder in all specimens but one, and
this exception is the outlier, where the pin was broken. Figure 8.22a,b shows
photographs of the top view of a tested igniter, Fig. 8.22c gives a bottom
view into the ring's borehole. In Fig. 8.22b, the pin is partly and in Fig. 8.22c
completely pulled out of the igniter. In no case has a rupture at the ring-
glass interface been observed. Conclusion: type-I igniters break in pull-out
tests predominantly due to lack of adhesion at the pin-glass interface.
The analysis results for type-II (glass solder) igniters are totally different.
The pins of all tested igniters were broken, no pin had been (completely)
pulled out. Figure 8.23 shows a photograph of a broken pin. Note that for

a) b)

Ceramic
solder

Fig. 8.22. Tested type-I igniter. (a, b) Top view,


(c) bottom view into the ring's borehole. In (b)
the pin is partly, in (c) completely pulled out
428 8. Glass Products Under Mechanical and Thermal Loads

.2 mm.

Fig. 8.23. Broken pin of a type-II igniter (tension force


F ~ 400N)

high tension forces, fracture of the pin is very likely. The estimated tensile
stress in the pin is

O"tension = f / (7r r2) ::::0 510 MPa , (8.3)

where r is the radius of the pin (r = 0.5mm, see Table 8.1). Under such high
tensile stresses, fracture of the pin must be expected.
Figure 8.24a gives the top view of a type-II igniter. The top face is flat.
This is the case for 47 out of 50 tested specimens. In the specimen shown
in Fig. 8.24b, the pin has been slightly pulled into the hole (~z ::::0 0.1 mm).
This effect was observed in three of the 50 specimens. The physical reason
for the movement is still unclear. It is probably related to a small plastic flow
of the pin material. Conclusion: type-II igniters break in pull-out tests by a
rupture of the pin.

a) b)

Glass
solder

c)

Pin
Glass

Fig. 8.24. Tested type-II igniter. (a, b) Top view


(tension force: F ~ 400 N), (c) view from the bot-
tom into the bore; no fracture visible
8.1 Strength Optimization of Airbag Igniters 429

FEA of the Pull Tests

Stress Distribution for Adhering Interfaces. We use the two-dimensional


model discussed in Sect. 8.1.1 for finite-element analysis. In accordance with
the experiments, a tensile force f is applied to the pin. The ring is supported
by constraining the vertical displacement of the ring's lower edge (indicated
by triangles in Fig. 8.25a). Figures 8.25 and 8.26 show the radial and shear
stresses in the device for f = 200 Nand f = 400 N, respectively.
Figures 8.27 and 8.28 show the normal and shear stresses at the pin-glass
and ring-glass interface. We restrict our discussion of stresses to values of
z > 0.3 mm because, owing to the sharp pin-glass and ring-glass junction,
FEA results are incorrect at z -:r 0 (see Fig. 8.8 for the definition of z). This
region is indicated in Figs. 8.27 and 8.28 by shading. The corresponding data
for the unloaded device (J = 0) are also shown for comparison.
We see from Figs. 8.27 and 8.28 that the normal stress at the interface
decreases for increasing load, whereas the tangential stress increases. In all
cases, the compressive stress is still high (Iar I > 350 MPa for the pin-glass and
> 200 MPa for the ring-glass interface) in large regions along the interface.
We see further that the influence of the tensile force on the shear stress
is stronger for the pin-glass than for the ring-glass interface. This difference
can be explained by a simple estimation of the mean additional stress from
the tension on the ring:

-- f { 11 MPa for ring-glass


(8.4)
I'rrz I = 21f r h ~ 30 MPa for pin-glass '

where r is the radius of the pin or ring, respectively, and h is the height of
the sealing glass, see Table 8.1). The different influence of the tensile force on

a) r- -

---
ur /MPa
- 550 b)
---
Trz IMPa
- 100

-- --
-470 -40
- 390 20
- 310 80
1'-' Ir""
- 230 140

--
I'
=
t=l
=
-
- 150 200

=
c::J c::J
c::J
-70 260
10 320

200 N •••• 90
200 N
380

Fig. 8.25. (a) Radial stress and (b) shear stress for f = 200N
430 8. Glass Products Under Mechanical and Thermal Loads

= ---
ar IMPa 'rz IMPa
a) - ,....."",.- ._ _ ~ _ -576 b) _ -130

=
--
- 70

=
-469
- 362 - 10

=
=
- 255
-148
50
110

= --41
= 170

-- -
I:=J
I:=J
66 =
= 230
290
173
280 350
400 N

Fig. 8.26. (a) Radial stress and (b) shear stress for f = 400N

200 ~----------------------------------~
Ring-glass interface

100

nI
a. o
~
....
o - 100

- 200
400 N
-3oo+-----!~..,--------..,--------.,--------.,--­
o 0.5 1.0 1.5 2.0
z/mm
Fig. 8.27. Normal and shear stress at the ring-glass interface

the shear stress simply results from the different outer radii of the pin and
sealing glass.
The stress distribution at the interfaces does not disclose the reason for
the fracture of the type-I devices for f = 200 N. The reason may be gathered
from Fig. 8.29, where the first principal stress distribution in the device is
plotted. Here, we see that for f = 200 N a stress of 25 MPa, and for f = 400 N
a stress of about 40 MPa occurs. We conclude that the ceramic of type-I
devices fractures because of the high first principal stress and that the glass
of type-II devices withstands a stress of 40 MPa.
8.1 Strength Optimization of Airbag Igniters 431

1oo ,---~~--~~~----------------,

~ - 100
~
...
ti - 200

-300

-400 +P
:....;i"-'-'=:..:....;..c.r-:.c..:.:.:~_--,_ _ _--.-_ _ _---,_----'
o 0.5 1.0 1.5 2.0
z I mm
Fig. 8.28. Normal and shear stress at the pin-glass interface

, 25 MPa

---
a(/MPa

---
- 200
, 40 MPa

----
a( fMPa
- 200

---
- 125 -80
-50 40
25 160
100 280

---
175 400
= =

-
250 =
= 520
325 640
200 N 400 400 N 760

Fig. 8.29. First principal stress for f = 200 Nand 400 N

Stress Distribution at Ripped Interfaces. We focus on the question of whether


or not the glass-metal interfaces are compressed when the pin is pulled. The
procedure is the same as described in Sect. 8.1.1 for analyzing the stress
distribution in the devices.
Figure 8.30a, b shows the radial stress in the device for a de-adhered pin-
glass and ring-glass interface, respectively (compare with Figs. 8.3a and 8.9).
The deformation of the device again appears disproportionately pronounced
due to the magnification of the displacement by x30.
Figure 8.31 shows the stress along the ripped pin-glass interface for f = 0
and 200 N. The shaded area indicates the region with incorrect FEA results
432 8. Glass Products Under Mechanical and Thermal Loads

---- -
or /MPa b) or fM Pa

---
a)
-530 - 307
-461 - 257

---
- 392 - 208
- 323 - 158

---
-254 - 109

---
- 184 - 59
- 115 -10
-46 40
23 200 N 90

Fig. 8.30. Normal stress for bottom-up ripping, magnification of displacement


x 30; (a) pin- glass, (b) ring- glass interface freed

100 . - - -- -------------------------------,

o
f=200 N
f=0
t1. - 100
E
~
-200

- 300
Pin-glass interface
frict ion IJ =0.3
-400 +-------.--------r------~~~~~~~
o 0.5 1 .0 1 .5 2 .0
z fmm
Fig. 8.31. Normal and shear stress for a slipping pin- glass interface

caused by the sharp ring- glass junction. The pressure level of > 200 MPa is
sufficient for tightening. Note that the shear stress is low.
Figure 8.32 shows for the ring- glass interface also a diminution of the
pressure on the glass when the pin is pulled. As with the pin-glass interface,
the pressure level of > 200 MPa is sufficient for tightening. The optimized
airbag igniters are capable of tightening even ripped pins or rings.
8.1 Strength Optimization of Airbag Igniters 433

100 ~--------------------~--~~--~---;
Ring-glass interface

o mn ~3 m0
f == 0
f == 200 N trz )J
8!.
E... -100
t:i"

- 200

- 300 -I-----~:::::===::::;:::=----,---------,----.J
o 0.5 1 1.5 2
z Imm

Fig. 8.32. Normal and shear stress for a slipping ring-glass interface

8.1.4 Push-Out Tests

The robustness of airbag igniters is also assessed in push-out tests. The ex-
perimental equipment resembles that of the pull-out test (see Fig. 8.19), but
now pin and sealing glass are pushed from the top (ground face) by a cylin-
drical steel punch (diameter 2.5 mm). An Instron testing machine has been
used with the jig moving from the top. The testing machine operates in the
force-controlled mode with a load rate of 100 N/s.
Table 8.3 gives the statistical data for the failure force. All data points
are taken into account, irrespective of the different failure reasons. We first
treated the normal and the Weibull distribution on fracture data. As the X 2
values in Table 8.3 show, the normal distribution is obviously unsuitable for
describing the failure data of the type-II specimens. We therefore restrict our
discussion to the Wei bull distribution. Figure 8.33 shows the fracture data
displayed 3.'l a Weibull plot. The dots depict the data points and the straight
lines are Wei bull approximations. We see that the fracture statistics of the
type-I and type-II specimens are very similar.
Let us suppose as, a worst case, that during the explosion (ignition) a
gas pressure of p = 2 kbar (= 2000 atmospheres ~ 200 MPa) is generated at
the igniters' top sides. This corresponds to a resultant force on the sealing
glass/ceramic and pin of about

f = 7fr2 p ~ 1200 N (8.5)


(r is the radius of the ring's bore, r = 1.355 mm). The fracture probabil-
ity of the devices is calculated from the data in Table 8.3. For the Wei bull
distribution (8.2) we get

m] {2 X 10- 12 type-I (8.6)


F = 1- exp [(- fife) ~ 3 X 10- 13 type-II .
434 8. Glass Products Under Mechanical and Thermal Loads

Table 8.3. Statistical data for the failure pressing force

N
N
N

N
N

0.01 .J-.'-£_--~---~---~--~-----1
4000 4500 5000 5500 6000 6500
fiN
Fig. 8.33. Fracture probability versus pressing force

These fracture probabilities are negligible, for both cases.

Fractographic Failure Analysis

In principle, fractured specimens of type-I and type-II igniters look similar.


Figures 8.34 and 8.35 show examples of broken type-I and type-II igniters,
view on the top face. It can be clearly seen that the sealing ceramic, respec-
tively the sealing glass, is pushed into the bore. The depth differs: for the
type-I igniters it is 0.3 ... 0.7mm, for type-II igniters 0.2mm. This difference
is not significant.
In the ceramic solder of type-I devices, fractures are visible at the top face
(the ceramic is not even transparent). Such fractures occur regularly in type-
I igniters, in contrast to type-II igniters. This fact might be important for
8.1 Strength Optimization of Airbag Igniters 435

Fig. 8.34. Type-I igniter


(force f ~ 5000 N)

Glass
solder

Fig. 8.35. Type-II igniter


(force f ~ 5000 N)

considerations concerning the potential damage mechanisms during ignition


of a device.
Figure 8.36 shows two views from the bottoms into the bores. They reveal
that both the ceramic and the glass sealings break on the bottom side. Con-
clusion: the devices break at very high forces; fractographic analysis yields
similar results for type-I and type-II igniters.

8.1.5 Pressure Tests

The normal operation of an airbag igniter is a detonation in proximity to a


person. The detonation causes high pressure on the top face of the igniter.
To avoid personal injury, any splitting-off of parts of the igniter must be
prevented. The adequate reliability test for such a failure is to apply high

a) b)

Fig. 8.36. View from the bottom side, fracture sealings (f ~ 5000 N). (a) Type-I
igniter (ceramic sealing), (b) type-II igniter (glass sealing)
436 8. Glass Products Under Mechanical and Thermal Loads

(hydrostatic) pressure to the top face of the igniter and determine the maxi-
mum pressure that can be applied without a failure (so-called push-out tests).
Figure 8.37 shows a testing apparatus consisting of a pressure cell and
customary devices for generating and measuring the pressure. Water is used
as liquid for pressure transmission. The pressure is increased manually.
Failure of igniters is defined by leakage. We find that for type-I igniters
leakage starts at 1880 bar, and for type-II igniters at 2410 bar (sample size:
N R::: 20; lowest pressure for leakage in a sample). In no case has catastrophic
damage (push-out of the pin and/or the sealing glass) been observed, i.e., the
fracture pressure of both devices is sufficiently high.

Summary

In our two-step FEA analysis, we first analyzed simplified axial-symmetric


(2D) models to achieve coarse optimization. The geometrical data and the
material parameters (metals and solders) have been varied. The data from
the coarse simulation were transferred to 3D models for refinement. FEA
yields detailed information about the mechanical behavior of air bag igniters
in "normal operation" and about failure cases.
The robustness of the igniters with respect to handling and ignition loads
(i.e., the explosion) has been assessed in pull-out, push-out, and pressure
tests. The experimental results show that the strength of the improved igniter
is remarkably higher than that of a predecessor.
• The pull-out strength of the pin is increased by a factor R::: 2, and the
push-out probability of the glass sealing has been reduced to R::: 1/10.
• The optimized igniters have passed robustness tests without failure or
failure propagation (catastrophic damage). Development of leakage or
splitting-off of components (in the case of an explosion) are not to be
expected.

Fig. 8.37. Pressure cell


8.1 Strength Optimization of Airbag Igniters 437

• The fundamental design of the compressive glass-to-metal seal precludes


the propagation of cracks through the glass volume. This is completely un-
derstood from the finite element analysis. The optimized igniters (damaged
in tests by excess loads) have a superior defect structure. For example, they
are tight even when the solder fractures.

8.1.6 Appendix: Statistical Procedure

In the following mathematical procedure we fit the measured strength values


by Weibull and normal distributions. Whereas the Wei bull distribution is
preferred in fracture mechanics, the normal distribution is commonly used in
statistics. For the normal distribution, the probability density P(x) and the
cumulative distribution F(x) read [8.6,7]

P (x) =
1
V27rb exp -
[(X 2b- xo)]
2 '
(8.7)

F (x) = ~ [1 + erf ( x ~~o )] , (8.8)

where Xo is the mean value and b the standard deviation. For the Wei bull
distribution the values are

P(x) = : (:J m-l exp [- (:J m] , (8.9)

F(x)=l-exp [-(:J m
] , (8.10)

where Xc is called the characteristic value and m the Weibull modulus. Fig-
ures 8.38 and 8.39 sketch examples for normal and Wei bull distributions.
In general, the probability density is more easily handled with the nor-
mal distribution, whereas the cumulative distribution function more easily
handled with the Wei bull distribution. For a rough estimation, the relations
between the characteristic value and modulus of the Weibull distribution and

1.6 a) 1.0 b)

1.2 0.8
0.6
a.. 0.8 u.
0.4
0.4
0.2
0 0
-1 0 1
x /xo
2 3 -1 o 1 2 3
x /xo

Fig. 8.38. Probability density and cumulative distribution of normal distributions


438 8. Glass Products Under Mechanical and Thermal Loads

1.6 a) 1.0 b)

1.2 0.8 -------------,---

0.6
Q. 0.8 u..
=2 0.4 m=4
0.4 0.2

0 0
-1 0 1 2 3 -1 0 1 2 3
x/xc x/xc

Fig. 8.39. Probability density and cumulative distribution of Weibull distributions

between the mean value and standard deviation of the normal distribution
can be given with

Xo ~ Xc r m+1)
(---:;;;:- (8.11)

and

(8.12)

respectively. Both, normal and Wei bull distributions are used for extrapola-
tions and for defining confidence regions. For the normal distribution, this
can be done in accordance with textbooks on statistics. For the Weibull dis-
tribution, we apply the DIN (ISO) 53303.7 standards.

8.2 Stiffness and Weight Optimization of a Reticle Stage


for Optical Lithography

Volker Seibert
Introduction

The production of microprocessors and memory chips is a strongly growing


market with a constant demand for faster, more efficient, and cheaper in-
tegrated circuits. In wafer steppers (see Fig. 8.40) and wafer scanners, the
structures of the subsequent semiconductor components are projected by an
optical system from an evenly illuminated mask (termed reticle) onto the
wafer, which is coated by a photoresist. This production step is termed op-
tical lithography. Approximately every five to eight years, the increase in
production due to the size reduction of the circuits reaches a limit. As a con-
sequence, larger wafers and larger reticles are needed. Instead of the circular
8.2 Optimization of a Reticle Stage for Optical Lithography 439

Fig. 8.40. Global view of a wafer stepper unit (see www.zeiss.de)

wafers with a diameter of 200 mm and the corresponding quadratic (6 x 6)"


reticles used up to now, the development of wafers with a diameter of 300mm
and of (9 x 9)" reticles is planned for the near future. Figure 8.41 shows the
scheme of a reticle stage.
The transition towards larger dimensions requires a complete reconstruc-
tion. This section will present the development of a new design for a reticle
stage appropriate for the new requirements.

8.2.1 Requirements for a (9 X 9)" Reticle Stage

The projection lens projects the reticle structures onto the wafer demagni-
fied, for example, by a factor of 4. In a wafer stepper the wafer is moved
under the optics by the size of the image field after each exposure process,
so that the wafer is exposed completely step by step. In order to handle ef-
fectively the ever-increasing areas, in the next-generation wafer scanners the
projection lens does not expose the whole image in one step, but the reticle
is scanned continuously by a smaller rectangular slit, which is then projected
onto the wafer. During the scanning, the reticle is moved over the fixed slit.
The accelerations of the reticle stage due to the horizontal moving by linear
motors typically are 50 m S-2. This process can excite vibrations that lead to
imaging faults on the wafer during the exposure, i.e., rejects are produced.
The requirements of the system lead to two fundamentally opposed opti-
mization objectives: the minimization of the moving mass and the maximiza-
440 8. Glass Products Under Mechanical and Thermal Loads

Vacuum suction area

-< 350
560

CutA- B

I -----,----,I
, - - - , - - - I 1;06~
A
450
Fig. 8.41. Scheme of a reticle stage

tion of the dynamical stiffness. Our aim was to develop a new reticle-stage
design that meets the geometric and kinematic requirements listed below.
The kinematic boundary conditions finally depend on the properties of the
available motors.

• Geometric requirements:
~ exterior dimensions approximately 560 x 450 x 66 mm 3 ;
~ diameter of the reticle fixing pot 350 mm, wall thickness in the fixing pot
20mm;
~ bore diameter for the z actuator approximately 100 Illm;
~ symmetrical structure with respect to the longitudinal axis:
~ flange for the linear motor;
~ vacuum suction area for the reticle;
~ minimum wall thicknesses 2: 5 mm .
• Kinematic requirements:
~ 1st bending eigenfrequency around the x axis > 500 Hz,
~ mass of the reticle stage < 14 kg.
8.2 Optimization of a Reticle Stage for Optical Lithography 441

For different reasons, the 1st bending eigenfrequency around the x axis is
the decisive one here. It refers to the final reticle stage including all additional
masses (the reticle itself and the motors for the movement).

8.2.2 Design of a Prototype

First, a prototype was developed without mathematical methods, simply on


the basis of the experiences with smaller reticle stages. As a first step, a
component of glass ceramics with a simple geometry was produced. The raw
block of dimensions 560 x 450 x 66 mm 3 had a mass of approximately 42 kg. In
order to fulfil the condition of a maximum mass of 14 kg, the material had to
be reduced by about 2/3. The next step was a modification of this geometry
by established principles of design for lightweight construction structures of
glass ceramics (mirrors for satellites, etc.). In order to reduce the weight,
quadrangular recesses were milled into the block. For increased stiffness, these
recesses were equipped with horizontal stiffening ribs. Figure 8.42 shows the
top and bottom of the resulting prototype, and Fig. 8.43 represents a cross
section with and without stiffening rib.
The dynamical properties of this prototype were tested by experimental
modal analysis. The measurements carried out in this analysis only concerned

Quadrangular

Fig. 8.42. Prototype (top


and bottom)
442 8. Glass Products Under Mechanical and Thermal Loads

Fig. 8.43. Recess with and without stiffening rib

the reticle stage without the parts to be added later (linear motor, reticle, and
z motors). The second column of Table 8.4 shows the measured frequencies
of the first four eigenmodes.
The additional masses lead to lower eigenfrequencies. In order to estimate
this influence, the system can be reduced to a simple bar model with a point
mass at one boundary. On the condition that the additional mass does not
influence the stiffness, this results in a frequency of the 1st bending around
the x axis of fwith mass = 405 Hz. Compared to the measured value of 694 Hz
without additional masses, this means a drastic reduction and an evidently
lower frequency than the required 500 Hz. Here, we decided to optimize the
component by means of FEM (= finite-element method).
First, the existing prototype was modeled and analyzed by FEM. Table 8.4
shows the computed and the measured eigenfrequencies of the prototype with-
out additional masses. They correspond very well. For the relevant bending
eigenfrequency, in particular, the deviation is < 2%. Figure 8.44 illustrates
the respective computed eigenmodes. Backed by these promising results, we
used our simulation method to carry out the remaining optimizations.

8.2.3 FEM Optimization Without Additional Masses

Before attacking an optimization problem, the relevant parameters must be


identified and their principal influence must be analyzed. The strongest ide-
alization of the reticle stage without additional masses is a one-dimensional
bar model (Euler-Bernoulli bar), which can be used for an easy estimation
of the first bending eigenfrequency. We use the following symbols: E is the
modulus of elasticity, I the geometrical moment of inertia against bending, A

Table 8.4. Measured and computed eigenfrequencies of the prototype

Eigen- Measurement Simulation Description


frequency
no. Hz Hz
1 461 466 1st torsional eigenmode
2 694 707 1st bending around the x axis
3 1120 1222 1st bending around the y axis
4 1300 1249 torsion around the y axis and
bending around the x axis
8.2 Optimization of a Reticle Stage for Optical Lithography 443

Fig. 8.44. Shapes of eigenmodes


(- bending axis),
above: torsional eigenmode,
h = 466Hz,
center: bending eigenmode around
the x axis, h = 707 Hz,
below: bending eigenmode around
the y axis, h = 1222 Hz

the area of the cross section, and L the free bending length. On the condition
that EI = const. and that there are no exterior load effects, the free bending
eigenmodes are described by the following equation of motion:

Elz"" + pAz = 0 . (8.13)


The computation of the eigenfrequencies is based on the assumption that
the component is not subject to any compulsory conditions at its boundaries,
i.e., that the boundaries are free. In this case, we have the following solution
for the 1st bending eigenfrequency (here, ,\ is the 1st eigenvalue):

W = ('\L)
2
V(EI
p;w = 4.73 V(EI
2
p;w. (8.14)

As a qualitative optimization rule, we can conclude: in order to increase


the bending eigenfrequency, the geometrical moment of inertia must be in-
creased and the cross-sectional area or the total mass must simultaneously
be reduced.
The starting point for the optimization was the prototype design. First,
only the bore sizes and the position of the quadrangular recesses were mod-
ified. However, the values of the resulting bending eigenfrequencies always
varied only between 690 Hz and 730 Hz. Therefore, the design was fundamen-
tally changed:
444 8. Glass Products Under Mechanical and Thermal Loads

• The lateral profiles with rectangular bores were substituted by closed box
profiles, which have a larger geometrical moment of inertia against bending
(additional red area in Fig. 8.45) and a smaller mass due to the missing
transverse links. The wall thicknesses remain unchanged at 5 mm.
• The quadrangular recesses were substituted by triangular ones. The ad-
vantage can be illustrated by means of a simple framework (see Fig. 8.46).
If a framework of four hinged rods is subject to a load, it is deformed into
a parallelogram. A framework of three rods, however, is more stable.
• In the back section, additional ribs (drawn in red in Fig. 8.47) were posi-
tioned tangentially from the reticle fixing pot across the boring for the z
motor to the motor flange. An additional rib is positioned along the longi-
tudinal axis. Because of its shape, this structure is also called a w-structure.
Owing to lack of space, a transverse rib (drawn in blue in Fig. 8.47) was
added in the front section in order to improve stiffness.
• The motor flange was integrated into the reticle stage. This construction
has the advantage of reducing the total length (including the linear motor)
and, thus, the free bending length.
• Further design features are the stiffening ribs at the top of the box profiles.
These are fixed by adhesion after the construction process is finished.
• The borings for the z actuators were moved, so that the surrounding ring
was integrated as a supporting component (connected with the lateral
channels and the ring around the reticle fixing pot) into the reticle stage.

Profile with rectangular bores

100 ob j'jdbo7J1 0 tm

Closed box profile

1'- - - -- - -d---l- - -l 0
Fig. 8.45. Profile variants

Undeformed Deformed

DJ~ot'vJ
, r:or~
t jJO;ot V
I ' I~or+

Fig. 8.46. Deformation of framework structures


8.2 Optimization of a Reticle Stage for Optical Lithography 445

Integration 01
motor flange

Fig. 8.47. Optimized structure

Figure 8.47 shows the structure optimized by these approaches. The FEM
analysis of the optimized reticle stage yielded a bending eigenfrequency larger
by almost 30% compared to the prototype; see Table 8.5 and Fig. 8.48. At
the same time, the mass was reduced by 2.2 kg.

8.2.4 FEM Analysis With Additional Masses

Figure 8.49 shows the final optimized geometry with all additional masses:
• The reticle is in the center of the fixing pot and is fixed in its position by
a partial vacuum. The mass of the reticle is approximately 1 kg.

Table 8.5. Comparison between the prototype and the optimized reticle stage

Eigenfrequency (Hz) number Mass (kg)


1 2 3 4
Prototype (simulation) 466 707 1222 1249 14.0
Optimized structure 561 914 1137 1344 11.8
Changes in % +20 .6 +29.3 -6.9 +7.6 -15.7

Fig. 8.48. Bending eigenmode


around the x axis, h = 914 Hz (see
also movies on CD)
446 8. Glass Products Under Mechanical and Thermal Loads

Linear motor

z motor (idealized
as point mass)

Fig. 8.49. Optimized structure including reticle, linear motors, and z motors

• The two linear motors are positioned at the right end of the reticle stage.
The mass of a motor is approximately 4 kg.
• The z actuators are only modeled by point masses. The mass per z actuator
is approximately 0.5 kg.
The sum of the additional masses is approximately 10 kg, which is nearly
the same value as the mass of the reticle stage (11.8kg). Principally, this leads
to a distinct reduction of the relevant bending eigenfrequency. However, an
FEM analysis shows that, if all additional masses are accounted for, the
bending eigenfrequency of the optimized component is 512 Hz, which is still
well above the required 500 Hz (Fig. 8.50).
A comparison of the costs and the development times proved simula-
tion to have decisive advantages over the purely experimental approach. The
construction of just one prototype took approximately six months and cost
approximately DM 60 000. The entire FEM optimization with approximately
25 model variants, in contrast, required only two man-months. Additionally,
the models also provided useful information about
• the thermal behavior,
• the deformation due to dead weight,
• the design of the control circuit position.

Fig. 8.50. Bending eigenmode


around the x axis, h = 512 Hz
References 447

The reticle-stage geometry we described in this section has been patented


(DE 19756486 C1).

References

8.1 Schott Glas: "Electronic packaging - glass-to-metal seals", brochure (Lands-


hut 1998)
8.2 Test regulation "Robustness test plan for header subassembly", TP 0575,
Rev. A
8.3 F.E. Grubbs: "Procedures for detecting outlying observations in samples",
Technometrics 11, 1-25 (1969)
8.4 A.H. Jager: Statistik mit Mathematika - Methoden und Anwendungen
(Springer, Berlin, Heidelberg 1997)
8.5 V.I.C. Barnett: Outliers in Statistical Data, 3rd edn. (Wiley, Chichester, UK
1994)
8.6 E. Kreyszig: Statistische Methoden und ihre Anwendungen, 7th edn. (Van-
denhoek & Ruprecht, Gottingen 1988)
8.7 S. Brandt: Datenanalyse, 3rd edn. (BI Wissenschaftsverlag, Mannheim 1992)
9. Simulation and Test of the Spinning Process
Applied to Platinum Metals
Seth Nzahumunyurwa, Hans-Jurgen Hoffmann, Hans Oechsner,
Joachim Disam

9.1 Necessity to Shape Materials


Solid materials must necessarily be given a specified shape in order to be used
in technical applications. A major part of engineering science and technology
is related to the different shaping techniques. The most important criteria for
the choice of the respective technique are the necessary precision, expenditure
and waste, or economical aspects for short. Many different shaping techniques
are described in [9.1-4] with emphasis on metals because of their widespread
use; for the shaping of non-metallic materials the reader is referred to special
literature, such as [9.5] for glasses.
Roughly, one discriminates between the shaping in the viscous and in the
solid state of materials. As a consequence, one needs to know the viscosity-
strain and viscosity-temperature curves in the first case, and the stress-strain
curves in the second. Characteristic of the viscous state is the possibility of
deformation and shaping even under low forces or stresses, whereas in the
solid state deformation is achieved essentially by stresses exceeding the yield
stress. Thus, there is no linear relation between the deformation and the ap-
plied stresses in the solid state, which renders the description of deformation
of metallic solids more difficult. In this chapter, a special technique, the spin-
ning of metals of the platinum group, will be considered. Such metals are
widely used in the glass industry. The spinning process is quite economical
to produce pieces with rotational symmetry. However, it requires some ex-
perience and sometimes several trials, which might increase production costs
and cause waste of material. Because the production costs of metals of the
platinum group are high, one is interested in avoiding unsuccessful trials.
Therefore, a simulation strategy has been developed and practically verified
to reliably calculate successful spinning curves and to avoid failures, which
will be described in the subsequent sections. It will be demonstrated by the
shaping of crucibles starting from sheets of Pt metals.

9.2 Qualitative Description of the Spinning Process


The metal-forming process "spinning" is applied to thin circular sheets of
metals to produce, by a continuous bending and pressing process on a spin-

H. Loch et al. (eds.), Mathematical Simulation in Glass Technology


© Springer-Verlag Berlin Heidelberg 2002
450 9. Spinning Process Applied to Platinum Metals

ning lathe, hollow pieces with a rich variety of shapes with rotational symme-
try (see Fig. 9.1). A circular blank of sheet metal (1) is clamped by a pressing
device (9) to a rotating spinning chuck (3), which is mounted to a spindle (7).
The blank rotates with the spindle (7) and the pressing device (4), which is
supported by the tailstock (9), while a spinning roller (2) mounted on an x-z
support (8) is pressed to it and shifted back and forth along spinning curves
(usually in a plane passing through the axis of rotation), until its flange (10)
fits the spinning chuck (3). A series of possible spinning curves are also shown
in Fig. 9.1. Sometimes one uses a rotating counter-pressing plate (5). The rim
of the spun workpiece can be finished by a cutting device (6).
Several kinds of damage (e.g. , wrinkles, cracks, and shearing [9.6]) may
occur in the deformed blank if the spinning parameters are not adjusted cor-
rectly. This causes premature termination of the spinning process. To avoid
such damages, a general theory and strategy are required to compute and
simulate the optimum movements of the spinning roller for a designed piece
and given properties of the blank material. This is important for production
on demand, such as in the glass industry, where devices made of Pt are used
in contact with the glass melt [9.7]. For such premature materials, failures
and waste of material have to be avoided. To reach this goal, an optimization
concept for the calculation of spinning curves, which is based on the plasticity
theory of von Mises for plastic-rigid materials [9.1,8,9]' has been developed
and applied experimentally to Pt materials. Using the spinning curves com-
puted with the presented method, a steering program has been written and
applied to control the movement of a spinning machine without the need of
a practical learning process.

1 blank of sheet metal


2 spinning roller
3 spinning chuck
4 pressing device
5 counter-pressing plate
6 cutting device
7 spindle
8 x-y support
9 tailstock
10 flange of the workpiece

Fig. 9.1. Spinning lathe and its essential parts


9.3 Essential Assumptions for the Modeling of the Spinning Process 451

9.3 Essential Assumptions for the Modeling of the


Spinning Process
In the present chapter, we will focus on the optimization strategy to calculate
the spinning curves based on [9.10, ll], whereas the material aspects will be
published elsewhere. In a second step, the calculated spinning curves have
to be implemented practically using the steering program of the spinning
lathe. The spinning curves are rather general, whereas the steering program
generally allows different ways for the realization with respect to the rotation
speed, the speed of the movement of the roller support and the pressing
force of the roller, for example, as long as the parameters are compatible
with the specifications of the respective machine. This implementation of
the steering program requires a second step of optimization with additional
practical conditions and simplifications if necessary, which has to be done
separately.
The spinning roller is moved in a plane defined by the x-y support. Due
to the rotation of the blank, it generates a spiral track on the surface of the
blank with a shift per revolution or reference shift (feed), D.s. It is reasonable
to adjust D.s smaller than one half of the front radius, RL, of the spinning
roller. The normal shift (deformation of the blank normal to its surface) after
a single revolution, D. Uh, should remain smaller than one half of the thickness
of the blank, h. With D.uh too large, the risk to create wrinkles increases,
because high compressive stresses and strains arise in the blank. In order to
avoid local stress discontinuities, the calculated spinning curves should be
continuous. The mechanical properties of the blank material are assumed to
be isotropic and free of texture and stress before deformation.
For the following, the spiral track on the surface of the blank is replaced
by a sequence of concentric and contracting (perpendicularly to the surface
of the blank) circles, each of which represent:::; approximately the contact
area during one revolution. This approximation is allowed, if D.s is chosen
small enough. Then, the movement of this "deformation ring" corresponds to
the integral effect of the spinning roller for each revolution completed or to a
stepwise axially symmetrical deformation of the blank instead of a continuous
local deformation at the contact area of the spinning roller. This :::;implification
enables the optimization of the spinning curves using the presumed integral
principal strains due to the deformation ring of varying diameter instead of
the principal stresses. It certainly does not allow us, however, to compute and
simulate the actual local stresses due to the spinning roller, which might cause
buckling instabilities, and due to the torque at the border of the clamped
center area of the blank, which might cause shearing. This is avoided in
practice, however, if D.s is sufficiently small.
With these assumptions, practical solutions of the spinning curves will be
derived by a discrete model. The quantities and parameters of the discrete
model can be seen in Fig. 9.2: Because of the rotational symmetry, a cylindri-
cal coordinate system (r, 8, z) is introduced, with the z coordinate coinciding
452 9. Spinning Process Applied to Platinum Metals

r
,1\j=O
I
N VN j>0
N-1

Center plane

i
i-1

r'

3
i = 2 z'
1
o . _. _. ->z

Fig. 9.2. Discretization of a blank with rotational symmetry around the axis of
rotation z in the initial state j = 0 and in a later state j > O. For explanation of
the symbols see text

with the axis of rotation and the radius r representing the distance from that
axis. The origin of the z coordinate axis is located in the center of the middle
plane of the blank.
The successive spinning curves are numbered by the index j, where j = 0
is the original non-deformed blank. By convention, the spinning curves from
the axis of rotation towards the rim of the blank are designated positive (j +),
those spinning curves from the rim towards the center are negative Ci - ).
For the discretization, the blank is segmented into annular elements with
a rectangular cross section characterized by the index i. Before the deforma-
tion starts, the blank is assumed to be plane and all annuli have the same
width, RiO = £0, and the same thickness, hiO = ho . Their radii, rij, are taken
as the respective inner radius of their middle plane (r in Fig. 9.2). During
the deformation, the cross sections of all annular elements shall remain rect-
angular, but their dimensions and their angles of inclination with respect to
the z axis, aij, change.
The numbering of the annular elements i = 0 ... N starts at the rotational
axis and proceeds to the circumference. Thus, in the initial state

riO = i £0 and ri+l,O = ri,O + £0 = (i + 1)£0 . (9.1)


9.4 General Relations for the Model of the Spinning Process 453

Consequently, the volume of the first element is VO,j and that of the last one
VN,j' The volume Vi,j of each individual annular element remains constant
during the whole spinning process, i.e.,

Vi,j = Vi,o . (9.2)

The angle of inclination of different annular elements, aij, changes during


the spinning process. However, because of the continuity two consecutive
elements always have their common contact circle in their middle plane. This
corresponds to the following contact conditions

rHl,j = ri,j + £i,j sin ai,j and Zi+1,j = Zi,j + £i,j cos ai,j , (9.3)

wherein £ij and a;j are the width and the inclination angle of the cross section
of the ith annular element of the jth spinning curve.

9.4 General Relations for the Model of the


Spinning Process
Each annular element in Fig. 9.2 is deformed in three dimensions, which are
characterized by the unit vectors ee, ee, and eh along the axis of the or-
thogonal curvilinear coordinate system 8, £, and h, also shown in Fig. 9.2.
Because the deformations are rather large, the principal true strains are de-
fined by 'Pp = In(sp/so), wherein sp and So are the lengths after and before
the deformation in any of the orthogonal directions. For symmetry reasons
each principal strain component points in the direction of the relevant unit
vector of the coordinate system. The first invariant of the strain tensor of a
plastic, isotropic incompressible material is

h = 'PEe) + 'Pe + 'Ph = 0 , (9.4)

which corresponds to the conservation of the volume of each annular element i


of the spinning curve j, or

In (2r i ,j + £i,j sin ai,j) + In (£;,j) + In (hi,j) = 0 . (9.5)


2ri,0 + £;,0 £;,0 hi,o
Equation (9.5) is the logarithm of the ratio of the volumes before and after
the deformation of the annular elements. These volumes are calculated by
applying the first of Guldin's rules, for example,

I t( ,Vi]') - I (27f (r;,j + ei,jS~"'i,j) £t,]'h t,]') - 0


n Vi,o - n 27f (r;,o + Ci;t) £i,O hi,o - ,

wherein £;,oh;,o and £i,j hi,j are the generating cross sections of the rotating ar-
eas, and 27f [ri,o + (£;,0/2)] and 27f [ri,j + (£i,j sinai,j/2)] are the pathlengths
454 9. Spinning Process Applied to Platinum Metals

of the center of gravity of the generating cross sections during one revolution.
It seems that there are four parameters, ri,j ,£i,j ,sin ai,j, and hi,j, to fulfil
(9.5) and to optimize the spinning curves. However, the number of degrees
of freedom is further reduced by additional conditions. For this purpose, we
consider the ratio of the principal strains
'Ps
'/= - , (9.6)
'Pc
which characterizes the maximum plastic flow under simultaneous deforma-
tion strains into the directions es and ec. (The indices i and j have been
omitted in (9.6) and in some of the following equations.) The practical sig-
nificance of,/ is seen from Fig. 9.3, which shows schematically the forming
limit diagram of an isotropic material [9.3]. The limiting forming curve gives
the onset of necking in sheet metal. Three different modi of deformation with
their '/ are sketched in Fig. 9.3. For the simple tensile test '/ = 0.5. One
can see that the maximum deformation without necking is rather low. In the
ideal stretching test one has '/ = 1. One can see that a larger deformation is
possible, before necking occurs. The largest deformation without necking is
possible in the case '/ = -1, corresponding to a tension into ec and a simul-
taneous compression into es of the same amount. One might assume that
'/ < -1 would be even more favorable to utilize the material for deformation
without failure. However, in this case one is at risk of too large compres-
sive stresses causing circumferential buckling instabilities and wrinkles in the
flange of the blank. Therefore, one should aim at '/ ;:::0 -1, which reduces the
number of degrees of freedom in (9.5). Consequently, we have

'Ps = '/'Pc and 'Ph = -('Ps + 'Pc) = -(1 + '/)'Pc . (9.7)


Now we take advantage of the second invariant of the strain tensor [9.1]

I
2 = 3"2 (2 2 2) 2
'Ps + 'Pc + 'Ph = 'Pv , (9.8)

~I

,,
,
" ,
j + 1',
"/

, /
/
/

j ..... ,
, // Y = 1
Y= -1 j-1,', // Ideal
Ideal '" \ // stretch test
deep drawing j = 2', \ \ , /
/

j = 1.....,\
/

/
/

--L-~---L---L--~-7~~---L--~--~~e
,
j=O

Fig. 9.3. Typical forming limit curve (solid curve) in a forming diagram cp£
f(cpe) according to [9.3]
9.5 Approximations 455

with the equivalent strain 'Pv. Inserting (9.7) into (9.8) yields
3
'P~ + ,'PE + (1 + ,)2'P~ = "2'P~ (9.9)

Replacing 'Pc = In B one obtains from (9.9)

B = Bi j
,
= exp (J3
2 )1
'Pv
+, +,2
), (9.10)

(wherein 'Pv = 'Pvi,j and, = ,i,j may depend also on and j) and the
geometric parameters of the annular elements from (9.7)
fl . . -
~',J -
fl.
~"
oB··
',J'
h··
',J
= h·" oB-CH,) and
(9.11)
ri ,J + -£i,j.
2 SIn 0i 'J'
_ (
- ri ,0 + -£i,O.
2 sm 0i '0
) I
B,
"1
J' .

£i,j and hi,j depend on 'Pv and, (implicitly through Bi,j), whereas ri,j de-
pends on £i,j and Oi,j in addition. The dependence of ri,j on £i,j can be
eliminated by the contact condition (9.3), which can be rewritten as

£HI,O) B,i+l,j £HI,O " B


( rHI,O + -2- HI,j - -2- smoHI,j HI,j
(9.12)

This equation can be simplified further with the constant initial width of the
annular elements £i,O = £0 according to (9.1) into

. 0HI,j B i+I,j + sm
sm . Oi,j B i,j = 2 (.Z + "2
3) B HI,j ,i
+1 ,J - 2 (.Z + "2
1) B,i,j
i,j' (913)
.

Equation (9.13) is the general condition to be fulfilled for the spinning curves
in the case of equal width of the annuli. Nevertheless, it is not possible to
calculate optimum spinning curves without further assumptions on Bi,j or
'Pvi,j and,. One can assume that 'Pvi,j and, are monotonous functions of i
and j, However, , will be given a constant value for a given spinning curve,
because its possible variation is rather small, and 'Pvi,j can be represented as
a power series, of which we use for the first approximation the first (constant)
term, only, and for the second approximation both the first and the second
terms. Both cases will be treated in the following sections.

9.5 Approximations
9.5.1 First Approximation: Quasi-Homogeneous Deformation

In the simplest case, 'Pvi,j and, are assumed to be constant for all annular
elements i of a given spinning curve, i.e.,
456 9. Spinning Process Applied to Platinum Metals

CPvi,j = CPv,j, 'Yi,j = 'Yj and Bi,j = Bj . (9.14)

With these relations, (9.13) is simplified to


.
sm ai+l,j + sm
.
ai,j =
2B'Yj-l
j with - 1 ~ 'Yj . (9.15)

From this equation, one can determine recursively the angles of inclination of
the annular elements. From two consecutive equations, however, one obtains
sin aH2,j = 2BT- 1 -sin aHl,j = sin ai,j, with the result that for all elements
at most only two different angles of inclination occur, which depend on the
index i being even or odd. In general, both angles are different, which is
not useful, because the spinning curve is not monotonous in this case. Only
for a suitable choice of the starting angle does one obtain the same angle of
inclination ai,j = aj for all i. Then, the spinning curves are straight lines.
From (9.15) one obtains

(v'3
V1 +'Yj+'Y]
Sinai,j = BT- 1 = exp (-yj - 1) CPv,j ) (9.16)
2

With (9.16) one obtains from (9.11)

(9.17)

and the principal strain


To
= In -.!..!l. = 'Yj In B j
0

CPei,j (9.18)
Ti,O

for all i. Using (9.7) and (9.11) one obtains the relations

.
Smaij=--=
To'l,J fo
0 (fo'l,J
-
o)'Yj - 1
=
(TO't,l
-
o)'Y~~l =-
(ho1,,) o)~
, Ti,O fi,j fo Ti,O ho

=exp (_v'3 (-yo -l)cpv


(9.19)

V
J ,J.
0 )

2 1 +'Yj+'Y]

Because sin ai,j = sin aj is valid for all i, the other quantities in (9.19)
must also be independent of i, such as hi,j = h j and fi,j = f j . Thus, the
thickness of the spun range (or flange) of the blank is independent of the
position of the annular element. Therefore, this approximation causes a quasi-
homogeneous deformation of the blank. However, to spin large blanks one
needs large stresses near the rim of the blank, which can cause bending and
buckling instabilities. One might assume that the bending can be avoided
by a counter-pressing device. In this case, however, one also reduces the
compressive stresses needed to reduce the diameter of the blank. Thus, the
9.5 Approximations 457

conditions of this approximation are obviously less favorable for blanks of


large diameter.
In this respect, it might be interesting to consider the case of ideal deep
drawing with 'Y = -1. According to (9.19)

.
Sinaij
ri,j Ro R
= - - = ( -.!..:2
.)-2 and
, ri,O Ri,j Ro (9.20)
hi,j = ho .

Hence, the thickness of the flange ho remains constant during the whole spin-
ning process, which is controlled simply by increasing l{Jv,j step by step with
the number of spinning curves j and by maintaining the value of 'Y constant
during the whole spinning program. The spinning process is continued until
the spun material begins to constrict (forming the limit curve of Fig. 9.3) or
the flange fits the spinning chuck. Attention has to be paid to a perfect tran-
sition of the positive and negative spinning curves one into another on the
spinning chuck and at the rim of the blank. This will necessarily cause dif-
ferent values of 'Y- for negative spinning curves j_ and for positive spinning
curves j+.
An example of calculated spinning curves based on this requirement is
shown in Fig. 9.4. Two positive spinning curves j+ and (j + 1)+, of which
the angles of inclination have been computed using 'Y = 'Y+ = -1, are shown
together with a negative spinning curve j_ between them, touching j+ at its
outer border and (j + 1)+ at the spinning chuck. From there, the angle of incli-
nation aj+ of j_ is between aj_ and a(j+1)_, leading necessarily to a different
value of 'Y-. In this case, it is possible that 'Y- < -1 and h_ > ho, increasing

L,
r

Fig. 9.4. A negative spinning curve j_ between two positive spinning curves j+
and (j + 1)+ for the first approximation. The arrows show the direction of shifting
the spinning roller
458 9. Spinning Process Applied to Platinum Metals

the tendency of wrinkle creation in the flange of the blank, if the original
blank is thin and large. If this happens, the first approximation cannot be
used, and an extended approximation will be needed. The dashed curves in
Fig. 9.4 show possible intermediate states of a flange during transition from
j+ to j _ or j_ to (j + 1)+, respectively.
A numerical example showing the behavior of the main equations (9.15)
and (9.16) with and without application of the optimization criteria will be
performed in Sect. 9.6. From experimental tests, however, a practical lim-
itation of the 1st approximation is given for large Pt blanks, although for
small blanks it has been applied reliably and with high accuracy. With the
studied Pt materials the limiting diameter was about 190 mm for a thickness
of h ;::::; 1.0 mm. For larger diameters, two failures have been observed. First,
owing to large shear stresses, the clamped area was sheared at its border (see
Fig. 9.5a), if the diameter of the clamped area was too small (in our case
smaller than 40mm). Second, even if the clamped area was not too small,
rosette-shaped wrinkles evolved in the flange of the blank (see Fig. 9.5b) due
to compressive stresses in the (9 direction and simultaneously too low ten-
sile stresses in the £ direction. This shape of wrinkles has also been reported
for the deep drawing process [9.12,13]. To overcome these disadvantages, the
following extended approximation has been used.

9.5.2 Second ApproxiInation: Linearly Decreasing Deformation

If the equivalent strain is constant for a spinning curve, the bending moments
and the stresses necessary for the strains become larger with increasing radius,
which may be the origin of instabilities and failures. Thus, it is favorable to
reduce the strain and consequently the stresses during the spinning with
increasing radius. Considering the equivalent strain 'Pv,j as a Taylor series,
the first approximation corresponds to the first constant term, whereas the

a)

Fig. 9.5. Typical damage in a spinning process using the first approximation:
(a) spun disc of PtRh10 sheared at (1) adjusting 'Pv ::::; 0.3 (2rN = 190mm and
ho = 1.0 mm); (b) rosette-shaped wrinkles in a spun disc of Pt (2rN = 298 mm and
ho = 0.7mm)
9.5 Approximations 459

second approximation includes a second term, which depends linearly on the


radius from the rotational axis of the blank. Assuming equidistant widths of
the annular elements one defines

lPvi,j = lPvk,j + D.lPv,j ( Ni -_ kk ) for k::; i ::; N , (9.21 )

wherein D.lPv,j < 0; k and N are the indices of the annular elements adjacent
to the spinning chuck and at the outer border of the blank, respectively
(Fig. 9.2). The ratios of the strains (9.6) will still remain constant for all
rings of a given spinning curve j
lPei,j
li,j = -- = Ij
lPRi,j

Inserting (9.21) into the substitution (9.10) one can factorize

B i,j i-k
= B j,l B j,2 (9.22)

with

and Tlj,' ~ exp (_V;_3 -Vr:.1=:~:=~=jk=+='=J)


(9.23)

which are constant for a given j.


The data of the spinning curves are obtained by inserting (9.22) and (9.23)
into the general equation (9.13). One arrives at the relation [9.10,11]

.
smaHl,j
B
j,2
.
+ Smai,j = 2B'":1'J-
),1
1 B(i-k)(-1'J-1)
),2
(9.24)
X ((B],j2 -1) (i - k) + ~ (Bj ,2 + 1)).

Using (9.24) one can calculate the parameters of the optimum spinning curves
for the second approximation. First, it is shown that (9.24) yields, in the
general case, a sequence of elements with alternating slopes. This is done by
expressing sin ai+2,j as a function of sin ai+l,j, which is subsequently replaced
by the similar equivalent form from (9.24). Collecting terms yields

.
sm aH2,j -
- B- 2 .
j,2 sm ai,j + 2Blj-1 (i-k)(lj-1)
j,l B j,2
(Blj
j,2 -
1) B-j,22
(9.25)
X ((B],j2 -1) (i - k) + B],j2 + ~ (B j ,2 + 1)) .

Thus, the slopes of the annular elements with even and odd indices i may be
alternating, which contradicts the requirement for smooth and monotonous
460 9. Spinning Process Applied to Platinum Metals

spinning curves. In order to obtain monotonous solutions we have to assume


sin aHl,j = sin ai,j, similar to the procedure of the first approximation, and
obtain

sin ai ,J. = B-:Yi - 1 B(i-k)("'Ij-l)


J,1 J,2
(1 + 2 (B'Yj
J,2
- 1) (B
J,
2 + 1)-1 (i - k)) . (9.26)
This result is the general optimized solution for the 2nd approximation fulfill-
ing the requirements. In contrast to the 1st approximation (9.16), it depends
explicitly on the index i of the annular element and consequently on its ra-
dius rio However, in general, D..ipv,j < 0 with the consequence 0 < B j ,2 < 1
for technical reasons mentioned above. Because, for increasing radius, (i - k)
can be very large, one must choose D..ipv,j in such a way that 0 ::; sin ai,j ::; 1.
With (9.26) and (9.11) one determines the radius

(9.27)

Equation (9.27) is obviously more complicated than (9.17) from the first ap-
proximation, as it contains two factors that depend explicitly on the number i
of the annular element. Now the spinning curves are no longer straight lines,
except for either Bj2 = 1 (which corresponds to the 1st approximation) or
just for the single element when i = k. Especially in the case of ideal deep
drawing with I = -1, the principal strains are

and the parameter

The angles of inclination are obtained from

(9.28)

According to the conditions, the spinning curves are bent. Nevertheless, the
thickness of the whole flange remains constant in this caSe.

9.6 A Practical Example for the


First and Second Approximations

Applying the results from Sects. 9.4 and 9.5, We show as a practical example
the results for spinning a circular blank with a radius r N = 100 mm and a
9.6 A Practical Example for the First and Second Approximations 461

thickness ho = 1.0 mm to form a crucible with constant (independent of the


axial coordinate z) diameter 2rc = 2 x 25.4 mm. An equidistant width of
the initial annular elements £0 = 0.2 mm has been chosen, corresponding to a
maximum number N = r N/£0 = 500. The maximum index io of the clamped
center area is io = 127; consequently k > 127 for all spinning curves. For the
following calculations, we have used:

k = 200, 'Pvk = 0.6, D.'Pvk = -O.l'Pvk, and 1 = 0.8 . (9.29)

With these data, B j ,1 and B j ,2 have been calculated according to (9.23) as a


function of 1 (Fig. 9.6). The denominator of the argument in the exponential,
which is quadratic in I, has a minimum for 1 = -0.5. Consequently, B j ,1
is maximum and B j ,2 is minimum at 1 = -0.5, because 'Pvk,j > 0 and
D.'Pv,j < O. In addition, B j ,2 :::::; I, because the argument of the exponential is
very small, compared to unity, and negative.
The behavior of the iteration equation (9.15) of the first approximation,
for which D.'Pvk = 0, is represented in Fig. 9.7. The first six values of sin Goj
are plotted as a function of the number i ?: k of the annular elements using
1 = -0.8 for different starting angles Gok. It is confirmed that this general
series is indeed alternating. A monotonous spinning curve is achieved by
using the optimum starting angle Gok,opt = 21.13°, only. Then, one obtains
a straight spinning curve with a constant angle of inclination Goi = Gok,opt.
This angle fulfils exactly the optimized equation (9.16). Since the sequence
(9.24) also alternates generally, its plot (not shown here) looks similar to the
curves of Fig. 9.7, whereas the optimized solution given by (9.26) is a smooth
monotonous curve.
Using the same starting parameters the sequences (9.15) of the 1st ap-
proximation and (9.24) of the 2nd approximation are compared with their

r--...,----__,_-__,_--~-~-__,_-__,_-_._ 0.99995
1.8

0.99990
1.6

0.99985 riJ"i

0.99980
1.2 . 'Pvk = 0.6
!:1'Pv =- 0.06
k =200, N =500
1.0+---i---+--r----i---i----r--t---+ 0.99975
-2.5 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5
Strain ratio y

Fig. 9.6. Bj,l and B j ,2 as a function of "( for '{!vk = 0.6 and f1,{!v = -0.06
462 9. Spinning Process Applied to Platinum Metals

1.0

0.8

0.6

~- 0.4
c:
'iii
0.2

0.0

-0.2

200 201 202 203 204 205


Number i of the ring element

Fig. 9.7. sinai (sine of the angle of inclination ai) from the sequence (9.151) as a
function of the ring number i for different starting angles ai

respective optimized sequences (9.16) and (9.26). The corresponding results


are shown in Fig. 9.8, where the upper and lower data points result from
the even and odd indices i - k of a non-optimum starting angle of 40°. The
data points from the 1st approximation (9.15) do not depend on i, whereas
the ones from the 2nd approximation (9.24) are increasing with i. The re-
sults from the optimized relations are shown in the center of Fig. 9.8. In
this case, alternation between the even and odd data points is absent, and

0.8 .---------:-------:--------~
a N = 46·
a k~=~4~02:·~-~~~---t:::_-- ;-k even
0.6

: 2nd approximation
t:l a N Opt = 24.6·
c: 0.4
'iii a '<l .Eel
~~2~ 1 . 1~·~===========­
1 st approximation: IIOet = 21 .1
0.2
a N- l = 6.8'
a k+l =4.5· ;-k odd
U t =4.5"
0.0 """'-----.-- - -- - r - -- - - r -- - -, - ------l
200 300 400 500 600
k N

Fig. 9.S. Comparison between the non-optimized sequences (9.15) and (9.24) and
the optimum sequences (9.16) and (9.26) of the 1st and 2nd approximation (pa-
rameters: , = -0.8, 'Pvk = 0.6, 6'Pv = -0.06)
9.6 A Practical Example for the First and Second Approximations 463

all data points fit a single monotonous and smooth curve according to the
expectations, horizontal for the 1st and rising for the 2nd approximation.
To show in some detail how the simulated spinning curves change during
the spinning process, three successive spinning curves j - 1, j, and j + 1 have
been computed and plotted in Fig. 9.9. The equivalent strains are ipvk,j-l =
0.4 for j - 1, ipvk,j = 0.6 for j, and ipvk,j+1 = 0.8 for j + 1. The starting
indices k j - 1 , k j and k j + 1 have been calculated using (9.17) and (9.27) with
rc = 25.4mm and the parameters (9.29). The dotted curves correspond to
the 1st approximation, where D..ipv,j = 0, whereas the dashed and the solid
ones belong to the 2nd approximation with D..ipv,j =I- O. It is clearly seen
that the spinning curves show a strong deviation from the 1st approximation
with decreasing D..ipv,j' As has been mentioned already, the risk of wrinkling
decreases for large blanks with decreasing D..ipv,j < O. Therefore, D..ipv,j is the
parameter of choice to minimize the torques near the rim of the blank and
the shear stresses at the border of the clamped center.
For simplicity, the front radius of the spinning roller RL has been assumed
to be very small, causing sharp transitions at i = k in Fig. 9.9. However,
this has no effect on the generality of the results. All important data of the
calculated spinning curves for D..ipv,j = -0.3ipv,j are summarized in Table 9.1.
Accordingly, the angle of inclination aN in the outer region is always larger
than ak at the spinning chuck, a behavior typical of the spinning curves
calculated according to the 2nd approximation.

j" 0
100 y" - 0.8
IPYk, j - 1 ,, 0.4
IPYk, i ,, 0.6 j- 1
80
IPYk, j + 1 " 0.8
j+1
E 60

40

r I-..L...""""",,- - - l{Jy" - 0.3I{Jy (2 nd approx.)


20 C il'Yk, j _ 1 il'vk, j + 1
- - - - D.I{Jy ,,- 0.1l{Jy (2ndapprox.)
6I{Jy" 0.0 (l S \ approx.)
0 k-4----r--~----r---.---_r--~----~
o 20 40 60 80 100 120 140
Zj/mm

Fig. 9.9. Three spinning curves (j - 1), j, and (j + 1) on a spinning chuck with
constant radius rc for different values of t::"cpv
464 9. Spinning Process Applied to Platinum Metals

Table 9.1. Calculated values of the spinning curves for 6<pv = -0.3<pv

Spinning curve , <pv rk


mm
k Ctk CtN

j-1 -0.8 0.4 25.41 172 30.4° 42.6°


j -0.8 0.6 25.41 200 21.2° 33.8°
j+1 -0.8 0.8 25.41 233 14.9° 27.1 °

9.7 Experimental Observations and Discussion

The optimization procedure of the spinning process has been applied to Pt


materials. The simplest 1st approximation has been successfully applied to
small blanks with a radius TN < 100 mm, only. Otherwise rosette-shaped
wrinkles or shear occurred, leading to a premature termination of the spinning
process. Nevertheless, the advantage of the 1st approximation was that the
thickness of the flange remained quasi-constant during the whole spinning
process, yielding workpieces that delivered samples for further investigations
of the material properties after spinning [9.10].
It has been observed experimentally that PtRhlO shears easier than Pt,
Ptlr1 or ODS-Pt. For this reason, it was interesting to compare the onset
of shear in PtRh10 for the 1st and 2nd approximations. The onset of shear
is illustrated in Fig. 9.5a for the 1st approximation. Using the second ap-
proximation, a blank of the same quality was spun adjusting the following
parameters: b.<pv = -0.5<p1J and Ij = - 0.6. As seen in Fig. 9.10a, shearing
occurs also using the 2nd approximation, but at a later stage. For the 1st
approximation the shear was observed at an angle C\'.k = 34° , whereas for the

Fig. 9.10. Two discs spun on a spinning chuck to form a crucible with constant
radius rc = 22.5 mm using the 2nd approximation (initial diameter and thickness
of the blank: 2rN = 190mm and ho = 1.0mm). (a) Sheared disc of PtRhl0 (1)
spun using I:!.<pv = -0.5<pv and '] = -0.6. (b) Disc of ODS-Pt with an overspread
of the flange (2) using I:!.<pv = -0.5<pv and '] = -0.8
References 465

2nd approximation an angle of CXk = 25° has been achieved. The thickness of
the flange was quite uniform, hflange :::::0 0.7mm over a large range. However,
it increased from 0.7 mm to hrim = 1.1 mm within the last 15 mm near the
rim of the flange.
The detrimental shear is considerably reduced, however, by applying the
2nd approximation to calculate the spinning curves. Using the same blank
geometry and spinning parameters, other materials, such as ODS-Pt, have
been fully spun, without damage, to fabricate cylindrical crucibles of a total
length of 220mm with an inner radius in the range of 22.5-25mm and a
maximum equivalent strain of CPv :::::0 1.8 at the rim. It is worth mentioning
that the cause of shear can be considerably reduced further by increasing the
global number of the spinning curves, particularly for the 2nd approximation.
A further significant result of the experiments is that nO wrinkles occurred
On applying the 2nd approximation. Thin and large blanks with a thickness
of ho = 0.5 mm and a radius of up to r N = 200 mm have been spun On a
small spinning chuck with rc ::; 25 mm, but in all cases nO wrinkles occurred.
Changing the spinning parameter Ij from -0.6 to -0.8, One observes an
overspreading of the flange at its rim as seen in Fig. 9.lOb for ODS-Pt. With
the spinning process progressing, more and more material is overspreading,
until the process has to be terminated prematurely. In order to obtain a
spinning curve as long as the flange of the blank, 6.cpv = !(r) must be
correctly chosen during the whole spinning process. Because this requires
calculation of the stresses in the blank, it is generally easier to determine
6.cpv as a function of 1 experimentally. For Pt materials the combination
6.cpv = -0.5cpvk and 1 = -0.6 has been found to be optimal. With these
parameters almost all kinds of known damages could be avoided, except the
shear at the border of the clamped center area of large blanks in some rare
cases. Obviously, the onset of shear at the border of the clamped area depends
On the spun material.
In summary, failure is in general only avoided if the global parameter 6.cpv
is correctly chosen as a function of f. For different materials, the suitable
combination of 1 and 6.cpv has to be determined experimentally. This is
necessary because the present approach does not take into account the local
stresses explicitly.

References

9.1 K. Lange (Ed.): Umformtechnik, Handbuch fur Industrie und Wissenschaft,


Vol. I, Grundlagen, 2nd edn. (Springer, Berlin, Heidelberg 1984)
9.2 K. Lange (Ed.): Umformtechnik, Handbuch fur Industrie und Wissenschaft,
Vol. 2, Massivumformung, 2nd edn. (Springer, Berlin, Heidelberg 1988)
9.3 K. Lange (Ed.): Umformtechnik, Handbuch fur Industrie und Wissenschaft,
Vol. 3, Blechbearbeitung, 2nd edn. (Springer, Berlin, Heidelberg 1990)
466 9. Spinning Process Applied to Platinum Metals

9.4 K. Lange (Ed.): Umformtechnik, Handbuch fur Industrie und Wissenschaft,


Vo!' 4, Sonderverfahren, ProzejJsimulation, Werkzeugtechnik, Produktion,
2nd edn. (Springer, Berlin, Heidelberg 1993)
9.5 W. Giegerich, W. Trier (Eds): Glass Machines (Springer, Berlin, Heidel-
berg 1969)
9.6 H. Dierig, M. Kleiner: "Tendenzen beim Drucken mit CNC und AC", Bander,
Bleche, Rohre 29, 18-23 (1988)
9.7 Edelmetall Taschenbuch der Degussa AG, 2nd edn. (Huthig, Heidelberg 1995)
9.8 R. v. Mises: "Mechanik der plastischen fest en Korper im plastischen de-
formablen Zustand", Gottinger Nachr., Mathem.-Phys. K!. 1913, 582-492
(1913)
9.9 R. v. Mises: "Mechanik der plastischen Formanderung von Kristallen",
Z. Angew. Math. Mech. 8, 161-185 (1928)
9.10 S. Nzahumunyurwa: Optimierungskonzepte und materialwissenschaftliche
Untersuchungen zum Umformen von Pt- Werkstoffen durch Drucken, PhD
Thesis, University of Kaiserslautern, Department of Physics (GCA-Verlag,
Herdecke 1999)
9.11 S. Nzahumunyurwa, H.-J. Hoffmann, H. Oechsner, J. Disam: "Optimization
strategy of the spinning process", Mater. Sci. Eng. Techno!. 32, 827-836
(2001)
9.12 M. Meier: "Die Faltenbildung im Flansch von rotations-symmetrischen Tief-
ziehteilen, Institut fUr Umformtechnik" (Schriftenreihe 1) ETH Zurich (1986)
9.13 T.X. Yu, W. Johnson: "The buckling of annular plates in relation to the
deep-drawing process", Int. J. Mechanical Science 24, 175-188 (1982)
List of Contributors

Ton Backx Henry Eisermann


IPOS Technology BV, Schott Glas, Mainz 1
Bosscheweg 145a, E-mail: henry.eisermann
5282 WV Boxtel, @schott.com
The Netherlands
E-mail: ton.backx@ipos.nl Ludwig Frank
Schott Glas
Ruud G.C. Beerkens Christoph-Dorner-Str. 29
LG. Philips Displays Nederland BV, 84028 Landshut,
Zwaanstraat 1, Building TZ, Germany
P.O. Box 807, E-mail: ludwig.frank@schott.com
5600 AN Eindhoven,
The Netherlands, Ulrich Fotheringham
E-mail: ruud.beerkens Schott Glas, Mainz 1
@lgphilips-displays.com E-mail: ulrich.fotheringham
@schott.com
Andreas Bensberg
Schott Glas, Mainz 1 Christoph Hermanns
E-mail: andreas.bensberg Schott Glas, Mainz 1
@schott.com E-mail: christoph.hermanns
@schott.com
Matthias Brinkmann
Schott Glas, Mainz 1 Hans-J urgen Hoffmann
E-mail: matthias.brinkmann Institut fur Werkstoffwissenschaften
@schott.com und -technologien: Glaswerkstoffe,
Technische Universitiit,
Joachim Disam Englische Strasse 20,
Schott Glas, Mainz 1 10587 Berlin,
E-mail: joachim.disam Germany
@schott.com E-mail: hoffmann.glas@tu-berlin.de

1 HattenbergstraBe 10, 55122 Mainz, Germany


468 List of Contributors

Kurt Holt berg Horst Loch


Schott Glas, Mainz 1 Schott Glas, Mainz 1
E-mail: kurt.holtberg@schott.com E-mail: horst.loch@schott.com

Thoralf Johansson Sandro Manservisi


Schott Glas, Mainz 1 LIN-Dienca,
E-mail: thoralf.johansson University of Bologna,
@schott.com Via dei Colli 16,
40136 Bologna,
Thomas Korb Italy
Schott Glas, Mainz 1 E-mail: sandro.manservisi
E-mail: thomas.korb@schott.com @maiLing.unibo.it

Harald Kriimmel Kurt N attermann


Schott Glas, Mainz 1 Schott Glas, Mainz 1
E-mail: harald.kruemmel E-mail: kurt.nattermann
@schott.com @schott.com

Wolf Kuhn Seth N zahumunyurwa


Stein Heurtey, Institut fur Werkstoffwissenschaften
PB 69, und -technologien: Glaswerkstoffe,
91002 Evry Cedex, Technische Universitiit,
France Englische Strasse 20,
E-mail: wskuhn@stein-heurtey.fr 10587 Berlin,
Germany
Clemens Kunisch E-mail: seth.n@tu-berlin.de
Schott Glas, Mainz 1
E-mail: clemens.kunisch Hans Oechsner
@schott.com U niversitiit Kaiserslautern,
Fachbereich Physik,
Ulrich Lange Erwin-Schrodinger-StraBe,
Schott Glas, Mainz 1 Gebiiude 46,
E-mail: ulrich.lange@schott.com 67653 Kaiserslautern,
Germany
Frank- Thomas Lentes E-mail: hoe@rhrk.uni-kl.de
Schott Glas, Mainz 1
E-mail: frank-thomas.lentes
@schott.com

1 HattenbergstraBe 10, 55122 Mainz, Germany


List of Contributors 469

Volker Seibert
Schott Glas, Mainz 1
E-mail: volker.seibert
@schott.com

Norbert Siedow
Fraunhofer-Institut fUr Techno- und
Wirtschaftsmathematik (ITWM),
Gottlie b-Daimler-StraBe,
Gebiiude 49,
67663 Kaiserslautern,
Germany
E-mail: siedow@itwm.fhg.de

Peter Thomas
Schott Glas, Mainz 1
E-mail: peter.thomas@schott.com

Gunter Weidmann
Schott Glas, Mainz 1
E-mail: guenther. weidmann
@schott.com

Translations by:
Translating Agency,
Petra Schulz,
Dipl.-Translator,
Januarius-Zick-StraBe 73,
54296 Trier,
Germany
Phone: +49-(0)651-7103855
E-mail: schulz.petra@gmx.de

1 HattenbergstraBe 10, 55122 Mainz, Germany


Sources of Figures and Tables

We are indebted to the following editors and authors, respectively, for the kind
permission to reproduce copyrighted materials.

Material Source Original Publisher

Fig. 2.21 [2.96] Deutsche Glastechnische Gesellschaft,


Fig. 2.23 [2.97] Mendelssohnstr. 75-77, 60325 Frankfurt, Germany

Fig. 2.26 [2.107] Deutsche Glastechnische Gesellschaft,


Mendelssohnstr. 75-77, 60325 Frankfurt, Germany

Fig. 2.33 [2.149] Glass Service, Inc.,


Rokytnice 60, 75501 Vsetin, Czech Republic
Index

absorption coefficient 263,289 blow molding 293


absorption thermometry 262 boundary conditions 31
accuracy of the computation 170 boundary layer 313
active thermal conductivity 290 boundary-element method (BEM)
adaptive mesh refining 389 385
airbag igniters 413 breaking strength 408
all-electric furnace 41 brittle solid 408
approximation bubble ascension 52
- Eddington-Barbier 268 bubble release 87
- - Improved 278 buckling 454
- Rosseland 268
axisymmetric TV screen 321
calibration of glass tubes 349
batch casting process 327
- blanket 31, 56, 90 chill ripples 326
- modeling 110 closed-blanket melting 98
- - validation 123 cold-top electric melter 96
- morphology 87 combustion reactions 35
batch model 74 combustion space 32
- Amse and Oda 113 compaction of granular materials 82
- Austin and Bourne 112 computational fluid dynamics (CFD)
- Beerkens 120 17
- Carvalho and Nogueira 118 computer-algebra 2
- de Waal 117 continuous charging 90
- Faber, Beerkens, and de Waal 118 continuum approach 95
- Fuhrmann 111
continuum equations 6
- Hilbig 115
controlled variables (CVs) 139
- Hilbig and KirmBe 115
- Hrma 113 convection flows 23
- Kawachi and Kawase 121 correctness 5
- Leyens 112 corrosion rate 69
- Lixin and Chengxu 119 crack
- Prasad et al. 121 - elongation 402
- Pugh 110 - mode 408
- Schill 2D 11 7 - opening 410
- Schill 3D 119 - path 411
- Trier and Voss 111 - tip 400
- Ungan and Viskanta 116 critical process conditions in glass
- Wang et al. 121 melting 21
- Wu and Viskanta 115 crown temperatures 67
- Yamamura et al. 120 cut elongation 396
Biot number 109 cutting of glass sheets 381
474 Index

Damki:ihler number 106 asymptotics 212


Danner process 304 finite-dimensional 229
defect source tracking 22 general 210
deformation - thickness optimization
- of cubes 177 -- 2D 215
- of cylinders 183 - - numerical tests 220
- of spheres 186 float process 293
- statistics 187 flow
depleted layer 65 - characteristics 27
direct heating with frequencies in the - fields 169
MHz range (Eden) 126 - of glass melts 8
discontinuous charging 90 - patterns 20
dispersed blanket 99 - ripples 326
dispersion statistics 177 foam formation 22,70
dissociation 38 forward simulation 3
dissociation correction function 38 Fourier number 109
dissolved gases 56 fracture probability 425
disturbance variables (DVs) 139,141 fracture toughness 409
DLC 600 machine 405 free sheet 397
downdraw process 293 Fresnel reflection 247,264
dynamic behavior 29 full-body cut 381
dynamic modeling 71
glass pipe flows 193
- model
eigenfrequency 442,446
coarse 195
electric boosting 41
paradox result 197
electric heating 193
refined 195
- of platinum tubes 209
temperature profiles 196
electrodynamics 126
velocity profiles 196
emissivity 295
- multi-stage heating 200
energy balance 93
- single-stage heating 198
energy conservation 101
- - critical final temperature 200
equilibrium concentrations 77
- - instability 200
evaporation 61
- stability phenomenon 197,203
- modeling 62
glass temperature Tg 12
glass-mold system 239
failure analysis 427,434 glass-to-metal seal 413
fiber drawing 302 grain dissolution 84
fictive temperature Tr 11 - diffusive and convective 85
fining 24 - reactive 84
- antimony oxide 57 granular batch mixtures 78
- index 60 Grashof number 51
- sulfate 54, 55
finite impulse response (FIR) 143 Hagen-Poiseuille law 195
finite-element method (FEM) 385 heat conduction 7, 33, 70, 311
flanges - phononic 240
- 3-ring/spoke 229 heat generation 386
- combined contour and thickness heat transfer 87
optimization 228 - active heat conductivity 257
- contour optimization approximation of the boundary layer
-- 2D 224 313
- shape optimization 209 coefficient 109,241,261
2D 212 - Fresnel reflection 247
-- 3D 212 - heat-flux profiles 252
Index 475

in casting experiments 286 melting


intensity spectrum 248 factor 59
laboratory experiments 259 flux density 106
- physics and mathematics 245 - high-frequency 126
pressing 316 index 59
- radiative contributions 255 kinetics 25,100
- sample computations 252 reactions 87
heat transfer in hot forming 239 mesh generation 318
coefficient of radiative heat transfer
method
255
asymptotic 294
interface conditions 249
- opaque materials 240 - formula-analytic 2
radiation intensity 251 - moving body 170
radiation leakage 250 press-blow 307
- radiative flux 247,251 - remeshing 318
- semi-transparent glass 240 - Runge-Kutta 172
- temperature profiles 252 sliding mesh 170
- time dependence of heat content - volumes-of-fluid (VOF) 308,317
253 microcrack cut 381
- time dependence of temperature mixing mechanism 165
253 mixing process 165
high-viscous cords 66 model
homogeneity of glass melts 165 combustion 37
homogenization 24, 166 - computational fluid dynamics (CFD)
Hooke's law 10 17
hot forming 239,262 - detailed combustion space radiation
- spectral imaging 279 35
- discrete transfer (DTM) 37
Improved Eddington-Barbier 290
fiber- and tube-drawing 302
index thermometry 262
inductive heating 126 finite impulse response (FIR) 143
industrial simulation 4 radiation 37
intensity spectrum 248,277 - simple combustion space radiation
interaction matrix 151 34
interactive system 150 - soot 38
interface conditions 249 - thin-layer 294
interfacial area 165 - Tool-Narayanaswamy 339
inverse problems 2 - Trouton 298
- turbulence 36
Joule effect 129 model-based process control 17,18,29
model-predictive control (MPC) 142
laser heating 410 - limited dynamic performance 149
laser modes 383 - restricted applicability 150
- with constraints 147
manipulated variables (MVs) 139
- without constraints 143
mass conservation 99
material parameters modeling
- for inductive heating 134 - dynamic 71
- phenomenological 4 - mathematical 2
mathematical modulus of elasticity 339
- machine 2 mold surface temperature 242
- modeling 2 mold surfaces 240
- simulation 2,3 molecular simulation 4
Maxwell equations 12, 126, 133 momentum conservation 104,106
476 Index

N avier-Stokes equations 169,294 - ANSYS 209,385,402


Newtonian flow behavior 33 - FIDAP 308
Newtonian fluid 310 - GAMBIT 318
number - POLYFLOW 301,318
- Biot 109 Pt materials 450,464
- Damk6hler 106 - ODS-Pt 464
- Fourier 109 - PtRh10 464
Grashof 51 pull-out tests 424
Nusselt 194 push-out tests 433
Peclet 108
Prandtl 206
- - modified 296 radiation heat transfer 33, 37
- Reynolds 50,206 radiative energy transport 102
- - modified 296 radiative flux 247,251
- Schmidt 51 Raman thermometry 262
- Sherwood 50 reaction enthalpy 77
- Stokes 296 reaction kinetics 79
Nusselt number 194 reactive sintering 83
reconstructed temperature profile 278
optical lithography 438 redox reactions 55
optimization 443 refractory attack 67
optimization criteria 150 refractory loss 69
remeshing algorithms 318
paddle stirrer 189 remote optical sensing 262
particle paths 169 - methods 262
particle tracing 46 absorption thermometry 262
Peclet number 108 - - index thermometry 262
phase diagrams 80 - - Raman thermometry 262
physical models 19 - - spectral pyrometry 262
plate glass 293 remote sensing technique 286
platinum pipes 193 residence time 23,25,47,59, 173
plunger temperature 331 response time 150
Poisson ratio 339 reticle stage 439,446
polyvalent ions 23 Reynolds number 50, 206
pot melting 96 ripples
Prandtl boundary layer 243 - chill 326
Prandtl number 206 - flow 326
predicting the response 72 - press 326
press ripples 326 Runge-Kutta method 172
pressing 316,343
- of stems 307
- process 317,320 sagging 343
process sand-grain dissolution 48, 58
- blow molding 293 - rates 48
- Danner 304 Schmidt number 51
- downdraw 293 semi-transparent media 102
- float 293 shear modulus 341
- overflow-fusion 293,305 Sherwood number 50
- pressing 317, 320 simulation
- spinning of optical fibers 293 forward 3
- Vello 304 - industrial 4
program mathematical 3
- ABAQUS 340, 345 - molecular 4
Index 477

sintering Technical Committee 21 (TC 21) 74


liquid phase 83 Technical Committee 25 (TC 25) 319
- of cullet 83 temperature
- reactive 83 - conductivity 311
solubility of gas 52 - distribution 44, 292, 388
soot formation 39 - field 20,45
spectral imaging 279 - gradient 388
- glass gob 280 - profile 276, 286
- temperature maps 283 thermal
- thermo-image 280 - conductivity
spectral pyrometry 262, 263, 286 - - apparent 288, 291
- and thermocouple measurement - - phononic 288
278 - radiation 7
- inverse problem 266 - shrinkage 415
- Eddington-Barbier approximation thermochemical reactions 76
268 thermocouples 276
- laboratory experiment 275 thin-layer flow 293
- regularization method 269 thin-layer model 294
- sample computations 272 - validation 300
- simple approximation 268 time-dependent response 18
spectral thermocamera 282 tolerance limits 139
spinning of metals 449 Tool-Narayanaswamy model 339
spinning of optical fibers 293 tracer experiments 173
Stefan problem 96 trajectory 22,57
sticking 418 transition range 9
sticking temperature 240 Trouton model 298
stiffness 441 tube drawing 302
stirrer 165 turbulence 36
- ideal 173
Stokes number 296 ultra-thin plate glass 295
strain tensor 9
streamlines 172
validation 2
stress
- distribution 391 Vello process 304
- field 408 vibrations 439
- intensity factors KJ 396 virtual laboratory 2
- levels 382 viscoelastic behavior 339
- maximum 393 viscosity 193
- relaxation 339 Vogel-Fulcher-Tamman (VFT) curves
- tensor 9 204
- von Mises 415
striae 166 Wei bull distribution 425, 433
structural relaxation 9 wrinkles 454
submodeling technique 399
sucked sheet 399 yield stress 449

You might also like