Discrete Element Methods PDF

DISCRETE ELEMENT
METHODS
Simulations of Discontinua:
Theory and Applications
Antonio Munjiza
Conference Technical Team:
Guillermo Gonzalo Schiava D’Albano
Tomas Lukas
Esteban Rougier
© 2010 by Queen Mary, University of London
The editor and publisher of this book have used their best efforts in preparing this book.
The editor and publisher make no expressed or implied warranty of any kind with regard
to any material in this book. The editor and publisher should not be liable in any event for
incidental or consequential damages in connection with using the material in this book.
ISBN: 978-0-9551179-8-5
CONTENTS
OPENING LECTURE
From Nanoscience to Astrophysics-Towards Discrete Populations Based
Virtual Experimentation 2
PLENARY LECTURE
Multi-core Strategies for particle methods 11
KEYNOTE LECTURES
DEM Validation – a Simple Problem 19
Coupled Discrete Element and Fluid Modelling with Non-Spherical Grains 25
Modelling of Coupled Multi-Physics in Discrete Systems 31
1. ALGORITHMS AND SOLVERS

MRCK_3D Contact Detection Algorithm 37
Polyhedra on the Cheap 46
An Event Driven Element Modelling Approach 51
Development of a Rotational Resistance Model in the Discrete Element
Method 57
Description of Rotation in the Movable Cellular Automaton Method 63
A Bounding Box Search Algorithm for DEM Simulation 69
A Spring System Equivalent to Continuum Model 75
Resolving the Indeterminacy of Vertex-Vertex Contact in the 2D
Discontinuous Deformation Analysis 86
LIGGGHTS – a New Open Source Discrete Element Simulation Software 92
2. COUPLED SOLUTIONS
Discrete Thermal Element Modelling of Heat Conduction in Spherical
Particle Systems 100
Implementation of Combined Single and Smeared Crack Model in 3D
Combined Finite-Discrete Element Analysis 102
Modelling of Reacting Discrete Particles in Continuous Fluid Flow:
an Energy Technology Perspective 108
Implementation of Tangential Force in 3D Discrete Element and Combined
Finite-Discrete Element Methods 115
Development of Virtual Geoscience Simulation Tools, VGeST Using the
Combined Finite Discrete Element Method, FEMDEM 121
Coupled Gas-Particulate Discharge from Bucket Elevators 127
Large Scale Discrete Element Modelling of Fine Particles in a Fluidized Bed 133
A Preliminary Study on Modelling Liquid-Solid Interaction Using Smoothed
Particle Hydrodynamics 139
A Coupled DEM/CFD Study of Suction Filling 147
A Combined Contact Model in CDEM and its Application in Blasting
Engineering 153
Computational Investigation of the Dispersion of Cohesive Aggregates 159
3. VERIFICATION AND VALIDATION

Simulating and Optimizing of a Ball Mill by Comparison between Numerical
Discrete Element Method (DEM) and Experimental Method 166
Numerical and Experimental Investigation of Progressive Development of
a Granular Pile of Binary Size Pellets 167
Particle Shape Effects in Medical Syringe Needles – Experiments and
Simulations for Polymer Microparticle Injection 176
Experimental Validation of Polyhedral Discrete Element Model 182
Modelling the Trajectory of 3D Sensor-Shapes in Discharging Hoppers Filled
with Monosized and Binary Mixtures: an Experimental and DEM Comparison 188
Comparing Experimental Measurements of Mill Lifter Deflections with 2D
and 3D DEM Predictions 194
Numerical Analysis of Cold-Formed Sigma Steel Beams 199
4. PARALLELISATION, GPU, OPTIMISATION

Parallelization of an Open-Source FEM/DEM Code Y2D 206
Wet-Mixing of Powders, a Large-Scale GPU Implementation 212
Grid Parallel Computing System for Calibration of DEM Material Models 222
A GPU Accelerated Continuous-Based Discrete Element Method for
Elastodynamics Analysis 228
Large Scale GPGPU Implementation of the Discrete Element Method
Applied to Modelling the Environment in the Positron Emission Particle
Tracking Experiment 239
Calibration of DEM Material Models Using Optimisation Method in a Grid
Parallel System 243
5. FRACTURE AND FRAGMENTATION, CRACKS

Development of Discrete Element Method for Simulation of Deformation
and Fracture of Heterogeneous Elastoplastic Materials 250
Fracture Behaviour of Highly Porous Ceramics 256
Numerical Strategy for Discrete Fine Cracking Description with Continuous
Model 262
Implementation of 6-Node Element in DFEM for Quasi-Brittle Materials 268
Rock Impact Modelling Using FEM/DEM 269
6. GRANULAR MATERIALS, POWDERS AND NANOPOWDERS
The 3D Splash Behavior for the Impacting of One Particle on a Particulate
Packing 276
Fast and Stable Simulation of Granular Matter and Machines 282
Evolution of Structure in Granular Materials 288
DEM Simulation of Ball Indentation on Cohesive Powders 294
Statistics of Internal Structure of a Granular Pile Generated by the Particle
Expansion Method 300
Simulating Granular Material Behaviour to General Loading Paths with DEM 306
Enhanced Vibrational Granular Mixing 312
Discrete Element Modelling of Cylindrical Cavity Expansion of Granular
Materials 318
Effect of Gradation on the Constant Volume Cyclic Behaviour of Granular
Media 324
Applying DEM to Understanding Jamming in Systems of Non-Spherical
Grains 330
Combining DEM and 3D Imaging of Real Granular Systems 336
Discrete Element Modelling of Compaction Behaviour of Agglomerated and
Aggregated Nanopowder 342
Agglomeration Dynamics of Magnetic Nanoparticles in Simple Shear Flow 348
Modelling Non-Spherical Particle Breakage in DEM Simulations 354
7. SOILS, ICE, ROCK, LANDSLIDES, EARTHQUAKES

Coupled DEM-LBM Simulation Of a Soil Fluidisation Problem 361
Fold Development in Compressed Multi-Layers Modelled with FEMDEM 367
Discrete Modelling of Geomaterials Under Extreme Loading 373
Modelling the Effect of Narrow Blade Geometry on Soil Failure and
Draught Force Using Discrete Element Method 379
The Application of the Hybrid Stress Blasting Model to Improving the
Understanding of Wall Control Blasting 382
Discrete Element Simulation of Rock Cutting with Evaluation of Tool Wear 388
Modelling Punch-Through Tests with 2D FEM-DEM 394
Combined Finite Discrete Element Simulations of a Floating Ice Sheet
Failing Against an Inclined Structure 400
One Dimensional Compression of Sand-Silt Mixtures Using 2D DEM 406
Using Discrete Element Methods to Model Soil – Machine Interaction 412
The Development and Application of Stochastic Block Shape Model
in Continuum-Based Discrete Element Method 416
Adaptive Discrete/Finite Element Coupling for Rock Cutting Process
Simulations 428
Failure Mechanism about RSA under Equal Stress Boundary Conditions 434
Using DEM for the Assessment of K0 in Soils 440
Theoretical Investigation of Regularities of Model Fault Zone Mechanical
Response to Low-Amplitude Dynamic Mechanical Actions 446
The Use of Discrete Element Methods on the Dynamic Analysis of
Multi-Drum Ancient Structures 452
Simulating P-Wave Propagation in DEM 458
8. OTHER APPLICATIONS
A Case Study of Impact on Glass Using the Combined Finite-Discrete
Element Method 465
3-Distinct Element Analysis of Head Top Coal’s Stability Control during
Mechanized Top-Coal Caving in Steep Thick Seam 470
DEM Modelling of Particle Flow in a Turbula Mixer 471
Modelling Breakage Environment in Tumbling Mills Using DEM and
Analyzing the Outputs 477
Vehicle Dynamics on Off-Road Terrain 484
Charge and Structure Behaviour in a Tumbling Mill 490
Analysis of Particle Charging Behaviour via Rotating Chute of Blast Furnace
by Using Discrete Element Method 496
Study on Cutting Forces of SiC Machining Process with Pre-Stressed Using
DEM Simulation 502
DEM Simulation of Particle Motion in a Paddle Mixer 508
Prediction of Particle Breakage in Agitated Dryers: a Combined DEM and
Experimental Approach 514
Simulation of the Confined Compression Test of Iron Ore Pellets Using
Random Distributed 3D Multi Particle Finite Elements 520
On Application of Symbiotic Cellular Automaton Method for Simulation
of the Mechanical Response of Lignite 526
Modelling Flexible Material Using EDEM: Calibration and Industrial
Applications 532
DEM Analysis on Particle Behaviour in the Course of Sinter Mixture
Charging 538
Modelling Emergency Egress from a Public Facility 544
DISCRETE ELEMENT METHODS
OPENING LECTURE
Simulations of Discontinua – Theory and Applications 1

FROM NANOSCIENCE TO ASTROPHYSICS –

TOWARDS DISCRETE POPULATIONS BASED
VIRTUAL EXPERIMENTATION
Antonio Munjiza
University of London
This is an opening lecture for the DEM5 International Conference on

Discrete Element Methods held in London on 25th-26th August 2010.
The aim of this lecture is to list and elaborate on some key challenges
and opportunities that need to be addressed by the next generation of
discontinua based simulation tools.
FINITE ELEMENT METHODS

Computational Mechanics and Computational Physics in general have gone a
long way from their humble beginnings when “large scale” simulations comprised
16 or 32 finite elements.
From these humble beginnings over decades the researchers have first moved
from linear to nonlinear problems. These were followed by introducing ever
greater complexity into the simulations in order not only to increase accuracy but
also to address the “real nature” of a particular problem.
The ambition was always to model more and more complex problems and
processes. Very often added complexity of new problems required development
of new computational methods.

In the early days finite difference methods were mostly employed. The next
logical step was the introduction of finite element methods. These were initially
introduced mostly for the reason of better approximation of boundary conditions.
Problems involving localizations, discontinuities, softening, fracture, etc.
stretched the capabilities of the finite element methods to its limits. These posed
new challenges to the finite element methods. In the process the finite element
mesh, which started as a great advantage of finite element based simulation tools,
became an obstacle in addressing some of these challenges.
Thus, both in solids and fluids a shift towards Lagrangian formulations, mesh-
free approaches, particle based approaches and similar were developed.

In parallel to the finite element community computational methods such as
DDA, DEM, MD were developed for particular problems.
Initially, the developments were completely independent from each other. For
instance the DEM community was developing algorithms completely independ-
ently from MD community.
Eventually the bridges were built between these tools. Also bridges towards the
main stream methods of continuum such as the finite element methods were con-
structed and combined approaches developed such as FEM/DEM and fluid cou-
pling.
TOWARDS DISCRETE POPULATIONS METHODS

In building the bridges between different methods, it is soon realized that the
boundary element method can be simply seen as a subset of the finite element
method. In turn the finite element method can bee seen as a subset of the mesh-
less methods.
In addition it can be demonstrated that the finite element method can be re-
duced to a special case of generalized discrete element methods. The finite ele-
ments simply become discrete elements and are interconnected at nodal points.

Now, when one thinks this way, a completely new approach is opened in both
the way the finite element method is implemented and applied.
Molecular dynamics can also be seen as a subset of discrete element methods
and discrete element methods can be seen as generalization of molecular dynam-
ics approaches.
One can also say that the discrete element methods are a special case of “dis-
crete populations” simulations. If we start thinking this way, there is no reason for
not extending the application of discrete element methods to modeling financial
markets or crowds or even biological populations.
By thinking this way we soon arrive to a general analogy: the “population” is a
mesh of finite elements, or a pile of discrete elements or a group of atoms and
molecules or market participants or terrestrial bodies, etc.
EMERGENT PROPERTIES
In the above light we can think of results of discrete element simulations as
emergent properties. For the case of simple linear finite element method, the in-
teraction between discrete entities (finite elements) is reduced to linear interaction
between the neighboring elements at nodes of the finite element mesh. The emer-
gent property is simply stress distribution.
In molecular dynamics problems the individual discrete entities are for instance
atoms. Interaction between these atoms gives the problem its “constitutive real
world” qualities. These through complex interaction patterns in time and space
produce emergent properties.
For instance interacting atoms of argon at low temperature will produce spheri-
cal blobs of atoms. These are nothing but liquid droplets of argon. These droplets
constantly exchange atoms with the neighboring argon gas. However, they pre-
serve spherical shape, which is known to be due to the surface tension.
The nice thing is that the molecular dynamics model employed never explicitly
mentions the surface tension; the surface tension together with liquid droplets ap-
pears “out of nothing” as an emergent property.

GENERALISATION OF THE CONCEPT OF EMERGENT PROPERTY

In the light of the above discussion, it is not difficult to observe that much of
the world around us functions this way.
For instance take brain cells; interaction between these produces complex pat-
terns in time and space. These in turn are manifested as “intelligence” in a similar
way that the motion of argon atoms closely attracted to each other manifests itself
as a droplet of liquid.
The concept can be generalized to complex behavior of biological populations,
but also to complex emergent properties associated with human populations such
as markets, crowds, unrests, etc.
In principle the same simulations approach applies to all of these:
Take individual discrete entity, describe its interaction with other discrete enti-
ties and finally extract emergent properties.
VIRTUAL EXPERIMENTATION
One can say that the generalized concept of discrete populations leads to a
paradigm shift in how we see the real world.
In the past scientists would rely on mathematical model for a physical problem,
usually through a set of governing differential equations. Computational tools
were then developed to “discretise” these equations in the spatial domain. In other
words computational methods were a way of solving a particular set of PDEs.
With discrete populations concepts, we are moving away from the set of partial
differential equations and replacing these with a set of interaction rules between
discrete entities – discrete elements.
In a sense the discrete populations system becomes a parallel virtual world and
simulation becomes a virtual experiment, while the results of the simulation bring
new quality.
Classic simulations bring results in terms of quantity, but are not able to bring
“new quality”. Virtual experiment, much like a real experiment surprises us with

an emergent qualitative property such as a new chemical compound or a new state

of matter or a qualitatively new phenomena of different nature depending on the
experiment.
Figure 1: Droplets of liquid argon from a nanoscale virtual experiment
Figure 2: Planetesimal scale virtual experiment

Virtual experiments can be performed on any scale. On the nano and atomic
scale they are mostly designed to mimic the real nature of the problem as shown
in Figure 1.
On other scales including the terrestrial scales they can be a “model” of a real
world. For instance in Figure 2 a growing planetesimal is simulated as an assem-
bly of 1016 rigid, smooth spherical particles of radius 50 m and mass 5.2·108 kg.
The particles are held together by gravity alone without friction or any other dis-
sipative forces.
KEY CHALLENGES
There remains much work to be done on modeling complex physics associated
with discontinua simulations in general. Among these are complex fracture pat-
terns, dynamic fracture, complex interaction laws such as rock joints, liquid
bridges, nano-scale interactions, field based interactions, coupling, complex appli-
cations, etc.
Another important front are further developments of key algorithmic proce-
dures. These involve processing interactions, detecting contacts, visualization
tools, etc.
The third front is addressing the affordability of large scale simulations. During
the 1980s and 1990s one could afford to do nothing on this front. Doing nothing
in actual fact meant waiting for few years and getting a faster processor on your
desktop.
The problem is that this avenue is for the time being not available any longer.
Instead of getting a faster processor on your desk, one is getting a larger number
of processors.
HARDWARE DILEMA
By getting more processors one is confronted by the computer architecture di-
lema. A completely new layer of complexity is added to the simulations; in addi-
tion to having interacting particles, we have interacting processors as well.

These interacting processors either transfer big chunks of data between each
other or compete for access to the same data. In both cases, the CPU is no prob-
lem any longer. The problem becomes memory.
In a sense there is a parallel to the 1980s. In those days the biggest problem
was RAM availability, i.e. size.
These days RAM size is no problem, but access to it is. If a number of proces-
sors compete for the access to the same RAM, the CPU performance is destroyed
by the waiting for RAM.
In a similar way, if processors need to transfer a large amount of data between
different processors, most of the time will go into copying chunks of RAM from
one place to an other.
There have been a number of attempts to resolve these in terms of both hard-
ware and software. Good examples are MPI, Thread Building Blocks, Open MPI,
CUDA, etc.
GPU AS AN EXAMPLE OF WHAT COULD BE

GPUs are massively parallel architectures built for processing graphics. These
architectures are relatively inexpensive and efficient in terms of energy input
needed to process large amount of data.
The speed of these architectures is mostly achieved by a clever way of dealing
with how memory is handled. This way the memory bottleneck is largely elimi-
nated.
By using GPU architecture instead of CPU, some complex simulation prob-
lems have been speeded up 100s even 1000s times.
The problem is that the above statement is valid only for special applications
and cannot be generalized to all simulations.
A good example is medical imaging applications where extremely efficient so-
lutions are not difficult to devise. Some applications of molecular dynamics bene-
fit a lot from the GPU approach.

On the other hand, general discrete element simulations with all their complex-
ity are not easy to map onto the relatively rigid architecture of the GPU. Neverthe-
less, the above mentioned examples give us some confidence that we are entering
the era of exponential growth of parallel architecture approaches.
It is most likely that parallel architecture designers will have to take some input
from simulation community in order to address the key challenges of massively
parallel scalable architectures.
If successful, it will provide another revolution in both scale and complexity of
simulations.
CONCLUSIONS
This is the 5th conference in the series organized by my predecessors. It has
presented us with a good number of high quality contributions ranging from engi-
neering, industrial and scientific applications to novel algorithmic developments
some of which go as far as paradigm shifting and/or addressing the problems solu-
tion of which would provide a qualitative leap forward.
ACKNOWLEDGMENTS
I would like to thank the International research community for trusting me with
organizing the conference.
Also, I would like to thank my local team of hard working organizing commit-
tee, secretaries, finance people, www people, publications people, catering people
without whose help it would not be possible to have this conference.
At the end I would like to thank all the participants, authors, reviewers, key-
note speakers and the plenary speaker for all the effort and good will.

PLENARY LECTURE

Multi-core Strategies For Particle Methods
John R. Williams, David Holmes and Peter Tilke*

Massachusetts Institute of Technology, *Schlumberger-Doll Research Laboratory
This paper discusses the implementation of particle based numerical methods
on multi-core machines. In contrast to cluster computing, where memory is dis-
tributed across machines, multi-core machine can share memory across all
cores. Here general strategies are developed for spatial management of particles
and sub-domains that optimize computation on shared memory machines. In
particular, we extend cell hashing so that cells bundle particles into orthogonal
tasks that can be safely distributed across cores avoiding the use of “memory
locks” while still protecting against race conditions. Adjusting task size pro-
vides for optimal load balancing and maximizing cache hits. Additionally, the
way in which tasks are mapped to execution threads has a significant influence
on the memory footprint and it is shown that minimizing memory usage is one
of the most important factors in achieving execution speed and performance on
multi-core. A novel algorithm called H-Dispatch is used to pipeline tasks to
processing cores. The performance is demonstrated in speed-up and efficiency
tests on a smooth particle hydrodynamics (SPH) flow simulator. An efficiency
of over 90% is achieved on a 24-core machine.
INTRODUCTION
Recent trends in computational physics suggest a rapidly growing interest in par-
ticle based numerical techniques. While particle based methods allow robust han-
dling of mass advection, this comes at the computational cost of managing the
spatial interactions of particles. For problems involving millions of particles, it is
necessary to develop efficient strategies for parallel implementation. A variety of
authors have reported parallel particle codes on clusters of machines (see for ex-
ample Nelson et al [1], Morris et al [2], Sbalzarini et al [3], Walther and Sbalza-
rini [4], Ferrari et al [5]). The key difference between parallel computing on clus-
ters and on multi-core is that memory can be shared across all cores on today’s
multi-core machines. However, there are well known challenges of thread safety
and limited memory bandwidth. A variety of approaches to programming on
multi-core have been proposed to-date. Concurrency tools from traditional cluster

computing, like MPI [6] and OpenMP [7] have been used to achieve fast take-up
of the new technology. If one runs an MPI process on each core, the operating
system guarantees memory isolation across processes. Communication of infor-
mation across process boundaries requires MPI messages to be sent, which is rela-
tively slow compared with direct memory sharing.
In response to this, Chrysanthakopoulos and co-workers [7, 8] have imple-
mented multi-core concurrency libraries using Port based abstractions. These
mimic the functionality of a message passing library like MPI, but use shared
memory as the medium for data exchange, rather than exchanging serialized
packets over TCP/IP as is the case for MPI. Such an approach provides extensibil-
ity in program structure, while still capitalizing on the speed advantages of shared
memory. In an earlier paper, Holmes et al. [9] have shown that a programming
model developed using such port based techniques provides significant perform-
ance advantages over MPI and OpenMP. In this paper, we apply the programming
model proposed in [9] to the parallelization of particle based numerical methods
on multi-core.
PARTICLE BASED NUMERICAL METHODS
Mesh based numerical methods have been the cornerstone of computational phys-
ics for decades. Here, integration points are positioned according to some topo-
logical connectivity or mesh to ensure compatibility of the numerical interpola-
tion. Examples of Eulerian mesh based methods include finite difference (FD)
and the lattice Boltzmann method (LBM), while Lagrangian examples include the
finite element method (FEM). While powerful for a wide range of problems, mesh
limitations for problems involving large deformation and complex material inter-
faces has led to significant developments in meshless and particle based method-
ologies. For such methods, integration points are positioned freely in space, capa-
ble of advecting with material in a Lagrangian sense. For methods like molecular
dynamics (MD) and the discrete element method (DEM), such points represent
literal particles (atoms and molecules for MD and discrete grains for DEM), while
for methods like dissipative particle dynamics (DPD) [10], smooth particle hydro-
dynamics (SPH) [11], wavelets [12], and the reproducing kernel particle method
(RKPM) [13], the particle analogy is largely figurative. For such methods, parti-

cles provide positions at which to enforce a partition of unity. By partitioning

unity across the particles, continuity can be imposed without a defined mesh, al-
lowing such methods to represent a continuum in a generalized way. In Eulerian
mesh based methods, such as FD and LBM, continuity is inherently provided by
the static mesh, while for Lagrangian mesh based approaches like FEM, continu-
ity is enforced through the use of element shape functions. The partition of unity
imposed on mesh-free particle methods can be seen to be a generalization of shape
functions for arbitrary integration point arrangements. From Li and Liu [14]
...meshfree methods are the natural extension of finite element methods, they pro-
vide a perfect habitat for a more general and more appealing computational
paradigm - the partition of unity.
The advantage of partition of unity methods is that any expression related to a
field quantity can be imposed on the continuum. Where, for a bounded domain -
in Eucledean space, a set of nonnegative compactly supported functions, sums to
unity. Correspondingly, the value of some field function can be determined from
its value at all other points via
f ( x )   ak  ( x  k )
k 0 ,1, 2... n
The function f (x) can be related to any physical field expression; hydrodynamic,
mechanical, electrical, chemical, magnetic etc. Such versatility is a key advantage
of particle methods.
Strategies for Programming on Multi-Core By eliminating the need for ghost
regions and high volume to surface ratio sub-domains, shared memory enables a
near arbitrary selection of spatial divisions by which to distribute a numerical
task. Correspondingly, it becomes convenient to repurpose a spatial division strat-
egy already in use in one form or other in most particle based numerical codes.
Spatial hashing techniques have been used widely to achieve algorithmically effi-
cient interaction detection in particle codes (see for example DEM [15, 16]).
SPATIAL HASHING IN PARTICLE METHODS
The central idea behind spatial hashing is to overlay some regular spatial structure
over the randomly positioned particles e.g. an array of equally sized cells. We can
then perform spatial reasoning on the cells rather than on the particles themselves.

Spatial hashing assigns particles to cells or ‘bins’ based on a hash of particle co-
ordinates. The numerical expense of such an algorithm is O (N) where N is parti-
cle number [15,16]. A variety of programmatic implementations of hash cell
methods have been used (for example linked list). In this work a dictionary hash
table is used where a generic list of particles is stored for each cell and indexed
based on an integer unique key to that cell, i.e. key  ( k  ny  j )  nx  i where
nx and ny are the total number of cells in the x and y dimensions and i, j and k are
the integer cell coordinates in the x, y and z dimensions.
Managing Thread Safety The cells also provide a means for defining task pack-
ages to the cores. By assigning a number of cells to each core we assure “task or-
thogonality” in that each core is operating on different set of particles. Traditional
software applications for shared memory parallel architectures have utilized locks
to avoid thread contention. We note that a core may “read” the memory of parti-
cles belonging to surrounding cells but may not update them. We execute a single
loop in which global memory for both previous and current field values ( vtn and
vt n1 ) is stored for each particle. Gradient terms can then be calculated as func-
tions of values in previous memory, while updates are written to the current value
memory, in the same loop. In parallel, minimizing the frequency of so called syn-
chronization points has advantages for performance and we utilize this “rolling
memory algorithm”. This allows the previous and updated terms to be maintained
without needing to replace the former with the latter at the end of each step and
thus, ensuring a single synchronization per step.
Counter Intuitive Design In a typical SPH simulator, two operations must be

done on particles in each cell per step, separated by a synchronization, the first to
determine the particle number density of each particle, and the second to perform
the field variable updates. Interacting particles must be known for each of these
two stages. Using a standard SPH formulation, the performance of two structure
variations are compared in this case study: A. Interacting particles are determined
in the first stage for all cells and stored in a global list for use in the second, and
B. Interacting particles are determined as needed in each stage for each cell, i.e.
twice per cell per time step. Recalculation of interacting particles means they need

only be kept in a local thread list that is overwritten with each newly dispatched
cell. While differences in execution memory are to be expected of the two code
versions, the differences in execution time are more surprising. For low core
counts (< 10 cores), as would be expected, the single search variant (A) solves
more quickly than the double (B) due to less computations. After this point, how-
ever, the double search (B) is shown to provide marked improvements in speed
over (A) (up to 50%). This can be attributed to better cache blocking of the second
approach and the significantly smaller amount of data experiencing latency when
being loaded from RAM to cache. The fact that such performance gains only
manifest when more than 10 cores are used, suggests that for less than 10 cores,
RAM pipeline bandwidth is sufficient to handle a global interaction list.
Fig 1. (a) memory vs. Number of Particles (b) Speed-up vs. Number of Cores
The MIT SPH simulator (see http://geonumerics.mit.edu ) has been validated for
single and multi-phase flow. Figure 2 shows results for two phase Rayleigh-
Taylor instability [17].

Fig 2. Rayleigh-Taylor instability example showing extremely complex fluid

phase interface geometries; clear liquid is water, green liquid is oil. [9].
REFERENCES
1. M. T. Nelson, W. Humphrey, A. Gursoy, A. Dalke, L. V. Kale, R. D. Skeel, K.
Schulten, NAMD: a parallel, object-oriented molecular dynamics program, Inter-
national Journal of High Performance Comp. Applications 10 (1996) 251-268.
2. J. P. Morris, Y. Zhu, P. J. Fox, Parallel simulations of pore-scale flow through
porous media, Computers and Geotechnics 25 (1999) 227-246.
3. I. F. Sbalzarini, J. H. Walther, M. Bergdorf, S. E. Hieber, E. M. Kotsalis, P.
Koumoutsakos, PPM - a highly efficient parallel particle-mesh library for the
simulation of continuum systems, Journal of Computational Physics 215 (2006)
566-588.
4. J. H. Walther, I. F. Sbalzarini, Large-scale parallel discrete element simula-
tions of granular flow, Engineering Computations 26 (2009) 688{697.

DEM5-2010, London, 25-26 August 2010 Page 7
5. A. Ferrari, M. Dumbser, E. F. Toro, A. Armanini, A new 3D parallel SPH

scheme for free surface flows, Computers and Fluids 38 (2009) 1203-1217.
6. W. Gropp, E. Lusk, A. Skjellum, Using MPI: Portable Parallel Programming
With the Message-Passing Interface, MIT Press, Cambridge, 1999.
7. X. Qiu, G. C. Fox, H. Yuan, S.-H. Bae, G. Chrysanthakopoulos, H. F. Nielsen,
Parallel data mining on multicore clusters, in: 7th International Conference on
Grid and Cooperative Computing GCC2008.
8. G. Chrysanthakopoulos, S. Singh, An asynchronous messaging library for C#,
in: Proceedings of the Workshop on Synchronization and Concurrency in Object-
Oriented Languages, OOPSLA 2005, San Diego, 2005, pp. 89-97.
9. D. W. Holmes, J. R. Williams, P. Tilke, An events based algorithm for distrib-
uting concurrent tasks on multi-core architectures, Computer Physics Communi-
cations 181 (2010) 341-354.
10. M. Liu, P. Meakin, H. Huang, Dissipative particle dynamics simulation of
pore-scale multiphase fluid flow, Water Resources Research 43 (2007) W04411.
11. J. J. Monaghan, Smoothed particle hydrodynamics, Annual Review of As-
tronomy and Astrophysics 30 (1992) 543{574.
12. W. K. Liu, Y. Chen, Wavelet and multiple scale reproducing kernel methods,
International Journal for Numerical Methods in Fluids 21 (1995) 901{931.
13. W. K. Liu, S. Hao, T. Belytschko, S. Li, C. T. Chang, Multi-scale methods,
International Journal for Numerical Methods
14. S. Li, W. K. Liu, Meshfree Particle Methods, Springer-Verlag, Berlin Heidel-
berg, 2004.
14. A. Munjiza, K. R. F. Andrews, NBS contact detection algorithm for bodies of
similar size, International Journal for Numerical Methods in Engineering 43
(1998) 131-149.
16. J. R. Williams, E. Perkins, B. Cook, A contact algorithm for partitioning N
arbitrary sized objects, Engineering Computations 21 (2004) 235-248.
17. A. M. Tartakovsky, P. Meakin, A smoothed particle hydrodynamics model
for miscible flow in three-dimensional fractures and the two-dimensional
Rayleigh-Taylor instability, Journal of Computational Physics 207 (2005) 610-
624.

KEYNOTE LECTURE

DEM VALIDATION – A SIMPLE PROBLEM
Colin Thornton1, Sharen Cummins2, Paul Cleary2

1
School of Chemical Engineering, University of Birmingham, Birmingham, B15 2TT, UK
2
CSIRO Division of Mathematics, Informatics and Statistics, Private Bag 33, Clayton, Vic 3168,
Australia
Using a variety of contact force models reported in the literature we

compare the rebound kinematics obtained for both elastic and inelastic
oblique collisions of a sphere with a target wall. For elastic impacts, it
is shown that a simple linear spring model can produce excellent
agreement with the results of the much more complicated Hertz, Min-
dlin and Deresiewicz contact mechanics based model, provided that an
appropriate value of the ratio of the tangential to normal spring stiff-
ness is used. For inelastic impacts, reasonable agreement is obtained
when viscous dissipation models are compared with plastic dissipation
models provided that the coefficient of restitution is not set too low.
However, the investigation did reveal that some contact force models
are intrinsically wrong and reasons for this are discussed.
INTRODUCTION
Attempts to quantitatively validate DEM simulations by comparing with expe-
rimental data are often frustrated by uncertainty in terms of the experimental data
and the fact that frequently the simulated particles are spheres and the experimen-
tal particles are non-spherical. Even with the advent of very sophisticated experi-
mental techniques such, as PIV, MRI and PEPT, comparisons are normally re-
stricted to the spatial and temporal particle velocity distributions and not detailed

local events – in other words the general overall behaviour. However, the overall
behaviour in these systems is dominated by the applied external field and the
boundary conditions imposed rather than the physics at the particle scale. There-
fore, such comparisons may provide good correlations rather than rigorous valida-
tions against experiment.
An important aspect of DEM simulations, the significance of which is not fully
understood, is the sensitivity to the various contact force models that have been
proposed and implemented. In order to compare the consequences of using differ-
ent contact force models we examine the simple problem of a single spherical par-
ticle impacting a planar target wall at different impact angles and with no initial
particle spin. Both elastic and inelastic collisions are considered. The effects of
different contact force models on the behaviour of a sphere impacting a wall obli-
quely have been reported by Schäfer et al [1] but, since then, other models have
been proposed.
ELASTIC IMPACTS
It is now well established that for interactions between two elastic spheres or
an elastic sphere and a planar surface the normal interaction follows the theory of
Hertz, see Johnson [2], and the tangential interaction is provided by the theory of
Mindlin and Deresiewicz [3]. This combination of interaction rules will be re-
ferred to as the HMD model and will be used as the benchmark against which
other models will be compared. The solution for the rebound kinematics has been
illustrated by numerical implementation of the HMD model [4], DEM simulations
[5], FEM simulations [6] and experiments [7, 8, 9]. The excellent agreement be-
tween the various data sets has been demonstrated in [6, 10].
In addition to the HMD model, we examine a simpler version by combining
the Hertzian model for the normal stiffness with the ‘no-slip’ theory of Mindlin
[11] for the tangential stiffness. We label this the HM model. This model differs
from the H-MDns model suggested by Di Renzo and Di Maio [12] which we also
examine. We also consider the model proposed by Langston et al [13] and used by
[14], which is labeled the LTH model. Finally, we examine a linear spring (LS)
contact force model, originally suggested by Cundall and Strack [15]. This model

assumes that, during contact, the two interacting bodies are connected, both nor-
mally and tangentially, by linear springs.
INELASTIC IMPACTS
The most common contact force model, used for both normal and tangential in-
teractions, is the linear spring-dashpot model introduced by Walton [16], here re-
ferred to as the LS+D model. We also consider a non-linear spring variant in
which we add normal and tangential dashpots to the HM model and label this as
the HM+D model. These dashpot models dissipate energy through viscous means.
Other normal interaction models dissipate energy plastically via the use of differ-
ent loading and unloading spring stiffnesses.
In order to mimic plastic deformation as observed in experiments, Walton and
Braun [17] introduced a so-called ‘partially latching spring’ model in which the
normal unloading stiffness was higher than the normal loading stiffness. For the
tangential interaction, Walton and Braun [17] used a non-linear stiffness that was
crudely based on the Mindlin and Deresiewicz [3] loading curve. The same model
was used by Stronge [18] for the normal interaction but a linear spring was used
for the tangential interaction. We choose to examine the simpler model proposed
by Stronge [18] but due to the significance of the normal interaction we label it
the WBS model.
Three other models that incorporate plastic dissipation are also considered. The
most complex one is to combine the theory of Thornton [19] for the normal inte-
raction with the theory of Mindlin and Deresiewicz [3], which is labeled the TMD
model. A non-linear simplified version of the TMD model is to combine the
theory of Thornton [19] with the no-slip theory of Mindlin [11] and is labeled the
TM model. Finally, a linearised version of the TM model is examined that is es-
sentially the same as the model used by Luding [20] for non-adhesive contacts.
This model is labelled the LTM model.
RESULTS AND DISCUSSION

In the case of elastic impacts it was found that the results of the simple LS
model were in excellent agreement with the results obtained using the much more
complex HMD model provided that an appropriate value of the ratio of the tan-

gential to normal stiffness is used. The HM model performed less well but the re-
sults were not unreasonable and were better than the corresponding results ob-
tained using the HM-Dns model proposed in [12]. The LTH model predictions
were very different from the results predicted by all the other models. More de-
tails of the results are provided in [21].
In the case of inelastic impacts, all the models were in reasonable agreement
provided that the normal coefficient of restitution was not set too low, not less
than 0.5. However, the trends obtained with the WBS model were opposite to all
the other models as the coefficient of restitution was decreased. More details are
provided in [22].
Most of the models used for this simple impact problem were observed to give
reasonable results at least for low energy dissipation. However, it should be re-
membered that the linear springs in the LS, LS+D, WBS and LTM models were
calibrated to give the best possible agreement with the HMD and TMD models. In
general DEM simulations, even in the collisional regime, spring stiffnesses cannot
be so reliably calibrated a priori. The LS+D, WBS and LTM models use constant
pre-specified coefficients of restitution that are independent of the relative impact
velocity, although this can be somewhat modified as suggested in [17]. Also, the
impact duration is independent of the relative impact velocity in the LS, LS+D,
WBS and LTM models. These simplifications could be quite significant for sys-
tems near the jamming transmission when enduring contacts start to evolve.
Consequently, the HM, HM+D and TM models may be more attractive alterna-
tives to the complicated and computationally expensive HMD and TMD models.
This is because (i) the model parameters are based on real material properties,
Young’s modulus and Poisson’s ratio, (ii) there is no need for any calibration of
the normal spring stiffness (iii) the force evolution is more realistic and (iv) the
impact duration is velocity dependent. However, for systems with enduring con-
tacts, further comparative studies are required.

REFERENCES
1. J. Schäfer, S. Dippel, D. E. Wolf, ‘Force schemes in simulations of granular
materials’, J. Phys. I France, 6, 5-20, 1996.
2. K. L. Johnson, Contact Mechanics, Cambridge University Press, 1985.
3. R. D. Mindlin, H. Deresiewicz, ‘Elastic spheres in contact under varying
oblique force’, Trans. ASME J. Applied Mechanics, 20, 327-44, 1953.
4. N. Maw, J. R. Barber, J. N. Fawcett, ‘The oblique impact of elastic spheres’,
Wear, 38, 101-14, 1976.
5. C. Thornton, Z. Ning, C. Y. Wu, M. Nasrulla, L. Y. Li, ‘Contact mechanic and
coefficients of restitution’ In Granular Gases (T. Pöschel, S. Luding, eds.),
Springer Berlin, pp 184-94, 2001.
6. C. Y. Wu, C. Thornton, L. Y. Li, ‘Coefficient of restitution for elastoplastic
oblique impacts’, Advanced Powder Technology, 14, 435-48, 2003.
7. N. Maw, J. R. Barber, J. N. Fawcett, ‘The role of elastic tangential compliance
in oblique impact’, Trans. ASME J. Lubrication Technology, 103, 74-80, 1981.
8. S. F. Foerster, M. Y. Louge, H. Chang, K. Allia, ‘Measurements of the colli-
sional properties of small spheres’, Physics of Fluids, 6, 1108-15, 1994.
9. A. H. Kharaz, A. D. Gorham, A. D. Salman, ‘An experimental study of the
elastic rebound of spheres’, Powder Technology, 120, 281-91, 2001.
10. C. Y. Wu, C. Thornton, L. Y. Li, ‘A semi-analytical model for oblique impacts
of elastoplastic spheres’, Proc. Roy. Soc. London A, 465, 937-60, 2009.
11. R. D. Mindlin, ‘Compliance of elastic bodies in contact’, Trans. ASME J. Ap-
plied Mechanics, 16, 259-68, 1949.
12. A. Di Renzo, F. P. Di Maio, ‘Comparison of contact-force models for the
simulations of collisions in DEM-based granular flow codes’, Chemical Engi-
neering Science, 59, 525-41, 2004.
13. P. A. Langston, U. Tüzün, D. M. Heyes, ‘Discrete element simulation of
granular flow in 2D and 3D hoppers: dependence of discharge rate and wall

stress on particle interactions’, Chemical Engineering Science, 50, 967-87,

1995.
14. Y. C. Zhou, B. D. Wright, R. Y. Yang, B. H. Xu, A. B. Yu, ‘Rolling friction in
the dynamic simulation of sandpile formation’, Physica A, 269, 536-53, 1999.
15. P. A. Cundall, O. D. L. Strack, ‘A discrete numerical model for granular as-
semblies’, Géotechnique, 29, 47-65, 1979.
16. O. R. Walton, ‘Application of molecular dynamics to macroscopic particles’,
Int. J. Engineering Science, 22, 1097-107, 1983.
17. O. R. Walton, R. L. Braun, ‘Viscosity, granular-temperature, and stress calcu-
lations for shearing assemblies of inelastic, frictional disks’, J. Rheology, 30,
949-80, 1986.
18. W. J. Stronge, ‘Planar impact of rough compliant bodies’, Int. J. Impact Engi-
neering, 15, 435-50, 1994.
19. C. Thornton, ‘Coefficient of restitution for collinear collisions of elastic-
perfectly plastic spheres’, Trans. ASME J. Applied Mechanics, 64, 383-6,
1997.
20. S. Luding, ‘Cohesive, frictional powders: contact models for tension’, Granu-
lar Matter, 10, 235-46, 2008.
21. C. Thornton, S. J. Cummins, P. W Cleary, ‘An investigation of the compara-
tive behaviour of alternative contact force models during elastic collisions’,
Powder Technology, submitted.
22. C. Thornton, S. J. Cummins, P. W Cleary, ‘An investigation of the compara-
tive behaviour of alternative contact force models during inelastic collisions’,
Powder Technology, to be submitted.

COUPLED DISCRETE ELEMENT AND FLUID

MODELLING WITH NON-SPHERICAL GRAINS
J. E Hilton and P. W. Cleary

CSIRO Mathematics, Informatics and Statistics, Clayton South, Victoria, Australia
We present a coupled discrete element and fluid method in which the

particle shape can range from ellipsoidal to cuboidal, based on a su-
perquadric formulation. Our method takes into account factors such as
the local particle orientation and porosity. The method is applied to
systems including fluidised beds and pneumatic conveying. We show
that shape plays an important role in the dynamics of each of these
systems.
INTRODUCTION
Most grains possess some degree of shape, but much of the computational
modelling to date using the Discrete Element Method (DEM) has used the ap-
proximation of spherical grains. The shape of a grain can significantly influence
the dynamics of a granular system 1 . The coupling of fluid dynamics to DEM has
extended the capability of the method to a wide range of two-phase gas-solid or
gas-liquid systems, such as fluidised beds and pneumatic conveying. The influ-
ence of grain shape on these systems has not, however, been fully investigated.
We present a method for coupling a fluid model to DEM with shaped particles
and show applicability of our method to these systems. Importantly, we show that
particle shape alters the dynamics of each system, meaning that a spherical ap-
proximation may not result in an accurate simulation of the true behavior of these
systems.

DEM was first formulated by Cundall and Strack 2 . We use a superquadric

DEM formulation allowing granular systems with a wide range of grain shapes to
be simulated 3 . This is coupled to an Eulerian scheme for fluid flow through
packed particle beds. The fluid-particle coupling forces from the lift and drag
forces are based on empirical formulae and are modified, where necessary, to ac-
count for the shape of the particle.
a) b) c)
Fluid Velocity
Prolate Ellipsoid
Particle Level 1
Sphere
Oblate Ellipsoid
Level 2
Projected Face Area
Level 3
Cuboid
Fig 1. a) The projected face area used in drag calculations b) Voidage fraction
calculated using Octree method c) Shapes represented using superquadrics
Drag Forces. The drag force is the dominant fluid-particle interaction force in
the system. The drag force exerted on a particle in a multi-particle system with
corrections due to Di Felice 4 is:
2
FD = 12 cd ρ u r ε − χ A⊥u r
(1)
χ = 3.7 − 0.65 exp − { 1
2 (1.5 − log Re ) }
2
where ρ the fluid density, ur the local relative flow velocity and ε the local voi-
dage fraction. The local Reynolds number is given by Re = 2rρ|ur|/η where r is
the equivalent spherical radius of the particle and η is the fluid dynamic viscosity.
We use a drag coefficient cd for non-spherical particles given by Holzer et al. 5 ,
based on correlations to experimental data. The drag coefficient uses the projected
cross-sectional area of the considered particle perpendicular to the flow, which is
calculated for each particle at each timestep from the local incident relative flow

vector, as shown in Fig. 1a. The area is calculated by summation the projection of
discretised rim of the super-quadric particle onto a plane orthogonal to the flow
vector. Particles are also subject to Stokesian rotational drag, given by TD =
8πηr3ωr.
Lift Forces. Lift forces are crucial for systems with gas flow over small parti-
cles. Shear (Saffman) lift and rotational (Magnus) lift forces are both included in
our model. The shear lift force, FLS, on an isolated spherical particle is given by a
generalisation of Saffman's 1D form 6 by Drew 7 with an empirical correction
given by Mei 8 :
ur ⋅ D
FLS = 6.46cls r 2 ρμ
D
⎪
cls = ⎨
(
⎧e − Re10 − 0.3314 α e − Re10 + 1
) Re ≤ 40 (2)
⎪ 0.0524 α Re Re > 40
⎩
where the rate of deformation tensor D = (∇ur+∇urT) and α = |D|r/|ur|. A spherical
particle subject to a relative spin ωr = ½ωf - ωp, where ωf is the local fluid vortic-
ity and ωp is the particle spin, is also subject to a Magnus lift force, FLM, with an
empirical correction given by Lun and Lui 9 :
ωr × ur
FLM = 12 clmπ r 2 ρ
ωr
⎧⎪ 2r ω r Re ≤ 1 (3)
clm = ⎨
(
⎪⎩2r ω r 0.178 + 0.822 Re
−0.522
) Re > 1
Fluid dynamics. We use constitutive equations for gas flow through a porous
bed for modelling the system, given by Kafui5. If the gas density is assumed con-
stant the expressions can be re-formulated for the superficial gas velocity u′ = εu:
∂ε
= −∇ ⋅ u′ (4)
∂t

∂u′ 1 u′ 1
+ ( u′ ⋅∇ ) u′ + u′∇ ⋅ = − ⎡⎣ε∇p + f fp + ∇ ⋅ ( ε τ ) ⎤⎦ + ε g
∂t ε ε ρ
where p is the fluid pressure, the stress tensor τ takes the usual form and the fluid
particle interaction body force, ffp, is the volume averaged force from the particle
drag on the fluid. These equations are discretised on a Cartesian grid over the
simulation domain. The porosity is explicitly calculated from the DEM simulation
at each timestep using a recursive Octree approach, as shown in Fig. 1b 10 .
APPLICATIONS
Fluidised beds. A series of simulations was carried out of fluidised beds filled
with particles of varying shape: Spherical, for benchmarking and validation,
prolate ellipsoidal, oblate ellipsoidal and cuboidal, shown in Fig. 1c. The pressure
drop measured over the bed was compared to the empirical Ergun relation 11 .
1.6
and simulated pressure gradient, kPa
1.4
u
Difference between Ergun
1.2 Oblate Ellipsoid
1.0
0.8
Cuboid
u
0.6
Prolate Ellipsoid
0.4
0.2
Spherical
0.0
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0
Fluidisation velocity, u Spheres Cuboids
Fig 2. Difference between fluidised bed pressure drop from the Ergun relation
and from simulations with shaped particles (left). Comparison of a bed of spheri-
cal and cuboidal particles with an inflow gas velocity of 3 m/s, 0.15s after the start
of the simulation (right).
It was found that the pressure drop for shaped particles was much greater than
predicted using the Ergun relations. Furthermore, the pressure drop was greatest
for the oblate ellipsoids, which were found to have the highest average projected

face area in the flow direction. The difference between the simulated and pre-
dicted pressure drop is shown in Fig. 2 (left). The fluidization velocity of the bed
was also found to be lower for shaped particles. Fig. 2 (right) shows a comparison
between a bed of spheres and a bed of cuboidal particles after 0.15s with an in-
flow gas velocity of 3 m/s. The cuboidal bed has fully fluidised, but the spherical
bed is still within an un-fluidised bubbling regime.
Pneumatic conveying.
A long duct of square cross section with periodic boundary conditions was
used to asses the influence of particle shape on pneumatically conveyed particles.
∇p ∇p
0.05s
0.10s
0.15s
0.20s
0.25s
0.30s
Spheres Prolate Ellipsoids
Fig 3. Pneumatic conveying of spherical (left) and prolate ellipsoidal particles

(right) with an applied pressure gradient of 2400 Pa/m.
The duct was half filled with either spherical, cuboidal or a series of ellipsoidal
particles with different aspect ratios and driven down the pipe using a pressure
gradient imposed on the gas. If a pipe is blocked by particles a ‘slug’ can form in
the pipeline, which is a slow moving plug of material filling the entire bore of the
pipe. We found slugs only formed for particles with an aspect ratio close to
spherical. Fig. 3 shows a comparison between pneumatically conveyed spheres

(left) and prolate ellipsoidal particles with the same volume (right) at the same
imposed driving pressure gradient of 2.4 kPa/m. The spherical particles form a
slug, but the ellipsoidal particles form a dilute flow of airborne particles over a
sheared lower layer. No slugs were found to form for either prolate or oblate ellip-
soids with high aspect ratios. The behaviour of the system can be explained using
a force balance between the driving pressure through the bed and the frictional
resistance between the slug and the base of the duct. Beds composed of shaped
particles form denser packings, increasing the pressure gradient over the slug.
This increase in pressure gradient is sufficient to destabilise slug formation.
CONCLUSION
We have combined the Discrete Element Method for shaped particles with a
fluid solver. We have applied the method to a range of systems and found that
shape plays a very significant role in each system. Our results open new possibili-
ties for simulation of fundamental and applied granular flow systems and high-
light the limitations of the spherical approximation in the Discrete Element
Method.
REFERENCES
1
Cleary, P.W., Sawley, M.L. (2002), App. Math. Mod., 26, p89
2
Cundall, P. A. and Strack, O. D. L. (1979), Geotechnique, 29, p47
3
Cleary, P.W. (2004), Engineering Computations, 21, p169
4
Di Felice, R. (1994), Int. J. Multiphase Flow, 20, p153
5
Hölzer, A. and Sommerfeld, M. (2008), Powder Technology, 184, p361
6
Saffman, P. G. (1964), J. Fluid Mech, 22, p385
7
Drew. D. (1976), Archive for Rational Mechanics and Analysis, 62, p149
8
Mei, R. (1992), Int. J. Multiphase Flow, 18, p145
9
Lun, C. K. K. and Liu, H. S. (1997), Int. J. Multiphase Flow, 23, p575
10
Hilton J. E., Mason L. R. and Cleary P.W. (2010), Chem. Eng. Sci., 65, p1584
11
Ergun, S. (1952), Chem. Eng. Prog., 48, p89

MODELLING OF COUPLED MULTI-PHYSICS IN

DISCRETE SYSTEMS
Y. T. Feng, K. Han, D. R. J. Owen

Civil and Computational Engineering Centre, School of Engineering
Swansea University, SA2 8PP, UK
Abstract
The paper presents a numerical framework for modelling the cou-
pled multi-physics phenomena in particle systems, including thermal
and magentic fields. It is a further extension of the combined Lattice
Boltzmann and discrete element method developed for solid particle
transport in fluid flows.
INTRODUCTION
The discrete element method (DEM) has provided a powerful numerical tool
for modelling mechanical behaviour of particle systems. In a wide range of
applications in science and engineering involving solid particles, other physical
phenomena, such as fluid, thermal, radiation and electro-magnetic, may also
be present. The complexity involved in coupled multi-physical phenomenon
imposes a real challenge to develop effective numerical modelling strategies
for particle systems. Over the last few years, the authors have developed the
so called combined Lattice Boltzmann (LBM) and Discrete Element approach
[1, 4] which appears to provide an ideal numerical simulation framework par-
ticularly for solid particle transport in fluid flows, in which LBM is used to
model incompressible fluid flows; DEM is to model the motion of and the inter-
action between the solid particles; and the hydrodynamic interaction between

the fluid and particles are accounted for by utilisation of an immerse bound-
ary technique. The LBM can also be applied to model heat transfer in a fluid
system[2]. The current work presents the further extension of the methodology
to couple additional physical fields including thermal and electric/magnetic.
Discrete Thermal Element Modelling of Heat

Transfer in Particle Systems
In our recent work [3, 5], a novel numerical methodology termed the discrete
thermal element method (DTEM) is proposed the modelling of heat conduction
in systems involving a large number of circular particles. The basic formulation
is outlined below.
Consider a particle in an assembly which is in contact with n neighboring
particles. (Unknown) heat flux distributions are imposed over the n isolated
contact zones on the boundary of the particle, and the reminder of the particle
boundary is assumed fully insulated. Each contact zone, which is assumed to
be an arc, is described by the position in terms of an angle θ of its middle
point, and the contact angle α that determines the contact arc length.
The average temperature Ti over the i-th contact arc can be obtained as
n h
X Qj Z αi Z αj
∆θij + θ − φ i
Ti = − ln sin dφdθ + To (1)
j=1 4πκαi αj −αi −αj 2
where Qj is the resultant flux on the j-th arc; To is the average temperature
of the particle; and ∆θij = θi − θj . The above equation can be expressed
compactly as
n
X
Ti = hij Qj + To (i = 1, ..., n) (2)
j=1
where
1 Z αi Z αj
∆θij + θ − φ
hij = − ln sin dφdθ (3)

4πκαi αj −αi −αj 2

or in an explicit approximation form
1 3 αi2 αi4 αi6 αi8

hii = − ln αi + + + + + (4)
πκ 2 36 2700 79380 252000
and
1
hij = −
ln | sin(θi − θj )/2| (5)
πκ
By introducing the element temperature and flux vectors,
Te = {T1 , ..., Tn }T , and Qe = {Q1 , ..., Qn }T

h i
and the element thermal resistance matrix He = hij as well as an n × 1
n×n
T
vector with all components of unity, e = {1, ..., 1} , Eq. (2) can be expressed
in matrix form as
Te − eTo = He Qe (6)
Pn
By utilising the global heat flux equilibrium condition i=1 Qi = 0, Eq. (6)
can be rewritten as
Ke Te = Qe (7)
where
c − g gT /κ ;
Ke = K c = H−1 ;
K c e, κ = eT g
ge = K
e e e e e e e e e
is the so-called element conductivity matrix.

Eq. (7) is essentially the heat conduction equation in discrete form for the
particle, which defines the discrete thermal element. An important feature of
(7), and thus this new thermal element, is that it has an identical form as a
finite thermal element. In other words, each circular particle can be viewed as
a special shaped finite thermal element, although both approaches stem from
different principles. The immediate consequence is that the subsequent proce-
dure in this discrete thermal element method to model the assembled particle
system can follow the same steps as used in FEM. For transient analysis, a
simplified version of the DTEM, termed the pipe-network model, is proposed

in [5] in which each particle is replaced by a simple thermal pipe-network con-

necting the particle centre with each contact zone associated with the particle.
The entire algorithmic structure of the pipe-network model is compatible with
the discrete element method, leading to an effective scheme for simulating
thermal-mechanical coupling problems. The numerical examples are given in
[3, 5]. The method can also easily be combined with the thermal LBM to
model heat transfer (including forced or natural convection) in a particle-fluid
(gas) system.
The above methodology has also be extended to 3D spherical particles [7].
Modelling Magnetorheological Fluids

A magnetorheological (MR) fluid is a type of smart fluid which is composed of
micrometer-sized particles suspended in a carrier fluid. Without an externally
applied magnetic field, a MR fluid is usually Newtonian. Upon application of
an external field, the particles acquire an magnetic dipole moment, which, due
to dipolar interactions, induces particle aggregation to form chainlike struc-
tures aligned with the applied magnetic field, perpendicular to the direction
of the fluid flow. This change in the suspension microstructure greatly al-
ters the rheological properties of the fluid due to the restricted fluid motion.
The apparent viscosity of the fluid can be increased up to several orders of
magnitude and the fluid becomes non-Newtonian. The rheological behavior of
the fluid could be described as that of a classical Bingham fluid - a material
with a yield stress that does not flow unless the applied stress exceeds the
yield stress. Additionally, the response of a MR fluid to the external magnetic
field is usually rapid (in milliseconds). These features have given rise to many
control-based applications, ranging from dampers and clutches in mechani-
cal engineering, shock absorbers in the automotive and aerospace industry, to
finishing in optics.
The combined lattice Boltzmann and discrete element approach [1, 4] has
proved effective in simulations of particle transport problems. The approach
can be extended to model MR fluids if the magnetic interaction between the
particles can be appropriately considered. The whole issue has been compre-

hensively addressed in [6], together with the discussion of all the computational
issues involved.
References
[1] Y. T. Feng, K. Han, D. R. J. Owen. Coupled lattice Boltzmann method
and discrete element modelling of particle transport in turbulent fluid
flows: Computational issues. International Journal for Numerical Meth-
ods in Engineering, 72 (9): 1111-1134, 2007.
[2] K. Han, Y. T. Feng and D. R. J. Owen. Modelling of Thermal Contact
Resistance Using Thermal Lattice Boltzmann Method. Journal of Inter-
national Journal of Thermal Sciences. 47(10): 1276-1283, 2008
[3] Y. T. Feng, K. Han, C. F. Li, D. R. J. Owen. Discrete thermal element
modelling of heat conduction in particle systems: Basic Formulations.
Journal of Computational Physics. 227: 5072-5089, 2008.
[4] Y. T. Feng, K. Han, D. R. J. Owen. Combined three-dimensional lattice
Boltzmann method and discrete element method for fluid-particle prob-
lems: computational issues and experimental assessment. International
Journal for Numerical Methods in Engineering, 81(2):229-245, 2010
[5] Y. T. Feng, K. Han, D. R. J. Owen. Discrete thermal element modelling
of heat conduction in particle systems: Pipe-network model and transient
analysis. Powder Technology, 193(3):248-256, 2009.
[6] K. Han, Y. T. Feng and D. R. J. Owen. Three Dimensional Modelling
and Simulation of Magnetorheological Fluids. International Journal for
Numerical Methods in Engineering, Published Online: May 25 2010, DOI:
10.1002/nme.2940
[7] Y. T. Feng, C. F. Li, K. Han, D. R. J. Owen. Discrete thermal element
modelling of heat conduction in spherical particle systems. The current
proceedings, 2010.

ALGORITHMS AND SOLVERS

MRCK_3D CONTACT DETECTION

ALGORITHM
E. Rougier †
Geophysics Group, Los Alamos National Laboratory, Los Alamos, New Mexico,
USA 87545
A. Munjiza
Department of Engineering, Queen Mary, University of London, U.K, E1 4NS.
Large-scale FEM-DEM and DEM simulations involving contact

of a large number of separate bodies need an efficient, robust
and flexible contact detection algorithm. In this work the
MRCK_3D search algorithm is outlined and its main CPU
performances are evaluated. One of the most important aspects
of this newly developed search algorithm is that it is applicable
to systems consisting of many bodies of different shapes and
sizes.
INTRODUCTION
The Combined Finite-Discrete Element Methods (FEM-DEM) and the
Discrete Element Methods (DEM) are characterized by systems consisting of a
large number of separate (distinct) bodies often called discrete elements. Discrete
elements are usually free to move in space and time and thus interact with each
other. Routine FEM-DEM and DEM problems may involve millions of
interacting discrete elements. In this context a fast, efficient and flexible, i.e. able
to deal with particles of general shape, contact detection algorithm is necessary. In
other words, the effective solution of large-scale real world problems relies upon a
robust contact detection algorithm being employed. However, the optimal contact
detection algorithm is usually dependent upon the problem to be solved. The
properties of different contact detection algorithms make them suitable for
different types of problems such as dense packing and loose packing or quasi-
static problems (where relative motion of individual bodies is restricted) and
†
Correspondence to E. Rougier, Geophysics Group, Los Alamos National Laboratory, MS-D408,
Los Alamos, New Mexico, USA 87545. email: [email protected]

dynamic problems1,2,4,5. Real problems, where contact detection is a big issue, are
dynamic problems, consisting of large numbers of discrete elements that are free
to move significantly3. In this context, one of the more important requirements for
a contact detection algorithm is the minimization of the total detection time T, i.e.
the total CPU time needed to detect all couples close to each other. This is a key
issue when real world problems are to be simulated. In this paper the CPU time
performance of the MRCK_3D contact search algorithm is presented for an ideal
case involving particles of general shape.
MRCK_3D SHORT DESCRIPTION

The MRCK_3D is a space based contact detection algorithm developed for the
general FEM-DEM framework, but also applicable to DEM simulations. In the
following paragraphs a short description of the algorithm is given. The detailed
description of this algorithm is outside of the scope of this work and will be
presented in future publications.
The first step of the algorithm is to decompose the simulation space onto cells,
as shown in Figure 1. Given any two particles, one called contactor and the other
one target, the contact search problem is reduced to determine whether the
contactor and the target share at least one cell. If they do, the couple is reported
back to the interaction routine for processing the actual contact.
(4, ny - 1, 6) (nx - 1, ny - 1, 0)
ymax
(nx - 1, 7, 2)
y (6, 4, nz - 1)
x
zmin
(0, 0, nz - 1)
z
ymin zmax
xmin xmax
Figure 1. Cell based space decomposition.

The following step in the MRCK_3D algorithm is to discretize the boundaries
of the target particle onto a set of points and then to render those points onto the

cells. The point rendering process is performed by simply calculating the

integerized coordinates of the points as follows:
 x  xmin   y  y min   z  z min  (1)
ii  int  i  ; ji  int  i  ; k i  int  i 
 d   d   d 
where ii, ji, ki are the integerized coordinates of the point i; xi, yi and zi are the
actual coordinates of point i; xmin, ymin and zmin are the minimum coordinates of the
system and d is the cell size. The integerized coordinates of the target points are
stored into three arrays. These arrays are ordered according to the MR space
ordering criterion1,2. The first ordering, istep = 0, is done using a Binary Sort
routine. For istep > 0, i.e. when the particles have moved, the ordering of the
integerized coordinate arrays is done using the MR-Linear Sort criterion, i.e.:
exploiting the temporal coherence of the list.
It is worth noting that in this work all the particles are discretized onto
tetrahedrons. However, the algorithm is also prepared to handle particles
discretized into hexahedrons. If needed, further extensions to more complex
particle shapes can also be implemented.
The rendering of the contactor particle is done on a tetrahedron by tetrahedron
basis. First each tetrahedron is expanded outwards by a buffer distance given by
dbuff  3d (2)
in order to ensure that no contact is missed during the search. Following the
expansion, the tetrahedron is divided onto regions along the z axis. Each region is
decomposed onto slices along the y axis. Finally, each slice is decomposed onto
columns along the x axis, see Figure 2. The MR spatial ordering criterion is
exploited to improve the performance of the search algorithm. For each region the
maximum and minimum integerized z coordinates are calculated,
k min , k max (3)
The target z integerized coordinate array is searched using binary search routines
to determine whether there is at least one point with integerized coordinate k that
satisfies
k min  k  k max (4)

Expanded
Tetrahedron
kmax
imin
z kmin imax
y
jmin jmax
x
Region Slice Column
Figure 2. Rendering: Decomposition of a tetrahedron onto regions, slices and
columns.
If the answer is no, the rendering process for the region is stopped, i.e. there is
no contacting couple to report for this region, and the next region is processed.
However, if there is at least one point that satisfies Eq. (4), the rendering process
for the region has to continue.
The region is divided onto slices along the y axis, see Figure 2, and a process
similar to the one described above is followed, i.e. the maximum and minimum y
integerized coordinates for the slice are calculated
jmin , jmax (5)
The target y integerized coordinate array is searched using binary search
routines to determine whether there is at least one point with integerized
coordinates j and k that satisfy Eq. (4) and Eq. (6)
jmin  j  jmax (6)
If the answer is no, the rendering process for the slice is stopped, i.e. there is
no contacting couple to report for this slice, and the next slice is processed.
However, if there is at least one point that satisfies Eq. (4) and Eq. (6), the
rendering process for the slice has to continue.
Each slice is divided onto columns along the x axis, see Figure 2. The
maximum and minimum x integerized coordinates for the column are calculated
imin ,imax (7)

All of the points that satisfy Eq. (4), Eq. (6) and Eq. (8) are reported as
potential contacts to the interaction routine.
imin  i  imax (8)
It is worth noting that during the search process only the target cells are
remembered. The contactor cells are calculated on the fly and discarded once the
search for each contactor tetrahedron is finished.
NUMERICAL EXAMPLES
The performance evaluation of the MRCK_3D is done with the help of a
cubical hollow raster of particles of general shapes placed inside a rigid spherical
container, as shown in Figure 4. In this work, four different types of particles of
general shapes were selected as shown in Figure 3.
A B C D
Figure 3. Particles of general shape. Particle A: 128 tetrahedrons, Particle B: 184

tetrahedrons, Particle C: 160 tetrahedrons and Particle D: 187 tetrahedrons.
Figure 4. Left: Full 3D hollow raster. Right: Sliced view of the 3D hollow raster
inside the rigid spherical container.
The raster is centered with respect to the spherical container. There is no
initial overlap between the particles. Each particle is given an initial velocity
pointing to the center of the sphere. The speed of each particle is set to 100 m/s.

The system is left to relax, and as a result, a random pattern of collisions takes
place between the particles, as shown in Figure 5.
Figure 5. Time evolution of the collapsing of the raster of particles inside the rigid
spherical container. For the sake of clarity the spherical container was filtered out
from the frames.
The cell size needs to be specified to the MRCK_3D algorithm. The level of
sensitivity of the MRCK_3D search algorithm to the change of the cell size is
shown in Figure 6. In the Figure d is the cell size and lmin is the length of the
smallest edge on the system. As the cell size is increased, the performance of the
MRCK_3D approaches that of the Quadratic Search algorithm, i.e.: all the
particles are contained in one cell and all couples are reported as potential

contacts. On the other hand, if the cell size is chosen to be too small, a
degradation of the performance of the search algorithm is noticed. This is due to
the fact that the contactor has to be decomposed onto a much larger number of
cells.
In the simulations presented in this paper the cell size was set to the value
shown in Eq. (9)
d (9)
 0.1
lmin
The cell size shown in Eq. (9) is not the optimum point for the search
algorithm alone, but it is the optimum point for the combined search and
interaction algorithms.
The specific CPU time, i.e. seconds of CPU per Nstep and per Ntet (where Nstep
and Ntet are the total number of time steps run and the total number of tetrahedrons
present in the system), consumed by the search routines is shown in Figure 7.
From the figure it can be seen that the time consumed in rendering the target
points is almost an order of magnitude smaller than the time needed to render the
contactors.
160
140
Search
CPU Time [sec of CPU Clock]
120 Interaction
Search + Interaction
100
80
60
40
20
0
0.001 0.010 0.100 1.000 10.000
d / l min
Figure 6. CPU time as a function of the cell size.

3.0E-02
Total Search
Specific CPU Time (s / Ntet / Nstep )

2.5E-02 Render Target
Render Contactor & Find Contacts
2.0E-02
1.5E-02
1.0E-02
5.0E-03
0.0E+00
0.0E+00 5.0E+05 1.0E+06 1.5E+06
N tet
Figure 7. Specific CPU time as a function of the number of tetrahedrons. Time

taken in an Intel Xeon E5540 2.53 GHz CPU.
CONCLUSIONS
The newly developed MRCK_3D preserves excellent properties of NBS,
CGRID and MR search algorithms. It is worth noting that one of the main
advantages of MRCK_3D algorithm over NBS and MR algorithms is that it can
handle particles of any shape and size. This is a key aspect within the combined
FEM-DEM method framework. The combined FEM-DEM method is often
applied to real world problems where particles can undergo fracture and
fragmentation along the process, creating new sub-particles of very different
shapes and sizes. The MRCK_3D is designed to handle these real world
phenomena without difficulties.
The changes of the CPU time with the number of tetrahedrons present in the
system and also the sensitivity of the algorithm to the cell size have been
demonstrated. However, it is outside the scope of this paper to demonstrate some
other nice properties of the MRCK_3D algorithm, i.e: CPU efficiency as a
function of the packing density, etc.

ACKNOWLEDGEMENTS
The work presented on this paper was done as part of a collaboration with Ted
Carney and Earl Knight from the Geophysics Group (EES-17) of Los Alamos
National Laboratory, Los Alamos, NM, USA.
REFERENCES
1. A. Munjiza, E. Rougier and N.W.M. John, ‘MR linear contact detection
algorithm’, Int. J. Numer. Meth. Eng., 66(1): 46−71, 2006.
2. A. Munjiza, E. Rougier and N.W.M. John, ‘MR-linear contact detection
algorithm for fragmenting solids.’, Advances in Boundary Element
Techniques V, Lisbon Portugal, Published by EC Ltd, UK, 279−286, 2004.
3. A. Munjiza, The Combined Finite-Discrete Element Method, Published by
John Wiley and Sons LTD, New York, 2004.
4. A. Munjiza and K.R.F. Andrews, ‘NBS contact detection algorithm for
bodies of similar size’, Int. J. Numer. Meth. Eng., 43: 131−149, 1998.
5. J.R. Williams, E. Perkins and B. Cook, ‘A contact algorithm for partitioning
arbitrary sized objects’, Eng. Computations, 21(2/3/4): 235-248, 2004.

POLYHEDRA ON THE CHEAP
Mark A. Hopkins
US Army Cold Regions Research and Engineering Laboratory
A new approach for discrete element modeling using dilated

polyhedral particle shapes is discussed. This technique is suitable for
non-convex polyhedral shapes that are based on the dilation process
from mathematical morphology. An arbitrary polyhedral shape is
dilated by placing the center of a sphere of fixed radius at every point
on the surface. Although the dilated vertices become sphere segments
and the edges become cylinder segments, the polyhedra can be made
arbitrarily sharp by limiting the dilation radius. Defining a contact
between two dilated polyhedra is equivalent to determining which
spheres amongst the infinite number that cover the surface of each
object is in contact. The algorithm is based on the idea that contacts
between two dilated polyhedra can be grouped into three categories;
vertex on surface, vertex on edge, and edge on edge.
INTRODUCTION
Particle shapes used in discrete element modeling include disks and spheres,
ellipsoids and poly-ellipsoids, and polygons and polyhedra. Contact detection ranges
from trivial for disks and spheres to non-trivial for ellipsoids, polygons, and

polyhedra. Ellipsoidal contact detection has been much discussed. Hopkins (2004)
presents a simple approach for dilated ellipsoids. Peters et al. (2009) presents a
similar approach for contact detection between poly- or asymmetrical ellipsoids.
Cundall (1988) presents an algorithm that places a common plane that bisects the
space between a pair of polyhedra. If the particles are in contact then both intersect
the common plane. Nezami et al., (2006) present a faster approach for establishing a
common plane that they call the shortest link method. All of these methods are
iterative.
This work presents an approach that makes a trade-off between adding extra
particles in the form of a sphere at each vertex and a cylinder at each edge for a
trivial, non-iterative solution for the vertex on surface, vertex on edge, and edge on
edge contacts between two polyhedral particles. The penalty paid for this
simplification is a drastic increase in the number of possible contacts that must be
searched. However, for slowly deforming systems this penalty can be significantly
reduced by using appropriate filters and increasing the interval between successive
searches. The list of actual and pending contacts that are compiled during a search is
fairly optimal, especially when one considers the necessity for multiple contacts
points (or the equivalent) for the contact between two flat parallel particle surfaces.
Furthermore, since the contact detection works between vertices and edges and
triangular surfaces, the polyhedral particles may be non-convex as long as they are
comprised of triangular surfaces.
The second important feature of the approach discussed here is the use of dilated
particles. In the dilation process in mathematical morphology (Serra, 1986) an
arbitrary shape is dilated by placing the center of a sphere of fixed radius at every
point on the surface of the basic shape. For example dilating a line segment results in
a cylinder with hemispherical ends and dilating a two-dimensional flat disk results in
a three-dimensional disk. This feature ensures that the vertex or edge of one undilated
polyhedral particle never penetrates the surface of the other undilated polyhedral
particle. (If this happens the code crashes.) This means that we never have to ask
whether the vertex is inside the other polyhedral particle. We only have to ask

whether the closest point between the undilated polyhedral particles is less than the
sum of the two dilation radii. Similarly, we don’t have to ask whether an edge of one
undilated polyhedral particle is inside an edge of the other undilated polyhedral
particle merely whether the shortest distance between the lines is less than the sum of
the two dilation radii. This simplifies the process significantly.
CONTACT DETECTION AND MECHANICS

The code begins by reading in a list of polyhedral particles that includes the
location of the center of mass and orientation, the vertex vectors, the number of
triangular surfaces, edges, and vertices, and the indices of the vertices that define the
edges and triangular surfaces. Next the program defines a set of quasi particles for the
surfaces, edges, and vertices. The quasi particles are polygons, cylinders, and spheres,
respectively. The polyhedra are placed in an ordinary three-dimensional search grid.
In the search algorithm the spheres and cylinders examine the grid around their
location for neighboring polyhedra. For each neighboring polyhedra each possible
sphere on polygon, sphere on cylinder, and cylinder on cylinder contact is examined.
Various simple filters are effective in drastically reducing the number of pairs that are
examined.
In a sphere on polygon contact we ascertain whether or not a perpendicular from
the sphere to the plane of the triangular polygon lies within the triangle. If it is within
the triangle and the length of the perpendicular is less than sum of the two dilation
radii plus some epsilon then the pair (the sphere and the polygon) is stored in the list
of pending contacts. This process is implemented by dropping a perpendicular from
the sphere to the plane of the triangle. The in-plane location of the point is
transformed to the body coordinate frame of the triangle. The first edge vector of the
triangle is crossed into the vector from the first vertex of the edge vector to the point.
If the cross product is positive then the point is inside that edge. This process is
repeated for the other two sides. If all cross products are positive then the point lies
inside the surface and contact is possible. The normal from the surface to the sphere
defines the plane of contact. We are not concerned with contacts with the edge of the

triangle since these will be captured by the sphere-cylinder algorithm.

In a sphere on cylinder contact we
ascertain whether or not the shortest
distance between the sphere and the line
segment that defines the edge is less than
sum of the two dilation radii plus some
epsilon. If so then the pair (the sphere and
the cylinder) is stored in the list of
pending contacts. This process is
implemented by making a vector from the
midpoint of the cylinder axis to the
sphere. The component perpendicular to
the axis of the cylinder is subtracted
leaving the component parallel to the axis.
If this component extends beyond either
end of the cylinder it is truncated to that
end point. If the vector from end of the
resulting vector to the sphere is less than
sum of the two dilation radii plus some
epsilon then the pair (the sphere and the
cylinder) is stored in the list of pending
contacts. Furthermore, that vector defines
the plane of contact.
In a cylinder on cylinder contact we
ascertain whether or not the shortest
distance between the line segment that
defines one edge and the line segment that
defines the other edge is less than sum of Fig 1. Falling polyhedra.
the two dilation radii plus some epsilon. If
so then the pair (the two cylinders) is

stored in the list of pending contacts. This process is implemented using one of the
many existing algorithms (e.g., Sunday, 2010) for finding the closest distance
between two line segments in three-dimensional space.
The list of pending particle contacts between pairs of particles is cycled through at
each time step. Each pair is examined for contact using the algorithms described
above. The normal and tangential components of the interparticle force are calculated
from a modified Hertzian model with parallel viscous damping and a Coulomb
friction cap on the tangential force. The goal of this effort is to develop a model of a
triaxial cell filled with a model soil having polyhedral grains.
REFERENCES
1. Cundall, P.A., Formulation of a three-dimensional discrete element model, Int. J.
Rock Mech. and Mining Sci. and Geomech. Abstracts, 25(3): 107-116, 1988.
2. Hopkins, M.A., Discrete element modeling with dilated particles, Engng.
Comput., 21(2), 422-430, DOI: 10.1108/02644400410519866, 2004.
3. Nezami, E.G., Y.M.A. Hashash, D. Zhao, and J. Ghaboussi, Shortest link method
for contact detection in discrete element method, Int. J. Numer. Anal. Meth.
Geomech., 30: 783-801, 2006.
4. Peters, J., M.A. Hopkins, R. Kala, and R.E. Wahl, A poly-ellipsoid particle for
non-spherical discrete element method, Eng. Computations, 26, Issue 6, 645-657,
DOI: 10.1108/02644400910975441, 2009.
5. Serra, J., Introduction to Mathematical Morphology, Computer Vision, Graphics,
and Image Processing 35, 283-305, 1986.
6. Sunday, D., Distance between Lines and Segments with their Closes Point of
Approach, http://softsurfer.com/Archive/algorithm_0106/algorithm_0106.htm,
2010.

AN EVENT DRIVEN DISCRETE ELEMENT

MODELLING APPROACH
Y. T. Feng, K. Han, D. R. J. Owen

Civil & Computational Engineering Centre, School of Engineering, Swansea University, UK
Abstract: The current work presents an asynchronous/multiply time

stepping integration scheme for the simulation of dynamic systems
exhibiting multi-fracture and discrete features. It is a novel extension
of existing methodologies to the discrete element simulation and sev-
eral fundamental developments in both theoretical and implementation
aspects are achieved that are essential for the success of the approach.
It is shown that the resulting AVI/MTS discrete element method can
achieve a nearly optimal speed-up for typical discrete systems. How-
ever, the stability is the issue that needs to be further addressed
INTRODUCTION
Significant advances have been made over the last decade in the modelling of
practical problems that exhibit strong discrete phenomena by employing discrete
element solution strategies [1]. Typical examples include granular/particulate ma-
terials in the processing industry and geo-mechanics. The highly nonlinear contact
dynamic/impact nature of these systems dictates that explicit time integration,
such as central difference schemes, are almost the only feasible option to be em-
ployed to compute the system dynamic response. The stability issue associated
with an explicit time integration, however, imposes a severe constraint on the

maximum time step that can be used. In many discrete particulate systems, how-
ever, a wide range of particulate size distribution must be employed so that the
underlying physical features can be captured to a certain degree. Normally, the
smallest element controls the maximum time-step that can be employed, leading
to extremely intensive computations involved in the simulations.
There has been a continuous interest over the last two decades in the develop-
ment of time integration schemes that can inherently exploit different time scales
in finite element analysis of many science and engineering problems. In addition
to the cases of the finite element analysis of structural dynamic systems where the
sizes of elements can be significantly different, which lead to different dynamic
scales of the elements, multi-scale dynamic features also exist in problems where
highly heterogenous materials and/or different physical phenomena are present.
Since 1980’s, some effort has been made in the finite element community to
develop so called multiple/sub-cycling time step (MTS) methods [2] where larger
time steps may be used for the regions with larger finite elements. More recently,
a more flexible approach, termed the Asynchronous Variational Integrator (AVI)
[3], is proposed within the general framework of symplectic integrators for con-
servative (Hamiltonian) systems. Limited numerical examples have demonstrated
their potential for significantly reducing computations in structural dynamic
analysis. Note that similar MTS methods have also been proposed in molecular
dynamic simulations [4].
The current work attempts to extend both MTS and AVI to discrete element
simulations so that a substantial improvement in terms of computational costs can
also be gained for the problems concerned. However, the extension is not trivial,
due primarily to the nature of the continuously changing configuration of the dis-
crete systems. This is essentially equivalent to the development of an “event-
driven” solution approach for discrete elements. Several fundamental develop-
ments are therefore needed to overcome the difficulties. In addition, theoretical
and implementation issues will be identified that are essential for the success of
the approach. It will be shown that the resulting AVI/MTS based “event-driven”
discrete element solution approach can achieve a nearly optimal speed-up for typi-
cal discrete systems. However, the stability is the issue that needs to be further
addressed.

MTS/AVI APPROACHES FOR FINITE ELEMENT SYSTEMS

Node/Element Update: AVI is so called element-partition based while
MTS/subcycling approaches are node-partition based and there exist several vari-
ants. For their suitability for discrete systems currently concerned, only the non-
integer ratio constant-velocity subcycling scheme [2] is considered. The
node/element update procedures for both MTS and AVI are illustrated in Fig.1.
Importantly, both approaches can be viewed as an “event-driven” time stepping
scheme, whereby each time step is treated as an “event”.
4 4
3 ∆tm 3
4
2 ∆tm 2 3
∆t e 2
3∆tm
∆t e1
1 2∆tm 1 2
(a) (b)
Fig.1. Node/element update: (a) MTS/Subcycling – constant velocity; (b) AVI
Stability: One important issue associated with MTS and AVI is their numerical
stability in comparison with the original single time step counterpart (the central
difference algorithm). The stability is examined numerically for simple spring-
mass models as shown in Fig2. For the model (a), the critical time step associated
with each node (mass) is
∆t1cr = 2 / ω1 = 2 m1 k1 = 2; ∆t2cr = 2 / ω2 = 2 m2 k2 = 2 m2 (1)

Assuming m2 ≥ 1 , the actual time steps for the two nodes in MTS are chosen to be
∆t1 = λ∆t1cr ; ∆t2 = n∆t1 (interger n = [ω1 / ω2 ]) (2)
where 0 < λ < 1 is the time step factor. Fig.1(a) depicts the (capped) energy in-
crease of MTS for the system after 1000 time steps for different values of λ and
nodal frequency/time ratios ω1 / ω2 , revealing the existence of a complex unstable
region within the standard stable limit. A similar result is also obtained for AVI
for the model (b) and is shown in Fig.2(b). In this case, m1 + m2 = 1 . By varying
the ratio of m1 / m2 , different element frequency ratios can be achieved. Again a
complex unstable region is emerged. Note, however, that the unstable region can
be reduced by introducing the system damping.
These results indicate that the stability issue for both MTS and AVI is more
pronounced than the central difference scheme commonly used.
m1=1 m2 m1 m2
k1=1 k2=1 k3=1 k1=1 k2=1
(a) (b)
Fig.2. Numerical stability: (a) MTS/Subcycling – constant velocity; and (b) AVI

MTS/AVI BASED EVENT-DRIVEN APPROACHES FOR DSICRETE

ELEMENT SYSTEMS
Algorithmic issues
MTS and AVI methods for finite element systems cannot be directly applied to
discrete element cases without important modifications to be made to the discrete
element method. The difficulty lies in the feature of continuous changing configu-
rations of discrete systems which requires the constant update of the contact list
between discrete objects. As both MTS and AVI advance only a (small) part of
the objects at each step, a global search algorithm should be able to effectively
deal with this situation. Unfortunately, most of currently used contact search
methods are not suitable for this purpose and a new effective approach has been
development. In addition, the improved computational efficiency can be achieved
in MTS and AVI mainly by the reduction of the number of internal force compu-
tations, but at the expense of an increased overhead for additional data processing.
To maximise the speed-up several operations are agglomerated at each internal
force calculation.
Priority queue for AVI

Although AVI has an advantage over MTS for its flexibility to choose desir-
able time steps for elements, an efficient scheduling procedure, or a priority
queue, has to be applied to determine the order of element operation. The ele-
ments in the priority queue are ordered according to the next time when they are
to become active. The order of complexity of maintaining a priority queue is how-
ever N log N which may lead to a possible computational bottleneck for large
scale applications. To overcome this problem, a bin-type algorithm is imple-
mented which regroups the elements into a limited number of bins and thus avoids
the use of priority queue and also provides a unified implementation of MTS and
AVI.
NUMERICAL ILLUSTRATIONS
Both MTS and AVI approaches developed are numerically assessed by two test
cases: hopper filling and particle compaction, with 3000 discrete particles. The

element size is distributed evenly in the range of [1,10] and both disks and ellipses
are considered. The speed-up of the methods over the standard discrete element
approach (i.e. the central difference scheme based) is listed in Table 1. Note that
the maximum speed-up that can be achieved is about 3.40. A higher speed-up can
be expected for a wider range of particle size, while a lower speed-up will be
achieved otherwise. Also no significant difference is observed between MTS and
AVI, although the former is slightly more efficient than the latter.
Case No of Particle Speed-up

particles shape size MTS AVI
Case 1 3000 disk 2.15 2.01
ellipse 3.23 3.12
Case 2 3000 disk 2.26 2.12
ellipse 3.24 3.15
Table 1 : Speed-up of MTS and AVI for two test cases
REFERENCES
1. D. R .J. Owen, Y. T. Feng and E. A. de Souza Neto, “The modelling of multi-
fracturing solids and particulate media”. Int. J. Numer. Meth. Engng., 60(1):
317−340 (2004).
2. M. O. Neal and T. Belytschko, “Explicit-explicit sub-cycling with non-integer
time step ratio for structural dynamic systems”. Computers & Structures, 31:
871−880 (1989).
3. A. Lew, J. E. Marsden, M. Ortiz and M. West, “Variational time integrators”,
Int. J. Numer. Meth. Engng., Vol. 60(1):153−212 (2004).
4. T. Schlick, R. D. Skeel, A. T. Brunger, L. V. Kale, J. A. Board Jr., J. Hermans
and K. Schulten, Algorithmic challenges in computational molecular biophys-
ics. J. Comp. Physics. 151: 9-48 (1999)

DEVELOPMENT OF A ROTATIONAL RESISTANCE

MODEL IN THE DISCRETE ELEMENT METHOD
Yoshinori yamada1, Mikio Sakai1, Masao Tsuchiya2, Shuichi Hirayama2

1 The University of Tokyo, School of Engineering
2Kinzoku-Giken Co, Ltd.
The Discrete Element Method (DEM) is widely used in computational

granular dynamics. The DEM is a Lagrangian approach where indi-
vidual particles are calculated based on the Newton’s second law of
motion. Hence, the DEM enables us to investigate the granular flow
characteristics at the particle level. Spherical particles has been used
in most of the DEM simulations because the contact judgment is easy.
On the other hand, the particle shape might affect the powder flowa-
bility. When effect of the particle shape on the flowability cannot be
neglected, the particle shape was modeled by meshes or cluster of cal-
culated particles in the past study. However, the calculation cost of
these models became too expensive. As a result, these models could
not be applied in large scale powder systems. Thereat, a new modeling
which gives the effect of particle shape is developed in the present
study. This model is called rotational resistance model. The rotational
resistance based on the particle shape is modeled in the DEM. This
model can simulate the non-spherical particle behavior without exces-
sive calculation cost. The adequacy of this model is proved by expe-
riment in this study.

INTRODUCTION
Granular materials are used in many industrial processes, e. q., fluidized bed,
pneumatic conveying system. In these processes, it is important to control these
granular materials behavior. In order to investigate the complicated phenomena
related to the granular flows, numerical analysis is a promising approach.
The Discrete Element Method (DEM)[1] is widely used in many granular sys-
tems. The DEM is a simulation method that calculate each granular particle dy-
namics based on Newton's second law. Although the DEM can calculate granular
dynamics with a high resolution, it requires much calculation cost. Hence, in the
DEM simulation, granular particle shape is often assumed to be sphere in order to
reduce calculation cost. On the other hand, actual particle shape is non-spherical.
The particle shape might affect the flowability. The effect of the particle shape on
the flowability might be important. In the past studies, the non-spherical particle
was modeled by connecting some spherical particles[2] or using mesh[3]. These
modeling need excessive calculation cost, so that they could not be applied in
large scale systems.
In this study, we develop a rotational resistance model to calculate the non-
spherical particle behavior without excessive calculation cost. In this study, the
adequacy of the rotational resistance model is shown by numerical examples and
experiments.
DEVELOPMENT OF A ROTATIONAL RESISTANCE MODEL

In this study, we develop a rotational resistance model into DEM in order to
consider the non-spherical shape. In this model, we use spherical particle. The ef-
fect of the non-spherical particle is considered by the non-spherical particle pose.
The spherical particle has virtual vertexes (Fig. 1). The contact force act on the
nearest vertex. After that, the torque from contact force is calculated. If two par-
ticle contact condition is regarded to be surface to surface, we calculate additional
torque based on viscous damper. The difference between the ordinal spherical
model and the rotational resistance model is only a torque calculation.

Fig. 1 : Contact point and torque from contact force

Partition
plate
70mm
70mm
Fig. 2 : Initial condition of simulation
NUMERICAL SIMULATIONS AND EXPERIMENTS

In this study, we performed two-dimensional simulation of granular column
breaking.
The initial condition of simulation is shown in Fig. 2. The height of the col-
umns is 70mm and the length is 70mm. We moved the left-side partition plate to
upper side by acceleration 9.8m/s2.
We performed 2 case simulations about original spherical model and using a
rotational resistance model. Physical properties are shown in Table 1.
We performed 2 case experiments that correspond to their simulations. The
height and length of column is the same and the depth is 100mm. We chose Alu-
mina particle as spherical particles Case 1 and quartz sand as non-spherical par-
ticles Case 2.

Table 1 : Physical properties

Case 1 Case 2
Rotational Resistance model × ○
Number of vertex - 6
Particle diameter 3mm 3mm
Spring constant 1000N/m 1000N/m
Restitution coefficient 0.8 0.8
Friction coefficient 0.3 0.3
3
Density 3900kg/m 3900kg/m3
(a) Simulation result : Case 1 (b) Simulation result : Case 2

Fig. 3 : Typical snapshots of the simulation results

Figure 3 shows the snapshots of the simulation results. The mass of avalanche
granular of Case 2 is less than that of Case 1.
Figure 4 show the snapshots of the experiment results. The mass of avalanche
granular of Case 2 is less than that of Case 1.

(a) Experiment result : Case 1 (b) Experiment result : Case 2

Fig. 4 : Typical snapshots of the experiment results
Table 2 : Angle of repose
Case 1 Case 2
Simulation 19.8° 33.5°
Experiment 20.9° 32.3°
Table 2 shows the angle of repose of the simulation results and the experiment
results. In simulations, the angle of repose of Case 2 is 13.7° larger than the one
of Case 1. These results qualitatively show that the rotational resistance model can
simulate non-spherical particle behavior. In experiments, the angle of repose of
Case 2 is 11.4° larger than the one of Case 1. These results show that non-
spherical particles are less flowable than spherical particles.
The differences of the angle of repose between the simulation results and the
experiment results are 1.1° in Case 1 and 1.2° in Case 2. The differences are
enough smaller than one between the simulations and the experiments. These re-
sults quantitatively show that a rotational resistance model can simulate non-
spherical particle behavior.
Table 3 shows calculation time about Case 1 and Case 2. Case 2 took about

Table 3 : Calculation time

Case 1 Case 2
Calculation time 41.67s 54.67s
30% larger than that Case 1 did. This additional calculation costs seems to be
much less than previous model like using some particle or mesh to model the non-
spherical particle.
CONCLUSION
In the past studies, non-spherical particle is modeled by some spherical particle
or mesh, so that the calculation cost became excessive. In order to resolve this is-
sue, we developed a rotational resistance model to introduce the effect of the non-
spherical particle movement. The adequacy of this model was proved by the expe-
riments. The calculation time of simulation with or without the rotational resis-
tance model was measured and compared.
In the future, we are going to apply the rotational resistance model to large-
scale powder systems.
REFERENCES
1. Cundall, P. A. and O. D. L. Strack, ‘A Discrete Numerical Model for Granular
Assembles’, Geotechnique, 29, 47-65, 1979.
2. Matsusima, T. and H. Saomoto, ‘Discrete Element Modeling for Irregularly-
shaped Sand Grains’, Proc:NUMBE2002 : Numerical Methods in Geotech.
Eng., 239-246, 2002.
3. Latham, J. P., A. Munjiza, J. Mindel, J. Xiang, R. Guises, C. C. Pain, G. Gor-
man and X. Garcia, ‘Modelling of massive particulates for breakwater engi-
neering using coupled FEMDEM and CFD’, Particuology, 6, 572-583, 2008.

DESCRIPTION OF ROTATION IN THE MOVABLE

CELLULAR AUTOMATON METHOD
Aleksey Yu. Smolin, Nikita V. Roman, Serguei G. Psakhie

Institute of Strength Physics and Materials Science SB RAS, Tomsk, Russia
The problem of rotation description in particle methods is discussed. It

is shown that taking into account rotation either as a degree of free-
dom or via motion of the neighboring particles allows to describe
classical isotropic continuum. To describe more complex media it is
required to realize the both ways of rotation description and to consid-
er rotational degree of freedom as independent rotation.
INTRODUCTION
Rotation takes place in the nature at any scale level from elementary particles
to galaxies. In the movable cellular automaton (MCA) method, a discrete method
for simulation material behavior at meso- and macroscale, rotation is taken into
account as an independent degree of freedom of an automaton in addition to trans-
lation motion of its centroid [1]. When automaton size tends to zero, the MCA
method would describe some continuum. It was shown that using linear response
function for automata tends to the Hook's law for the result continuum when au-
tomaton size vanishes [2]. But rotation was not taken into account in that paper
and therefore it has been still undetermined what a continuum the MCA method
describes when automaton size vanishes: a classical one, where an elementary
point has no rotational degree of freedom or a micropolar one, where an elementa-
ry volume has nonzero inertia moment and is characterized by independent rota-
tion [3].

TAKING INTO ACCOUNT ROTATION AS INDEPENDENT DEGREE OF

FREEDOM IN THE FRAMEWORK OF MCA METHOD
Description of rotation in the MCA method is adopted from discrete element
method [4]. At that there is a moment of forces acting among automata and ap-
plied in their contacts in the right side of the Euler equation (for rotational veloci-
ty). Let us consider calculation of the force resisting shear deformation (main tan-
gential force) in MCA.
As is well known, any two points i and j of rotating rigid solid has the same ro-
tational velocity and their translational velocities are described by the formula
j i ij rij , where rij (r j ri ) , ri is radius-vector of the point i, ij is rota-
tional velocity of the rigid body. We can apply this formula for an automaton pair
i–j toget rotational velocity ij of the pair as a hole. If premultiply both sides of
the above equation by vector rij and neglect rotation about the pair axis (i.e.
ij rij 0 ), then one can obtain the following formula ij nij ( j i ) rij .
Instantaneous rate of shear strain of automaton i under its interaction with au-
tomaton j can be defined as a ratio of tangential velocity in the contact point to the
distance qij from the center of automaton i to its contact with automaton j:
sh
ij q ij ( ij i) nij q ij ij i nij , where i is rotational velocity of
automaton i. Shear strain of automaton i at time T is defined by the integral
sh
ij ij (t )dt ij i (1)
where ij is vector of rotation of the pair axis relative to the initial position.
For small deformation we can use Hook's law to compute the resistance force:
e
f ij (Gi ij Gj ji ) nij (2)
where Gi is the shear modulus of the material of automaton i. Eq. (2) describes the
pair tangential force produced by shear strain of each automata of the pair.

From equations (1) and (2) one can see, that in case of rotation of automata of
the pair with velocities of the same value but in the opposite directions ( i j ),
the shear resisting force (2) does not prevent such motion. But the medium, de-
scribed by such automata is subjected to deformation and has to resist it by the
torque proportional to i j . This means that for correct describing of motion of
the media, not an ensemble of particles, it is required to add such torque which
resists to relative rotation of the elements to the moment of inter-element forces.
TAKING INTO ACCOUNT ROTATION AS COOPERATIVE MOTION

OF MEDIUM ELEMENTS
Numerical methods of classical continuum mechanics give solution of motion
equations as displacement or translation velocities of mesh nodes. Rotations in the
medium can be computed from consideration of distortion of the mesh cells con-
taining the node examined. Obviously, this approach should allow computing ro-
tation in the MCA method also. To do so, rotation velocity of an element may be
considered as an average rotation velocity of all its neighbors interacting with it.
In 3D the problem how to determine such average rotation may be formulated as
follows. There are points j=1..N, defined by vectors rij , starting at the fixed point
i. Each point j moves relative to point i with velocity ij . This motion can be
represented as rotation of a solid having points j about point i. Real motion of
each point j is a sum of such average rotation of a solid and deformation of this
solid. The problem: to determine rotational velocity i of this solid.
Computations with rotation taken into account in such way showed that in this
case we have the same behavior as in case of taking into account rotation explicit-
ly. At least compression–tension tests, shear and the Lamb's problem did not re-
veal any significant dissimilarity.
ADVANTAGES OF TAKING INTO ACCOUNT ROTATION AS

INDEPENDENT DEGREE OF FREEDOM.
In spite of similarity of the results in both approaches it should be noted that
taking into account rotation as independent degree of freedom allows modeling

more complicated medium with implicit describing its internal structure. For ex-
ample, it was shown that stress tensor for such medium in general is a nonsymme-
trical one, as it should be for a Cosserat medium with bending torsion. More of
that, it is possible to get a constitutive equation for micropolar medium describing
a granular media based on computations by discrete element method (with explicit
rotation) [3].
Fig 1. Spherical wave of longitudinal rotation as vector field of rotational

velocities (at the right the cross section is shown)
Taking into account rotation by the both approaches simultaneously gives the
possibility to model a Cosserat medium by the MCA method if consider rotational
degree of freedom as independent rotation. Of course, this leads to necessity of
introducing additional elastic modulii, which are considered in the theory of mi-
cropolar media and can be defined experimentally. As is well known, a new type
of elastic waves are possible in such a medium, for example longitudinal rotation
wave. Fig. 1 shows development of spherical waves of longitudinal rotation, com-
puted by the MCA method. The wave is initialized by setting rotational velocity to
the automaton in the center of the cubic body. In spite of bell-shaped (positive)
initial signal the resulting wave consists of two regions of opposite rotation just as
a spherical longitudinal elastic wave consists of a compression region followed by
an expansion region.

Disadvantage of taking into account rotation as motion of surrounding neigh-

bors is an ambiguity in case of breaking one or more bonds with neighbors or in
case of formation of a new bond that may occur under severe deformation.
SIGNIFICANCE OF TAKING INTO ACCOUNT ROTATION FOR MCA

METHOD
Let us consider how significant is rotation for the MCA method. For this pur-
pose we analyzed the problem of elastic waves generation and propagation in a
half-space from the point source on the surface (the Lamb's problem) solved by
the MCA method with taking into account rotation and without it (Fig. 2). Ac-
cording to symmetry of the problem we considered a quarter of a steel cube with
the size of 0.25 m ( =7800 kg/m3, vP=5.95 km/s, vS=3.19 km/s. The automaton
size was 0.0025 m. Symmetric boundary conditions was applied on the planes
X=0 and Y=0, the other faces of the cube were free. Initial pulse of velocity was
applied in the origin of coordinates (shown in Fig. 2 by the arrow) as a “bell” of
sine curve during 5 s. The results in Fig. 2,b,c are shown at 40 s in Y=0 plane.
a b c
Fig. 2. Velocity field for the Lamb's problem solved by various models:
a,b – cubic packing with rotation of automata, c –cubic packing without rotation
As a result of such loading the longitudinal P and transverse (shear) S elastic
waves are formed at some distance from the source in the medium. Free surface of
the body cause generation of the conical C and Rayleigh R waves.
The computation shows that the result obtained without rotation taken into ac-
count depends on the packing of automata used. More of that, Rayleigh's wave in
that case vanishes and the shear wave propagates immediately after the longitu-

dinal wave with the same velocity, that is a qualitatively wrong result (Fig. 2,c). It
has to be noted that close packing in 2D has larger order of symmetry and pro-
vides qualitatively correct results, there is just a difference in wave velocities.
Taking into account rotation provides simulation result to be not depending on the
used packing of automata and physically correctness of elastic waves description
for continuum (Fig.2,a,b)
CONCLUSION
Thus, the study conducted shows that rotation of automata taken into account
as independed degree of freedom or via motion of neighboring automata allows
correct describing classical continuum by the movable cellular automata method.
To describe more complex media (for example Cosserat continuum) it is required
to realize the both ways and to consider rotational degree of freedom as indepen-
dent rotation.
ACKNOWLEDGMENTS
This study was supported by the Russian Foundation for Basic Research
(project 09-07-12016-ofi-m).
REFERENCES
1. Psakhie S.G., Dmitriev A.I., Shilko E.V., et al. ‘Method of movable cellular
automata as a new trend in discrete computational mechanics. I. Theoretical
description’, Physical mesomechanics, 3, No.2, 5–15, 2000. (in Russian)
2. Psakhie S.G., Chertov M.A., Shilko E.V. ‘Interpretation of the parameters of
the method of movable cellular automata on the basis of continuum descrip-
tion’, Physical mesomechanics, 3, No.3, 89–92, 2000. (in Russian)
3. Walsh S.D.C., Tordesillas A.A. ‘Thermomechanical approach to the develop-
ment of micropolar constitutive models of granular media’, Acta Mechanica,
167, No 3/4, 145–169, 2004.
4. Cundall P.A., Strack O.D.L. ‘A discrete numerical model for granular assem-
blies’, Geotechnique, 29, No.1, 47–65, 1979.

A Bounding Box Search Algorithm for DEM

Simulation
Laura E. Walizer and John F. Peters
US Army Engineer Research and Development Center, Vicksburg, MS
In Discrete Element Method (DEM) simulations, the most costly

operation performed by the program in terms of CPU time is of-
ten the process of identifying which pairs of particles are poten-
tially in contact. Program performance can especially be degraded
when the relative size difference between the smallest and largest
discrete elements is greater than a factor of 2 to 5. In this pa-
per, the authors evaluate the performance of a related but simpler
algorithm dubbed the ‘bounding box search method’. The bound-
ing box method entails identifying all cells which any part of a
target particle may occupy, listing the target particle as present
in those cells, and searching for potential contacts over the same
set of cells (the ‘bounding box’). To evaluate the performance of
the bounding box algorithm, timed simulations were performed
on systems with varying numbers of particles and particle size
distributions.
Introduction
The struggle to reduce the computational cost of contact detection in DEM simula-
tions has inspired the development of a whole range of potential solutions. While

substantial progress has been made, the performance of many search algorithms still
depends upon the size distribution of particles within the system, with performance
degrading significantly when the largest particles are more than 2 to 5 times the
size of the smallest particles. This limitation has detrimental implications for the
application of DEM to certain classes of problems, such as the study of granular
media [1]. The objective of the work presented here was to develop and implement
a contact algorithm that avoids performance degradation due to size distribution,
while still being conceptually simple and easy to implement. The proposed algo-
rithm has been dubbed the ‘bounding box’ method.
One of the simplest existing contact search algorithms is the basic spatial dis-
cretization algorithm (herein referred to as the ‘basic search’). In such an algorithm,
the simulation space is discretized into cells of equal size. Each particle is then
placed in a single cell based on the location of its center. A cell-based or particle-
based search is then performed, during which the program will cycle through all
cells (respectively, particles) in the system and look for potential contacts within
the 27-cell volume defined by the given cell or particle. While the simplicity of
such an algorithm has its advantages, it is subject to the requirement that all cells
must be larger than the diameter of the largest particle in the system. Thus, it is
vulnerable to performance degradation when particles in the simulated system are
not of similar size. Two of the most recent solutions to the contact search problem
include hierarchical or multi-grid methods [1, 2]. These methods could be consid-
ered descendants of the basic search algorithm as they discretize space into multiple
cell spaces with different cell sizes. The bounding box algorithm described below is
also closely related to this basic search algorithm. Therefore, a particle-based basic
search algorithm was employed by the authors as a benchmark for evaluating the
performance of the bounding box algorithm presented herein.
Methods and Results

By contrast to the basic search algorithm, the bounding box algorithm begins by
listing a given particle as present in every cell any part of the particle touches. This
placement is achieved by locating the particle’s center, then adding and subtracting
the particle radius to determine the maximum and minimum bounds, respectively,

in all dimensions. The particle is then added to the content list of all cells in the
indicated range. The range of bounds in the three coordinate directions define the
so-called ‘bounding box’. Because it is not possible for two particles to be in con-
tact if they do not occupy any cells in common, the bounding box also defines the
volume of the contact search in the bounding box algorithm. This use of the bound-
ing box to define the search volume allows the cell size to be based on the diameter
of the smallest particles, rather than the largest particles.
To assess the effect of the distribution of particle sizes on algorithm perfor-
mance, test systems were generated with three different particle size distributions.
Systems with a ‘bimodal’ size distribution consisted primarily of particles of the
smallest size (98% of particles had a radius of 0.19) and some particles (2%) of the
largest size (radius of 1.6), with very few particles (less than 1%) of any size in-
between. In systems with a ‘Gaussian-like’ distribution, a majority of the particles
(about 88%) had an intermediate radius of 0.79, particles with radii of 0.41 and 1.1
each accounted for approximately 10% of the system, and particles at the extreme
ends of the size range (0.19 and 1.6 radii, respectively) each accounted for about 1%
of the system. Systems with a uniform size distribution were comprised of equal
proportions of all five particle sizes.
Simulated system sizes were also varied. Five different system sizes were used,
with as few as 98 particles and as many as 56,000 particles. To ensure that the
algorithm was thoroughly tested, timed simulation runs were performed for each
unique combination of system size and particle size distribution. While examining
the effect of system size and particle size distribution, simulations based on the
basic search algorithm all discretized space into cubes with sides of length 2 (i.e., a
cell size of 2), while simulations based on the bounding box method all used a cell
size of 0.5.
Also, because the bounding box approach allows the user more latitude in select-
ing a cell size, simulation runs were performed to examine the effect of selected cell
size on runtime. A bimodally distributed system of 23,088 particles was used, and
the cell size was varied for each run. Selected cell sizes ranged from 0.3 (slightly
larger than the radius of the smallest particles) to 2 (the cell size used by the basic
search algorithm).
Total DEM runtime was recorded for all cases, and the runtime was normalized

by dividing by the number of particles and by the number of simulation timesteps.

The simulations were coded in FORTRAN 95 compiled using gcc, and run on a
quad-core Mac Pro.
Figure 1: Runtime per particle per simulation timestep versus number of particles for differ-
ing distributions. Dashed lines represent runs performed using the basic search algorithm,
and solid lines represent runs performed with the bounding box algorithm.
The results found for systems with varying system sizes and particle size dis-
tributions are presented in Figure 1. For the larger systems (of 10,000 to 20,000
particles or more), the greatest difference in performance between the basic and
bounding box algorithms was found when simulating a bimodal distribution. The
bounding box algorithm still outperformed the basic algorithm for large systems
with Gaussian-like and uniform distributions, but the differences in runtime be-
tween the algorithms were smaller for these systems. In general, the runtime ad-
vantage of the bounding box algorithm decreased with decreasing system size, and
for the smallest systems modeled, the performance of the bounding box algorithm
degraded significantly, particularly when modeling the bimodal distribution. This
performance degradation is attributed to the extra operations required by the bound-
ing box algorithm, such as computing the extents of bounding boxes, because the
contribution of these operations to overall runtime is expected to be more significant
as system size decreases.
When cell sizes were small (less than 0.5) runtimes increased with decreasing
cell size primarily due to an increase in the amount of runtime dedicated to gener-

ating the cell occupancy lists. Occupancy list arrays for these systems also became
larger (and therefore less memory efficient) with decreasing cell size. This was due
to the fact that particles will be present in an increasing number of cells as cell sizes
shrink relative to particle size.
Figure 2: Variation in runtime with respect to selected cell size for a system of 23000 particles
with a ‘bimodal’ type size distribution.
For large cell sizes (greater than 1) runtimes increased with increasing cell size.
The average number of particles per cell increased from 4.5 to nearly 12, and this
explains the increase in runtimes. However, the runtime leveled off around 6 × 10−6
seconds per particle per timestep as the cell sizes approached two. This runtime
was far below the value of 3.7 × 10−5 seconds obtained for the same system using
the basic search algorithm. This lower runtime for large cell-size appears to be due
to the fact that a bounding box is never made up of more than eight cells, and it
becomes increasingly likely that the bounding box has fewer cells as the cell size
increases Therefore, a bounding box search will never scale at more than O(8m×n)
while a basic search always scales at O(27m × n) for an n-particle system with an
average of m particles per cell.
The shortest runtimes were obtained with cell sizes between 0.5 (a little larger
than the diameter of the smallest particle considered) and 1. The average number
of particles per cell varied from 2.5 to 4.6 when cell sizes were in this range. These
results suggest that, while care must be taken in selecting cell size, there is a rel-
atively wide window of cell sizes (about 25% of the size range considered) that

will yield similarly good performance. Overall, it is concluded that the algorithm
performs best when cell sizes are between one and two times the diameter of the
smallest particles in the system, and given the leveling off of runtimes, it is other-
wise preferable to select too large a cell size than too small a cell size, for reasons
of both runtime and efficient memory use.
Conclusion
Some subtleties–such as the avoidance of double-counting contacts–exist in the im-
plementation of this algorithm as part of a simulation code. However, enough sim-
plicity remains that conversion of a basic algorithm to a bounding box algorithm is
quite straightforward, and a modularly designed simulation can contain both algo-
rithms as distinct subroutines. Implementation in parallel should also be possible.
The improvements in runtime relative to the basic search algorithm and the sim-
plicity of the bounding box algorithm suggest that the bounding box method is an
efficient and effective way to improve DEM simulation performance.
Acknowledgements
Any use of trade names is for descriptive purposes only and does not imply en-
dorsement by the U.S. Government. Permission to publish was granted by Director,
Geotechnical and Structures Laboratory, US Army Engineer Research and Devel-
opment Center. Approved for public release; distribution is unlimited.
References
[1] J. F. Peters, R. Kala, and R. S. Maier, “A hierarchical search algorithm for
discrete element method of greatly differing particle sizes,” Engineering Com-
putations, vol. 26, no. 6, pp. 621–634, 2009.
[2] K. He, S. Dong, and Z. Zhou, “Multigrid contact detection method,” Physical
Review E, vol. 75, p. 036710, 2007.

A SPRING SYSTEM EQUIVALENT TO

CONTINUUM MODEL
Shihai Li, Yanan Zhang, Chun Feng

Institute of Mechanics, Chinese Academy of Sciences
A new approach used to calculate the inner force and describe the
crack of element is demonstrated, with the name spring element
method (SEM). According to the energy functional and element stiff-
ness matrix, four types of springs have been gotten, with the name
normal spring, tangential spring, Poisson spring and pure shear spring
respectively. Based on thus springs above-mentioned, crack model is
presented, and two failure types (inner failure and boundary failure) of
element are given under tensile and shear criteria. Some numerical
cases are shown to demonstrate the validity and accuracy of SEM.
INTRODUCTION
Due to the development of numerical techniques, mechanical behavior of mate-
rial in continuous and discontinuous state could be simulated well. However, the
numerical method used to simulate progressive failure process from continuous
state to discontinuous state is still at the exploratory stage. It could be generally
classified into two branches about the method from continuum to discontinuum: a
block-based method and a particle-based method.
Block-based method is ground on the macroscopic mechanics, with the main
task to describe the microscopic crack rationally and precisely. Some skills have
been developed to simulate the initiation and extension of the crack, such as non-

continuous function, mathematical spring, block cut, grid regeneration, and so on.
[1-3]
Particle-based method is ground on the statistical mechanics, with the major

study emphasis on the equivalence between the particle aggregate and continuum
structure. Particle fluid code (PFC), lattice model (LM), particle model (PM), hy-
brid lattice particle model (HLPM) and spring network (SN) are some prevalent
methods based on statistical mechanics. [4-9]
Block-based method has a good advantage to simulate the failure feature on
engineering scale, but the formation mechanism of microscopic crack could not be
well modeled. On the contrary, particle-based method is suitable for simulation
the initiation and extension of micro-crack, but this approach could only simulate
the problem on laboratory scale. So, it is important to find a combined method
which combines the advantage of block-based method and particle-based method
together.
SPRING ELEMENT METHOD (SEM)

(1) Spring stiffness
A new method to construct the element stiffness matrix of three-node triangle
according to a special system of springs is presented. For a triangle of any shape,
the system includes six springs (Fig. 1).
3(c,w)
Syx
Syy
Sxyp
Sxys
1(0,0) 2(b,0)
y Sxx
Sxy
x
Fig 1. Spring system of three-node triangular element

They are the normal spring Sxx and the tangential spring Sxy along the bottom
edge, the normal spring Syy and the tangential spring Syx along the height direction
of the bottom edge, and a Poisson spring Sxyp related to the first two normal
springs and a pure shear spring Sxys related to the first two tangential springs. The
normal and the tangential springs follow Hooke's Law, and the direction of force
is opposite to the relative displacement; the direction of the Poisson spring force is
vertical to the one of displacement increment, and the energy of the Poisson
spring is the product of its force and its vertical displacement; The energy of the
pure shear spring is the product of shear stress and its vertical shear displacement.
The energy functional of the spring system is established (shown in equation 1).
Where, Kii , Kij , Kijp , Kijs (i, j  x, y; i  j ) are the spring stiffness; uij are the de-
formation of Sij (i, j  x, y ) ; ui , vi (i  1, 2,3) are displacements of element nodes;
m  c / b and n  1  m are coefficients.
By the variation of energy functional to coordinates of each node, the relation-
ship between force of each node and all the relative displacements of nodes is pre-
sented, thus, the stiffness matrix in the form of spring stiffness is obtained (shown
in equation 2). Compared with the finite element stiffness matrix, the spring stiff-
ness can be represented by the elastic modulus, Poisson's ratio and geometry pa-
rameters (shown in equation 3), and the relationship between spring stiffness is
shown in equation 4.
z
4(x4,y4,h)
1(0,0,0)
Szz
Sxx
y
Syy
3(c,w,0)
x 2(b,0,0)
Fig 2. Spring system of four-node tetrahedral element

By extending the similar method to three-dimensional problem, the spring sys-

tem of a tetrahedral element is built. Fig 2 shows the structure of a tetrahedral
spring system. According to the corresponding energy functional and element
stiffness matrix, the spring stiffness can be obtained by the same way (shown in
equation 5).
1 1 1 1
 K xx u xx2  K xy uxy2  K yy u yy
2
 K yx u yx
2
 K pxy uxx u yy  K sxy uxy u yx
2 2 2 2
(1)
u xx   u2  u1  ; u xy   v2  v1  ; u yy  v3  mv2  nv1; u yx  u3  mu2  nu1.
 F1x   Kxx  nnKyx n Kxyp  Kxys  Kxx  nmKyx mKxyp  nKxys nKyx Kxyp  u1 
F    
 1y  n Kxyp  Kxys  nnKyy  Kxy nKxyp  mKxys nmKyy  Kxy Kxys nKyy  v1 
F2x  K  nmK nK  mK  
    xx yx xyp xys Kxx  mmK yx mK xyp  mKxys mK yx Kxyp  u2  (2)
F2y   mKxyp  nKxys nmKyy  Kxy mKxyp  mKxys mmKyy  Kyx Kxys mKyy  v2 
F3x    
   nKyx Kxys mKyx Kxys Kyx 0  u3 
 
F3y   Kxyp nKyy Kxyp mKyy 0 Kyy  v3 
Et w Et b Et w
K xx  ； K yy  ； K xy  ；
2 1   2  b 2 1   2  w 4 1    b
(3)
Et b  Et Et
K yx  ； K xyp  ； K xys  .
4 1    w 2 1   2  4 1   
Kxy 
1  K ；K 
1  K ；Kxyp   Kxx Kyy；Kxys 
1  Kxx Kyy . (4)
xx yx yy
2 2 2
E (1   ) L j Lk
K ii  (i, j , k  x, y , z; i  j  k ; Lx  b, Ly  w, Lz  h)
(1   )(1  2  ) 6 Li
(5)
2(1   )
Kij  Kii ; Kijp  Kii K jj ; Kijs  Kij K ji .(i, j  x, y, z; i  j )
(1  2 )

(2) Calculation method

Suppose node displacements are known, the deformation of spring could be
obtained by equation 1, and then element strain could be gotten (shown in equa-
tion 6). The function to calculate the element stress is shown in equation 7, Where,
G,  is the shear modulus and lame constant of material,  kk is the normal strain
component and  ij is the Kroenecker delta symbol. Spring force is shown in equa-
tion 8, and node force is shown in equation 9. In the two-dimensional case, the
spring area could be gotten by equation 10, and where， t is element thickness.
Actually, Sii also could be gotten by Fij /ij (i, j  x, y;i  j) . There is a simple expres-
sion about spring area, which only relates to the element geometry. Once the ele-
ment is given, element strain could be gotten by equation 11, which is a unique
way for SEM to solve the strain.
 x   u xx / b  y   u yy / w  xy   u xy / b   u yx / w (6)
 ij  2G ij   kk ij (i, j  x, y) (7)
Fii  Kii uii  Kijp u jj Fij  Kij uij  Kijs u ji (i, j  x, y; i  j ) (8)
F1i  Fxi  Fyin F2i  Fxi  Fyim F3i  Fyi (i  x, y) (9)
S xx  Fxx /  x  wt / 2 S yy  Fyy /  y  bt / 2 (10)
 x  F xx / S xx  y  F yy / S yy  xy  F xy / S xx ( or  F yx / S yy ) (11)
The calculation method of 3-D element is similar to 2-D’s. The deformation of

spring is given by equation 12. Where, m and n are the interpolation coefficient of
the bottom edge, and N1, N2 and N3 are the interpolation coefficient of the under-
side. Element strain, spring force, node force, spring area and element stress are
shown in equation 13-17 respectively. To calculate the inner force of elements,
just the stiffness of the normal spring Kii are need (Kxx, Kyy are need in two-
dimensional case and Kxx, Kyy, Kzz are need in three-dimensional case).

uxi ui2 ui1 uyi ui3 mui2 nui1 uzi ui4  Nu

1 i1  Nu
2 i2  Nu
3 i3 (i  x, y, z;ux u,uy v,uz  w) (12)
i  uii / Li  ij  uij / Li  uij / Lj (i, j  x, y, z;i  j; Lx  b, Ly  w, Lz  h) (13)
Fii  Kiiuii  Kijpu jj  Kikpukk Fji  K jiu ji  Kijsuij (i, j, k  x, y, z;i  j  k) (14)
F1i  Fxi FyinFziN1 F2i  Fxi FyimFziN2 F3i  Fyi FziN3 F4i  Fzi (i  x, y, z) (15)
S xx  Fxx /  x  wh / 6 S yy  Fyy /  y  bh / 6 S zz  Fzz /  z  bw / 6 (16)
x  Fxx / Sxx y  Fyy / Syy z  Fzz / Szz xy  Fxy / Sxx xz  Fxz / Sxx yz  Fyz / Syy (17)
CHARACTERISTICS OF SEM
There are two characteristics about SEM: independence of element strain and
high efficiency of computation.
(1) Independence of element strain
To prove that the element strain is independent of the coordinate system of
spring system, two spring systems are given. One is b-w spring system and the
other is B-W spring system. (shown in figure 3)
y
3 
B C
3
W y
w
1 2 1 2
c x
b x
Fig 3. Spring systems of different bottom edge

Considering the relationship between B-W and b-w shown in the equation 18
and 19, the principal strain of two spring systems are expressed by node dis-
placements of b-w spring system, and the results are same.
Bw/sin Bc/cos W bsin BC bcos (18)
ui  ui cos  vi sin; vi  ui sin  vi cos(i 1,2,3) (19)
(2) High efficiency of computation

The time cost of SEM and FEM based on GPU are compared (table 1). SEM
owns high computational efficiency compared with FEM.
Method Two-dimensional Three-dimensional
FEM 456 1920
SEM 360 1032
Table 1. Comparison of time cost about SEM and FEM
CRACK MODEL
Crack model based on tensile criteria and shear criteria is demonstrated and
two crack types named inner crack and boundary crack are presented. When inner
crack takes place, a new block will turn up. When boundary crack occurs, a new
node will appear.
(1) Tensile crack model
There are two types for tensile crack models: they are (a) crack occurs inside of
element and (b) crack occurs along the edge. (Shown in figure 4)

a b
Fig 4. Tensile crack models
(2) Shear crack model
There are five types for shear crack model: they are (a) two directions of crack
inside of element through the same node, (b) two directions of crack inside of
element through different nodes, (c) two directions of crack along two edges (d)
one direction inside of element and one direction along the edge through the same
node, and (e) one direction inside of element and one direction along the edge
through different nodes. (Shown in figure 5)
a b
c d e
Fig 5. Shear crack models
NONLINEAR CONSTITUTIVE EQUATION IN SEM

SEM is based on the discrete springs, so the nonlinear constitutive equation
could be easily realized. One type of nonlinear coefficients associated with the
strain of spring is introduced (shown in equation 20), where, L1 is the length of the
bottom edge and L2 is the height. According to this type of nonlinear coefficients,

spring force could be calculated (shown in equation 21). The comparison between
numerical and theoretical result of uniaxial compression test is shown in figure 6.
104 xij / Li
Cij  1  e (i  x, y; j  x, y ) (20)
Fii  Cii K ii  uii  C jj K ijp  u jj Fij  Cij K ij  uij  C ji K ijs  u ji (i , j  x , y ; i  j ) (21)

1.2x106 Numerical result
1.0x106 Numerical result
Theoretical result
Axial Stress(Pa)
Theoretical result
Axial Stress(Pa)
9.0x105 Non-linear elasticity 8.0x105 Linear elasticity
6.0x105 6.0x105
4.0x105
3.0x105
2.0x105
0.0 0.0
0.0 3.0x10 -5
6.0x10 -5
0.0 1.0x10-5 2.0x10-5 3.0x10-5
Axial Strain Axial Strain
Fig 6. Comparison between numerical and theoretical result
DYNAMIC FAILURE OF CONCRETE WALL SIMULATED BY SEM

The thickness of concrete wall is 0.5 m, with the top and bottom constrained. A
steel ball with diameter 0.1 m impacts the wall with the speed 100 m/s. The pro-
gressive failure of concrete wall is shown in figure 7.
Fig 7. Progressive failure of concrete wall

CONCLUSION
Spring element method (SEM) is an approach to calculate the inner force of
element, with simple relationship between spring force and spring displacement.
In the calculation of triangle and tetrahedral element, the result of SEM and FEM
are the same. The feature that strain is unrelated with coordinate systems of spring
system is demonstrated. The computational efficiency of SEM is 30% higher than
FEM, and non-linear calculation could be easily realized In SEM. Based on the
inner force calculated by SEM, crack model is presented, and two failure types of
element are given. Some examples are presented to show the validity and accu-
racy of SEM, while further study would be needed on the failure model and crack
direction.
ACKNOWLEDGMENTS
Gratefully acknowledge the support from National Natural Science Foundation
of China (Grant No.10632100), the project of important research aspect of Chi-
nese Academy of Sciences (KJCXZ-YW-LDI) and the Chinese National 973 Pro-
ject (Grant No.2010CB731500).
REFERENCES
1. A Munjiza, The combined finite-discrete element method. John Wiley and
Sons, Ltd. 2004.
2. Shi GH. Three-dimensional discontinuous deformation analysis. Proceedings
of the 38th US rock mechanics symposium, Washington DC, 2001: 1421-1427.
3. T Belytschko, D Organ, C Gerlach. Element-Free Galerkin Methods for Dy-
namic Fracture in Concrete. Comp. Methods Appl. Mech. Engrg. 2000, 187,
385-399.
4. CUNDALL P A, Strack D L. A discrete numerical model for granular assem-
blies. Geotechnique. 1979, 29(1): 147-165.
5. Ostoja-Starzewski. M. Lattice models in micromechanics, Appl. Mech. Rev.,
2002, 55(1), 35-60.

6. Wang, G., Ostoja-Starzewski, M., Particle modeling of dynamic fragmentation

-I: theoretical considerations, Computational Materials Science, 2005, 33,
429-442.
7. G. Wang, A. Al-Ostaz, A.H.D. Cheng, P.R. Mantena. Hybrid lattice particle
modeling: Theoretical considerations for a 2D elastic spring network for dy-
namic fracture simulations. Computational Materials Science, 2009, 44(4):
1126-1134.
8. J. E. Bolander Jr, S. Saito. Fracture analyses using spring networks with ran-
dom geometry [J]. Engineering Fracture Mechanics. 1998, 61(5-6):569-591.
9. Norihiro Nakamura, Kenshiro Otsuki, Hiroyuki Nagahama. A spring-network
model of fault-system evolution. Computer Methods in the Ge-sciences. 1996,
15: 343-358.

RESOLVING THE INDETERMINACY OF VERTEX-

VERTEX CONTACT IN THE 2D DISCONTINUOUS
DEFORMATION ANALYSIS
Zhiye Zhao, Huirong Bao

Nanyang Technological University, School of Civil & Environmental Engineering, Singapore
Many numerical simulation systems, either continuum-based or dis-

continuum-based, require contact analysis. In the 2D discontinuous
deformation analysis (DDA), the contact reference edge for the ver-
tex-edge contact is uniquely defined, while the contact reference edge
for the vertex-vertex contact is indeterminate because of the disconti-
nuity of the outer normal vector of the boundary at the vertex. The in-
determinacy of the vertex-vertex contact will lead to an uncertainty in
the numerical analysis. This paper aims to develop suitable approach-
es in dealing with the two types of indeterminacy in the DDA. For the
first type of indeterminacy, it can be solved by an enhanced shortest
path method which employed a temporary vertex-vertex contact
spring. For the second type of indeterminacy, a complementary me-
thod is provided, which uses the trajectory of the vertex to find the en-
trance edge. A numerical example is used to verify the proposed ver-
tex-vertex approach.
INTRODUCTION
The two-dimensional discontinuous deformation analysis (DDA) method [1]
provides a useful tool for investigating the kinematics of blocky rock masses. The
DDA has drawn much attention in rock engineering applications recently due to

its unique feature in dealing with rock mass discontinuity, and many develop-
ments have been carried out to validate and to improve its performance [2].
The vertex-vertex contact indeterminacy exists in many numerical methods and
various approaches have been proposed to model the contact numerically [3-8]. In
the 2D DDA, there are two types of indeterminacy in the vertex-vertex contact
model. The first type of indeterminacy is referred as the genuine indeterminacy in
this paper because it originates from a true vertex-vertex contact. When the trajec-
tory of the hitting vertex of the contact pair passes its target vertex, the genuine
indeterminacy will arise.
C B
A D
D D A
A1 A2
E A
d1 d2
F
(a) (b) (c)

Fig. 1 Working principle of the shortest path method: (a) quasi V-V contact
case 1; (b) overlapped at the end of time interval; (c) quasi V-V contact case 2
The second type of indeterminacy is referred to as the pseudo indeterminacy in

this paper because it is a phenomenon belonging to particular numerical model-
ling methods when dealing with the quasi vertex-vertex contact. In the DDA, a
quasi vertex-vertex contact is finally transformed into a vertex-edge contact dur-
ing the open-close iteration in the framework of the penalty method. And the con-
tact fitness test is based on the geometry of overlapping at the end of the iteration.
However, the information provided by the geometry of overlapping is insufficient
for judging the entrance edge. For example, in Fig. 1b, Vertex A invaded into
Vertex D at the end of the time step, but the overlapping information is insuffi-
cient to obtain the proper entrance edge, from which Vertex A enters the bottom

block. The initial position could be one of the two cases shown in Fig. 1(a) and
Fig. 1(c). The shortest path method, as shown in Fig. 1(b), determines the en-
trance edge based on the distance between Vertex A and the two edges of the oth-
er block. But such a method may result in a wrong entrance edge.
SOLUTION TO THE GENUINE INDETERMINACY

Since the vertex-vertex contact is unstable and will degenerate into a vertex-
edge contact instantly when the collision occurs, a temporary vertex-vertex con-
tact spring is introduced to reflect this interaction process. The temporary vertex-
vertex contact spring will connect two vertexes like a hinge and will be removed
right after the potential vertex-edge contact is determined at the end of that itera-
tion. There are several open-close iterations in one time step and the adding-and-
removing temporary vertex-vertex contact spring is done in the first cycle of the
open-close iterations. Details about this method can be found in reference [9].
SOLUTION TO THE PSEUDO INDETERMINACY

To solve the inaccuracy of the shortest path method, a complementary method
is proposed here. Since every vertex has a trajectory during the time interval, the
trajectory of the vertex must intersect with the boundary of the target block before
invading into it. Therefore, by computing the intersection point, the entrance edge
can be easily found. This method is suitable for any vertex-vertex contact which
does not have any penetration at the start of the time step. If the trajectory happens
to pass the vertex point of the target block, it becomes the first type of indetermi-
nacy and the enhanced shortest path method can be used [9].
At the time instant that vertex P1 contacts edge P2P3, we have
(t )  0 (1)
Here,  (t ) is the area of triangle PP
1 2 P3 at time instant t. Solve Eq.(1) with re-
spect to t will obtain the time instant that the contact occurs. Eq.(1) is solved for
both potential reference edges. It is necessary to test whether the time instant t is
inside the time step and whether the intersection point lies on the edge instead of
the extending line. After obtaining the proper entrance edge, the vertex-vertex

contact is transformed into a vertex-edge contact and the open-close iteration will
continue.
A NUMERICAL EXAMPLE
A two-block system under vertical body force is at rest at the beginning of the
analysis (see
Fig. 2). The bottom block is fixed at two points. At the initial position, there is
a small gap between them, and the vertex of the upper block will land on the right
edge of the bottom block when they are in contact
The material properties for both blocks are as follows: mass density = 2.8×103
kg/m3; Young’s Modulus = 50MPa, Poisson’s ratio = 0.25, body force fx = 0N and
fy = -2.8×105N. The allowed displacement ratio is 0.1, and the maximum time in-
crement for each time step is 0.05s. It is assumed that there is no friction between
the interfaces.
In the first test, the time step is set as T = 0.05s and the penalty spring stiffness
is 40 times the value of Young’s Modulus, i.e. p = 40E. The original DDA code
produced a wrong result as shown in Fig. 3, whereas the new DDA code gives
correct results as shown in Fig. 4. In the second test, the time step is the same as
the first test and the penalty spring p= 80E, the original DDA code also gets the
correct results as the new code (similar to Fig. 4). In the third test, the penalty
spring p = 40E and the time step size is reduced to 0.025s, both DDA codes pro-
vide the same correct results.
1 1
0.02
0.1
3
4
3 5

Fig. 2 Geometry of the numerical example (unit: m)
(a) step=0 (b) step=1 (c) step=2 (d) step=3

Fig. 3 results from the original DDA code (p=40E, T=0.05s)

Fig. 4 results from the new DDA code (p=40E, T=0.05s)

Fig. 5 results from the original DDA code (p=40E, T=0.025s)
CONCLUSIONS
A new approach to deal with the vertex-vertex contact is introduced in this pa-
per. The shortest path method employed by the original DDA code is sensitive to

the choice of the DDA analysis parameters, such as the contact spring stiffness
and the time step. The proposed complementary method can resolve the pseudo
indeterminacy of vertex-vertex contact in the context of 2D DDA, which has veeb
validated by the numerical example. The new method can also be applied to the
numerical manifold method and other penalty based contact model for resolving
the indeterminacy of the vertex-vertex contact.
REFERENCE
1. Shi G, Discontinuous deformation analysis - a new numerical model for
the statics and dynamics of block systems, Ph.D. Thesis, 1988, University
of California.
2. MacLaughlin MM and Doolin DM, Review of validation of the disconti-
nuous deformation analysis (DDA) method. International Journal for Nu-
merical and Analytical Methods in Geomechanics, 2006. 30: p. 271-305.
3. Oden JT and Pires EB, Algorithms and numerical results for finite element
approximations of contact problems with non-classical friction laws.
Computers & Structures, 1984. 19(1-2): p. 137-147.
4. Chaudhary AB and Bathe K-J, A solution method for static and dynamic
analysis of three-dimensional contact problems with friction. Computers &
Structures, 1986. 24(6): p. 855-873.
5. Krishnasamy J and Jakiela MJ, A method to resolve ambiguities in corner-
corner interactions between polygons in the context of motion simulations.
Engineering Computations, 1995. 12: p. 135-144.
6. Kane C, Repetto EA, Ortiz M, and Marsden JE, Finite element analysis of
nonsmooth contact. Computer Methods in Applied Mechanics and Engi-
neering, 1999. 180(1-2): p. 1-26.
7. Feng YT and Owen DRJ, A 2D polygon/polygon contact model: algorith-
mic aspects. Engineering Computations 2004. 21(2/3/4): p. 265-277.
8. Cundall PA, UDEC 4.0 Manual - Theory and Background. 2004, ITASCA
Consulting Group, Inc.
9. Bao HR and Zhao ZY, An alternative scheme for the corner-corner contact
in the two-dimensional Discontinuous Deformation Analysis. Advances in
Engineering Software, 2010. 41(2): p. 206-212.

LIGGGHTS – A NEW OPEN SOURCE DISCRETE

ELEMENT SIMULATION SOFTWARE
Christoph Kloss1,* and Christoph Goniva1

CD Lab on Particulate Flow Modelling, Johannes Kepler University Linz
Altenbergerstr. 69, 4040 Linz
*[email protected]
In this work, we present the development of an open-source software

for modeling granular material by means of the Discrete Element Me-
thod. LIGGGHTS stands for ‘LAMMPS Improved for General Granu-
lar and Granular Heat Transfer Simulations’ and is based on
LAMMPS (‘Large Atomic and Molecular Massively Parallel Simula-
tor’), a successful open source Molecular Dynamics code by Sandia
National Laboratories for massively parallel computing on distributed
memory machines. We give a brief overview of implemented models
and features. Finally, we would like to focus on the simulation of
coupled granular-fluid systems with the CFD-DEM approach.
INTRODUCTION
The ‘Discrete Element Method’ (DEM) is an emerging method to simulate the
behaviour of granular materials. Open source codes represent an emerging alterna-
tive to existing commercial software. First, we will briefly present LAMMPS and
summarize LAMMPS-related work. Then, we outline which steps have been tak-
en to improve the granular models implemented in LAMMPS and illustrate the
capabilities of LIGGGHTS with a few examples. Finally, we would like to focus
on the simulation of coupled granular-fluid systems.

THE MD CODE LAMMPS

LAMMPS is an acronym for ‘Large-scale Atomic/Molecular Massively Paral-
lel Simulator’. It is a classical molecular dynamics solver written in C++ provid-
ing potentials for soft materials (bio-molecules, polymers), solid-state materials
(metals, semiconductors) and coarse-grained granular materials [1]. It can be used
to model atoms or, more generically, as a parallel particle simulator at the atomic,
meso, or continuum scale. LAMMPS runs on single processors or in parallel using
message-passing techniques and a spatial-decomposition of the simulation do-
main. The code is well developed and supports a wide variety of models handling
very large systems with good parallel efficiency. Its design allows for easy mod-
ification. Detailed descriptions can be found in the LAMMPS manual [24].
LAMMPS is distributed as an open source code under the terms of the GPL. To
name just a few examples, LAMMPS includes models for peridynamics [2], a
coupling interface to multibody software [3], and long-range electrostatics solvers
[4]. LAMMPS also offers GPU support for simulation speed-up.
For granular simulations, LAMMPS offers implementations for both linear
(Hooke) and non-linear (Hertz) granular potentials along with efficient algorithms
for detecting and calculating the pair-wise interaction forces. LAMMPS has been
used to study avalanching flow [5-7], static granular packings [8-10], granular
drainage [11-14], granular shear and its relation to constitutive modeling [15-19],
and granular segregation [20]. LAMMPS also allows to ‘overlay’ force fields,
meaning that e.g. an electrostatic force field (with point charges and/or dipole
moments) can be added to a granular potential easily.
THE DEM CODE LIGGGHTS

In this section, we would like to emphasize those models that have been im-
proved or added in LIGGGHTS with respect to the implementations in LAMMPS,
Movies of all this examples can be viewed on http://www.liggghts.com.
Contact Law
We have changed the implemented granular contact law to match the Hertz-
Mindlin and Linear Spring formulations published in [27] and [28]. Moreover, it
should be easy now for users to implement their own customized contact model.

Also, we added a macroscopic particle cohesion model described in [29]. Fig. 1.

shows an example where this cohesion model is used.
Fig. 1. Loosely packed granular column in a cylinder falling under gravity, with
and without macroscopic cohesion model. Left: Initial condition; Right: Result
Implementation of Import and Handling of Complex Geometries..
Based on work of Kremmer and Favier [26], we implemented a method to im-
port and handle complex wall geometry. Walls can be imported from STL CAD
files. Possible wall styles are hookean or hertzian with optional friction, shear his-
tory effects and cohesion. In Fig. 2., a particle charging simulation is shown.
Fig. 2. Example showing charging of particles

Implementation of a Moving Mesh Feature.
Many applications in granular flow require the constraining wall to be moving.
We added such a feature, which is illustrated in Fig. 3., where a bucket is digging
into a bed of particles.

Fig. 3. Example showing granular charging simulation

Implementation of a simple Granular Heat Transfer Model
A simple model for heat conduction described in [30] has been implemented.
An example is shown in Fig. 4. It is noted that heat convection often plays an im-
portant role in heat transfer of granular media, but an accurately modeling convec-
tion requires a coupled CFD-DEM approach as outlined in the next section.
Fig. 4. Example showing conductive heat transfer. Left: t=0, Temperature gradient
is imposed on a particle bed, Right: t=t1, Temperature is leveling gradually

CFD-DEM COUPLING
The CFD-DEM Method is an emerging method to model coupled fluid-
granular systems (e.g. [22],[23]). The motion of the particles is resolved with
DEM, and the CFD method is used to calculate the fluid flow. The authors have
already managed to implement a CFD-DEM approach using two commercial
software packages, which has been presented along with validation [21]. Here, we
want to focus on a fluidized bed application example that has been calculated
coupling LAMMPS to the OpenFOAM® environment [25].
Fig. 2: Geometry and velocity boundary condition and initial particle positions.
Fig. 3: Simulation result for the fluidized bed test case. Gas velocity contour
in a y-normal plane, particles coloured by particle velocity.

As discussed in this paper, the authors have succeeded allowing parallel com-
putation on distributed memory machines. A simple test case is presented here.
The granular and fluid flow of a small partly fluidized bed is calculated. On the
lower side of the box there is a non uniform, fixed value velocity inlet. Geometry
and boundary conditions are shown in Fig. 2. As it can be seen from Fig. 3, the
particles are accelerated by the gas flow and a region where the particles are flui-
dized is formed above the velocity inlet.
CONCLUSION
We presented an emerging open source DEM framework called ‘LIGGGHTS’
for the simulation of granular materials by means of the Discrete Element Me-
thod, which is based on LAMMPS, a very established software for performing
Molecular Dynamics (MD) simulations. We showed some examples of the fea-
tures ‘LIGGHTS’ offers for the simulation of granular materials by presenting
some examples. It is planned to release a final version of ‘LIGGGHTS’ in the
coming months on http://www.liggghts.com. Finally, we outlined a promising
open source approach to simulate coupled granular-fluid systems massively paral-
lel on distributed memory machines.
REFERENCES
1. S. J. Plimpton, ’Fast Parallel Algorithms for Short-Range Molecular Dynamics’, J. Comp. Phys., 117, 1-19, 1995
2. M. L. Parks, R. B. Lehoucq, S. J. Plimpton, S. A. Silling, ‘Implementing peridynamics within a molecular dynamics
code’, Comp Phys Comm., 179, 777-783, 2008
3. R. M. Mukherjee, P. S. Crozier, S. J. Plimpton, K. S. Anderson, ‘Substructured molecular dynamics using multibody
dynamics algorithms’, Intl. J. of Non-Linear Mechanics, 43, 1045-1055, 2008
4. S. J. Plimpton, R. Pollock, M. Stevens, K. S. Anderson, ‘Particle-Mesh Ewald and rRESPA for Parallel Molecular
Dynamics Simulations’, Proc. of the Eighth SIAM Conference on Parallel Processing for Scientific Computing, Min-
neapolis, MN, March 1997
5. L. E. Silbert, D. Ertas, G. S. Grest, T. C. Halsey, D. Levine, S. J. Plimpton, ‘Granular flow down an inclined plane:
Bagnold scaling and rheology’, Phys Rev E, 64, 051302, 2001
6. L. E. Silbert, J. W. Landry, and G. S. Grest, ‘Granular flow down a rough inclined plane: Transition between thin and
thick piles’, Phys. Fluids, 15, 1, 2003
7. R. Brewster, G. S. Grest, J. W. Landry, and A. J. Levine: ‘Plug flow and the breakdown of Bagnold scaling in cohe-
sive granular flows’, Phys. Rev. E, 72, 061301, 2005
8. L. E. Silbert, D. Ertas, G. S. Grest, T. C. Halsey, and D. Levine, ‘Geometry of frictionless and frictional sphere pack-
ings’, Phys. Rev. E, 65, 031304, 2002
9. L. E. Silbert, G. S. Grest, and J. W. Landry, ‘Statistics of the contact network in frictional and frictionless granular
packings’, Phys. Rev. E, 66, 061303, 2002

10. J. W. Landry, G. S. Grest, L. E. Silbert, and S. J. Plimpton, ‘Confined granular packings: Structure, stress, and forces’,
Phys. Rev. E, 67, 041303, 2003
11. H. P. Zhu and A. B. Yu, ‘Steady-state granular flow in a three-dimensional cylindrical hopper with flat bottom: mi-
croscopic analysis’, J. Phys. D, 37, 1497, 2004
12. C. H. Rycroft, M. Z. Bazant, G. S. Grest, and J. W. Landry, ‘Dynamics of random packings in granular flow’, Phys.
Rev. E, 73, 051306, 2006
13. C. H. Rycroft, G. S. Grest, J. W. Landry, and M. Z. Bazant, ‘Analysis of granular flow in a pebble-bed nuclear reac-
tor’, Phys. Rev. E, 74, 021306, 2006.
14. J. W. Landry, G. S. Grest, S. J. Plimpton, ‘Discrete element simulations of stress distributions in silos: Crossover from
two to three dimensions’, Powder Technology, 139, 233-239, 2004
15. X. Cheng, J. B. Lechman, A. Fernandez-Barbero, G. S. Grest, H. M. Jaeger, G. S. Karczmar, M. E. Mobius, and S. R.
Nagel, ‘Three-Dimensional Shear in Granular Flow’, Phys. Rev. Lett., 96, 038001, 2006
16. K. Kamrin, C. H. Rycroft, and M. Z. Bazant, ‘The stochastic flow rule: a multi-scale model for granular plasticity’,
Model. Simul. Mater. Sci. Eng., 15, 449, 2007
17. M. Depken, J. B. Lechman, M. van Hecke, W. van Saarloos, and G. S. Grest, ‘Stresses in smooth flows of dense
granular media’, Europhys. Lett., 78, 58001, 2007
18. C. H. Rycroft, K. Kamrin, and M. Z. Bazant, ‘Assessing continuum postulates in simulations of granular flow’, J.
Mech. Phys. Solids, 57, 828, 2009
19. L. E. Silbert, G. S. Grest, S. J. Plimpton, D. Levine: ‘Boundary effects and self-organization in dense granular flows’,
Physics of Fluids, 14, 2637-2646, 2002
20. J. Sun, F. Battaglia, and S. Subramaniam, ‘Dynamics and structures of segregation in a dense, vibrating granular bed’,
Phys. Rev. E, 74, 061307, 2006
21. C. Kloss,C. Goniva, G. Aichinger, S. Pirker, ‘Comprehensive DEM-DPM-CFD simulations: Model synthesis, expe-
rimental validation and scalability’, Proc. Seventh International Conference on CFD in the Minerals and Process In-
dustries, CSIRO, Melbourne, Australia, December 9-11, 2009
22. Zhu, H.P., Zhou, Z.Y, Yang, R.Y. and Yu, A.B., (2007), Discrete particle simulation of particulate systems: Theoreti-
cal Developments, Chemical Engineering Science, 62, pp. 3378-3396
23. Zhu, H.P., Zhou, Z.Y, Yang, R.Y. and Yu, A.B., (2008), Discrete particle simulation of particulate systems: A review
of major applications and findings, Chemical Engineering Science, 63, pp. 5728-5770
24. Steve Plimpton et. al., LAMMPS User Manual, Sandia National Laboratories, Albuquerque, USA, available online at:
http://lammps.sandia.gov/doc/Manual.html,
25. C. Goniva, C. Kloss, S. Pirker, ‘Towards fast parallel CFD-DEM: An Open-Source Perspective’, Proc. Open Source
CFD International Conference, Barcelona, November 12-13, 2009
26. M. Kremmer, J. F. Favier, ‘A method for representing boundaries in discrete element modelling - part I: Geometry
and contact detection’, Int. J. for Num.l Meth. Eng., 51 (12), 1407 – 1421, 2001
27. Y. Tsuji, T. Tanaka and T. Ishida, ‘Lagrangian numerical simulation of plug flow of cohesionless particles in a hori-
zontal pipe’, Powder Technology, 71, 239-250, 1992
28. A. Di Renzo, F. P.Di Maio, ‘Comparison of contact-force models for the simulation of collisions in DEM-based gra-
nular flow codes’, Chemical Engineering Science, 59, 525 – 541, 2004
29. N. Mitarai, F. Nori, ‘Wet granular materials’, Advances in Physics, 55, 1-45, 2006
30. B. Chaudhuri et al, ‘Modeling of heat transfer in granular flow in rotating vessels’, Chem. Eng. Sci., 61, 6348 – 6360,
2006

COUPLED SOLUTIONS

DISCRETE THERMAL ELEMENT MODELLING OF

HEAT CONDUCTION IN SPHERICAL PARTICLE
SYSTEMS
Y. T. Feng, C. F. Li, K. Han, D.R.J. Owen

Civil and Computational Engineering Centre, Swansea University, UK
Abstract: This paper is the further development of the discrete thermal ele-
ment method (DTEM) [1], for the effective modeling of heat conduction in
spherical particle systems. In essence, the temperature distribution in each particle
is not assumed isothermal and the actual distribution can be evaluated. The whole
development is analogy to the 2D case [1].
Consider a 3D spherical domain of radius R in a particle assembly and in ther-
mal contact with n neighboring spheres. Assume the spherical particle has an iso-
tropic thermal conductivity κ and is subjected to a prescribed heat flux (Neumann)
boundary condition outside the contact zones. Establish a local coordinate system
with the origin set at the centre of the domain. Then the temperature distribution
on the surface can be expressed by a boundary integral in terms of a Green func-
tion and the heat flux function on the surface.
By further assuming a flux distribution function over each thermal contact
zone, a linear algebraic system of thermal conductivity equations for each particle
can be derived in terms of the average temperatures and the resultant fluxes at the
contact zones with its neighboring particles.
Thus each particle is treated as an individual element with the number of (tem-
perature) unknowns equal to the number of particles that it is in contact with. The
element thermal conductivity matrix can, in principle, be evaluated and is entirely

dependent on the characteristics of the contact zones. This resulting element

shares the same form and properties with its conventional thermal finite element
counterpart.
The computational challenges, in comparison with the 2D formulation, lie in
the effective and accurate evaluation a triple integral. Various numerical integra-
tion quadratures have to be employed to achieve the goal. The numerical evalua-
tions reveal that for a similar contact angle [1] as in a 2D case, the average tem-
perature of the zone is much closer to the average temperature of the whole
sphere.
In addition, similar to the 2D case [2], the element thermal conductivity matrix
can be simplified so that the formulation can readily be used to handle transient
heat transfer problems.
Numerical results are provided to assess the accuracy and effectiveness of the
methodology developed.
REFERENCES
1. Y. T. Feng, K. Han, C.F. Li and D. R. J. Owen. ‘Discrete Thermal Element
Modelling of Heat Conduction in Particle Systems: Basic Formulations.
Journal of Computational Physics. 227(10): 5072-5089, 2008.
2. Y. T. Feng, K. Han & D.R.J. Owen, ‘Discrete thermal element modelling of
heat conduction in particle systems: Pipe-network model and transient analy-
sis’, Powder Technology, 193(3), 248-256, 2009.

IMPLEMENTATION OF COMBINED SINGLE AND

SMEARED CRACK MODEL IN 3D COMBINED
FINITE-DISCRETE ELEMENT ANALYSIS
Zhou Lei1, Mengyan Zang2, Antonio Munjiza1

1
Department of Engineering, Queen Mary, University of London
2
School of Mechanical & Automotive Engineering, South China University of Technology
Large-scale combined finite-discrete element simulations involve a

large number of separate bodies that interact with each other and in
general deform and fracture. In this context, the so-called combined
single and smeared crack model, which is based on actual representa-
tion of experimental stress-strain curves, was proposed and success-
fully implemented into a general purpose two-dimension (2D) com-
bined finite-discrete element code in the last decade. In this paper, the
combined single and smeared crack model is extended into three-
dimension (3D) in the context of the combined finite-discrete element
method. The validation of such extension is illustrated through a nu-
merical example.
INTRODUCTION
The combined finite-discrete element method, which merges the finite ele-
ment-based analysis of continua with discrete element-based transient dynamics,
contact detection, and contact interaction solutions, has already proven its effi-
ciency and reliability as a computational tool to solve the problems involving
transient dynamics of systems in which deformation and contact play an important

role. Now, this approach is a fast developing area of computational mechanics of

discontinua involving researchers and engineers from various disciplines [1].
One of the key advantages of the combined finite-discrete element method is
capable of simulating large-scale multi-fracture problems. The robust fracture al-
gorithm called combined single and smeared crack model was proposed for such
problems by Munjiza [2-3]. However, the combined single and smeared crack
model was aimed at mode I loaded cracks only in its original form and it has been
implemented in 2D.
In this work, the combined single and smeared crack model is extended into
three-dimension and is implemented into a general purpose three-dimension com-
bined finite-discrete element code Y3D. The 3D FEMDEM code, which contains
usual features of the finite element method, has the full capability for contact de-
tection and contact interaction.
In the remainder of the paper, algorithmic issues concerning the combined sin-
gle and smeared crack model in the context of the 3D combined finite-discrete
element method are presented together with a numerical example.
COMBINED SINGLE AND SMEARED CRACK MODEL IN 3D

The combined single and smeared crack model is based on actual representa-
tion of experimental stress-strain curves. It divides the stress-strain curves into
two parts: the strain-hardening part and the strain-softening part [1-3]. In the
strain-hardening part, no failure occurs in the material and the standard continuum
constitutive law dominates the system performance; while in the strain-softening
part, a single-crack model is used and the bonding stresses are taken to be a func-
tion of the separation vector δ.
At any point on the surfaces of a crack, the separation vector δ can be divided
into two components,
δ = δnn + δtt (1)

where n and t are the unit vectors in the normal and tangential direction of the
surface at such a point, δn and δt are the magnitudes of the components of δ,
respectively.

Accordingly, the traction vector p in the strain-softening stage are divided into
two components in the direction of n and t,
p = σn + τt (2)
where, σ and τ are the normal and tangential stresses, and can be calculated by
  zf t (3)
  zfs (4)
here, ft and fs are the tensile and shear strength of material. z is a heuristic
parameter depended both on the material and the seperation components [1-3],
z  1 
ab
e 
 a  b  1 D ( a  cb /(( a b )(1 a b ))) 
 a(1  D)  b(1  D) c  (5)
 
where a, b, c are material parameters chosen to fit a particular experimental curve,
and D is therefore parameter calculated through the separation components
2 2
   
D   n   t  (6)
  cn    ct 
here, δcn and δct are the normalized crack opening and sliding displacements
determined by the energy release rate Gf and the strengths. In the above equation,
D ≥ 1 means that the material is failure and no resistance exists; to represent such
a truth, D is thereof set to 1 as long as D ≥ 1, thus z as well as σ and τ are zero.
In actual implementation, the cracks are assumed to coincide with the finite
element boundary (surfaces in 3D); the separation of adjacent element surfaces is
assumed in advance through the topology of the finite elements. A so-called joint
element in the context of the combined finite-discrete element method is inserted
in advance between two adjacent elements to connect those two originally adja-
cent elements. Thus no two original finite elements share any nodes at the begin-
ning of calculation-the continuity between elements is enforced through the pen-
alty function method [1-3].

NUMERICAL EXAMPLE
In this section, a relatively simple example is shown to demonstrate the imple-
mented fracture model.
As shown in Fig 1, a beam is supported by 3 supports moving in opposite di-
rections (in direction y) with constant velocity of 0.1 m/s. The material properties
of the beam are: Young’s modulus E = 26 GPa, Position’s ratio v = 0.18, density ρ
= 2340 kg/m3, tensile strength ft = 3.15 MPa, and strain energy release rate Gf =
10 N/m. To constrain the cracks to be the mode I type, the shear strength of beam
is set large enough.
The displacements of the nodes, the initional positions of which are at points A
and B in Fig 1, are taken into consideration. In Fig 2(a), the displacement of all 8
nodes at point A are the same, which represents that the continuity between finite
elements is enforced through the penalty function method; while Fig 2(b) shows
that the material at point B undergoes both the strain-hardening stage (before 0.62
ms) and strain-softening stage (from 0.62 ms until to failure). Through the post-
processer, the total progressive collapse of the beam is observed. It presents that
the first crack appears at the bottom midspan point and propagates towards the top
until the beam eventually breaks into two. The process satisfies well with that
simulated by the 2D combined finite-discrete element code Y2D [2], and the frac-
ture pattern at some moment is shown in Fig 3.
Fig 1. Diagram and initial finite element mesh of the simply supported beam.

(a) Point A (b) Point B

Fig 2. Displacements of the nodes at point A and B.
Fig 3. Fracture pattern at some moment.
CONCLUSION
The combined single and smeared crack model is extended into three-
dimension and implemented into a general purpose 3D combined finite-discrete
element code Y3D. The extension enables both single and multiple cracks to be
modelled in 3D by using Y3D. Thus it is promising to use the 3D model to ana-
lyze progressive fracturing and fragmentation involving a large number of cracks.
However, the numerical examples with more complex load conditions should
be tested, and the influence of the parameters such as the penalty factor on the
analysis results should be studied in our future work.
ACKNOWLEDGMENTS
Zhou Lei would like to acknowledge the support from the State Scholarship
Fund of China (File No. 2009615030), with the help of which, he has the oppor-

tunity to pursue his study under the supervision of Professor A. Munjiza in Great
Britain as a joint PhD student. Zhou Lei is on leave from the School of Mechani-
cal & Automotive Engineering, South China University of Technology.
REFERENCES
1. A. Munjiza, The combined finite-discrete element method, Wiley, 2004.
2. A. Munjiza, K.R.F. Andrews, J.K. White, ‘Combined single and smeared
crack model in combined finite-discrete element analysis’, International jour-
nal for numerical methods in engineering, Volume 44, No.1, 41-57, 1999.
3. A. Munjiza, N.W.M. John, ‘Mesh size sensitivity of the combined FEM/DEM
fracture and fragmentation algorithms’, Engineering fracture mechanics, Vol-
ume 69, No.2, 281-295, 2002.

MODELLING OF REACTING DISCRETE

PARTICLES IN CONTINUOUS FLUID FLOW: AN
ENERGY TECHNOLOGY PERSPECTIVE
H. Kruggel-Emden1, S. Wirtz1, V. Scherer1, A. Munjiza2

1: Department of Mechanical Engineering, Ruhr-University Bochum, Universitaetsstrasse 150,
44801 Bochum, Germany
2: Department of Engineering, Queen Mary, University of London, Mile End Road, E1 4NS, Lon-
don, UK
Energy Technology is an engineering discipline dedicated to the effi-

cient, environmental friendly and safe provision of final energy to the
end user. This involves a broad range of processes which often have in
common that they occur at high temperatures and involve reacting
particle/gas systems. Modeling of these systems is important to gain
insight into details of the processes and for the improvement of ma-
chinery and process conditions. Among other methods discrete ele-
ment approaches play an important role.
INTRODUCTION
Energy Technology and Energy Process Engineering are a conglomerate of dif-
ferent engineering disciplines dealing with the safe, efficient, economic and envi-
ronmental friendly energy production, transformation, transport, storage and us-
age of various forms of energy.
Nature provides different forms of primary energy in the form of fossil (coal,
oil, natural gas), nuclear (uranium) and renewable sources (wind, hydro, solar, bio

mass, waste). Only part of the primary energy sources can be utilized directly.
Most have to be processed into e.g. coke, briquettes, fuel oil, gasoline, electricity
or heating and passed in this form to consumers; the involved transformation
processes are usually of complex nature.
For the current century studies predict a population growth from the current 6.5
to 9.2 billion people in the year 2050. Despite increasing energy efficiency this
will lead to a significant growth in energy consumption. Currently 80% of the
world's primary energy demand is covered by fossil fuels like coal, oil and natural
gas [1]. The combustion of fossil fuels leads to carbon dioxide emissions which
are released into the atmosphere and accumulate there. This process enhances the
natural greenhouse effect and leads to a hazardous warming of the climate [2].
Many scientific studies therefore come to the conclusion that a prompt reduction
of green house gas emissions especially from the most dominant sources like
power generation and other industrial processes is essential [2].
On the background of a growing world population and rising energy consump-
tion it is essential to enhance energy efficiency and to expand the use of non-fossil
energy sources. At the same time as part of the climate issue, a significant reduc-
tion of carbon dioxide emissions is required. This can be achieved through the
application of carbon dioxide capture technologies for fossil fuels, but in addition
an expansion of the use of renewable energy is required. Among renewable non-
fossil fuels biomass and waste are favorable options, as they are available in con-
siderable amounts, storable, and embeddable in the existing energy supply chains.
To meet the highlighted challenges the development and improvement of proc-
ess conditions as well as the scale up of energy utilization processes today must
and can rely on numerical simulations. In stationary energy conversion high tem-
perature particle/gas systems are predominant. Systems vary strongly in system
size, particle sizes, particle/gas ratio and the way of agitation. Conditions range
from dilute regimes as encountered for coal dust firing up to densely packed sys-
tems as found in moving bed equipment. In the following a review of possible
simulation approaches is given for the relevant system conditions and advantages
and disadvantages are briefly discussed.

ENERGY CONVERSION SYSTEMS

The provision of usable end energy to the user nearly always requires a se-
quence of dedicated conversion steps. In the context of fossil fuels an associated
combustion process requires high temperatures and large mass and energy fluxes.
Pulverized Thermal Conversion. For pulverized thermal conversion the solid
fuel has to be ground into very small particles before being ignited and burnt in
the boiler. The modeling of the associated processes can be performed in the
Euler/Lagrangian framework, slightly extending the single phase, time averaged
Navier-Stokes equations. Since the volume fraction of the particle phase is usu-
ally very low the particle/particle interactions may be neglected. Gas/particle and
particle/gas momentum, heat and mass transfer are represented through sub-
models defining corresponding fluxes among the phases considered. Simulations
of the steady state situation, considering a statistically adequate number of repre-
sentative particle tracks are usually sufficient for technical purposes. As an exam-
ple, the particle tracks and temperatures from a boiler simulation are shown in
figure 1a. Euler/Lagrange frameworks have been extensively applied to pulver-
ized coal combustion [3,4] and full scale boilers. Special attention in this context
gained the reduction of pollutants in low NOx-combustion systems [5]. A co-
firing of up to 20% biomass (by energy) in conventional boilers is possible result-
ing in significant carbon dioxide emission reductions. Numerical studies investi-
gated bio mass co-firing and sole biomass combustion with special focus on ap-
plicable sub-models [6].

m& s , ox m& fuel exhaust

m& air exhaust
Fuel-
Air- reactor
reactor
Buffer
m& s ,red
m& air
m& s ,buf m& s , feed m& fuel
a) b)
Fig 1. (a) Predicted tracks of particles in a coal fired boiler from Euler/Lagrangian
modeling, (b) Outline of an Eulerian framework for chemical looping combustion.
Fluidized Bed Thermal Conversion. Fluidized bed thermal conversion does

not allow the simplifications applicable to dilute systems, requiring a considerably
greater numerical effort. Due to this a variety of simulation approaches was de-
veloped. It can be distinguished between macroscopic balance models, multi-
phase continuum models and discrete element/CFD models.
Fluidized bed models utilizing the macroscopic balance approach date back to
the 1950s [7] and are still the basic design tool for these systems. A number of
review papers address macroscopic balance model developments; a recent review
was published by Mahecha-Botero [8]. Models can be distinguished regarding
many characteristics as the number of phases, the flow regime, the kind of bed
regions addressed, the type of reactions/processes considered, the mixing behav-
ior, the detailedness of the differential equations of mole- and energy balance. In
the review by Mahecha-Botero more than 26 different characteristics are outlined.
Macroscopic balance models are often very case specific. Only in very few inves-
tigations they are stated in three dimensions capturing transient behavior. Many

model assumptions are included which have to be derived from experimental in-
vestigations.
In multi-phase fluid dynamics the particle and fluid phase are modeled in the
framework of the Navier-Stokes equations using averaged quantities. Closure eq-
uations for the solid phase pressure and the solid phase viscosity have to be pro-
vided, they are derived from the kinetic theory of granular flow [9] which is an
extension of the classical kinetic gas theory. Although these models have been
frequently used for bubbling fluidized beds, mixing, down-flow reactors and
spouted beds, only few studies applied the model to reacting systems. This is due
to the fact that particle size shrinkage and different particle sizes as encountered in
combustion processes are difficult to capture in the Eulerian multi-phase frame-
work. Multi-phase fluid dynamics models can however be used e.g. for carbon
dioxide capture systems based on the chemical looping process involving a gase-
ous fuel. In the chemical looping process a solid carrier is providing the necessary
oxygen for the combustion as shown in figure 1b. Multi-phase fluid dynamics
models can be further enhanced if the Eulerian particle phase is linked with repre-
sentative Lagrangian particles [10]. Particle size distributions are realizable and
particle quantities like temperatures and species concentrations do not get blurred
through the averaged quantities available on the Eulerian level.
The discrete element method has been widely used for non-reacting fluidized
particle systems. For reacting systems only few examples can be found in litera-
ture which are two dimensional [11,12].
Moving Bed Thermal Conversion. Moving beds are relevant for gasification
reactors, cement kilns, down draft combustion and grate firing. Details on the
processes can be found e.g. in [13,14,15]. Models are available on different levels
of detail in a similar way as found for fluidized beds. Of importance in some
models is the assumption of spatial temperature and species profiles within the
particles. In contrast to other methods discrete element/CFD models can give de-
tailed insight into particle motion and conversion with the least number of as-
sumptions. Figure 2 gives details of a simulated lime shaft kiln modeled by
DEM/CFD.

Feed Stone
Energy
Cooling Air
a) b)
Fig 2. (a) Geometry of a lime shaft kiln and (b) steady state particle temperature
in a limestone calciner.
DISCUSSION AND CONCLUSIONS

Many different simulation methodologies are available for reacting fluid/gas
systems in energy technology. All frameworks have certain limitations. Contin-
uum models sometimes have the advantage to be able to address full size systems
and time scales. For discrete element models this is only true for exceptional
cases. Much more effort must be invested in comparing different approaches and
in the investigation of the ability of the models to complement each other.
REFERENCES
1. International Energy Agency, World Energy Outlook, Paris, France, 2007.
2. IPCC Fourth Assessment Report, Paris, France, 2007.
3. Smith PJ, Fletcher TH, and Smoot LD, Model for pulverized coal-fired reac-
tors, Proc. Combust. Inst 18, 1285–1293, 1981.
4. Zhang J and Zhou LX, Particle behaviour in a pulverised coal-fired sudden-
expansion combustor with coaxial jets, Fuel 80, 289–299, 2001.
5. Kurose R, Makino H, Suzuki A, Numerical analysis of pulverized coal com-
bustion characteristics using advanced low-NOx burner, Fuel 83 (6), 693-703,
2004.

6. Wang Y, Yan L, CFD Studies on Biomass Thermochemical Conversion, Int. J.

Mol. Sci. 9, 1108-1130, 2008.
7. Shen CY, Johnstone HF, Gas-Solid Contact In Fluidized Beds, AIChE Journal
1 (3), 349-354, 1955.
8. Mahecha-Botero A, Grace JR, Elnashaie SSEH, Lim CJ, Advances in model-
ing of fluidized-bed catalytic reactors: A Comprehensive Review, Chemical
Engineering Communications, 196 (11), 1375-1405, 2009.
9. Ding J, Gidaspow D, A bubbling fluidization model using kinetic theory of
granular flow, AIChE J. 32 (1), 523–538, 1990.
10. Snider DM, An incompressible three-dimensional multiphase particle-in-cell
model for dense particle flows, Journal of Computational Physics 170 (2),
523-549, 2001.
11. Zhou H, Flamant G, Gauthier D, DEM-LES simulation of coal combustion in
a bubbling fluidized bed Part I, Chem. Eng. Sci. 59 (20), 4193-4203, 2004.
12. Oevermann M, Gerber S, Behrendt F, Euler-Lagrange/DEM simulation of
wood gasification in a bubbling fluidized bed reactor, Partic.7 (4), 307-316,
2009.
13. Valipour MS, Saboohi Y, Modeling of multiple noncatalytic gas-solid reac-
tions in a moving bed of porous pellets based on finite volume method, Heat
And Mass Transfer 43 (9), 881-894, 2007.
14. Yin C, Rosendahl LA, Kaer SK, Grate-firing of biomass for heat and power
production, Prog. in Ener. and Comb. Sci. 34 (6), 725-754, 2008.
15. Bluhm-Drenhaus T, Simsek E, Wirtz S, Scherer V, A coupled fluid dynamic-
discrete element simulation of heat and mass transfer in a lime shaft kiln,
Chem. Eng. Sci. 65 (9), 2821-2834, 2010.

IMPLEMENTATION OF TANGENTIAL FORCE IN

3D DISCRETE ELEMENT AND COMBINED FINITE-
Jiansheng Xiang1, John-Paul Latham1, Ante Munjiza2,

1. Imperial College London, 2. Queen Mary, University of London
This paper describes the implementation of the tangential forces into

3D DEM and FEMDEM. Besides the classical Coulomb friction, the
authors have also implemented velocity and pressure dependent fric-
tion. Three verification tests show FEMDEM code implementation of
friction force in agreement with theoretical solutions. Finally, a rock
dumping simulation shows 3D FEMDEM code is capable of simulat-
ing complex multibody system with friction.
BACKGROUND
The discrete element method (DEM), sometimes called the distinct element
method, is becoming widely used in simulating granular flows. Pioneering work
in the application of the method to granular systems was carried out by Cundall
and Strack [1]. They developed the so-called 2D soft-sphere discrete element
method. However the capability of 2D DEM to model particulate phenomena of
practical interest is quite limited. An increase in computing power and the motiva-
tion from industrial applications has driven the surge in development of fully 3D
DEM. However, a problem has become apparent that may have been overlooked
when 2D DEM is extended to 3D DEM. This concerns the implementation of tan-
gential forces and these can’t be naturally extended from a 2D model. Contrary to
the normal force, the formulation of the tangential forces is either known as the

“History-Dependant Method" or “increment method" in the DEM community. In

2D models, the tangential direction can be easily decided by vertical clock-wise
normal direction or taking the direction of relative velocity in the tangential direc-
tion. But in a 3D model, two components are needed and an inappropriate imple-
mentation of friction force will lead to many problems and inaccuracy. Some re-
searchers have realized this problem e.g. Raji and Favier [2] employed a spherical
coordinate system to determine the tangential vector, but others haven’t noticed it
and still apply the 2D model approach.
An open source 3D FEMDEM code without friction was developed and used
to simulate stresses upon packing of arbitrary shaped rocks [3]. The friction force,
now implemented in this FEMDEM model is the subject of the present paper. Be-
sides the most frequently used classical law of Coulomb, various friction laws that
include static friction, velocity dependent and pressure-dependent friction laws
have been implemented.
MATHEMATICAL MODEL
The computation of the tangential component of the contact force, i.e. the fric-
tion forces, is an essential feature of any discrete element code. In the present pa-
per, we employ dry friction and Coulomb law which is described below.
Coulomb’s law
If the tangential force Ft is bigger than the friction force obeying the Coulomb-
type friction law, Ft ≥ µFn ,where µ is the coefficient of sliding friction, the parti-
cles slide over each other and the tangential force is calculated using the total
normal elastic contact force Fn : Ft = µFn . The normal contact force model can be
found in well documented DEM [1] and FEMDEM [4] publications.
In most current DEM codes, the classical Coulomb friction law in which the
friction coefficient is constant is applied. It has been found that for certain materi-
als in sliding contact, the coefficient of dynamic friction is not constant when rela-
tively low pressures act across the sliding surface, although pressure sensitivity is
negligible once high pressures are acting. The velocity dependence of friction is
perhaps better known, but often neglected in simulation technologies.

Implementation of tangential force

An elastic force-displacement model was used to calculate tangential forces.
The total displacement between two elements (two spheres in DEM) is initialized
to 0 once two elements are detected in contact. Then the tangential displacement
between the two overlapping elements is evaluated and incrementally increased
by the relative velocity between two elements. In order to calculate relative veloc-
ity, in FEMDEM it is assumed that the relative velocity of the contactor and the
target elements is the average velocity of the nodes describing those elements (in
DEM, the relative velocity can be exactly calculated at the contact point). In order
to calculate tangential force direction accurately, an arbitrary rotating axis system
is used and described in the full paper.
VERIFICATIONS OF MODELS
The FEMDEM simulation of a sliding cube is performed with length, 1 m,
density 2340 kg/m3, Young’s modulus 26GPa and Poisson’s Ratio 0.205.
Classical Coulomb friction law
10
Analytical results
9 Numerical results
7 Vt 2
L=
2 µg
Stop distance (m)
0
0 1 2 3 4 5 6
Velocity (m/s)
Fig. 1 Comparison of simulated and analytical results for stop distance of a

block as a function of Vt and µ for a classical Coulomb friction law.
The case of a cube which is paced on a horizontal plane with initial horizontal
velocities is carried out as a benchmark test for 3D FEMDEM with sliding fric-
tion. The stop distance L can be theoretically derived and is given by the equa-

tion, L = Vt 2 2µg , where µ is the coefficient of sliding friction, g is the gravity accel-
eration, and Vt is the initial horizontal velocity.
The cube is given an initial velocity Vt and then slows down due to the fric-
tional effect with the base. Finally it stops at a distance L. Results are compared
with the analytical solution (see Fig. 1). It shows that the numerical results are in
excellent agreement with the theoretical values.
Pressure-dependent Coulomb friction law
A pressure-dependent Coulomb friction law was also implemented and investi-
gated. The friction coefficient µ which is a function of contact pressure P, refer-
ence pressure p0, minimum µ m and maximum µ s_c sliding friction coefficient con-
ducted with different contact pressures P was able to accurately reproduce the
analytical expression, as shown in Fig. 2. It is worth mentioning that the friction
coefficient in the plot is derived from the ratio between normal and tangential
forces measured in the simulation for the different pressures.
0.45
Numerical results
Analytical results
0.40
0.35
0.30
n
Coefficient of friction
 p 
0.25 µ = µ m + ( µ s _ c − µ m ) 
 p0 
0.20
P0 = 2.6 × 10 3 Pa
0.15
n = −2
0.10
0.05
0.00
0.00E+00 5.00E+04 1.00E+05 1.50E+05 2.00E+05
Pressure (Pa)
Fig. 2 Comparison of simulated and analytical results for friction coefficient as

a function of contact pressure for the pressure-dependent Coulomb’s friction law.
Velocity-dependent Coulomb friction law
As there is a transition from a static to a dynamic frictional state, it is self-
evident that there is a velocity-dependence for the friction coefficient responsible

for actively resisting movement by sliding. A model considered appropriate for

capturing this behaviour is given in Fig. 3, in which the friction coefficient is cal-
culated by static friction coefficient µ s and sliding friction coefficient µ d. The plot
shows the transitions from peak to residual sliding friction for both directions of
motion is captured by the simulation and shown to follow the analytical velocity-
dependent Coulomb friction law used in the simulation.
0.5
Analytical results
Numerical results 0.4 µ = µ d + ( µ s − µ d )e − c V t
0.3 c=100
0.2
Coefficient of Friction
0.1
0
-6 -4 -2 0 2 4 6
Vt (m/s)
-0.1
-0.2
-0.3
-0.4
-0.5
Fig. 3 Comparison of simulated and analytical results for friction coefficient as

a function of constant horizontal velocity
EXAMPLE APPLICATION: ROCK DUMPING SIMULATIONS

The test conditions and set up for the rock dumping simulation were described
in [3] and the start position of dumping is as shown in the space filled array of
Fig. 4. When this rock dumping simulation is performed with zero friction it takes
many seconds to reach a state of rest as the only energy dissipation mechanism
implemented is by viscous damping of the elastic wave propagation. There is con-
siderable bouncing of the rock-like boulders and a high packing density is
achieved. The friction model implemented here is the classical Coulomb friction
law i.e. µ=constant. Setting the friction coefficient to 0.4 produces a rapid settling
to a state of rest, with a low packing density that preserves some of the original
boulder array structure (see Fig 4 left). The visualisation of the simulation is ex-
tremely realistic with dynamic stress levels captured throughout the process. The
simulation movie is now on the VGeST website (vgest.net/applications/silo-filling/).

Fig. 4 Rock dumping simulation, 5.3 metre wide bin, 288 rocks of 40 kg mass,
D = 0.31 m, friction coefficient µ = 0.4.
CONCLUSIONS
The implementation of the tangential forces into 3D DEM and FEMDEM has
been completed and will be described in the conference presentation and the full
paper. Besides the classical Coulomb friction, the authors have also implemented
friction models which account for some more complex observed frictional phe-
nomena such as dependence on relative velocity and contact pressure. A limited
series of simulations have been reported showing FEMDEM code implementation
of friction force in agreement with theoretical solutions. However, the authors are
under no illusions about the complexities associated with modeling frictional be-
haviour as it is very different for different application conditions. More work will
be needed on validation studies including physical experimental work.
REFERENCES
1. Cundall, P.A. and O.D.L. Strack, 1979. Discrete numerical model for granular
assemblies. Geotechnique, 29(1): p. 47-65.
2. Raji, A.O. and J.F. Favier, 2004. Model for the deformation in agricultural and
food particulate materials under bulk compressive loading using discrete ele-
ment method. I: Theory, model development and validation. Journal of Food
Engineering, 64(3): p. 359-371.
3. Xiang, J., Latham, J.P., Munjiza, A., Mindel., J. . 2008. Applications of the
combined finite-discrete element method. in Beijing DEM'08. 2008. Beijing.
4. Munjiza, A., 2004. The Combined Finite-Discrete Element Method: John
Wiley & Sons.

DEVELOPMENT OF VIRTUAL GEOSCIENCE

SIMULATION TOOLS, VGEST USING THE
COMBINED FINITE DISCRETE ELEMENT
METHOD, FEMDEM
John-Paul Latham1, Jiansheng Xiang1, Antonio Munjiza2

1. Imperial College London, Department of Earth Science and Engineering
2. Queen Mary College London, Department of Engineering
The VGeST is a suite of freely available Open Source simulation tools

based on the powerful combined FEMDEM discontinuum modelling
approach. Graphical user interface (GUI) functionality of leading pro-
prietary pre-processors can easily be exploited. The effort to set up in-
put files for the 2D and 3D FEMDEM solvers is minimized through
interactively completing drop-down templates and using meshing
tools in a pre-processor. In this paper we provide a taste of the work-
ing environment of VGeST, and an illustration of the power of the
FEMDEM codes Y2D and Y3D to tackle stresses and displacements
in blocky or layered systems and in addition, problems of a granular
nature which benefit from VGeST’s inclusion of a shape library of
virtual particles, P3D.
INTRODUCTION
Combined FEMDEM was proposed by Antonio Munjiza with the first working
FEMDEM code (the RG program written in C++) developed by Munjiza in 1990.
Combined FEMDEM allows the individual particle motion governed by DEM
formulations to be combined with an ability to discretise any particle into consid-

erably smaller deformable finite elements. By 2004, many of the key FEMDEM
algorithms had been published in journals. These developments and implementa-
tions were brought together and expanded upon in Munjiza’s textbook [1]. Impor-
tant advantages of FEMDEM over DEM models based on spheres, ellipsoids su-
perquadrics or clustered spheres include:
• complex particle shapes including non-convex and angular faceted ones
can be introduced,
• a vast range of alternative e.g. non-linear constitutive or internally fractur-
ing properties can be introduced for the individual particles,
• further field variables such as temperature can be added.
WHAT IS VGEST?
VGeST – a suite of Virtual Geoscience Simulation Tools for modelling discon-
tinuous systems, i.e. particulate, granular, porous, blocky, layered, fracturing and
fragmenting systems – is a computer software environment. A new website (and
name) was created to act as a forum from which users and developers can carry
forward research on code development and applications arising from the now
completed VGW (Virtual Geoscience Workbench) research project, a 5-year col-
laborative project developed on two sites (Imperial College London and Queen
Mary, University of London). The main objectives of VGW were to assemble
FEMDEM codes in both 2D and 3D, and most importantly, to make them avail-
able in Open Source format and to apply the codes to Geoscience. The suite of
simulation software, VGW v1.0 was launched on the Open Source repository,
Sourceforge.net, on March 30th 2009 and at the time of writing this continues to
be the current version available for users of VGeST. It is the first freely available
open source suite of 2D and 3D simulation tools based on FEMDEM. Complex
multi-body systems from which dynamic 3D stress components can be visualised
and interrogated can now be modelled. The software modules are managed
through Sourceforge and are available for research users and developers with
LGPL licenses. Fluid modelling coupled to FEMDEM was a key component of
the VGW project and this aspect will also be taken forward under VGeST.

V G eS T
released on Sourceforge.net
INPUT
Geometry Import Pre-processing
CAD GiD
MESHES
DFN B2D
B3D
P3D
SOLVERS
Discrete Solids Fluids
Y2D FLUIDITY
Y3D
CLDEM
VISUALISER & ANALYSIS TOOLS

MAYAVI, PARAVIEW, OPENDX
DOCUMENTATION
Manuals, Coding Standards, Examples,

Tutorials, Movies, Publications
Fig. 1 Contents of VGeST indicating software modules available for download

from Sourceforge.net in bold font.

VGEST COMPONENTS
To unblock the tedious error-prone data input phase for problem definition, the
considerable advantages of adopting the proprietary software, GiD as a pre-
processor (free for monthly trials and comparatively reasonably priced and devel-
oped by the International Centre for Numerical Methods in Engineering, CINME)
seemed to out way writing a customised graphics user interface (GUI) for
FEMDEM applications. Furthermore, off-the-shelf freeware post-processors Ma-
yavi and Paraview available from Sourceforge were also found to be excellent.
Mahabadi et al. [2] introduced a specifically designed GUI for geomechanical ap-
plications of the Y2D solver with a number of particular features, notably a het-
erogeneity module. Their user interface for 2D is named Y-GUI. There are there-
fore several options for users of the Y2D solver to build models.
The VGeST package contents are shown in Fig. 1. Several specific attributes of
the GiD package of great importance in harnessing the Y2D and Y3D solver tech-
nology are noted here:
• versatile handling of import and export file formats
• hierarchical geometry specification; point, line, plane, volume
• advanced meshing including unstructured and nonlinear tetrahedra
• customizable drop-down templates for problem condition setting
The input files required by the solver codes Y2D and Y3D are created in GiD us-
ing B2D and B3D. Drop-down templates acting as GUIs are integrated within
GiD. To get started with FEMDEM simulations:
• Download and install the VGW package from sourceforge.net
• Download and install GiD from gid.cimne.upc.es and Mayavi from ma-
yavi.sourceforge.net
• Copy the ‘FEMDEM’ folder which includes GiD templates (B2D.gid and
B3D.gid) to the GiD ‘problem types’ location
• Open GiD and Mayavi
• Build problem in GiD using templates
• Press ‘calculate’ and wait to view results in Mayavi

EXAMPLE: MULTI-BODY COLLISIONS IN 3D

A small multi-body system was simulated with www.VGeST.net [3] to demon-
strate the 3D FEMDEM code’s ability to model impacting complex-shape bodies.
This small system consists of two objects drawn by CAD (a sphere and a cube),
and two identical rock geometries obtained from a shape library of geometries
captured in our laboratory from pieces of rock aggregate using a laser scanner.
These rock geometries (seen as GiD NURB surfaces in Fig. 2) were chosen from
the shape library, P3D. Some VGeST applications are illustrated in Fig. 3.
(a) (b)
(c) (d)
Fig. 2 Multi-body problem: pre-processing in GiD and visualization in Mayavi (a)
typical template, (b) geometry: blue – lines, pink – surfaces, turquoise – volumes
(c) meshed volumetric elements (d) visualized 3D transient stress and motion.

(a) (b) (c)
(d) (e)
Fig. 3 FEMDEM applications [3]: (a) multilayer folding in alternately stiff and
weak layers, (b) rock slide avalanching, (c) silo discharging with stress chains and
bridging, (d) and (e) 200m cube of dissected rock mass subjected to boundary dis-
placements generating strong stress heterogeneity between different blocks [4].
REFERENCES
1. Munjiza, A., 2004. The Combined Finite-Discrete Element Method: John
Wiley & Son
2. Mahabadi O.K. Grasselli. G., Munjiza A. 2010. Y-GUI: A graphical user in-
terface and pre-processor for the combined finite-discrete element code, Y2D,
incorporating material heterogeneity. Computers & Geosciences 36, 241–252.
3. VGeST website: http://vgest.net, accessed 12.05.10.
4. Harrison, J.P. Latham, J.P. Xiang, J and Zhang, C. 2010. Heterogeneity of
stress field in a dissected rock mass modelled with the combined finite-
discrete element method. NARMS Conference, Salt Lake Cit, June 2010.

COUPLED GAS-PARTICULATE DISCHARGE FROM

BUCKET ELEVATORS
Matt Sinnott1, James Hilton1, William McBride2 and Paul Cleary1

1
CSIRO Mathematics, Informatics and Statistics, Melbourne, Australia
2
TUNRA Bulk Solids, University of Newcastle, Newcastle, Australia
The vertical conveying of bulk materials is often performed using

bucket elevators. Such devices use conveyor belts or chains to drag a
series of buckets carrying fine materials up and down the inside of an
enclosed chute. The speeds at which these buckets travel can displace
a large volume of air inside the chute, and the interaction between air
and bucket contents can lead to fines accumulation in the chute as well
as modify the discharge dynamics of the granular load. We present
here a first study of coupled gas-particle dynamics in the discharge of
such buckets.
INTRODUCTION
Bucket elevators are a common industrial device used throughout the bulk ma-
terials handling world. They offer a compact footprint for the vertical elevation of
a wide variety of bulk materials. There is very little restriction on the height of
elevation and tonnage rates are broadly scalable. A bucket elevator consists of a
conveyor belt (or chains) with a series of buckets bolted onto the belt. It is en-
closed in a vertical chute with the head-pulley at the top of the chute and beside a
splitter and exit chute that material is thrown into upon discharge.

In operation, the buckets within a bucket elevator pass through the lowermost
section of the elevator called ‘the boot’. The boot of the elevator is generally full
of the bulk material to be conveyed. This is supplied either from material being
fed to the downward strand or the upwards strand of the elevator and is added to
by spillage at the head pulley itself. Irrespective of the mode of feeding, at some
point soon after the boot of the elevator the buckets fill to a level commensurate
with the tonnage rate needed in operation. A previous Discrete Element Method
(DEM) study from McBride et al. (2009) investigated the effect of belt deforma-
tion (at the transition point from linear belt to head pulley) on material discharge.
The speed at which buckets are elevated upwards in the confined space of the
chute displaces a large volume of air. This gas may then enter the bucket, blowing
off surface layers of fine material which can settle into other buckets below or ac-
cumulate inside the boot. The gas will also apply a drag force on the stream of
discharging particles. This can modify particle trajectories and therefore influence
which particles are directed into the exit chute or fall back into the boot.
GAS-PARTICULATE COUPLED MODEL

We use DEM to model the granular material in the bucket and couple this to a
Finite-Volume implementation of a gas. The details of the gas model follow:
Drag Forces. The drag force is the dominant fluid-particle interaction force in
the system. The drag force exerted on a particle in a multi-particle system with
corrections due to Di Felice (1994) is:
FD = 12 cd ρ u r ε − χ A⊥u r
2
(1)
{
χ = 3.7 − 0.65 exp − (1.5 − log Re )
1
2
2
}
where cd is a drag coefficient, ρ the fluid density, ur the local relative flow veloci-
ty and ε the local voidage fraction. The local Reynolds number is given by Re =
2rρ|ur|/η where r is the equivalent spherical radius of the particle and η is the fluid
dynamic viscosity. We use the following drag coefficient for single particles given
by Holzer and Sommerfeld (2008), based on correlations to experimental data:

8  1 2  3 1
0.42 × 10 (
0.4 − log Φ )
0.2
cd =  + + + (2)
Re  Φ ⊥ Φ  Re Φ 4 Φ ⊥
3
where the sphericity, Φ , and the crosswise sphericity, Φ⊥, are both unity for spher-
ical particles. Particles are also subject to Stokesian rotational drag, given by TD =
8πηr3ωr.
Fluid dynamics. We use constitutive equations for gas flow through a porous
bed for modelling the system, given by Kafui et al. (2002). If the gas density is
assumed constant the expressions can be re-formulated for the superficial gas ve-
locity u′ = εu:
∂ε
= −∇ ⋅ u′
∂t
(3)
∂u′ 1 u′ 1
+ ( u′ ⋅∇ ) u′ + u′∇ ⋅ = − ε∇p + f fp + ∇ ⋅ ( ε τ )  + ε g
∂t ε ε ρ
where p is the fluid pressure, the stress tensor τ takes the usual form and the fluid
particle interaction body force, ffp, is the volume averaged force from the particle
drag on the fluid. These equations are discretised on a Cartesian grid over the
simulation domain. The porosity is explicitly calculated from the DEM simulation
at each timestep using a recursive Octree approach. The particle surface is
checked for intersection with a grid spanning the particle. Any part of the surface
within a grid cell is recursively subdivided down to some prescribed minimum
level. At the smallest cell size a linear approximation to the volume fraction is
used. Our method is fast and highly accurate.
The procedure for the numerical calculation is firstly to determine the porosity
at the new time step by calculating the volume distribution from the DEM par-
ticles. The velocity field is then calculated using Eq. (5), which gives the rate of
change of porosity as divergence of the velocity and the velocity field at the next
timestep. Eq. (5) is solved using a variation of the pressure correction method.

INFLUENCE OF GAS ON BUCKET DISCHARGE

The bucket was filled with a 5kg load of 200,000 DEM spheres with diameters in
the range 2.5-5mm with an equal mass distribution in each size class. We have
used a solid density of 1100 kg/m3, a coefficient of friction of 0.5 for particle-
particle and particle-boundary collisions, and a coefficient of restitution of 0.5 for
particle-particle and particle-boundary collisions. These correspond to reasonable
parameters for Polyethylene pellets. A spring constant of 2000 N/m was used, giv-
ing average particle overlaps < 0.2%. We specify a belt speed of 1.35 m/s for the
bucket motion. The bucket and chute assembly for the DEM simulation is shown
in Figure 1 at a couple of instances to show bucket motion and material discharge.
When the bucket reaches the head pulley, it tilts following the belt’s path around
the pulley. We observe material being centrifugally thrown from the rear of the
bucket. The bucket then fully tips over at the other side of the pulley and discharg-
ing material is thrown towards the splitter and exit chute.
Figure 1: Flow pictures demonstrating the discharge of the bucket where par-
ticles are coloured by speed and red corresponds to high speed flow.
The effect of the bucket and particulates motion on the gas flow field is shown
in Figure 2. We observe that there is a high speed gas flow dragged up behind the
bucket and strongly follows the motion of the bucket as shown in Figures 2a-c.
Gas is also pushed in front of the bucket and travels at the speed of the bucket,

except for a gas layer near the surface of the bucket that appears to be travelling
slower and is deflected upwards due to interaction with the shearing layers of dis-
charging material. In Figure 2d, the gas velocity vectors distort to follow the path
of the discharging granular stream rather than fully following the bucket.
(a) (b)
(c) (d)
Figure 2: Close-up view of the gas field interacting with the particulate flow.
We now compare two different DEM simulations, with and without the gas in-
cluded. Figure 3 shows the two cases superimposed where the case with gas is
coloured red and the no gas case is coloured blue. There are observable delays for
the discharging stream with coupled gas flow, even for the relatively large particle

sizes used here. Figure 3a shows the gas perturbing the surface material layer and
pushing it upwards away from the bucket. The discharging stream also exits the
bucket at a slightly different angle to the no gas case as shown in Figures 3a-b. In
Figure 3c-d, the bucket has fully tipped over and the bulk of the material centri-
fuges out and is thrown towards the exit chute and splitter. Here we see that the
drag on the particles by the gas is significant enough that the main discharging
stream for the case with gas can be observed lagging behind the case with no gas.
● DEM + gas
● DEM only
(a) (b) (c) (d)
Figure 3: Superimposed flow of 2 DEM simulations with (red) and without

(blue) the inclusion of the gas model.
CONCLUSION
A model of the interaction between gas in an enclosed chute and a high-speed
bucket elevator conveying and discharging granular spherical pellets has been
presented. We demonstrate the effect of the gas field on the discharging stream
out of the bucket and observe that the drag force on the particles is enough to
modify the discharge dynamics even for 2-5 mm particles.
REFERENCES
1. W. McBride, M.D. Sinnott and P.W. Cleary, CHOPS 09, 2009.
2. R. Di Felice, Int. J. Multiphase Flow, 20, p153, 1994.
3. A. Hölzer and M. Sommerfeld, Powder Technology, 184, p361, 2008.
4. D.K. Kafui, C. Thornton and M.J. Adams, Chem. Eng. Sci., 57, p2395, 2002.

LARGE SCALE DISCRETE ELEMENT MODELING

OF FINE PARTICLES IN A FLUIDIZED BED
M. Sakai, Y. Yamada, Y. Shigeto, K. Shibata, S. Koshizuka

Department of Systems Innovation, School of Engineering, The University of Tokyo
The Discrete Element Method (DEM) is widely used in computational

granular dynamics. The DEM is a Lagrangian approach where indi-
vidual particle is calculated based on the Newton’s second law of mo-
tion. Hence, the DEM enables us to investigate the granular flow
characteristics at the particle level. On the other hand, the DEM has a
difficulty to be used in large-scale powder systems because the calcu-
lation cost becomes too expensive when the number of particles is
huge. Consequently, we have developed a coarse grain modeling as a
large scale model of the DEM. The coarse grain particle represents a
group of original particles. The coarse grain model was applied in
typical gas-solid and solid-liquid two phase flows where the particle
size was relatively large. In the present study, the coarse grain model
is evolved to simulate fine particles by considering cohesive force,
namely the interparticle van der Waals force. The adequacy of the
coarse grain model is proved by comparing the simulation results of
original particle system. It is shown that the coarse grain model con-
sidering the interparticle force could simulate the original particle be-
havior precisely.
INTRODUCTION
It is well known that particle behavior depends on the properties. Geldart clas-
sified powders into four different groups based upon their fluidization behavior. In

these groups, the particles belonging to the Group B and D are regarded to be a
coarse particle. In the past studies, various numerical studies were made by focus-
ing on the Group B and D particles. The fluidized bed simulations were performed
by coupling Discrete Element Method (DEM) and Computational Fluid Dynamics
(CFD). Since the coupled DEM-CFD method was developed, various numerical
simulations were performed in fluidized beds. The DEM is a Lagrangian approach
where individual particle is calculated based on the Newton's second law of mo-
tion. The DEM enables us to investigate the granular flow characteristics at the
particle level. Therefore, the DEM solve lots of complicated problems related to
the granular flows. On the other hand, the numerical simulations of fine particles
belonging to Groups A or C were hardly performed unlike the Geldart B and D
particles. This is because the calculation cost becomes too expensive due to mod-
eling of various interparticle forces. Recent advances of computer performance
makes possible to simulate the behavior of the Geldart A particles, though the sys-
tem or number of the calculated particles is relatively small. Fundamental studies
of Geldart A particle behavior were made by the 2D and 3D DEM-CFD method.
In these studies, the cohesive particles were computed considering the interparti-
cle van der Waals force.
As addressed above, the computer simulation makes feasible to understand the
complicated phenomena related to fine particles. On the other hand, in the past
studies, small-sized fluidized beds were employed due to the drawback of such a
detailed description. Thereat, a large-scale discrete element model is developed to
simulate the fine particles. A coarse grain particle represents a group of the origi-
nal particles. In this study, 2D DEM-CFD simulations were performed to show
the adequacy of the coarse grain model. The adequacy of the model was proved
by the numerical examples, namely comparison of the result of the original parti-
cle system with those of the coarse grain model.
THE COARSE GRAIN MODEL

Originally, the coarse grain model was developed to simulate the non-cohesive
particles belonging to Geldart B and D in a large-scale system. In our previous
studies1), the contact force, drag force and gravitational force are taken into con-
sideration. A coarse grain particle represents a group of the original particles in
this model. In other words, the calculations are performed by using the coarse

grain particles literally instead of the original particles. In the present study, the
coarse grain model is evolved to consider the interparticle van der Waals force.
When the coarse ratio is l, namely size of the coarse grain particle is l times
larger than that of the original particle, l3 original particles exist in the coarse
grain particle. In this model, when the binary interaction of the coarse grain parti-
cle was made, l3 pairs of the original particles are assumed to interact simultane-
ously. The contact force acting on a coarse grain model is estimated by equating
the kinetic energy of the coarse grain particle with that of the original particles.
The drag force and the external force like gravitational force are modeled by bal-
ancing the coarse grain particle with the group of the original particles. Hence, the
drag force and external force acting on the modeled particle are proportional to the
number of the original particles. The van der Waals force is modeled that the po-
tential energy of the coarse grain particles is equated with that of the original par-
ticles. When the coarse grain particle i interact with j, binary interaction of all the
original particles is assumed to happen in the coarse grain particles.
Eventually, all the forces acting on the coarse grain particle is expressed as:
mCGM xCGM  F fCGM  VCGM p   FCCGM  FgCGM  FvwCGM
(1)
 l 3 F fO  l 3VO p  l 3  FCO  l 3 FgO  l 2 FvwO
where Ff, p, FC, Fg and Fvw indicate drag force, pressure, contact force, gravita-
tional force and van der Waals force. The van der Waals force is given by:
HrO
FvwCGM  l 2 n (2)
12hO
ALGORITHM
In this current study, the discrete particle equations of motion are first solved
yielding the particle positions and velocities. The porosity of each computational
cell is estimated based on the particle positions. The gas-phase hydrodynamic
equations are solved to give the gas velocity and pressure fields by using the
known values of particle velocity and the void fraction at the present time. This
procedure is iterated for each calculation cycle. A semi-implicit finite volume
method employing a staggered grid is used for discretising the incompressible

Navier-Stokes equation on a two dimensional Cartesian grid. The particle accel-

eration was calculated according to the contact force, drag force, gravitational
force and van der Waals force as mentioned above. The angular acceleration was
obtained from the torques.
NUMERICAL SIMULATIONS
The calculation domain was rectangular whose size was 60mm x 600mm. The
spherical particles were packed randomly as the initial conditions. Number of
grids was 15 x 150. The gas was injected from the bottom side by 0.1 m/s of su-
perficial velocity. The particle density was 1000 kg/m3. Coefficients of restitution
and friction were 0.9 and 0.3, respectively. The gas density and viscosity were 1.0
kg/m3 and 1.8×10−5 Pa s, respectively. The same values of the physical properties
were used in all the simulations. 8 cases of the calculation were performed in the
current study. In Case 1 to 4, the Hamaker number was set to be 1.0 × 10−20 J. In
Case 5 to 8, the Hamaker number was set to be 1.0×10−19 J. In Case 1 and 5, the
DEM-CFD simulations were performed in the original particle system, where par-
ticle size was 200 μm and number of the particles was 90000. The coarse grain
model was applied in Case 2, 3, 6 and 7. The coarse ratio was set to be 2.0 in Case
2 and 6. In these simulations, the diameter of the coarse grain particle was 400 μm
and number of the calculated particles was 22500. The coarse ratio was set to be
3.0 in Case 3 and 7. In these simulations, the diameter of the calculated particle
was 600 μm and number of the calculated particles was 10000. Effectiveness of
the coarse grain model was investigated in Case 4 and 8, where simply large-sized
particle was used without the coarse grain model. The size of the calculated parti-
cles was 600 μm in these calculations. The simulations were finished when the
actual time became 2.0 sec.

Figure 1 shows the typical snapshots of the simulation results in Case 1 to 4.
These pictures shows the particle behavior after 1.0 sec from the initial state.
Bubbles were observed in Case 1 to 3, whereas the particles hardly moved in Case
4. Quantitative evaluation of the coarse grain model was made by measuring the
bubble size, the bubble velocity and the bed height. The results oof the coarse
grain model were in good agreement with that in the original particle system.

Figure 2 illustrates the typical snapshots of simulation results in Case 5 to 8.

These pictures shows the particle behavior in 1.0 sec. Channeling was observed in
Cases 5 to 7, whereas the particles hardly moved in Case 8. Quantitative
evaluation of the coarse grain model was made by measuring the bed height. The
results obtained by the coarse grain model were in good agreement with that in the
original particle system.
It is shown that the coarse grain model could simulate the original particle be-
havior precisely. This is because all the forces acting on the coarse grain model,
namely the drag force, the contact force, the van der Waals force and the
gravitational force, could be modeled adequately.
CONCLUSION
The coarse grain model which is a large scale discrete element modeling of co-
hesive particles was developed in this study. The coarse grain model was applied
in a 2D fluidized bed. It is shown that the coarse grain model could simulate the
original particle behavior precisely.
ACKNOWLEDGMENTS
This study was supported in part by Global COE Program (Nuclear Education
and Research Initiative), MEXT, Japan.
REFERENCES
1. M. Sakai and S. Koshizuka ‘Large-Scale Discrete Element Modeling in Pneu-
matic Conveying’, Chem. Eng. Sci., 64, 533-539, 2009.

Fig 1. Typical snapshots of simulation results (Case 1 to 4)
Fig 2. Typical snapshots of simulation results (Case 5 to 8)

A PRELIMINARY STUDY ON MODELING LIQUID-

SOLID INTERACTION USING SMOOTHED
PARTICLE HYDRODYNAMICS
Jihoe Kwon, Heechan Cho, Hoon Lee

Seoul National University, Department of Energy Resources Engineering
As a preliminary study to model liquid-solid interaction using

smoothed particle hydrodynamics, a different approach was taken to
determine drag force acting on the rigid body. The developed SPH al-
gorithm was tested for flows past a circular cylinder at various Rey-
nolds numbers. Model is verified by comparing with existing experi-
mental and numerical results for wake length and drag coefficient.
Drag forces acting on the cylinder were determined by using the ve-
locity distribution of neighboring fluid particles on the surface of cy-
linder. Drag coefficient from the simulation showed good agreement
with standard experimental curve under the condition of of Reynolds
number range from 1 to 100.
INTRODUCTION
Liquid-solid flow systems are widely used in mineral processing industries in-
cluding milling, separation, and material transport processes; however mathemati-
cal and computational analysis of these systems presents numerous difficulties.
The most challenging problem in this area is how to analyze and model the inte-
raction of two phases, solid and liquid. The motion of solid particulate materials
affects the liquid flow and is also influenced by the flow fields through interactive
effects; thus, the moving particulate materials can be considered to create an un-

steady boundary condition for fluid flow. Thus, conventional CFD methods based
on spatial grids and meshes suffer from fundamental limitations in simulating this
changeable flow domain system.
In this study, we used SPH (Smoothed Particle Hydrodynamics) to model liq-
uid phase kinematics. SPH is a Lagrangian CFD method in which flows are ana-
lyzed by tracing the movement of fluid material points without using any spatial
grid. It was originally invented to solve astrophysical problems in three-
dimensional open space. In its early days SPH was applied to the analysis of
compressible flows [1, 2, 3]; at present, the application areas of SPH are being
expanded to various incompressible flow problems such as free surface flow [4 ,5],
Stokes flow [6], and flow through porous media [7].
Also there have been few studies on modeling the liquid-solid interaction using
SPH-DEM [8, 9] or SPH-SPH [10]. However, their works were based on the indi-
rect assumptions or using one-way coupled model which do not represent the inte-
raction of two phases.
In this study, a different approach was taken to model liquid-solid interaction
by directly calculating the stress and the pressure of fluid particles neighboring on
the surface of rigid body. The developed SPH algorithm was tested for two-
dimensional flows around circular cylinders at various Reynolds numbers. Drag
forces acting on the cylinder were evaluated and the drag coefficient from the si-
mulation was compared with standard experimental curve.
SIMULATION METHOD
The flow around a cylindrical particle was modeled in two dimensions. Simu-
lations were executed under the conditions of Reynolds number ranging from 1 to
100. The diameter of the cylinder was set at D=5 mm and v0, the bulk velocity of
water varied from 0.2 to 20 mm/s. In all cases, a problem domain with dimensions
of a height of 6 times the diameter and a length of 10 times the diameter was filled
with fluid particles, and the particles inside the cylinder were then excluded. All
the fluid particles were initially at rest, and an external acceleration was applied to
reach the bulk velocity of water, v0 after 5 s. Then, simulations were continued for
a further 5 s, after which the fluid velocity distribution and drag force acting on
the cylinder were evaluated.

The flow was assumed to be an unbounded external flow. To model the conti-
nuous flow, fluid particles crossing the outlet boundary were immediately tele-
ported to the inlet. Imaginary particles were located around the boundary of prob-
lem domain to set the fluid velocity at the boundary of the domain as , and also
set around the surface of the cylinder to satisfy the no-slip condition at the solid
boundary.
The focus of this work is in determining drag force using SPH based on
straightforward approach. In SPH, determination of drag force is relatively diffi-
cult compared to conventional CFD because interpolation points are not fixed on
the space. However, if the drag can be determined using SPH, it is possible to
model interaction between solid and liquid without any conflict. In conventional
CFD based on spatial grid, although we can calculate drag force straightforwardly,
it is very difficult to move solid body because it leads to the redetermination of
spatial grid for every timestep. In SPH, on the other hand, when drag force is de-
termined, it is relatively easy to change in position of immersed body. This is one
of the advantages which SPH has compared to conventional CFD for modeling
liquid-solid flow.
Takeda et al. [11] and Potapov et al. [8] have attempted to determine drag force
using SPH; however, their approaches were based on the indirect assumption that
the drag force acting on the rigid body can be approximated by the sum of interac-
tion forces between real fluid particles and imaginary particles in the body. To
determine drag force straightforwardly, we classified fluid particles into 72 groups
with their angle θ in the cylindrical coordinate system (5° for each). The closest
particle to the surface of cylinder was determined for each group. or the stress
that acting on the cylindrical surface, can be expressed as
s = (s n + s q ) (1)
where the normal stress s n and the tangential stress s q are as below:
¶vr 3 mv
s n = t rr + p × n = - m ( 2 ) + [ p0 - rgR cosq - ( 0 ) cosq ] × n (2)
¶r 2 R

¶ vq
s q = t rq = - m[r ( )] (3)
¶r r
Derivative terms were evaluated using the linear approximation of fluid veloci-
ty from the solid surface to the neighboring fluid particles. The term is
related with buoyancy, and the integration of p0 along the surface should be zero.
The drag force acting on the infinitesimal area dA is
v r 3 mv 0 1 v
f d = { -2 m - ( ) cos q × n - m[ R ( q )]}dA (4)
Dr 2 R Dr r
Dr = R - r (5)
v
where R is radius of the cylinder, and r , q , v r and q are coordinates and ve-
locity vectors of neighboring fluid particles. There are 72 neighboring fluid par-
ticles; thus the infinitesimal area dA is approximated as below:
pRDz
dA @ DA = (6)
72
Solving the above equations algebraically, the drag force acting on the cylinder
can be evaluated.
RESULT AND DISCUSSION

Simulation results for the various Reynolds numbers are shown in Fig.1. With
alteration of Re, clearly different flow patterns are observed. In the case of low-Re
laminar flow (Re=1), streamline goes around the surface and no wake occurs. In
the cases of flows under the conditions of Re≥10, boundary layer separation
points were formed and gradually moved upstream with the increase of Reynolds
number. Wake length increases with the increase of Reynolds number. Fig.2
shows the wake length as a function of Reynolds numbers compared against the
results from previous experimental and numerical studies. Simulation result cor-
responds to the result from previous studies with acceptable degree of accuracy.

(a) (b)
(c) (d)
Fig.1. Vector expression of fluid velocity at various Re values: (a) Re = 1, (b)
Re=10, (c) Re=30 and (d) Re=100
CD from the simulation at various Re values were evaluated and compared with
the standard experimental curve (Fig.3). Resolution studies showed that with the
increase of Reynolds number, resolution required to model the flow increases. For
a low-Re flow (Re=1), CD from simulation with 8,872 particles shows good
agreement with experimental data. Under the conditions of Re=3 and 10, 35,500
particles are required to get a result that corresponds to the experimental data with
acceptable error. Simulation result converged to the standard experimental curve
when 142,008 particles are used under the conditions of Re=30 and 100. With the
increase in resolution, simulation result shows a good convergence and corres-
ponds well to the standard experimental result.

2.5
Coutanceau & Bouard
D/L ,htgneL ekaW sselnoisnemiD

2.0 Present (SPH)
Dennis & Chang
Fornberg
He & Doolen
1.5
1.0
0.5
0.0
0 10 20 30 40
Reynolds Number, Re
Fig.2. Wake length as a function of Reynolds numbers: compared against the

experimental data by Coutanceau & Bouard [12] and numerical results by Dennis
& Chang [13], Fornberg [14] and He & Doolen [15]
103
Standard Experimental Curve

SPH (with 8,872 particles)
SPH (with 35,500 particles)
102 SPH (with 142,008 particles)
D
C ,tneiciffeoC garD
101
100
10-1
0 1 10 100 1000
Reynolds Number, Re
Fig.3. Comparison of the Drag Coefficient for a circular cylinder with standard
experimental curve from Hoerner [16]

CONCLUSION
As a preliminary study to model liquid-solid interaction, drag force acting on a
circular cylinder were modeled under the conditions of Reynolds number range
from 1 to 100 using SPH. Model is verified by comparing with existing
experimental and numerical results for wake length and drag force. Drag
coefficient from computational result shows good agreement with the standard
experimental curve. Resolution studies showed that with the increase of Reynolds
number, higher resolution is required to model the flow.
The result of this preliminary study are promising and indicate that SPH may
be an useful tool for modeling liquid-solid interaction. Ongoing work includes
flow around a square cylinder and settling of spheric particle in three dimensions.
REFERENCES
1. L.B. Lucy, ‘A numerical approach to the testing of the fission hypothesis’, As-
tronomical Journal, 82, 1013-1024, 1977.
2. R.A. Gingold, J.J. Monaghan, ‘Smoothed particle hydrodynamics: Theory and
application to non-spherical stars’, Monthly Notices of the Royal Astronomical
Society, 181, 375-389, 1977.
3. J.J. Monaghan, ‘Smoothed particle hydrodynamics’, Annual Review of Astro-
nomical and Astrophysics, 30, 543-574, 1992.
4. J.J. Monaghan, ‘Simulating free surface flow with SPH’, Journal of Computa-
tional Physics, 110, 399-406, 1994.
5. A. Colagrossi, M. Landrini, ‘Numerical simulation of interfacial flows by
smoothed particle hydrodynamics’, Journal of Computational Physics, 191,
448-475, 2003.
6. J.P. Morris, P.J. Fox, Y. Zhu, ‘Modeling low Reynolds number incompressible
flows using SPH’, Journal of Computational Physics, 136, 214-226, 1997.
7. J.P. Morris, Y. Zhu, P.J. Fox, ‘Parallel simulations of pore-scale flow through
porous media’, Computers and Geotechnics, 25, 227-246, 1999.

8. A.V. Potapov, M.L. Hunt, C.S. Campbell, ‘Liquid–solid flows using smoothed
particle hydrodynamics and the discrete element method’, Powder Technolo-
gy, 116, 204-213, 2001
9. P.W. Cleary, M. Sinnott, R. Morrison, ‘Prediction of slurry transport in SAG
mills using SPH fluid flow in a dynamic DEM based porous media’, Minerals
Engineering, 19, 1517-1527, 2006
10. H.H. Bui, K. Sako, R. Fukagawa, ‘Numercial simulation of soil-water interac-
tion using smoothed particle hydrodynamics(SPH) method’, Journal of Ter-
ramechanics 44, 339-346, 2007
11. H. Takeda, S.M. Miyama, M. Sekiya, ‘Numerical simulation of viscous flow
by smoothed particle hydrodynamics’, Progress of Theoretical Physics, 92,
939-957, 1994
12. M. Coutanceau, R. Bouard, ‘Experimental determination of the main features
of the viscous flow in the wake of a circular cylinder in uniform translation.
Part 1. Steady flow’, Journal of Fluid Mechanics, 79, 231-256, 1977.
13. S. C. R. Dennis, G. Z. Chang, ‘Numerical solutions for steady flow past a cir-
cular cylinder at Reynolds number up to 100’, Journal of Fluid Mechanics, 42,
471-489, 1970.
14. B. Fornberg, ‘A numerical study of steady viscous flow past a circular’, Jour-
nal of Fluid Mechanics, 98, 819-855, 1980.
15. X. He, G. Doolen, ‘Lattice Boltzmann method on curvilinear coordinates sys-
tem: flow around a circular cylinder’, Journal of Computational Physics, 134,
306-315, 1997.
16. S.F. Hoerner, Fluid-Dynamic Drag, published by the author, 1958.

A COUPLED DEM/CFD STUDY OF SUCTION

FILLING
Yu Guo, Chuan-Yu Wu, Colin Thornton

School of Chemical Engineering, University of Birmingham, Birmingham, B15 2TT, UK
A numerical method based on coupling the Discrete Element Method

(DEM) with Computational Fluid Dynamics (CFD) was established to
simulate suction filling of powders. The effect of suction on the pow-
der flow behaviour was examined. The numerical analysis was found
to be in good agreement with previous experimental observations,
demonstrating that the coupled DEM/CFD was capable of modeling
the rapid and smooth flow process during suction filling.
INTRODUCTION
Suction filling is a typical stage in some manufacturing processes, such as in
pharmaceutical tabletting and powder compaction in ceramics and powder metal-
lurgy. During suction filling, a movable punch, which initially occupies the die
cavity, moves downwards to create a partial vacuum environment below the pow-
der mass. The induced pressure gradient ‘sucks’ the powder into the die. Previous
experimental studies showed that the powder flow rate could be significantly im-
proved in suction filling [1]. In this study, suction filling was numerically ana-
lyzed using a coupled DEM/CFD approach, which has been proved to be robust in
modeling powder flow in the presence of air [2, 3]. The effect of suction kinemat-
ics on the powder flow behavior was then investigated by comparing suction fill-
ing with conventional gravity filling.

THE COUPLED DEM/CFD METHOD

The Discrete Element Method (DEM) is employed to model the dynamics of solid
particles. The translational and rotational motions of each particle are governed by
Newton’s equations of motion. The particles considered in this study are assumed
to be elastic and frictional, and the particle interactions are according to classical
contact mechanics [4], in which the theory of Hertz is used to determine the nor-
mal force and the theory of Mindlin and Deresiewicz is used for the tangential
force. The air-particle interaction force can be described in the form
fapi = −vpi ∇p + vpi ∇ ⋅ τ + ε fdi (1)
where vpi is the volume of particle i , and p , τ , ε and fdi are the local air pres-
sure, viscous stress tensor, void fraction and drag force, respectively. The air is
treated as a continuum and modelled using Computational Fluid Dynamics
(CFD), in which the continuity and momentum equations
∂ ( ερa )
+ ∇ ⋅ ( ερa u ) = 0 (2)
∂t
∂ ( ερ a u )
+ ∇ ⋅ ( ερ a uu ) = −∇p + ∇ ⋅ τ − Fap + ερ a g (3)
∂t
are solved to obtain the air density ρa and air velocity u. In Eq. (3), the fluid-
particle interaction force per unit volume, Fap , is obtained by summing up the
fluid-particle interaction forces fapi acting on all the particles in a fluid cell, nc ,
and dividing by the volume of the fluid cell ∆Vc, thus
Fap = (∑ nc
f
i =1 api ) ∆V
c (4)
The detailed description of this coupled DEM/CFD approach was provided in [5].
A 2D numerical model of suction filling is shown in Fig. 1a. The top of the
punch, which is modeled using a physical wall, is initially located at the level of
the die opening. During the filling process, the punch moves down at a constant
velocity (vp), allowing the deposition of powder into the die. A die of dimensions
2×4 mm is filled by a monodisperse powder system with a particle diameter of 50
µm and a particle density of 1500 kg/m3 from a 5 mm wide shoe. The spherical

particles are assumed to be elastic with Young’s modulus of 8.7 GPa and Pois-
son’s ratio of 0.3. The physical walls are also elastic with Young’s modulus of
210 GPa and Poisson’s ratio of 0.3. Both interparticle and particle-wall friction
coefficients are set to 0.3. The powder bed is color-banded in order to visualise
the macroscopic flow patterns.
A CFD scheme is incorporated to consider the effect of suction. Fig. 1b shows
the schematic diagram of the computational mesh of the fluid field and the bound-
ary conditions. The field is partitioned uniformly using identical cells. The die and
shoe walls are treated as impermeable static wall boundaries without slip (denoted
as Cell 3). The top of the system is open to the atmosphere, so that continuous air
flow boundary (Cell 6) is assigned. During the suction filling process, the air
flows into the die with the downward movement of the punch but it can not pene-
trate the top of punch. Hence, the cells, which are located just below the top of
punch, are assigned as impermeable moving wall boundary cells (Cell 2). On this
boundary, the normal air velocity is equal to the punch velocity (vp) and the tan-
gential air velocity is zero (no-slip boundary). The cells below the moving wall
boundary in the die are temporarily inactive, and they will be activated gradually
as the punch moves downwards.
Shoe
Top of punch
Die
Inactive cells
vp
Punch
(a) (b)
Fig. 1 (a) The numerical model for suction filling with a stationary shoe and (b)
schematic diagram of the computational mesh of the fluid field and surrounding
boundary conditions.

The procedure to update the fluid cells in the die region during suction filling is
illustrated in Fig. 2. As the top of punch moves down from the current layer to the
layer below, the newly approached cells are activated as interior fluid cells (1). At
the same time, the cells immediately below the interior fluid cells are set to be im-
permeable moving wall cells (Cell 2). This procedure continues until the top of
punch reaches the bottom of the die, and then the punch stops and the imperme-
able static wall boundary with zero air velocity (Cell 3) is specified at the bottom.
Top of punch vp
vp
Fig. 2 The update procedure of the fluid cells in the die region.

Figure 3 shows a comparison of the powder flow patterns during gravity filling
in air and suction filling with a punch velocity of vp=276 mm/s. The cumulative
mass of particles delivered into the die as a function of filling time is plotted in
Fig.4 for the different filling processes. During gravity filling in air, the air can be
entrapped inside the die and the build-up of air pressure hinders the flow of pow-
der into the die (Fig.3a). Two air bubbles are then formed close to the die walls,
resulting in a thin powder flowing stream at the centre. Therefore, a much lower
filling rate is obtained for gravity filling in air (Fig.4). As the air bubbles rise up,
the deposition of powder is inhibited further (Fig.3a) and some particles can even
be blown out from the top of die, leading to a slight decrease in mass of particles
in the die (Fig.4). When the air bubbles move to the shoe region, the powder can
flow into the die again and the mass of particles in the die increases but at a slow-
er rate. During suction filling, a continuous draw of powder plug into the die is
obtained (Fig.3b), and a much higher filling rate is achieved by suction filling
(Fig.4).

30.7 ms 56.2 ms 97.2 ms 230.1 ms
(a)
3.4 ms 6.8 ms 11.1 ms 17.9 ms
(b)
Fig.3 Power flow patterns during (a) gravity filling in air and (b) suction filling
with vp= 276 mm/s.
Fig.4 Time evolutions of cumulative Fig. 5 Air pressure distribution at

mass of particles deposited into the die. t= 11.1ms during suction filling.

The air pressure distribution at a specified time instant of t=11.1 ms during

suction filling is shown in Fig. 5. It is observed that a lower air pressure environ-
ment is developed in the die region. A pressure gradient is then created with the
air pressure decreasing from the top to the bottom, which facilitates the powder to
flow into the die, so that a very high filling rate is obtained in suction filling
(Fig.4).
CONCLUSIONS
A coupled DEM/CFD method has been employed to model suction filling. It is
found that a lower air pressure environment is created below the powder mass
with the downward motion of punch. As a result, the filling rate is significantly
improved with suction filling, as observed experimentally.
ACKNOWLEDGMENTS
This work is funded by EPSRC through an EPSRC Advanced Research Fel-
lowship awarded to CYW (Grants No: EP/C545230 and EP/C545249).
REFERENCES
1. S. Jackson, I.C. Sinka, A.C.F. Cocks, ‘The effect of suction during die fill on a
rotary tablet press’, European Journal of Pharmaceutics and Biopharmaceu-
tics, 65, 253-256, 2007.
2. Y. Guo, K.D. Kafui, C.-Y. Wu, C. Thornton, J.P.K. Seville, ‘A coupled
DEM/CFD analysis of the effect of air on powder flow during die filling’,
AIChE Journal, 55, 49-62, 2009.
3. Y. Guo, C.-Y. Wu, K.D. Kafui, C. Thornton, ‘Numerical analysis of density-
induced segregation during die filling’, Powder Technology, 197, 111-119,
2009.
4. C. Thornton, K.K. Yin, ‘Impact of elastic spheres with and without adhesion’,
Powder Technology, 65, 153-165. 1991.
5. K.D. Kafui, C. Thornton, M.J. Adams, ‘Discrete particle–continuum fluid
modelling of gas-solid fluidised beds’, Chemical Engineering Science, 57,
2395-2410, 2002.

A COMBINED CONTACT MODEL IN CDEM AND

ITS APPLICATION IN BLASTING ENGINEERING
Chun Feng, Shihai Li, Xiaoyu Liu

Institute of Mechanics，Chinese Academy of Sciences
To improve the efficiency of contact detection and simplify the calcu-

lation steps of contact forces, a new type of vertex-face & edge-edge
combined contact model is presented. According to thus model, the
contact force could be calculated directly (the information of contact
type is not necessary), and the failure judgment could be executed
straight (each face vertex and face edge own their characteristic areas).
A numerical case in blast engineering is demonstrated to show the va-
lidity and accuracy of this model.
INTRODUCTION
Continuum-based Discrete Element Method (CDEM) is a numerical approach
to simulate the progressive failure of geological mass [1], while false contact and
block embedding will take place when the large translation and rotation of blocks
occurs. To solve the problem above-mentioned, a combined contact model is in-
troduced to CDEM.
The direct method, common plane method, penetration edge approach, and in-
cision body scheme are the 4 typical way to detect the contact relationship be-
tween blocks [2-6], while the final destination of contact detecting is to calculate
the contact force. Four steps should be taken if any method above-mentioned is
adopted: (1) contact state judgment (if contact or not); (2) contact type determina-

tion; (3) contact area calculation; and (4) normal and tangent contact force solving.
Each step costs lots of calculating time, so it is an important thing to give an effi-
cient contact model and simplifying the computation steps.
CONTACT DETECTION METHOD

To find the contact relationship between blocks, vertex-face and edge-edge
combined contact model is established. Different from other contact detection
methods, vertexes and edges of each block is set on the face (shown in Figure 1),
named face vertex and face edge respectively. For an example, a cubic contains 24
face vertexes and 24 face edges. The distance from face vertex to block vertex is
about 1%-10% of the distance from block vertex to face center, and each face
edge consists of two face vertex. According to this approach, vertex-face and
edge-edge contact could be easily founded.
Face center
Face edge
Face vertex
Block vertex Block edge
Fig 1. Face vertexes and face edges in a block
Figure 2 shows the relationship between mother face and target face. Because
the face vertex is set on the face, when search the target face for face vertex A,
face A is the mother face, block A is the mother block, Face B is the target face,
block B is the target block, and the interpolated node in face B locates the same
position with face vertex B. When search the target face for face vertex B, the op-
posite situation takes place.

Face Vertex A
Face vertex B
Face A Face B
Block A Block B
Contact spring
Fig 2. Relationship between mother face and target face
For searching the vertex-face contact, face vertex linked list is created. When
the distance from face vertex (from mother block) to face (from quasi target
blocks) is less than the limit value and the face vertex is also in the quasi target
face, this face is considered the target face and the interpolated node, interpolated
coefficient is calculated then. To avoid face vertex imbedding into a target block,
once a face vertex is inside a block, search the face nearest to the face vertex im-
mediately and set this face as target face, then calculate the interpolated node, in-
terpolated coefficient.
For searching the edge-edge contact, face edge linked list is created, and the
main steps are similar to the vertex-face contact model. The key step of edge-edge
contact model is to determine the distance and intersection point between two dis-
senting beelines.
DEMOLITION OF COOLING TOWER UNDER BLASTING

The height of this cooling tower is 123.2m, with the bottom diameter 90.7m
and top diameter 54.1m. The property of the block is shown in table 1, and the
property of contact is the same as the block.
The collapse of the cooling tower after blasting lasts about 11 s, the same as
the actual process. The numerical result is shown in figure 3, and the actual proc-
ess is shown in figure 4.

Density Elastic modulus Poisson ratio Cohesion Friction Tension Dilation
kg / m3 GPa / MPa Deg MPa Deg

2550 38 0.25 6 45 4 15
Table 1. Parameter of block
(a) 1 s after blasting (b) 3 s after blasting
(c) 5 s after blasting (d) 7 s after blasting

(e) 9 s after blasting (f) 11 s after blasting

Fig 3. Numerical result
(a) 1s after blasting (b) 3s after blasting (c) 5s after blasting
(d) 7s after blasting (e) 9s after blasting (f) 11s after blasting
Fig 4. Actual process

The collapse process of numerical simulation is consistent with the actual.

When the bottom surface of cooling tower touches the ground, the further move-
ment of this part is prohibited. While the opposite part still keeps the same speed
because of inertia, then the compression line occurs and compression-shear failure
takes place.
CONCLUSION
A combined contact model is introduced, which consists of vertex-face contact
model and edge-edge contact model. Different from other models, the vertex and
edge is set on the face and owns the characteristic area.
Based on the contact model above-mentioned, the numerical study about the
demolition of cooling tower under blasting has been carried out. The numerical
case not only shows the validity and accuracy of the combined contact model, but
also explains the failure mechanism of cooling tower under dynamic loads.
REFERENCES
1. Shihai Li, Xiaoyu Liu, Tianping Liu, et al. Continuum-based Discrete Element
Method and its Applications [C].Beijing, DEM’08, 2008:147-170.
2. S. Amir, Roozbeh, Kianoosh. Three-dimensional discontinuous deformation
analysis (3-D DDA) using a new contact resolution algorithm [J]. Computers
and Geotechnics, 2008, 35: 346-356.
3. P. A. CUNDALL. Formulation of a Three-dimensional Distinct Element
Model Part I. A Scheme to Detect and Represent Contacts in a System Com-
posed of Many Polyhedral Blocks [J]. Int. J. Rock Mech. Min. Sci. & Geomech.
Abstr, 1988, 25(3): 107-116.
4. Itasca Consulting Group Inc. 3DEC-3D distinct element code [R]. Minneapo-
lis, USA: Itasca Consulting Group Inc., 1987.
5. Y. M. Cheng, W. S. Chen, Y. H. Zhang. New Approach to Determine Three-
Dimensional Contacts in Blocks System: Penetration Edges Method [J]. Int. J.
Geomech. 2006, 6(5): 303-310.
6. Wang Jianquan, Lin Gao, Liu Jun. Static and dynamic stability analysis using
3D-DDA with incision body scheme [J]. Earthquake Engineering and Engi-
neering Vibration, 2006, 5(2): 273-283.

COMPUTATIONAL INVESTIGATION OF THE

DISPERSION OF COHESIVE AGGREGATES
Graham Calvert, Ali Hassanpour and Mojtaba Ghadiri

Institute of Particle Science and Engineering, SPEME, University of Leeds, Leeds, LS2 9JT, UK
The aerodynamic dispersion of bulk powders is important for a num-

ber of applications including particle characterization, and the delivery
of drugs via the lungs using Dry Powder Inhalers (DPIs). In this pa-
per, the Distinct Element Method (DEM) and Computational Fluid
Dynamics (CFD) are used to investigate the dispersion of cohesive
aggregates in a uniform fluid flow. As expected, with increasing co-
hesion, it is progressively more difficult to disperse a loose aggregate.
However, once the relative particle-fluid velocity is beyond a thresh-
old, dispersion occurs quickly and approaches a completely dispersed
state asymptotically. A dimensionless relationship between the fluid
drag force and cohesion force, given by the JKR model, which in-
cludes the Weber number, is introduced and linked to the dispersion
behaviour.
INTRODUCTION
The aerodynamic dispersion of bulk powders is important in many industries
such as pharmaceutical, bulk chemical and food. Currently, dispersion in the gas
phase is receiving much attention with respect to particle characterisation and
drug delivery via the lungs from dry powder inhalers (DPIs). The ability to con-
trol dispersion of a wide range of powders is an area of great interest and im-
portance. However, it is recognised that the complete dispersion of fine cohesive
powders is difficult due to the relatively large interparticle attraction forces, name-
ly van der Waals, electrostatics and liquid bridge [1], compared to separating
forces. Consequently, a great amount of energy is necessary to deform and disin-
tegrate particle clusters completely to their primary constituents.

Recently, Calvert et al. [2] reviewed aerodynamic dispersion of cohesive pow-

ders and highlighted the current state of theoretical and experimental understand-
ing. Loosely aggregated particles suspended in fluid flows experience several
kinds of forces caused by rapid acceleration, deceleration, turbulent eddies, impact
on surfaces etc.
In this paper the behaviour of a single loose aggregate cluster consisting of a
number of particles that are accelerated in a uniform flow field is investigated.
Furthermore, a dispersion model based on the balance between adhesive and dis-
ruptive forces, with the latter caused by the fluid drag force acting on the loose
aggregate, is introduced.
DISPERSION MODEL
Aggregate dispersion occurs when the separating force acting on it exceeds the
force which is binding the primary particles. Considering the motion of a sphere
in a fluid, the drag force acting on an aggregate is given by:
 DAgg
2
F  CD'  f ur2 (1)

4
where F is the drag force acting on the particle, DAgg is the loose aggregate diame-
ter, ρf is the fluid density, CD’ is the drag coefficient, and ur is the velocity of the
fluid relative to the particle. In the range of aggregate size and fluid velocity of
interest for dispersion, the particle Reynolds number is within the Newtonian re-
gime; therefore, the drag coefficient can be regarded as a constant (CD’ = 0.44).
To investigate the effect of surface energy on loose aggregate dispersion be-
haviour, the model of Johnson et al. [3] (JKR model) is used to define the pull-off
force necessary to break an interparticle contact, POFF,
3
POFF   Γ R * (2)
2
where Γ and R* are the interface energy and the reduced radius of the two particles
in contact, respectively. For surfaces of the same material the interface energy, Γ
= 2γ, where γ is the particle surface energy. Furthermore, if the loose aggregate is

composed of monodispersed particles R* = R / 2 = d / 4, where d is the primary

particle diameter.
At this point, the dispersion ratio, DR, is introduced and defined as the ratio of
the number of broken bonds to the initial number of bonds. DR = 1 means no
bonds are present and the cluster has fully dispersed.
N 0  Nt
DR  (3)
N0
Additionally, the dispersion ratio, DR, is proportional to the ratio of the fluid
drag acting on the spherical loose aggregate to the cohesive interparticle bond:
2  f DAgg
2
ur2
DR  CD' (4)
3 Γd
Now considering the terms in Eq. (4), it is possible to introduce a dispersion in-
dex, DI, by dropping the constant terms and grouping the terms which define the
Weber number:
 f DAgg ur2  DAgg   DAgg 

DI     We   (5)
Γ  d   d 
Therefore, the dispersion ratio is given by:
  DAgg 
DR  f  We   (6)
  d 
Hence, it is of interest to see whether the DI, which contains the Weber num-
ber, is related to particle dispersion in an accelerating fluid flow field.
SIMULATION DETAILS
To investigate the relationship described by Eq. (6), 500 particles were ran-
domly generated in a spherical volume and brought together by applying a gravi-
tational force to a point in space until a stable coordination number was achieved.

The resulting spherical cluster has an aggregate diameter of 0.966 mm. Surface
energy and particle friction was then progressively introduced between particles to
the required level. Subsequently, the gravitational force was gradually removed
from the particle assembly. The dispersion of this particle cluster has been inves-
tigated for a range of interface energies, 0.1 – 1.0 J/m2. Once the clusters are pre-
pared, they are subjected to a uniform flow field with various fluid velocities. The
particles and fluid properties used are shown in Table 1.
Diameter (μm) 100
Density (kg/m3) 2500
Stiffness (N/m) 7.9×104
Contact damping (-) 0.16
Friction coefficient (-) 0.3
Fluid density (kg/m3) 1.225
Fluid viscosity (Pa.s) 1.8×10-5
Table 1. Particle and fluid properties.

The dispersion ratio as a function of the relative velocity between the particles
and the dispersing fluid is shown in Fig 1 for a range of interface energies. At low
relative velocities the dispersion ratio is not sensitive to the fluid force, but once a
threshold relative velocity is exceeded, the dispersion ratio quickly rises and ap-
proaches unity in an asymptotic fashion. As intuitively expected, with an increase
in interface energy it becomes increasingly difficult to break bonds and disperse
the particle assembly.

1.0
0.8
0.1 J/m²
0.6 0.2 J/m²
DR (-)
0.3 J/m²
0.4 0.4 J/m²
0.2 0.5 J/m²
1.0 J/m²
0.0
0 20 40 60 80 100 120 140 160
Relative velocity (m/s)
Fig 1. Relationship between dispersion ratio and relative velocity between the
fluid and particles for different values of interface energy after 3×10-4 s.
In Fig 2, the dispersion ratio, DR, has been plotted as a function of the dimen-
sionless dispersion index, DI, for different interface energies after 3×10-4 s. It is
seen that there is a remarkable unification of the data obtained for the different
interface energies. This indicates that the DI can be used to define the onset of
cohesive aggregate dispersion due to acceleration by a fluid flow.
1.0
0.1 J/m²
0.8 0.2 J/m²
0.6 0.3 J/m²
DR (-)
0.4 J/m²
0.4 0.5 J/m²
1.0 J/m²
0.2
0.0
1 10 100 1000 10000
DI (-)
Fig 2. Relationship between dispersion ratio and relative velocity between the
fluid and particles for different values of interface energy after 3×10-4 s.

CONCLUSION
Dispersion by acceleration in a uniform flow field is seen to be effective when

the relative particle-fluid velocity is beyond a threshold value. In this case,
dispersion of the aggregate occurs quickly and approaches a completely dispersed
state asymptotically. The effect of particle surface energy has been investigated
and as expected a greater relative velocity is necessary to disperse the aggregate
with increasing particle bond strength.
To analyse the effect of surface energy and loose aggregate diameter on parti-
cle dispersion, a model has been developed that relates the fluid resistance force
acting on a sphere to the bond strength binding the particles. This approach leads
to a dimensionless dispersion index which incorporates the Weber number for the
aggregate and a second parameter for the size ratio between the aggregate diame-
ter and primary particle diameter. The results of the numerical simulations have
shown a good correlation between the number of broken contacts and the particle
surface energy and loose aggregate diameter, as defined by the dispersion index,
DI = We(DAgg/d).
ACKNOWLEDGMENTS
The financial support from the Engineering and Physical Science Research
Council (EPSRC) and Malvern Instruments is greatly appreciated.
REFERENCES
1. J. Visser, Van der Waals and other cohesive forces affecting powder fluidiza-
tion, Pow Technol, 58(1): 1-10, 1989.
2. G. Calvert, M. Ghadiri and R. Tweedie, Aerodynamic dispersion of dry pow-
ders: A review of understanding and technology. Adv Pow Technol, 20(1): 4-
16, 2009.
3. K. L. Johnson, K. Kendall and A. D. Roberts, Surface energy and the contact
of elastic solid, Proceedings of the Royal Society of London, Series A: Math-
ematical, Physical and Engineering Sciences, 324(1558): 301-313, 1971.

VERIFICATION AND VALIDATION

Simulating and optimizing of a ball mill by comparison between

numerical discrete element method (DEM) and experimental method
B. Arabzadeh a, V. Hasanzadeh a, A. Farzanegan b

a
M. Sc. student of mineral processing engineering. Department of Mineral
Processing Engineering, Mining Engineering Faculty, Tehran University, Iran.
b
Professor of mineral processing engineering. Department of Mineral Processing
Engineering, Mining Engineering Faculty, Tehran University, Iran.
Nowadays, valuable simulations in types of mills for approaching to

optimized efficiency have been drastic concern for mineral processing
engineers in industrial plants. During the last ten years, a lot of studies have
done on simulation and optimization of some types of mills. In this paper,
the purpose is studying a three dimensional transparent ball mill in
laboratory. The features of the mill are considered similar to “bond mill”.
And using lucent liner makes a proper condition for assessing ball
movements in length section of the mill. Then, the simulation of mill by
discrete element method (DEM) is done; and, by comparison between
numerical and experimental simulations the validation of DEM has been
proved. Afterwards, optimizing of mill features has done by using DEM.
The time for optimizing of mill features by numerical method is not as long
as by experimental studies; therefore, we considered all features
combinations for reaching to an adequate and optimized condition in our ball
mill.
Key words: ball mill; transparent; lucent liner; discrete element method.

NUMERICAL AND EXPERIMENTAL INVESTIGATION OF

PROGRESSIVE DEVELOPMENT OF A GRANULAR PILE OF
BINARY SIZE PELLETS
Yaowei Yu* and Henrik Saxén
Thermal and Flow Engineering Laboratory, Department of Chemical Engineering, Åbo Akademi University,
Biskopsgatan 8, FI-20500 Åbo, Finland. *Corresponding author: E-mail: [email protected]
Granular conical piles consisting of a large number of particles of different size are
frequently encountered in industry, and size segregation is known to occur as the piles
are being formed. This work studies the shape, repose angle and size segregation of a
pile of binary size pellets in small-scale experiments as well as computationally using
the discrete element method (DEM). The simulations were found to be consistent with
the experiments, which indicate the applicability of DEM for predicting the
conditions at pile formation of spherical particles.
INTRODUCTION
The storage and conveying of granular materials in sandpiles and stockpiles are
important industrial and engineering operations in dealing with granular matter, and
the behavior (rolling, percolation, segregation, etc.) of the granular materials is of
interest in the field of soil mechanics, chemical engineering, powder technology,
pharmaceutical engineering, environmental sciences and mineral processing. A deeper
understanding of the static and dynamic mechanisms of pile formation could
contribute remarkably to the development and enhance the efficiency of various
production technologies.
Along with significant progress in the computational power and algorithms, the
discrete element method (DEM) is becoming a mature and feasible numerical method
for analyzing discontinuous media since its original development by Cundall and
Strack.[1] Today DEM calculations can be made even on a personal computer.
Recently, DEM has been extensively applied to simulate different granular flows in
the industries, modeling drum mixers[2], fluidized beds[3] and hopper charging and
discharging flow.[4] Herrmann[5] proposed a simple 1-D model to calculate the

reposed angles of a conical pile considering the segregation of different types of

particles. Matuttis[6] modeled granular piles in two dimensions using the molecular
dynamics method with convex polygons and pointed out that the repose angle is
dependent on the size dispersion of the particles and the minimum pressure appears
under the apex of the pile. Other researchers, including Zhou et al.[7], Brockbank et
al.[8], Majmudar,[9] Cates et al.[10], Zuriguel[11] and Oron[12] focused on the force
(stress or pressure) distribution in the pile. Although many researchers have studied
conical[5] and wedge-shaped piles[13] the shape and repose angles of the pile, as well
as the size distribution along the radial direction have not attracted much attention in
DEM studies. In order to understand the behavior of the particles in the pile in more
detail, experimental work combined with numerical modeling should be applied.
DISCRETE ELEMENT METHOD

In the DEM, the inter-particle contact model applied is shown in Fig. 1. The contact
is composed of a spring and a dashpot which correspond to the elastic and plastic
nature of particles in the normal direction, respectively. In the tangential direction, the
model consists of a slider, a spring and a dashpot. All particles in a granular system
undergo translational and rotational motions which are generally described by
Newton’s second law of motion.
Fig.1 Simulation model for interaction forces between two contacting particles
The motion of each particle is tracked, and interaction with other particles or
boundaries is considered in the simulation. Forces (normal elastic force, tangential
elastic force, tangential force, normal force and gravitational force) and torques are
calculated based on interaction of particles and the physical properties of the entities,

including the hardness of particles, expressed with a spring, and the particle energy
dissipation, expressed with a dashpot.[14] The hardness of the particles and the
dashpot are related to the Young’s modulus and the coefficient of restitution,
respectively. The friction between entities is defined with a Coulombic type of friction,
limited below by the maximum static friction and implemented with a friction factor
based on the physical properties. The particle forces and torques are calculated which
lead to new particle movement and new positions. Additionally, all the forces and the
torques involved in motion are given by the EDEM[15] software applied in this study.
The general procedure for each run of simulation of the formation of the pile was
as follows: As for the shape simulation, 5501 spherical particles (1.5 mm diameter)
and 3116 spherical particles (3.5 mm diameter) were mixed uniformly and adequately
and charged without overlap into the hopper before the simulation began. The
particles were then allowed to flow out under gravity (with the gravitational
acceleration g = 9.81m/s-2 in the negative z axis direction) from the exit of the hopper
at the bottom, and formed a pile on a horizontal steel plane until the hopper was
emptied. In order to reduce size segregation of particles during charging into the
hopper, a special filling method was applied, namely the layer (fines over coarse) as
shown in Fig. 2B. As for size distribution simulation a transparent plane (plexiglass, II
in Fig. 2A) was inserted along the radial direction and the particles were discharged
into one side of the transparent plane through an inclined plane (I) at the exit of the
bottom of hopper. The simulation was conducted with a constant time step (4×10-5 s)
and the numerical solution required about 3 hours of computation time on a personal
computer.
EXPERIMENTAL METHOD AND EQUIPMENT

The experiments were divided to two parts: the evolution of pellet pile and size
distribution in the pile. The evolution of the pellet pile was observed by using a high
speed digital video recorder as shown in Fig. 2.C. Before the experiment, 5501
spherical pellets (1-2 mm in diameter) and 3116 ones (3-4 mm in diameter) were
mixed fully and put into the hopper. A stopper at the bottom was removed and the
particles exit the bottom and formed a pile on the steel plane, as shown in Fig. 2.C.
The video camera was used to film the pile formation process.
The shape of the pile was characterized by two parameters, namely the diameter

of the bottom of the pile (D) and the height of the apex from the bottom (H). One
ruler in the horizontal and another in the vertical
A B C
Fig. 2 Geometry of calculation area (A: whole structure of simulation, B: fines over
coarse in hopper, C: schematic of experiment setup)
direction (cf. Fig. 2.C) were used to measure D and H, respectively. From the
diameter and the height, repose angle (a) of the conical pile can be calculated from
equation 1. When the experiment was completed, the pictures of evolution were
obtained from the video camera and were compared with the simulated results.
𝐻𝐻
𝑎𝑎 = atan �𝐷𝐷/2� (1)
The procedure of determining the size distribution along the radial direction of the
pile is the same as in the simulation of this part. The camera recorded the video of the
pile along the radial direction. The pile was separated into five subregions along the
radial direction as shown in the uppermost panels of Fig. 5. For every subregion, j, the
length of which was chosen as six times of the large size particle (i.e., 6d2), the mass
ratio of the particles with small diameter was calculated as
𝑚𝑚 s 𝑁𝑁
𝑥𝑥𝑗𝑗 = 𝑚𝑚 𝑀𝑀+𝑚𝑚 ; j=1, 2, …,5 (2)
l s 𝑁𝑁
where ms and ml are the mass of one small pellet and one large pellet, respectively,
and N and M denote the number of small and large pellets, respectively, and their
values were obtained by calculating their numbers on the pictures.
COMPARISON BETWEEN SIMULATION AND

EXPERIEMNTAL RESULTS
The evolution of the pellet pile profiles can be observed from the photos taken

during the experiments and by images from the simulations, as shown in Fig. 3a-f,
corresponding to different times, in which the left profiles (A) are those of the
experiment and the right ones (B) are those of the simulations. In the simulation, 67
particles (62 small and 5 large ones) came out the calculating area and the situation is
similar to what was observed for the experiment, where 58 (7 large and 51 small)
were not recorded by the video camera. From the photographs in Fig. 3a-f, the
experimental flow rates out of the hopper is seen to rise at the beginning (t = 0.5 s)
and keep almost the same value with the simulation before t = 3.2 s and the rate
reduces clearly for t > 3.2 s. The changes of the simulated flow rates are similar to
those of the experiment. Comparing subfigures A and B, the profiles of the piles are
quite similar between experiment and simulation from t = 1.2 s to t = 5.21 s, except at
0.5 s: The diameter of the bottom of the pile (D) at different times has almost the
same values in the simulation and the experiment. Thus, the snapshots indicate a good
consistency in terms of the shape of pile between experiments and simulations.
A comparison of the diameter, width and repose angle of the pile for the
experiments and the simulations is presented in Fig. 4. The diameter, width and
repose angle all increase with the growth of the pile. The repose angle reaches its
maximum at t = 3.2 s, the diameter its maximum at t = 5.2 s and the height at t = 4.2 s.
This indicates that pellets slide along the inclined surface of cone and the inclined
surface collapses at t = 3.2 s, so the repose angle decreases and the diameter grows.
Comparison between experiment and simulation shows good agreement between
repose angles, and also for pile width and diameter, although some occasional
deviations are observed for the latter ones, e.g., at t = 2.2 s and at t = 3.2 s. Still, the
results indicate overall agreement between experiments and simulations.
Figure 5 shows a comparison between video images of experiments and simulation
screenshots studying the size distribution of the pellets along the radial direction at
different times. In the beginning, there are only individual coarse particles distributing
along the surface of the transparent glass. The fines start to flow out from the exit of
the hopper and pile on the large particles at t = 2.1 s. With increasing number of small
particles, they mainly percolate into the space between the coarse ones and distribute
in the centre part of the pile (cf. t = 2.1 s and 3.5 s). At the same time, there is not
much space for large particles to percolate into the pile, so they slide along the free
surface and primarily distribute at the bottom and the free surface of the pile. This is
consistent with general observations of segregation is piles and also with what was

proposed by Drahun and Bridgwater.[16].
6.29 cm
3.55 cm
a. (t = 0.5 s) b. (t = 1.2 s)
8.18 cm
7.32 cm
c. (t = 2.2 s) d. (t = 3.2 s)
9.97 cm
10.05 cm
e. (t = 4.2 s) f. (t = 5.2 s)
Fig. 3 Comparison of pellet piles in experiments and simulations (Left: experiment, Right:
simulating)
There is almost the same distributions of fines and coarse in the pile in the
experiments and simulations. An overall comparison of size segregations is plotted in
Fig. 6.
Figure 6 plots the size segregation of pellets along the radial direction at t = 15.4 s
for the experiment and at t = 15.9 s for the simulation. The mass fraction of fine
particles represents only about 30% at the bottom of the pile while it is about 70% at
the apex. Conversely, the shares of large particles are 70% and 30% at the pile
bottom and apex, respectively. Comparing the results for the five subregions (cf.
topmost part of Fig. 6), the experimentally determined fractions of fine particles at the

left part of the pile are larger than the corresponding simulated values, while the
opposite holds true for the right part of the pile.
1.0
The height and width (H:cm)
10 0.7 0.9
Exp.
0.6 0.8 Sim
Fine mass fraction (x,%)

8 0.7
0.5
Repose angle (a:arc)

0.6
6 0.4 0.5
4 0.3 0.4
0.3
0.2
2 0.2
H in Exp. H in Sim
D in Exp. D in Sim 0.1 0.1
a in Exp. a in Sim
0 0.0 0.0
0 1 2 3 4 5 Left bottom Mid Centre Mid Right bottom
Time (s) Position of pile
Fig. 4 Comparison of H, D and a of pellet pile Fig.6 Size segregation along the radial direction
in experiment and simulation at t = 15.4 s for experiment and t = 15.9 s for simulation.
This may be due to some asymmetry in the flows of the experiment. However, on the
whole, there is a good agreement between the experiments and simulations, which
indicates that DEM can be used to predict the properties of a pellet pile.
CONCLUSTIONS
The present work has used the discrete element method (DEM) and experimental
methods to investigate the behavior of binary size pellets at the formation of a pile.
The shape of pile (the height of the apex and the width of the bottom) as well as the
size segregation along the radial direction were measured in the experiment and
compared with results from DEM simulations. From the presented results the
following conclusions can be drawn.
1. The discrete model used for the binary size pellet pile prediction was
successfully validated through comparison between results from experiments
and simulations. It was demonstrated that DEM can be applied to predict the
evolution of pile formation, including the outflow of the particles from the
feeding hopper.
2. In the pile formation process, the fines mainly percolate into the space
between the coarse particles and concentrate in the centre part of the pile,

while coarse particles slide along the free surface, move toward the bottom of
the pile and primarily distribute at the bottom and on the free surface of the
pile.
a. t = 0.7 s in experiment, t = 0.7 s in simulation b. t = 2.1 s in experiment, t = 2.1 s in simulation
c. t = 3.5 s in experiment, t = 3.5 s in simulation d. t = 5.3 s in experiment, t = 5.4 s in simulation
e. t = 8.6 s in experiment, t = 9.5 s in simulationg. t = 10.0 s in experiment, t = 14.3 s in simulation
h. t = 15.4 s in experiment, t = 15.9 s in simulation j. Size distribution

Fig. 5 Comparison between experimental and simulation size distribution pellets
along the radial direction at different times. The topmost panels show the radial
regions studied in Fig. 6.

REFERNCES
[1] Cundall PA, Strack ODL. A discrete numerical model for granular assemblies. Geotechnique.
1979;29(1):9.
[2] Stewart RL, Bridgwater J, Zhou YC, Yu AB. Simulated and measured flow of granules in a bladed
mixer—a detailed comparison. Chemical engineering science. 2001;56(19):15.
[3] Kaneko Y, Shiojima T, Horio M. DEM simulation of fluidized beds for gas-phase olefin
polymerization. Chemical engineering science. 1999;54:13.
[4] Nguyen VD, Cogné C, Guessasma M, E. Bellenger a JF. Discrete modeling of granular flow with
thermal transfer: Application to the discharge of silos. Applied thermal engineering. 2009 Discrete
modeling of granular flow with thermal transfer: Application
to the discharge of silos;29:8.
[5] H.J.Herrmann. Shapes of granular surfaces. Physica a. 1999;270:6.
[6] Matuttis H-g. Simulation of the pressure distribution under a two-dimensional heap of polygonal
particles. Granular matter. 1998;1:9.
[7] Zhou YC, Xu BH, Zou RP, Yu AB, Zulli P. Stress distribution in a sandpile formed on a deflected base.
Advanced powder technology. 2003;14(4):10.
[8] Brockbank R, Huntley JM, R.C.Ball. Contact force distribution beneath a three-dimensional
granular pile. Journal de physique II. 1997;7:12.
[9] Majmudar TS, Behringer RP. Contact force measurements and stress-induced anisotropy in
granular materials. Nature. 2005;435(23):4.
[10] Cates ME, J.P.Wittmer, Bouchaud JP, Claudin P. Development of stresses in cohesionlesspoured
sand. Philosophical transactions of the royal socity. 1998;356(22):27.
[11] Zuriguel I, Mullin T. Effect of particle shape on the stress dip under a sandpile. physcial review
letters. 2007;98:4.
[12] Oron G, H.J.Herrmann. Exact calculation of force networks in granular piles. physical review E.
1998;58(2):11.
[13] Li E, Bagster DF. A new block model of heaps. Powder technology. 1990;63:7.
[14] Renzo AD, Maio FPD. Comparison of contact-force models for the simulation of collisions in
DEM-based granular flow codes. Chemical engineering science. 2004;59:17.
[15] WWW.dem-solutions.com. DEM Solutions Limited.
[16] Drahun JA, Bridgwater J. The mechanisms of free surface segregation. Powder technology.
1983;36:15.

PARTICLE SHAPE EFFECTS IN MEDICAL

SYRINGE NEEDLES – EXPERIMENTS AND
SIMULATIONS FOR POLYMER MICROPARTICLE
INJECTION
Mark Whitaker1, Paul Langston2, Steve Howdle1, Barry Azzopardi2

1. University of Nottingham, Chemistry
2. University of Nottingham, Faculty of Engineering, Process & Environmental Division
Experiments on injection of PLGA polymer particles have shown that

jamming is a problem under some conditions. It is postulated that par-
ticle shape could be a significant factor. 2D DEM simulations of cir-
cles, and semi-circles approximated by rounded polygons, are used to
demonstrate the effect of shape on jamming. Preliminary experiments
on glass spheres and polymer hemi-spheres show good agreement
with simulations of normally distributed particle sizes.
INTRODUCTION
Medically approved polymers are commonly used to make drug encapsulated
microparticles for subcutaneous injection into the human body1. Such particles are
usually spherical and are made by double emulsion2 and spraying techniques3. Of-
ten overlooked is the administration of these microparticles, via hypodermic sy-
ringe with a medical gauge needle. Large needle sizes are often employed to
eliminate blockages; however this approach is cautious, when perhaps smaller
sizes can be used to improve patient comfort. Therfore, the effects of particle
shape and size need to be understood when they pass through a needle orifice.

Two types of particle shape are investigated, spherical and hemi-spherical; to

see what effect they have in terms of jamming. Both these shapes have initially
been made using the Particles from Gas Saturated Solution Technique, PGSS,
where supercritical CO2 under mild temperatures and pressures has been used to
atomise the polymer Poly(lactic-co-glycolic acid), PLGA, into microspheres
though the aid of a nozzle4. These particles can be further processed to make he-
mispheres. A hemi-spherical particle has the benefit of secondary plasticisation,
which reduces its internal pore size giving a better release characteristic over a
spherical one.
Both types of particles have blocked under initial injection testing with narrow
gauge needles (25 & 23G) compared with larger ones (21 & 19G). A balance
must be struck between using the narrowest needle to minimize discomfort for the
patient and the largest available particle size that can be injected. Generally, the
larger the particle, the more likely the encapsulated drug will remain in the poly-
mer- preventing a dose dump effect after injection.
This study compares results of polygon DEM 2D simulations with experiments
on syringe flow in terms of whether jamming occurs. It is recognized that 2D has
significant limitations but previous studies5,6 have shown it to be useful in initial
assessments. Comparable 3D simulations are being developed. Several techniques
to model non-spheres are evolving in DEM as reviewed by Dziugys and Peters7.
POLYGON DEM MODEL

The basic form of this DEM model6 is fairly standard with a linear spring and
dashpot contact mechanics approach and has been applied to hopper flow. It is
essentially a 2D model with a fixed particle thickness in the 3rd dimension. An
example of the polygon DEM model used here to model the hemi-spheres is
shown in Fig 1. It should be noted that the vertices are “rounded”. This is to facili-
tate the model and to acknowledge that real granular particles do not contain per-
fect vertices. In this study cohesion is negligible, contact damping is assumed
quite high due to the interstitial liquid, and a moderate value of friction is as-
sumed. Gravity flow is used to replicate the syringe flow. The side walls of the
DEM vessel are at 4° to vertical measured on a needle cross-section.

Fig 1. Snapshot of 2D DEM model replicating syringe flow showing blockage

of Circles and Semi-Circles in 23G Needle, and close-up of modeling polygon.
EXPERIMENTS
Injection Testing Procedure: The injections must perform within the accept-
able user limits (10N) for a 1mL syringe8; if this value was exceeded then the in-
jection was deemed a blockage. A TA HD Plus texture analyser (Stable Micro
Systems, Goldaming, UK) was used to measure the force of injection as a func-
tion of syringe plunger movement.
The particles were suspended in an injection vehicle (Water with 1.5% Car-
boxymethyl cellulose, 0.9 % Sodium Chloride), with a dose of 200 mg/mL. The
temperature of the injection vehicle was controlled at 10 °C giving a viscosity of
0.13 ±0.01 Pa.s. Three repeats were taken for each sample tested.
Modelling and Experiment: Monodisperse and normal distributions were
simulated using DEM. Distributions were then experimentally tested with a 23G
needle. To achieve quick results glass spheres were used in a variety of sieved dis-
tributions, whereas polymer hemi-spheres were made using the secondary plasti-
cisation process. The size range of the nominal sieved fractions were more accu-
rately measured using laser diffraction and were approximated by normal distribu-
tions truncated to describe 80% of the particle population in the simulation.

Fig 2 Left: SEM of glass spheres Sieved in the nominal range 75-80 µm.
Right: PLGA Hemispheres Sieved in the nominal range < 100 µm.
MODEL AND EXPERIMENTAL RESULTS

Mono-sized simulation. Table 1 shows the critical particle size to just cause
jamming for the simulations of near mono-sized systems. These show that the
semi-circles are more likely to block in terms of particle size based on area ie
amount of material in the particle. The size ratio needle to particle is fairly con-
stant.
Needle Circles Semi-Circles
Gauge B (µm) dp (µm) B/dp dp (µm) B/dp de (µm) B/de
25 260 67 3.88 88 2.96 62 4.19
23 340 91 3.73 112 3.04 79 4.30
21 510 127 4.02 168 3.04 119 4.29
19 690 222 3.11 247 2.79 175 3.94
Table 1. DEM predictions for blockage of circles and semi-circles for four
common medical gauge needles. B: Needle inner diameter. dp: Critical particle
diameter. de: Equivalent area circle diameter for semi-circles.

Normal Particle Size Distribution. Tables 2 and 3 show a comparison of ex-

perimental and 2D DEM results of spherical and hemi-spherical particles. These
show a very good agreement for the glass spheres and reasonable agreement for
polymer hemi-spheres.
Nominal Sieve Range d10 (µm) d90 (µm) DEM Experiment
(µm)
Set I: 75-80 50.32 85.47 F 3*F

Set II: 75-80 62.25 87.66 B 3*B
Table 2 Comparison of DEM and Experiment for a range of glass sphere distribu-
tions through a 23G needle – d10 10% of particles below this size, d90 90% of par-
ticles below this size, F indicates flow, B indicates blockage, 3 * F indicates 3 re-
peated experiments all showing flow.
Nominal Sieve Range d10 (µm) d90 (µm) DEM Experiment
(µm)
<100 35.75 108.94 F 2 *F & 1*B

100-150 35.97 143.96 B 3*B
Table 3 Comparison of DEM and Experiment for a range of polymer hemi-sphere
distributions through a 23G needle.
CONCLUSION
The flow and jamming phenomena in the syringe injection of polymer micro
particles is reasonably represented by the 2D circle and polygon DEM simula-
tions. DEM simulations of near mono-sized particles showed that the semi-circles
are more likely to jam as expected from previous experiments. Experiments here
injecting glass spheres and polymer hemispheres gave reasonable agreement with
the DEM. Further work is required using 3D DEM.
ACKNOWLEDGMENTS
We would like to thank BBSRC and Critical Pharmaceuticals for financial
support for this project.

REFERENCES
1. M.J. Rathbone, J. Hadgraft, M.S. Roberts, M.E. Lane, Modified-Release Drug
Delivery Technology, 2nd ed. 2008.
2. Zhang , J.X., and Zhu, K. J. , An improvement of double emulsion technique

for preparing bovine serum albumin-loaded PLGA microspheres. Journal of
Microencapsulation, 2004, Vol. 21, No. 7, Pages 775-785.
3. J. Kerc, S. Srcic, Z. Knez, P. Sencar-Bozic, Micronization of drugs using su-

percritical carbon dioxide. International Journal of Pharmaceutics, 182(1), 33-
39, 1999.
4. H. Jianyuan, M.J. Whitaker, B. Wong, G. Serhatkulu, K.M. Shakesheff, S.M.
Howdle, Plasticization and spraying of poly (DL-lactic acid) using supercriti-
cal carbon dioxide: Control of particle size, Journal of Pharmaceutical
Sciences, 93(4), 1083 -1090, 2004.
5. F.Y. Fraige, P.A. Langston, G.Z. Chen, „Distinct element modelling of cubic
particle packing and flow‟, Powder Technology, 186, 224-240, 2008.
6. F.Y. Fraige, P.A. Langston, A.J. Matchett, J. Dodds, „Vibration induced flow
in hoppers: DEM 2D polygon model‟, Particuology, 6, 455-466, 2008.
7. A. Dziugys, B. Peters, „An approach to simulate the motion of spherical and
non-spherical fuel particles in combustion chambers‟, Granular Matter, 3,
231-265, 2001.
8. ISO 7886-1 1993 (E), Sterile hypodermic syringes for single use - Part 1: Sy-
ringes for manual use.

EXPERIMENTAL VALIDATION OF POLYHEDRAL

DISCRETE ELEMENT MODEL
Stuart Mack1, Paul Langston2, Colin Webb1, Trevor York3
1. School of Chemical Engineering and Analytical Science, University of Manchester
2. University of Nottingham, Faculty of Engineering, Process & Environmental Division
3. School of Electrical and Electronic Engineering, University of Manchester
Experiments on small scale 3D hopper flow of polyhedral dice are

compared with DEM simulations of rounded polyhedra. Reasonable
agreement is shown in terms of flowrate and critical orifice required
for flow. All DEM parameters were set up to match the particle prop-
erties prior to undertaking hopper experiments.
INTRODUCTION
DEM is now widely used to model granular flow. Most models consider par-
ticles as spheres and produce useful results, but an increasing number model non-
spheres1. This can extend the applicability of the technique and aid our under-
standing of the important physics. This paper summarises a method in which par-
ticles are modelled as polyhedra with rounded edges and vertices.
Laboratory experiments are very useful in helping to show the validity of mod-
els such as DEM2. Other methods include coding checks, comparison with other
types of models such as continuum methods, conservation of energy and compari-

son with full-scale industrial plant. This paper makes a direct comparison between
the model and non-cohesive polyhedral dice with no “fitting-parameters”.
Several techniques to model non-spheres are evolving in DEM granular flow as
reviewed by Dziugys and Peters1. These include ellipsoidal particles, sphero-
cylinders, superquadrics and hyperquadrics, polar shape in 2D, skeleton shape,
polygons, composite particles- “glued” spheres or other shapes, or overlapping
spheres modelled as a single particle, and virtual space method in which the field
of interest is described by a large number of cells. Each particle encloses a number
of cells. Contact occurs when one or more cells are assigned to more than a single
particle.
Investigations in the flow of polyhedral particles provide a basis for extending
our work on the flow of „sensor-shapes‟ in a discharging hopper filled with
spheres5.
POLYHEDRAL DEM MODEL

The basic form of this DEM model is fairly standard with a linear spring and
dashpot contact mechanics approach3. Fraige et al4 describes a DEM polygon
model and its application to hopper flow. The polyhedral model is an extension to
the polygon model. An example of the polyhedron DEM model is shown in Fig 1.
It should be noted that the vertices are “rounded”, and hence the edges are corres-
pondingly rounded. This is to facilitate the model and to acknowledge that real
granular particles do not contain perfect vertices.
For each type of polyhedron the user specifies the vertex positions, vertex ra-
dius and. triangulation data. The particle can be imagined as a series of triangles
with cylindrical edges and spherical nodes. Irregular polyhedra can be modelled
but each particle must be convex. A cohesion distance is defined surrounding the
outer surface. The material density is specified and the program calculates numer-
ically the mass, centre of gravity and moment of inertia tensor. Spherical particles
can be included.
For every pair of close particles (tested by intersection of enclosing spheres)
the model tests for intersection of nodes, edges and triangle faces. For a contact
the normal direction is then readily determined. The interaction may comprise a

number of contact points at which a normal force is applied. In such a case the
area of contact is calculated and the cohesive force is modified as a function of
contact area in a similar manner to the 2D model4. More details will be published
later.
Fig 1. Snapshot of polyhedron DEM model of discharging hopper (side walls

only shown for clarity)
HOPPER EXPERIMENTS
Experiments were conducted on a small 3D hopper constructed of perspex us-
ing 322 polyhedral dice. There is negligible cohesion and adhesion.
Fig. 2 LEFT: Photograph of polyhedral particles with curved edges, from top
left to bottom right going across the page: octahedron, pentagonal trapezohedron,
cube, icosahedron and tetrahedron. RIGHT: Photo of 3D hopper.

The principal hopper and particle data are shown below.

Particle No. Volume cm3 Density gcm-3 Side length cm
Pentagonal 90 2.73 1.65
trapezohedra
Octahedra 31 2.57 1.56 1.76
Tetrahedra 29 1.05 2.19 2.08
Icosahedra 81 5.02 1.27 1.32
Cube 91 3.83 1.17 1.56
Friction coefficient 0.28
Coeff of restitution 0.54
Hopper half angle α 30, 60, 90
Silo dimensions cm H = 60, W = 2.5, D = 45
Orifice length B cm 12
Orifice width cm 2.5
Table 1 Principal hopper and particle data
The friction and damping coefficients were obtained by conducting tests
on a cube. The cube was chosen as it was easier to drop on to its face. The friction
coefficient was obtained by placing the cube on one of the side pieces from the
3D hopper and increasing the angle of the incline until it started to slide. The
damping coefficient was obtained by dropping the cube from a height of 50cm
onto a Perspex base and recording the rebound height. The hopper flow experi-
ments were repeated ten times at orifice widths between 6cm and 12cm so that a
critical orifice length at which blocking occurred could be deduced.
RESULTS
Fig 3 shows a comparison of the experimental and DEM results for hopper dis-
charge for B=12cm. The experiments discharge somewhat faster, the 90o case is
similar. Fig 4 shows the proportion of complete experimental discharges. At
smaller B and generally larger α, blockages are more likely to occur, similar to
that shown in Fig 5. Initial DEM simulations for α= 60o indicate that the critical

orifize size for flow is about 10cm. Further simulations are required including
sensitivity studies.
Fig. 3 The fraction of particles discharged against time for α= 30, 60 & 90o
B=12cm LEFT: the first three experiments, RIGHT: DEM
Fig. 4 The number of complete experimental discharges at each hopper half

angle and orifice length
Fig. 5 Example of hopper blockage for α= 60o B=10cm, LEFT: experiment,

RIGHT: DEM pojected 2D graphic.

CONCLUSION
Reasonable agreement between DEM simulations and practical experiments on
polyhedral dice has been shown using DEM data prior to hopper experiments, il-
lustrating that the Discrete Element Method has the potential to predict the flow of
many granular materials used in industry, such as in minerals processing.
ACKNOWLEDGEMENTS
The financial support from the UK‟s EPSRC is gratefully acknowledged.
REFERENCES
1. A. Dziugys, B. Peters, „An approach to simulate the motion of spherical and
non-spherical fuel particles in combustion chambers‟, Granular Matter, 3,
231-265, 2001
2. F.Y. Fraige, P.A. Langston, G.Z. Chen, „Distinct element modelling of cubic
particle packing and flow‟, Powder Technology, 186, 224-240, 2008.
3. B.N. Asmar, P.A. Langston, A.J. Matchett, J.K. Walters, „Validation tests on a
distinct element model of vibrating cohesive particle systems‟, Computers and
Chemical Engineering 26, 785-802, 2002
4. F.Y. Fraige, P.A. Langston, A.J. Matchett, J. Dodds, „Vibration induced flow
in hoppers: DEM 2D polygon model‟, Particuology, 6, 455-466, 2008
5. S. A. Mack, P. A. Langston, C. Webb, T. A. York, „Modelling the trajectory of
3D sensor-shapes in discharging hoppers filled with monosized and binary
mixtures: An experimental and DEM comparison‟, Abstract submitted to
DEM5 The Fifth International Conference on Discrete Element Methods, 25-
26 August 2010, London.

MODELLING THE TRAJECTORY OF 3D SENSOR-

SHAPES IN DISCHARGING HOPPERS FILLED WITH
MONOSIZED AND BINARY MIXTURES: AN
EXPERIMENAL AND DEM COMPARISON
Stuart Mack1, Paul Langston2, Colin Webb1, Trevor York3

1. School of Chemical Engineering and Analytical Science, University of Manchester
2. Process and Environment Division, Faculty of Engineering, University of Nottingham
3. School of Electrical and Electronic Engineering, University of Manchester
Experimental and DEM simulations have been conducted for the flow of
simple 3D individual “sensor-shapes” in a small 3D hopper. Flow of these indi-
vidual particles replicating sensors in monosized spheres and binary mixtures of
spheres having equal density and different density are compared.
INTRODUCTION
Storing granular materials, especially for the food industry, for many weeks or
months is necessary in order to provide sufficient food for an ever growing global
population. The staple foods of many countries include wheat. Wheat grain is
used in the making of bread, biscuits and some breakfast cereals and is stored in
huge silos or in grain mountains. If the wheat is subjected to warm, damp condi-
tions it can start to spoil, which makes it unfit to be used in food production.
It is proposed to place wireless sensor particles in silos or grain mountains to
determine the distribution of temperature and humidity of the wheat grain and
transmit to a central computer.

Fig 1. Sensor particles interacting with each other in a silo filled with wheat.
Experimental and DEM simulations are being conducted to determine what
density and shape these sensor particles should have, so that they move in synch
with the surrounding grain. To simplify the problem, both computationally and
experimentally, spherical particles are being used for the bulk material.
Rotter[1]et al. placed free cells in rectangular silos to determine the pressures
during fill. Chehata[2] et al. have investigated the flow around a fixed cylinder
placed in a rectangular silo. Tüzün and Nedderman[3] observed granular flow
around fixed triangular and square inserts.
DEM MODEL
The basic form of the DEM model uses the linear spring-dashpot contact me-
chanics developed by Langston[4] et al.. To reduce the run time the „en masse‟ fill
method [4] is used. The program has been modified to include the sensor particles
modelled as cylinders or polygon prisms which maintain their major axis orienta-
tion.
HOPPER EXPERIMENTS
Practical experiments are being conducted in a small 3D hopper with three dif-
ferent sizes of particle and three different simple “sensor-shapes” having three dif-
ferent densities (Fig. 2).

Fig. 2 Photographs of: small 3D silo; sensor-shapes, made of metal and perspex
and a hollow set; bulk particles
The “sensor-shapes” are placed at twelve different points in the hopper, so that
comparisons can be made between their trajectories and the trajectory of a single
particle. From the comparisons deductions can be made about how the sensor
density and shape affect the flow of the “sensor-shapes”.
DEM AND EXPERIMENTAL COMPARISON
DEM simulations and practical experiments of single particles rolling and slid-
ing down inclines have been compared to determine particle model parameters.
Comparisons between simulations and experiments to determine the wall stresses
and mass flow rate in the hopper will be published in a future paper. The principal
hopper and particle data is given in Table 1.

Particles Shapes
black white yellow Density gcm-3 Volume cm3
Diameter cm 0.396 0.485 0.587 cuboids
Density gcm -3 1.29 1.29 1.16 metal 7.9 7.3
Particle-Particle perspex 1.2 7.3
sliding friction 0.3249 0.3249 0.3249 triangular prisms
rolling friction 0.0042 0.0022 0.0011 metal 7.9 4.3
Particle-Wall perspex 1.2 4.3
sliding friction 0.3057 0.3057 0.3057 cylinders
rolling friction 0.0415 0.0212 0.0210 metal 7.9 7.5
perspex 1.2 7.5
Particle stiffness Nm -1 1110 1400 1488 Shape-wall sliding friction 0.176
Coeff of restitution 0.85 0.81 0.82 All other friction coefficients as for particles
Angle of repose
between rig sides 26° 27° 28° Hopper
Angle of repose in heap 20° 19° 19° Height, Width, Length cm 100, 2.5, 40
No.of particles N 50000 Hopper half angle α 60°
Table 1 Principal hopper, sensor-shape and particle data used in DEM simula-
tions.
Experimental and DEM comparisons of three starting positions of the shapes
and the metal and perspex shapes are being conducted. Simulations of the flow of
the shapes in discharging hoppers have only at present been conducted with the
white monosized spheres (Fig.3) and a binary mixture of 80% white and 20% yel-
low. Figure 4 shows DEM and experimental comparison for the path followed by
the perspex shape against time where z is the height of the shape from the orifice
and x is the horizontal distance of the shape from the orifice. Figure 5 shows the
variation in the angular velocity with time for the perspex cuboid shape in DEM.
Figure 6 shows the horizontal force component Force X and the vertical force
component Force Z exerted on the perspex cuboid in the DEM simulation of a
discharging hopper.

Fig. 3 DEM simulation and practical experiment showing a square sensor-

shape flowing in a discharging silo filled with monosized particles.
Fig. 4 Comparison between experiment and DEM of trajectory followed by

perspex cuboid
Fig. 5 Angular velocity against time for the square perspex shape in DEM

Fig. 6 The horizontal and vertical forces exerted on the perspex cuboid in the
discharging silo from the DEM simulation.
CONCLUSION
Reasonable agreement in flow trajectory has been shown between the practical
experiment and DEM simulation. DEM has the potential to model the flow of
wireless sensors in discharging hoppers, and model applied forces.
ACKNOWLEDGMENTS
The financial support from the UK‟s EPSRC is gratefully acknowledged.
REFERENCES
Figure 1 has been reproduced from the Wireless Sensor Networks for Industrial
Processes website „www.wsn4ip.org.uk/schematic‟.
1. Rotter,J.M., Brown,C.J., Lahlouh,E.H. „Patterns of wall pressure on fill-
ing a square planform steel silo.’Engineering Structures 24 135-150 2002
2. Chehata,D., Zenit,R., Wassgren,C.R. „Dense granular flow around an
immersed cylinder.‟ Physics of Fluids. Vol. 15, No. 6, 1622-1631, 2003
3. Tüzün,U. and Nedderman,R.M. „Gravity flow of granular materials
round obstacles: I – Investigation on the effects of inserts on flow pat-
terns in a silo.‟ Chem. Eng. Sci. Vol. 40, Issue 3, 325-336, 1985
4. Langston,P., Tüzün,U., Heyes,D.M. „Discrete element simulation of
granular flow in 2D and 3D hoppers: Dependence of discharge rate and
wall stress on particle interactions.‟ Chem. Eng. Sci., Vol. 50, No. 6, 967-
987 1995

COMPARING EXPERIMENTAL MEASUREMENTS

OF MILL LIFTER DEFLECTIONS WITH 2D AND 3D
DEM PREDICTIONS
Johanna Alatalo1, Bertil I. Pålsson1, Kent Tano2

1
Division of Mineral Processing, Luleå University of Technology, SE-97187 Luleå, Sweden
2
Technology and Business Development, LKAB, SE-98381 Malmberget Sweden
The grinding process has high power consumption and as a

result, it is also costly. It is a difficult process to monitor since
there are many variables that affect the end product.
Information of the events occurring in the charge can be
achieved by measuring the lifter deflection and with simulation
it is possible to illustrate what take place in the charge.
However, for simulations to be reliable it demands that they are
verified against process data. At LKAB R&D facility there is a
pilot mill with a lifter deflection measuring system mounted
inside one of the rubber lifters. A series of experiments with a
steel media charge were performed and this provides an
opportunity to validate simulation results. The measured lifter
deflection signal is used to compare with signals from 2 and 3
dimensional DEM simulations of the pilot mill. The resulting
signals show that the 3-dimensional case displays a better
profile and the difference of toe and shoulder angles are less
than in the 2-dimensional case. This means that the simulations
are more reliable when they are performed in three dimensions
and that the DEM simulations can be used to increased the
knowledge of the events inside a mill and its charge.

INTRODUCTION
Milling is the process of reducing the particle size distribution of the extracted
ore. It demands much energy and it is a very expensive part of the process.
Therefore, it needs to be optimised as much as possible. However, there is a lack
of knowledge of what happens inside the charge of a tumbling mill. In principle it
is impossible to use direct measurements in the charge since the environment is
too harsh, but it is possible to use indirect methods of measuring, for example,
deflection of lifters, variation in energy consumption or surface vibration on the
mill.
The total process of grinding is relatively difficult to monitor, since there are a
large number of variables to consider, for example the flow rates, percentage of
solids or filling, and of course, the ore itself. These variables will strongly affect
the end product. This also means that there is a possibility to improve the process
and to adapt it to the feed product stream.
To increase the understanding of the grinding process, DEM simulation might

be very helpful to illustrate the events occurring in the charge. Note, the
simulations can only be considered reliable, if they are validated with real process
data.
EXPERIMENTAL
LKAB mines iron ore in the north of Sweden and has at Malmberget a R&D-
facility from where process data were obtained.
The pilot mill measures, 1.414 m in diameter and is 1.22 m in length. It has 12
rubber lifters and a strain gauge detector is embedded in one of them. It is a grate-
discharge mill.

Fig 1. Inside the pilot mill, here seen with a pebble charge
In a series of experiments with a steel media charge, consisting of balls with a

radius between 5 – 15 mm, the pilot mill filling was varied between 25 and 35 %.
Its rotational speed was either 73 or 78 % of critical speed. To get stable grinding
conditions, the chosen feed material was a hematite pellet feed, with a d50 of
35 µm and 5200 kg/m3, and feed rate was kept constant at 1.5 ton/h.
The strain gauge detector, a Continuous Charge Measuring system (CCM)

comes from Metso Minerals. The detector measures the deflection of the lifter as
it passes through the charge and from the signal, the toe and shoulder angles as
well as degree of filling can be calculated.
SIMULATION
The 2-dimensional simulations were performed in Pfc2D at LKAB [1] and
show a difference between the simulated and measured toe and shoulder angles.
One explanation is the fact that the simulation were done in 2D [2]. Noticeable is
that the lifter deflection was simulated directly in the simulation, using a
“bendable” lifter.
Here, the 3-dimensional simulations are performed in EDEM. A 10 cm slice

of the pilot mill is simulated. To get a more reliable simulation, the sides of the
mill are set as periodic boundaries. The deflection is calculated from the

experienced force on the lifter. The same material data and parameters have been
used for both simulations.
RESULTS
The resulting deflections from the experiments and the simulations can be
seen in figure 2. In this case, the mill filling is 35 % by volume and the rotational
speed is 73 % of critical speed. The signals from the simulations are low pass
filtered.
0,8
Normalised deflection
0,6
0,4
0,2
0
0 30 60 90 120 150 180 210 240 270 300 330 360
Rotation angle (degrees)
Fig 2. Graph of the lifter deflection signals for the measured (blue), 2D
simulation (green) and 3D simulation (red) signals.
The difference of degrees in that occurred in the 2-dimensional events is not

as large for the 3-dimensional event, which means that the 3D simulations are
more reliable. The simulated signal in the 3D case shows a better profile than the
2D case.

CONCLUSIONS
The simulation in 3D gives a better representative value and can be used to
increase the knowledge of the events inside a mill and its charge. The absence of
slurry in the simulation is of course a limitation.
FUTURE WORKS
A new series of experiments were performed at the LKAB R&D facilities at
Malmberget in March 2010. At this time, the charge consisted of pebbles and the
rotational speed, percentage of filling and volume %solids were varied. Also,
additional experiments with varying types of pebbles were done. At this stage, the
result from this is being analysed and new simulation will be made.
ACKNOWLEDGEMENTS
Thanks to LKAB for financial support and for making the experiments
possible.
REFERENCES
1. K. Tano, Continuous Monitoring of Mineral Processes with Special Focus on
Tumbling Mills – A Multivariate Approach, Luleå University of Technology,
2005.
2. P. W. Cleary, R. Morrison and S. Morrell, ‘Comparison of DEM and
experiment for a scale model SAG mill’, International Journal of Mineral
Processing, 68, 129-165, 2003.

NUMERICAL ANALYSIS OF COLD-FORMED

SIGMA STEEL BEAMS
Qiang Liu, Jian Yang and Long Yuan Li

University of Birmingham, School of Civil Engineering
A nonlinear finite element (FE) based model was developed to simu-

late the structural behaviour of simply supported cold-formed steel
(CFS) Sigma beams under mid-span point loading. The FE model
shows good agreement with the test result in representing the actual
moment-rotation responses of the Sigma sections at internal supports.
It can be therefore explored to conduct quasi-plastic design for con-
tinuous CFS Sigma beams at internal supports.
INTRODUCTION
A nonlinear finite element (FE) model has been developed by using general
purpose software ANSYS [1] to study the structural behaviour of cold-formed
steel (CFS) continuous Sigma beams at internal supports, Numerical results were
then validated by test data.
In the test, the beam was subjected to the central point load to represent the
reaction from the internal support. The nominal dimensions of the tested speci-
mens are summarized in Table 1.
ELEMENT TYPE, MESHING AND MATERIAL PROPERTIES
A shell element, i.e. SHELL181, was chosen to model the Sigma beam.
Bolts that used to connect the Sigma section and supporting cleats were modelled

with elastic shell element, i.e. SHELL63. Since mesh pattern can directly influ-
ence the convergence and the accuracy of FE results, the shape and size of the
element in the present FE modelling were determined through extensive trial and
error process. The best result was achieved through the following discretization
approach, i.e. 2 elements in the lip, 8 elements in the flange, 8 elements in the
outer-web, 10 elements in the inner-web and 1 element in the corner. All elements
have quadrilateral shape. As demonstrated in Fig. 1(a), the aspect ratio of the ele-
ments near the middle span and two ends, where high stress gradient was present,
was kept between 1 and 2. The aspect ratio of the remaining elements was mostly
chosen between 1 and 3. Areas around bolt holes in the outer-web were discre-
tised into regular quadrilateral elements [see Fig. 1(b) and (c)] by using the quad-
morphing mesher.
Section D (mm) F (mm) L (mm) O (mm) I (mm) S (mm) r (mm) t (mm)
20012 200 62.5 20 45 110 16 4 1.2
20016 200 62.5 20 45 110 16 4 1.6
20025 200 62.5 20 45 110 16 4 2.5

Table 1. Nominal cross sectional dimensions
Material property of the beam was modelled by using a multilinear stress-strain

relationship with Von-Mises yield criteria. The nonlinear material model was de-
fined by the true stress and the strain, which were derived from tension coupon
tests. The Poisson’s ratio υ of the elements was assumed as 0.3. The elastic mate-
rials for bolts and cleats were modelled by using E = 210000MPa and υ = 0.3.
LOADING AND BOUNDARY CONDITIONS
Loads were directly applied through four bolt holes near the middle span.
Each hole equally shares the total load. The circular edge of each hole was prede-
fined by 4 arcs in the geometry modelling process. The bolt hole was assumed to
receive loading in the bottom quarter of the circular edge, which can reflect prac-
tical contact details between the loading bolt and the bolt hole. Four semi-rigid
regions were created on loading edges to avoid stress concentration. All cleats

were rationally treated as rigid bodies due to their significantly higher stiffness in
comparison with that of the cold-formed sections. Full lateral restraints therefore
were imposed to the Sigma beam at the middle span and both ends, respectively.
In the mid-span plane of the Sigma beam, the longitudinal movement of two
nodes (in the upper and the lower out-webs respectively) were restrained to glob-
ally immobilize the member.
(a) Overall view (b) Meshing near loading points (c) Meshing near supports
Fig 1. Finite element meshing
The bolt connecting the beam to supporting cleats was specifically simu-
lated by using a 10mm thick wedge shaped block. The central node on the block
was transversely restrained. The nodes on the top quarter of the bolt hole and the
exterior circular surface of the block were coupled together. The main benefit of
doing this is that it introduces no additional rotational restraint about the major
bending axis and significantly reduces both stress concentration and local distor-
tion at the bolt-beam interface.
INITIAL IMPERFECTIONS
The introduction of geometric imperfections into the FE modelling was
based on the published studies [2-4]. Three basic imperfection shapes determined
from elastic buckling analysis are presented in Fig.2, i.e. the distortional buckling,
local buckling and lateral torsional buckling mode shape. The initial imperfections
in the FE modelling were included by the linear superposition of individual linear
buckling modes. Since the first buckling mode from elastic buckling analysis is
normally critical on structural behaviours, the corresponding buckling mode was
selected as one of the imperfection components for each specimen. Other modes

were included only if the ratio of respective buckling load to the critical load was
within 2. For local mode and distortional mode, the amplitudes of imperfection
were specified as 0.5t and t (t is the cross-sectional thickness), respectively, which
represent approximately 50% of the probability of exceedance in the sampling
data recorded from real samples [2]. The amplitude of the lateral torsional imper-
fection mode is suggested in EC3 [5], i.e. the unrestrained span length over 1000.
However from trial analysis, it was found that this value resulted in overly con-
servative predictions. Hence in this FE modelling, the amplitude of the lateral tor-
sional imperfection mode was assumed to be the unrestrained span length over
1500.
(a). Distortional buckling mode: antisymmetric and symmetric modes
(b). Local buckling mode (c). Lateral torsional buckling shape

Fig 2. Typical imperfection modes
COMPARISON WITH EXPERIMENTAL RESULTS
The full loading range in the FE modelling was obtained by using the arc-
length method. Moment-rotation responses obtained from the FE simulations
aaaare presented in Fig. 3, together with the test results. In Fig.3, the test speci-
mens have been assigned a 3-part ID symbol, e.g. 24-20012-C indicates the beam
span (2.4m), section (200mm deep) and thickness (1.2mm), type of connection (C
indicates continuous connection at internal supports).
Fig.3 shows that the developed FE model produces excellent predictions in
linear flexural stiffness and ultimate moment of the tested beams. It is worth not-
ing that in the FE model, the boundary condition at both ends was effectively
modelled to be simply supported. The good agreement between the predicted and
tested results therefore indicates that the partial rigidity provided by two bolts at
supports is negligible in practice.

25.0
Bending moment (kNm) 20.0 24-20012-C FEM

24-20012-C test
24-20016-C FEM
15.0
24-20016-C test
24-20025-C FEM
10.0 24-20025-C test
40-20025-C FEM
40-20025-C test
5.0
0.0
0 1 2 3 4
Rotation(Deg)
Fig 3. Moment-rotation curves

The FE model also provides good predictions in modelling the post-failure behav-
iours of the beams. As shown in Fig. 4, the post-failure patterns of 24-20025-c
and 40-20025-c obtained from the tests and the predictions are almost coincident
except that the 40-20025-c test specimen experienced lateral torsional buckling in
the right half span but in the FE simulation the lateral tortional buckling occurred
in the left half span. Due to the nature of symmetry, the failure could occur ran-
domly in either span.
CONCLUSIONS
The structural behaviour of a simply supported Sigma beam under mid-span
point load was simulated numerically by using FEM. This is to study the pre-and
post-failure moment-rotation behaviour and localized plastic collapse of a con-
tinuous beam at the internal support. The selected element type, meshing pattern,
material properties, implementation of the boundary and loading conditions and
the introduction of the initial imperfections employed in the modelling process
have been described in details. Validated by the experimental results, the numeri-
cal modelling has been demonstrated a good capability in reproduction of the
moment-rotation responses of the Sigma sections under mid-span point load. The
developed FE model therefore can be utilized to study the structural behaviour of
continuous beams at internal support, such as the moment-rotation relationship,

the flexural stiffness of beams, the moment capacity under point load, post-failure
response and the localized plastic collapse mechanism.
(a) Experimental result and FE simulation result (24-20025-c)
(b) Experimental result and FE simulation result (40-20025-c)

Fig 4. Post-failure deformations
REFERENCES
[1] Commands Reference, Elements Reference, Operations Guide, Basic Analysis
Guide, Theory Reference for ANSYS.
[2] Schafer, B.W., Peköz, T., “Computational modelling of cold-formed steel:
Characterizing geometric imperfections and residual stresses”, Journal of Con-
structional Steel Research, 1998, 47(3), pp.193–210.
[3] Dubina D., Ungureanu V., “Effect of Imperfections on Numerical Simulation
of Instability Behaviour of Cold-Formed Steel Members”, Thin-Walled Struc-
tures, 2002, 40(3), pp.239-62.
[4] Sarawit, A. T., Kim, Y., Bakker, M. C. M., and Peköz, T., “The finite element
method for thin-walled members—applications”, Thin-Walled Structures, 2003,
41(2), pp.191–206.
[5] European Committee for Standardisation., “Eurocode 3: Design of steel struc-
tures: Part 1.3: General rules – Supplementary rules for cold-formed thin gauge
members and sheeting’s”, 2006.

PARALLELISATION, GPU, OPTIMISATION

PARALLELIZATION OF AN OPEN-SOURCE
FEM/DEM CODE Y2D
Tomas Lukas, Antonio Munjiza

Department of Engineering, Queen Mary, University of London
This paper presents static domain decomposition based parallelization

strategy of an open-source combined finite-discrete element code
Y2D. MPI is adopted as the message passing library. An implementa-
tion of parallel FEM/DEM code on the 500 node in-house parallel
cluster is summarized and demonstrated using benchmark examples.
INTRODUCTION
The combined finite-discrete element method merges finite element tools and
techniques with discrete element algorithms. Finite element-based analysis of con-
tinua is merged with discrete element-based transient dynamics, contact detection
and contact interaction solutions. Thus, transient dynamic analysis of systems
comprising a large number (from a few thousands to more than a million) of de-
formable bodies which interact with each other and in this process can break, frac-
ture or fragment, becomes possible. The computational requirements for simula-
tions of systems comprising a large number of deformable bodies (millions of par-
ticles) are enormous and the computational power that can be achieved with a sin-
gle-processor system is limited by physical and architectural bounds. Thus the use
of high-performance parallel computers is indispensable.
In this work a geometric approach to domain decomposition has been adopted.
It divides the computational domain by exploiting the location of objects in the

simulation. To demonstrate parallelization strategy used, computational domain is

discretised into a static rectangular grid of sub-domains of the same size. Each
sub-domain is assigned to a single processor in the cluster.
Existing single processor FEM/DEM program is written in C (Y2D code) and
from wide range of libraries for parallel computing message-passing interface
(MPI) has been chosen as the parallelization library.
PARALLELIZATION STRATEGY
Status of elements. Elements can be divided into several categories depending
on their location within sub-domain. Constant strain triangular elements located
inside sub-domain (Fig 1a) are marked as internal (A). Elements located at the
border between two sub-domains are shared by two processors, see Fig 1a.
a) A – internal
B – original
C – copy A
B C
1st processor 2nd processor
b) 3rd processor 4th processor
E E D – original
E – copy
D E
1st processor 2nd processor
Fig 1. Status of constant strain triangular elements located a) at the border be-
tween two sub-domains, b) at borders between four sub-domains.
If two nodes of element are located inside sub-domain, then the element is
marked as an original (B) on corresponding processor while on second processor

is marked as a copy (C). At last, elements located at the borders between four sub-
domains are shared by four processors and marked as originals (D) at one proces-
sor (sub-domain which is an intersection of two planes each containing two nodes
of element) and copies (E) on remaining three processors, see Fig 1b. The reason
for keeping originals and their copies is a possibility of their contact with internal
elements located on two or four different processors.
Transition from continua to discontinua in the combined finite-discrete element
method occurs through fracture which is introduced through discrete crack model
[1] represented in Y2D code as a four-node joint element holding together triangle
elements. Status of joint element can be determined analogically to triangle ele-
ments by its location within sub-domain with addition of using also the centre of
joint in case two nodes of element are inside sub-domain and two nodes outside.
Nodal and contact forces. Each time step during time integration nodal forces
are calculated and later forces arising from contact between discrete elements are
added to nodal forces. Nodal forces for both triangle and joint elements are calcu-
lated only for internal elements and originals. Since elements located around bor-
ders are shared by two or four different processors, contact forces between origi-
nals and copies must be divided by two or four with exception of contact when
one element is located at horizontal and second element at vertical border. In this
case the contact force is unique because corresponding counterparts of these two
elements are located on two different processors and therefore this contact cannot
occur anywhere else. Contact force is divided by two if first element is located on
two and second element on four processors. Contact force of internal element in
contact with any other stays always intact, i.e. is “divided” by 1.0.
After calculation of nodal and contact forces, total forces of elements located
around borders between sub-domains are exchanged between corresponding proc-
essors.
This is performed in two stages. In the first stage messages containing forces
for elements located at right and left borders are exchanged in horizontal direc-
tion, see Fig 2a. In the second stage forces for elements located at top and bottom
borders are exchanged in vertical direction as shown in Fig 2b. There is no com-
munication in diagonal direction. Forces received in each stage are added to their
corresponding forces on receiving processor. Forces of elements shared by four

processors are exchanged in both horizontal and vertical direction resulting in a

sum of forces from all four processors, see Fig 2c.
a) b) c) 3 3
6 7 8 6 7 8 ∑F
i =0
i ∑F i
i =0
2 3
3 4 5 3 4 5 3 3
∑ Fi
i =0
∑F i
i =0
0 1 2 0 1 2 0 1
Fig 2. Send-receive operation in a) horizontal direction, b) vertical direction; c)

forces of elements shared by four processors after communication.
Moving of elements. After force exchange the moving equation is solved re-
sulting in new positions of elements. It is therefore necessary to update status of
all originals and copies and internal elements close to each border. A singly con-
nected lists of internal elements located in the proximity of each border are as-
sembled during contact detection. These lists are used for checking new positions
of internal elements. Originals and copies saved in the element database are sepa-
rated from internal elements and it is therefore possible to check their positions
directly. Elements which left sub-domain must be deleted from database and new
elements which moved into sub-domain from neighbouring processors are re-
ceived.
Communication. For now parallel code is using MPI-1, which means that in-
put and output is performed only by I/O processor. The rest of the processors must
send their output to I/O processor. The content of each message is converted into
string of characters before sending and converted back upon receiving to ensure
portability and simplicity.
Numerical examples. Parallel code was tested on a CPU with quad-core Intel
Xeon W3570 and 12 GB of RAM. Testing was done on a box filled by different
numbers of discrete elements with initial velocity 200 m/s in random directions.
Dimensions of the box were scaled with increasing number of elements (size of
elements stays the same). All other input parameters stayed the same. Due to ac-

cumulation of rounding errors a small difference between results calculated by

sequential and parallel code occurs. Results calculated for 1024 and 10000 dis-
crete elements are displayed in Fig 3.
Fig 3. A box filled by 1024 (top) and 10 000 (middle) discrete elements calcu-
lated by a single processor code at time steps 0 and 3000. Zoom box (bottom) for
10 000 discrete elements.
CPU time for different numbers of elements

4000
3500
3000 1 processor
Time [s]
2500 2 processors
2000 1 processor
1500
2 processors
1000
500
0
0 5000 10000 15000 20000
Number of elements [-]
Fig 4. CPU times for different numbers of elements.

Recorded simulation times for six different numbers of discrete elements are
plotted in Fig 4. With increasing number of elements simulation time for a single
processor code is not growing linearly which is also a cause for a speed-up bigger
than 2 for higher numbers of elements.
CONCLUSION
Speed-up values for different numbers of discrete elements are distorted by the
fact that simulation time is not growing linearly with increasing number of ele-
ments. Further improvement of the parallel code will be achieved by introducing a
buffer around borders between sub-domains [4] which means that status of ele-
ments doesn’t have to be checked each time step which also reduces number of
communications and the need for new contact detection in case new elements are
received or old elements deleted.
REFERENCES
1. A. Munjiza, The combined finite-discrete element method, Wiley, 2004.
2. P.S. Pacheco, Parallel programming with MPI, Morgan Kaufmann Publishers,
1997.
3. D.R.J. Owen, Y.T. Feng, ‘Parallelised finite/discrete element simulation of
multi-fracturing solids and discrete systems’, Engineering Computations, 18,
No.3/4, 557-76, 2001.
4. F. Wang, Y.T. Feng, D.R.J. Owen, ‘Parallelisation for finite-discrete element
analysis in a distributed-memory environment’, Int. J. Comp. Eng. Sci., 5,
No.1, 1-23, 2004.

WET-MIXING OF POWDERS, A LARGE-SCALE

GPU IMPLEMENTATION
Radeke, A.C., Khinast, J.G.

Research Center Pharmaceutical Engineering GmbH, A-8010 Graz, Austria
The operation which is addressed by our DEM simulations is wet-

mixing. Here a powder is filled into a mixing device. Afterward fluid
is continuously sprayed from a nozzle onto the powder whilst mixing.
We use porous catalyst particles. These particles absorbing fluid over
the process time, from the spray zone on top of the bed as well as from
contacting particles having a fluid film caused by over-saturation.
With increasing time the powder is reaching a saturated state. The in-
fluence of the spray zone location on mixing homogeneity is studied
by DEM simulations.
INTRODUCTION
Processes in pharmaceutical applications are often complex. For example
powders are fluidized, sprayed and dried in one operation. There particles moving
in hot air whilst a liquid is sprayed which forms layers on the particle surface. In
other processes powders and bio-active ingredients have to be mixed until a ho-
mogeneous blend is achieved. Often the desired functionality of pharmaceutical
products restricts the choice of engineering processes. One cannot expose heat
sensitive ingredients in a hot air fluidized bed.
Usually the grain size of powders in the pharmaceutical industry is around 50
microns. This means to have billions of particles in only one Liter of powder. In
order to make an 1 : 1 validation of our simulation in the future we have chosen

fine glass beads of 360µm in average size. This material is available for experi-
ments and fits tightly our computational limitations in terms of the problem size.
THE RPD CODE

In order to get closer in terms of the powder resolution we switched to a new
technology. Specific demands of a research project at Rutgers University led to
the development of a new enhanced DEM code using the (at this time) newly
available CUDA technology [6, 7, 11]. The program development of the RPD
(Rutgers Particle Dynamics) code started in April 2008. The code is still under de-
velopment and is intended as a feasibility study.
At this time with RPD one can simulate millions of particles on consumer
hardware. The simulation can be directly visualized by a GUI with key controlling
of display features. Because of the computational strength, the preparation of
large simulations is very convenient. One first simply incorporate hundred thou-
sand particles for testing purposes of the desired physical process. This yields
particle flow simulations at interactive speed. Thereafter RPD can be executed
with the maximum possible amount of particles. This maximum number of
particles depends on the memory size of the equipped graphics card. The simula-
tions presented in this paper are executed on an NVIDIA GTX 285 2GB graphics
board. At the present development stage of RPD it is possible to simulate about
two millions of particles per Giga Byte of graphics memory.
IMPLEMENTATION
Our numerical model is a 'soft' particle approach where spherical particles ex-
ert forces due to their mass m and geometrical overlap δ at the contact points c.
The total force F and torque M acting on a particle
F =∑ f c mg (1)
c
M =∑ r× f c (2)
c
is the summation of contact forces fc with other particles or boundaries and gravit-

ational force. Considering Newtons second law, at time t we obtain an accelera-

tion a(t) which can be used for an explicit time step integration with a time step dt
in order to obtain new velocities v(t + dt) and positions x(t + dt) for each particle
in the system. Here, the classical Verlet integration scheme is employed for posi-
tions [12]
a t =F /m (3)
x tdt=2xt −x t−dtat  dt 2 (4)

and velocities obtained by central dierence approximation
1
v tdt=  x tdt −x t−dta t dt. (5)
2dt
The implementation accounts for translational as well as for rotational degrees
of freedom. Both are handled separately by the application of the Verlet algorithm
(3 - 6) and an Euler integration (7) in the latter case. Total torque M and moment
of inertia I are used to obtain the angular acceleration ̇t
M
̇t= (6)
I
and the new angular velocity at time t + dt
tdt =t ̇t dt. (7)

In order to obtain the new particle orientation q(t + dt) the integration of the angu-
lar velocity is performed afterward. Here, q(t + dt) is a quaternion representation
of particle orientation, details can be found in [9]. The implemented force laws are
linear in normal and tangential direction and follows the implementation sugges-
ted in [5].
f n=k n  n− Dn ˙n− f ch (8)

⃰
f t =min k t t −D t ̇t , ∣Fn∣ (9)
where µ is the static Coulomb friction coecient, kn, kt are spring constants and Dn,
Dt account for viscous damping in normal and tangential direction respectively.
The cohesive force fch is only active in the case of liquid bridges. The tangential
overlap t⃰ =t t−t t−dt  allows for restoring forces if the contact had been
existed already in the previous time step. Analogous to sliding friction rolling res-
istance is implemented.
PARALLELIZATION
The RPD code uses the common cell space logic [1, 10]. The cell space
implementation is flexible, so particles with moderate size ratio up to 1:5 or arbit-
rary size distribution within that limit can be simulated.
In contrast to MPI or OpenMP based approaches where large subsets of cells
are distributed to (CPU) processor cores, the parallelization is implemented on
particle level using the 'Compute Unified Device Architecture' (CUDA) program-
ming technique. So all equations (1-9) are executed by each single particle simul-
taneously.
NUMERICAL PRECISION
GPUs originating from computer graphics and became more an more powerful
to satisfy the demands of 3D games. There the performance is expressed in terms
of frame rates and smooth interactive handling whilst such a game is running. In
order to display the result (the rendered frames) fast, the accuracy of the rendered
images are sufficient by processing the data with single precision (SP) numeric on
SP hardware. Using this hardware for numerical simulations leads to problems in
the implementation of the algorithms caused by the lower amount of valid digits
in the floating point representation [4]. As an intermediate solution we scale our
geometry according to Froude theory [8]. Though, the Verlet algorithm is used
widely in real molecular dynamics software packages which run at SP with com-
parable accuracy to DP. There different length scales and time scales apply (in the
nano meter and femto seconds) which exudes to resolve this problem by unit scal-
ing or better numerical techniques [2].

SPRAY EXTENSION
Two reservoirs are assigned to each particle: A film reservoir and an absorp-
tion or wetting reservoir. The extension for spraying is divided into three parts.
First, a nozzle is modeled as a point source of a bunch of rays which are run-
ning through the powder bed. The intersected particles are detected and the
particles nearest to the nozzle are forming the top layer. By defining a nozzle
flow-rate the particles in this top layer receive evenly distributed fluid and the
film reservoir is filled when the particles are hit. In order to keep the spray pattern
stationary onto the bed surface, this spray zone projection has to be executed fre-
quently (about 100x / s). The CUDA implementation outperforms the CPU ver-
sion with a factor of 500.
Second, the particles can absorb fluid from the film reservoir.
Third, in case of over-saturation the fluid content of the film reservoir can be
transfered by a liquid bridge model. So whilst a wet contact is lasting, the film
fluid is equilibrated between contacting particles.
A detailed description of the algorithms is beyond the scope of this paper and
will be given in a separate paper.
APPLICATION EXAMPLE
The device which was used for our simulations is a bladed mixer as sketched in
Figure 1. As a convective mixer it consists of a stationary vessel and a rotating
impeller. These types of mixers are used for materials that tend to segregate or ag-
glomerate, and hence, cannot be mixed in tumbling blenders. Bladed mixers have
been extensively investigated both numerically and experimentally by Stewart et
al. [3, 13].
The inner diameter of this lab-size mixer is 2R = 0.1m, the working area of the
impeller covers a volume of about 0.2L (see Table 1)for details.

Figure 1: Schematic of the pitched blade mixer geometry (left) and four million
particles filled into the blade mixer. Spray zone green colored on top of the
powder bed and only already wet particles in opaque visualization colored in
green and grey (right).
Two simulations starts from an identical spatial distribution of four million
particles inside the blade-mixer. All boundary conditions kept constant except the
position of the spray nozzle. We perform the mixing of a moderate speed of 19
rpm. Our simulations also include cohesion in the case of film liquid on the
particle surface on both, particle- and wall contacts. The cohesion is modeled as a
simple superimposition of sevenfold weight force of the particle. The application
of the spray algorithms is demonstrated in Fig (1). After ten revolutions our stat-
istical analysis leads to the recommendation to prefer the central spray position in-
stead of the outside location of the spray zone Fig. (2).

Figure 2: Simulation after ten revolutions and continuously fluid spraying from
a fixed nozzle position: central (left image) and outside (right image). Particles
colored according to its fluid content (increasing blue) and green if the wear a flu-
id film which can be observed in the spray zone area.
Figure 3: Mixing analysis of the two spray locations (central = black, outside =
red). Mixing and saturation is better and higher resp. for the case of the central ra-
dial location. (dashed lines).

With increasing radius the circumferencial speed is increasing too. Therefore

and because of the known effect of diffusive mixing on the blade tips, one would
expect that if fluid is sprayed onto this region, this would also lead to better wet-
mixing. Our simulation results predicts the opposite. The throughput of particles
is obviously higher for the central nozzle position. Residence time distribution
data (not shown in this paper) conforms this, although the circumferential speed in
that region is much slower than on the blade tip.
GEOMETRY AND SIMULATION PARAMETERS

Parameter S. Value Unit
Blade width BW 25 mm
Blade diameter BD 95 mm
Blade angle BA 45 deg
Hub diameter HD 22 mm
Hub height HH 20 mm
Shaft diameter SD 12 mm
Stator diameter DS 100 mm
Table 1. Mixer: dimensions of the simulated mixing device.
Parameter S. Value Unit

Density ρ 2500 kg/m3
Normal stiffness kn 1.05105 Nm-1
Tangential kt 8.25104 Nm-1
stiffness
Cohesion fch 7˟weight N

Restitution e 0.78
coefficient
static friction µ 0.3 mm
(PP)
static friction µ 0.2 mm
(PW)
Table 2. DEM simulation parameters.
CONCLUSION
A massively-parallel approach for the implementation of DEM using CUDA
was presented. The take-home message is: CUDA and GPU allows for significant
increase of computational speed. Intelligent programming technique with respect
to the GPU memory enables for large-scale simulations on consumer graphics
hardware.
REFERENCES
1. M. P. Allen and D. J. Tildesley. Computer Simulation of Liquids. Oxford Uni-
versity Press, Oxford, 1987.
2. Kevin J. Bowers, Edmond Chow, Huafeng Xu, Ron O. Dror, Michael P. East-
wood, Brent A. Gregersen, John L. Klepeis, Istvan Kolossvary, Mark A. Mor-
aes, Federico D. Sacerdoti, John K. Salmon, Yibing Shan, and David E. Shaw.
Scalable algorithms for molecular dynamics simulations on commodity
clusters. In SC '06: Proceedings of the 2006 ACM/IEEE conference on Super-
computing, page 84, New York, NY, USA, 2006. ACM.
3. G.R. Chandratilleke, A.B. Yu, R.L. Stewart, and J. Bridgwater. Effects of
blade rake angle and gap on particle mixing in a cylindrical mixer. Powder
Technology, 193(3):303-11, 10 August 2009.
4. D GOLDBERG. What every computer scientist should know about floating-

point arithmetic. COMPUTING SURVEYS, 23(1):5-48, MAR 1991.

5. S. Luding. Collisions & contacts between two particles. In H. J. Herrmann, J.-
P. Hovi, and S. Luding, editors, Physics of dry granular media – NATO ASI
Series E350, page 285, Dordrecht, 1998. Kluwer Academic Publishers.
6. NVIDIA. "cuda compute unified device architecture, programming guide ver-
sion 1.0". 2008.
7. NVIDIA. "on accelerating financial applications using cuda gpu technology,
hilton new york, u.s.". The HP and NVIDIA Seminar at Security Industry
Technology Show (SIFMA), Jun 2008.

GRID PARALLEL COMPUTING SYSTEM FOR

CALIBRATION OF DEM MATERIAL MODELS
Elson Mourao, David Curry, Richard LaRoche

DEM Solutions Ltd
DEM material model calibration is a method to adjust DEM model pa-

rameters to find the best match between the results of simulations and
physical bulk particle tests. Many hundreds of simulations may be re-
quired in order to determine the optimum parameter values for the
DEM material model. This paper presents a Grid Parallel approach to
implementing the calibration process involving carrying out a series of
different simulations in parallel on a grid of compute nodes, and its
application in bulk material characterization for the purposes of DEM
simulation.
INTRODUCTION
DEM material model calibration is a method where DEM model parameters
which characterize the bulk particles; i.e. the parameters in the particle contact
model and the size and shape distribution of the particles, are adjusted to obtain
the best match between the results of simulations using the material model and
physical bulk particle tests. The objective is to match the bulk behavior in the vir-
tual test with that observed in the physical bench test within an acceptable level of
accuracy.
Current methods for determining suitable material model parameters are based
on physical tests sometimes also using comparison of process-scale simulation

results against observed behavior, to tune the model. These methods typically
generate thousands of virtual tests which can take several days to run even if each
virtual test by itself takes a small amount of time. A system is required that mini-
mizes the time needed to run the virtual experiments. Here we outline a grid paral-
lel DEM material model calibration system which allows such experiments to run
concurrently and in a highly scalable way such that the runtime is linearly im-
proved the more computation power is added to the system.
GRID PARALLEL DEM CALIBRATION SYSTEM

Model calibration is an exhaustive process; depending on the model and the
optimization technique many hundreds of EDEM simulations may need to be run
for a single model calibration. By running these simulations in parallel, the time to
perform an optimization could be greatly reduced. If enough computational power
is available, the whole optimization could be executed in approximately the same
time it takes to run a single test simulation. We have developed an optimization
process that employs a grid parallel computer network to run simulations in paral-
lel.
The grid parallel system is a computational grid where a number of computers
are used to process each simulation in parallel while one controls the optimization
process. As the processing nodes finish processing a simulation, they send the re-
sult to the optimization engine to determine if the optimal result is found. If the
optimal result is not found a new set of parameters is sent to the processing node
to re-run the simulation with. The process repeats until the optimal result is
achieved.

Fig. 1 Grid Parallel System
Grid Parallel Speed-Up

Consider a simple angle of repose test to analyze the effect of inter-particle
static and rolling friction on the angle of repose of a pile of particles (Fig. 2). At a
resolution of 100-by-100 parameter values, 10,000 different simulations would be
required to fully define the parameter space. If each simulation executed on a sin-
gle computer takes 1 hour to complete, then the full set would take 416 days to
complete. When executed on a small Grid Parallel System containing only 8
nodes it will complete this analysis in 52 days.

Fig. 2 Angle of Repose Test.
APPLICATIONS OF THE GRID SYSTEM

This approach is well suited to calibration methods such as design of experi-
ment (DOE) and parameter optimization, both of which can require large numbers
of individual simulations to be performed. DOE’s are particularly straightforward
as the optimization engine simply generates a predefined list of parameters which
are then distributed to the nodes in turn until the whole set is processed. Optimiza-
tion is more complex approach and different optimization methods have been im-
plemented in this system. This includes techniques such as Local Gradient Search,
Exhaustive Search, Simulated Annealing and Genetic Algorithms [1].
Example application of Exhaustive search.
Previously the Angle of repose example was used to demonstrate the number
of simulations required to determine an optimum parameter combination for the

DEM Material Model. The following example, of calculating Torque in a Shear

Mixer is a further refinement of this method and is based on a series of progres-
sively higher resolution exhaustive searches. Fig. 4 shows the mapped parameter
space for such a method. This is a 4-level optimization which has reduced the
number of total simulations from the 425 that would be required for a single high-
resolution exhaustive analysis, to only 100. This yielded a result accurate to
99.5% of the experimental result in just 7 hours on a small Grid System.
Fig. 3 Example Application of Exhaustive Search
More advanced optimization methods such as simulated annealing can further

reduce the number of simulations required to achieve convergence on the opti-
mum DEM Material Model parameter set [1].
CONCLUSION
DEM material model calibration is required to support the use of DEM simula-
tions in the engineering design process. Calibration usually involves carrying out

a large number of simulations to sample the parameter space and test the results
against observables from physical bulk tests. The system presented describes a
grid parallel system to distribute the simulation workload across many compute
nodes to significantly reduce the time taken to perform DEM material model cali-
bration. It has been demonstrated to be applicable to DOE and optimization meth-
ods.
REFERENCES
1. D. Curry, et al, ‘Calibration of DEM Material Models Using Advanced Opti-
mization Methods in a Grid Parallel System’, Submitted to DEM5-2010.

A GPU ACCELERATED CONTINUOUS-BASED

DISCRETE ELEMENT METHOD FOR
ELASTODYNAMICS ANALYSIS
Zhaosong Ma, Chun Feng, Tianping Liu, Shihai Li
Institute of Mechanics，Chinese Academy of Sciences，Beijing 100190，China
The GPU (Geographic Processing Units) computing is a new tech-

nique which makes use of VGA cards to provide high-performance
computing capacity with very low costs. This paper presents the Con-
tinuous-based Discrete Element Method (CDEM) realized on the GPU.
Using a NVIDIA GTX-285 VGA card, the computing speed achieves
an average 650 times speedup ratio (vs. Intel Core-Dual 2.66 GHz
CPU). To parallelize the CDEM algorithm, the clone node force re-
freshing process is separated from the elemental calculation, and is re-
placed by a “Node Group” force assignment process, which ensures
the data independence in parallel execution.
INTRODUCTION
The Continuum-based Distinct Element Method (CDEM) is an approach to
simulate the progressive failure of geological mass, which is mainly used in land-
slide stability evaluation, coal and gas outburst analysis, as well as mining and
tunnel designing.
CDEM is the combination of Finite Element Method (FEM) and Distinct Ele-
ment Method (DEM), It contains two kinds of elements, blocks and interfaces
(Fig 1). A discrete block is consisted of one or more FEM elements, all of which

share the same nodes and faces. An interface contains several normal and tangent
springs; each connects nodes in different blocks. Inside the block the FEM is used,
while for the interface, the DEM is adopted.
Various constitutive models can be used in the block, including the linear elas-
tic model, Drucker-Prager model, the block breakage model and the discrete
spring model. For the interfaces, the linear elastic model and the linear crack
model can be used. To simulate the randomness of dimension and the geometrical
features of geo-material, the random joint model is used. A fluid-solid interaction
model is introduced to CDEM to study the influence of the seepage flow in frac-
tures.
Interfa
Block 1
Block 2 Block 1 Block 2
3 ce
Interface 1 Inte
rfac
e2
Block 3
Block 3
Block 1------5 FEM elements Contact 1------4 normal and 4 tangent springs
Block 2------1 FEM element Contact 2------4 normal and 4 tangent springs
Block 3------5 FEM elements Contact 3------4 normal and 4 tangent springs
Fig 1. Blocks and interfaces for 8 nodes hexahedron

CDEM is an explicit time history-analysis FEM/DEM approach on finite dif-
ference principles and forward-difference approximation is adopted to calculate
the progressive process through a time marching scheme. During the calculation,
the dynamic relaxation method is used to achieve convergence in a reasonable pe-

riod time with small time steps, and the convergence is reached when the total
magnitude of the kinetic energy is minimized.
Fig 2 shows the main process to solve a classical geological problem.
Start
Node Acceleration
Damping
Node Velocity Damping Force
Equation
Node Displacement
No
Constitutive Equation
Block Force Interface Force
Node Force
Kinetic Energy
Is reached the Yes

Finish
convergence ?
Fig 2. The process to solve a geological problem

The constitutive equation of block is show in Equation 1 (just for liner elastic
e e
model), where {F }i denotes the vector of node force of element i, {u}i means
e
the vector of node displacement of element i, and [ K ]i represents element stiff-
ness matrix of element i.
{F }i  [ K ]i {u}i
e e e
(1)
The constitutive equation of interface is show in Equation 2 (just for liner elas-
j j
tic model), Where Fn and Fs denotes the normal and tangent force of spring j,

K n and K s means the normal and tangent stiffness of spring j, d n and d s

j j j j
represents the normal and tangent displacement increment of spring j.

 Fn j   K n j  d n j
 j (2)
 Fs   K s  d s
j j
For simulating the failure of interface, liner crack model is adopted. During
elastic stage, the same as liner elastic model, equation 2 is used to calculate the
spring force in the interface. If the spring force exceeds the strength of the inter-
face, the crack on interface begins and the Mohr-Coulomb criterion (Equation 3)
is adopted. It is positive when Fn is compression, and negative when Fn is tensile.
In Equation 3, T denotes the tension,  means inner friction, and C represents
cohesion.
 If  Fn j  T then Fn  Fs  0, T  0
 (3)
 If Fs  Fn  tan( )  C Fs  Fn  tan( )  C , C  0
j j
then
To dissipate the excessive energy in block system, the Rayleigh damping is in-
troduced, which contains two kinds of damping, the mass proportional damping
(  ) and the stiffness proportional damping(  ). The damping force is calculated
e
from Equation 4, where { f }i means the vector of damping force of element i,
e e
[ M ]i denotes the element mass matrix of element i, and {v}i represents the vec-
tor of velocity of element i.
{ f }i   [ M ]i {v}i   [ K ]i {v}i
e e e e e
(4)
SOLUTION ALGORITHM IN PARALLEL

Unlike general CDEM algorithm which uses clone node strategy to ensure
node consistency among elements, the parallel algorithm does the same work by
means of a “nodal force group” strategy. The difference is out of reason of access
conflict. For general CDEM, a node’s force is cloned to the associated nodes just
when the node’s force calculation is done. This works well in serial execution, but
may cause data access conflict in parallel execution, for various threads may clone

their forces to a same node at the same time. Differently, the parallel algorithm
uses a data structure to temporarily store the forces calculated. The nodal forces
will be redistributed in a separate procedure after all of the node’s force calcula-
tion is done, as shown in Fig 3.
Preprocessing
Copy data to GPU
Element calculation Element calculation … Element calculation
Global synchronization
Nodal force redistribution Nodal force redistribution … Nodal force redistribution
Meet iteration conditions?
Yes
Post processing
End
Fig 3. Flowchart of CDEM Parallel Calculation

The element calculation procedure is designed according to the specialty of
GPU. As is known, a GPU consists of a set of SIMT (single-instruction, multiple-
thread) multi-processors, which are mapped to CUDA blocks[1]; each multi-

processor has an instruction unit and several scalar processor cores, which are
mapped to CUDA threads, and each scalar thread executes independently with its
own instruction address and register state. Accordingly, the element calculation
processes are divided into parallel CUDA blocks, each of which contains certain
number of elements. In this work, a CUDA block contains 32 elements for best
performance. As is shown in Fig 4, the CUDA blocks are further divided into
threads, each of which processes one node.
` Block 1 Block 2 Block m
Element 1 Element 2 Element 3 Element 4 Elem. n-1 Element n
…
Nodes Nodes Nodes Nodes Nodes Nodes
Threads Threads Threads
Thread 1 Thread 2 Thread 3 Thread nn

Nodal Calculation Nodal Calculation Nodal Calculation … Nodal Calculation
…... …... …... …...
Fig 4. Thread Division Scheme for Element Calculation

* Practically in the program, each block contains 32 elements.
Similarly, the nodal force redistribution procedure is designed following the

block-thread architecture, namely, each CUDA block contains certain number of
nodal force groups as threads, as is shown in Fig 5.
The number of nodal force groups in a CUDA block should be properly deter-
mined to obtain best performance, which is discussed in the following section.

` Block 1 Block 2 Block nb

Element Element Element
Group 1 Group 2
…
Element Element Element
Group 3 Group 4
Threads Threads Threads
Thread 1 Thread 2 Thread 3 Thread ng

Element Element Element Process Process Process … Process
Forces Forces Forces Forces
Fig 5. Nodal Force Group
PERFORMANCE OPTIMIZATION
The program is developed and tested using NVDIA CUDA Toolkit 2.1, on a
GeForce GTX285 GPU. The properties of GTX285 GPU are shown in Table.1.
Total amount of global memory 1073414144 bytes
Number of multiprocessors 30
Number of cores 240
Total amount of constant memory 65536 bytes (64k bytes)
Total amount of shared memory per block 16384 bytes, 16 banks
Total number of registers available per block 16384 (16k bytes)
Warp size 32
Maximum number of threads per block 512
Maximum sizes of each dimension of a block 512 x 512 x 64
Maximum sizes of each dimension of a grid 65535 x 65535 x 1
Maximum memory pitch 262144 bytes

Texture alignment 256 bytes

Clock rate 1.48 GHz
Concurrent copy and execution Yes
Table 1. Properties of GTX285 GPU
The 8 nodes isoperimetric element is chosen as element model.

To maximize parallel execution, the number of threads per block is carefully
chosen. As is recommended by NVIDIA, the best results will be achieved when
the number of threads per block is a multiple of 64, and for ideal performance,
exceed 192 threads per block is recommended. In practice, best performance is
achieved when the number of elements per block is 16 or 32. Because of the 8
nodes element model, the number of threads per block is equal to Equation 5.
256, N e  32
Nt  Ne  8   (5)
128, N e  16
Where Nt is the number of threads per block; Ne is the number of elements per
block. When problem scale is small (1000 elements), 16 elements per block is a
little faster; for large-scale problems, 32 elements per block is a little better. The
test results are shown in Table.2:
Problem scale 16 elements per block 32 elements per block
1000 elements, 8000 steps 0.460 sec 0.478 sec

Problem scale 16 elements per block 32 elements per block
Table 2. Running Time: 16 vs. 32 Elements per Block

The effective bandwidth of each memory space depends significantly on the

memory access pattern. As described in CUDA Programming Guide, global
memory bandwidth is used most efficiently when the simultaneous memory ac-
cesses by threads in a half-warp (during the execution of a single read or write
instruction) can be coalesced into a single memory transaction of 32, 64, or 128
bytes. To achieve maximum global memory bandwidth, the data structures are
realigned to thread. Besides, shared memory is used to store nodal forces, for the
shared memory space is much faster than the local and global memory spaces, and
nodal forces are reused among threads.
EVALUATION
A simple 1000 elements model is used to verify the algorithm. The model con-
sists of a 10x10x10 hexahedron elements mesh, with the nodes on the bottom con-
strained, and gravity applied. Fig 6 shows the result comparison between the CPU
version and the GPU version. Table 3 shows the time cost comparison.
Another instance is a fortress of the Great Wall, which contains 41232 hexahe-
dron elements. The results are shown in Fig 7.
a. CPU version b. GPU version

Fig 6. Result Comparison between the CPU Version and the GPU Version

Fig 7. z Result of the Fortress
Problem Scale CPU GPU
1000 elements brick, 8000 steps 252 sec 0.478 sec

41232 elements Fortress, 4000 steps 95 min 6.10 sec
Problem Scale CPU GPU
1000 elements brick, 8000 steps 252 sec 0.478 sec
41232 elements Fortress, 4000 steps 95 min 6.10 sec
Table 3. Performance Comparison between CPU and GPU
The results of the CPU version and the GPU version are not the very same.
Average error is less than 0.01%. This is because the parallel algorithm has a dif-
ferent accumulation order from the serial one. The accumulation occurs in the
nodal force redistribution procedure. The CPU version accumulates the nodal
forces one by one in the order of the elements, while the GPU version has a dif-

ferent order, which depends on the generation of nodal force group. Besides, the
Arithmetic Logical Units on CPU and GPU have different respective error bounds.
REFERENCES
1. NVIDIA, NVIDIA CUDA Programming Guide,
http://developer.download.nvidia.com/compute/cuda/2_1/toolkit/docs/NVIDI
A_CUDA_Programming_Guide_2.1.pdf, 2008.
2. Li SH, Zhao MH, Wang YN, Rao Y, ‘A New Numerical Method for DEM-
Block and Particle Model’, International Journal of Rock Mechanics and Min-
ing Sciences, 2004, No.41(3), 436, 2004.
3. Liu Tianping, Li Shihai, Liu Xiaoyu, Cheng Kai, ‘A Continuum-based Dis-
crete Element Method of Stochastic Block Shape for Analyzing Stability of
Cataclastic Rock Mass Slope’, Beijing DEM’08, 2008.9.





CALIBRATION OF DEM MATERIAL MODELS USING

OPTIMISATION METHOD IN A GRID PARALLEL
SYSTEM
David Curry, Elson Mourao, Richard LaRoche

DEM Solutions Ltd
The predictive power of a DEM simulation depends on the degree to

which the parameter values in the chosen DEM material model produce
the desired level of approximation of physical bulk particle behaviour. A
good way of finding the best parameter values is through calibration using
simulated and physical bulk tests to characterize the bulk behaviour of
interest. This paper describes how advanced optimization methods
combined with simulation using a grid parallel network of computers can
be used calibrate a DEM material model.
BACKGROUND
Handling and processing of granular materials is a significant factor in the cost of
production in many industries. Very often relatively small percentage improvements
in performance can produce a substantial increase in production efficiency, a
reduction in product waste, or an increase in through-put. Such improvements can be
difficult to achieve using a purely empirical approach to equipment and process

design which limits innovation due to constraints of time and cost required for
physical testing, and reliance on correlations specific to particular equipment and
operating regimes.
The accuracy of the approximation of the mechanical and other properties at the
bulk scale is determined by the mechanical and inertial properties of the particles as
they interact with each other and their environment. Parameters such as particle size
and shape, coefficients of friction, shear modulus, etc. all make up the DEM material
model for a particular bulk solid material. In common with other numerical
techniques the material models provide a means of scaling the relationships between
particle-scale and bulk-scale properties.
Current methods for determining suitable material model parameters are based on
physical tests sometimes also using comparison of process-scale simulation results
against observed behaviour, to tune the model. In many cases, experimentally-derived
material properties will not reproduce the desired material behaviour in a DEM
simulation to the level of accuracy that is required. Calibration using process-scale
measurements can be very time consuming, and is not independent of the particular
equipment design and operating conditions used in the test. This paper focuses on the
application of the simulated annealing method to calibrating a DEM material model
using a grid parallel system [1].
DEM MATERIAL MODEL CALIBRATION METHOD

The material calibration method combines laboratory bench tests, EDEM
simulation, DOE analysis and application of optimization techniques. The process
varies the critical simulation parameters in order to achieve the optimized parameter
values that produce simulations that accurately match the experiments.
Fig 1 shows a schematic of the method.

Fig 1. Schematic of Optimization Process.

At the heart of this approach is the optimization routine. This performs multiple
EDEM simulations and compares the predicted results to the objective function. As it
progresses, each variable is changed according to the optimization algorithm and the
sensitivity of each material property until the difference between the prediction and
the objective function is minimized.
The nature of such a method is that many simulations are required to ensure that
the closest possible match is achieved. If a large number of parameters are considered
to significantly influence the observable then the number of simulations required to
complete an optimization can reach many 1,000’s. The effort required can be reduced
through use of a small test; however it is still a significant cost in terms of time if the
simulations are performed in serial. Consequently, the optimization methods
implemented have all been selected and coded to make use of a gridded cluster of
computers to significantly reduce processing time [1].
SIMULATED ANNEALING APPROACH

Simulated annealing is a method that is based on a local search; however it
introduces an element of probability (based on a temperature and cooling parameter)

that enables the algorithm to escape local minimums [2]. To illustrate this approach,
an example of parameter calibration for determining an angle of repose of particles
through variation of inter-particle static and rolling friction is used. The work was
performed using a gridded hardware configuration and automated angle of repose
calculation tools. The required repose angle was defined, and the resulting global
optimum identified. A starting point was selected that would result in a false optimum
for a local, gradient based method. Suitable optimization parameters were set (based
on a sensitivity study the simulated annealing inputs) and the simulated annealing
approach commenced. Fig 2 shows an example results from the optimization
analysis. The path taken is shown in black and the start (red square) and end points
(blue square) are also shown. The green square indicates the optimal solution.
Fig 2. Left: Successful simulated annealing optimization of static and rolling

friction parameters. Right: A failed optimization run.
Fig 3: Left shows a successful path where the optimum result was found in 743
simulations. This equates to 3.8 days using the gridded computing system and a
simulation time of 1 hour per individual run. An exhaustive search performed using
standard, serial computing hardware would require 416 days to produce the same
result from 10,000 simulations.

Since the simulated annealing method is based on a probability, starting the

optimization from the same location with all optimization parameters fixed will result
in a different solution. Analysis of the standard implementation indicated that the an
optimum would be found achieved 50% of the time. Inspection of failed results
demonstrated that the main cause was the algorithm becoming trapped not in a local
minimum, but at the extremity of the parameter space (Fig 3: Right).
The simulated annealing algorithm was refined to introduce a check that limits the
number of iterations allowable at the extremes of a parameter. If this limit is reached,
then the algorithm will restart from a randomly chosen location in the parameter
space. Fig 3 shows the path, including restart event, consisting of 835 simulations (=
4.3 days) taken by the advanced simulated annealing algorithm. Introducing the
restart capability refinement increased the success ratio from 50% to 95-100% for
simulated annealing.
Fig 3. Path taken by simulated annealing algorithm with restart enabled.

SUMMARY
Incorporating DEM simulations early in the design process can have major
impacts on revenue in terms of savings and improved margins. Savings can be found
in reduced design cycles, faster time-to-market, more efficient operation, and reduced
prototyping and testing costs. Greater profits can be found by reducing product costs
by eliminating over-designed parts and inefficient processes. Technically, DEM
software such as EDEM provides the simulation tools needed. However to achieve
maximum technical and financial benefits from DEM simulation attention needs to be
given to calibration of DEM models of bulk materials to provide the necessary
quality of engineering data at the required scale.
With access to a gridded computing system, advanced global optimization
procedures can be applied to converge on DEM Material Model parameters using
simulations of physical bench tests from which the same observables as those
measured are extracted. These advanced methods, in conjunction with the grid
capability, greatly reduce the simulation effort required to calibrate the material and
interaction parameters that define a DEM Material Model of a real-life bulk system.
REFERENCES
1. E. Mourao, ‘Grid Parallel Computing System for Calibration of DEM Material
Models‘, submitted to DEM5, 2010.
2. ‘How to Solve It: Modern Heuristics‘, 2nd Edition, Michalewicz & Fogel,
2004.

FRACTURE AND FRAGMENTATION, CRACKS

DEVELOPMENT OF DISCRETE ELEMENT

METHOD FOR SIMULATION OF DEFORMATION
AND FRACTURE OF HETEROGENEOUS
ELASTOPLASTIC MATERIALS
Sergey G. Psakhie, Evgeny V. Shilko, Alexey Yu. Smolin,

Sergey V. Astafurov, Artem Yu. Panchenko*
*Tomsk State University, Tomsk, Russia
Authors propose an approach to realize relations of elasticity and plas-

ticity models, which are conventionally written in terms of
stress/strain tensor components, within the framework of discrete ele-
ment approach.
INTRODUCTION
Discrete element method (DEM) is known to be a reliable tool for simulation
of deformation and fracture of condensed media of various nature. The term DEM
combines various realizations of the approach to modeling the solid by ensemble
of interacting bodies (particles) having final size and defined shape. Evolution of
ensemble is governed either by Newton’s equations of motion or system of equa-
tions for potential energy minimization [1]. An important problem in DEM is
formulation of particle interaction, which provides mechanical response of par-
ticle ensemble conforming to simulated material or media. In spite of considerable
progress in DEM, up to now two-particle forces are mainly used [1,2]. This may
lead to a series of difficulties. Particularly, use of pair interaction makes additional

problems in correct simulation of irreversible strain accumulation in ductile mate-

rials.
Analysis of state of the art shows that described problems can be solved even
on regular packing of equal size elements with use of many-body interaction
forces. An approach to building the expressions for normal and tangential interac-
tion of discrete elements simulating isotropic elastoplastic medium is proposed in
the paper. The approach is realized within the framework of two-dimensional ver-
sion of the movable cellular automaton (MCA) method [3], which can be consi-
dered as one of realizations of discrete element approach.
DESCRIPTION OF ELASTOPLASTIC MEDIUM WITH MCA

FORMALISM
Proposed approach is based on use of conventional models of elasticity (or
elastoplasticity) for formulation of normal and tangential interaction of discrete
elements (movable cellular automata). Unlike conventional DEM formalism [2],
within the framework of the MCA method rheological properties are assigned to
individual element (or kind of elements) and not to the interacting pair. Because
of this, each element is characterized by personal parameters of elasto-plastic re-
sponse.
Conventional models of isotropic elasticity and elastoplasticity are formulated
in terms of stress () and strain () tensor components, whereas within the
framework of DEM main local parameters are element interaction force and rela-
tive displacement vectors. Authors propose an approach to formalization of inter-
relation between local values of stress/strain tensor components and
force/displacement vectors in pairs of interacting cellular automata (elements). It
is based on the following four approximations (quasi-two-dimensional realization
of the MCA method is considered below):
1. In the local coordinate system of considered pair i-j (Fig. 1) strain tensor
components yy and xy are associated with normal (ij) and shear (ij/2) relative
displacements normalized by automaton size. Their values are calculated by in-
crements:

rij qij q ji
 ij      i  j    j i 
d i  d j  2 di 2 dj 2
(1)
ij
Vshear t
 ij    i  j    j i 
rij
where symbol  hereinafter indicates increment of corresponding parameter dur-

ing one time step t, rij is the distance between mass centers of automata i and j
(Fig. 1), qij and qji are the distances from mass centers of automata i and j to the
center of contact area (qij+qji=rij), d is size of automaton, i(j) and j(i) are strains of
automata i and j in i-j pair, i(j) and j(i) are contributions of automata i and j to
the total value ij.
2. In the local coordinate system of the pair i-j stress tensor components yy
and xy are associated with specific normal (ij) and shear (ij) interaction forces
as follows:
ij
Fnorm   ij S ij (2)
ij
Ftan g   ij S ij
where Sij is area of contact between automata i and j.

3. Association for stress tensor component xx cannot be found considering i-j
pair independently of surrounding automata. In the simplest case it can be defined
following the procedure in [4]. Nevertheless, more correct approximation is based
on stress homogenization procedure [2,5]:
Ni
 qik S ik cos ij ,ik cos ij ,ik  ik  sin  ij ,ik ik 

1
 ixx j   (3)
Vi k 1
where Ni is number of neighbors of automaton i, Vi is the volume of automaton i,

cos(ij,ik) is defined as shown in the Fig. 1 b.

4. Stress tensor component zz could be found using plane stress state
(  izz j   0 ) or plane strain state (  izz j  
K 
2K  
 
 ij   ixx j  ) approximation. In
both cases  ixz j    iyz j   0 .
hij j
n
k
di i j dj m i ij,ik
g l X
rij
a b
Fig 1. Definition spatial parameters of the pair i–j (a) and
instantaneous local coordinate system (b)
Described statements give the possibility to calculate normal and shear forces
of interaction of cellular automata whose ensemble simulates isotropic linearly
elastic medium.
Specific normal force (ij=ji=yy) in i-j pair, as well as strains i(j) and j(i), are
calculated in an incremental fashion by rewriting Hooke’s law for diagonal stress
tensor component yy:
   i j     j i 
 ij   ji  i  i  j   1  i  m   j  j i   1  j  m (4)
 Ki   Kj 
 
where i and Ki are elastic constants of material of the automaton i, mean stresses
 im j  and  mj i  are taken from the previous time step
 
(  im j    ij   ixx j    izz j  3 ,  mj i  is calculated in the same way).

Specific tangential (shear) interaction force in i-j pair (ij=ji=xy) as well as

strains i(j) and j(i) are calculated following rewritten Hooke’s law for nondiagonal
stress tensor element xy:
i j
ij   ji   i  j    j i  (5)
2 2
Proposed association of specific forces and relative displacements of automata
(discrete elements) with local values of stress and strain tensor components seems
to be general for all realizations of DEM and provides potential possibility to im-
plement various models of elastoplastic or viscoelastoplastic media. In particular,
the authors applied the proposed approach to realize deformation plasticity theory
with elastic unloading (DPT) and incremental plasticity theory (IPT). Implementa-
tion of the DPT is based formally on the same expressions (1)-(5) which were
used for the model of linear elasticity. The main difference is connected with vari-
able value of  modulus within the framework of DPT (ii(j)). Current (instan-
taneous) value of i(j) is calculated from assigned relationship  int    int  for
material of automaton i (when following the loading surface) or possesses its elas-
tic value (when being inside the loading curve). Stress intensity for automaton i in
i-j pair is calculated as follows:
 iint j  
1
2
       
i j
xx ij
2
ij   izz j           
2 i j
zz
i j
xx
2
 6 ij2 (6)
Implementation of IPT is carried out similarly, however expressions (4)-(5) are

replaced by the other relations.
Testing of the developed approach showed that an ensemble of interacting
movable cellular automata demonstrates macroscopically isotropic mechanical
response even in case of uniform packing of equal-size elements. Comparison of
the results of a series of tests with results of FDM simulation with use of the same
plasticity models showed their close fit. This argues for correctness of the pro-
posed approach and gives grounds for its application to simulation of heterogene-
ous elastoplastic materials and media.

CONCLUSION
The authors proposed an approach to build associations between the compo-
nents of stress/strain tensor of simulated medium and the inter-automaton
forces/displacements. The proposed associating allows one to rewrite relations of
the applied model of plasticity (these are conventionally written in terms of
stress/strain tensor components) in terms of forces and displacements or their in-
crements. The developed approach provides potential possibility to realize various
models of elastoplastic or viscoelastoplastic media in the framework of DEM.
Furthermore, it allows one to get isotropic deformation pattern even on regularly
packed particle ensembles. Additional advantage of the proposed approach is the
possibility of direct use of conventional multiparametric fracture criteria (Huber-
Mises-Hencky, Drucker-Prager, Mohr-Coulomb, Podgorski etc.), which are writ-
ten in tensor form.
ACKNOWLEDGMENTS
The investigation has been carried out within the Project No.2 of RAS Presi-
dium Program No.11, at partial financial support of the RFBR Grant No. 09-05-
00968-а and Russian science support foundation.
REFERENCES
1. L. Jing, O. Stephansson, Fundamentals of discrete element methods for rock
engineering: theory and applications, Elsevier, 2007.
2. D. Potyondy, P. Cundall, ‘A bonded-particle model for rock’, Int. J. Rock
Mech. Min. Sci., 41, No.8, 1329-64, 2004.
3. S. Psakhie et al.‚ ‘Movable cellular automata method for simulating materials
with mesostructure’, Theor. Appl. Fract. Mech., 37, No.1-3, 311-34, 2001.
4. S. Psakhie et al., ‘Interpretation of the parameters of the method of movable
cellular automata on the basis of continuum description’, Phys. Mesomech., 3,
No.3, 89-92, 2000.
5. A. Love, A Treatise on the Mathematical Theory of Elasticity, Cambridge Un-
iv. Press, 1927.

FRACTURE BEHAVIOUR OF HIGHLY POROUS

CERAMICS
Xiao-xing Liu and Christophe L. Martin

Laboratoire SIMaP – GPM2, Grenoble-INP, BP46, 38042 Saint Martin d’Hères cedex, France.
Email: [email protected]
A discrete element model for the elastic and fracture behaviour of

highly porous ceramics is presented. It uses elastic force-displacement
laws to represent the bonds formed between particles during sintering
and accounts for bond fracture. Using numerical microstructures gen-
erated by a sintering model, we compute the effective strength of these
microstructures in tension and compression. Simulation results are
quantitatively compared to experimental data on alumina ceramics at
various porosities.
INTRODUCTION
Applications of porous ceramics are numerous: thermal isolators, filters, elec-
trodes for SOFC or SOEC (Solid Oxide Fuel or Electrolyser Cells), and mem-
branes. Fundamentally, there is an inherent contradiction for these applications
between the optimal microstructure for functional performance (high porosity)
and mechanical performance (low pore volume). These porous ceramics are ob-
tained by partial sintering at high temperature of powders leaving typically 20 to
50% porosity. Thus, the discrete character of the ceramic material needs to be
taken into account when modelling its mechanical behaviour. To our best knowl-
edge, the only attempt to use DEM for simulating the fracture of highly porous

ceramics was described in [1] on complex mullite/alumina mixtures. These au-

thors enriched the model developed in [2] to incorporate the fracture of bonds be-
tween sintered particles. The model presented in this paper accounts for the elastic
and fracture behaviour of the solid bonds formed between the discrete particles
that have sintered together. The microstructure that mimics the porous ceramic is
obtained from a sintering numerical procedure described in [3].
MODEL DESCRIPTION
Particles that have partially sintered are modelled as truncated spheres overlap-
ping geometrically (Fig. 1). The normal and tangential forces between two bonded
particles with elastic constants (E, ν) and bond radius ab are given by:
E 2E
Nb = f N ab u N , Tb = − f au , (1)
1 −ν 2 ( 2 − ν )(1 + ν ) T b T
where the f N and fT describe the interaction between neighbouring bonds [2] and
uN and uT are the normal displacement (tensile or compressive) and the accumu-
lated tangential displacement, respectively. Note that bonds transmit resisting
moments, M N and M T , in the normal and in the tangential direction respectively.
R1+ R2-hb R1+ R2-hb+uN
R2
ab
MT
R1 MT
MN Nbb
T
Nb
Tb
(a) (b) (c)
1. The three possible stages of a solid bridge forming a bond between two particles. (a) Initial
status. (b) Broken bond. (c) Resumed contact after fracture.

Bond failure occurs when the tensile or the shear stress is larger than a critical
value σ c , determined from the singular stress field along the edge of the bond [4]:
E Γ
σc = , (2)
1 − ν π ab
2
where Γ is the bond toughness that we approximate to the surface energy contri-
bution ( Γ ≈ 2γ s =1 J.m-2). Elastic properties are taken from the fully dense alu-
mina material (E=400GPa). The contact law of a broken bond between two parti-
cles that have resumed contact has been detailed in [5]. It allows the broken bond
geometry to be taken into account in a simplified manner (Fig. 1c).
ELASTIC AND FRACTURE BEHAVIOUR

Two planes bonded to the particles are used to impose a tensile or a compres-
sive strain to the 3D rectangular samples which are made of 10,000 particles typi-
cally. Fig. 2 shows the typical fracture patterns in tension (a crack growing from
the free surface) and in compression (shear bands developing at 45°). The axial
stress is calculated from the total reaction forces on the planes
Fig 2. Typical fracture behaviours in tension (left) and compression (right).

Colours indicate the number of broken bonds per particle.

a) b)
Fig 3. a) Typical stress-strain curves in tension and compression for 0.72 relative
density (28 % porosity) and b) Evolution of fracture stress in tension and com-
pression with relative density.
The typical stress-strain curves for tension and compression tests are shown in
Fig. 3a, demonstrating that the material behaviour is symmetric in tension and
compression in the elastic zone but that the behaviour is much more brittle in ten-
sion than in compression, as it should. Systematic tests were carried out at various
relative densities. Fig. 3b shows that strength increases as density increases (or as
porosity decreases). It should be noted that high temperature sintering of ceramics
proceeds by an enlargement of solid bonds between particles (the ab radius) and
by an increase in the average number of contacts per particle. These two phenom-
ena explain the resulting strength of the porous ceramics at room temperature.
Fig. 3b also shows a comparison with four-point bending tests on alumina [6].
The comparison is rather satisfying considering that only the elastic constants
were fitted to the experimental data. Thus, the value of the bond toughness, which
we have approximated to twice the typical surface energy for ceramics
( Γ ≈ 2γ s =1 J.m-2), leads to a very good agreement. This would indicate that frac-
ture at the length scale of the submiconic size bonds between ceramic particles,
which are typically 1 µm in size, does not operate by any dissipative mechanisms
like plasticity, crack tilting or twisting. Rather, the only work needed at this length

scale would be the one used to create two free surfaces. Note also that this micro-
scopic value of the toughness is much smaller than the macroscopic toughness at
the macroscopic scale (typically closer to 40 J.m-2 from experimental data).
CONCLUSIONS
We have presented a realistic model for the fracture of porous ceramics using
DEM. Building from a contact law derived for sintered contacts, we show that
such a model is able to represent both qualitatively and quantitatively the behav-
iour of highly porous ceramics that have been partially sintered from powders.
The advantage of building a model with realistic contact laws which introduce
measurable material parameters (elastic constants typically) rather than artificial
contact stiffness is that it allows for a critical comparison with experimental data.
This comparison allows realistic scenarios on the fracture mechanics at the micro-
scopic scale to be proposed.
Finally, it should be noted that it would be simple to apply this model to intro-
duce initial defects in the numerical sample. Notched samples, for example, would
allow partially sintered samples to be tested within the framework of fracture me-
chanics by calculating the fracture toughness of the numerical sample at the mac-
roscopic length scale.
ACKNOWLEDGEMENT
The Agence Nationale pour la Recherche (ANR) is greatly acknowledged for fi-
nancial support (ANR-07-PANH-009 project).
REFERENCES
1. H Fujita, G Jefferson, R M McMeeking, F W Zok. ‘Mullite/alumina mixtures
for use as porous matrices in oxide fiber composites‘. J. Am. Ceram. Soc. 87,
261 - 267 (2004).
2. G. Jefferson, G.K. Haritos, R.M. McMeeking, ‘The elastic response of a cohe-
sive aggregate - A discrete element model with coupled particle interaction’ J.
Mech. Phys. Solids, 50, 2539-75, 2002.

3. C L Martin, H Camacho-Montes, L Olmos, D Bouvard, R K Bordia. ‘Evolu-

tion of Defects During Sintering: Discrete Element Simulations‘. J. Am. Ce-
ram. Soc. 92, 1435 - 1441, 2009.
4. L B Freund, E Chason. ‘Model for stress generated upon contact of neighbor-
ing islands on the surface of a substrate‘. J. Appl. Phys. 89, 4866-4873, 2001.
5. C. L. Martin, D. Bouvard, G. Delette, ‘Discrete element simulations of the
compaction of aggregated ceramic powders’ J. Am. Ceram. Soc., 89, 3379-87,
2006.
6. T Ostrowski, J Rödel. ‘Evolution of Mechanical Properties of Porous Alumina
during free Sintering and Hot Pressing‘. J. Am. Ceram. Soc. 82, 3080-3086,
1999.

NUMERICAL STRATEGY
FOR DISCRETE FINE CRACKING DESCRIPTION
WITH CONTINUOUS MODEL
Arnaud Delaplace1 , Benjamin Richard2 , Cécile Oliver1 ,

Frédéric Ragueneau1
1
LMT Cachan (ENS Cachan/CNRS/Universite Paris 6/UniverSud Paris)
61, avenue du Président Wilson, 94235 Cachan, France.
2
University Paris-Est, Laboratoire Central des Ponts et Chaussées, France
This study deals with a combined continuous/discrete model

for describing cracking in brittle heterogenous materials like
concrete. The proposed idea is to use a macro description at
the structure level, and a discrete model as a reanalyzing tool
of the damage area with the objective of providing a fine crack
description. This numerical strategy is exposed and applied
to a single cracking test. An extension will be proposed to a
reinforced structure, for which the interface between concrete
and rebars plays an important role in cracking phenomena.

INTRODUCTION
The description of damage zones in large scale structures is nowadays possible
with a Finite Element code and an appropriate damage model. But a fine
description of cracking (crack pattern, crack opening, crack tortuosity) has
become a new requirement, especially when dealing with structure durability.
A numerical strategy, based on (i) a full resolution at a macroscale with a
continuous model and (ii) a reanalysis (seen as a post-treatment) of the damage
zones at a meso-scale with a discrete model is proposed. After a summary of
the method, a numerical illustration on the classical Geers beam is proposed.
An extension to reinforced structure is under progress and will be presented
during the conference.
NUMERICAL STRATEGY
The main assumption of the proposed strategy is to suppose that the continu-
ous model used at the macroscopic scale (the structure scale) is pertinent and
gives a reliable description of the damage areas. In the following, a damage
model expressed within the framework of the irreversible processes thermody-
namics is used [1, 2]. It is dedicated to the description of concrete behavior,
taking into account the dissymmetry between the behaviors in tension and in
compression. The discrete model used to reanalyze the damage zone is based
on a Voronoi particle assembly, linking by brittle beams [3, 4]. This model has
been successfully applied to the study of brittle material, either in 2D or 3D
[5, 6].
Once the computation at the macroscopic scale is done, a sub damage
region R is reanalyzed at the mesoscopic scale. The loading condition is an
imposed displacement field computed from the macroscale analysis and applied
to the boundary ∂Ru of R (figure 1). A Neumann condition is considered on
the free boundary (i.e. σ.n = 0 where σ is the stress tensor and n the normal
of the considered boundary).

1
11
F ∂Ru 10
F
11
Figure 1. A scheme of the Geers beam, with the reanalyzed region R. Bold
lines represent the boundary ∂Ru where the imposed displacement is applied.
NUMERICAL EXAMPLE
The strategy is applied to the study of the crack propagation on a single notch
beam [7, 8]. We focus on the central part of the beam (figure 1), where the dam-
age grows around the notch. A first analysis is performed at the macroscale,
leading to the damage map represented on the left side of figure 3. For dif-
ferent loading steps, the discrete analysis is performed using the interpolated
displacement field computing at macroscale. An example of such imposed
displacement field is shown on figure 2.
Figure 2. The discrete mesh superimposed with the finite element mesh in
the deformed elastic configuration.

The iterative process is necessary when dealing with non linear behavior.
The initial state of each mesoscale analysis is taken as the cracking pattern
computed during the previous analysis. As it will be discussed, an important
point is that the number of reanalyses at mesoscale could be lower than the
number of steps performed at macroscale. As expected, one can observe a
good agreement between the damage zone ((figure 1)-left) and the cracking
pattern ((figure 1)-right), although the macro- and micro- models are based
on opposite assumptions. Some considerations about the distance between the
two models will be addressed.
Figure 3. The cracking pattern obtained with the proposed strategy.

CONCLUSION
We described a numerical strategy to obtain a fine description of cracking on
a large scale computation. The main feature of this strategy is that it can be
seen as a post-treatment and does not need a non intrusive modification of the
macroscale analysis. It can be applied to brittle material as well as reinforced
material as it will be presented during the conference.
References
[1] F. Ragueneau, C. La Borderie, and J. Mazars, “Damage model for concrete
like materials coupling cracking and friction, contribution towards struc-
tural damping: first uniaxial application,” Mechanics Cohesive Frictional
Materials, vol. 5, pp. 607–625, 2000.
[2] B. Richard, F. Ragueneau, C. Cremona, L. Adelaide, and J. Tailhan,

“Isotropic continuum damage mechanics for concrete under cyclic load-
ing: stiffness recovery, inelastic strains and frictional sliding,” Engineering
Fracture Mechanics, vol. in press, 2010.
[3] J. G. M. Van Mier, M. R. A. Van Vliet, and T. K. Wang, “Fracture mech-

anisms in particle composites: statistical aspects in lattice type analysis,”
Mech. Mater, vol. 34, pp. 705–724, 2002.
[4] G. A. D’Addetta, F. Kun, and E. Ramm, “On the application of a discrete

model to the fracture process of cohesive granular materials,” Granular
Matter, vol. 4, pp. 77–90, 2002.
[5] A. Delaplace and R. Desmorat, “Discrete 3d model as complimentary nu-

merical testing for anisotropic damage,” International Journal of Fracture,
vol. 148, pp. 115–128, 2007.

[6] A. Delaplace, “Tensile damage response from discrete element virtual test-
ing,” Geomechanics and Geoengineering, vol. 4, pp. 79–89, 2009.
[7] E. Schlangen, Experimental and numerical analysis of fracture processes in

concrete. PhD thesis, Delft University of Technology, 1993.
[8] M. G. D. Geers, R. de Borst, and R. H. J. Peerlings, “Damage and crack

modeling in single-edge and double-edge notched concrete beams,” Engi-
neering Fracture Mechanics, vol. 65, no. 2-3, pp. 247 – 261, 2000.

Implementation of 6-node element in DFEM for quasi-brittle materials
Xu Chunhui Li Mingrui Y uan Li

College of Science, China Agricultural University, Beijing 100083, P. R. CHINA
Abstract A new 6-n ode fi nite deformation el ement is in troduced i n t he d iscrete fin ite e lement
method (DFEM) for treating th e quasi-brittle materials. Com paring the num erical precision,
convergence r ate and t he ti me consum ption of the 6- node element w ith th ose o f the 3-n ode
element, it demonstrates t he worthiness to be adop ted in application though we h ave to use the
3-point integration instead. In the implementation to the strain-softening materials of anisotropic
rotating sm eared cra ck m odel t he 6-node e lement ac tually shows higher ef ficiency a nd m ore
precise than the 3-node element while using an even larger time integration step. A new method of
node splitting and explicit contact searching algorithm of the 6-node element are introduced.
Key word: 6-node element, Quasi-brittle materials, Rotating smeared crack model

ROCK IMPACT MODELLING USING FEM/DEM
Andrea Lisjak, Giovanni Grasselli

Geomechanics Research Group, Lassonde Institute, Department of Civil Engineering, University
of Toronto, Canada
The accurate simulation of rock i mpact physics is a pivotal aspect of

modeling rockfall phenomena. In the current study we present the i m-
plementation of a new dissipative contact interaction algorithm in the
combined fin ite discrete-elem ent method (FEM/DE M) Y-code. The
numerical results show how the energy loss process during impact can
be reproduce d and different coefficients of restituti on obtained. Fi-
nally, the effect of fracturing is analyzed.
INTRODUCTION
Over the past three decades m any expe rimental and num erical studies have
been carried out aimed at understanding the mechanics of rockfall. Ultimate scope
of these stu dies is the predic tion o f traje ctories and velo cities of f alling rock
blocks, and energies associated to rockfalls.
Several energy dissipation mechanisms may contribute to the observed velocity
reduction of a boulder during its fall. Neglecting air friction and reduction in slope
angle, all causes of energy loss are rela ted to th e rock-s lope intera ction such as
plastic deform ations at contact, rate depe ndent m aterial behavi or, elastic waves,
crushing, fracturing, friction. Since the en ergy dissipation during a rockfall event
is strictly related to the contact between colliding bodies, the impact model plays a
key-role in the num erical sim ulation of the phenom enon. The m ost widely
adopted approach is to use lum ped-mass models in which so–called coefficients

of restitution account for energy loss dur ing each im pact. The m ain limitation of
these num erical tools is the general inabil ity to consider factors such as boulder
shape, rotational velocity, block fractur ing and fragm entation and subsequent
fragments interaction.
ENERGY DISSIPATION IN FEM/DEM

In order to overcom e the lim itations m entioned above and correctly sim ulate
rockfalls, the com bined finite-discr ete elem ent m ethod (FEM/DEM) Y-code
(Munjiza, 2004) was employed. To specifica lly account for the loss in kinetic en -
ergy during the impact of a rock block against the sl ope, a new energy dissipation
mechanism was added to the three alre ady implemented in the FEM/D EM code.
In general, the total amount of energy dissipated by a system can be expressed as:
Ediss = Edamp + Wfric + Wfrac + Econt (1)

where Edamp is the energy dissipated through viscous dam ping; Wfric is the f ric-
tional work; Wfrac is the work done to propagate fractures; Econt is the energy dis-
sipated at the contact.
Viscous da mping acts by reduc ing the am plitude of elas tic oscilla tions in th e
continuum body. In Y-code viscous dam ping is added as an additional term in the
finite element stress strain constitutive law.
Frictional dissipation occurs at the inte rface between discrete elem ents and, in
Y-code, it is im plemented as a function of the f riction coefficient, normal stres s
and sliding distance.
During the fracturing process fracture nuc leation and growth dissipates energy
through (i) the creation of new surfaces and (ii) plastic deform ations ahead of the
crack tips. In order to account for both processes, a combined single and sm eared
crack m odel (Munjiza, 2004) is im plemented in FEM/DEM code. A new algo-
rithm has been added to calcu late the energy dissipated through fracturing based
on the integration of the product of nodal ve locities and forces acting on joint co-
hesive elements with respect to time. Ma terial parameters that ultim ately control
the amount of energy spent for fracturing ar e fracture energy release rate; tensile
strength; cohesion; friction angle.

The original Y-code does not include a ny feature for dissipating energy at the
contact. Thus, a new algorithm wa s im plemented as discussed in the following
section.
DISSIPATIVE CONTACT INTERACTION ALGORITHM
To simulate rockfall phenomena, the introduction of the last term in Eq. 1 was
necessary since, as seen in the field, most of the kinetic energy in rock impact is
lost during the compression-rebound process at the contact. Therefore, the original
contact interaction algorithm im plemented in Y-c ode (Munjiza et al., 2000) was
modified based on the model proposed by An et al. (2007).
In Y-code the interaction of two dis crete elements is based on a penalty func-
tion method: contacting elements are allowed two penetrate each other resulting in
contact stresses proportional to the over lapping area through a normal penalty co-
efficient. If the pen alty coefficient, wh ich ultimately represents the con tact stiff-
ness, is constant during bot h loading and unloading the m odel elastic behavior is
reproduced and energy is preserved. Howe ver, if during the restitution phase the
stiffness is varied, damping can be intr oduced into the m odel by numerically re-
moving kinetic energy from the element contacting couples.
Specifically, the elas tic–power damping contact model illustrated in Fig 1 was
implemented. In its im plementation, two c ontact parameters are needed. The first
is the initial normal penalty coefficient, p, which represents the com pressive stiff-
ness of the boulder-slop e interface and aff ects the rate at w hich con tact stresses
are developed. The second is the power dam ping exponent, B, which plays a cru-
cial role in contro lling the energy d issipation: the higher th e value the large r the
amount of energy dissipated.
NUMERICAL EXAMPLE
In orde r to illustrate the newly implem ented contac t a lgorithm, the m odel
shown in Fig 2(a) was created. It consists of circular boulder having a radius of
0.1m and impacting vertically against a rigid ground. Since the acceleration of
gravity is neglected only inertial forces are considered. The density of the boulder
is 2,700kg/m 3. The kinetic energy of the boulder before im pact, KE b, is equal to
4.1kJ. The normal penalty coefficient, p, of the rock-ground interface is 40MPa.

Fig 1. Elastic – power damping contact constitutive relationship.
Fig 2. (a) Model used to evaluate the energy dissipation during impact and (b)
kinetic energy of the boulder for different values of the power damping exponent.

In a f irst se t of sim ulations the bou lder was not allowed to fracture, viscous
damping was set to zero and therefor e the effect of varying the exponent B of the
contact model was studied. Fig 2(b) presents the kinetic energy of the boulder as a
function of time for B values equal to 1, 2, 5, 10 and 20. The energy and velocity
coefficient of restitution are reported in Table 1. By changing B the amount of ki-
netic energy removed from the system can be adjusted in order to obtain coeffi-
cients of restitu tion suc h as those experim entally m easured (ie, Azzoni et al.
1995).
Power damping exponent, B RKE (KEafter/KEbefore) RV (Vafter/Vbefore)
1 0.93 0.96
2 0.58 0.76
5 0.28 0.53
10 0.16 0.40
20 0.09 0.30
Table 1. Kinetic energy and velocity coefficients of restitution for different values
of the contact power damping exponent B.
Finally, the sim ulation of rock fragm entation during rock fall events is im por-
tant because it can p roduce small fragments having trajectories and velocities dif-
ferent from those of the im pacting block. The effect of fra cturing on impact be-
havior was analyzed in a sim ulation with B=5 while allowing rock breakage. The
following rock strength param eters were used: friction angle = 45 o; cohesion =
12MPa; ten sile strength = 5MPa ; f racture energy release rate = 20J /m2. Fi g 3
shows how upon i mpact a number of fragments detach from the boulder and start
moving in horizontal direction. From an energy point of view, by allowing frag-
mentation in the m odel, a further reduc tion of total kinetic energy was observed
with respect to the elastic case. Approxim ately 0.5 kJ were dissipated because of
the fragm entation process and friction be tween broken rock fragm ents, 30% of
which was spent to create new fractures.

Fig 3. Impact simulation with fracturing boulder.
CONCLUSIONS
This study shows that the new modified FEM/DEM code is capable of simulat-
ing the energy loss that is observed during rock fall im pacts. Friction, fracturing
and non-conservative forces acting during ro ck impact are directly taken into ac-
count and a calibration based on experim ental coefficients of restitution is possi-
ble. Future work will focus on the simulation of in-situ rockfall tests.
ACKNOWLEDGMENTS
The Y2D FEM/DEM code was originally developed by Antonio Munjiza from
Queen Mary - University of London.
REFERENCES
1. A. Munjiza, The combined finite-discrete element method, J. Wiley & Sons,
2004.
2. A. Munjiza, K.R.F. Andrews, ‘Penalty function m ethod for com bined finite-
discrete element systems comprising large number of separate bodies’, Inter-
national Journal for Numerical Methods in Engineering, 49, 1377-1396, 2000.
3. B. An, D.D. Tannant, ‘Discrete elem ent method contact model for dynam ic
simulation o f inelastic rock im pact’, Computer & Geosciences, 33, 513-521,
2007.
4. A. Azzoni, M.H. de Freitas, ‘Experim entally gained param eters, decisive for
rock fall analysis’. Rock Mechanics and Rock Engineering, 28, 111-124, 1995.

GRANULAR MATERIALS, POWDERS AND

NANOPOWDERS

THE 3D SPLASH BEHAVIOR FOR THE

IMPACTING OF ONE PARTICLE ON A
PARTICULATE PACKING
Mao Xing1,2, Chuanyu Wu2, Michael J. Adams2

1.Mining Technology Institute, Taiyuan University of Technology, Taiyuan, PR China
2.School of Chemical Engineering, University of Birmingham, Birmingham, UK
The 3D splash process induced by impacting of one particle was

simulated by DEM and the distributions of the speed and angle
of ejected particles were analyzed. It was found that the ejection
number is proportional to the total absolute momentum of the
ejected particles and the mean angle between the ejection direc-
tion and horizontal plane is unexpectedly independent of both
impact speed and incident angle and is around 80º, the distribu-
tion of ejection speed is approximately an exponential function.
INTRODUCTION
The impact-ejection process on a particulate packing is critical to understand
the aeolian sand transport. Many experimental observations and numerical simula-
tions [1~4] have indicated that sand grains are ejected from the sand bed by the
impact of saltating grains. Anderson and Haff [1] ever developed a numerical
model for aeolian sand transport on the basis of the impact-ejection scheme. They
presented a splash function by analyzing the simulated results of the ejection
speeds and angles statistically. Wind tunnel experiments [2~4] were also carried
out to investigate the probability distribution of the ejection state and its relation-

ship with the impact state of the saltating particles. However, these researches are
all restricted to 2D analysis. For exploring further realistic characteristics of sand
transport, it is essential to study the impact-ejection process in 3D space.
Recently, Beladjine et al. [5] and Ammi et al. [6] researched the impact-
ejection process for sand transport experimentally with a 3D mono-sized granular
packing impacted by one particle at various speeds and directions. The 3D dy-
namic behavior was revealed by their experiments, but their results are restricted
by using given type of mono-sized particles with given physical property in the
experiments. In order to obtain further insight into the interaction between the par-
ticles and their ejection mechanism and to obtain a general splash function for a
wide range of sand grain size, the 3D numerical simulation with discrete element
method is a powerful approach. This is the main purpose of this paper.
EQUATIONS FOR PARTICLE-PARTICLE INTERACTION

The simulation was performed using the DEM code in University of Birming-
ham. The interactions between contiguous particles were modeled according to
the theories of Hertz for the normal interaction and Mindlin & Deresiewicz for the
tangential interaction, see Thornton & Randall [7], Thornton & Yin [8].
SIMULATION METHODOLOGY
Bed preparation
The simulation region is a rectangular box with the dimension of
0.015×0.015×0.015 m, and a 3D granular bed was formed in its lower part with
50000 mono-sized particles and has the dimension of 0.015×0.015×0.006 m. The
particle is spherical and its size is 300 µm. The mechanical properties of the parti-
cles used are: Young’s modulus E=54 GPa, Poisson’s ratio ε=0.3, and the coeffi-
cient of friction between the particles µ=0.3. One particle with the identical size
and mechanical properties was generated in the upper part of the simulation re-
gion above the bed and was shot onto the bed at given speeds and angles.
Test procedure
The incident particle impacted onto the packing at various speeds in the range
of 3~10 m/s and at various angles in the range of 40°~80°. Representatively, at

each impacting condition, the simulation was repeated for several times at various
positions to obtain more than 500 ejected particles. The ejected particles were
identified only if the particles reach the height greater than one particle diameter
above the bed. The evolution of inter-particle contact force and the velocity of
each particle in the simulation region were calculated. The x, y, z components of
the initial speeds of the splashed particles were recorded. We define two angles
characterizing ejected particles (see Fig 1), which are the ejection angle α measur-
ing the angle between the ejection direction and the horizontal plane, and the azi-
muthal angle β representing the deviation of the ejection direction from the inci-
dent plane. The angle α is restricted to vary between 0° and 90°, while β can vary
between −180° and +180° (β=0 corresponds to the shooting direction).
Fig.1 Schematic diagram of the impact and ejection directions
RESULTS AND DISCUSSIONS

Number and speed of ejected particles
Fig. 2(a) shows that the number of ejected particles is proportional to their total
absolute momentum. Fig. 2(b) shows the mean ejection speed slightly increases
with the increasing impact speed. This means that the increase of the impact en-
ergy mainly motivates more particles to eject but slightly affect their mean ejected
kinetic energy.
Distributions of ejected particles

350 0.5
(b)
mean ejection speed (m/s)

(a)
total absolute momentum
300
0.4
250
(10 kg m s )
-1
0.3
200
150 0.2
-8
100
0.1
50
0 0.0
50 100 150 200 250 300 350 2 3 4 5 6 7 8 9 10 11
ejection number impact speed (m/s)
Fig.2 relationship between ejection number and momentum at the impact speed of
10 m/s (a) and mean ejected speeds as functions of impact speeds (b)
Fig. 3(a) is the mean ejection angle as a function of impact speed and impact an-
gle. It is surprising that the mean ejected angle α is independent of both impact
speed and impact angle, and is a constant around 80º. The ejected angles vacillate
around 80º in a small range. This is coincident with Ammi et al.’s experiment re-
sult [6] except that they reported the ejection angle is around 60ºdue to their par-
ticles with different packing geometry from that in this simulation. The mean
azimuthal angle β is about 0º and the ejection direction distributes symmetri-
cally around the incident plane. It was also found that for different impact speeds
90
85 Vi=10 m/s
Vi= 8 m/s
mean ejection angle
80 0.1 Vi= 6m/s

Vi= 5m/s
75 fitted line
Density
70
65 0.01
60
55
(a) 1E-3 (b)
50
2 3 4 5 6 7 8 9 10 11 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50
impact speed (m/s) ejection speed (m/s)
Fig.3 mean ejection angle as a function of impact speed (a)

and the distribution of ejection speed (b)

and angles, the form of the probability distributions of the ejection speed are all
similar. Fig.3(b) are the probability distributions of ejection speeds which can be
fitted by exponential functions.
CONCLUSIONS
3D DEM simulation about the splash process on a granular bed was per-
formed and the ejection behavior was analyzed statistically. This model is proved
to be valid for the investigation of the impact-ejection process by the published
experimental data. It is believed that the increase of impact speed slightly in-
creases the mean ejection speed but significantly increases the number of ejected
particles which is proportional to their total absolute momentum at a given impact
speed. It is confirmed that the angle between the mean ejection direction and the
horizontal plane is a constant around 80º, the ejection directions distributes sym-
metrically around the incident plane. The probability distribution of ejection
speeds can be fitted by an exponential function.
ACKNOWLEDGMENTS
This work was financially supported by the National Natural Science Founda-
tion of China (Grant No. 50706031) and the Natural Science Foundation of Shan-
xi province in China (Grant No. 2008021005)
REFERENCES
1. R. S. Anderson, P.K. Haff, ‘Simulation of eolian saltation’, Science, 241: 820-
823,1988
2. Z. B. Dong , X. P. Liu, F. Li, H. T. Wang, A. G. Zhao, ‘Impact-entrainment re-
lationship in a saltating cloud’. Earth Surface P rocesses Landforms. 27:641–
58, 2002
3. M. Rice,B. B. Willetts, I. K. McEwen, ‘An experimental study of multiple
grain-size ejecta produced by collisions of saltating grain with a flat bed’. Se-
dimentology. 42, 695-706, 1995
4. B. T. Werner, P. K. Haff, ‘The impact process in eolian saltation: two dimen-
sional studies’, Sedimentology, 35:189-196, 1988

5. D. Beladjine, M. Ammi, L. Oger, A. Valance, ‘Collision process between an

incident bead and a three-dimensional granular packing’, Physical Re view E ,
75(061305):1-12, 2007
6. M. Ammi, L. Oger, D. Beladjine, A. Valance, ‘Three-dimensional analysis of
the collision process of a bead on a granular packing’. Physical Review E ,
79(021305):1-9, 2009
7. C. Thornton, W. Randall, ‘Applications of theoretical contact mechanics to sol-
id particle system simulation’. In M. Satake, J. T. Jenkins (Eds.), Microme-
chanics of granular materials, pp. 133–142, 1988
8. C. Thornton, & K. K. Yin, ‘Impact of elastic spheres with and without adhe-
sion’, Powder Technology, 65(1–3), 113–123, 1991

FAST AND STABLE SIMULATION OF GRANULAR

MATTER AND MACHINES
Claude Lacoursière1, Martin Servin1, and Anders Backman2

1. Umeå University. 2. Algoryx Simulation.
Fast and stable simulation of granular matter and machines is achieved

through a unified model for particles, fluids and rigid bodies based on
constraint multibody systems and using dynamic resolution.
INTRODUCTION
We present mathematical models and numerical techniques designed to address
the challenges of real-time simulator training for earth moving equipment involv-
ing heavy vehicles and granular matter. The same techniques have also been ap-
plied to mix granular matter with a nearly incompressible SPH model of water.
The central elements in our technique are a unified model based on constrained
multibody systems including point particles and rigid bodies, a dynamic resolu-
tion technique based on merging and splitting elementary bodies. The numerical
integration is based on a time-discrete variational formulation of analytic mechan-
ics, which is closely related to the Rattle and Shake [3] solvers. We have modified
these by introducing stabilized linear approximations to avoid solving the non-
linear equations exactly [5]. Frictional contact forces are modeled using a linear
complementarity (LCP) formulations similar to those already well-known in this
field [8]. Our deviation from these models is mainly in the solution technique,
which mixes a direct LCP solver for computing normal forces and a Gauss-Seidel
(GS) process to solve for the frictional ones. The SPH model includes kinematic

constraints for incompressibility and for the boundary conditions. The latter are
formulated as non-penetration conditions producing also buoyancy for immersed
bodies. Another novelty of our overall method is the introduction of dynamic
resolution using a merge-split technique based on local analysis of contact forces
and complementarity conditions. A significant result is the real-time simulation of
a tractor, which can shovel granular matter consisting of moderately small ele-
ments integrated in an interactive 3D application used for operator training. The
conventional computational techniques for granular materials are based either on
modeling the system as distinct particles [7] or using continuum mechanics with
specific constitutive laws for granular matter and discretization using finite ele-
ments or mesh-free methods. In turn, the discrete element techniques are phrased
either as penalty methods, e.g., the Hertz contact model, requiring very small time
steps. Complementarity formulations of contact models solved using Gauss-Seidel
iterations [9] can use larger time steps, but have slow, linear convergence. By con-
trast, interactive applications put strong constraints on speed and time steps. There
are usually roughly 10 milliseconds to compute a 1/60 second update. With this
budget, there is usually only time for a single step with size h = 1/60 s. In turn the
stable simulation of machines demands precise solutions to maintain stability.
One additional reason for the choice of our integration method is that it is solidly
anchored in discrete time mechanics and is in fact much more stable than higher
order explicit methods. It is also necessary to explicitly compute the Lagrange
multipliers with good precision, using a direct method for the most part, or with a
good preconditioner. Our experience has shown that neither explicit penalty
methods nor GS based solver can be tuned to provide sufficient stability.
THEORY AND METHODS

Our starting point is the descriptor form of multibody dynamics. We write M
for the system mass matrix, G for the constraint Jacobians, x for the generalized
position variables v for the velocities, and for the Lagrange multipliers. The
constraint indicators are g(x) and these can be either equality or inequality condi-
tions.

After discretization and linearization, our time stepping method requires the
solution of the following Mixed Linear Complementarity Problem (MLCP)
Hz + b = w+ w
G T v
0 z l w+ 0 , with H = M and z = . (1)
G
0 w z w 0
Vectors l and u are lower and upper bounds, and vector b contains stabilization
terms to avoid numerical drift away from the in constraint surface. The slack vari-
ables w+, w- are discarded once the solution is computed. The inequalities and or-
thogonality should be understood component wise. is a non-negative diagonal
matrix which protects against constraint degeneracy and introduces a small
amount of elasticity. We typically use perturbations which are eight orders of
magnitudes less than the masses. Details of this model are presented elsewhere,
along with a description of our direct solver which is based on a non-smooth
Newton method [5,4]. It is still impractical to solve system (1) directly for very
large matrices. Reasonable approximations of Coulomb frictional forces also
prove impractical to solve directly and therefore, we have resolved to a split itera-
tive method in which normal forces are computed directly and frictional ones are
processed using a GS process. For SPH fluids, we have used both GS and precon-
ditioned Conjugate Gradient solvers with success [1] to produce very near incom-
pressibility as well as mixed simulation involving boundary conditions and buoy-
ancy.
The H matrix for coupled system has the form of nested dissection shown in Fig.
1 where the blocks, from left to right, “fluid”, “rocks” and “vehicle” each have the
matrix form shown in Eqn. (1). For the latter, examples of the sparsity pattern of
the H matrix are shown in Fig. 1 for each of the subsystems (20.000 fluid parti-
cles, 30 rocks and 12 machine parts). The “contact” blocks involve pairwise con-
straints couplings between elements of the subsystems and consist of relatively
few equations. The fluid matrix in Fig. 1 scales linearly with the number of parti-
cles. This is because each particle brings in a fixed number of element per row,
around 25 for the SPH kernel we use. For the rigid rock model, each body brings a
number of contacts (at worst 20) related to the shape of the geometry. A good di-

rect solver such as SuperLU [6] scales linearly for such problems. Note that both
the rock pile and the loader are overdetermined.

Fig 1. Schematic illustration of the system matrix H and its sparsity patterns.
The dynamic resolution algorithm is illustrated in Fig. 2. Contacting bodies are
merged into rigid super-bodies, the filled geometries, if the relative velocity at the
contacts are below a given threshold and contact forces are all sufficiently posi-
tive. A merge event is marked in the figure with a dashed line. Conversely, any
geometry exposed to an impacting, sliding or separating contact is split from the
super-body it belongs to. Split and merge events are propagated using GS itera-
tions ordered as a breadth first traversal starting at the transition points. Splits and
merge operations are recursive so that meta bodies can be agglomerated and split.
Elementary bodies are atomic and cannot be fractured. The merge-split operations
preserve the important invariants, namely, total mass, inertia, momentum, angular
momentum. The energy is preserved up to the accuracy implicitly set by the
thresholds.
Fig 2. Illustration of the dynamic resolution technique.
THEORY AND METHODS

Still frames from our earth moving training simulator are shown in Fig. 3. On a
laptop with 64 bit 3.03 GHz processor and 4 Gb RAM a single simulation with a
pile of 500 rocks and a tractor made of 12 bodies and 15 joints, the computational

time was between 2-3.5 s per 1 s of simulation time in single thread mode. The
variation depends on the amount of interaction between the vehicle and the pile.
The time step is 1/60 0.017 s, and the rocks are represented as composite
spheres with mean size of 0.6 m, and realistic mass density. The numerical solver
takes roughly 75% of the time and collision detection dominates the remaining
time. Using dynamic resolution, the computational time per simulated second is
reduced to 0.1 s when the pile is at rest, 0.8 when running the load bucket into the
pile and interacting with 150 bodies and 1.6, or when driving over the rocks with
the vehicle. For these examples, the split-merge procedures take up to 5% of the
computational time while it improves the speed by 100-2000%, depending on the
amount of interaction.
Fig 3. Image from prototype system, final wheel loader simulator and samples
from simulation involving a mix of rigid particles and SPH fluid.
The implementation has room for further optimization and acceleration. It is pos-
sible for instance to use graphical processing units (GPGPU) to perform some of
the iterative computations, part of work in progress. A detailed analysis of
GPGPU for these systems and different combinations of direct and iterative
solvers will be presented elsewhere. A two-dimensional example of coupling be-
tween our SPH based nearly incompressible fluid and immersed pebbles is shown
in Fig. 3. For a 3D configuration, not shown here, our GS based implementation
on the laptop described above can handle 1000 particles in real-time [1]. Prelimi-
nary results show roughly a factor 100 speedup when applying a parallel conju-
gate gradient solver on GPGPU, i.e., 100.000 particles in realtime. The particular
2D example in Fig. 3 has 222 rigid particles, 861 non penetration contacts and
1235 fluid elements was run with time step 10 ms and with computational time
100 ms per time step, still on the same laptop. With iterative solvers and paralleli-
zation on streaming processors the computational speeds for constraint based

methods are steadily improving. For instance, Tasora and Negrut [9] report 1000
bodies at 0.43 s per 0.01 s time step in the serial version and a speed-up by a fac-
tor seven to 0.06 s per time step when running on GPGPU.
ACKNOWLEDGMENTS
This research was supported by HPC2N, SSF grant A3 02:128, UMIT Lab and
ProcessIT. Contributions were also made by N. Melin at Algoryx and K. Bodin.
REFERENCES
1. K. Bodin, C. Lacoursière, and M. Servin, Constraint fluids, submitted to IEEE
Transactions of Visualization and Computer Graphics, (2010).
2. R. W. Cottle, J.-S. Pang, and R. E. Stone, The Linear Complementarity Prob-
lem, Computer Science and Scientific Computing, Ac. Press, New York, 1992.
3. E. Hairer, C. Lubich, and G. Wanner, Geometric Numerical Integration, vol.
31 of Spring Series in Computational Mathematics, Springer-Verlag, 2001.
4. C. Lacoursière, Ghosts and Machines: Regularized Variational Methods for
Interactive Simulations of Multibodies with Dry Frictional Contacts, PhD the-
sis, Dep. of Computing Science, Umeå University, Sweden, June 2007.
5. C. Lacoursière, Regularized, stabilized, variational methods for multibodies,
in The 48th Scandinavian Conference on Simulation and Modeling, P. Bunus
and C. Führer, eds., Linköping University Electronic Press, 2007, pp. 40–48.
6. X. S. Li, An overview of superlu: Algorithms, implementation, and user inter-
face, ACM Trans. Math. Softw., 31 (2005), pp. 302–325.
7. T. Pöschel and T. Schwager, Computational Granular Dynamics, Models and
Algorithms, Springer-Verlag, 2005.
8. A. Tasora and M. Anitescu, A convex complementarity approach for simulat-
ing large granular flows, Journal of Comp. and Nonlinear Dynamics, (2010).
9. A. Tasora and D. Negrut, Parallel algorithm for solving complex multibody
problems with streaming processors, Int. J. Comp. Vis. and Biomech., (2009).

Evolution of Structure in Granular Materials

John F. Peters, David A. Horner, Laura E. Walizer, Raju Kala
US Army Engineer Research and Development Center, Vicksburg, MS
The discrete element method provides a valuable research tool for

understanding the evolving internal state of granular media as it is
described by phenomenological models. By comparing individual
particle motions to those projected from continuous affine motion
a diffusive law emerges, providing a means to relate macro-scale
continuum motion to micro-scale motion and allowing constitu-
tive equations to be build up from micromechanical laws. The
laws thus described are those for a local continuum assumption
although it is noted that the scheme can be extended to existing
non-local formalism. Recent results from DEM simulations sug-
gest that significant non-affine, but non-diffusive motions can also
arise, possibly as precursors to localization.
Introduction
Since its introduction by Cundal and Strack [1], the discrete element method has
evolved as a tool for fundamental investigations in granular mechanics by provid-
ing information on particle motion and contact forces at a level of detail not possible
in physical tests. This same period has seen a concomitant rise in micromechani-
cal studies of granular media, DEM being the principal tool of investigation. In

micromechanical models, the onerous computations required by a particle compu-

tation are replaced by a continuum idealization built around constitutive equations
that are created from an understanding of the micromechanics. A difficulty in any
attempt to devise a constitutive equation based solely on micromechanical princi-
ples is relating the evolution of the internal state with the continuum deformation
(e.g. [8]). In an investigation from nearly a decade ago Horner [3] found an evo-
lution law based on modeling the change is of contact statistics with a diffusion
law. The approach showed promise for developing micromechanical constitutive
laws but at the time was preliminary and therefore not published in the open liter-
ature. This paper presents Horner’s approach along with recent DEM simulations
that illustrate some non-diffusive motions that can be captured by the DEM.
Analysis
The analysis begins with the definition of average stress σ ij within a representative
volume V at a point x ∈ V
1 Z
σ ij (x) = σij (x − x0 )dV 0 . (1)
V V
For a granular media, the integration is taken over the solid particles. The average
stress within each particle is given by
Np
1 X
σ pij = fic xcj , (2)
Vp cp
where the superscript p is the particle number, Vp particle volume, and Np the num-
ber of contacts for the particle. The average stress within the representative volume
therefore can be computed as a volume weighted average of the particle stresses,
which implies a double summation. The averaging can be reduced to a single sum-
mation be simply expanding the sums and pairing the terms associated with each
common contact. The result is
Nc
1 X
σ ij = f cX c. (3)
V c i j

The sum is now taken over all of the contacts in the volume, Nc . The vector Xjc
connects centers of particles sharing the contact. Note that although the averaging
process removes all spatial content from the data, the summation still references
specific contacts. To create a completely statistical interpretation the contacts are
grouped into bins consisting of sectors having equal angles, ∆θ (in two dimensions).
b
Within each bin there are pb (θ)Nc contacts directed in the average direction X i (θ)
b
subjected to an average contact force f i (θ). Defining Nc as the number of contacts
per volume, the average stress becomes
X b b
σ ij = N c pb f i X j . (4)
b
Inasmuch as the size of the bins can be made arbitraily small, the statistical descrip-
tion is identical to the direct summation. However, to the extent that the statistical
parameters vary with the continuum deformation in a regular manner, their evolu-
tion can be described through phenomenological relationships that are tied to mi-
cromechanical measurements. To this end we note that the rate of change of average
stress is given by
X ˙b b
σ̇ ij = N˙ c
X b b X b b
pb f i X j + N c ṗb f i X j + N c pb f i X j . (5)
b b b
b
Therefore the evolution laws of interest are for N c , pb , and f i .
A common procedure is to connect the continuum motion to that of the par-
ticle centers through an affine mapping given, in the case of small strain and in
the absence of rotation by ∆vic = xcj ˙ij . The particle centers do not, in fact, fol-
low an affine mapping, as Horner found when observing how pb should vary if
particle motions were affine. The comparisons of contact directions expected from
affine motion and those actually observed indicated that particle centers follow paths
that have strong random components. The contact forces likewise show the effects
of random motion of the contacts leading Horner to speculate that an evolution
equation for the contacts might follow the form of an advection-diffusion equation,
whereby the advection modeled the change in contact force driven by the affine
motion and the non-affine component followed a diffusion law
2 b
b b ∂ f
f˙ i = f˙ i |affine + D 2i (6)
∂θ

The constitutive equations that follow from combining Equation (4) with (6) fit the
averaged particle data well [3], although it should be noted that the constitutive
equations do adhere to the local continuum assumption. It is now commonly ac-
cepted that constitutive equations for granular media should have a non-local char-
acter to avoid numerical pathologies in solutions [6] and to better capture size ef-
fects. The diffusive nature of particle motions is well documented (e.g. [2]) along
with theoretical treatments of its dissipative nature (e.g. [5], [7]). These works
clearly point to spatial diffusion leading to non-local laws in which boundary con-
ditions must be prescribed for the diffusive variables.
However, not all non-affine motions need to be diffusive; for as shown by Vala-
nis [7], patterned deformations can arise from instabilities in the sub-scale con-
stitutive behavior. An example of such behavior can be observed in simulations
of biaxial tests by subtracting the averaged affine motion from the individual par-
ticle motions, leaving a residual non-affine motion (e.g. [4]). Examples of three
cases of boundary conditions are shown in Figure 1. Smooth platens, whereby the
particle-platen contact friction is zero, creates complex rotational fields. In the case
of friction on all platens, a symmetric pattern of four rotation zones emerge. By
making a single platen frictional, rotational patterns with mirror symmetry emerge.
For the two cases with platen friction, the symmetric rotational patterns gave way
to localized shear banding.
A number of constitutive theories have been developed to explain the complex
behavior of granular materials. Because these theories involve “internal” variables,
not measured in continuum experiments, it is difficult to build an empirical foun-
dation for their application. Virtual experiments using the discrete element method
allows the observation of motion and forces needed to create that foundation.
Acknowledgements
Funding for this research was provided by Military Engineering Basic Research
Project Fundamental Investigation on Bridging Between Microscale Phenomena
and Macroscale Performance. Any use of trade names is for descriptive purposes
only and does not imply endorsement by the U.S. Government. Permission to pub-
lish was granted by Director, Geotechnical and Structures Laboratory, US Army

Figure 1: Simulations of biaxial test with various platen friction configurations. Platens
motions were specified such that constant volume was maintained. Arrow directions indicate
particle motions after removing affine motion imposed by platens; color indicates degree of
particle rotations. Symmetric localized shear banding eventually emerged in the two cases
with platen friction.

Engineer Research and Development Center. Approved for public release; distribu-
tion is unlimited.
References
[1] Cundall, P. A. and Strack, O.D.L. (1979) “A Discrete Numerical Model for
Granular Assemblies,” Géotechnique, Vol. 29, No. 1, pp. 48–64.
[2] Bridgewater. J. (1993), “Mixing and Segregation Mechanisms in particle

Flow,” in Granular Matter, SpringerVerlag, Anita Mehta, Editor, pp. 161–
193.
[3] Horner, D. A. (2000) Application of DEM to Micro-Mechanical Theory

for large Deformations of Granular Media, PhD dissertation, University of
Michigan at Ann Arbor (see also ERDC technical report ERDC/GL TR-00-
7).
[4] Kuhn, M. (1999) “Structured Deformation in Granular Materials,” Mechan-

ics of Materials, Vol. 31, pp 407–429.
[5] Valanis, K. C., Peters, J. F. and Gill, J. (1993). “Configurational Entropy,

Non-Associativity and Uniqueness in Granular Media”, Acta Mechanica, Vol
100, pp. 79-93.
[6] Valanis, K. C. and Peters, J. F. (1996). “Ill-Posedness of the Initial and

Boundary Value Problems in Non-Associative Plasticity,” Acta Mechanica,
Vol. 114, pp. 1-25.
[7] Valanis, K. C. (1998) “A Gradient Thermodynamic Theory of Self-

Organization,” Acta Mechanica, Vol. 127, pp. 1-23.
[8] Walsh, S., Tordesillas, A., and Peters, J. F. (2007) “Development of Mi-
cromechanical Models for Granular Media: The Projection Problem,” Gran-
ular Matter, Vol. 9, pp 337–352.

DEM SIMULATION OF BALL INDENTATION ON

COHESIVE POWDERS
Massih Pasha, Ali Hassanpour, Mojtaba Ghadiri

Institute of Particle Science and Engineering, SPEME, University of Leeds, Leeds LS2 9JT, UK
An approach based on indentation, which can be carried out on very

small quantities, is considered for assessing the flowability of cohe-
sive powders. In this method the yield stress of weak powder com-
pacts is estimated by indentation hardness measurements. To be able
to measure the plastic yield stress by this method, the ratio between
the flow stress, inferred by indentation (i.e. hardness) and the yield
stress needs to be established. In the present work an attempt is made
to simulate the indentation process and uniaxial compression test us-
ing Distinct Element Method (DEM) to investigate the correlation be-
tween indentation characteristics and powder flowability.
INTRODUCTION
Industrial processes involving powder blending, transfer, storage, feeding,
compaction and fluidisation all require reliable powder flow [1]. There are a num-
ber of cases which involve dealing with small amounts of loosely compacted
powders. These include filling and dosing of small quantities of powders in cap-
sules and dispersion for dry powder inhalers and dry particle sizing. In other cas-
es, the availability of powders is an issue. For instance in nuclear and pharmaceu-
tical industries, the amount of testing powder is limited due to ionising radiation
for the former and toxicity and cost of drugs for the latter [2].

There exist a number of test methods for evaluation of flow behaviour of pow-
ders such as uniaxial compression test, shear test, raining bed and the Sevilla
powder tester [3]. Most of these test methods require a relatively large amount of
powder and measuring the flow behaviour at relatively high compaction stresses.
Hassanpour and Ghadiri [2] introduced a testing method by ball indentation which
can be performed on small amounts of loosely compacted powders. In the present
paper an attempt is made to evaluate the flowability measurement of cohesive
powders using ball indentation method. The indentation process and unconfined
compression test are simulated using the Distinct Element Method (DEM) for par-
ticles consolidated to different levels. The correlation between indentation charac-
teristics and flow behaviour of powders (unconfined yield stress) are investigated
by comparing indentation results with those of unconfined compression.
INDENTATION PROCEDURE
In the indentation process, different samples of powders are consolidated into a
cylindrical die to the pressure in the range that forms weak tablets. The die must
be made of low friction materials in order to reduce the effects of wall friction.
The weakly formed tablets are then indented using a spherical indenter and the
depth/load cycle is recorded from which the hardness of the bulk powder can be
calculated. During loading, the load is increased at a specified rate until a desired
maximum load is reached. Then the load is decreased to zero at the same rate.
During unloading, the elastic deformation of the sample will recover. Hardness is
given as the ratio of maximum indentation load to projected area of the impres-
sion (Equation 1).
Fmax
H (1)
A
The projected area can be calculated as follow,
A    dh  h2  (2)
where d is the diameter of the indent, and h is the depth of impression [2].

In indentation test, during formation of the local plasticity zones around the in-
denter, the volume of the powder bed present in a yield condition is surrounded by
an elastically deformable region. This leads to an increase in the local yield
strength (i.e. hardness) [5]. A linear relationship is usually considered:
H  CY (3)
where H is the hardness, Y is the yield stress and the proportionality factor C is
known as constraint factor. It is important to relate hardness to yield stress, since
the flow behaviour is defined based on the yield strength. Wang et al. [4] has con-
cluded that indentation hardness and unconfined yield stress have a linear rela-
tionship with pre-consolidation pressure for a number of materials. This corrobo-
rates the linear relationship between yield stress and hardness. The constraint fac-
tor for a number of testing powders was also determined, and it was concluded
that it is independent of the pre-consolidation pressure but is material dependant.
For particle assemblies, the constraint factor would depend on single particle
properties which need to be analysed by the Distinct Element Method (DEM)
simulations [4].
DEM SIMULATION OF INDENTATION PROCESS

The EDEM code from DEM Solutions, using the Hertz-Mindlin contact model,
is used for simulations of the indentation process. A linear cohesion model is used
to model cohesion force between particles which is calculated as follow:
F  kA (4)
where A is the contact area and k is a cohesion energy density with units Jm-3. The
material properties used in the simulations are summarised in Table 1.

Property Particles Die (Geometry)

Diameter (mm) 1 45
Density (kg.m-3) 2500 7800
Poisson’s Ratio 0.25 0.3
Young’s Modulus (GPa) 75 182
Interface Energy (J.m-2) 3.46 -
Table 1. Material properties used in DEM simulations
70,000 particles with normal size distribution (standard deviation = 0.1384) are
generated inside the cylindrical die. This has provided a bed height of 45 mm. The
powder bed is then consolidated to different pressure levels. The consolidated bed
is indented using a spherical indenter with a diameter of 15mm. Figure 2 shows
the simulated hardness values for a range of indentation load (Fmax).
16
14
12
Hardness (kPa)
10
0
0.6 0.8 1 1.2 1.4 1.6
Indentation Load (N)
Fig 1. The relationship between hardness and maximum indentation load for 5kPa
pre-consolidation
It can be seen that hardness does not change significantly with indentation load.
This shows that the powder bed is not consolidated during indentation process. If
the bed is consolidated, the hardness value will be different and it will not be rep-
resentative of the pre-consolidation of interest. The correlation between hardness
and pre-consolidation stress is illustrated in Figures 2.

15
14
Hardness (kPa)
13
12
11
10
3 4 5 6 7 8 9 10 11
Pre-consolidation Pressure (kPa)
Fig 2. The relationship between hardness and pre-consolidation pressure for in-
dentation load of 0.8 N
It can be seen that hardness increases linearly with pre-consolidation pressure.
The uniaxial compression test is also simulated using the same cylindrical die and
material properties. The unconfined yield stress results as a function of pre-
consolidation are shown in Figure 3.
1.37
1.36
Unconfined Yield Stress (kPa)
1.35
1.34
1.33
1.32
1.31
1.3
1.29
1.28
1.27
3 4 5 6 7 8 9 10 11
Pre-consolidation (kPa)
Fig 3. The relationship between unconfined yield stress and pre-consolidation

pressure

It can be seen that there is a linear correlation between unconfined yield stress and
pre-consolidation, similar to the case of indentation.
CONCLUSION
DEM simulations of ball indentation for a range of indentation loads and pre-
consolidation pressures showed that hardness is independent of the indentation
load but linearly increases with pre-consolidation pressure. Uniaxial compression
test was also simulated for a range of pre-consolidation pressures. The results
showed that unconfined yield stress also increases linearly with pre-consolidation
pressure. It can be concluded that there is a correlation between hardness and un-
confined yield stress.
ACKNOWLEDGMENTS
This work described here has been financially supported by the Engineering
and Physical Sciences Research Council (EPSRC).
REFERENCES
1. Prescott J.K., and R.A. Barnum. Pharmaceutical Technology, 2000. 24(10): p.
60.
2. Hassanpour A., and M. Ghadiri, Particle & Particle Systems Characterization,
2007. 24(2): p. 117.
3. Castellanos A., J.M. Valverde, M.A.S. Quintanilla, Kona, 2004, 22: p. 66.
4. Wang C., A. Hassanpour, and M. Ghadiri, Particuology, 2008. 6(4): p. 282.
5. Kozlov G.V., V.D. Serdyuk, and V.A. Beloshenko, Mechanics of Composite
Materials, 1995. 30(5): p. 506.

STATISTICS OF INTERNAL STRUCTURE OF A

GRANULAR PILE GENERATED BY THE
PARTICLE EXPANSION METHOD
Chi-Yan Lo1, Malcolm Bolton1, Yi-Pik Cheng2

1
Department of Engineering, University of Cambridge, UK
2
Department of Civil and Environmental Engineering, University College London, UK
We present a systematic investigation of the influence of a common

sample preparation method (Particle Expansion Method) has on the
mechanical and structural properties of granular packings. The method
produces distinctively different attributes from natural sedimentary
samples, in terms of both the contact force distributions at small forces
and the geometric internal ordering. The artificial fabric induced by
this technique may pose serious implications in the accurate modelling
of granular matter which are often overlooked by DEM users.
INTRODUCTION
Unlike ordinary solids and liquids, contact forces inside a static granular pack-
ing are organised into highly inhomogeneous force chains with stresses localised
along the force network. Their statistics and distribution are particularly important
as it determines the internal structure and bulk properties of the material. Previous
studies based on experiments, molecular dynamics, Monte Carlo simulations and
thermodynamics analogies [1-2] have revealed a common form of bimodal char-
acter of the internal contact network: for normal forces Fn greater than the mean
value <Fn> of the system the spatial force distribution P(Fn) decays exponentially
while at weak force networks (Fn << <Fn>) a power law function is portrayed.

This paper is concerned with the investigation of the influence of sample

preparation method on the internal fabric of the numerical packing. A common
preparation strategy in DEM, known as the particle expansion method [3], is
tested to justify its suitability for realistic sample generation. The force distribu-
tion and contact anisotropy of the resulting sample are studied and compared with
the theoretical statistical model. We also examine the dependency of contact net-
work on the system size.
DEM SIMULATIONS
We perform 3D DEM simulations of the problem using the software PFC3D.
The software uses explicit time-stepping scheme in which contacts between parti-
cles are soft and Hookean and a Coulomb frictional sliding model applies at tan-
gential directions. For quasi-static simulations, a local mechanical damping func-
tion is used to eliminate possible accelerative motion. A more comprehensive de-
scription of the code can be found at [3].
r
Normal and shear stiffness 4 × 106 N/m
of particles
Normal and shear stiffness 4 × 109 N/m
of column
a = h/r Radius of particles 1.0 mm
h
Friction coefficient, μp (par- 0.577, 0.0
ticles) and μw (walls)
Porosity 0.4
Fig 1. Geometry. Table 1. Modelling parameters.
To prepare the sample using the particle expansion method, a population of

particles with initially reduced radii is randomly created in a cylinder. Particles are

then gradually expanded to their final radii value until a target porosity is reached
within the sample. Substantial cycling follows to dissipate locked-in stresses until
an equilibrium packing is achieved. Compared to other preparation methods, the
particle expansion method enables modellers to prepare their DEM samples
within reasonable computational times, but its representation is susceptible as the
packing methodology is fundamental different from naturally sedimentary granu-
lar materials found in nature.
Cylindrical samples of varying aspect ratio a (ranging from 3 to 9) are gener-
ated as in Fig 1, with the micromechanical parameters listed in Table 1. Particles
are packed in the cylindrical tube which rests on a horizontal ground. With r set at
10 mm, samples of different a are created by varying h. The black lines in Fig 1
indicate the normal contact force chains in a sample. Note that in all our simula-
tions gravity is activated such that vertical stresses increase with depth. This is
evident in Fig 1 in which the contact force chains get more heavily loaded at
deeper sections.
FORCE DISTRIBUTION
Fig 2. PDF of normalised normal Fig 3. PDF of normalised tangen-

forces. tial forces.

Fig 2 and 3 show the semilogarithmic plots of the probability distribution func-
tion (PDF) of normalised normal and shear contact forces in the packings. We
find that the contact forces exhibit an exponential decay at large forces, irrespec-
tive of the aspect ratio. This is consistent with the findings in [1-2, 4], but a power
law relationship is absent for contact forces of magnitude below the mean in our
case. In fact the exponential function seems to be applicable to the small force re-
gion and a unimodal distribution character is portrayed for samples created by the
particle expansion method.
Fig 4. PDF of stress ratios. Fig 5. Side and elevation views

of spatial distributions of contact
centres with η ~ 0 (top) and η
~0.577 (bottom) in the column.
A closer examination to the probability distribution P(η), where η = Fs/Fn, re-

veals the heterogeneity of the stress ratio at the contacts in the sample. Fig 4
shows the histogram of η and their distribution collapses into a single curve irre-
spective of the sample size. The two extremes η ~ 0 and 0.577 evolve from the
boundary effects and the frictional characteristics of the system respectively. This
can be proved by plotting the spatial distribution of the contact points correspond-

ing toη ~ 0 and 0.577 in the column. As shown in Fig 5, the contacts with nearly
zero η are mostly located along the cylindrical wall surface. This is because the
wall is frictionless and the contact forces are principally perpendicular to the cyl-
inder surface. On the other hand, mobilised contacts (i.e. those with η at the Cou-
lomb failure threshold) occur in relatively unloaded regions and are sparsely dis-
tributed. This is in agreement with the usual findings that sliding tended to occur
at contacts that carried small forces [5].
For intermediate values of η ∈ [0.05, 0.55], the probability distribution P(η) is
a non-uniform function and varies from 2% to 8%. This is likely to be attributed
to the preparation method and history of the sample. For a naturally pulviated
sample, Radjai and co-workers [1] discovered that P(η) displays a uniform distri-
bution for all η except at the extremes. This may imply that samples created from
the particle expansion method are geometrically different and possess an internal
structure dissimilar to natural earth materials.
CONTACT ORIENTATION
Fig 6. PDF for the contact normal ori- Fig 7. Rose diagram for the pro-
entation from the vertical. The contact jected contact normals orientation
angles φ and θ are defined in the inset. from the positive x-axis.

The internal fabric of the samples can be probed by looking at the probability
distributions of the contact orientations. Fig 6 and 7 display the statistics of the
angles between the contact normals and vertical (φ) as well as the angles between
the projected contact normals on the x-y plane and the positive x-axis (θ). Clearly
the granular structure is highly anisotropic, with P(φ) peaks at 60º (Hexagonal
close-packed orientation for a monodisperse sample) and 90º (mostly from parti-
cle-wall contacts which are principally horizontal). Also as shown in the rose dia-
gram for P(θ), the particle expansion method induces no bias in the distributions
of θ. This is close to the findings from a pulviated DEM sample reported in [4].
CONCLUSION
This paper examines the statistical micro-structure of monodisperse samples
prepared by the particle expansion method. Despite being a computationally effi-
cient algorithm for sample generation, we show that samples prepared in this way
possess distinctive granular structure from sedimentary specimens: 1) absence of
power law distributions for contact forces below average; 2) non-uniformity in
contact stress ratios η. DEM users must therefore consider seriously the likely ef-
fects from sample preparation techniques to fit their simulation purposes and ob-
jectives.
REFERENCES
1. F. Radjai, M. Jean, J-J. Moreau, S. Roux, ‘Force distribution in dense two-
dimensional granular systems’, Phys. Rev Lett., 77, No.2, 274-77, 1996.
2. D. M. Mueth, H. M. Jaeger, ‘Force distribution in a granular medium’, Physi-
cal Review E, 57, No.3, 3164-69, 1998.
3. PFC3D Version 3.1 User Manual, Itasca Consulting Group, 2005.
4. G. Marketos, M. D. Bolton, ‘Quantifying the extent of crushing in granular
materials: A probability-based predictive method’, J. Mech. Phys. Solids, 55,
No.10, 2142-56, 2007.
5. C. Thornton, S. J. Antony, ‘Quasi-static deformation of particulate media’,
Phil. Trans. R. Soc. Lond. A, 356, No. 1747, 2763-2782, 1998.

SIMULATING GRANULAR MATERIAL

BEEHAIVOUR TO GENERAL LOADING PATHS
WITH DEM
Xia Li, Hai-Sui Yu

Nottingham Centre for Geomechanics, University of Nottingham
A numerical technique is developed to simulate the elementary stress-

strain behavior of granular material to general loading paths under qu-
asi-static conditions. The technique has three features: (1) the repre-
sentative element has a polyhedral (polygonal) shape to minimize
boundary arching effect; (2) the finite strain definition is adopted for
accurate boundary control; (3) the stress controlled loading, the strain
controlled loading and the mixed loading can be easily implemented.
INTRODUCTION
Multi-scale investigation on the stress and strain relations of granular materials
is an important application of discrete element method in soil mechanics (Cundall
and Strack, 1979). As the current research interest has been shifted to general
loading paths (Ishihara and Towhata, 1983). A numerical technique of simulating
the elementary behavior of granular materials to general loading paths is needed.
STRESS AND STRAIN FOR GRANULAR MATERIALS

To serve the purpose of multi-scale investigation, expressions of the macro-
scale stress and strain tensors in terms of the micro-scale forces and displacements
are derived (Li et al., 2009). On the micro-scale, the boundary tractions are esti-

mated from the boundary forces exerted at discrete boundary points, i.e.,
v∫ xi p j dS = ∑ xi Fj , Hence, the stress tensor could be expressed as:
β β
S β ∈S
1⎛ ⎞
σ ij = − ⎜ ∑ xiβ Fjβ + ∑ X iP G Pj ⎟ . (1)
V ⎝ β ∈S P∈V ⎠
in which Fi β is the boundary force exerted at discrete boundary points β with the
coordinates of xiβ ; GiP is the volumetric force of granular particle P acting at the
center of gravity X iP . This expression is used to calculate the stress tensor from
the boundary forces.
To give complete information on specimen deformation and for precise boun-
dary control over a large range of strain levels, finite strain definition is adopted
for the simulations, rather than infinitesimal strain definition. The Biot strain ten-
sor ε is used to characterize the specimen strain state for its simplicity. The Biot
strain ε is given in the indicial notation as:
ε IJ = δ IJ − U IJ = δ IJ − RkJ −1FkI = δ IJ − δ Jk FkI . (2)
For two dimensional spaces, the expression of strain tensor could be written as:
φJK
ε IJ = −
V
∑X
ΔL∈B
K Δu I (3)
⎛ 0 1⎞
in which φIJ = ⎜ ⎟ is the 2D permutation tensor, X K stands for the centre of
⎝ −1 0 ⎠
the boundary segment and ΔuI represents the relative displacement between the
two vertices of the boundary segment.
APPLYING GENERAL LOADING PATHS

The geometry of each boundary wall is described by its center coordinates x O
and its normal direction n O . Stain boundary condition is controlled by specifying
translational and rotational velocities of boundary walls in each loading step. The

movements of the boundary walls are controlled synchronously. To achieve a

strain increment Δε IJ , we have:
δ xi = FiJ δ X J = δ iK (δ JK − ε JK ) δ X J (4)
Fixing the origin O of the specimen, to apply a strain increment Δε IJ within a

time interval Δt , the velocity assigned to the wall centre should be:
viw = Δxiw / Δt = δ iK Δε JK X Jw / Δt (5)
in which X Iw is the coordinate of wall center in the undeformed configuration.

And the rotational velocity of the wall segment w should be:
1 δ iK Δε JK N Jw
ω ≈−
w
(6)
Δt δ iK (δ JK − ε JK ) N Jw
i
where N Iw represents the wall normal direction in the undeformed configuration.

The stress-boundary is applied using a servo-controlling mechanism. By anal-
ogy with an isotropic elastic constitutive relationship, a strain increment can be
determined according to:
1 +ν ν
( Δε ) I = ( Δσ )i − Δσ ii
E E (7)
(α Δε ) I = (α Δσ ) I
where E and ν are Young’s modulus and Poisson’s ratio. In the new calculation
cycle, the stress increment Δσ ij = σ ijt − σ ij is updated, and a new strain increment
is applied according to the above procedure. By repeating doing so, the specimen
stress state gradually approaches the target stress state.
TYPICAL NUMERICAL SIMULATION RESULTS

Numerical simulations have been carried out in analogy to the conventional
testing procedure in soil laboratory. Typical simulation results are presented in
this section. The specimen preparation is shown in Fig. 1.

(a) Boundary walls (b) Disk generation (c) Clump generation(d) Prepared sample
Fig. 1 Specimen preparation
The numerical simulation result of biaxial shearing is given in Fig. 2. The ob-
served material behavior is in good qualitative agreement with laboratory results
on Toyota sand (Verdugo and Ishihara, 1996).
1200
1200
1050 e0=0.240
Deviatoric Stress, q (kPa)
1000
Deviatoric Stress, q (kPa)
e0=0.226
900
e0=0.182
800 750
600 600
450
400
e0=0.240 300
e0=0.226
200 150
e0=0.182
0 0
0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 22.5 25.0 0.16 0.18 0.20 0.22 0.24
Deviatoric Strain, εq (%) Void Ratio, e
Fig 2. Numerical simulation result of biaxial shearing

For rotational shearing, the mean normal stress p and the deviatoric stress ra-
tio η are constant, while the principal stress direction nσ keeps changing. The
simulation results of a dense specimen sheared in vertical direction up to η = 0.7 ,
and then subjected to stress rotation are presented here. The stress path is plotted
in Fig. 3. And the observed material behavior is shown in Fig. 3. Significant vo-
lume contraction due to stress rotation has been observed as reported in literature

(Yang et al., 2007). It is also interesting to observe that as stress rotation contin-
ues, the strain trajectory in the deviatoric plane approaches a circle.
(a) Variation of stress component; (b) Stress trajectory.

Fig. 3. Stress components in stress rotation.
(a) Strain trajectory; (b) Variation of strain component.

Fig. 4. Strain components in stress rotation.

CONCLUSION
This paper presents a numerical technique to carry out simulations of the ele-
mentary behavior of granular materials using discrete element method. The devel-
oped numerical technique is suitable to apply general loading path so as to inves-
tigate the general constitutive relationship at multi-scales. Here, we use polyhe-
dral/polygonal specimens to minimize the influence of boundary arching effect.
The finite strain definition is adopted for accurate boundary control. A servo-
mechanism of stress boundary condition is developed and could be easily com-
bined with strain boundary condition to apply mixed loading conditions. The qua-
litative agreement between the numerical simulation results and the laboratory
observations validate the developed numerical technique as useful tools to carry
out numerical simulations with DEM for exploring the fundamental particle-scale
mechanism and for the development of the constitutive relations of granular mate-
rials.
ACKNOWLEDGMENTS
The financial support from the UK’s Engineering and Physical Science Re-
search Council (EPSRC) and the University of Nottingham is gratefully acknowl-
edged.
REFERENCES
CUNDALL, P. A. & STRACK, O. D. L. A discrete numerical model for granular
assemblies. Geotechnique, 29, 47-65, 1979.
ISHIHARA, K. & TOWHATA, I. Sand response to cyclic rotation of principal
stress directions as induced by wave load. Soils and Foundations, 23, 11-
26, 1983.
LI, X., YU, H. S. & LI, X. S. Macro-micro relations in granular mechanics.
International Journal of Solids and Structures, 46, 4331-4341, 2009.
VERDUGO, R. & ISHIHARA, K. The steady state of sandy soils. Soils and
Foundations, 36, 81-91, 1996.
YANG, Z. X., LI, X. S. & YANG, J. Undrained anisotropy and rotational shear in
granular soil. Geotechnique, 57, 371-384, 2007.

ENHANCED VIBRATIONAL GRANULAR MIXING
J. E. Hilton and P. W. Cleary

CSIRO Mathematics, Informatics and Statistics, Clayton South, Victoria, Australia
The mixing of granular material is an important process in many in-

dustrial applications. A container of granular material can be mixed by
vibration of the base of the container, while the walls are held fixed.
We show that the rate of mixing in this system can be greatly en-
hanced by the inclusion of a vertical plate, offset from a wall by a dis-
tance, d. The inclusion of this offset plate breaks the symmetrical
granular convection of the system causing enhanced mixing. We vary
this spacing length and show that although the system has a complex
dependency on the length of spacing, the mixing rate is always en-
hanced compared to a system without a vertical offset plate.
INTRODUCTION
The mixing of different types of granular material plays an important part in
many industrial applications such as chemical engineering, agriculture and phar-
maceuticals 1 . Many different means of mixing have been designed and employed
for different processes. Here, we investigate the mixing rate of a vibrational mixer
consisting of a container of material with a base plate subject to periodic vibra-
tion. The behaviour of a granular system subject to a vertical sinusoidal oscillation
has been well studied 2,3 . It is known that convection cells form in the granular bed
with particles moving down the walls, towards the centre of the bed up through
the central axis of the bed, along the upper layer of the bed and down the walls,
completing the cycle. This process can be use to mix granular material, where the

mixing rate shows a strong dependence on the vibrational frequency 4 . We show a

means to greatly enhance the mixing rate is by breaking the symmetry of the sys-
tem by the inclusion of an offset vertical plate within the mixing container.
Offset Plate
y
A0 sin (2πft) d A0 sin (2πft)
x Plate Separation
z
Fig 1. Schematic set up of vibrational mixer without offset plate (left) and with
offset plate (right).
METHODOLOGY
We use the Discrete Element Method (DEM) for simulating our system, which
is a Lagrangian particle scheme first formulated by Cundall and Strack 5 . The par-
ticle-particle contact force is determined by particle overlap information using a
soft-sphere linear spring, dashpot and slider approximation.
Our system consists of a square container as shown in Fig. 1. The boundaries
are frictional side walls in the x directions, an open top in the positive y-direction,
a vibrating base in the negative y-direction and periodic boundary conditions on
the particles in the z-directions. The distance between the side walls in the x-
direction is set to be 5 cm, the width of the container in the z-direction is set to be
2 cm and the container is filled to a height of 5 cm with particles of 2 mm diame-
ter ± 0.1 mm to prevent crystallization effects. The particles are marked as one of
two types split equally in the vertical direction, as shown in Fig. 1 with light col-
oured particles on the upper half and dark coloured particles on the lower half.
The friction between the particles and the container walls is set to μ = 0.5, the

normal and tangential coefficients of restitution are set to 0.5 and the spring con-
stant is set to 10,000, sufficient to give a maximum overlap of less than 1% of the
particle diameter. The particle density is set as 1000 kg/m3.The base plate is vi-
brated with a vertical offset of the form A = A0sin(2πft), where the vibration fre-
quency, f, is fixed at 400Hz for all simulations, with an amplitude, A0, of 0.1mm.
The distance of the offset plate, d, is varied and the mixing rate measured. The
mix state is determined by superimposing a grid over the simulation and ascertain-
ing the mixing fraction M within each grid cell, where M = 0 represents an un-
mixed state (the cell is only filled with particles of one type) and M = 1 represents
a fully mixed state (the cell is filled with an equal distribution of particles of both
types). The calculations for the mixing rate are given in Cleary and Sinnott1.
RESULTS
The motion of the particles in a system with no offset plate is shown in Fig. 2
(left). The upper layer of the bed (light coloured particles) rapidly moves down
the walls of the container until it reaches the vibrating plate (at ~ t = 10 s). This
vertical layer of particles then begins to curl under the particles initially at the
base of the bed (dark coloured particles) causing these particles to form a vertical
central column which is ‘pinched off’ by the light coloured particles (at t ~ 20 s).
This motion also causes the central portion of the bed to rise causing a hump at
the surface. The dark particles reach the surface of the bed at ~ t = 30 s, where-
upon they begin to roll down the hump of the surface (at ~ t = 40 s) and begin to
convect down the walls of the container (at ~ t = 50 s). When the dark coloured
particles reach the vibrating base of the container they follow the same behaviour
as the lighter coloured particles, rolling inwards towards the centre of the bed (at ~
t = 60 s onwards). The stretching and folding of the particle layers is the mecha-
nism by which the mixing of the particles takes place in the system. However, the
equal downward convection of particles causes a static region along the central
vertical axis of the bed. The lighter coloured particles ‘pinch’ the darker particles
but never meet in the centre; a thin strand of darker material can be seen running
along the central vertical axis in the final frame shown at t = 80 s. This suggests
that better mixing would be achieved by breaking the symmetry of the system –
which is achieved by the introduction of a vertical offset plate.

t = 0s t = 10s t = 20s t = 0s t = 10s t = 20s
t = 30s t = 40s t = 50s t = 30s t = 40s t = 50s
t = 60s t = 70s t = 80s t = 60s t = 70s t = 80s
Fig 2. Particles mixing in a container with no offset plate (left) and with an off-
set plate at d = 1 cm (right) from time 0 to 80 s. The approximate granular flow is
marked with arrows.
The motion of the particles with a vertical offset plate included in the system is
shown in Fig. 2 (right). The dynamics of the particles when this plate is included
is very different to the dynamics with no plate. The upper layer of particles moves
rapidly down the side walls and also down each side of the offset plate. However,
the particles between the offset plate and the nearest wall (on the left hand side of
the assembly) have a greater net downward motion that between the offset plate
and the furthest wall (on the right hand side of the assembly). This causes a
greater downward convective motion of particles which pushes the lighter col-
oured particles far underneath the layer of darker coloured particles. Comparison
of the systems with and without the offset plate at ~ t = 10 s shows that the darker
coloured particles have already reached the surface of the bed by this time when
the plate is included. The bed surface also forms a hump, but this is offset to the
right hand side for this case. By ~ t = 20 s the darker coloured particles have
rolled down the left hand side of the assembly and along the vibrating base, caus-
ing a second fold in the granular layers to form. A third fold can just be seen at ~ t
= 30 s, but becomes difficult to make out by ~ t = 40 s. This stretching and folding

of the granular material continues until the system is, to the eye, well mixed from
t = 50 s onwards. Comparison of the systems with and without the offset plate
clearly shows a great increase in the rate of mixing when the offset plate included.
This is due to the broken symmetry giving different convective contributions on
the two sides of the system.
100
90
80 No Plate
4 mm
70
6 mm
Mixing %
60 8 mm
10 mm
50
12 mm
40 14 mm
16 mm
30 18 mm
20 20 mm
10
0
0 20 40 60 80 100 120 140
Time (s)
3.5
2
Mixing Constant, α ×10
3.0
2.5
2.0
1.5
1.0
0.5
0.0
0 5 10 15 20
Plate Distance from Wall, d (mm)
Fig 3. Mixing percentage for the system against time (top) and parameter α
from an exponential fit to the mixing data (bottom) for a range of offset plate
separations, d.
To quantify the rate of mixing the system the mixing percentage (M × 100) is
plotted against time in Fig. 3 (top) for a range of offset plate separations. The in-
clusion of a plate always increases the mixing rate, but the increase does not ap-
pear to be a simple function of the plate separation. If the mixing fraction, M, is

assumed to of the form M = (1 – e-αt) + β, the parameters α and β can be deter-

mined from a fit to the data. The parameter α gives the rate of mixing but there is
no clear trend to the data pattern, as shown in Fig. 3 (bottom). This sensitivity of
mixing to a small change in system parameter is a hallmark of chaotically mixed
fluid systems. However, using optimal plate positioning an increase in the mixing
rate of ~600% can be achieved.
CONCLUSION
The mixing rate of a vibrational granular mixer can be greatly enhanced by the
addition of a vertical plate offset from a container side wall by a offset length d.
The mechanism for the mixing is a non-symmetrical stretching and folding of the
layers of the granular material due to increased convection on the side of the as-
sembly with the offset vertical plate. Although the mixing rate will also depend on
particle size, friction, vibrational amplitude and frequency, this mixing mecha-
nism is not expected to change. The separation of the plate from the side wall does
not have a simple relation to the overall mixing rate, although this rate always in-
creases when compared to an assembly without an offset plate. Our results are of
interest from both a theoretical aspect, as well as for possible application in indus-
trial granular mixing.
REFERENCES
1
Cleary, P. W. and Sinnott, M. D. (2008), Particuology, 6, p419
2
Knight, J. B., Ehrichs, E. E., Kuperman, V. Y., Flint, J. K. , Jaeger, H. M and
Nagel, S. R., Phys. Rev. E. (1995), 54, p5726
3
Zeilstra, C. Collignon, J. G., van der Hoef M. A, Deen, N. G., Kuipers, J. A. M.
(2008), Powder Tech., 184, p166
4
Li-Shin, L. and Shu-San, H. (2007), Powder Tech., 184, p31
5
Cundall, P. A. and Strack, O. D. L. (1979), Geotechnique, 29, p47

DISCRETE ELEMENT MODELLING OF

CYLINDRICAL CAVITY EXPANSION OF
GRANULAR MATERIALS
Yan Geng, Hai-Sui Yu, Glenn McDowell

The University of Nottingham, Department of Civil Engineering
Cavity expansion in soil or rock is a fundamental problem in theoreti-

cal geomechanics primarily because it provides a useful and simple
tool for modeling many complex geotechnical problems. This paper
describes the simulation of cylindrical cavity expansion of granular
materials to investigate their mechanical response under monotonic
loading using the discrete element method (DEM). The influence of
the micro-properties of the granular material on the macroscopic
stress-strain behaviour during cavity expansion observed in the DEM
simulation is described. These simulations show that the particle prop-
erties have a great effect on the soil behaviour.
INTRODUCTION
Granular media such as sand composed of discrete particles exhibit very com-
plex macroscopic mechanical responses to externally applied loading. The dis-
crete nature makes the constitutive relationship very complex. Interpretation of
tests on real granular media, such as sand, is difficult because the stresses inside
the sample can not be measured and must be estimated from the boundary condi-
tions. The Discrete element method (DEM) is an alternative view of granular ma-
terials to investigate their mechanical behaviour at the microscopic level. The dis-
crete element method was first developed by Cundall (1971) for rock mechanics

and is being increasingly used to simulate the mechanical behaviour of granular

materials (Cunall&Strack, 1979; Ting et al., 1989; Rotherburg&Bathurst, 1992).
DEM provides the possibility to investigate the mechanical behaviour of granular
materials at both micro and macro levels. Some researchers developed a numeri-
cal quasi-static model for granular media based on displacement methods of struc-
tural mechanics and finite element method (FEM). However, the method was not
able to follow large displacement and rearrangements which are fundamental
characteristics of state changing processes of granular media. DEM is better in
modelling a discontinuous material than other numerical tools such as FEM and
BEM (boundary element method). It considers the particle-scale interaction, al-
lowing the micro-mechanics of soil response to be analysed in detail. In DEM, the
interaction of the discrete particles is monitored contact by contact and the motion
of the particles modeled particle by particle.
Cavity expansion analysis focuses on the cavity pressure needed to expand a
cavity in soil by a certain amount. The problem of expansion of cylindrical cavi-
ties has been dealt with by many investigators in the last 50 years, in connection
with its potential applications to a number of practical problems in mechanics and
engineering. For the granular material, the discrete element method which is
based on discrete properties is an alternative tool for studying the cavity expan-
sion. This numerical analysis technique starts with basic constitutive laws at inter-
particle contacts and the macroscopic response of particle assemblage under load-
ing can be provided. Many researchers focused on this problem. The boundary
condition problems like bearing cavity of granular ground have been studied by
Ting et al (1989). Huang and Ma (1994) were the first to use DEM in deep pene-
tration simulations in sand. In their paper, the DEM-BEM simulation technique is
described. The granular soil mass near the penetration was simulated by the dis-
crete element method (DEM). However, the soil mass was simulated using the
boundary element method (BEM) in the far field, where the strain is expected to
be small. The cone penetration mechanism can be evaluated without the complica-
tion caused by boundary effects using this method. Later, the study on soil-
penetrometer interface friction has been described by Jiang, Yu and Harris (2006).
A two-dimensional discrete element method has been used to simulate the deep
penetration tests on a granular ground. DEM has been developed to investigate the
particle breakage and rearrangement around piles by Lobo-Guerrero and Vallejo

(2005). The production of particles of different sizes can be caused by particle

crushing. It can induce stress relaxation and a lower resistance to pile penetration.
In this paper, DEM is used to simulate cylindrical cavity expansion in granular
material. Several cylindrical cavity expansion simulations were conducted with
different initial porosities. The code PFC3D (particle flow code in 3 dimensions) is
used.
MODELLING PROCEDURE
The DEM code PFC3D was used to simulate cylindrical cavity expansion on
samples consisting of spheres (Figure 1). A quarter of fan-shaped sample was
generated and loaded by its internal boundary using a constant strain rate. The ve-
locity of its external boundary was automatically controlled by a servo-
mechanism to keep a constant external pressure (Figure 2).
Stresses on the walls are determined from the particles contact forces acting on
them. The minimum particle radius for each sample was 0.075m and maximum
radius was 0.1m. The particle friction coefficient is 0.7 and the normal stiffness
and shear stiffness of particles are 1×109N/m. Six walls were generated as sample
boundaries. The wall friction coefficient is 0 and wall normal stiffness and shear
stiffness are 1×109 N/m. Around 7000 Particles were generated to fill into this
space for each sample. The samples were generated by the particle radius expan-
sion method. After the sample is generated, the sample stress state is adjusted to
Pi=Pe=P0. P0 is the initial cavity pressure. Pi and Pe are defined as the internal and
external pressures. The loading was carried out by moving the internal wall at a
rate of 0.01m/s after the sample reached the initial state. The sample thickness was
fixed by front and back walls to simulate the plane-strain condition. Two other
walls normal to x and y axes are fixed since it is an axial symmetry of the x and y
axes structure. Four samples are generated of different initial porosities. The ini-
tial porosities are calculated by measurement sphere which is a function available
in PFC3D. Using the measurement sphere, the porosity of each sphere can be ob-
tained. The spheres are generated many enough to cover the whole sample. The
finial porosity is decided by using the average value of these porosities. In order
to obtain the effect of particle fricition coefficient, there are four different particle
fricition coefficients are used for the sample of initial porosity=0.36.

Pe
b-a
x
a
Pi
Fig 1. A sample of spherical particles Fig 2. Schematic illustration

PFC3D can be used to simulate the soil behaviour under different conditions.
The parameters of the particles hold an important place in DEM simulation. The
selection of particle parameters has significant influences on the sample behav-
iours. The effect of particle friction coefficient is given in Figure 3 (a). Figure 3
(b) shows the results of four simulations carried out from initial porosity, n, of
0.36, 0.38, 0.40 and 0.42, in terms of P/P0 against a/a0. This pressure-expansion
relationship plotting presents sample initial porosities effect on the cavity expan-
sion curve. The initial cavity pressure for these samples is 20MPa and b0/a0 is
circa 6. The various initial porosities of these samples can result from the different
particle friction coefficients during the generation of each sample. Once the initial
state of the sample is reached, the particle friction coefficient is changed to the
same value 0.7. For the sample of initial porosity=0.36, four different particle fric-
tion coefficients (0.1, 0.3, 0.5, 0.7) are used. When the sample of initial poros-
ity=0.36 is obtained, change the particle friction coefficient during each loading.
So every time the same sample is used before loading. The results in Figure 3(a)
show that the friction coefficient between the particles has a large effect on the
peak of cavity pressure. The maximum cavity pressures increase when the particle

friction coefficient increases. It can be seen from Figure 3(b) that the peak cavity
pressure ratio (P/P0) varied from 3.1 for the densest sample to 2.2 for the loosest.
The corresponding range of internal radius (a/a0) at peak pressure was 1.13 (dens-
est) to 1.28 (loosest). The final cavity pressure ratio, at 1.7 internal radius, was
about 2 in all cases. Meanwhile, it is found that the maximum cavity pressures
increase significantly while the value of initial porosities decreases.
(a) (b)
Fig 3. The cavity expansion curves
CONCLUSIONS
In this paper, a numerical analysis of cylindrical cavity expansion in a granular
material is investigated with the aid of DEM. While the analytical approach works
at a macroscopic level, the numerical method has the advantages of providing the
mechanical behaviour of the granular assembly at both the macro- and micro-
scales. The cylindrical cavity expansion stress-strain curves of cohesionless
granular material under monotonic loading are studied.

The key contribution is that it has been shown that the cavity expansion of
granular materials can be modelled using DEM, providing micromechanical in-
sight into the behaviour.
Particle parameters have a considerable effect on the simulated behaviour. The
maximum cavity pressures increase when the friction coefficient of particles in-
creases. Another procedure which permits the generation of clumps of spheres
will be developed to obtain the particle shape effect on the soil behaviour during
cavity expansion.
DEM is a useful method to simulate cylindrical cavity expansion in a granular
material using spherical or non-spherical particles. These simulation results need
to compare with the experimental results.
REFERENCES
1. P. A. Cundall, ‘A computer model for simulating progressive large scale move-
ments in blocky rock systems’, Proceedings of the Symposium of the Interna-
tional Society of Rock Mechanics, Nancy, France, Vol.1, Paper No. П-8, 1971.
2. P. A. Cundall and O. D. L. Strack, ‘A discrete numerical model for granular
assemblies’, Geotechnique, 29, 47-65, 1979.
3. A. B. Huang and M. Y. Ma, ‘An analytical study of cone penetration tests in
granular material’, Canadian Geotechnical Journal, 31, 91-103, 1994.
4. M. J. Jiang, H.S. Yu and D. Harris, ‘Discrete element modeling of deep pene-
tration in granular soils’, Int. J. Numer. Anal. Meth. Geomech., 30, 335-361,
2006.
5. S. Lobo-Guerrero and L. E. Vallejo, ‘DEM analysis of crushing around driven
piles in granular materials’, Geotechnique, Vol.55, 8, 617-623, 2005.
6. L. Rothenburg and R. J. Bathurst, ‘Micromechanical features of granular as-
semblies with planar elliptical particles’, Geotechnique, 42, 79-95, 1992.
7. J.M. Ting, B.T. Corkum, C. R. Kauffman and C. Greco, ‘Discrete numerical
model for soil mechanics’, Journal of Geotechnical Engineering, ASCE, 115,
379-398, 1989.

EFFECT OF GRADATION ON THE CONSTANT

VOLUME CYCLIC BEHAVIOR OF GRANULAR
MEDIA
Abbas Soroush, Behrooz Ferdowsi*

Amirkabir University of Technology, Department of Civil and Environmental Engineering
Tehran, Iran
*
[email protected]
Gradations of granular assemblies have notable effects on their ma-

croscopic behavior under different loading conditions. This paper stu-
dies the contribution of the gradations of two assemblies of particulate
media on their macroscopic strength under cyclic constant-volume
loading condition. It also discusses what happens to the distribution of
forces of particles during the loading cycles as an important characte-
ristic of their micro-scale behavior. Discrete Element Modeling
(DEM) have been used for simulation of the assemblies. Results of the
study show an agreement between the micro- and macro-scale beha-
vior of the two assemblies.
INTRODUCTION
Interparticle contact friction, packing density and polydispersity are characte-
rized as the major contributors to the macroscopic strength of granular assemblies
[1]. This research focuses on the effect of gradation on the macro- and micro-
scale behaviors of granular assemblies under cyclic Constant-Volume (C.V.) con-
ditions which represents the undrained behavior of granular soils. Previous stu-
dies have shown that the degree of polydispersity in densely packed assemblies

has negligible effect on the macroscopic strength, but in less densely packed spe-
cimens, a greater degree of polydispersity would result in higher macroscopic
strength [1]. O’Sullivan & Cui (2009) showed that non-uniform specimen has
higher values of particles movement and thus energy dissipation during load re-
versals, however they observed an increase in stress ratio (q/p) in uniform speci-
men compared to the non-uniform specimen [2].
In this paper, results of monotonic and cyclic C.V. loading of two granular as-
semblies are presented. Thereafter, the distribution of forces which is believed to
have an important link to the observed macro-scale behavior are presented and
analyzed. The micro-scale behavior shows signs of higher amount of localization
in the non-uniform specimen by progression of the loading cycles resulting in
sooner loss of strength, while the monotonic strength of non-uniform specimen is
higher.
DISCRETE ELEMENT ANALYSIS

In this study, a series of C.V. monotonic and cyclic triaxial tests on uniform
and non-uniform assemblies of spherical particles have been simulated by the
Discrete Element Method (DEM) [3]. The uniform specimen is consisted of N =
4000 particles having radius of r = 0.25 mm, and the non-uniform specimen in-
cludes equal number of particles (N=1500) with 0.2, 0.25 and 0.3 mm radiuses
(totally 4500 particles). After generating the specimens by randomly distributing
the particles in the box, a strain controlled loading was applied to prepare samples
with the same isotropic confining pressures of 55 kPa. Monotonic and sinusoidal
cyclic strain-controlled loadings were applied to the specimens. The interaction of
the spheres was modeled using the Hertz contact law.
MACRO-SCALE BEHAVIOR
Figure 1-a shows the stress-strain behavior of the uniform and non-uniform as-
semblies. It is obvious from the figures that the non-uniform specimen has
slightly higher deviatoric stress. This is consistent with the observations of [1] that
the polydispersity increases the undrained strength of medium dense to loose gra-
nular media.

The variation of degradation index and induced pore water pressure with the
number of cycles for the uniform and non-uniform assemblies are shown in Fig.
1-b. The degradation index (δ) introduced by [4] quantifies the degradation of the
shear strength and deformation modulus during cyclic loading. For a cyclic strain-
controlled triaxial test with a given cyclic strain ( c ), δ is defined as:
GSN  dN /  c  dN
   (1)
GS 1  d 1 /  c  d1
where GSN and GS 1 are the secant deformation modulus at cycles 1 and N, respec-
tively and  dN and  d 1 are, respectively, the maximum cyclic deviatoric stresses at
these cycles. A lower value of δ indicates a higher amount of degradation. Fig. 1-b
shows that the non-uniform specimen loses strength faster than the uniform spe-
cimen by progression of the cycles, especially during the 4th to 6th loading cycles.
30
q, Deviatoric stress, (kPa)
non-uniform
uniform
Uniform
Non-uniform
20
(b) (a)
non-uniform
10
uniform
0
0 5 10 15 20 25 30
Strain (%)
Figure 1. (a) Stress-strain behavior of granular assemblies under monotonic load-

ing. (b) Variation of degradation index and pore water pressure vs. loading cycles.
STUDY OF MICRO-MECHANICAL PARAMETERS

Probability distribution of normal forces during the first six loading cycles in
the uniform and non-uniform specimens are illustrated in Fig. 2. The distribution
of normal forces of the uniform specimen has a Gaussian distribution in the first
cycle (specially at low force values). In the subsequent cycles, the force distribu-
tion gradually tends to be more alike exponential distribution. Similar behavior
can be observed for the variation of force distribution plots in the non-uniform

specimen. A Gaussian distribution of forces can be observed up to the third cycle

and then the force distribution gradually changes to exponential (decreasing) dis-
tribution. It can also be derived from the force distribution plots that the forces
lower than the average contact force increases more rapidly in the non-uniform
specimen by continuation of the loading cycles. Previous studies have revealed
that the Gaussian type of distribution is mostly common for homogenous force
chain structures (homogenous force distribution) which usually happens in iso-
tropic confined specimens [5]. Transition from Gaussian to exponential distribu-
tion is indicative of an increase in the anisotropy of force distribution and locali-
zation. Makse et al. (1998) introduced the participation number for quantification
of the localization in force distribution of the granular media as follows [5]:
1
 M 
   M  qi2  (2)
 i 1 
M
in which, M is the number of contacts between the spheres, qi  f i /  f j (3)
j 1
where fi is the magnitude of total force at every contact. From Equation (2),
  1 indicates a limiting state with a spatially homogeneous force distribution.
On the other hand, in the limit of complete localization, we would have   0 as
M   . The values of  have been computed for the initial state and the next
six loading cycles in both of the uniform and non-uniform specimens and are pre-
sented in Fig. 3. The participation number decreases from 0.748 and 0.71 at initial
isotropic compression state in uniform and non-uniform specimens respectively to
0.55 and 0.49 at the moment of maximum applied strain of 6th cycle, showing a
considerable increase in the localization and anisotropy of the forces distribution
in the assemblies. The decrease of participation number confirms the gradual tran-
sition from Gaussian to exponential distribution of normal forces while the rate of
decrease is slightly more in the non-uniform specimen from the 4th to 6th cycles.
This indication to the increase of the anisotropy of force distribution confirms the
formation of a very disperse contact force network in the 6th cycle in both of the
assemblies (not shown here). Moreover, this complies with the gradual decrease
of the deviatoric stress of the specimens.

Uniform
specimen
Non-Uniform
specimen
Figure 2. Distribution of normal forces for uniform and non-uniform assem-

blies. Fn expressed as Fn/<Fn> and Ft/<Fn>, respectively for six first shear cycles.
<Fn> represents the average of normal forces of all contacts.

0 .7 5
Participation number
0 .7 u n ifo rm
n o n -u n ifo rm
0 .6 5
0 .6
0 .5 5
0 .5
0 .4 5
0 1 2 3 4 5 6
N (N u m be r of cycle s)
Figure 3. Variation of participation number at initial state (isotropic compression,

N=0) and six subsequent loading cycles.
CONCLUSION
Results of monotonic and cyclic C.V. loading of two uniform and non-uniform
granular assemblies were presented in this paper. Although the monotonic C.V.
strength of the non-uniform specimen is slightly more, the micro-scale behavior
shows signs of higher amount of localization in the non-uniform specimen by
progression of the loading cycles resulting in sooner loss of strength in the cyclic
C.V. loading of the non-uniform assembly compared to the uniform specimen.
REFERENCES
1. M. Muthuswamy, A. Tordesillas, ‘How do interparticle contact friction, pack-
ing density and degree of polydispersity affect force propagation in particulate
assemblies?’, Journal of Statistical Mechanics – Theory and Experiment, 06
No. 9, P09003, 2006.
2. C. O’Sullivan, L. Cui, ‘Micromechanics of granular material response during
load reversals : Combined DEM and experimental study’, Powder Technology,
193, 289-302, 2009.
3. P.A. Cundall, O.D.L. Strack, ‘A discrete numerical model for granular assem-
blies’, Geotechnique, 29, No. 1, 47-65, 1979.
4. I.M. Idriss, R.D. Singh, R. Dobry, ‘Nonlinear behaviour of soft clays during
cyclic loading’, Journal of the Geotechnical Engineering Division, 104, No.
12, 1427-47, 1978.
5. H.A. Makse, D.L. Johnson, L.M. Schwartz, ‘Packing of compressible granular
materials’, Physical Review Letters, 84, No. 18, 4160-63, 1998.

APPLYING DEM TO UNDERSTANDING JAMMING

IN SYSTEMS OF NON-SPHERICAL GRAINS
Gary W. Delaney, Paul W. Cleary and Anish A. Bhuta
CSIRO Mathematics, Informatics & Statistics, Private Bag 33, Clayton
South, Vic, 3168, Australia
We investigate the properties of jammed disordered packings of

three dimensional non-spherical grains using the Discrete Ele-
ment Method. We quantify how various aspects of the grain’s
shape, including its symmetry and surface curvature, affect the
properties of the resulting packing when the system jams. We
consider both monodisperse packings and mixtures composed of
particles with different shapes and sizes.
INTRODUCTION
Granular materials are all around us, from naturally occurring sands, soils and
stones, to the aggregates used in concrete. They are important in a diverse range
of industries, including food production, mining and pharmaceuticals. Understand-
ing the structural properties of granular packings composed of differently shaped
particles is a key fundamental aspect of the study of granular materials [8, 4, 5].
It also has many important applications, from simple matters of designing efficient
shapes for transportation, to understanding the complex structural properties and the
inter-grain force network in real granular assemblies [1]. It also has potential appli-
cations in areas of material design and in the production of high strength concretes
[6].
In this work we investigate granular packings composed of a wide range of
particle shapes [3]. We quantify how particular aspects of the particle’s shape, in-

cluding its surface curvature and symmetry, affects the properties of the packing
when the system jams. We also consider packings composed of mixtures of grains
with different shapes and sizes.
METHOD
We investigate the packing properties of 3D superellipsoids defined by:

x m y m z m
+ + = 1,
a b c
where we will refer to m as the shape parameter, and a, b and c are the semi-major
axis lengths. We will consider biaxial grains, setting a = b and defining the aspect
ratio α = c/b. For m = 2, we recover the general formula for an ellipsoid, while
larger values of m generate increasingly cubic shapes. For equiaxed grains (α = 1)
we get a progression of ”superballs”, that range from a sphere (m = 2) to a cube
(m = ∞). Other shapes that lie between ellipsoids and cuboids are obtained for
values of m > 2 combined with non-integer values of α (See Fig. 1). Superellip-
soids have previously been utilized in large scale industrial simulations, where they
have highlighted the important role particle shape plays in the behavior of granular
systems [1, 2].
We utilize a fully dynamic linear spring Discrete Element Method (DEM) sim-
ulation to model the interaction between particles [1]. At each iteration we obtain
the contact locations and linear overlaps between particles and hence solve for the
forces and torques on each particle. Further details of our DEM technique are given
in Refs [2, 1]. The particles in our simulations have unit density and are frictionless.
We utilize a large spring constant to ensure very low inter-particle overlaps of less
than 5 × 10−4 d, where d is the major axis length.
Our simulations are seeded at a low packing fraction (Φ < 0.2) in a unit peri-
odic box. We employ a dynamic particle expansion technique, in which the size of
the grains is slowly and uniformly increased at a constant volumetric growth rate γ,
while imposing a condition of constant kinetic energy. The simulation has a well
defined end point, where the average particle overlap diverges sharply as the system
jams. This technique is similar to that defined by the Lubachevsky-Stillinger algo-
rithm for hard particles [7]. The final density and the resulting structural properties

Figure 1: Monodisperse packing of ellipsoids (m = 2, α = 1.25) (left) and superel-

lipsoids (m = 4, α = 0.4) (right).
of the packings is dependant on the volumetric growth rate γ used in the simulation.
Here, we choose a growth rate γ that generates random close packings of ellipsoids
(m = 2) and then apply that same growth rate to generate packings of superellip-
soids.
MONODISPERSE PACKINGS
Monodisperse packings in which all of the constituent particles are identical in

both size and shape are an interesting idealization of real world granular materi-
als. They are an intrinsically interesting problem, with applications in a diverse
range of chemical, biological and crystallographic areas [3].
We have considered monodisperse packings of a broad range of superellipsoids
(Figure 1). Two key properties of the grain shape are found to determine their pack-
ing behavior: their degree of anisotropy (aspect ratio) and their surface curvature.
The large regions of low surface curvature for high m values, readily generate large
amounts of face on face inter-grain contacts and thus large amounts of orientational
ordering. We have quantified the degree of orientational ordering using several
order parameters [3]. These include the traditional S2 parameter quantifying the

Figure 2: Bidisperse packing of ellipsoids (m = 2, α = 0.3) and superballs (m = 5,

α = 1). The packing fraction is plotted against the relative fraction of ellipsoids vs
cuboids in the packing. The densest packing is formed for monodisperse cuboids
(0.82) and the loosest for monodisperse ellipsoids (0.66).
degree of alignment of the grain’s major axes and also an averaged order parameter
Savg which takes account of any symmetry of the grain axes. Further details are
given in Ref [3].
The anisotropy of the grain shape (aspect ratio) has been found to have a very
interesting effect on the packing properties. For low m values (near ellipsoidal
shapes) where generally very disordered packings are formed, increasing anisotropy
is found to increase the density of the resulting packings. Contrarily, for high m
values (cuboidal shapes) where the packings are much more ordered, increasing
anisotropy is found to disrupt that order and consequently reduce the density of the
resulting packings.
MIXTURES OF SUPERELLIPSOIDS
We have considered a broad range of bidisperse superellipsoid packings. These in-

clude simple mixtures of differently shaped equal-volume grains, to mixtures com-
posed of particles with both large size and shape variations (An example of a pack-
ing composed of a mixture of ellipsoids and cuboids in shown in Figure 2). We
can investigate several key properties of such systems, including maximal density,

Ordered
Cuboids 0.9
0.9 0.85
0.8
0.75
Φ 0.8 0.7
0.7 0.65
0.6
0.6 RCP
spheres
5
4
2 mB
3 3
mA 4
5 2
Figure 3: Image on the left is of a bidisperse packing of equal-volume superballs

with m = 2 (spherical) and m = 5 (cuboidal). The packing fraction variation as the
shape parameter m of each of the species is varied is shown on the right.
average coordination number and the degree of ordering of the particles.

In Figure 3 we show results for packings composed of 50:50 mixtures of bidis-
perse equal-volume superballs (a = b = c). The superball’s shape parameter m
controls how cubic its shape is, and is plotted for each species on the horizontal
axes. The packing fraction of the resulting mixture is shown on the vertical axis.
The loosest packing is obtained for monodisperse spheres (mA = mB = 2) and cor-
responds to a random close packed arrangement with Φ = 0.64. All other superball
mixtures are found to pack more densely, with a monotonic increase in the pack-
ing fraction found as mA and mB are increased. The densest packings are obtained
for the highest m values, with the monodisperse packings of highly cubic particles
(mA = mB = 5) found to pack most densely. We have also quantified the orienta-
tional ordering in these packings and found that the increases in packing density are
directly correlated to increases in the ordering within the system.
CONCLUSION
We have examined the structural properties of a broad range of jammed packings

of non-spherical particles. The surface curvature and degree of anisotropy in the

particle shape are found to be key properties in determining the packing behaviour.
We have also considered packings composed of mixtures of grains with different
shapes and sizes. These allow us to understand how different particle shapes interact
with one another in a jammed state, which is highly important in the majority of real
world granular systems [3, 1].
The type of numerical packing model presented in this work also has interest-
ing applications in simulations of industrial systems. One such application that is
currently being pursued is in particle breakage simulations, where a broken particle
needs to replaced with a packing of daughter particles with a realistic shape and size
distribution [1].
References
[1] P. W. Cleary. Large scale industrial DEM modelling. Engineering Computa-

tions, 21(2/3/4):169 – 204, 2004.
[2] P. W. Cleary. DEM modelling of particulate flow in a screw feeder. Progress in
Computational Fluid Dynamics, 7(2-4):128–138, 2007.
[3] G. W. Delaney and P. W. Cleary. The packing properties of superellipsoids.
EPL (Europhysics Letters), 89(3):34002, 2010.
[4] G. W. Delaney, S. Hutzler, and T. Aste. Relation between grain shape and
fractal properties in random apollonian packing with grain rotation. Physical
Review Letters, 101(12):120602–4, 2008.
[5] A. Donev, I. Cisse, D. Sachs, E. A. Variano, F. H. Stillinger, R. Connelly,
S. Torquato, and P. M. Chaikin. Improving the density of jammed disordered
packings using ellipsoids. Science, 303(5660):990–993, Feb. 2004.
[6] H. Herrmann, R. M. Baram, and M. Wackenhut. Polydisperse packings. Brazil-
ian Journal of Physics, 33(3), 2003.
[7] B. D. Lubachevsky and F. H. Stillinger. Geometric properties of random disk
packings. Journal of Statistical Physics, 60(5):561–583, 1990.
[8] S. R. Williams and A. P. Philipse. Random packings of spheres and spherocylin-
ders simulated by mechanical contraction. Physical Review E, 67(5):051301,
May 2003.

COMBINING DEM AND 3D IMAGING OF REAL

GRANULAR SYSTEMS
Gary W. Delaney1 , T. Di Matteo2,3 and Tomaso Aste2,3,4
1 CSIRO Mathematical and Information Sciences, Private Bag 33,
Clayton South, Vic, 3168, Australia;
2 Department of Applied Mathematics, School of Physical Sciences,
The Australian National University, Canberra, Australia;

3 Department of Mathematics, King’s College London, The Strand,
London, WC2R 2LS, UK;

4 School of Physical Sciences, University of Kent, Canterbury, Kent,
CT2 7NH, UK.
Three-dimensional X-ray imaging has recently been applied to the

detailed examination of the internal geometrical structure of gran-
ular systems. We utilize this experimental data as an input into
Discrete Element Method (DEM) simulations. This allows us to
both improve on previous tomographic measurements of the geo-
metric properties of the system and to obtain information on the
dynamic properties of the system. This includes measurements of
the force network, stress paths and stability of the system, which
are otherwise not directly accessible from physical experiments.
INTRODUCTION
Two key tools that are utilized in the study of granular materials are computational
modelling using the Discrete Element Method and 3D tomographic imaging [1, 4].

In this work we combine these two technologies, utilizing real experimental data
obtained from 3D imaging of granular packings as the input for a DEM simulation,
which can then be used to obtain properties of the system that are not available from
physical experiments.
Bernal and Scott performed some of the earliest work attempting to discern the
internal structure of granular systems [2]. They carried out many painstaking exper-
iments using packings of spherical grains. These included pouring paint through the
packing, allowing it to dry, then examining the individual spots caused by contact-
ing spheres. Today, modern 3D imaging techniques can be used to look inside 3D
sphere packings and obtain their 3D geometric structure. Confocal microscopy has
proven useful in examining semi-transparent materials such as glass beads [7], but
is not applicable to the study of more complex non-transparent materials. The most
promising technique applied to date is x-ray computed tomography [6, 9], which
offers the possibility of exactly characterizing the positions, shape and size of each
of the grains in the packing. This has been used to extensively investigate very
large disordered sphere packings [1]. This work has produced the AAS database
of disordered packings comprising the coordinates (with precision better than 1%
of the sphere diameters) of more than three million spheres from 18 samples of
mono-sized acrylic and glass spheres prepared in air and in fluidized beds [8].
DEM SIMULATIONS
The Discrete Element Method (DEM) has been widely used to simulate granular
matter, powders, bulk materials, rock mechanics and many other systems that are of
great scientific and industrial importance [3]. We employ the Hertzian interaction
law and incorporate a viscoelastic dissipation into the collision process, with the
normal force between two contacting spheres given by
3/2 1/2
Fn = kn ξn − γn ξn vn (1)
where kn is the spring stiffness, ξn is the overlap, γn is a dissipation parameter related
to the coefficient of restitution, and vn is the relative normal velocity. The normal
d 1/2 E
spring stiffness is given by kn = 3(1−ν 2 ) , where d is the grain diameter, E is Young’s

Figure 1: Combined tomographic and DEM reconstruction of a sphere packing.

The inner spheres that are part of the DEM simulation are shown in red, with the
fixed boundary spheres rendered translucent. The whole sample is shown (left) and
also the sample with a section removed (right) to show the internal structure of the
packing.
modulus and ν is Poisson’s ratio. We model the tangential or shear interaction

between the spheres by
1/2
Fs = −min(|ks ξn ξs − γs vs |, |µFn |) · sign(vs ) (2)
where ξs is the tangential displacement, µ is the kinematic friction coefficient be-
tween the spheres and vs is the relative tangential velocity.
DEM RELAXATION OF AN XCT SAMPLE
We consider a set of 4 packings of mono-sized acrylic beads packed in a cylindrical

container with an inner diameter of 55 mm and filled to a height of ∼75 mm. These
cover a range of densities φ and correspond to samples B (φ = 0.5971), D (φ =
0.6258), E (φ = 0.6288) and F (φ = 0.6367) in Ref [1]. We use realistic physical
parameters for acrylic beads having: radius r = 0.795 mm, Young’s modulus E =
3.2 GPa, Poisson ratio ν = 0.3, density 1150 kg/m3 , and inter grain static friction
coefficient µ = 0.28.

The experimental packings are all at static mechanical equilibrium. The co-
ordinates of the grain centers obtained from XCT analysis are used as the initial
configuration in our DEM simulation. The spheres located at the sides and the bot-
tom of the sample (within four diameters of the boundaries) are fixed in position and
top of the system is left open to allow for the possibility of small expansions of the
system during the relaxation process. The DEM simulation is then performed on the
N ' 7000 unfixed spheres in the central region of the sample, with the boundaries
provided by the outer fixed spheres (See Figure 1).
In a typical simulation, the spheres re-arrange from the initial positions by mov-
ing only a small fraction of their diameters to settle into the final configuration
where force and torque are balanced and the system rests in mechanical equilib-
rium. The final average height of the grains is within 0.1-0.2% of the initial average
height and the average displacement of the centers of the spheres during the relax-
ation process is less than 5% of the sphere diameters. Large values of the normal
and tangential dissipation parameters γn and γt are used to ensure the relaxation is
purely quasi-static and we have verified that during the relaxation process grains
keep an average number of contacts larger than 4. Essentially, the relaxation dy-
namics consists in a re-equilibration of the elastic stress trapped in the grain over-
laps in the initial configurations which are unphysical and unbalanced and are due
to the experimental uncertainties on the grain positions, grain polydispersity and
a-sphericity.
From the experimental bead pack, we have thus produced a simulated ideal
mono-disperse sphere packing, a ‘virtual packing’, that matches the original geom-
etry of the experimental bead pack and that can now be used to compute with pre-
cision several static and dynamical properties which are otherwise not directly ac-
cessible from experiments (e.g. force network, avalanches precursors, stress paths,
stability, fragility, etc.).
FORCE DISTRIBUTION
Figure 2 shows the distribution of the normal forces at the inter-grain contacts mea-
sured in an internal part of the sample 4 sphere diameters away from the boundaries.
We see that the majority of the forces in the system are carried by only a small frac-
tion of the inter-grain contacts, and thus a large fraction of contacts are carrying

0.25
0
10
Complementary Cumulative Distribution

0.2 −1
10
Relative Frequency
−2
10
0.15
−3
10
0.1
−4
10
0 5 10 15 20
fn / <fn>
0.05
0
0 1 2 3 4 5
fn / <fn>
Figure 2: Distribution of the normal forces at the inter-grain contacts for samples
B (red circles), D (green squares), E (blue diamonds) and F (magenta triangles).
The inset shows the complementary cumulative distribution plotted on a log-linear
scale.
only small forces. The shape of the distributions shows an exponential drop off
and this can be seen clearly in the straight lines present in the inset showing the
complementary distribution plotted on a log-linear scale.
This force information is one of the key additional measurements that we can
obtain from our combined DEM and 3D imagining approach. We have utilized this
to study the in-depth nature of the forces at the individual grain scale, examining
how the local geometric properties (using measures such as the Voronoi volume and
local coordination number) affect the structure of the force network [5].
CONCLUSION AND OUTLOOK
We have outlined the details of a novel approach to combine Discrete Element Sim-
ulations with real experimental data from tomographic imaging. This approach has
the potential to provide the key bridge between the two main methodologies, ex-
perimental and computational, that are being used to explore and better understand
granular matter. By more closely brining together these investigations, we can ob-
tain a more complete and confident picture of the properties of granular matter;
providing the key information necessary to formulate and test new theories.
In future work we will explore the combination of high speed tomography and

other bulk imaging techniques with DEM during dense granular flow and at the
jamming transition. The role of grain shape in granular systems is also of great
interest [4, 5] and an extension of our technique to consider three-dimensional non-
spherical grains will also be pursued.
Acknowledgements - Many thanks to T.J. Senden and M. Saadatfar for their
work in obtaining the tomographic data.
References
[1] T. Aste, M. Saadatfar, and T. J. Senden. Geometrical structure of disordered

sphere packings. Phys. Rev. E., 71:061302, 2005.
[2] J. D. Bernal and J. Mason. Packing of spheres: Co-ordination of randomly
packed spheres. Nature, 188(4754):910–911, December 1960.
[3] P W Cleary. Predicting charge motion, power draw, segregation and wear in
ball mills using discrete element methods. Minerals Engineering, 11(11):1061–
1080, November 1998.
[4] Gary W. Delaney, Stefan Hutzler, and Tomaso Aste. Relation between grain
shape and fractal properties in random apollonian packing with grain rotation.
Physical Review Letters, 101(12):120602–4, 2008.
[5] Gary W. Delaney, T. Di Matteo, and Tomaso Aste. Combining tomographic
imaging and DEM simulations to investigate the structure of experimental
sphere packings. submitted to Soft Matter., 2010.
[6] GW Delaney, S Inagaki, T Di Matteo, and T Aste. Virtual experiments on
complex materials - art. no. 68020G. Complex Systems II, 6802:G8020–G8020,
2008.
[7] M. M. Kohonen, D. Geromichalos, M. Scheel, C. Schier, and S. Herminghaus.
Oncapillary bridges inwet granular materials. Physica A, 339:7–15, 2004.
[8] See. http://wwwrsphysse.anu.edu.au/granularmatter/, 2006-2009.
[9] M. Tsukahara, S. Mitrovic, V. Gajdosik, G. Margaritondo, L. Pournin, M. Ra-
maioli, D. Sage, Y. Hwu, M. Unser, and Th. M. Liebling. Coupled tomog-
raphy and distinct-element-method approach to exploring the granular media
microstructure in a jamming hourglass. Physical Review E (Statistical, Nonlin-
ear, and Soft Matter Physics), 77(6):061306–11, June 2008.

DISCRETE ELEMENT MODELLING OF

COMPACTION BEHAVIOUR OF AGGLOMERATED
AND AGGREGATED NANOPOWDER
Avinash Balakrishnan, Patrick Pizette and Christophe L. Martin

Laboratoire SIMaP – GPM2, Grenoble-INP, BP46, 38042 Saint Martin d’Hères cedex, France.
Email: [email protected]
In the present work, Discrete Element Method (DEM) is used to study

the effect of particle size in aggregated and agglomerated ceramic
powders during compaction. Agglomerates and aggregates are mod-
elled as porous systems consisting of 5000 monosized particles held
by adhesive forces and strongly bonded by solid necks, respectively.
Series of 100 crushing tests are carried out to obtain the characteristic
strength of a single agglomerate or aggregate with particle sizes of 10,
100 and 1000 nm. Stress-strain curves indicate that the agglomerates
behave plastically while aggregates are brittle. It is also observed that
with decreasing particle size, agglomerates and aggregates exhibit in-
creasing strength. Simulation of the uniaxial compaction of agglomer-
ated and aggregated powders demonstrates that below a critical parti-
cle size, adhesive forces play an important role thus explaining the dif-
ficulty to compact nanosized powders.
INTRODUCTION
Agglomeration (sticking of particles) and aggregation (particles bonded by
fused necks from a high temperature process) are common phenomena observed
during nanopowder synthesis. If these agglomerates and aggregates are too strong,

their survival may retard the powder compaction leading to a non uniform pore-
size distribution and internal cracks which act as strength-limiting flaws for the
compact. Although microscopy techniques have been developed to characterize
the microstructure of these porous systems; their mechanical properties can not be
retrieved easily, especially for submicronic particles.
DEM has been successfully employed to predict the mechanical properties of
these porous systems [1-3]. However, the effect of particle size and of the resul-
tant adhesive forces (van der Waals forces of attraction) on their mechanical be-
haviour has not yet been fully explored. Thus, the present study utilizes DEM
(dp3D code developed at Grenoble-INP), to study the compaction behaviour of
both agglomerates and aggregates made of submicronic particles. DEM is used to
model porous spherical aggregates and agglomerates consisting of 5000
monosized particles. Three different primary particle sizes of 10, 100 and 1000
nm are considered.
MODEL DESCRIPTION
For agglomerates, particles in contact follow the DMT (Derjaguin- Muller-
Toporov) model [4]. The tensile force, Nf, necessary to break the adhesive bonds
between two particles of radius R with surface energy γ is given by
N f = 2πγ s R (1)
while the repulsive term is given by the classical Hertzian term.
In case of an aggregate, the contact law between bonded particles is derived
from the finite element calculations of Jefferson et al. [5]. Two types of contacts
are involved: i) particles are bonded together by fused neck with bond strength σb
and contact radius ab=0.3R and ii) once the bond has failed in tension or shear, the
particles come under the influence of adhesion. In this case, if particles approach
each other, a contact can be reactivated by taking the history of the existed bond
into account. This broken bond may then transmit compressive load as before the
fracture and may transmit tensile forces according to eqn (1). Owing to the nano-
meter size of ab that holds particles together in aggregate, σb was assumed to be of
the order of E/100 where E=200 GPa is the Young’s modulus. Bending and shear-
ing moments are also included in the aggregate model.

Fig 1. a) Crushing of a) agglomerate and b) aggregate between two platens at

increasing strain. Stress strain profile of c) plastically deformed agglomerate and
d) aggregate exhibiting brittle fracture. The particle size in both cases is 100 nm.
More details on the aggregate model can be obtained in previous works [3,6].
Friction forces are considered for non-bonded contacts by using a simplified
Hertz-Mindlin model applied on the repulsive part of the contact force. For both
models, the surface energy (eqn (1)) is set to 1 J.m-2. For both agglomerates and
aggregates, porosity is approximately 40 %.
CRUSHING TESTS AND CLOSE DIE COMPACTION

Fig 1 (a) shows the typical evolution of an agglomerate crushed in between two
rigid plates. Friction coefficient between the particles and the platens was set to
0.5 to ensure no slippage at the interface. Contrasting to Fig 1 (a), Fig 1 (b) shows
the crushing evolution of an aggregate, which exhibits brittle fracture by breaking
typically in two or three large chunks. Stress strain profile of agglomerate (Fig
1(c)) shows a peak stress, indicating the fracture of some adhesive bonds, after
which the stress decreases to a non-zero plateau-like value at a certain axial strain.
Fig 1(d) shows the typical stress-strain profile for the aggregate. The aggregate
behaviour is brittle and the stress becomes nearly zero for strain larger than 5%. It
should be noted that for aggregate with primary particles of 10nm, the microstruc-
ture evolution during crushing and stress strain profile resembles agglomerate.

Fig 2. a) Weibull plot relating the fracture probability Pf to the peak stress σp
for agglomerate and aggregate b) Characteristic strength σ0 for aggregates and ag-
glomerates (calculated from 100 crushing tests) as a function of particle size.
Series of 100 numerical crushing tests were conducted on agglomerates and
aggregates generated from different random seeds. The obtained peak stress data
(σp) is analyzed with the median ranking method to yield the Weibull parameters.
The cumulative probability of fracture, Pf for an applied stress σ is given by:
Pf = 1 − exp ⎡⎣ − (σ / σ 0 ) ⎤⎦
m
(2)
where σ0 and m are Weibull parameters.

Weibull modulus m increases with decreasing particle size (Fig 2a), thus indi-
cating a decrease in strength spread with decreasing particle size. Fig 2 (b) shows
clearly the particle size effect on agglomerate strength. For aggregates, size effect
arises only when size reduces from 100 to 10 nm. For such sizes, strongly adhe-
sive bonds replace solid bridges that have broken. In the present study, the particle
size at which this transition occurs is 20 nm as depicted in Fig 2b.
To simulate a close die compaction of agglomerated or aggregated powder, a
representative volume element made of 10 spherical agglomerates and aggregates
as described above is used. In each case agglomerates and aggregates are located

randomly in a periodic simulation box such that they should not contact each
other. The procedure for densification has been explained in detail elsewhere [7].
The end of this first stage leads to a density of the order of 0.30-035 also known
as tap density. The interparticle friction is set to 0.5. Fig 3 (a&b) show the micro-
structural evolution of aggregates and agglomerates under compaction. Both mi-
crostructures exhibit some indication of anisotropy with some aggregates clearly
more deformed in the axial direction. For aggregates, particles that are bonded and
fail come under adhesive interactions as per eqn (1). Thus, particle rearrangement
within the powder reduces as inter particle adhesion becomes dominant for
smaller primary particle size. This can be observed for both aggregates and ag-
glomerates as larger stresses are needed to obtain a given relative density when
nanosized primary particles are used. This is clear from Fig 3c where the evolu-
tion of the compaction stress for aggregates and agglomerates packings is plotted
for various particle sizes.
Fig 3. Microstructural evolution of a) agglomerates and b) aggregates during

compaction c) Compaction Stress Vs Relative density for different particle sizes.

CONCLUSIONS
The present work uses DEM to study the compaction behaviour of agglomer-
ated and aggregated nanopowders. Aggregates are brittle when their primary par-
ticles are ≥ 100nm, while agglomerates tend to behave plastically. As the primary
particle size decreases to nanoscale values (10 nm), the collective rearrangement
of agglomerates and aggregates (and of primary particles) is hindered by adhesive
forces during close die compaction. Thus, the stress needed to compact an ag-
glomerated and aggregated powder to attain an optimal green density is increased
with smaller particle size, explaining the difficulty to consolidate nanopowders.
ACKNOWLEDGEMENT
IFCPAR/CEFIPRA is acknowledged for financial support (project 3608-3).
REFERENCES
1. J Nam, J J Lannutti. ‘Localized Densification during the Stage II-III Transition:
Compaction Efficiency at High Pressures‘. J. Am. Ceram. Soc. 87, 557-564, 2004.
2. D Golchert, R Moreno, M Ghadiri, J Litster. Effect of granule morphology on break-
age behaviour during compression. Powder Technol. 143-144, 84-96, 2004.
3. C. L. Martin, D. Bouvard, G. Delette, ‘Discrete element simulations of the compac-
tion of aggregated ceramic powders’ J. Am. Ceram. Soc., 89, 3379-87, 2006.
4. B. V. Derjaguin, V. M. Muller, Y. P. Toporov, ‘Effect of contact deformations on the
adhesion of particles’ J. Colloid Interface Sci., 53, 314 - 26, 1975.
5. G. Jefferson, G.K. Haritos, R.M. McMeeking, ‘The elastic response of a cohesive
aggregate - A discrete element model with coupled particle interaction’ J. Mech.
Phys. Solids, 50, 2539-75, 2002.
6. A Balakrishnan, P Pizette, C L Martin, S V Joshi, B P Saha. Effect of particle size in
aggregated and agglomerated ceramic powders. Acta Mater. 58, 802 - 812 2010.
7. C L Martin, R K Bordia. Influence of adhesion and friction on the geometry of pack-
ings of spherical particles. Phys. Rev. E, 77, 031307, 2008.

AGGLOMERATION DYNAMICS OF MAGNETIC

NANOPARTICLES IN SIMPLE SHEAR FLOW
Eldin Wee Chuan Lim

Department of Chemical & Biomolecular Engineering, National University of Singapore
Magnetic nanoparticles have been observed recently to exhibit inter-

esting behaviours and clustering patterns under the influence of an ex-
ternally applied magnetic field. This phenomenon may have important
future applications in the biomedical and pharmaceutical industries.
However, the interactions between such magnetic nanoparticles and
an applied magnetic field are complex and not well understood at pre-
sent. In this study, a computational approach applying a modified Dis-
crete Element Method (DEM) is used to simulate the phenomenon of
clustering of magnetic nanoparticles. For application towards systems
of nano-sized particles, the additional effects of van der Waals forces,
electrical double layer forces and Brownian forces have been incorpo-
rated into the conventional DEM code. Furthermore, the effects of
particle-particle and particle-field magnetic interactions have also
been taken into account. To our knowledge, this work represents the
first attempt in the research literature at extending the DEM method-
ology towards the study of magnetic nanoparticles. The modified
DEM approach developed in this study represents a promising compu-
tational technique for fundamental studies of nanoparticle mechanics
in general and is expected to gain widespread popularity and impor-
tance in the near future.

INTRODUCTION
The study of magnetic nanoparticles has gained considerable interests among
research workers in recent years. The potential range of application that these
novel magnetic nanomaterials can offer is gradually being recognized and contin-
ues to be explored. At a more fundamental level, one of the most important and
widely investigated aspects of magnetic nanoparticles is the formation of self-
organized microstructures under the influence of an externally applied magnetic
field. A suspension of magnetic nanoparticles in a fluid medium can generally be
considered as a single magnetic domain with macroscopic properties that are de-
pendent on the properties of individual nanoparticles as well as the interactions
between them.1 Although it is well-established that the magnetization of a mag-
netic fluid is related to the arrangement of the suspended magnetic nanoparticles,
which in turn arises due to the effects of interactions between various types of
forces present such as Brownian and dipole-dipole interactions for example, cur-
rent understanding of the kinetics, dynamics and resulting microstructure of the
nanoparticle aggregation process is far from complete.
In this work, we report the first application of a modified version of the Dis-
crete Element Method (DEM) towards the simulation of magnetic nanoparticle
agglomeration with and without an external magnetic field. The various types of
interparticle forces that are important in the dynamics of the agglomeration
process, such as dipole-dipole interactions, van der Waals forces and Brownian
effects, are taken into account through the incorporation of the respective force
models into the classical DEM model. The effects of an external magnetic field
and shear flow on the propensity of such magnetic nanoparticles to agglomerate
and the microstructure of the resulting clusters or chain-like assemblies formed
are investigated computationally.
COMPUTATIONAL MODEL
Discrete Element Method. The molecular dynamics approach to modeling of
particulate systems, otherwise known as the Discrete Element Method (DEM), has
been applied extensively for studies of flow behaviours in various types of multi-
phase systems. The translational and rotational motions of individual solid par-
ticles are governed by Newton’s laws of motion:

= ∑ (f c ,ij + f d ,ij ) + f dd ,ij + f vdw ,ij + f e ,ij + f B,i

N
dv i
mi (1)
dt j=1
dωi N
Ii = ∑ Tij (2)
dt j =1
where mi and vi are the mass and velocity of the ith particle respectively, N is the
number of particles in contact with the ith particle, fc,ij and fd,ij are the contact and
viscous contact damping forces respectively, fdd,ij is the dipole-dipole interaction
between particles i and j in the presence of an applied magnetic field, fvdw,ij and fe,ij
are the van der Waals interaction and electrostatic repulsion between particles i
and j respectively, fB,i is the random force arising due to Brownian effects, Ii is the
moment of inertia of the ith particle, ωi is its angular velocity and Tij is the torque
arising from contact forces which causes the particle to rotate. The effect of gravi-
ty is neglected in the present study. A linear spring-and-dashpot model is imple-
mented for the calculation of contact and damping forces.
Long Range Interactions. In the presence of an externally applied magnetic
field, magnetic nanoparticles can be considered as magnetic single-domains with
πd 3i
a permanent magnetic moment, µi proportional to their volume, µ i = Mi ,
6
where Mi is the intensity of magnetization. By the superparamagnetic magnetiza-
tion law for a monodisperse, colloidal ferrofluid,1
Mi 1
= coth α i − (3)
φs M d αi
π µ o M d Hd 3i
where α i = , µo is the permeability, Md is the saturation magnetiza-
6 kT
tion of the bulk magnetic solid, H is the magnetic field strength, k is the
Boltzmann’s constant, T is thermodynamic temperature and φs is the volume frac-
tion of solid present. The anisotropic dipole-dipole interaction energy Edd,ij is then
given by1

µi ⋅ µ j 3 
 3 − 5 (µ i ⋅ rij )(µ j ⋅ rij )
1
E dd ,ij = (4)
4πµ o  rij rij 
where µi and µj are the magnetic moments of particles i and j respectively and rij is
the displacement vector between the two particles. The dipole-dipole force of inte-
raction acting on particle i is then derived from fdd,ij = -∇Edd,ij. The van der Waals
forces of interaction and electrostatic repulsion between particles are calculated as
follows:2
Ha d i6
f vdw ,ij = ni
( )
6 h ij2 + 2d i h ij 2 (h ij + d i )
(5)
where Ha is the Hamaker constant and hij is the surface-to-surface separation dis-
tance along the line of centers of particles i and j. The electrostatic repulsion be-
tween particles due to the so called double layer forces is described by the DLVO
theory. According to this theory, the electrostatic potential takes the form of a
Yukawa repulsion:2
Z 2 λ B  exp(κR i )  exp(− κrij )

2
U ij =   (6)
β  1 + κR i  rij
where Z is the surface charge, λB is the Bjerrum length and is approximately 0.7
nm for water at room temperature (300 K), κ-1 is the Debye-Huckel screening giv-
en by κ2 = 4πλBn and β-1 = kT is the thermal energy scale at absolute temperature
T. The electrostatic force of interaction between particles i and j is then obtained
by fe,ij = ∇Uij. For the size of nanoparticles simulated, it is pertinent to consider
the random forces arising due to Brownian effects. The algorithm for simulating
Brownian forces is similar to that for generating a Gaussian white noise process:2
d 
f B,i = 12πµ f  i kT∆t n i
'
(7)
2 

where ∆t is the time step used in the simulation and ni’ is a unit vector with a ran-
dom direction.
Simulation Conditions. 5480 spherical nanoparticles of diameter 70 nm and
density 1000 kg m-3 were simulated within a microfluidic channel of length 20 µm
and width 5 µm. A time step of 0.1 ns was applied. At the start of each simulation,
the positions of all nanoparticles were assigned randomly within the computation-
al domain such that no overlap between any two nanoparticles occurred. The
magnetic nanoparticles were allowed to agglomerate within the microfluidic
channel for 10-3 s and then shear flow within the channel was simulated for 10-3 s.
Periodic boundary conditions were applied to the inlet and outlet of the channel.
Shear rates of 2.0 × 102 s-1 and 2.0 × 103 s-1, in the presence and absence of an ex-
ternal magnetic field, were applied.

Fig. 1 and 2 show that an applied shear rate of 2.0 × 102 s-1 in a microfluidic
channel has minimal effect on the agglomeration behaviours of the nanoparticles.
However, with a higher shear rate of 2.0 × 103 s-1, the agglomerates formed both
in the absence and presence of an external magnetic field are observed to undergo
considerable deformation. The original network structures of the agglomerates are
disrupted by the shearing forces and the onset of agglomerate breakage can also
be observed. The case where an external magnetic field is applied shows that the
strength of the magnetic forces is now insufficient to maintain the vertically
aligned orientation of the nanoparticle agglomerates. This represents the first ap-
plication of the modified DEM algorithm for computational studies of the agglo-
meration dynamics of magnetic nanoparticles under the effects of shear as well as
magnetic fields within microfluidic channels.
(a) (b)
Fig 1. Agglomeration patterns of magnetic nanoparticles in a microfluidic channel

after being subjected to shear flow for 10-3 s physical time in the absence of an
external magnetic field. The shear rates were (a) 2.0 × 102 s-1 and (b) 2.0 × 103 s-1.
(a) (b)
Fig 2. Agglomeration patterns of magnetic nanoparticles in a microfluidic channel
after being subjected to shear flow for 10-3 s physical time in the presence of an
external magnetic field. The magnetic field was applied in the vertical direction.
The shear rates were (a) 2.0 × 102 s-1 and (b) 2.0 × 103 s-1.
CONCLUSIONS
The agglomeration dynamics of magnetic nanoparticles was investigated com-
putationally in this study. A modified Discrete Element Method (DEM) which
takes into account various types of long range interactions which are important for
the size of nanoparticles considered was developed. These include magnetic di-
pole-dipole interactions, van der Waals interactions, double layer forces and
Brownian forces. To our knowledge, this work represents the first attempt in the
research literature at extending the DEM methodology towards the study of mag-
netic nanoparticles.
ACKNOWLEDGMENTS
This study has been supported by the National University of Singapore under
Grant No. R-279-000-275-112. The award of a Lee Kuan Yew Postdoctoral Fel-
lowship by the National University of Singapore is gratefully acknowledged.
REFERENCES
1. R. E. Rosensweig, Ferrohydrodynamics, Cambridge University Press, 1985.
2. W. B. Russel, D. A. Saville, W. R. Schowalter, Colloidal Dispersions, Cambridge
University Press, 1989.

MODELLING NON-SPHERICAL PARTICLE

BREAKAGE IN DEM SIMULATIONS
Gary W. Delaney1 , Paul W. Cleary1 , Matt D. Sinnott1 and Rob D.
Morrison2
1 CSIRO Mathematical and Information Sciences, Private Bag 33,
Clayton South, 3168, Australia.
2 Julius Kruttshnitt Mineral Research Center, the University of
Queensland, Isles Rd, Indooroopilly, Qld, 4068, Australia.
Comminution processes are a critical component in many indus-

tries including mineral processing, pharmaceuticals and food pro-
cessing. DEM is a key tool that can be applied to model and better
understand such processes. We extend a previous implementa-
tion of DEM particle breakage that utilized spherical particles to
consider non-spherical particles. Our new implementation uses
super-quadric particles, where daughter fragments with realistic
size and shape distributions are packed inside a bounding parent
super-quadric.
INTRODUCTION
Communition processes in which granular particles are broken down into smaller
sizes are essential in numerous industries, including mineral processing, quarry-
ing, food processing and pharmaceuticals [6, 1]. DEM implementations of particle

breakage using spherical particles have demonstrated that DEM can offer insights
into the communition processes occurring inside complex pieces of industrial ma-
chinery [1, 2]. We have implemented a new DEM particle breakage model using
non-spherical particles, where particles can be broken and replaced with a packing
of daughter fragments that have a desired size and shape distribution. This offers
several advantages and allows for more realistic modelling of complex communi-
tion processes. We demonstrate this new method with examples of low and high
energy particle breakage.
In our DEM simulation, we utilize a linear spring dashpot model and represent
our particle using super-quadrics (superellipsoids) define by:
x m y m z m
+ + = 1, (1)
a b c
which allows us to consider a wide range of particle shapes ranging from ellipsoids
(m = 2) to cuboids (m → ∞). Details of the model are given in Refs [3]. Superel-
lipsoids have previously been utilized in large scale industrial simulations of ball
mills [3], screw feeders [2] and sample cutters [4, 2], where they have highlighted
the important role particle shape plays in the behavior of granular systems.
MODELLING PARTICLE BREAKAGE
Previous DEM models utilizing spherical particles have incorporated simple but
effective dynamic breakage models [1], where breakage occurs due to both high en-
ergy impacts in which the cumulative energy absorption during a collision exceeds
a size dependent breakage energy, and compressive fracture where the total instan-
taneous force applied to a particle exceeds a specifie size dependent compressive
stress threshold.
In Figure 1 we show an example of a simulation of spherical particle breakage
in an industrial jaw crusher. Particles fl w in from the top and fall into the jaw
crushing device. Smaller particles fall through the device and are carried away by
the conveyor belt below. Larger particles become trapped in the jaws of the crusher
and are broken by the large compressive forces. When a particle is broken, it is
replaced with a suitable packing of daughter particles. In the jaw crusher example,
the smaller of these particles will then fall out of the device onto the belt below.

a b c
Figure 1: Spherical DEM simulation of particle breakage in an industrial jaw
crusher.
Where daughter particles are still too large to exit the crusher, they will undergo
further breakage events.
NON-SPHERICAL PARTICLE BREAKAGE
In our new extension to this model we are able to simulate breakage of non-spherical
particles. This requires solving a complex packing problem [5], where we must gen-
erate a dense arrangement of daughter fragments in the space occupied by the parent
particle. We employ dynamic packing techniques similar to those described in Ref
[5], to generate a dense packing inside the bounding parent super-quadric. The
packing can be composed of daughter fragments with any desired range of super-
quadric shapes and come from any desired size distribution. Realistic fragment
distributions from drop weight or similar calibration tests can be used to ascertain
the appropriate distribution of fragment sizes.
In our procedure we begin by identifying a particle that will undergo a break-
age event (Figure 2a). We then initialize a set of small daughter particles inside
the parent superellipsoid (Figure 2b). The daughter particles are then uniformly
increased in size (such that the desired size distribution is maintained) while per-
forming a DEM like dynamic simulation in which the grains collide and interact
with one another (Figure 2c) [5]. As we wish to obtain the densest possible pack-

a b
d c
Figure 2: Example of breakage of a non-spherical particle.
ing, no inter-grain friction is used during the packing process and particles can thus
smoothly slide against one another. A point is reached where the inter-grain over-
laps of the daughter particles diverges as the particles jam together in their fina
configuratio (Figure 2d). The energies of the daughter particles are set using in-
formation from the parent particle’s breakage event and then the DEM simulation
proceeds as normal.
Figure 3a shows a breakage example that is comparable to what would be ex-
pected in a reasonably high energy impact, where a particle shatters into several
moderate fragments and a large number of fin particles. The example shown uti-
lizes daughter particles with a range of shape parameters from m = 2 to m = 5, a
range of aspect ratios (relative sizes of the three grain axes a, b and c) from 1.0 to 2.0
and a wide distribution of particle sizes, with the particles’ diameters ranging from
2.5% to 15% of the parent’s diameter. These various parameters and the choice
of distribution can be calibrated from experimental investigations of real particle
breakage events.
The dynamic packing process can be optimised in several ways. Figure 3 shows
how the packing fraction varies during the particle expansion process. Here, we
initialize the particles at a low packing fraction of φinitial = 0.05. An initial rapid
growth rate at a constant dφ/dt is performed, where the particles are uniformly

a c
Figure 3: (a) Variation in packing fraction φ during the packing process for a high
energy breakage case. (b,c) Example of a low energy breakage case where one
primary fragment and a number of fine fragments are formed.
expanded, while maintaining the desired size distribution. This continues until
φswitch = 0.60, when the rate is reduced by an order of magnitude for the fina stage
of the packing process. The initial fast growth rate is used for computational ef-
ficien y, followed by a reduced growth rate to ensure that a dense fina packing is
formed. It has previously been shown that denser fina packings are formed for
lower growth rates [5]. The values of φinitial and φswitch can be chosen such that the
packing can be generated in the minimum amount of time, while still achieving a
dense packing. In particular much larger values than φinitial = 0.05 can generally
be used and the choice of the φswitch value can be made based on the expected fina
packing fraction value. The possibility of utilizing a non-linear expansion rate is
also being explored.
In the example shown in Figure 3a, a fina packing fraction of φ = 0.72 was
obtained when packing the particles within the parent superquadric. Larger packing
fractions can generally be obtained by utilizing larger numbers of particles that
cover a wider distribution of particle sizes.
This is illustrated with the example of a low energy particle breakage in Figures
3b and 3c, where one primary fragment is formed with a large number of fine
particles. The packing in Figure 3b is formed by 72 particles with a range of sizes
from 4.5% to 35% of the parent diameter and has a packing fraction of φ = 0.62.
If we reduce the minimum fragment size and thus include an additional number
of smaller particles, then we can obtain a higher packing fraction. This is shown in

Figure 3c, where 184 particles with sizes ranging from 2.25% to 35% of the parent’s
diameter are used, and a packing of φ = 0.76 is obtained. The minimum resolved
fragment size can be chosen such that the particle numbers in the simulation remain
within a practical range.
CONCLUSION
We have described a novel DEM implementation which incorporates a new method

for the simulation of non-spherical particle breakage. This overcomes a key limita-
tion in previous implementations that relied on a purely spherical shape description.
The packing technique that we employ to generate a dense arrangement of daughter
particles within the broken parent is sufficientl general that we can consider any de-
sired size and shape distribution. The method we have introduced will offer several
key advantages in particle breakage simulations in industrial communition devices,
where realistic particle shapes can be used for the input material and when a particle
is broken, progeny with a realistic size and shape distribution can be produced.
References
[1] P. W. Cleary. Recent advances in dem modelling of tumbling mills. Minerals

Engineering, 14(10):1295–1319, Oct. 2001.
[2] P. W. Cleary. DEM modelling of particulate fl w in a screw feeder. Progress in
Computational Fluid Dynamics, 7(2-4):128–138, 2007.
[3] P. W. Cleary. The effect of particle shape on simple shear fl ws. Powder Tech-
nology, 179(3):144–163, 2008.
[4] P. W. Cleary and G. Robinson. Evaluation of cross-stream sample cutters using
three-dimensional discrete element modelling. Chemical Engineering Sciences,
63(11):2980–2993, June 2008.
[5] G. W. Delaney and P. W. Cleary. The packing properties of superellipsoids.
EPL (Europhysics Letters), 89(3):34002, 2010.
[6] O. Hlungwani, J. Rikhotso, H. Dong, and M. H. Moys. Further validation of
DEM modeling of milling: effects of liner profil and mill speed. Minerals
Engineering, 16(10):993–998, Oct. 2003.

SOILS, ICE, ROCK, LANDSLIDES,

EARTHQUAKES

COUPLED DEM-LBM SIMULATION OF A SOIL

FLUIDISATION PROBLEM
Xilin Cui, Jun Li, Andrew H.C. Chan, David N. Chapman
A recently developed computational technique coupling Discrete

Element Method (DEM) with Lattice Boltzmann Method (LBM) is
used to simulate soil fluidisation surrounding leaking pipes. Domain
decomposition is adopted so as to make parallel programming possi-
ble in the future. An optimum contact detection algorithm similar to
no binary search (NBS) is also presented in terms of CPU efficiency
and memory requirement. Numerical results indicate the coupled
DEM-LBM technique is able to capture the physical behaviour of the
fluid-particle system for the pipe leakage problem considered.
INTRODUCTION
The problem of leakage from pipes has recently stimulated great interest to re-
searchers not only because it is of special importance in the underground envi-
ronment, but also due to its complexity and the research challenges it poses. In
order to capture the physical behaviour of the problem involving soil erosion and
fluidisation in detail, a fine resolution of fluid-particle interaction is necessary.
This demand has encouraged the emergence of an innovative numerical technique
coupling Discrete Element Method (DEM) and Lattice Boltzmann Method (LBM)
[1]. The coupled technique directly simulates both fluid and solid phases, and it
enables one to capture the physical behaviour of the fluid-particle system in detail.

METHODOLOGIES
Soft-Particle DEM. DEM provides an effective tool for micro-scale analysis
of the behaviour of granular materials. In DEM, granular material is considered as
an assembly of separate particles. When the soft-particle approach is used, slight
overlap between particles is allowed, and the particles interact with one another
only through such contacts. In the present work, for the reason of simplicity, a
contact is viewed as a linear spring-dashpot system. Both normal and tangential
contact forces can be obtained through the deformation history at the contact. The
motion of a single particle is governed by the Newton’s second law in the form of
the following dynamic equations,
    
ma  cv  Fc  Fb  Fh (1)

d  
I  c  T
dt (2)
  
where Fc , Fb and Fb denote, respectively, the total contact force, body force, and
hydrodynamic force computed from the fluid phase with m, c, c’, and I being the
mass, translational and rotational viscous damping coefficients and moment of

inertia of the particle. T denotes the total torque applied on the particle.
LBM with Large Eddy Simulation (LES). LBM is used as an effective alter-
native to conventional macroscopic methods for fluid flow simulations. It is a
time-stepping procedure, based on microscopic kinetic models. In the classic
LBM, the fluid domain is divided into a rectangular or cubic lattice with uniform
spacing. Fluid is viewed as packets of micro-particles residing on the lattice nodes.
During each time step, particles are allowed either to remain in the same node, or
to travel to their nearest neighbouring nodes with corresponding discretised veloc-

ities ei (i  0,..., n) . In order to simulate fluid with high Reynolds numbers, the
Large Eddy Simulation (LES) technique has been implemented with LBM [2].
The governing LBM equation in a turbulence filtered form is given as follows,
~   ~  1 ~  ~
f i ( x  ei t , t  t )  f i ( x , t )  * [ f i ( x , t )  f i eq ] (3)


~ ~eq
where f i are filtered density distribution functions, and f i are a set of filtered
equilibrium distribution functions relating to the discretised velocity and the ma-
croscopic variables, including fluid density and flow rate. The effect of the unre-
solved scale, which is principally responsible for energy dissipation through visc-
ous forces, is considered by a relaxation time  * relating to the turbulence viscosi-
ty t .
Coupling of Solid and Fluid Phases. The Immersed Moving Boundary (IMB)
scheme is adopted for fluid-particle inter-phase treatment. At each lattice node,
the LBM equation is modified using a weighting function depending on the vo-
lume fraction of each particle that intersects the volume surrounding the lattice
node (see Fig.1). In such a way, the fluid flow influenced by the presence of solid
particles can be taken into account. On the other hand, smooth and accurate hy-
drodynamic force and torque acting on a moving particle can also be obtained.
The detailed implementation of this coupling scheme is illustrated in Noble and
Torczynski (1998) [3].
Volume
surrounding
by particle the lattice
boundary Ub node
Moving particle
Figure 1. Volume fraction used in IMB

Domain Decomposition and Contact Detection. In order to bring a large
scale simulation into reality, parallel programming is expected in the future.
Therefore, the computational domain is firstly decomposed into separate sub-
regions, and computations can be carried out on each region independently. The
fluid interface scheme regarding adjacent regions is detailed in Jun et al (2010) [4].
In DEM simulations with a great number of contacting particles involved, con-
tact detection process usually consumes a considerable amount of time. Among
various detection algorithms, the direct mapping scheme (DMS), which is a linear
method, is capable of providing the most CPU efficiency. However, in such a

scheme, a huge amount of computer memory may be required, and adversely lim-
its its applications. An algorithm called no binary search (NBS) for discrete par-
ticles with similar size was proposed in 1998 [5], so as the memory required can
be significantly reduced.
The present work carries out the contact detection with a similar idea to NBS.
Each sub-region is further divided into cells with the same size, and the size of
individual cells is set to equal the diameter of the largest particle. Direct mapping
of all the discrete particles is performed onto regions and cells. A singly con-
nected list of particles is assigned for each region, and the particles mapped onto
the region are listed in an ascending numerical order. An issue needs to be ad-
dressed here is that an overlap area is adopted at the interface of the regions with a
width of the maximum particle diameter. If a particle is centre-located within the
overlap area, it is mapped to both of the adjacent regions, so as to meet the ro-
bustness requirement of the detection algorithm. With a loop performed over all
the particles in a particular connected list, these particles are individually further
mapped onto the cells belonging to the particular region. Contact detection is the-
reafter carried out by checking the particles mapped to a particular cell against
others in the neighbouring cells.
In this scheme, the CPU time spent on contact detection is proportional to the
total number of particles; and the maximum memory requirement in a 2-D case is
given as,
M  4N  R  N   C (4)
where M is memory required for RAM, and N , R, N , C indicate total particle
number, total sub-region number, maximum number of particles within one region,
and total cell number belonging to each region, respectively.
SIMULATION RESULTS
The capability of the coupled DEM-LBM technique is verified by a 2-D simu-
lating of a complex soil fluidisation phenomenon with 10,000 particles. A densely
packed sand bed with different particle diameters is settled at the bottom of a rec-
tangular container, and a pressure-driven fluid is injected in the middle of the con-
tainer base. Thereafter, the sand bed undergoes gradual fluidisation.

The computational domain without space decomposition is divided into a

400*600 lattice grid with the lattice spacing h  0.25mm . In this 2D case, an
assumption of the fluid flow paths out of plane is needed to allow permeability to
be considered. To comply with this assumption, an effective fluid diameter, which
is a certain percentage of the real particle diameter used in DEM, is adopted in
LBM calculation [6].
(a) T=1.25s (b) T=3.75s (c) T=5s
(d) T=6.25s (e) T=7.5s (f) T=8.75s

Figure 2: 10,000-particle sand bed fluidisation with blow-out failure

The results from the numerical analysis of the fluidisation with blow-out fail-
ure of the sand bed are shown in Fig.2. From the results, one can identify that the
blow-out behaviour of the particles and the general pattern of the fluid appear to
be realistic.
CONCLUSIONS
The results from the numerical analysis indicate that the coupled DEM-LBM
technique is able to capture the physical behaviour of the fluid-particle system in
detail. This technique can be regarded as a very promising tool for the investiga-
tion of soil erosion and fluidisation surrounding leaks in pipes. Further validation
against experimental results will be carried out in further work.
REFERENCES
1. B. K. Cook, D. R. Noble, and J. R. Williams. A direct simulation method for
particle-fluid systems. Engineering Computations, 21(2-4), 151-168, 2004.
2. K. Han, Y.T. Feng, and D. R. J. Owen. Coupled lattice Boltzmann and dis-
crete element modelling of fluid-particle interaction problems. Computers and
Structures, 85(11-14): p. 1080-1088, 2007.
3. D. R. Noble, and J. R. Torczynski. A lattice Boltzmann method for partially
saturated cells. Int. J. Mod. Phys., C, 9: p. 1189–1201, 1998.
4. J. Li, X. Cui, A. H. C. Chan, and J. Bridgeman. An innovative method coupl-
ing Darcy fluid flow and lattice Boltzmann method. Proceedings of the 18th
Annual Conference of the Association of Computational Mechanics in Engi-
neering, 29-31 March 2010, University of Southampton, UK ed. A. Zervos,
p.191-194.
5. A. Munjiza, and K. R. F. Andrews. NBS contact detection algorithm for bo-
dies of similar size. Internation Journal for Numerical Methods in Engineer-
ing. 43, 131-149, 1998.
6. D. F. Boutt, B. K. Cook, B. J. O. L. McPherson, and J. R. Williams. Direct
simulation of fluid-solid mechanics in porous media using the discrete element
and lattice-Boltzmann methods. Journal of Geophysical Research, 112,
B10209, 2007.

FOLD DEVELOPMENT IN COMPRESSED MULTI-

LAYERS MODELLED WITH FEMDEM
John-Paul Latham, Jiansheng Xiang and Adrian Shelley

Imperial College London, Department of Earth Science and Engineering
Folded rock strata provide ubiquitous structural evidence that com-

pressional tectonic regimes have been active in most parts of the
world at various periods in Earth history, yet mechanical explanations
for the variation in styles of these fold structures are highly specula-
tive. Geologists have constrained typical bulk geological strain rates
during folding to be between 10-13-10-15 sec-1 but face a huge range of
possible rheologies and ductile styles of deformation including sinu-
soidal, concentric-like, chevron, kink-like and fault-related forms.
Most of the breakthroughs in the understanding of geological folding
in the 60’s and 70’s came about through experiments with physical
analogue models combined with predictions from analytical theories,
many of which were based on elastic theory or linear and power law
viscous theory valid for small strains only. Nowadays, structural ge-
ologists increasingly turn to numerical modelling methods to simulate
the growth of initial instabilities into large finite strain structures.
FEMDEM modelling which combines the finite element method FEM
(for the internal deformation) and the discrete element method DEM
(for the discrete interactions between sliding layers or folding fabric
elements) has a potential advantage here over FEM as it can combine
brittle deformation and discrete slip with bending and shearing inter-
nally within multilayers. This paper illustrates preliminary results of
plane strain FEMDEM simulations of multilayer deformation which
were performed using the www.VGeST.net suite of simulation tools.

INTRODUCTION
Compositional layering, mineral textures or a series of parallel discontinuities
are often the key components in tectonic folds as they tend to be the mechanically
active elements. To a first order approximation, all three components can be rep-
resented by mechanically homogeneous materials of anisotropic rheology. One of
the most intriguing folding theories is Biot’s (1965) theory for internal buckling of
homogeneously anisotropic materials. When applied to a competence contrast be-
tween regular alternations of layers, two bulk moduli for resistance to shortening
and resistance to shear govern the form of instability. For compression parallel to
layering Biot’s small strain elastic theory predicts a sinusoidal buckling (Type 1)
instability with axial planes growing normal to the compression. Fig. 1a is a geo-
logical example of this type of instability after the folds have amplified.
When the layering is at 45 degrees to the principal compression, the theory in-
dicates conditions to be ideal for localised shear (Type 2) instability to develop.
These propagate along characteristic directions at approximately ±45 degrees to
the compression direction. The original figure presented by Biot in 1965 showed
this Type 2 solution to be expressed by a deformation pattern that resembled what
geologists termed a kink-band, such as seen in Fig. 1c. The conditions which pro-
mote localised kink-band formation in preference to diffusive buckling modes
have been investigated experimentally by several geologists, who also referred to
Biot’s theory, e.g. Cobbold, Cosgrove and Summers (1971) and Cosgrove (1976).
They proposed that for layer parallel shortening a high mechanical anisotropy
combined with locally inclined layer attitude could lead to local modulus inver-
sion sufficient to kick off tabular Type 2 instabilities. Honea and Johnson (1976)
also performed a series of experiments with an elastic sheet bilaminate with high
competence contrast (Fig. 1b) as well as experiments with identical elastic sheets
separated by frictional contacts, (Fig. 1d). They associated kink-banding with the
overcoming of frictional resistance locally on interfaces at low angles to the prin-
cipal compression together with plastic failure in hinges to create the localized
sharp forms for the tabular kinked zones. A high competence contrast, they there-
fore suggested, would favour sinusoidal buckles, amplifying into concentric and
chevron forms with box-shapes becoming necessary at planar boundaries for high
strains. Our simulations are directed towards this debate and the various interpre-
tations of conditions that favour either sinusoidal buckling or kinking.

a b
c
d
Fig 1. Expression of instability in multilayer systems compressed parallel to
the layering. Top: Sinusoidal Buckling Instability often interpreted as Biot’s Type
1 Instability. Bottom: Conjugate reverse kink-bands sometimes interpreted as
Biot’s Type 2 Instability. Left: Examples from nature; (a) in a mica fabric from
Mesozoic metasediments, after Cosgrove (1976); (c) in a layered phyllite, West
Cork, Ireland after Cobbold et al. (1971). Right: Examples from experiments with
compressed elastic sheets by Honea and Johnson (1976); (b) stiff photoelastic rub-
ber/soft gelatin with Modulus ratio ~45, equal thickness layers lubricated with
silicone grease; (d) thin strips of pure gum rubber in frictional contact.

(a) 4.7
(b) 5.3
(c) 5.9

(d) 8.2
(e) 5.3 (f) 5.9
Fig. 2 FEMDEM modelling of the growth and locking up of fold structures with
percentage shortening as indicated. The simulated folds are developed in an elas-
tic bilaminate multilayer with strong and weak layers having a competence con-
trast of 20 and allowing interlayer slip. A Coulomb friction coefficient of 0.6 is
applied. (a) to (d) Left; Von Mises (VM) stress components and deformation,
Right; details also including velocity vectors. (e), (f) show details of the mesh,
indicating the magnitude and location of slip between the individual layers near
the stage of explosive amplification at about 5% bulk shortening.
NUMERICAL MODEL OF MULTILAYER BUCKLING IN A

BILAMINATE OF ALTERNATING STIFF AND WEAK LAYERS
The bilaminate model consists of 50 layers representing a 25m thick stack of
50 m long rock layers. Alternating stiff and weak layers of E1 =10 GPa and E2 =
0.5 GPa, have a Poisson’s ratio of 0.325 and Coulomb friction coefficient of 0.6

between layers. To achieve significant shortening of ~20%, constant inward dis-

placement rates of 20 ms-1 parallel to layering and constant outward displace-
ments set to 2 ms-1 were applied. Material properties are time-independent. An
initial perturbation of part circular form was introduced at the lower boundary.
The progressive deformation is shown in Fig. 2. Run-times were about 3-4 hours.
By 5.3% shortening, a period of significant amplification begins for the fold at
the centre of the domain and limb rotation is clearly indicated. Five antiformal
hinges can just be seen and ‘flow cells’ have propagated to influence almost the
entire model domain. The progressive fold development follows the deformation
observed in Honea and Johnson’s (1976) bilaminate experiments (Fig 1d) almost
identically. Discrete slip has clearly occurred on certain layer interfaces and to
differing degrees, but increasingly after about 2% shortening (Fig. 2e, f).
CONCLUSIONS
High mechanical anisotropy of the compressed system investigated in this pre-
liminary study has favored the initiation of Type 1 i.e. sinusoidal buckling.
FEMDEM handles complex behaviour including slip between layers, boundary
conditions, initial perturbations in layer geometries and material properties, layer
thickness ratios, interlayer friction, and material properties including failure
within layering. Further developments will be reported at the conference.
REFERENCES
1. Biot, M.A., 1965. Mechanics of Incremental Deformations. Wiley, New York,
N.Y., 505 pp.
2. Cobbold, P. R., Cosgrove, J. W., & Summers, J. M. 1971. Development of In-
ternal Structures in deformed anisotropic rocks. Tectonophysics (12), 223-253.
3. Cosgrove, J. W., 1976. The formation of crenulation cleavage. Journal of the
Geological Society 1976; v. 132; p. 155-178.
4. Honea, E. & Johnson, A.M., 1976. A theory of concentric, kink and sinusoidal
folding and of monoclinal flexuring of compressible, elastic multilayers. IV.
Development of sinusoidal and kink folds in multilayers confined by rigid
boundaries. Tectonophysics, 30: 197-239.

DISCRETE MODELING OF GEOMATERIALS

UNDER EXTREME LOADING
T. Tran1,2, P. Marin1, L. Scholtes2, F.-V. Donzé1,2

1
Laboratoire 3SR, Université Jospeh Fourier, Grenoble Université, France
2
Earth Science and Resource Engineering, QCAT, CSIRO, Pullenvale, Australia
A numerical model based on a three dimensional Discrete Element

Method (DEM) has been used to study the behavior of cohesive geo-
materials such as concrete, under a high-confining pressure (up to 650
MPa). At this range of pressures, irreversible compaction of the mate-
rial occurs and needs to be considered. Within the discontinuous na-
ture of the model, a local constitutive law has been developed to re-
produce this phenomenon quantitatively. The computational imple-
mentation has been carried out in the Discrete Element and open
source code YADE: http://yade.wikia.com [1].
INTRODUCTION
During extreme loadings, such as impacts or blasting, the material is subjected
to various states of stress. Near the loading zone, irreversible compaction is pro-
duced, whereas farther from this location, the material experiences compression
with a moderate triaxial stress state [2]. The computational analysis of such events
must be able to capture the key features of the material response under such loads:
tensile cracking, compression failure, the effect of confinement on the ultimate
stress and compaction.

The complex constitutive behavior of cohesive geomaterials, arising from ex-

tensive micro-cracking and macro-cracking, is difficult to characterize in terms of
a continuum formulation. Discrete models, which take directly into consideration
the physical mechanisms and the influence of the concrete aggregate structure,
offer an interesting tool to model fracture, but they still present limitations when
confining pressures, encountered for example during impact loading, reach values
on the order of 1 GPa. When dealing with non deformable discrete elements at the
structure scale, the local behavior needs to integrate local compaction process in
order to obtain correct quantitative simulations.
FORMULATION OF THE MODEL

The algorithm used in the present Discrete (or Distinct) Element Method
(DEM), involves two steps. In the first one, interaction forces are computed when
elements slightly interpenetrate each other. In the second step, Newton second law
is used to determine, for each Discrete Element (DE), the resulting acceleration,
which is then time integrated to find the new element positions. This process is
repeated until the simulation is finished.
For small deformations, cohesive frictional geomaterials exhibit a linear elastic
response. To reproduce this behavior, linear elastic interaction forces between the
discrete elements are sufficient and lead to small simulation times. In the present
model, the interaction force F which represents the action of DE a on DE b may
be decomposed into a normal force Fn and a shear force Fs which may be classi-
cally linked to relative normal and incremental shear displacements respectively,
through the stiffnesses, Kn and Kn2 in the normal direction and Ks in the tangential
direction.
The normal interaction force can be calculated through the updated local con-
stitutive law, which is shown in Fig. 1, and can be split into two parts, the com-
pressive and the tensile components. In the compressive part, we propose a more
complete formulation of Fn in order to take into account the compaction process
which occurs at a lower scale than the discretization size of the model. Note that
when the interaction’s deformation reaches an elastic deformation limit εmax which
is related to the interaction distance D1, a hardening-damage response is consid-
ered (section BC in Fig. 1).

Fig 1. Normal interaction force between two discrete elements

At this step of the interaction, the normal interaction force is characterized by a
nonlinear stiffness Kn2, which is an exponential function of the deformation and it
is controlled by three parameters ζ1, ζ2 and ζ3. This stiffness is expressed as,
⎡ ⎛ ζ (ε -ε max ) ⎞ ⎤
K n2 = K n ζ1 * ⎜ e 2 ⎟ + ζ3⎥ (1)
⎣⎢ ⎝ ⎠ ⎦
where the maximum elastic deformation is defined by,
Deq − D1
ε max = (2)
Deq
So, the complete normal interaction force becomes,
F = K (D − D ) + K (D − D ) (3)
n n eq 1 n2 1 ab
In the tensile part, the normal interaction force can also be computed. How-
ever, the stiffness may be modified by a softening factor ζ while the softening be-
havior occurs after the normal force has reached its maximum value, Fn,max. When
the rupture occurs (Dab > Drupture), the interaction forces are set to zero. The shear
force Fs is calculated by updating its orientation which depends on the orientation

of the direction passing through the two centroids of the interacting DE, and by
adding the increment of shear force.
A modified Mohr-Coulomb model has been used in the model. For a given in-
teraction, the maximum normal interaction force Fn,max is defined as a function of
the tensile strength T. The maximum shear interaction force Fs,max is characterized
by the normal force Fn, the cohesion C, the contact frictional angle Φc and the in-
ternal frictional angle Φi. When the new contacts appear during the simulation, the
tensile strength T is set to zero and their maximum shear force will only be fric-
tional.
Finally, to reproduce quantitatively the behavior of a granular material when
spherical discrete elements are used, the interaction between DE must transmit a
moment [3] which controls the rolling occurring during shear displacement. Do-
ing so, the sliding process increases and the resulting friction angle can reach val-
ues corresponding to those measured for concrete materials.
PREDICTION FOR HIGH CONFINING PRESSURE

The numerical model has been set up (see Fig. 2) to simulate the experimental
tests carried out by Gabet [4] on concrete samples. For this purpose, the local pa-
rameters of model were first identified using uni-axial compressive – tensile tests,
hydrostatic tests at 650 MPa and the tri-axial test carried out at 50 MPa.
Once calibrated and without changing the values of the local parameters, the
numerical model is used to simulate tri-axial tests at other confining pressures.
Four tri-axial tests have been simulated for the following confining pressures: 100
MPa, 200 MPa, 500 MPa and 650 MPa (see Fig. 3). They are compared with data
obtained from the experimental campaign. In all cases, the same initial numerical
parallelepiped sample has been used. Both, experimental and numerical results are
presented with lateral-axial strain/axial stress components and volumetric curves.
The results indicate that in the hydrostatic phase, the numerical results are very
close to the experimental measurements (Fig. 3(a)). In the differential-stress
phase, the simulated results are overall comparable with the experimental ones.

Fig 2. Configuration of tri-axial test

Experimentally, for the 100 MPa test, a stagnation of the axial pressure is ob-
served, without reaching a peak as is the case in the 50 MPa test. Then, for the
500 MPa and 650 MPa tests, the stiffness curves gradually decrease without
changing slope. All the numerical curves show a stress peak and before it occurs,
a decrease of the stiffness is observed in both the model and the experiments.
Fig 3. Stress-strain (a) and mean stress-volumetric strain curves for 50MPa, 100
MPa, 200 MPa, 500 MP and 650 MPa confining pressures. Solid and dotted lines
correspond to the numerical and experimental results respectively.
As expected, the numerical and experimental volumetric deformations on the
hydrostatic phase seem to fit very well (see Fig. 3(b)). However, when the con-
traction-dilatation transition occurs at the same mean stress value, the amplitudes

of the maximum contraction of the numerical volumetric curves are lower than the
experimental ones. This is directly linked to the observations made on the stress-
strain curves.
CONCLUSION
Within the Discrete model, when the confining pressure is high, the rearrange-
ment of the DEs within the numerical model, which simulates the collapse of po-
rosity, is no longer possible, since the DE cannot be deformed and only overlap-
ping is permitted. Thus, the only way to simulate the whole compaction process
numerically is to control how the overlapping between DEs will evolve; this is
done by adding the non-linear hardening step to the compressible part of the inter-
action forces. This is the key point to get a quantitative macroscopic description of
concrete (i.e. here, stress-strain or volumetric curves) under extreme static load-
ings.
REFERENCES
1. J. Kozicki, F.V. Donzé, ‘YADE-OPEN DEM: an open--source software using

a discrete element method to simulate granular material’, Engineering Compu-
tations, 26(7), 786-805, 2009
2. W, Shiu, F.V. Donzé, L. Daudeville, ‘Compaction process in concrete during
missile impact: a DEM analysis’, Computers and Concrete, 5(4), pp. 329-
342, 2008
3. J.P. Plassiard, N. Belheine, F.V. Donzé, ‘A spherical discrete element model:
calibration procedure and incremental response’, Granular Matter, 11(5),
1434-5021, 2009
4. T. Gabet, Y. Malecot, L. Daudeville, ‘Triaxial behavior of concrete under high
stress: Influence of the loading path on compaction and limit states’, Cement
and Concrete Research, 38, 403-412, 2008.




THE APPLICATION OF THE HYBRID STRESS

BLASTING MODEL TO IMPROVING THE
UNDERSTANDING OF WALL CONTROL BLASTING
Ewan J. Sellers
AEL Mining Services, South Africa
The Hybrid Stress Blasting Model is a continuum - lattice - discrete

element code for the analysis of rock blasting. The multiphysics
coupling of detonation and rock models are briefly described. The
benefits of numerical modelling for reducing the number of test blasts
required for the selection of optimal blasting parameters are discussed
by considering the case study of wall control blasting at a large open
pit diamond mine. The model indicates that the interactions of the
rock fragments during the blast are much more complex than pre-
viously thought and emphasizes the criticality of correct and accurate
inter hole timing.
INTRODUCTION
The HSBM model has been developed by an international consortium of min-
ing and explosive companies to model blasting processes from detonation, dy-
namic wave propagation and rock fragmentation to muck pile formation (Furtney
et al, 2009). The new code uses a unique combination of discrete and continuous
numerical techniques Here, the model is applied to a case study of wall control
blasting on an open pit mine in Botswsana (Oageng, 2008). In this case study, a
number of blasts were taken before the optimal solution was determined. Such
tests involve expensive and major administrative efforts to bring new explosive

products onto the mine as well as taking time to arrive at a final design. The avail-
ability of a numerical model to simulate the effect of many of these blasting pa-
rameters would significantly speed up the time to determine the optimal blast de-
sign. This paper describes insights from validation studies of the HSBM model
that, though done in retrospect, will provide the knowledge and confidence to ap-
ply the HSBM to limit the number of costly test blasts in future wall control blast-
ing applications.
HYBRID STRESS BLASTING MODEL (HSBM)

The numerical aspects are described in detail by Furtney et al, (2010). A conti-
nuum method is used to represent the near-borehole area and the detonation
process. The detonation process and axial borehole flow are modeled with a “pro-
grammed burn” approach in which the velocity of detonation, heat of reaction and
product-phase behavior are supplied by an external code (Braithwaite et al, 2010).
This continuum zone is coupled to a DEM representation of the rock body, which
models the wave propagation and initial fragmentation through to muck pile for-
mation. The DEM representation uses a lattice-type method in which rock is
represented as a collection of randomly located point masses connected by springs
with a tensile breaking strength. “Quiet” lattice boundaries are used to represent a
far-field rock mass; “flexible” lattice boundary conditions are used to model exist-
ing broken rock. A simplified diffusion law based gas flow model describes the
flow of high-pressure detonation gas into the newly created fractures. The solving
logic is designed for multicore processors on a windows pc and is encapsulated in
a user-friendly graphical interface that displays results during code execution. The
code includes built-in blast geometries, a customizable non-ideal detonation mod-
el, a stochastic joint generator and advanced graphics.
CASE STUDY OF WALL CONTROL BLASTING

In 2005, a large open pit mine in Botswana was experiencing stability prob-
lems in the pit walls as a result of over breaking by incorrect blasting in very weak
Kimberlite rock with a compressive strength that ranges from 8MPa to 40MPa. A
project to implement specialized wall control blasting techniques was successfully
accomplished (Oageng, 2008). As shown in Fig 1, the interventions significantly
improved the condition of the final walls, but required a large number of test

blasts that varied the blasting parameters. Wall control blasting techniques involve
firstly creating a tensile pre-split along the line of the final wall with closely
spaced blast holes that contain low quantities of explosive, usually decoupled
from the side of the blast holes. Then, production blasts are taken to remove rock
until only 3 lines of holes remain in front of the presplit. The blasting of these last
three lines is known as a trim blast and requires special care in the selection of
blasting parameters to produce good wall control by limiting the explosive energy
transmission and associated damage in the final wall.
a) b)
Fig 1. Final walls for soft kimberlite rock a) before and b) after the project
(Oanageng, 2008)
MODELLING OF DAMAGE DUE TO TRIM BLAST

In the first model, only the effect of changing the explosive distribution be-
tween holes is considered. A simple model (shown in Fig 2) consisting of a single
column of holes and a section of the presplit (shown on plan in Fig 3) is consi-
dered. This model uses about 600 000 nodes in 1Gb of memory and took 15
hours to run until 200ms. Fig 4a shows that damage (indicated as broken bonds in
the DEM lattice) in the case where the trim holes all have a same amount of ex-
plosive penetrates the presplit crack into the final wall. In contrast, when the ex-
plosive mass is reduced towards the back of the blast (Fig4b) the presplit holes
remain undamaged and protect the rock in the final wall position.

Fig 2. Model showing line of presplit holes and a line of 3 blast holes forming
the trim blast
a) b)
Fig 3. Plan view of blast. Solid line indicates the free face. Chain dotted line
indicates section model with presplit lime (blue dots). The dotted box defines the
extent of the plan model. a) shows original 33-100 timing and b) final 75-42 tim-
ing with expected throw direction indicated by arrows
a) b)
Fig 4. Final damage state in section for the case of a) uniform explosive mass
in all 3 holes and b) decreasing the explosive mass towards the presplit

PLAN MODEL OF INTER-HOLE TIMING IN TRIM BLASTS

Fig 3 indicates the position of the second model used to develop an understand-
ing of the effect of inter-hole timing. The first 3 holes near the corner of a face are
considered. The arrows in Fig3 indicate the conventional understanding of blast
throw as being predominantly in the direction perpendicular to the timing con-
tours. Slower timing along the free face leads to a flatter arrow Fig 5b and hence
less reaction back into the remaining rock mass. However, the velocity vectors in
a sectional cutaway of the second model show surprising complex particle interac-
tions that differ significantly as the inter-hole timing is varied.
Fig 5. Complex particle directions after 150ms as indicated by velocity vectors

for a) 33-100 and b) 45-75 timing

CONCLUSION
Selection of the correct blasting parameters leads to better risk management,
pit stability, improves mine planning, extends the life of mine and increases the
net present value of the project. The Hybrid Stress Blasting Model is a continuum
- lattice - discrete element code for the analysis of rock blasting. By considering
the case study of wall control blasting at a large open pit diamond mine it has
been shown that suitable blasting parameters can be selected upfront to minimize
the number of costly test blasts required to improve the blasting practices on
mines. The model has improved our understanding of the effect of inter-hole tim-
ing on the interactions between the rock fragments.
ACKNOWLEDGMENTS
Acknowledgments to J.K. Furtney, P.A. Cundall, M Braithwaite, K Oageng, R
O Joseph and I Munyadzwe for their technical input.
REFERENCES
1. J.K. Furtney, P.A. Cundall, G.D. Chitombo. Developments in numerical mod-
eling of blast induced rock fragmentation: Updates from the HSBM project.
Rock Fragmentation by Blasting , Proceedings of the 9th International Sympo-
sium on Rock Fragmentation by Blasting, FRAGBLAST9, Granada, Sanchi-
drián (ed). Taylor & Francis Group, London, 335-342, 2010
2. M. Braithwaite, G.J Sharpe & G.P. Chitombo Simulation of real detonations
as an energy source term for the Hybrid Stress Blasting Model, Proceedings of
the 9th International Symposium on Rock Fragmentation by Blasting,
FRAGBLAST9, Granada, Sanchidrián (ed). Taylor & Francis Group, London,
327-333, 2010
3. K Oageng, R O Joseph and I Munyadzwe. Perimeter control at Orapa Di-
amond Mine. The Southern African Institute of Mining and Metallurgy Surface
Mining Conference. 137-152. 2008.

DISCRETE ELEMENT SIMULATION OF ROCK

CUTTING WITH EVALUATION OF TOOL WEAR
Jerzy Rojek1, Carlos Labra2, Eugenio Oñate2

1
Institute of Fundamental Technological Research, Polish Academy of Sciences, Warsaw, Poland
2
International Center for Numerical Methods in Engineering (CIMNE), Barcelona, Spain
This paper presents numerical modelling and simulation of rock cut-

ting processes. The rock is modelled using spherical discrete elements.
Estimation of tool wear has been included into the numerical model.
The discrete element model has been calibrated by simulation of the
UCS and Brazilian tests. Rock cutting with a single point attack pick
of a roadheader has been studied experimentally and numerically. A
good qualitative and quantitative agreement of numerical results with
experimental measurements has been found out.
INTRODUCTION
Variety of rock-cutting technologies is used in civil as well as in mining engi-
neering. Design of cutting tools and setting parameters of cutting operations re-
quires knowledge about the cutting process and its parameters. Cutting forces and
tool wear are important factors characterizing a cutting process. Experimental
studies of cutting processes are expensive and require a lot of tests to check dif-
ferent tool designs and their performance for different process parameters. Theo-
retical evaluation of the cutting force is not an easy task. Simple analytical mod-
els, like those developed by Evans [2] or by Nishimatsu [5], can provide a very
approximate estimation of cutting forces only. Numerical methods based on the

continuum models, like finite element methods, have serious problems in model-
ling discontinuities of the material occuring during rock cutting [4]. The discrete
element method is commonly regarded as a suitable tool to model cutting proc-
esses. Due to large computational effort discrete element of rock cutting model-
ling is often limited to 2D case [3]. Two-dimensional models, however, cannot
predict all the components of cutting forces. Here a three-dimensional model is
presented. The discrete element model of rock cutting is extended on the evalua-
tion of tool wear.
BASIC ASSUMPTIONS OF THE ROCK CUTTING MODEL

A numerical model of rock cutting has been developed within the authors’ own
implementation of the discrete element method [6, 7]. The system consisting of a
tool and rock sample is considered in the model. The rock material is represented
as a collection of rigid spherical particles interacting among themselves with con-
tact forces. The contact force between two particles F can be decomposed into
normal and tangential components, Fn and FT, respectively
F = Fn + FT = Fn n + FT (1)
where n is the unit vector normal to the particle surface at the contact point. The
contact forces Fn and FT are obtained using a constitutive model formulated for
the contact between two rigid spheres. In the present formulation the elastic per-
fectly brittle model is used. Initial bonding for the neighbouring particles is as-
sumed. When two particles are bonded the contact forces in both normal and tan-
gential directions are calculated from the linear relationships:
Fn = k n u n , FT = k T u T (2)
where kn and kT are the interface stiffness in the normal and tangential direction,
respectively, un – normal relative displacement, uT – tangential relative displace-
ment. The tensile and shear contact forces are limited by the tensile and shear in-
terface strengths, Rn and RT, respectively:
Fn ≤ Rn , FT ≤ RT (3)

Cohesive bonds are broken instantaneously when the interface strength is ex-
ceeded either by the tangential contact force or by the tensile contact force. In the
absence of cohesion the normal contact force can be compressive only and tan-
gential contact force can be nonzero due to friction. In the present formulation the
Coulomb model of friction is used.
A rock cutting tool is treated as a rigid body with a surface discretized with tri-
angular facets. The tool-rock interaction is modelled assuming Coulomb friction
model. The tool-rock contact model is enriched by evaluation of the tool wear.
The wear rate w& is calculated using the Archard formula [1]:
p n vT
w& = k (4)
H
where pn is the contact pressure, vT – the slip velocity, H – the hardness of worn
surface, and k – a dimensionless wear parameter.
DETERMINATION OF ROCK MODEL PARAMETERS

A set of micromechanical parameters yielding required macroscopic rock
properties has been determined using the methodology developed by Huang [3]
based on the combination of the dimensional analysis with numerical simulation
of the standard laboratory tests for rocks, unconfined compression test (Fig. 1a)
and Brazilian test (Fig. 1b).
a) b)
Fig 1. Results of numerical simulation of laboratory strength tests: a) uncon-
fined compression test, b) Brazilian test

SIMULATION OF ROCK CUTTING WITH A ROADHEADER PICK

Validation of the rock cutting model has been carried using experimental re-
sults obtained in a laboratory test performed in the laboratory of Sandvik Mining
and Construction (Fig. 2a). The tests consisted in cutting of a sandstone block
a) b)
Fig 2. Laboratory rock cutting test: a) experiment (courtesy of Sandvik Mining
and Construction GmbH, Zeltweg, Austria), b) numerical simulation
with a rotating roadheader pick. Mechanical properties of the rock have been de-
termined experimentally and are the following: Young modulus E = 18690 MPa,
Poisson ratio ν = 0.23, compressive strength σc = 127 MPa and tensile strength σt
= 12 MPa.
The test has been analysed using a three dimensional discrete element model.
Rock sample has been discretized using 71,200 spherical particles. For the rock
considered the following set of micromechanical parameters has been found for
the DEM model: contact stiffness in the normal direction kn = 5.4·106 N/m, con-
tact stiffness in the tangential direction kT = 2.16·106 N/m, cohesive bond
strengths in the normal and tangential direction, Rn = RT = 86 N. The results of
numerical simulation are shown in Fig. 2b. Splitting of chips typical for brittle
rock cutting can be seen.

The three components of cutting forces obtained in simulation are plotted in

Fig. 3. Numerical forces are compared with experimental average measurements.
Quite a good agreement can be observed.
The tools used in the laboratory tests of rock cutting have special tips made of
copper, a soft material which is easily worn. This allowed us to obtain visible
wear effects after few cuts (Fig. 4a). Evaluation of wear was included in the
analysis. Evolution of the tool wear predicted by the analysis is shown in Fig. 4b.
It can be seen that the wear pattern obtained in simulation agrees very well with
the worn area observed in the tools used in laboratory tests.
Fig 3. Rock cutting forces – comparison of numerical results with experimental

average values
a) b)
Fig 4. Wear pattern on the tool surface: a) laboratory test, b) simulation

CONCLUSION
The three-dimensional discrete element model of rock cutting is capable to rep-
resent correctly complexity of a rock cutting process. A good qualitative and
quantitative agreement of numerical results with experimental measurements has
been found out in the validation of the model developed in the present work. The
model developed can be employed in the design of rock cutting tools and proc-
esses.
ACKNOWLEDGMENTS
The authors acknowledge partial funding by the EU project
TUNCONSTRUCT (contract no. IP 011817-2). Special thanks are given to
Hubert Kargl and Jan Akerman from Sandvik Mining and Construction GmbH,
Zeltweg, Austria, for providing experimental results.
REFERENCES
1. J.F. Archard, Contact and rubbing of flat surfaces, J. Appl. Phys., 24, 981–988,
1953
2. I. Evans, The force required for pointed attack picks, Int. J. Min. Engng., 2,
63–71, 1965.
3. H. Huang, Discrete Element Modeling of Tool-Rock Interaction, PhD Thesis.
University of Minnesota, 1999.
4. J. Jonak, J. Podgórski, Mathematical model and results of rock cutting model-
ling, Journal of Mining Science, 37, 615–618, 2001.
5. Y. Nishimatsu, The mechanics of rock cutting, Int. J. Rock Mech. Mining Sci.,
9, 261–270, 1972.
6. E. Oñate, J. Rojek, Combination of discrete element and finite element meth-
ods for dynamic analysis of geomechanics problems, Comput. Meth. Appl.
Mech. Eng., 193, 3087–3128, 2004.
7. J. Rojek, E. Oñate, F. Zarate, J. Miquel, Modelling of rock, soil and granular
materials using spherical elements, Proc. of the 2nd European Conference on
Computational Mechanics ECCM-2001, Cracow, Poland, 26-29 June 2001.

MODELLING PUNCH-THROUGH TESTS WITH 2D

FEM-DEM
Arttu Polojärvi ∗ and Jukka Tuhkuri
Aalto University, School of Science and Technology
Department of Applied Mechanics
P.O.Box 14300, FI-00076 Aalto
e-mail: [email protected]
Discontinuous models have been successfully used in numerical

modelling of ice rubble, ice pile-up processes, ridge keel defor-
mation and other discrete problems in ice mechanics. This paper
describes modelling of ridge keel punch through tests by using a
two dimensional combined finite-discrete element method. In a
ridge keel, the ice blocks are usually partly frozen together. The
freeze bonds between the ice blocks are modelled by using cohe-
sive bonds between the blocks. Failure of the bonds due to ten-
sion and shear is allowed. The model was applied to simulation
of ridge keel punch through tests. Example simulation runs with
different bond strengths were performed. Preliminary simulation
results were compared to full-scale results.
INTRODUCTION
Ice ridges are common and important features in Northern seas. Ridge structure
can be divided into three main parts. The visible part above the waterline and level
ice is called sail. Close to the waterline lies the consolidated layer. Under the
consolidated layer lies a rubble pile of ice which can be extensive in size. This part
∗
corresponding author

Direction of
indentor motion indentor platen
frozen layer
keel formed by ice blocks
Fig. 1: A sketch of a punch through test. An indentor platen is pushed down into
the ice rubble formed by ice blocks partly frozen together.
is called the keel. The keel can be consolidated, unconsolidated or a mixture of

these.
In a punch through test, a flat indentor is ran down into the ice rubble (see Fig
1). From the dimensions of the test set up and measurements during the experiment
some material properties for the rubble can be derived. Research has been done on
punch through tests in full-scale, laboratory and through modelling (see eg. [4]).
In the research reported in this article FEM-DEM [3] was used to model punch-
through tests of a ridge keel. Individual ice blocks forming the rubble are mod-
elled as deformable continuum. The blocks interact through pairwise contacts. The
freeze bonds connecting individual ice blocks are implemented using initially rigid
cohesive elements. Hence, before the initation of the cohesive crack growth the
freeze bond behaves as part of the continuum. A failure criterion related to stress
components is used on bond surfaces to initiate crack growth. The material failure
within the ice blocks is limited to occur on freeze bonds.
MECHANICS
In the simulations contact forces were derived using penalty function and poten-

tial contact force method [3]. The deformation of ice blocks is caused by inertial
forces, contact forces, cohesive forces, and buoyant force due to water.
The deformation of individual material elements within an ice block is assumed
to be small. Material behaviour is assumed to be linear and isotropic until failure.
Further, plane stress assumption is used. Material damping in the simulations is
viscous. Constant strain triangles elements were used. A detailed description of the
determination of internal forces is presented in [3].
The freeze bonds in this work were modelled using initially rigid cohesive bonds
i.e. each cohesive bond is generated into the initially continuous mesh at the onset
of fracture process as failure criterion at a bond point is reached. After initiation,
cohesive crack growth starts and can lead to true crack.
The approach to determine the cohesive force during cohesive crack growth is
adopted from [2]. The difference here is that the crack initiation and growth can
also start in compression due to shear or with pure shear. In this work, the bond
positions and further potential fracture sites were defined a priori. This choice was
made to study the effect of bonding.
Failure criterion F for sea ice by [5] was used to initiate the crack growth. Ac-
cording to the criterion, for the allowed stress states
σb τ2
F = + b2 − 1 < 0. (1)
σcr τcr
In the previous equation, σcr and τcr are the critical normal and shear stresses, re-
spectively, for the material. As a stress state with F ≥ 0 is reached at a bond point,
cohesive crack growth initiates. The cohesive crack growth follows linear softening
law.
SIMULATON RESULTS AND DISCUSSION
In Fig 2a, an example of an initial configuration of a punch-through test simula-
tion is shown. To generate the initial ice rubble, rigid blocks with random transla-
tional and rotational velocities were released underwater. After releasing, the blocks
were allowed to float due to the buoyant force to form the rubble mass. Freeze bonds
were generated after a quasi static rubble was achieved on the contacting parts of

(a)
(b)
Fig. 2: Initial configuration used in the example simulation: (a) a snapshot from the
simulation and (b) a sketch illustrating the dimensions, used coordinate system, and
boundary conditions (dashed line on rigid boundaries). In (b) w is domain width, h
keel depth, and d indentor width.
the surfaces of the blocks. Rubble was meshed after the freeze bond generation.
The indentor was rigid.
In Fig 2b, an illustration of the example simulation used here with boundary
conditions, the coordinate system, and the main dimensions is shown. In the sim-
ulation, the domain width w = 32 m, keel depth h = 4 m, and the indentor width
d = 4 m. A rigid boundary condition was used on the top surface of an elastic layer
presenting frozen layer of ice on top of the rubble.
For the ice blocks and the elastic layer, the material parameters were chosen
after first year sea ice with Young’s modulus E = 5 GPa and Poisson ratio ν = 0.3.
The critical normal and shear stress values for bonds were varied between σcr =
12.5 . . . 50 kPa and τcr = 6.25 . . . 25 kPa, respectively. Fracture energy value 15
J/m was used for bonds.
In Fig 3, a sequence of snapshots from an example simulation run with σcr = 50
kPa and τcr = 25 kPa is shown. Further, in Fig 4 the indentor force Fi for the same

yi = 0.1 mm yi = 1.0 mm
yi = 1.5 mm yi = 9.3 mm
-50 kPa 0 kPa 50 kPa

Fig. 3: Snapshots of a simulation with σcr = 50 kPa and τcr = 25 kPa with various
indentor displacments yi . The colormap in each snapshot refers to the stress com-
ponent σyy . Bond sites with true crack are presented with black lines. Subdomain
{−10 < x < 10, −5 < y < 1} m of the total simulation domain is shown. The grid
size in the snapshots is 1 × 1 m.
90
80
70
yi = 1.5 mm
60
y = 1.0 mm
i
50
Fi [kN]
40
yi = 9.3 mm
30
20
10
yi = 0.1 mm
0
0 5 10 15
yi [mm]
Fig. 4: Fi (yi )-graph from the simulation with σcr = 50 kPa and τcr = 25 kPa. The
markers in the Figure refer to the snaphsots shown in Fig 3.

simulation is plotted as a function of indentor displacement yi . Markers in Fig 4

correspond to the displacements in snapshots of Fig 3. As seen from F (yi )-graphs,
the snapshots are from time steps of the local peak values of Fi .
The first assumption for ridge keel shear strength τk is usually the assumption of
cylindrical (or in two dimensions, rectangular) failure of the keel under the indentor:
τk = Fi,M AX /2h, where Fi,M AX is the maximum indentor force. From the example
simulations τk values 6.4 to 11 kPa, highest strength giving the highest value, were
achieved. These values are on the range of measured values which are from 1.8 up
to 18 kPa [1]. The measured peak forces occur with fairly small values of yi . This
is partly due to the chosen boundary condition.
CONCLUSIONS
In the paper a combined finite discrete element method simulation of ridge keel
punch through tests was described. The model was applied to a small example case
of a ridge keel punch through test with various bond strengths.
REFERENCES
[1] J. Heinonen. Constitutive modeling of ice rubble in first-year ridge keel. Doc-
toral Thesis, TKK, 2004. VTT Publications 536, Espoo, Finland, 2004.
[2] J. P. Morris, M. Rubin, G. Block, and M. Bonner. Simulations of fracture and

fragmentation of geologic materials using combined fem/dem analysis. Inter-
national Journal of Impact Engineering, 33(1-12):463–473, 2006.
[3] A. Munjiza. The combined finite-discrete element method. Chichester, England.

John Wiley & Sons Ltd., 2004.
[4] A. Polojärvi and J. Tuhkuri. 3D discrete numerical modelling of ridge keel

punch through tests. Cold Regions Science and Technology, 56(1):18–29, 2009.
[5] H. L. Schreyer, D. Sulsky, L. B. Munday, M. D. Koon, and R. Kwok. Elastic-

decohesive constitutive model for sea ice. Journal of Geophysical Research,
111(C11S26), 2006.

COMBINED FINITE DISCRETE ELEMENT

SIMULATIONS OF A FLOATING ICE SHEET
FAILING AGAINST AN INCLINED STRUCTURE
Jani Paavilainen∗ , Jukka Tuhkuri and Arttu Polojärvi
Aalto University, School of Science and Technology
Department of Applied Mechanics
P.O.Box 14300, FI-00076 Aalto
e-mail: [email protected]
This paper presents some results of the 2D numerical simulations

of ice rubble formation process against an inclined structure. The
process was simulated with a two dimensional combined f nite
discrete element method (2D FEM-DEM). In the method, a level
ice sheet and its fracture were modelled by using the f nite ele-
ment method. The contact forces between the colliding ice blocks
broken off the level ice sheet were calculated by using the discrete
element method. Observations made from the simulated process
indicate that the pressure peaks on the structure are highly local-
iced even in two-dimensions.
INTRODUCTION
Failure of ice cover and rubble pile build-up against an inclined marine structure
(lighthouse, bridge pier, etc.) is a fragmentation process where an initially intact
ice sheet breaks into discrete ice blocks which then accumulate in the ice-structure
interface and thus affect the failure process. It is believed, that in order to understand
the failure of ice and ice loads on structures, the ice-structure interaction should
∗
corresponding author

be studied as a process and simulated. One method for simulating the ice failure
processes is the combined f nite discrete element method (FEM-DEM), see e.g. [4].
This paper presents some results of 2D numerical simulations of ice rubble for-
mation process against an inclined structure obtained by using the combined f nite
discrete element method. First the used method is brief y presented and then some
results of the simulations are examined.
COMBINED FINITE-DISCRETE ELEMENT METHOD
In the method, the ice sheet and its fracture are modelled by using the non-
linear Timoshenko beam and cohesive crack models. The contact forces of broken
ice in the rubble are calculated using the soft contact model. A comprehensive
presentation of the method can be found in [6].
Intact ice sheet and the pieces that are broken from it are modelled with non-
linear Timoshenko beam elements in a f nite element scheme.. The basic kinematic
assumption of the Timoshenko beam state that the cross sections remain plane but
not necessarily normal to the mid surface during deformation. In addition, small
strain assumption, large displacements and a viscous damping model are utilized in
the model.
Before fracture initiation, ice is assumed to behave as linear elastic and the linear
constitutive equation for homogeneous and isotropic material is used for def ning
the stresses. After fracture initiation, the above constitutive relation changes and the
stresses are def ned according to the cohesive crack model. The brittle-ductile and
the quasi-brittle models are the two basic types of cohesive crack models. Of these
two models, the quasi-brittle model is used as it appears to be more appropriate for
sea ice [1].
In addition to f nite elements, interacting ice beams are also composed of one
or more discrete elements. The contact detection and the calculation of the con-
tact forces are handled by using these discrete elements. The contact detection is
based on the algorithm presented in [5] and the forces between contacting discrete
elements are calculated with an elastic-viscous-plastic normal force model and in-
cremental Mohr-Coulomb tangential force model as described in [3].

Equations of motion for the whole system can be presented in the form
Mẍ = fint + fcol + fext (1)
where M is the diagonal mass matrix containing masses and moments of inertia of
the discrete elements, x contains the positions xi , yi and rotations θi of all discrete
elements in the system, fint are the internal forces obtained from the Timoshenko
beam model, fcol is a vector containing forces and moments from the collisions and
fext is a vector of external forces. The ẍ denotes the second time derivative of x.
In the method, the equations of motion (1) are solved with the central difference
scheme.
SIMULATIONS
Overview. A total of 60 FEM-DEM simulations of the process of ice rubble
formation against an inclined structure were made. The ice sheet thickness, ice–ice
friction, inclination angle of the structure were the main parameters varied. In the
simulations, a f oating ice sheet was pushed at its left end with a constant velocity
vp against an inf nitely rigid inclined structure. In the left end of the ice sheet, about
95–175 m from the structure, a viscous damping boundary condition was applied.
A snapshot of a simulation with ice sheet thickness h = 0.5 m is shown in Fig. 1. At
the end of a simulation, the number of discrete elements was between 1062–1927.
Results. The simulated process of ice failure and rubble pile formation in front
of an inclined structure was sensitive to initial conditions and to the values of the
simulation parameters. These sensitive process details include, for example, in-
stantaneous ice force on the structure and locations of individual ice blocks in the
rubble.
In the simulations, the rubble pile build-up process was continued for such a
long time that most of the extreme load events in the process are assumed to have
occurred. This is supported by fact that all the observed phenomena behind the ex-
treme load cases repeated itself in different simulations. Therefore, it was possible
to analyze the phenomena behind these different extreme load events in more de-
tail. From a simulation with ice sheet thickness h = 0.5 m, one of these extreme
load events was chosen for closer investigation. This event is shown in Fig. 1. The

time 5744.4 [s]

10
y [m]
−5
−10
−15
60 65 70 75 80 85 90 95 100 105 110
x [m]
Fig. 1: A snapshot of an extreme load event from a simulation with ice sheet thick-
ness h = 0.5 m. Structure (black), water (light gray), ice (ends of each ice block
light gray and ice in between white, individual elements dark gray). Red lines in
the rubble shows contact forces above 5% of the maximum contact force between
the ice blocks. Thicker line means higher contact force.
instant around this chosen extreme event was carefully analyzed with the help of
animations, snapshots and pressure plots from the simulation.
Fig. 1 illustrates how the force is transferred from the moving ice sheet to the
inclined structure. The ice sheet is hitting the rubble on top of the pile in about the
middle of the rubble pile. The red contact force chain coming from the ice sheet
changes direction downwards and spreads a bit at this point. Nevertheless, at this
shown extreme load event, the major contact force chain is seen clearly to hit the
structure in a narrow area.
From Fig. 2 it can be seen that pressure peaks on the inclined structure are highly
localized during this examined load event. Localized pressure peaks, high pressure

1
2
0.9
0
0.8
−2
0.7
−4
0.6
y/h
−6
0.5
−8
0.4
−10
0.3
−12
0.2
−14
0.1
5741 5742 5743 5744 5745 5746 5747 5748 5749 5750 5751
t [s]
Fig. 2: Nondimensional pressure distribution on the inclined structure around the

extreme load event at time t = 5744.4 s. Nondimensional pressure presented by the
color scale is def ned by p/pmax , where pmax is the maximum pressure on the shown
time interval. Only pressures above 10% of the maximum pressure are shown.
zones, are observed also during compressive intact ice sheet failure, see e.g. [2].
Simulation results thus show that the same kind of high pressure zone behavior is
present in the case of ice rubble pile and structure interaction and that it is present
also in two dimensional situations.
CONCLUSIONS
A 2D combined f nite discrete element method to model ice-structure interac-
tion problem was presented. In the method, ice sheet and its fracture was modelled
according to f nite element method by using non-linear Timoshenko beam and co-
hesive crack models. The contact forces between the colliding ice blocks were

calculated by using the discrete element method. Observations made from the sim-
ulated process indicate that the pressure peaks on the structure are highly localiced
even in two-dimensions.
REFERENCES
[1] Z. Bažant and Z. Murphy. Scaling of sea ice fracture – part I: Vertical penetra-
tion. Journal of Applied Mechanics, 69:11–18, 2002.
[2] J. Dempsey, A. Palmer, and D. Sodhi. High pressure zone formation during
compressive ice failure. Engineering Fracture Mechanics, 68(17-18):1961–
1974, 2001.
[3] M. Hopkins. Numerical simulation of systems of multitudinous polygonal

blocks. Technical Report 92-22, Cold Regions Research and Engineering Lab-
oratory, CRREL., 1992. 69 p.
[4] A. Munjiza. The combined finite-discrete element method. Chichester, England.

John Wiley & Sons Ltd., 2004. 333 p. ISBN 0-470-84199-0.
[5] A. Munjiza and K. Andrews. Nbs contact detection algorithm for bodies of
similar size. Int. J. Numer. Meth. Engng., 43:131–149, 1998.
[6] J. Paavilainen, J. Tuhkuri, and A. Polojärvi. 2d combined f nite-discrete element

method to model multi-fracture of beam structures. Engineering Computations,
26(6):578–598, 2009.

ONE DIMENSIONAL COMPRESSION OF SAND-

SILT MIXTURES USING 2D DEM
Nguyen Hop Minh, Yi Pik Cheng

University College London, Dept. of Civil, Environmental & Geomatic Engineering
Two-dimensional Distinct Element Method (DEM) is used to simulate

the deformational characteristics of mixtures with particle size in the
range of sand and silt subjected to one dimensional loading. The influ-
ence of the amount of small particles (fines) and their contact stiffness
are studied.
INTRODUCTION
One dimensional (1D) compression testing of an element of granular soil is
very common in geotechnical engineering due to its simplicity and its ability to
approximate deformation under many loading conditions. Soil specimen is placed
within a rigid steel ring without lateral deformation as the sample is compressed
vertically. Typical test results show that compression curves of soil samples pre-
pared at different initial densities converge on a same straight line at high stress
when plotted in (e,logv’) space; this straight line is called the Normal Compres-
sion Line (NCL) in the literature. It should be noted that the laboratory testings
are normally conducted with reconstituted soils, e.g. clean sands or silts. Experi-
mental data for intermediate graded samples, i.e. mixtures with particles from
both sand and silt size ranges, is less well known even though it is not uncommon
to find sand-silt mixtures in the field. Behavior of soils with intermediate grad-
ings can be complex and deviated from the general deformation framework suita-
ble for pure sands and/or silts1. For mixture of sand with finer particles, the initial

void ratio and the slope of the NCL are affected by percentage of fines 2,3. In this
study, two dimensional (2D) DEM is used to simulate 1D compression of mixture
of sand with silt (non-plastic fine) in order to investigate the effects of contact
characteristics and amount of fines on the macroscopic behavior of the mixture.
ONE DIMENSIONAL COMPRESSION OF 2D DEM ASSEMBIES

DEM samples are prepared by creating assemblies of 2D disks comprising dif-
ferent proportions of sand-size and silt-size particles, the Particle Size Distribution
(PSD) of which are given in Fig. 1a. Firstly, disks of different sizes are generated
inside a square box. The numbers of disks are specified based on the adopted PSD
and the percentage of uniform-size fine (silt) such that the initial void ratio (e, =
area of void/area of solid) is 1.0 to obtain very loose assembly. Summary on the
number of particles generated is given in Table 1. Note that sand in Table 1 speci-
fies particles with d > 0.4 mm, whereas silts are particles with d = 0.1 mm. During
sample preparation, isotropic compression is applied with a temporary particle
frictional coefficient of 0.0 and 0.5 to create the initially dense and loose samples
at about 100 kPa, measured by a measurement circle. The frictional coefficient is
then returned to 0.5 for the subsequent 1D compression. Here, compressions, ei-
ther isotropic or 1D, are carried out by applying velocity to every disks and walls.
The velocity is calculated from a specified strain rate and the current position of
particle (or wall). The process, however, is applied in stages. In each stage, at
first, all disks and walls move according to the applied velocity in a specified
number of time steps, then all the velocities are zeroed and the system are relaxed
until unbalanced force becomes negligible when the data is recorded. The simula-
tion is conducted by using PFC2D (ver. 4) 4. Interaction between particles is mod-
eled by using simple linear contact model 4. Input parameters are given in Table 2.
Fine (silt) content
Number of disks
0% (sand) 5% 30% 60% 80% 100% (silt)
Sand (big) 1465 1392 1023 583 292 0
Silt (small) 0 2864 17188 34377 45836 57295
Total 1465 4256 18211 34960 46128 57295
Table 1. Number of 2D DEM particles generated for different fine contents

Parameter Value
Particle density ( kg/m3) 2650
Particle friction, µ 0.5
Contact elastic modulus, E ( N/m2) 1x109
Particle stiffness ratio, ks/kn 1.0
Initial sample size (mm) 30x30
Wall friction, µwall 0.0
Wall stiffness, kwall 100kn (average ball stiffness)
Table 2. Input parameters for 2D DEM simulation
Figure 1b plots the variation of minimum initial specific volume, ini (=1+e),
obtained for dense samples against %fine. It is shown that firstly, ini reduces as
%fine increases from 0% to 30%, then ini increases again as %fines are in the
range of 30% to 100%. This tendency actually agrees with experimental result by
Lade&Yamamuro2, the trend of reducing ini with %fine indicates small particles
filling voids in between bigger particles; the increasing trend indicates big par-
ticles being pushed apart from each other and eventually become floaters inside
matrix made by small particles2. Compression curves for various %fines are given
in Fig. 2. The curves of the dense and loose samples converge on a unique NCL
at high stress similar to the experimental data of sands. It should be noted that ex-
perimentalists sometimes (not always) observed that the NCL of sand-silt mix-
tures depend on initial void ratio, which leads to non-convergence of the NCL 1.
This behavior, however, is not seen here.
100 1.24
a b
Specific Volume,  (=1+e)
80 1.22
1.20
60
% Finer
1.18
40 1.16
Sand 1.14
20
Silt 1.12
0 1.10
0.01 0.1 1 10 0 10 20 30 40 50 60 70 80 90 100
Particle size, mm Fine Content (%)
Fig 1. a) DEM particle size distributions b) Initial specific volume with % fines

1.25
Loose,
0%
Dense,
1.20 0%
Specific Volume,  ( = 1+e)

Loose,
5%
1.15 Dense,
5%
Loose,
30%
1.10 Dense,
30%
Loose,
1.05 60%
Dense,
60%
1.00 Loose,
100%
1.E+04 1.E+05 1.E+06 1.E+07 1.E+08 1.E+09
Dense,
Vertical Stress (N/m2) 100%
Fig 2. One dimensional compression curves for different fine contents

Slope of the NCLs (Cc) plotted against %fines in Fig. 3a shows similarity with
Fig. 1a. As the voids among sand particles are fully filled by silts for %fine =
20%~30%, the big and small particles together resist the applied load. The assem-
bly is less compressible, which can be reflected by a flatter NCL or smaller Cc.
The decrease of Cc with %fine from 0% to 10% indicates sand is losing its domin-
ance when fines are added to pure sand sample, although the overall behavior in
this range can still be justified as sand dominated. Similarly, the increase of Cc
with %fine from 40% to 100% reflects mixture behavior getting more and more
dominated by the presence of larger proportion of fines in the mixture.
0.35 6
a
Compression Index, Cc
0.3 5 b
Coordination Number, Z
0.25 4
3
0.2
2 Z_ini (dense)
0.15 Z_ini (loose)
1
Z_final (dense)
0.1 0 Z_final (loose)
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Fine Content, % Fine Content, %
Fig 3. a) Variation of Cc with % fine b) Growth of Z in 1D compression

1.20
5% (So)
5% (N)
Specific volume,  (= 1+e)

1.15
5% (St)
30% (So)
1.10 30% (N)
30% (St)
60% (So)
1.05
60% (N)
dense samples 60% (St)
1.00
1.E+04 1.E+05 1.E+06 1.E+07 1.E+08 1.E+09 1.E+10
Vertical stress (N/m2)
Fig 4. Compression curves for softer (So), stiffer (St) and normal (N) fines.
Figure 3b describes growth of the coordination number, Z, for dense and loose
samples during 1D compression. Initial gap of Z between dense and loose cases
seems to be erased at the end of compression. Note that for frictional uniform
disks, 3 ≤ Z ≤ 4 is considered as the minimum value so that the sample can act as
a stable system in a solid-like manner 5. In Fig. 3b, both initial and final coordina-
tion numbers (i.e. Z_ini and Z_final) for 5% and 10% fines are smaller than 4,
which are also lower than other cases. Further microscopic analysis, however,
showed that Z_ini calculated separately for sand and silt particles of 5% fine are
3.83 and 0.7 for dense case as well as 2.73 and 0.21 for loose case. Similarly,
Z_final = 6.75 (sand) and 2.16 (silt) for dense sample, = 6.54 (sand) and 1.92 (silt)
for loose sample. This clearly demonstrates that the behaviors of 5% fine samples
are dominated by sand particles. The fines contained inside void space have very
little contact with their neighbors. The authors also observed similar results in
other 3D simulations of gap-graded DEM assemblies 6. Figure 4 shows the effects
of changing particle stiffness of fines. Stiffer (St) and softer (So) fines are created,
respectively, by applying contact elastic modulus (E) of 5 times higher and 5 time
smaller than the value of E in Table 2 for silt particles. It seems there is no signifi-
cant change in the slope of NCLs for varied stiffness but the NCL is shifted to-
wards higher stress range for “St samples” and vice versa for “So samples”. The
original curve of normal (N) stiffness locates in between the two “St” and “So”

cases. Besides, the gap between “St” and “So” curves for 5% fine is quite narrow
because sand, not silt, dominates the behavior in this case. It then comes at no
surprise that the gap becomes wider for other cases of higher %fines in Fig. 4 as
the sample behavior is growingly dominated by silts particles with increasing
%fine over 30%.
CONCLUSIONS
Behavior of silty sands under 1D compression can be classified as sand-
dominated, sand&silt-dominated and silt-dominated depending on fine percen-
tage. Changing particle stiffness of fines in the mixture, hence, can have different
effects to the behavior of samples under these three different categories.
ACKNOWLEDGMENTS
We gratefully acknowledge the sponsor from an EPSRC grant (EP/F036973/1)
and the contribution of Messrs. Suliaman Zaheer and Ali Rahal in this article.
REFERENCES
1. P.W. Mayne, M.R. Coop, S. Springman, A-B. Huang and J. Zornberg, ‘Geo-
material behavior and testing’, Proc. 17th Intl. Conf. Soil Mechanics & Geo-
tech. Engrg., Alexandria, Egypt, 2009.
2. P.V. Lade, J.A. Yamamuro, ‘Effects of nonplastic fines on static liquefaction
of sands’, Canadian Geotechnical Journal, 34, No. 6, 918-928, 1997.
3. M.R. Coop, J.H. Atkinson, ‘The mechanics of cemented carbonate sands’,
Géotechnique, 43, No. 1, 53-67, 1993.
4. Itasca, Particle flow code in 2 dimensions (PFC2D) version 4, Minnesota,
USA, 2008.
5. M.V. Hecke, ‘Jamming of soft particles: geometry, mechanics, scaling and
isostaticity’, J. Physics: Condensed Matter, 22, 033101, 2010.
6. N.H. Minh, Y.P. Cheng, ‘One Dimensional Compression of Gap-graded Mix-
tures in DEM’, Proc. Intl. Sym. Geomech. & Geotech.: Micro to Macro IS
Shanghai, 2010 (in print).

USING DISCRETE ELEMENT METHODS TO

MODEL SOIL – MACHINE INTERACTION
Mustafa I. Alsaleh
Caterpillar Inc., Applied Research
Numerical modeling for soil-machine interaction has been of high in-

terest to Caterpillar engineers and analysts. Such models are essen-
tially used to build virtual models for the whole machine system and
the surrounding environment. Therefore, it is very important that two
soil-machine interactions can be captured accurately; in other words,
soil force prediction and material flows should be as close as possible
to real life measurements. Several numerical techniques have been de-
veloped and made available in the literature to help model dynamic
soil response; for instance, finite element method (FEM), discrete ele-
ment method (DEM) and mesh free methods (MFM) have shown
good success in capturing such behavior. Each of the above methods
has it’s own advantages and limitations, however. The DEM models
can be considered to be the most general techniques to model discrete
solid systems, such as soil masses. Caterpillar DEM code
(ROCKS3DTD) has been under development and use for the last two
decades. It has been used to model wide range of earth moving opera-
tions. The model formulations and material parameters have been
validated using laboratory and full machine tests data. This paper in-
tends to discuss the features of the code and will give overview for
some of the applications.

INTRODUCTION
Caterpillar Inc. researchers have been developing and using a discrete element
code for the last fifteen years; the code is called ROCKS3DTD. The code uses an
advanced computationally efficient contact detection algorithm and can handle
any particle shape specified by the user. The contact frictional and normal forces
are computed using the well-known Hertzian contact model for cohesionless ma-
terials. Additional algorithms are implemented to treat cohesive-like bonds when
modeling fine-grained materials. Cohesion is modeled as a cohesive pillar that
bond neighboring particles together; when the pillar’s strain threshold is exceeded
the bond between the neighboring particles is broken [1]. Recently, Cosserat rota-
tion has been added to the code kinematics along with particle shape indices de-
scribed by Alsaleh et. al [2]. This additional degree of freedom enabled the code
to capture more of the microstructural properties of the material being modeled
(angularity, size, spherecity etc.). Primarily ROCKS3DTD is used to model par-
ticulate force responses and material flows. It is capable of linking to full machine
models using in-house built codes for modeling machine dynamics, tire-ground
interaction, machine hydraulics, etc.
Another feature of ROCKS3DTD is its capability of interacting with tracked-
type tractors to pass proper forces to the machine through the track shoes. The
machine tools can be treated as either rigid and/or flexible bodies. The code is
parallelized to take advantage of mutli-threaded processors. The usability, simula-
tion speeds and accuracy of ROCKS3DTD compared to other commercial codes
have been found superior.
As acknowledged by many researchers, it is very challenging to obtain DEM
model parameters that best represent real materials, specially, when dealing with
fine-grained materials. Researchers and analysts at Caterpillar Inc. have been suc-
cessful in defining engineered procedures to map these micro quantities to mate-
rial physical and macro quantities. Both small scale laboratory testing and full
machine testing, are being utilized to develop micro-macro parameter mapping
functions. The particle size for instance, a very important DEM parameter, must
be chosen carefully. Choosing the particle size for a given model will always have
a great trade-off between simulation accuracy and computational cost; special at-
tention is given to this matter. The particle size distribution is established for a

given model in a way to ensure highest simulation accuracy at the lowest compu-
tational cost. This way, the model parameters (micro-mechanical properties such
as friction, stiffness, etc.) can be linked to macro properties to achieve better
physical representation.
ROCKS3DTD can predict the dynamic forces and flows of different discrete
system geometries under dynamic loading. As previously mentioned, the contact
parameters are micromechanical parameters that are very difficult to physically
measure and very challenging to evaluate since it is almost impossible to represent
the actual shape and size of real materials. Real materials are very complex to
mimic in terms of shape, size, and size distributions.
RESULTS AND DISCUSSIONS:

Earthmoving operations such as; digging, dozing, ripping dumping etc. had
been successfully modeled using ROCKS3DTD coupled with multi-body dynamic
models to model machine dynamics. Both rigid and flexible bodies are considered
whenever necessarily for machine bodies. Machine performance, payloads, struc-
tural life can be reasonably predicted using the coupled models, which helps the
engineers and analysts enhance/change machine design for better performance and
longer life. Models have been validated extensively using full machine test data
and sometime scaled laboratory models..
CONCLUSION
ROCKS3DTD has been successfully used across Caterpillar business units to
understand machine systems, when they interact with geomaterials, virtually. The
code is robust and general enough to model wide range of earthmoving opera-
tions. Materials parameters are calibrated using different techniques, such as sim-
ple laboratory testing.
ACKNOWLEDGMENTS
The author would like to thank the applied research department for the con-
tinuous support.

REFERENCES
1. Alsaleh MI and Hofestetter KW, 2007 “A Numerical Study for Geoma-
terials Shear Strength Components using Discrete Element Models”,
under review
2. Alsaleh, M. I., Alshibli, K. A., and Voyiadjis, G. (2006). “The Influ-
ence of Micro-Material Heterogeneity on Strain Localization in Granu-
lar Materials”, ASCE International Journal of Geomechanics Volume 6,
Issue 4, pp. 248-259

THE DEVELOPMENT AND APPLICATION OF

STOCHASTIC BLOCK SHAPE MODEL IN
CONTINUUM-BASED DISCRETE ELEMENT
METHOD
Tianping Liu, Jingjing Lu, Shihai Li, Baojuan Qiao

Institute of Mechanics，Chinese Academy of Sciences，Beijing 100190，China
Stochastic block shape model (SBSM) in continuum-based discrete element me-

thod (CDEM) is developed to simulate the progressive failure of landslide. The
paper compares SBSM with the limit equilibrium method (LEM) and the finite
element method (FEM) on analyzing the stability of landslide in detail, and gives
the advantage of the method in this field. Then it simulates different damage
forms of the earth-rock aggregate slope with different distribution of strength. It
also simulates the Liangshuijing landslide happened in November 2008 occurred
along the Changjiang River, and obtains the corresponding landslide phenomena.
INTRODUCTION
Due to the important role of landslide on huge disasters, the analysis of slope
stability is very important. The most popular methods to analyze the landslide are
the Limit Equilibrium Method and the Finite Element Method. Duncan summa-
rized LEM [1], and Chen Zuyu developed three—dimensional LEM [2]. Zhu
Dayong theorized the LEM by solving the equation of higher degree [3], and
Zheng Yinren introduced the application of the method of strength reduction by
FEM and compared the differences among the results based on different strength
criterion [4]. Jing Lanru concluded the character and function of FEM and DEM
[5]. All of these methods have their own limitations in analyzing the progressive

failure from continuous to discontinuous statement of the geological body. Aimed

at giving more reasonable simulation of the progressive failure, this paper devel-
ops SBSM [6-8] and concludes the differences among the LEM, FEM and SBSM,
which shows the advantage of SBSM.
Earth-rock aggregate slopes are widely distributed in China, which can be quite
suitably simulated by SBSM. According to it, this paper simulates typical slopes
by the SBSM, and analyzes the failure law.
PARISON BETWEEN SBSM AND LEM

In LEM, the safety factor K is the ratio of the sum of anti-sliding force and
sliding force (Formula 1), which assumes all the points on the potential sliding
surface are in the critical failure status simultaneously. With the same hypothesis,
there is almost no difference between the value of K calculated by SBSM and
LEM. Take the example of a common slope, the safety factor K is 1.084 calcu-
lated by LEM (Fig.1) and 1.086 calculated by SBSM, and the slope is stable.
n
∑f anti-sliding force, i
K= i =1
n (1)
∑f i =1
sliding force, i
However, the points on the sliding surface cannot fail simultaneously, which is
different from the assumption of LEM. In SBSM, on the potential sliding surface,
each point has its own K i which is the ratio of anti-sliding force and sliding force
of this point (Formula 2). Take the previous slope for instance (Fig.2), the points
on the potential sliding surface whose safety factors are less than 1.0 will fail,
which can change the whole stress distribution, and the slope will be not stable.
f anti-sliding force, i
Ki = (2)
f sliding force, i

potential sliding surface
2
1.8
Ki of every point
1.4
1.2
1 Ki = 1
0.8
K=1.084 0.6
0.2
0
0 10 20 30 40 50 60 70
X coordinate of point
Fig.1 The result of LEM Fig.2 The result of SBSM
FEM can only indirectly reflect the stress distribution of slope on the sliding
surface by K, and cannot analyze the progressive failure. However, SBSM can
directly give the comprehensive stress distribution of the slope, and simulates the
instability caused by progressive failure.
COMPARISON BETWEEN SBSM AND FEM

In FEM, the Transfixion of plastic strain is usually considered as the instability
criterion of landslide, which can give especially good result of soil slope (Fig.3).
However, it cannot be used in the simulation of the collapse of rock slope of rock
mass and earth-rock aggregate slope. In SBSM the occurrence and development
of fractures can be simulated easily, it can provide great advantage in rock slope
collapse simulation (Fig.4).

Fig.3 The result of FEM Fig.4 The result of SBSM
SIMULATION OF EARTH-ROCK AGGREGATE SLOPES

In order to build the earth-rock aggregate slope model, this paper randomizes
the regular geometric mesh, and sets the material parameter according to the soil-
rock ratio. Therefore, the earth-rock aggregate slope model is shown in Fig. 5-6.
Fig. 5 Earth-rock aggregate slope Fig. 6 Randomized geometric model

A typical slope is built as shown in Fig. 7. Below the sliding surface is the bed-
rock, the strength of which is very high. Above the sliding surface is the earth-
rock aggregate, of which the soil and rock distribute uniformly, and the rock vol-
ume is forty percent of the total volume. (Unit: m)

Earth-rock aggregate
Bedrock
Sliding surface
Fig. 7 Earth-rock aggregate slope model

There are three kinds of block material in the numerical model: rock mass, soil
and bedrock. Thus, there should be three kinds of interface parameters: rock
mass, soil and bedrock. If the interface is between different kinds of blocks, its
parameter values should take the lower ones. The detail parameter values are
shown in Table 1.
Variable Density Interface Young Poisson Tensile Cohesion Friction
modulus ratio strength angle
/kgm-3 g/mt-2 E/pa υ T/pa C/pa /°
Rock mass
Rock mass 2.1×103 9.8 3×1010 0.3 2×103 2×104 30
interface
Soil 1.8×103 9.8 5×109 0.3 Soil interface 1×103 5×103 20
Bedrock 2.7×103 9.8 5×1010 0.3 Bedrock interface 5×106 2×106 45
Table 1 Parameter values

The progressive failure displacement nephogrames are shown in Fig. 8-11.
Fig. 8 Displacement nephogram Fig. 9 Displacement nephogram

at the 60,000th step at the 200,000th step

Based on various combination simulations, as is shown above, to the slope
without weak soil layers, it can be concluded that the landslide happens after the
cracks appearing between soil and rock mass rather than along some sliding sur-
face integrally.
Considering that there may be a weak soil layer between the earth-rock aggre-
gate and bedrock, this paper also simulates this kind of condition. Using the same
model shown in Fig. 7, a weak soil layer is assumed between the earth-rock ag-
gregate and bed rock, and the detail parameter values are shown in Table 2.

Variable Density Interface Young Poisson Tensile Cohesion Friction

modulus ratio strength angle
/kgm-3 g/mt-2 E/pa υ T/pa C/pa /°
Rock mass
Rock mass 2.1×103 9.8 3×1010 0.3 2×103 3×104 40
interface
Soil 1.8×103 9.8 5×109 0.3 Soil interface 1×103 5×103 20
Soft soil interface 1×103 1×103 15
Bedrock 2.7×103 9.8 5×1010 0.3 Bedrock interface 5×106 2×106 45
Table 2 Parameter values

The progressive failure displacement nephogrames are shown in Fig. 12-15.



As is shown above, to the slope with weak soil layers, it can be concluded that
the landslide happens along the weak layer together with the increasing of cracks
in the earth-rock aggregate, which is quite different from the previous conclusion
without weak layers. Thus, the progressive failure of the earth-rock aggregate
slope can be obtained.
SIMULATION OF THE LIANGSHUIJING LANDSLIDE

The Liangshuijing landslide is located in the Yunyang county, slope zone on
the right bank of the Changjiang River (Fig.16), which has great influence on peo-
ple and shipping. We have already got sufficient database of the slope by the geo-
logical survey and the landslide monitoring.
Landslide boundary
Sliding direction
Fig. 16 Liangshuijing landslide

The Liangshuijing landslide is earth-rock aggregate slope, which is very suit-

able to be simulated by SBSM .One of the most important phenomenons of the
slope is that there is a large fracture on the landslide boundary and a bulging frac-
ture in the middle of the sliding mass. The paper gives corresponding simulation
of the progressive failure of the slope that fractures occurred in the stable slope
after the water level was up to 172 meters together with the strength of the geo-
logic body reduction (Fig. 17- 19).
Fig.17 Stable slope
Fig. 18 Fractures occurring

Fracture
on the landslide boundary
Bulging fracture
in the middle of the sliding mass
Fig. 19 Fractures phenomenon up to now

CONCLUSION
This paper develops SBSM to analyze the landslide. And the results can be
summarized as follow:
(1) SBSM can well simulate the progressive failure of the landslide.
(2) SBSM can simulate the instability caused by progressive failure, which
cannot be analyzed by LEM.
(3) SBSM can easily simulate the occurrence and development of fractures,
and provide great advantage in rock slope collapse simulation, which is dif-
ficult for FEM.
(4) This paper simulates the Liangshuijing landslide, and obtains the corre-
sponding landslide phenomena, which has been regarded as important deci-
sion-making basis in the governance of the landslide.
ACKNOWLEDGMENTS
Gratefully acknowledge the support from National Natural Science Foundation
of China (Grant No.10632100), the project of important research aspect of Chi-
nese Academy of Sciences (KJCXZ-YW-LDI) and the Chinese National 973 Pro-
ject (Grant No.2010CB731500).
REFERENCES
1. Duncan J M, ‘State of the Art: Limit Equilibrium and Finite-Element Analysis
of Slopes’, Journal of Geotechnical Engineering, No.22/7,577～596, 1996．
2. Chen Zuyu, MI Hongliang, WANG Xiaogang, ‘A three—dimensional limit
equilibrium method for slope stability analysis’, Chinese Journal of Geotech-
nical Engineering, No.23/5, 525-529, 2001．
3. Zhu Dayong , QIAN qihu, ‘Rigorous And Quasi-rigorous Limit Equilibrium
Solutions of 3D Slope Stability And Application to Engineering’, Chinese
Journal of Rock Mechanics and Engineering, No.26/8, 1513-1528, 2007．

4. Zheng Yingren, Zhao Shangyi, Deng Chujian, et al. ‘Development of Finite

Element Limit Analysis Method and Its Applications in Geotechnical Engi-
neering’, Engineering Science, No.12/8, 39-61, 2006．
5. Jing Lanru, ‘A review of techniques advances and outstanding issues in nu-
merical modeling for rock mechanics and rock engineering’, Internati0na1
Journal of Rock Mechanics and Mining sciences, No.4O/3, 283-353,2003．
6. Liu Tianping, Li Shihai, Liu Xiaoyu, ‘A Continuum-based Discrete Element
Method of Stochastic Block Shape for Analyzing Stability of Cataclastic Rock
Mass Slopec, Beijing DEM’08, 2008.
7. Li Shihai, Zhao Manhong, ‘A New Numerical Method for DEM-Block and
Particle Model’, International Journal of Rock Mechanics and Min-ing Sci-
ences, No.41/3, 436, 2004.
8. Liu Bingshan, ‘The researh of Maoping landslide on the failur mechnism and
the law of evolution’, Doctor’s Dissertation institute of mechanics CAS 2007
9. Tian Zhennong, ‘Discrete Element Method Simulation and Experimental
Study of Blasting Effect on Rock Mass’, Doctor’s Dissertation institute of me-
chanics CAS 2006

ADAPTIVE DISCRETE/FINITE ELEMENT

COUPLING FOR ROCK CUTTING PROCESS
SIMULATIONS
Carlos Labra1, Jerzy Rojek2, Eugenio Oñate1

1
International Center for Numerical Methods in Engineering, Spain
2
Institute of Fundamental Technological Research, Poland
In the last years, the Discrete Element Method (DEM) has arisen as an
interesting alternative to the continuum based methods for the
simulation of problems characterized by strong discontinuities.
Nevertheless, the high computational cost of this technology hinders
its application to practical problems. For this reason new coupling
techniques are developed in order to obtain efficient numerical models
[4]. The possibility of the coupling between DEM and the Finite
Element Method (FEM) can be a good solution in some cases, but the
definition of each domain can be predefined in the generation of the
model. A new adaptive coupling technique between DEM and FEM is
presented, in order to avoid the definition of the DEM domain and
minimizing the computational cost. The idea is to use DEM to analyze
just the region of the solid wish presents fracture or damage, whereas
FEM is used in the rest of the domain. These region or sub-domain
can change during the simulation process, resulting in a progressive
modification of the coupling definition. This technique is applied to
geomechanical problems, such as rock cutting processes [5]. In this
case, the rock is modelled using DEM/FEM coupling, and the tool as
rigid body or FEM, depending of the analysis required.

INTRODUCTION
The DEM based on spherical and cylindrical particles is widely recognized as a
suitable tool for the modelling of geomaterials [1, 2], advantages of this method
are clearly seen in modelling problems characterized by the strong discontinuities
e.g. rock fracture during excavation. The main problem in the use of DEM for
large simulations is the computational cost involved. In the present work, a
coupling scheme of the DEM with the finite element method is presented.
The requirement of predefine the sub-domains with each methods is solved using
an adaptive algorithm for the definition of the DEM sub-domain. This allows the
use of DEM just in the areas whit higher values of stress, where fracture can
occur.
DISCRETE ELEMENT FORMULATION

The discrete element model assumes that material can be represented by an
assembly of rigid particles interacting among themselves. The overall behaviour
of the system is determined by the cohesive/frictional contact laws. The contact
law can be seen as the formulation of the material model on the microscopic level.
Cohesive bonds can be broken, which allows us to simulate fracture of material
and its propagation. Basic formulation of the discrete element formulation using
spherical or cylindrical particles was first proposed by Cundall and Strack [1].
The present work use the formulation developed in CIMNE by Rojek et al. in
[3], where the translational and rotational movement is described by means of the
standard equations of rigid body dynamic.
x D = FD
M D &&
(1)
& D = TD
J Dω
where xD represent the position vector of the particle, ωD its angular velocity, MD
a diagonal matrix with the particle mass in the diagonal, JD a diagonal matrix with
the particle moment of inertia, and FD and TD the resultant forces and moment
about the particle central axes, respectively, considering the sum of all forces and
moment applied to the particle due to external loads, contact interaction with the
neighbour particles, and contact interaction with other obstacles. These equations
are integrated considering the central difference scheme.

DISCRETE/FINITE ELEMENT COUPLING SCHEME

One of the main problems of DEM is the computational cost of the
simulations. In a real case, a large number of particles are required, and the
analysis of each contact between them, wish leads to long computational time. In
order to solve the situation different solutions can be considered. The
combinations of DEM with other simulation technologies, as FEM, allow an
efficient to create efficient models, minimizing the computational cost, taking
advantages of each method, and avoiding their weak sides. The idea is to use the
DEM in the sub-domain (ΩD) where the fracture occurs, and the FEM can be used
in the other part, where the behaviour can be represented with a continuum based
scheme.
Finite element formulation. In the case of the finite element method, the so-
called explicit dynamic formulation is used. The explicit FEM is based in the
solution of discredited equations of motion written in the current configuration in
the following form where MF is the mass matrix, rF the vector of nodal
displacements, Fext and Fint the vectors of external loads and internals forces,
respectively. Similarly to the DEM algorithm, the central difference scheme is
used for the time integration of Eq.(2).
x F = FFint + FFext
M F && (2)
Coupling Algorithm. It is assumed that the DEM and FEM can be applied in
different sub-domains of the same body. The DEM and FEM sub-domains overlap
each other. The common part of the sub-domains (ΩDF) is the part where both
discretization types are used with gradually varying contribution of each
modelling method, similarly to the idea used by Xiao and Belytschko in [6]. The
coupling of DEM and FEM sub-domains is provided by additional kinematical
constraints. Interface discrete elements are constrained by the displacement field
of overlapping interface finite elements. Additional kinematic relationships can be
written jointly in the matrix notation as follows:
χ = rD − N rD (3)
where N is the matrix containing adequate shape functions. The set of equation of
motion for the coupled system with the penalty scheme is

⎡M F 0 0 0 ⎤ ⎧ &&rF ⎫ ⎧FFint − FFext + NT k DF χ ⎫

⎢ ⎥⎪ ⎪ ⎪ ⎪
⎢ 0 M DU 0 rDU ⎪ ⎪
0 ⎥ ⎪&& FDU ⎪
⎨ ⎬ =⎨ ⎬ (4)
⎢ 0 0 M DC 0 ⎥ ⎪&&rDC ⎪ ⎪ FDC − k DF χ ⎪
⎢ ⎥⎪ ⎪ ⎪ ⎪⎭
⎣ 0 0 0 J F ⎦ ⎩ω D ⎭ ⎩
& TD
where the subscript DC and DU are the constrained and unconstrained discrete
elements, respectively, and kDF the vector of the discrete penalty functions.
Adaptivity of the DEM/FEM interface. In a real case, the zone wish requires
the use of DEM can change during the simulation process. With the standard
DEM/FEM coupling, both sub-domains are predefined during the discretization
process. This requires knowing where the fracture will appear and the use of DEM
even when the solid have a linear elastic behaviour. In order to make more
efficient the use of the DEM during the simulation, an algorithm for the change of
the sub-domains is used. The idea is start the simulation using finite elements.
When some area with a stress value higher than a predefined limit state is
achieved, the FEM is replaced with DEM and a new DEM/FEM interface is
defined. The progressive changes of the simulation method make more efficient
these of both methods.
The criteria for the change of simulation method is defined with a predefined
limit stress value (σ*) in the finite elements, as shown Eq. (5).
σ >σ *
(5)
When a change of domain occurs, the kinematic variables of the finite element
should be projected into the new discrete elements, and the behaviours of both
methods should be equivalents. In order to ensure the geometrical compatibility in
the discrete elements used progressively during the simulation, a discretization of
the full domain with cylindrical/spherical particles is generated and stored for its
posterior use. The discrete elements activated in a change of domain are defined
by the elements to be changed, and the particles in the new overlap sub-domain
required in the coupled scheme. The particles in the new overlap region are
selected using a distance function [7] based in the new boundary elements
generated in the FEM sub-domain.

SIMULATION OF ROCK CUTTING PROCESSES

The simulation of different rock cutting processes has been simulated using the
adaptive discrete/finite elements coupling. The coupled model has been used in
the rock subjected to fracture, while the cutting tool is discretized as a rigid body
or with finite elements, depending of the analysis required.
(a) t = 0.00 s (b) t = 0.06 s (c) t = 0.12 s
Fig 1. Rock cutting process with a disc cutter. Evolution of coupling interface.
In the Figure 1 the evolution of a rock cutting process with a disc cutter is
shown. In the Figure 2, a comparison of the normal force obtained in the case of
use full discrete elements in the rock, the fixed coupling scheme and the adaptive
algorithm is depicted.
The difference of the average value is negligible, and the computational cost
involved in the adaptive coupling case 40% lower than the standard coupling
scheme.
Fig 2. Comparison of normal force in rock cutting process with the coupling
schemes.

CONCLUSION
The coupling scheme using discrete and finite element methods can be an
interesting approach for the simulation of fracture in geomaterials, minimizing the
computational cost. The adaptively of the DEM/FEM interface can improve the
efficiency in the use of both methods. The technique has been used in the
simulation of rock cutting processes. The obtained results are consistent with the
values obtained in a standard discrete elements simulation.
REFERENCES
1. P.A. Cundall and O.D.L. Strack. A discrete numerical method for granular
assemblies, Geotechnique, 29,47-65, 1979.
2. C. Labra, J. Rojek, E. Oñate, F. Zárate. Advances in discrete element
modelling of underground excavations, Acta Geotechnica, 3, 317-322, 2008.
3. J. Rojek, E. Oñate, F. Zárate, J. Miquel. Modelling of rock, soil and granular
materials using spherical elements. Proc. of the 2nd European Conference on
Computational Mechanics EC CM-2001, Cracow, Poland, 26-29 June 2001.
4. E. Oñate, J. Rojek. Combination of discrete element and finite element
methods for dynamic analysis of geomechanics problems. Comput. Meth.
Appl. Mech. Engrg., 193, 3087-3128, 2004.
5. J. Rojek, E. Oñate, H. Kargl, C. Labra, J. Akerman, U. Restner, E. Lammerr,
F. Zarate. Prediction of wear of roadheader picks using numerical
simulations, Geomechanics and Tunnelling, 1, 47-54, 2008.
6. S.P. Xiao, T. Belytschko. A bridging domain method for coupling continua
with molecular dynamics. Comput. Meth. Appl. Mech. Eng., 193, 1645-1669,
2004.
7. R. Elias, M. Martins, A. Coutinho. Simple finite element-based computation
of distance functions in unstructured grids. Int. J. Numer. Meth. Engng., 72,
1095-1110, 2007.
8. J. Rojek, E. Oñate. Multiscale analysis using a coupled discrete/finite element
model. Interaction and Multiscale Mechanics, 1, 1–31, 2007.

FAILURE MECHANISM ABOUT RSA UNDER EQUAL

STRESS BOUNDARY CONDITIONS
Fan yongbo, Li shihai, Feng chun, Hou yuefeng
Institute of mechanics, Chinese Academy of Sciences, The key laboratory of Environmental
mechanics
RSA (Rock and Soil Associate) is inhomogeneous and discontinuous, the stress
distribution inside of materials is uneven while external forces are applied on them. We
carried out numerical calculation with CDEM (Continuum-based Discrete Element
Method) procedure, the result is that the deformation of RSA is inhomogeneous under the
equal stress boundary conditions and accurately reflect the internal structure of materials,
at the same time, we carried out the flexible triaxial test (it can provide the equal stress
boundary conditions by flexible head and flexible base), the result of experiment is in
good agreement with the numerical simulation result.
INTRODUCTION
The landslide involving rock and soil Associate (abbreviation, RSA) [1]
account for a large proportion of geological hazards, eighty percent of landslide is
this kind in Three Gorges Reservoir Area, such as Xintan landslide、Huanglashi
landslide etc[2][3]. There are six RSA landslide among Seven large-scale landslide
along Sichuan-Tibet Highway in Qinghai-Tibet Plateau, accounting for eighty-six
percent of the total landslides[4]; The landslide involving Rock and Soil Associate
has caused 20 billion yuan loss to the state, seriously affecting the national
economy、people's normal life safety. Therefore, the systematic research for this
type of landslide has a special theoretical and practical value.
MECHANICAL RESPONSE OF MATERIALS UNDER DIFFERENT
LOADING CONDITION
RSA is inhomogeneous and discontinuous, under external loading, because
of the differences in stiffness and strength about soil and rock, the sample will
appear inhomogeneous deformation, top and bottom of the sample will be uneven,
in other words, the stress state of each point is different, according to the theory of
rigid loading, the stress state of each point is the same, This is not consistent with
the actual situation. Triaxial compressive test machine is based on the idea of
uniform loading, in other words, it is loading with equal stress boundary condition,
it can study deformation and failure discipline about inhomogeneous and
1
discontinuous materials.
RESEARCH OF DEFORMATION CHARACTERISTICS ABOUT RSA
UNDER TOTALLY FLEXIBLE BOUNDARY LOADING
SAMPLE PREPARATION
According to the standard of soil test method（GB/T50123-1999）, Analysis
of particles test is carried out, there are five type of particle diameter:
d<0.074mm 、 0.074<d<0.25mm 、 0.25<d<0.5mm 、 0.5<d<1mm 、 1<d<2mm,
physical map and the content is as illustrated in Table 1.
Table 1 Range and content of particle diameter
Diameter of the soil 1<d<2 0.5<d<1 0.25<d<0.5 0.074<d<0.25 d<0.074
particles（mm）
content（%） 24.4 26.8 17.9 22.0 8.9
All the samples are remodeling samples, with all the same initial conditions
(including each grain group、compaction characteristics), with the different
confining pressure (0、0.2MPa、0.3MPa、0.4MPa), the same lithology (granite)、
the same size.
THE RESULTS OF TRIAXIAL COMPRESSIVE TEST WITH FLEXIBLE
BOUNDARY LOADING
At first, we give a brief description about the triaxial stress and platen press,
the triaxial stress is the contact stress between the flexible head and the sample;
the platen stress is the contact stress between the rigid plate and the top of the
flexible head, in other words, the platen stress is the traditional rigid contact stress,
this stress can’t reflect the heterogeneity of the RSA, the difference between the
triaxial stress and he platen stress is clear. The following figure 1 is the curve of
triaxial stress-strain.
a The curve of triaxial b The curve of triaxial c The curve of triaxial

stress-strain (0.2MPa) stress-strain(0.3MPa) stress-strain(0.4MPa)
Fig. 1 The curve of triaxial stress-strain
2
Table2 Bearing capacity about RSA under different confining pressure

The number of the Confining pressure Loading platen stress triaxial stress
sample （Mpa）（KN）（Mpa）（Mpa）
RSA-1 0.2 66.2 0.94 0.653
RSA-2 0.3 89.5 1.27 0.94
RSA-3 0.4 113.5 1.55 1.086
EQUAL STRESS BOUNDARY CONDITION LOADING BASED ON
CDEM
SIMULATION MODEL
We have conducted the numerical simulation with CDEM, according to the
symmetry of the model, we can take one-forth of the whole model, the numerical
model is illustrated in fig. 2. the top and bottom of the model (①face、②face)with
x and y direction constraint, symmetry plane (③face 、 ④face) with normal
constraint. Appling axial loading at the top and bottom of the model, and applying
the confining pressure around the model.
②
clay
granite
60cm
Size of granite：
③ ④ length：10cm
wide：10cm
high：20cm
①
15cm 15cm
Fig. 2 Numerical model

MECHANICAL PARAMETERS
By uniaxial compression test, we can obtain the physical and mechanical
3
parameters about soil and rock, as illustrated in table 3, using Mohr-Coulomb

model to have a numerical simulation.
Table 3 The parameter of material
poisson's Young's Tensile Internal
density Cohesion
material ratio modulus strength friction angle
kg/m3 / GPa MPa MPa °
clay 1580 0.35 0.015 0.003 0.015 25
granite 2500 0.25 38 3 10 45
NUMERICAL SIMULATION
We select the top and middle of the face to monitor z-direction stress、
z-direction strain、z-direction displacement, the monitoring points are illustrated in
the fig. 3.
1 7
2
8
3 9
4
5 11 10
6 12
（a）the top of the face （b）the middle of the face

Fig. 3 The map of arrangement of monitoring points
ANALYSIS OF THE RESULT
loading-displacement curve
0.02 0.00 0.00
0.00
-0.01 -0.03
-0.02
-0.02 测点 7 -0.06 测点1
-0.04 测点 8 测点2
测点1 -0.09
竖向位移
-0.03 测点 9 测点3
-0.06 测点2
测点10 测点4
测点3 -0.04 -0.12
-0.08 测点11 测点5
测点4 测点6
测点5 测点12 -0.15
-0.10 -0.05
测点6
-0.12 -0.18
-0.06
0.3 0.6 0.9 1.2 1.5
-0.14 0.2 0.4 0.6 0.8 1.0 1.2
0.2 0.4 0.6 0.8 1.0 1.2
c、top of the face（0.3MPa）
a、top of the face（0.2MPa） b、middle of the face（0.2MPa）
4
0.00 0.00 0.00

-0.01 -0.01
-0.05
-0.02 测点 7 -0.02
测点1 测点 7
测点 8 -0.10 测点 8
-0.03 测点 9 测点2
测点3 -0.03 测点 9
测点10 -0.15
-0.04 测点4 测点10
测点11 -0.04
测点5 测点11
-0.05 测点12
-0.20 测点6 测点12
-0.05
-0.06
-0.25 -0.06
0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 0.6 0.9 1.2 1.5 1.8 2.1 0.6 0.9 1.2 1.5 1.8 2.1
d、middle of the face（0.3MPa） e、top of the face（0.4MPa） f、middle of the face（0.4MPa）

Fig.4 The curve about loading-displacement under different confining pressure
We can see from the fig. 4 that the vertical displacement of sample is
gradually increasing with the loading increasing; the amount of sample swelling
about middle and top of the face is increasing with the confining pressure is
increasing, and the displacement of top face is much higher than the
corresponding point in the middle face, and more, the displacement of the central
measuring point is higher than the point at the edge of the sample, this is agree
with the experiment result.
stress-strain curve
1.6x106 2.1x106 测点1
2.5x106 测点2
1.4x106 1.8x106 测点 7 测点3
测点 8 2.0x106 测点4
1.2x106 1.5x106 测点 9 测点5
1.0x106 6
测点10 1.5x106 测点6
1.2x10 测点11
8.0x105 测点1 测点12
测点2 9.0x105 1.0x106
6.0x105 测点3 6.0x105
4.0x105
测点4 5.0x105
测点5 3.0x105
2.0x105 测点6 0.0
0.0 0.00 0.07 0.14 0.21 0.28 0.35 0.42 0.49
0.00 0.03 0.06 0.09 0.12 0.15
0.00 0.08 0.16 0.24 0.32
a、top of the face（0.2MPa） b、middle of the face（0.2MPa） c、top of the face（0.3MPa）

3.0x106 3.2x106 3.6x106
测点 7 测点 7
2.8x106 3.2x106
2.5x106 测点 8 测点 8
测点 9 2.4x106 2.8x106 测点 9
6
2.0x10 测点10 2.0x106 2.4x106 测点10
应力（Pa）
测点11 测点11
1.6x106 2.0x106
应力
1.5x106 测点12 测点12

测点1 1.6x106
1.2x106
测点2
1.0x106 8.0x105 1.2x106
测点3
4.0x105 测点4 8.0x105
5.0x105
0.0 测点5 4.0x105
0.0 测点6
0.00 0.05 0.10 0.15 0.20 0.00 0.08 0.16 0.24 0.32 0.40 0.48 0.00 0.05 0.10 0.15 0.20
d、middle of the face（0.3MPa） e、top of the face（0.4MPa） f、middle of the face（0.4MPa）

Fig. 5 The curve about stress-strain under different confining pressure
We can see from the figure 5(a)、5(b), the peak-stress of each point is
different, the value of test point at the edge of sample is the lowest, about 0.4MPa,
the value of test point at the middle of sample is about 0.7MPa, on the whole, the
5
peak-stress of the sample is about 0.7MPa. Similarly, we can see from the figure
5(c)、5(d)、5(e)、5(f) , the peak-stress of the sample is about 0.9MPa and 1.1MPa
when confining pressure is 0.3MPa and 0.4MPa. The test point at the middle of
the sample appears strain hardening, but the test point at the edge of the sample
appears strain softening; The stress value at the middle of the sample is straight up,
and the value of strain is zero, so it is reasonable that we study the heterogeneity
of the RSA under equal stress boundary condition. All the result of numerical
simulation has a good agreement with the experiment result.
CONCLUSIONS
By triaxial compressive test with flexible boundary loading testing machine
and CDEM numerical simulation, we can obtain the following results:
(1)We can study the mechanical behavior of RSA or fractured rock under
flexible boundary loading testing machine providing fully flexible boundary
conditions.
(2)Under fully flexible boundary conditions, the top of the RSA will be
inflated, but the bottom of the RSA will be depressed, this phenomenon just
reflects the inhomogeneous characters of RSA.
(3)The result of numerical simulation has a good agreement with laboratory
experiment result.
REFERENCES
[1]Li Xiao, Liao Qiulin, He Jianming. In - situ Tests and Stochastic Structural
Model of Rock and Soil Aggregate in the Three Gorges Reservoir Area [J].
International Journal of Rock Mechanics and Mining Sciences, 2004, 41 (3): 494.
[2]Changjiang Water Resources Commission Comprehensive Survey. Exploring
report of geological engineering during treatment and planning stage about Bai
yi’an landslide in Fengjie county.1992.12
[3]Yin yueping, Zhang guiru, Chen baosun. The study of cause mechanism about
loosing deposit of new Wushan county in Three Gorges Reservoir, Journal of
Engineering Geology, 2000, 8(3):265-271
[4]Research about resurrection mechanism and trend forecast of landslides[PhD
thesis]. Beijing: Graduate School of Chinese Academy of Sciences, 2004
6
USING DEM FOR THE ASSESSMENT OF K0 IN

SOILS
Daniel Barreto
Edinburgh Napier University, School of Engineering and the Built Environment
This paper explores the possibility of using DEM for gaining further
understanding of the evolution of K0 in soils. In spite of the inherent
approximations involved when using numerical methods, the results
appear to lie within the scatter produced by experimental results. In
addition, micro-scale analyses of the evolution of soil fabric demon-
strate that the evolution of K0 is related to the observed macro-scale
response.
INTRODUCTION
The in-situ state of stress in soils is of major significance in geotechnical prob-
lems. While the vertical effective stress (’v) is easily estimated from the soil pro-
file, the horizontal stress (’h) is highly dependent on depositional history. Nor-
mally, the at-rest coefficient (K0 = ’h /’v) is used. Numerous laboratory and in-
situ tests can be performed to estimate the magnitude of these stresses.
As pointed out by Mayne & Kulhawy (1982), numerous researchers have ad-
dressed this issue achieving varying degrees of success as many factors are in-
volved in the experimental K0 data that cannot be quantitatively analysed. For ex-
ample: different test methods, equipment and personnel, sample disturbance ef-
fects, time and aging effects, inherent anisotropy, etc. It is expected that an addi-
tional difficulty in such analysis is provided by the particulate nature of soils.

The Distinct Element Method (DEM) was proposed by Cundall & Strack
(1979) and uses and explicit finite difference approach to solve the dynamic equi-
librium of individual particles. DEM has been extensively used in order to simu-
late laboratory testing of soils. Although some difficulties have been found possi-
bly due to particle shape effects (i.e. Ji et al, 2009; Barreto, 2009), DEM results
have been validated with experimental results on spherical particles (e.g. Cui et al
2007).
In this study, a set of three-dimensional DEM simulations on assemblies of
spheres were used to study the evolution of the K0 value. Consequently, the ad-
verse effects caused by different testing methods, personnel, as well as those re-
lated to sample disturbance, time effects, etc. can be isolated. Furthermore, the
effects of inherent anisotropy were quantified.
STRESS HISTORY AND THE K0 COEFFICIENT

2000
OCR = 1
Vertical effective stress (kPa) - 'v
1600
Virgin loading
1200 Unloading
OCR = 2
800
OCR = 3 Re-loading
OCR = 4
400
0
0 400 800 1200
Horizontal effective stress (kPa) - 'h
Fig 1. Typical stress path for soil under K0 conditions

Consider a homogeneous soil deposit with horizontal ground surface. Fig. 1 il-
lustrates a typical stress path. The virgin loading portion is associated with sedi-
mentation and normal consolidation. Any reduction (unloading) in the effective
overburden stress caused by erosion or excavation results in over-consolidation of
the soil. Clearly, during unloading the over-consolidation ratio (OCR = ’vmax/’v)

has a significant effect on the value of K0. If the soil is reloaded the stresses will
evolve as shown in Fig. 1.
Several expressions can be found in the literature to calculate the value of the
K0 coefficient. Mayne & Kulhawy (1982) collected and analysed a database com-
prising tests from over 170 soils and extended Jaky’s relationship (K0 = 1 - sin)
in order to estimate the at rest coefficient during virgin compression, unloading
and subsequent reloading:
 OCR  3  
K 0  1  sin  
OCR
 
1sin    1 
 (1)
 OCRmax  4  OCRmax 
where  is the angle of shearing resistance and the OCR is calculated as:
 v'
OCR  max
 '
vcurrent
(2)
1.0
K0 = 1 - sin 
0.8 r2 = 0.39
0.6
DEM
K0
0.4
0.2 Laboratory and

in-situ measurements
0.0
0.0 0.2 0.4 0.6 0.8 1.0
sin 
Fig 2. Observed relationship between K0 and sinfor cohesionless soils during

virgin compression (modified after Mayne & Kulhawy, 1982)

The accuracy of Equation (1) can be assessed if only the data for granular ma-
terials and virgin loading is considered. Fig. 2 illustrates that the scatter of the re-
sults is significant. Note that for over-consolidated conditions the reliability of the
expression reduces due to the even more limited data available.
With the recent advances in computational systems a vast amount of informa-

tion can be collected without major difficulty if DEM is used. Fig. 2 also shows
the results of two DEM simulations presented in the following section. Clearly the
numerical results lie within the scatter of the experimental data available. There-
fore, the potential of using numerical methods to provide further understanding
into the particle scale interactions affecting the K0 value is evident.
DEM SIMULATIONS
A series of simulations involving different stress histories were performed on

polydisperse assemblies of spheres using periodic boundaries. The behaviours of
two assemblies with different initial fabrics but identical initial solid density were
evaluated. The first assembly was isotropically compressed to 50 kPa using a ser-
vo-control algorithm. The second assembly was initially compressed isotropically
as the first one, but then it was pre-sheared under axi-symmetric compression be-
fore being re-consolidated back to an isotropic stress of 50 kPa. These procedures
created an assembly with isotropic and anisotropic fabrics (with a bigger propor-
tion contacts in the vertical direction), respectively.
Macro-scale response. After specimen generation, both assemblies were then

subjected to identical combinations of virgin loading, unloading and re-loading
(while 2=3=0), as illustrated in Figure 3a. Note that unloading and re-loading
cycles were simulated starting from different OCR values. Interestingly, Fig. 3a
shows that the difference in the K0 value caused by having a different initial fabric
is of the same order of magnitude as the difference between the isotropic speci-
men and the empirical relationship. Therefore, the effects of inherent anisotropy
that were impossible to quantify from experiments might be responsible for a sig-
nificant part of the scatter in the results available.

2000 0.036
Vertical effective stress (kPa) - 'v OCR = 1
Deviatoric fabric (d = 1-3)

1600 0.032
1200 0.028
OCR = 2
800 0.024
OCR = 3
OCR = 4
400 DEM (Anisotropic) 0.020
DEM (Isotropic)
Empirical relationship
0 0.016
0 400 800 1200 1 2 3 4
Horizontal effective stress (kPa) - 'h Overconsolidation ratio - OCR
a) b)
Fig 3. a) Macro-scale results for different stress histories on inherently iso-
tropic and anisotropic fabrics and b) micro-scale response for isotropic specimen
Micro-scale response. The evolution of soil fabric for these different stress
histories was also monitored. As means of quantifying the evolution of soil fabric,
the fabric tensor (ij) was evaluated:
1 Nc
 ij 
Nc
n n
k 1
i j (4)
where Nc is the number of inter-particle contacts in the assembly ni = nj are the

Cartesian components of the unit contact normal between the two particles in con-
tact. In particular the deviatoric fabric (d) defined by Thornton (2000) as the dif-
ference between the major and minor eigenvalues of the fabric tensor was calcu-
lated and it is presented in Fig. 3b for the isotropic specimen.
In Fig. 3b only the over-consolidated states are presented, therefore after initial
virgin loading the fabric of the specimen has increased from 0 to about 0.034 (i.e.
it has become anisotropic. Then after unloading the fabric tries to become isotrop-
ic before increasing its anisotropy during re-loading. It can be observed that the

evolution of d against OCR is somehow similar and inversely proportional to the

evolution of the macro-scale response shown in Fig. 3a.
CONCLUSION
The results presented in this study demonstrate that the scatter available from
experimental measurements of K0 can be partly explained by the evolution of soil
fabric. Furthermore, it has been demonstrated that the observed macro-scale be-
haviour can be also explained by this evolution of fabric.
The success of K0-OCR relationships is demonstrated by the success of existing
geotechnical structures and extensive additional research is needed to further un-
derstand the factors affecting such relationships. This would immediately lead to
the use of lower factors of safety and perhaps more economical geotechnical de-
signs. However, the biggest challenge still remains in how to reconstruct the geo-
logical stress history of soils in-situ.
REFERENCES
1. Barreto, D. ‘Numerical and experimental investigation into the behaviour of
granular materials under generalised stress states’. PhD thesis, Imperial Col-
lege of Science, Technology and Medicine. Univeristy of London, 2009.
2. Cui, L., O'Sullivan, C. and O'Neill, S. ‘An analysis of the triaxial apparatus
using a mixed boundary three-dimensional discrete element model’, Géotech-
nique, 57, No. 10, 831-844, 2007.
3. Cundall, P. A. & Strack, O. D. L. ‘A discrete numerical model for granular
assemblies’, Géotechnique, 29, No. 1, 47-65, 1979.
4. Ji, S., Hanes, D. M. & Shen, H. H. ‘Comparisons of physical experiment and
discrete element simulations of sheared granular materials in an annular shear
cell’. Mechanics of Materials, 41, No. 6, 764-776, 2009
5. Mayne, P. W. & Kulhawy, F. H. ‘K0-OCR Relationships in Soil’. Journal of
the Geotechnical Engineering Division, ASCE, 108, No. 6, 851-872, 1982.
6. Thornton, C. ‘Numerical simulations of deviatoric shear deformation of granu-
lar media’, Géotechnique, 50, No. 1, 43-53, 2000.

THEORETICAL INVESTIGATION OF
REGULARITIES OF MODEL FAULT ZONE
MECHANICAL RESPONSE TO LOW-AMPLITUDE
DYNAMIC MECHANICAL ACTIONS
Sergey V. Astafurov, Evgeny V. Shilko, Alexandr S. Grigoriev*,

Artem Yu. Panchenko*, Sergey G. Psakhie
*Tomsk State University, Tomsk, Russia
Theoretical analysis of general features of fault zone mechanical re-

sponse to low-amplitude dynamic mechanical actions is carried out.
Results make an important contribution to development of an ap-
proach to estimation of proximity of stress state of faults and large
cracks in rock massifs to critical value wherein seismogenerating un-
steady slip can take place.
INTRODUCTION
An important direction in mechanics is investigation of features of mechanical
response of geological media. An important feature of geomedium is that its
fragments are in complex stressed state. Relation of internal stresses to strength
characteristics of interfacial regions (faults and cracks) appreciably defines de-
formation and relaxation capacity of the mechanism concerned with relative block
displacement. As the shear stress at an active block boundary reaches limiting
(threshold) value, its local deformation mode can change qualitatively from slow
deformation (creeping) to dynamic deformation (referred to as unstable sliding).

Note that according to modern notion, acts of dynamic block sliding are seismo-
generating events those magnitude can reach 6-7 [1]. Thus, an urgent task in geo-
mechanics is to develop methods of estimating the local stress state at active inter-
faces of fragments of rock massifs or the earth’s crust. Theoretical studies as well
as experiments on prestressed rock samples and fragments of plane discontinuities
in rock massifs revealed an important effect consisting in deformation response of
geomedium to dynamic perturbations of stress state in form of irreversible relative
displacement of blocks [1]. This allowed different authors to formulate the idea
about the possibility of estimation of shear stress level in the fragments of active
fault zones by measuring deformation response on testing dynamic actions. Nev-
ertheless to transform this idea into practical approach it is necessary to have in-
formation about the character of connection between shear stress level in interfa-
cial zone and deformation response to dynamic perturbation of local stress state.
This information could be obtained in particular on the base of computer-aided
simulation of fragments of block interfaces being under complex stress condition.
In the present paper this study was done with use of the movable cellular automa-
ton method (MCA) [2]. This method is a variant of particle method and over last
decade successfully employed for studying the peculiarities of deformation and
fracture of indurated, unconsolidated and loose media.
PROBLEM STATEMENT
For solving the problem posed, a two-dimensional structural model of the in-
terface between elements of a block medium was developed in the framework of
MCA method (Fig. 1а). The model includes the interface itself and contiguous
block fragments (hereinafter referred to as blocks). The model blocks were as-
sumed to be monolithic, brittle and of high strength (curve 1 in Fig. 1b). The me-
chanical response of interfacial region featured irreversible strain accumulation
and simulated the mechanical properties of a cataclasites (curve 2 in Fig. 1b). For
mathematical description of elastoplastic response of the interface and of the
blocks in the framework of the MCA method, the model described in [3] was
used. The lower surface of the specimen was rigidly fixed while external load was
applied to the upper surface (Fig. 1a). Dissipative and “inertial” influence of sur-
roundings of simulated fragment of interface was modeled by “applying” an addi-
tional viscous force to the automata of the lateral surfaces. The initial stress state

of the system was specified by applying an external force with normal (Fy) and
tangential (Fx) components to the upper surface of the model specimen (Fig. 1a)
with subsequent relaxation of the system. In the calculations, we used different
values of normal force Fy, whose specific value (hereinafter referred to as y) va-
ried in the range from 20% to 40% of the yield stress of interface material. On
completion of the relaxation stage, deformation response of model interface to
dynamic pulse testing loading by P- and S-waves was studied.
a b
Fig. 1. Structure and loading pattern of the model interface (a) and response
functions (b) of the automata of the blocks (1) and interfacial region (2).
Fig. 2. Dependence of force fx of resistance to the shear displacement of active

(upper) block of the model interface on the value of displacement lsh. In the dia-
gram shown, shear displacement is normalized to the interface layer thickness h.
Normal stress (y) value is 20% of the yield stress of interface material.

In the work, we considered different shear stress levels at the interface for
which fx=Fx/Fxmax, where Fxmax is the limiting (maximum) shear resistance force
at considered value of Fy. Note that shear resistance curve has a profile typical of
interfaces in rock massif fragments and is characterized by three main stages: a
quasielastic stage (Fig. 2, stage I), a quasiplastic stage (stage II) and a stage
beyond the ultimate strength (stage III). Testing pulse dynamic action on the sam-
ple being under loading with forces Fx and Fy was carried out by “instantaneous”
(for T=310-5 s) assigning a sinusoidally changing velocity Vy=-VAsin(2t/T) in Y-
direction (compressive P-wave) or Vx=VAsin(2t/T) in X-direction (S-wave) to
the cellular automata of the upper surface (VA=0.001 m/s in both cases).
DISCUSSION OF SIMULATION RESULTS

It is common knowledge that even small dynamic exposure gives rise to irre-
versible relative block displacements if shear stress level (fx) at the interface ex-
ceeds some threshold value. This threshold is below the point of “formal” transi-
tion from quasielastic to quasiplastic response stage. The observed effect is con-
nected with inhomogeneous distribution of stresses in interface region that is a
consequence of inhomogeneous distribution of mechanical properties. Stress state
of some areas of the interface can be beyond the yield point, even though integral
value of applied shear stress is relatively small. This leads to accumulation of ir-
reversible deformation in these areas during elastic wave propagation.
Simulation results showed that the path length lx of the upper block along the X
direction varies over order of magnitude with the shear stress level fx, and this var-
iation has a clearly defined nonlinear character (the value of initiated displacement
lx is presented in dimensionless form х=lx/h obtained by normalization of lx to in-
terface thickness h). Fig. 3a shows the curves х(fx) for the case of pulse loading
by P-wave for different values of normal stress y. These curves have three main
regions. In the first region, which is far from the point of formal transition to qua-
siplastic response of the interface (stage I in Fig. 2), value of dynamically initiated
displacements is nearly constant. In the second region (near the formal point sepa-
rating stages I and II in Fig. 2) х increases greatly. In the third stage (in the range
of quasiplastic response of the interface) the value of initiated displacements be-
comes “stable”. Thus, as the shear stress level approaches the formal boundary of

quasielastic-to-quasiplastic response transition, qualitative variations occur in de-

formation response of the interface to dynamic loading. Revealed effect is impor-
tant for formulation of the approach to diagnostics of stress state of block inter-
faces in block-structured geomedium.
a b
Fig. 3. Reduced pulse-initiated shear displacement х of the upper block versus the
shear stress level fx (a - compressive P-wave loading; b - S-wave loading). Value
of normal stress y amounts to 20% (curve 1) and 40% (curve 2) of the yield
stress of interface material. Vertical dashed line symbolically separates stages of
quasielastic and quasiplastic response of studied interface.
Note that character of interface deformation response defines to a great extent
not only by shear stress level fx, but by applied normal load Fy and by type of in-
cident wave as well. Comparing of curves 1 and 2 (Fig. 3 a) obtained for different
values of normal stress y shows that under large normal load the region of great
increase of initiated shear displacement extends far in the stage of quasiplastic in-
terface response. Effective slope of this region therewith decreases nearly propor-
tional to increase of y. (see curves 1 and 2 in Fig. 3a). Hence character of change
of block boundary deformation response to testing pulse actions by compressive
P-waves with shear stress level is appreciably defined by normal stress value.
At the same time, pulse testing actions with use of elastic S-waves looks more
promising. Fig. 3 b shows curves of deformation response of the model block
boundary to dynamic exposure to S-waves at two values of applied normal stress
y. It is seen that dependencies х(fx) at different y are close to each other. The

region of great increase of х is strongly pronounced in all cases, and its comple-
tion is associated with the point of “formal” transition from quasielastic (I) to qua-
siplastic (II) response stage. Hence characteristics of interface shear deformation
response to dynamic exposure to S-waves insignificantly depend on value of nor-
mal stress y (at least within the range of “moderate” values of y).
CONCLUSIONS
Simulation results allowed authors to propose the way to development of an
approach to diagnostics of shear stress level at sections of active faults of structur-
al blocks in rock massif. The approach could be based on long-time monitoring of
dynamics of natural and man-caused displacements in studied part of active block
boundary. It is shown that both theoretical and experimental study of deformation
response of real fragments of fault zones to testing pulse dynamic exposure could
be carried out with use of local sources of pulse perturbations of stress state,
which generate mainly S-waves. In that case obtained data about interface defor-
mation response could be interpreted the most unambiguously and will reliably
characterize the shear stress level, i.e. the proximity of shear stress at the interface
to the current value of stress corresponding to beginning of the stage of “macros-
copic” irreversible deformation.
ACKNOWLEDGEMENTS
The investigation has been carried out at financial support of the RFBR Grant
No. 09-05-00968-а, RAS Presidium Project No.11.2, Grant of the President of
Russian Federation (MK-130.2010.5) and Russian Science Support Foundation.
REFERENCES
1. E. Shilko et al., ‘On the possibility of shear stress estimation at interfaces of
block media’, Phys. Mesomech, 13, No. 1-2., 25-31, 2010.
2. S. Psakhie et al.‚ ‘Movable cellular automata method for simulating materials
with mesostructure’, Theor. Appl. Fract. Mech., 37, No.1-3, 311-334, 2001.
3. E. Shilko et al., ‘An approach to modeling of elastic-plastic continuum within
the framework of discrete element approach’, Proc. of APM 2009, 587-594,
2009.

THE USE OF DISCRETE ELEMENT METHODS ON

THE DYNAMIC ANALYSIS OF MULTI-DRUM
ANCIENT STRUCTURES
Loizos Papaloizou, Petros Komodromos

University of Cyprus, Department of Civil and Environmental Engineering
The seismic behaviour of ancient monumental structures with mono-

lithic or multi-drum classical columns and colonnade systems in two
rows is investigated. In particular, the Discrete Element Method
(DEM) is utilized in the study of ancient columns under strong ground
excitations, by simulating the individual rock blocks as distinct bod-
ies. A specialized software application is developed, using a modern
object-oriented programming language, in order to enable the effec-
tive simulation of multi-drum columns and colonnades under har-
monic and earthquake excitations.
INTRODUCTION
Strong earthquakes are common causes of destruction of ancient monuments,
such as classical columns and colonnades. Ancient classical columns of great ar-
chaeological significance can be abundantly found in high seismicity areas in
Eastern Mediterranean. Multi-drum columns are constructed of stone blocks that
are placed on top of each other, usually without connecting material between the
individual blocks. The seismic behaviour of these structures exhibits complicated
rocking and sliding phenomena between the individual blocks of the structure that
very rarely appear in modern structures.

In ancient Greece the temples formed the most important class of buildings
erected during that era and can be classified into three "Orders of Architecture",
the Doric, Ionic and Corinthian order. An "order" in Greek architecture consists of
the column, including the base and the capital, and the entablature (Fig. 1). The
entablature is divided into the architrave (lower part), the frieze (middle part) and
the cornice (upper part). The differences among these three orders refer on the
dimensions, proportions, mouldings and decorations of the various parts.
Cornice
Entablature Frieze
Architrave
Capital
Shaft
Column
Fig. 1: Architecture of a typical classical monument.
Today, the remains of most of these temples are often limited to series of col-
umns with an entablature or only an architrave, and in some cases only standalone
columns. The investigation of the seismic behaviour of such monuments is scien-
tifically interesting, as it involves complicated rocking and sliding responses of
the individual rock blocks. The understanding of the seismic behaviour of these
structures contributes to the rational assessment of efforts for their structural reha-
bilitation and may also reveal some information about past earthquakes that had
struck the respective region. Ancient monuments, compared to modern structures,

have been exposed to large numbers of strong seismic events throughout the many
centuries of their life spans. Those that survived have successfully withstood a
natural seismic testing that lasted for several centuries. Thus, it is important to un-
derstand the mechanisms that enabled them to avoid structural collapse and de-
struction during strong earthquakes. Since analytical study of such multi-block
structures under strong earthquake excitations is practically difficult, if not impos-
sible for large numbers of blocks, while laboratory tests are very difficult and
costly, numerical methods can be used to simulate their seismic response.
Α very extensive review of the literature on the usage of numerical methods for
the analysis of monuments until 1993 was published by Beskos (1993/94) [1]. The
dynamic behaviour of infinitely rigid bodies during horizontal excitations was
studied by Housner (1963) [2], while, later on, other researchers [3]-[7] investi-
gated further, both analytically and experimentally, the required conditions to
overturn rigid bodies. Such structures can be simulated utilizing the Discrete Ele-
ment Methods (DEM), which have been specifically developed for systems with
distinct bodies that can move freely in space and interact with each other with
contact forces through an automatic and efficient recognition of contacts.
Research efforts to use the DEM in the simulation of ancient structures have al-
ready shown promising results, motivating further utilization of this method. Re-
cent research work based on commercial DEM software applications [3,8], dem-
onstrated that the DEM can be reliably used for the analysis of such structures,
although they reported a sensitivity of the response to small perturbations of the
characteristics of the structure or the excitation. However, similar sensitivity has
also been observed in experiments with classical columns [9]. Hence, it is impor-
tant to perform large numbers of simulations with varying earthquake characteris-
tics and design parameters to properly assess and interpret the simulation results.
A custom-made DEM software has been specifically designed and imple-
mented to enable efficient performance of large numbers of numerical simulations
with varying parameters, modelling these structures with independent distinct
bodies, as they are constructed in practice. Such simulations allow us to assess the
influence of different earthquake characteristics as well as the various mechanical
and geometrical parameters of these structures on their seismic responses.

NUMERICAL ANALYSIS
In ancient Greek temples the internal columns supported an upper row of
smaller columns carrying the timber roof, whose remains can be seen today in
several ancient Greek monuments, such as the temple of Aphaia in Aegina.
T=35.62 sec T=36.37 sec
T=37.07 sec T=37.05 sec
Fig. 2: Mexico City Earthquake (scaled 1.5 Times).

Fig. 2 and Fig 3. show snapshots of the response of such systems for the Ath-
ens and the Mexico City Earthquakes, respectively. The results show that the fre-
quency content of the ground motion affects significantly the response. The dis-
placements of the upper level columns in respect to the displacements of the lower
level column are affected by the frequency content of the excitation. The maxi-
mum acceleration that is required to overturn the system of columns increases as
the excitation frequency decreases.
For the Mexico City earthquake, which has low predominant frequencies, the
upper level columns overturn. On the other hand, for the Athens Earthquake,

which has higher predominant frequencies, and for a much larger maximum ac-
celeration the upper level columns are not as much affected.
T=4.475 sec T=5..225 sec
T=6.850 sec T=12.20 sec
Fig. 3: Athens Earthquake (scaled 9 Times).
CONCLUSION
A major contribution of this research work is the development of a flexible ob-
ject-oriented software that is specifically developed for the numerical investiga-
tion of the response of multi-drum columns and colonnades under harmonic and
earthquake ground excitations. The developed application is based on the DEM,
utilizing contact detection algorithms, object-oriented design and programming, as
well as Design Patterns to address effectively the specific problem, while main-
taining the necessary flexibility and extensibility of the program.
REFERENCES
1. D. Beskos, ‘Use of finite and boundary elements in the analysis of monuments
and special structures’, Association of Civil Engineers of Greece, 217, 1993.

2. G.W Housner, ‘The behavior of inverted pendulum structures during earth-

quakes’, Bulletin of seismological Society of America, 53, 403-417, 1963.
3. I. Psycharis, J. Lemos, D. Papastamatiou, C. Zambas and C. Papantonopoulos,
‘Numerical study of the seismic behaviour of a part of the Parthenon Pronaos’,
Earthquake Engineering & Structural Dynamics, 32, 2063-2084, 2003.
4. A. Pompei, A. Scalia, M.A. Sumbatyan, ‘Dynamics of Rigid Block due to
Horizontal Ground Motion’, Engineering Mechanics, 124, 713-717, 1998.
5. N. Makris and J. Zhang, ‘Rocking response of anchored blocks under pulse-
type motions’, Engineering Mechanics, 127, 484-493, 2001.
6. G.C. Manos, M. Demosthenous, V. Kourtides, A. Hatzigeorgiou, ‘Dynamic
and earthquake behavior of models of ancient columns and colonnades with or
without energy absorbtions systems’, Proceedings of Second Greek National
Conference on Earthquake Engineering and Seismology, 1, 257-276, 2001.
7. P. Komodromos, L. Papaloizou, P. Polycarpou, ‘Simulation of the response of
ancient columns under harmonic and earthquake excitations’, Engineering
Structures, 30(8), 2154-2164, 2008.
8. C. Papantonopoulos, I. Psycharis, D. Papastamatiou, J. Lemos, H. Mouzakis,
‘Numerical prediction of the earthquake response of classical columns using
the distinct element method’, Earthquake Engineering and Structural Dynam-
ics, 31, 1699-1717, 2002.
9. H. Mouzakis, I. Psycharis, D. Papastamatiou, P. Carydis, C. Papantonopoulos,
C. Zambas, ‘Experimental investigation of the earthquake response of a model
of a marble classical column’, Earthquake Engineering and Structural Dy-
namics, 31, 1681-1698, 2002.
10. P. Komodromos, L. Papaloizou, P. Polycarpou, ‘Simulation of the response of
ancient columns under harmonic and earthquake excitations’, Engineering
Structures, 30(8), 2154-2164, 2008.

SIMULATING P-WAVE PROPAGATION IN DEM
G. Marketos
Currently part of Imperial College London, Department of Civil and Environmental Engineering
Formerly part of the National Oceanography Centre, Southampton, NSRD Geology and Geophys-
ics Research Group
The Results from a Discrete Element Method simulation of wave

propagation through a dry pack of frictionless spheres are presented
here. Even though damping was suppressed, velocity dispersion and
apparent attenuation of the input signal were observed. In order to
clarify the reasons behind this the displacement time histories of the
individual grains will be shown. The grain-scale mechanism responsi-
ble for this phenomenon will then be discussed.
INTRODUCTION
The propagation of seismic waves is becoming increasingly important for a
number of applications, ranging from hydrocarbons exploration in geophysics to
site investigations in geotechnics. In addition to its speed, the values of attenua-
tion (damping) of the propagated wave can be used as a second parameter to bet-
ter constrain the geophysical inversion problem. However, the highly frequency-
dependent wave attenuation mechanisms in granular materials are poorly under-
stood, causing problems in the interpretation of field (low-frequency) or ultrasonic
(high-frequency) laboratory data. Discrete Element Method (DEM) simulations
have been successfully used in the past for studying wave propagation through
granular packs (e.g. Mouraille and Luding, 2005 and Makse et al, 2004) as they
can allow for the connection of grain-scale processes to particular aspects of the

observed macroscopic response. Detailed results from a DEM simulation will be

presented here with the aim of clarifying aspects related to attenuation and damp-
ing in granular packs.
THE SIMULATION
The simulation presented here was conducted using the commercial code
PFC3D and forms part of a wider study into wave propagation through sands. The
granular sample used contained 8943 spheres with radii uniformly distributed be-
tween 1 and 2 mm. The interactions between spheres were modelled by linear
springs of a stiffness of 4x106N/m in both the normal and shear directions. In the
simulation presented here the frictional coefficient of the grains was set to zero
and all damping was removed. There were therefore no frictional dissipation and
no energy losses due to damping. The observed response could therefore be di-
rectly linked to the granular nature of the material.
Fig 1. The input velocity at the top (excited) wall.

A compressional signal was propagated through the sample by moving the top
rigid confining wall. The velocity of the excited wall was set as a half sinusoid
cycle after which the wall was kept stationary (see Fig. 1). The disturbance thus
induced travelled through the sample as layers of grains were sequentially set into

motion around their equilibrium position. After some time the disturbance reached
the stationary, bottom receiving wall, and was measured as a perturbation around
the equilibrium stress value. The stress histories for the exciting and receiving
walls are plotted in Figure 2.
Fig 2. The stress time histories for the exciting (top) and receiving (bottom)
wall.
One can observe here that the movement of the top wall induced a stress pulse
immediately at the excited wall, as this was forced to move into the granular ma-
terial. The bottom wall was initially unaffected by this. The first signs of the dis-
turbance were picked up after 0.080 ms, while the peak stress was registered after
0.145 ms. One would be tempted to extract velocities for the propagating distur-
bance; as the sample height was 6.02 cm these can be calculated as 753 m/s for
‘first-break’ of the signal and 439 m/s from peak to peak.
The difference between the two velocity values above is striking- a commonly
used approach is to assume that the granular medium is linear and compute the
velocities (and attenuations) for each individual frequency contained in the input
signal. The value of the wave velocity thus computed will be highly frequency-
dependent. High levels of this so-called velocity dispersion are commonly as-
sumed in geophysics to be linked to high levels of attenuation of the propagating

waves. This leads to a paradox here, as there is no damping and no inter-grain

friction in the simulation, and so one would expect no attenuation. The reason for
this apparent attenuation and the observed widening of the transmitted stress pulse
must therefore be sought elsewhere. Individual grains will be probed for dis-
placement with the aim of clarifying the picture.
GRAIN DISPLACEMENTS
As one can observe in Figure 3 the displacement of grains that were contacting
the top wall (where the perturbation was applied) were initially very similar. If the
grains were rigidly attached to the wall there would have been zero relative dis-
placement between them. However, the rapid downwards movement of the wall
induced the grains to move initially into it. This caused the rapid stress increase
observed in Figure 2. Some displaced into the wall more than others; the differ-
ence in the amplitudes must be related to the resistance to movement offered by
the granular pack in the vicinity of the top grains (combined with the effect of dif-
ferences in the mass of the grains). However as seen in Figure 3 the initial shapes
of the time-displacement curves were very similar, reaching the peak at similar
times and so, initially containing similar frequencies.
Fig 3. The displacement time histories of 8 grains that are in contact with the
top wall. Note that the displacement of the moving top wall has been subtracted.

Figure 3 demonstrates how the first layer of grains was excited. As discussed
by Marketos and Bolton (2010) the presence of the top wall disrupts the granular
structure in its vicinity. The displacement of individual grains for layers below
will quickly become less orderly as the packing moves to a more random one and
the pathways by which the excitation travels to them will vary. In addition two or
more different paths by which the excitation travels through the sample might
converge on a single grain. The displacement of the grain will be then given by
the superposition of the displacement time histories transmitted to it. As there is
no reason to expect that the arrival times through two different paths to be the
same the effect of superposing a number of pulses centred at different times is to
create broader pulses. This can be nicely observed in Figure 4 below, which plots
the time histories for displacements of 8 randomly selected grains that are in con-
tact with the bottom wall.
Fig 4. The displacement time histories of 8 grains that are in contact with the
bottom, stationary wall.
The broadening of the displacement pulses in turn manifests itself as a broad-
ening in the stress signal picked up at the bottom receiving wall. As there was no
frictional dissipation and no damping in the simulation presented here the results
clearly show that there is another mechanism by which frequency dispersion can

be created inside a granular sample. This is due to the fact that there are a large
number of pathways through which a signal can be transferred through a granular
sample.
The difference between the ‘peak to peak’ and ‘first break’ velocities can also
be explained now. The time at which ‘first break’ occurs is governed by the short-
est distance by which the disturbance can propagate through the granular pack.
This will most probably coincide with an efficient pathway for carrying stress (a
strong-force chain). On the other hand the peak of the received signal corresponds
to the point in time when the largest number of transmitted pulses reaches the re-
ceiving wall. This can then be said to relate to the most probable length of path-
way available to the propagating wave.
ACKNOWLEDGMENTS
This work was funded by the Natural Environment Research Council through
the Oceans 2025 Theme 5.4 (Fluid flow) project. Support from Dr Angus Best
and Jeremy Sothcott is gratefully acknowledged.
REFERENCES
1. H.A. Makse, N. Gland, D.L. Johnson, L. Schwartz, ‘Granular packings:
Nonlinear elasticity, sound propagation, and collective relaxation dynamics’,
Physical Review E, 70, 061302, 2004.
2. G. Marketos, M.D. Bolton, ‘Flat boundaries and their effect on sand testing’,
Int. Journal for Numerical and Analytical Methods in Geomechanics, 34, 821-
837, 2010.
3. O. Mouraille, S. Luding, ‘Sound propagation in dense, frictional granular ma-
terials’, Proceedings of the 5th International Conference on Micromechanics
of Granular Media, Powders and Grains, Stuttgart, 319-322, 2005.

OTHER APPLICATIONS

A CASE STUDY OF IMPACT ON GLASS USING THE

COMBINED FINITE-DISCRETE ELEMENT
METHOD
Xudong Chen, Andrew HC Chan and Jian Yang

Glass is being widely used as structural elements due to the appealing

feature in both architectural and functional aspects. In order to inves-
tigate how glass fractures into pieces under impact and the subsequent
interactions between fragments, a new numerical analysis method
called the combined finite-discrete element method (FEM/DEM) was
employed in this paper. The corresponding 2D FEM/DEM program Y
was used to analyse the behaviour of glass under impact. Studies were
performed through numerical example and discussion. Numerical re-
sults were presented to demonstrate the applicability of the Y program
in the analysis of glass fracture analysis.
INTRODUCTION
Due to the appealing aesthetic appearance and transparency, increasing glass
applications have emerged in civil engineering in structural use. One example is
the Grand Canyon Skywalk in Arizona, the United States. The glass observation
deck of the Skywalk lets tourists gaze deep into the abyss without any barriers.
Another case of glass structures is the Sears Tower Glass Platform. With its four
10 foot by 10 foot compartments protruding 4.3 feet from the building's 103rd
floor observation room, the sky balcony provides a completely transparent space
to generate the sensation of hovering over Chicago.

As can be observed from the above cases, glass has served the role of structural
elements which was usually the task of steel and concrete. This new application
requires accurate analysis of the behaviour of glass structures as well as their
components under both static and dynamic loading. Since the use of static analy-
sis method is relatively well-developed [1,2], the main concern from engineers is
dynamic performance, mainly due to the brittle nature of glass [3]. For these glass
structures high in the sky, one of the threats will be the hit caused by windborne
debris, which should be analysed with nonlinear dynamic behaviour.
The combined finite–discrete element method (FEM/DEM) is a newly devel-
oped numerical method aiming at investigating failure, fracture and fragmentation
in solids. This method combines the advantages of both finite element method and
discrete element method. Finite elements are meshed into the discrete particles to
obtain more accurate results. Discretised contact solutions are used for both con-
tact detection and contact interaction. Many other rationales of FEM/DEM can be
found in the book by Munjiza [4].
This is the first time that FEM/DEM is applied in the field of computational
analysis of glass impact mechanics. The aim of this paper is to study the impact
resistance and breakage behaviour of monolithic glass under hard body impact
through examples. Some significant aspects of FEM/DEM will be illustrated in
the numerical results and discussion section.
NUMERICAL STUDIES
Considering the case that a monolithic glass beam of 2m long 0.02m thick with
its two ends installed in rigid channels at both sides is impacted by a steel projec-
tile at the velocity of 60.85m/s. The beam is meshed with a relatively fine mesh
(1600 discrete elements in total), and there is no friction between the ends of the
glass beam and the surfaces of the channels.
Fig 1. The configuration of the glass beam system

Material properties are given as:

Glass Steel
Density 2500kg/m3 7800kg/m3

Young’s Modulus 70GPa 200GPa
Shear Modulus 30GPa 76.9GPa
Poisson’s Ratio 0.2 0.3
Table 1. Material properties of glass and steel
The transient responses of the beam were plotted by the M, which is a post-
processing tool of FEM/DEM program Y. Some representative diagrams are se-
lected to show the deflection and damage of the beam as follows:
a. t = 1ms
b. t = 3ms
Fig 2. Deflection and damage of glass beam during impact
It can be observed that at the beginning, the beam experienced a small amount
of local bending due to the dynamic impact of the projectile (Fig 2a). With the
elapsing of time, deflection waves were propagated from middle to the two ends
and cracks developed on and inside the surface of the beam (Fig 2b).
The following two figures showed the fracture of the beam at the impact point
at 1ms and 3ms, respectively. The cracks on and inside the beam surface can eas-
ily be observed (Fig 3a and 3b). According to figure 3, crack at t=3ms is less se-

vere than that at t=1ms. This should be attributed to the repeated deformation of
the beam and the cracks were made open and close again and again.
a. t = 1ms b. t = 3ms
Fig 3. The cracks at the impact point
The change of total kinetic energy of the projectile and glass beam was plotted
versus the time as:
30000
Total Kinetic Energy(J)
25000
20000
15000
10000
5000
0
0 0.01 0.02 0.03
Time(s)
Fig 4. The total kinetic energy versus time

According to Fig. 4, the total kinetic energy decreased with the increase of time
in general. At the very beginning, only the projectile contributes to the total ki-
netic energy as the beam is at rest. As time elapses, interaction occurred between
the projectile and the beam, which transferred the kinetic energy of the projectile

to the beam with the majority being absorbed by the fracture of beam and the total
kinetic energy decreased dramatically. Finally, the total kinetic energy was kept in
a constant level with some acceptable oscillations, which represents the absorp-
tion and release of energy due to the oscillating deflection of the glass beam.
CONCLUSIONS AND PROSPECTS

A numerical example of glass beam under impact is presented in the proceeding
section. Numerical results were discussed and demonstrated that the combined
finite-discrete element method program Y is applicable for analysing the glass
impact mechanics. The method for analysing two-dimensional monolithic glass
beam can be extended to multilayer and three-dimensional glass plate in the fu-
ture.
ACKNOWLEDGMENTS
We would like to thank Professor Ante Munjiza (Queen Mary, University of
London) provide us the FEM/DEM program Y, both 2D and 3D. The funding
provided by the University of Birmingham and China Scholarship Council which
enabled the research work to be carried out is gratefully acknowledged.
REFERENCES
1. A. Darkov, Structural Mechanics, Mir Publishers, 1979.
2. S. Timoshenko, S. Prokop, Theory of Plates and Shells, McGraw-Hill, 1959.
3. T.L. Anderson, Fracture Mechanics: Fundamentals and Applications, CRC
Press, 2000
4. A. Munjiza, The Combined Finite-Discrete Element Method, John Wiley &
Sons, 2004


DEM MODELLING OF PARTICLE FLOW IN A

TURBULA MIXER
M Marigoa,b, D L Cairnsa, M Daviesa, A Ingramb, E H Stitta

a
Johnson Matthey Technology Centre
b
University of Birmingham, Chemical Engineering
The three-dimensional motion of particles in a Turbula mixer was

calculated using discrete element modelling (DEM). Multiple-Positron
Emission Particle Tracking (PEPT) was used to determine the
movement of the Turbula mixer to set-up the numerical model. The
measured movement was mathematically reconstructed and directly
imported into EDEM so that the same movement is applied to the
modelled vessel. Good qualitative agreements have been found
between EDEM mixing simulations and experimental data from
Sommier et al. (2001). In particular initial studies on the mixing
mechanisms have been carried out.
INTRODUCTION
Many industrial operations involve solid mixing in the absence of any liquid.
These processes are fundamental and often critical in a wide range of industries
such as food, pharmaceutical and chemical.
The Turbula mixer, shown in Fig 1, is a laboratory scale mixer used in the
development or testing of new products. The Turbula mixer T2F has a capacity of
2 L and can hold almost any form of container. The rotational speed can be varied
from 22 to 100 rpm. The mixing vessel, located in the mixer chamber, is subjected

to intensive, periodically pulsating movements, which simulate the pattern of

agitation achieved by manual shaking (Sommier et al., 2001). In addition to the
traditional principal motions of rotation and translation there is a third
fundamental motion of inversion based on the Schatz inversion principle. This
inversion subjects the contents of the mixing container to two alternating,
rhythmic pulsating motions.
Fig 1: Turbula T2F shaker-mixer (W.A. Bachofen, Basel, Swizterland).

In this work, the mixing rate of particles with monomodal and bimodal size
distributions was quantified using a segregation index. Magnetic resonance
imaging (MRI) experiments (Sommier et al., 2001) revealed that segregation
occurs whenever a size difference between particles is present. Moreover, the
lower the rotation speed, the more segregated the final state was. In this work a
first attempt to model the motion of particles in the mixer has been carried out
using DEM. The results are compared with those of Sommier et al. (2001).
NUMERICAL MODEL DEM

The DEM used in this work is a commercially available package provided by
DEM-Solutions (EDEM version 2.0) and it is based on the algorithm developed
by Cundall and Strack in 1979. In this work the Hertz Mindlin (no slip) contact
model provided in the EDEM package is employed. This is based upon the work
of Mindlin (1949).
The Multiple-Positron Emission Particle Tracking technique has been used to
study the rotational and translational motions of the Turbula mixer. Three ion
exchange resin particles of different intensities were attached to three locations on
the Turbula shaft and their locations were followed over time. By knowing the
position of these three points the movement of the Turbula has been

mathematically reconstructed and imported into the EDEM simulations. A

cylindrical vessel with diameter 45 mm and length 80 mm was filled with
spherical particles (fill level 50%) and the movement of the vessel simulated at
different speeds, from 23 up to 69 rpm. In this work a diameter ratio of 1.4 has
been used. Fig 2(a) shows the initial loading pattern, the colours represent the 2
types of particles (red for 2 mm diameter and blue for 1.4 mm diameter particles).
Fig 2(b) shows a slice through the bed following 15 rotations at 23 rpm; a core of
small particles is apparent.
(a) (b)
Fig 2: (a) Initial loading pattern and (b) section through bed after 15 rotations for
bidisperse simulations at 23rpm and R=1.4.

The degree of mixing, and the time needed to achieve an acceptable mixing,
can be predicted by simulations. In this work a segregation parameter, proposed in
Stambaugh et al. (2004), is employed. The segregation index S’ is defined as:
C AA C BB
S'   (1)
C AA  C AB C BB  C AB
where Cij is the total number of contact between particles of type i and particles
of type j (with i and j equal to A or B).
Sommier et al. (2001) observed that segregation in the Turbula increases when
the particle diameter ratio differs from 1. It was also showed that the segregation
index decreases when the rotational speed increases but that at the highest speeds
the system remains segregated.

In Fig 3.a the segregation index from (Sommier et al., 2001) is reported as
function of rotation speed in the case when R=2.8. In Fig 3.b the steady state
segregation index, calculated by DEM simulation, for R=1.4 and after 15
rotations, is reported as a function of rotation speed. Good qualitative agreements
can be observed between the experimental results and the DEM simulations. The
increase of rotational speed leads to an improvement in the quality of mixing.
S e g r e g a tio n in d e x S '
2
1.8 R=1.4
1.6
1.4
1.2
1
0.8
0.6
-5 5 15 25 35 45 55 65 75
Rotational speed (rpm)
a) b)
Fig 3: Segregation index in function of rotation speed. a) Experiments from
[Sommier et al., 2001]. b) DEM simulations, value after 15 rotations.
Simulations with monosized particles have also been carried out in order to
investigate the mixing mechanisms. The segregation index S’ for various
rotational speeds reaches the asymptotic value corresponding to random mixing
S ' =1. However the number of rotations necessary to achieve this value changes
with rotation speed. Each curve has been fitted with an exponential law such as:
S 'fit  S  Ae N R / N mix (2)

Where NR is the actual number of rotations, Nmix is the characteristic number of
rotations and the asymptotic S defines the lower value for the segregation index
that in case of random mixing equals 1. By knowing the period of rotation at each
rotational speed the characteristic mixing time can be estimated. Both values are
reported in Fig 4.

Fig 4. Characteristic number of rotations and time as function of rotational speed.
CONCLUSIONS
A first attempt of discrete element modelling of the three-dimensional motion
of particles in a Turbula mixer was carried out by EDEM. The complex motion of
the Turbula mixer has been measured using Multiple Positron Emission Particle
Tracking (Multiple-PEPT) and it has been imported in the EDEM model.
The system has been studied under different conditions for non-cohesive
particles. Segregation due to particle size ratio has been shown. Good qualitative
agreements have been found between DEM simulations and experimental data
from Sommier et al., (2001).
The effect of rotational speed on mixing for monosized particles has been
investigated in terms of characteristic number of rotations, mixing time and also
dispersion coefficients. “Axial” and “transverse” dispersion coefficients have been
defined; in particular the minimum axial coefficient of dispersion, at 46 rpm,
corresponds to the maximum characteristic number of rotations. The non-
monotonic change in mixing efficiency with rotational speed might imply a
change of the bed behaviour. More investigations, not reported here, have been
carried out using PEPT in mixing studies on mono-sized glass beads in order to
validate the simulations. In particular, the PEPT experimental results have
confirmed the trends of the coefficients of dispersion obtained from the DEM
simulations.

ACKNOWLEDGEMENTS
MM would like to acknowledge the EU for financial support through the
Framework 6 Marie Curie Action "NEWGROWTH", contract number MEST-CT-
2005-020724, Johnson Matthey Plc for funding and supporting this research.
REFERENCES
1. R.D. Mindlin, Compliance of elastic bodies in contact, Journal of Applied
Mechanics, Volume 71, No.3, 259-268, 1949.
2. N. Sommier, P. Porion, P. Evesque, B. Leclerc, P. Tchoreloff, G. Couarraze,
Magnetic resonance imaging investigation of the mixing-segregation process
in a pharmaceutical blender, International Journal of Pharmaceutics, Volume
222, No.2, 243-258, 2001.
3. J. Stambaugh, Z. Smith, E. Ott, and W. Losert, Segregation in a monolayer of
magnetic spheres, Phys. Rev., E 70, 031304, 2004.

MODELING BREAKAGE ENVIRONMENT IN

TUMBLING MILLS USING DEM AND ANALYZING
THE OUTPUTS
N. S. Weerasekaraa, M. S. Powella, S. Coleb, J. Favierb

a
University of Queensland, JKMRC, 40 Isles Road, Indooroopilly, QLD 4068, Australia.
b
DEM Solutions Ltd., 49 Queen Street, Edinburgh EH2 3NH, UK
Corresponding Author Email: [email protected]
The present state of the art Discrete Element Method (DEM) is quite
capable of predicting the mechanical environment in full 3D for a
given mill configuration. A full 3D simulation providing a compre-
hensive prediction of bulk particle dynamics in a grinding mill is now
possible using the latest commercial DEM software tools. These DEM
codes provide a range of information of individual particle impact his-
tories, such as kinetic energy at the time of collision, normal and tan-
gential collision energy, contact angle, particle velocities, force, im-
pulse, etc. The ability to run a large number of DEM simulations and
data analysis provides the opportunity to resolve the breakage envi-
ronment in mills. This provides breakage models such as the Unified
Comminution Model (UCM) with a full range of information needed
on the mill mechanical environment for a wide range of mill operating
conditions and information need for liner wear modelling. Further-
more, this information provides an insight into the mechanics and
drivers of milling that in itself can be used in improving mill design
and operation. Therefore, in this work an attempt is being made to ex-
plore the breakage environment in mills using DEM techniques, and
how these techniques may be expanded to provide even more useful
data for mill and comminution device modelling.

INTRODUCTION
The motion of grinding media and the energy distribution have a profound in-
fluence on the comminution of particles in tumbling mills. The potential of the
discrete element method (DEM) allows numerical simulation of the dynamic in-
teraction of the mill charge media with the mill liners and lifters. It allows calcula-
tion of the trajectories of individual entities in the entire grinding charge as they
move in the mill and collide with one another and the mill. These impact energy
and force distributions can be used to calculate the grinding rate in mills, as pre-
sented by some workers in this area Mishra (1994), Cleary (2001), Datta and Ra-
jamani (2002), Herbst (2004), Powell (2006), Powell, Govender and McBride
(2008). The calculations are based on the fundamental laws of motion and can
take into account the geometry, dimensions, and material properties of each indi-
vidual steel ball, chunk of rock, mill liner, and the lifters (Mishra 2003).
A DEM simulation of a typical mill section with several thousands of particles
generates massive amounts of output data, running to several Gigabits. Even with
good sampling techniques (Cleary 2001) this generates files of close to a Giga-
byte. Although some DEM software provides a GUI based data analysis environ-
ment, their techniques either require a considerable amount of processing power
or are not capable of delivering the required inputs for the UCM kind of model-
ling or liner wear modelling.
This work brings together the power of DEM and an ability to analysis the out-
puts to explore and expand the understanding of the breakage environment in
mills and wear of liners.
SAG AND BALL MILL DEM SIMULATIONS

In this study several aspects of a SAG mill were investigated using EDEMTM.
A pilot scale SAG mill operating at 75% critical speed and drawing 9.94 kW of
power is used as the “BaseCase” for the DEM simulation campaign. This Base-
Case mill has an internal diameter of 1696 mm, length of 575 mm and has 11 lift-
ers with 50x50 mm cross-section. Three different particle size distribution were
used along with varying only the internal diameter from 1.6 m to 6 m while in-
creasing the cross sectional dimensions of the lifters in proportion to the change in

mill diameter and keeping the other conditions the same. This gives a matrix of
SAG mill DEM simulations
An industrial scale ball mill was simulated at two sizes, 4m and 8m diameter,
with all other parameters remaining unchanged. Real mill feed size distributions,
ball size distributions, and liner profiles were utilised. Taking advantage of the
symmetry along the mill and utilising average liner dimensions, a short slice of
the mill can be simulated to make the simulation time feasible. The results are
then multiplied by the ratio of the slice length to total mill length to give the total
collisions for the full size mill. This dramatically reduces the number of balls that
require simulation, from 4.5 million to 110 thousand for the 8m mill, and a reduc-
tion of simulation time by about 50-fold. As it is the simulation takes a number of
days, so such a reduction is essential to render the work feasible.
DATA ANALYSIS METHODOLOGY

A C/C++ based framework was employed in extracting and analysing the mas-
sive data sets generated though the DEM simulation (Weerasekara et al 2008).
The framework employs a two phase approach, where Phase I of this system in-
teracts directly with the DEM output and generates data sets which can be easily
managed and handled by the Phase II. The method adapted in Phase I will vary
slightly depending on the DEM output data structure of the DEM code used, but
the outputs are consistent for linking into the next stage of analysis. In this study
the selected collision data for Phase I analysis was written out via a custom mod-
ule coded using EDEM’s API (DEM Solutions, 2009). In Phase II, the data ex-
tracted and grouped in Phase I is further summarised depending on the require-
ments for further modelling work.

Using these DEM simulations, the energy dissipation associated with separate
particle interactions was calculated. Frequency distributions of these energy
losses, both in the normal and shear directions were determined from all the indi-
vidual collision events. Spectra were calculated for each type of collision event
and for each ball and rock size class. Energy spectrums for three different mill
sizes are compared in Fig 1. The techniques of logging collisions and calculating

collision energies is addressed in detail in Weerasekara and Powell (2008, 2010).

With the increase in mill size from 1.6 m to 3 m it was observed that smaller par-
ticles, less than 11 mm, experience higher normal and tangential energies, but
when the mill size further increase to 6 m the energies reduce considerably (Fig 1e
and f). It is also observed that the particles larger than 43 mm experience higher
normal and tangential energies as the mill size is increased. The particles in the
mid range between 30 – 15 mm seem to have an increasing trend with a peak at
the 3 m size and start to decrease gradually at the 6 m size.
Fig 1. Distribution of collision frequencies vs collision energy components in

the normal (column 1) and tangential (column 2) directions for the three mill sizes
(1.6 to 6 m from row 1 to 3)

The collisions in the ball mills are also summarised in the same fashion as for
the SAG mill. Every collision is summed up from the lowest energies to the high-
est energy collisions. The distributions for the four ball sizes over the four liner
profiles are presented in Fig 2. The four distinct energy distributions correspond
to the ball sizes, with the largest ones providing the highest energy collisions. The
collisions are expressed as per ball per second, so to obtain the total collision rate
the presented rate must be multiplied by the number of balls in each size class.
These collision spectrums can be utilized in predicting breakage rates in these
types of mills using breakage modelling techniques like the UCM.
Fig 2. Collision energy distribution for each ball size for the four liner geome-
try profiles
In modelling liner wear, the spectrum of loading on the front, s1, and rear, s2,
face of each wave liner are given in Fig 3. This shows the relative velocity and
pressure plots as a function of angular position in the mill. These are simply point
plots of all accumulated data. The pressure plot shows the range over which they
are in contact with the liner, and the bands represent loading by the different ball
size classes. The toe of the charge is at about 45° and the shoulder at about 220°.
As the wear model of Archard’s equation (Pirso 2004) is linear with the F*v*t
values (t being the duration of contact), the scatter plot data can simply be
summed to give the summarised totals, which could be used to predict the rate of
wear in liners.

Fig 3. Force and relative velocity distribution on selected set of liner segments
in the 8m ball mill
CONCLUSION
A campaign of DEM simulations were performed varying the mill type, size,
charge particle and ball size distribution, to explore and understand the breakage
environment in mills using DEM techniques. Using the large DEM outputs colli-
sion energy spectrums were calculated for estimating grinding rate. Then the nor-
mal force and relative velocity due to charge on the liner were derived for estimat-
ing liner wear rates. Using these tests the ability to analysis large data outputs of
DEM for the requirements for comminution modelling have being tested and
demonstrated.
ACKNOWLEDGMENTS
This work was carried out under the auspice and funding of the CSRP, and
with additional funding from the JKMRC

REFERENCES
1. P. W. Cleary, ‘Recent advances in DEM modelling of tumbling mills’, Miner-
als Engineering, Volume 14, Issue 10, 1295-1319, October 2001.
2. A. Datta, R.K. Rajamani, ‘A direct approach of modeling batch grinding in
ball mills using population balance principles and impact energy’, Int. J.
Miner. Process, 64 (4), 181–200, April 2002.
3. DEM Solutions Ltd, EDEM User Guide, 2009.
4. J.A. Herbst, ‘A microscale look at tumbling mill scale-up using high fidelity
simulation’, Int. J. Miner. Process, Volume 74S, pp.S299–S306, 2004.
5. B.K. Mishra, ‘A review of computer simulation of tumbling mills by the dis-
crete element method Part II—Practical applications’, Int. J. Miner. Process.
Volume 71. 95–112, 2003.
6. B.K. Mishra, R.K. Rajamani, ‘Simulation of charge motion in ball mills: Part
1. Experimental verifications’, Int. J. Miner. Process, 40, 171–186, 1994.
7. M.S. Powell, ‘The Unified Comminution Model - a conceptually new model’,
Proceedings of the XXIII IMPC. 3 - 8 Sep. 2006, Istanbul, Turkey, ISBN 975-
7946-27-3, pp. 1783-1788, 2006.
8. M.S. Powell, I. Govender and A.T. McBride, ‘Applying DEM outputs to the
unified comminution model’, Minerals Engineering, Volume 21, Issue 11, pp
744-750, 2008.
9. J. Pirso, M. Viljus, S. Letunovits, ‘Sliding wear of TiC–NiMo cermets’, Tri-
bology International, Volume 37, Issue 10, Pages 817-824, 2004.
10. N.S Weerasekara and M.S. Powell, The New Energy Logging from the Dis-
crete Element Method. In proceedings CSRP’08 2nd annual conference, Bris-
bane, Queensland, Australia, November 18-19, 2008.
11. N.S Weerasekara and M.S. Powell, Exploring the breakage environment in
mills with DEM. Accepted for publication In Proceedings XXV International
Mineral Processing Congress, Brisbane Australia, Sep 2010.

VEHICLE DYNAMICS ON OFF-ROAD TERRAIN
Matt Sinnott1 and Paul Cleary1

1
CSIRO Mathematics, Informatics and Statistics, Clayton, Australia
A 3D DEM model of a light truck was developed for the purpose of

investigating tyre-ground dynamics. Wheel motion and a three-
component independent suspension were included. As a preliminary
test for the model, the truck was driven over a flat, rocky terrain. The
effect of the loose granular topography on motion of the truck is re-
ported.
INTRODUCTION
The interaction of tyres with dry, flat rigid surfaces to propel a vehicle forward
is well understood as is evident from the advent of realistic driving simulators [1].
However, real off-road surfaces (soil/rocks/sand) consist of granular materials that
can compact under load [2], destabilize and flow around moving wheels, and ex-
perience large discontinuous deformations due to wheel-track sinkage. Such de-
formations are a challenge for FE models of the terrain. Horner et al. [3] devel-
oped a DEM model of a rigid, smooth wheel plowing through a terrain of 10 mil-
lion spherical particles. Nakashima et al. [4] have investigated the effect of tyre
tread on vehicle traction for a single wheel interacting with a coarse, 2D spherical
DEM model of sand. However, depending on the terrain it is not obvious that soil
deformations generated by each wheel can be isolated from that of the other
wheels. To date, no full vehicle model with multiple wheels interacting with rocky
ground has been attempted.

This is a preliminary study of a model of an AWD (All Wheel Drive) Toyota

Hilux truck driving in a straight line over a granular terrain. The model will be
demonstrated on very simple, almost flat, rocky ground but more complex topog-
raphy are straight forward to use and will be investigated in the future. The rock
shapes are represented by blocky, super-quadric (SQ) particles [5]. Independent
double-wishbone suspension and coil spring/shock absorbers have been included
for each wheel. The model contains all the elements needed to predict both the
nature of ground-vehicle engagement and the full resultant motion of the car.
VEHICLE DYNAMICS
(a) (b)
Figure 1: a) The CAD geometry of the Toyota Hilux model; b) An exploded view
of one Double-Wishbone suspension set and its connections to the chassis.
A CAD model of a Toyota Hilux truck (5 m length, 1.8 m wide and 1.6 m tall)
which is used in this study is shown in Figure 1a. The truck assembly consists of:
a chassis; independent Double-Wishbone suspension for each wheel; and wheels.
The tyres used here have a smooth tread. Steering, braking and drive-train control
systems, and more complex tyre tread patterns have not yet been included in the
model but will be implemented in the future.
The vehicle’s load is directed through the contact points between tyres and
ground (and not through the vehicle’s centre of mass). Therefore moments are
generated that cause the vehicle body to pitch forwards or backwards as the vehi-
cle changes speed, or roll from side to side when the wheels encounter pot-holes
or undulations in the road. This modifies the load distributed onto each wheel, and

in turn modifies the total tyre-ground traction available for vehicle propulsion.
Consolidation and compaction [2] (or even destabilization) of a loose terrain un-
der load and shear will further modify the tyre-ground traction. Models based on
simplified kinematics (with specified positions or motion of the wheels) are un-
able to capture the dominant dynamics that control motion.
The suspension, wheels and chassis are coupled together using pivot joints that
we model as zero-length spring and dashpot forces as shown in Figure 1b. The
colors represent the different connections. Red, dashed lines indicate where the
suspension component joints are connected to the chassis. The blue lines represent
the spring forces in the coil springs, and the gray line indicates the axle connec-
tion. The axle of each wheel is connected to the chassis gearbox via a single joint.
The upper and lower arms of the suspension are attached to different parts of the
chassis via pairs of joints that constrain the motion of each arm to a single rota-
tional degree of freedom. The other end of each arm is attached to a vertical strut
by more pivot joints (with the same rotational constraints as before). The simple
rectangular strut represents the connections of a normal suspension to the wheel
and is attached to the wheel by three joints along the axle to keep the strut posi-
tioned upright and parallel to the plane of the wheel. The bottom of the shock ab-
sorber is connected to the middle of the lower suspension arm by a pair of joints,
and the coil spring from the top of the shock absorber is modeled using four
springs of 275 mm in length attached to the chassis. This suspension is the same
for each wheel, and thus by four-fold symmetry gives 72 spring forces in the sus-
pension assemblies. Spring stiffnesses ranged from 106 – 107 N/m. A timestep of
2x10-5 s provides stable numerical integration for the assembly. The truck compo-
nents are all modeled as fully dynamic bodies and the wheels are set in motion by
specifying a constant angular acceleration.
The ground layer consists of 14,000 stably packed SQ particles (representing
coarse gravel with diameters between 10 and 30 mm, and shape index in the range
3 – 6. The truck is initially stationary and is supported by the contact forces be-
tween wheels and ground. A steady angular acceleration of 5 rad/s2 was applied to
each wheel. A static friction of 1.0 and dynamic friction of 0.8 were used for
tyre/rock and 0.6/0.4 for rock/rock interactions. The restitution coefficient was 0.5
for tyre/rock and 0.3 for rock/rock. For details of the DEM implementation used
here, see Cleary [6].

MOTION OF TRUCK DRIVING OVER GRANULAR TERRAIN

(a)
(b)
(c)
Figure 2: The truck motion over the rocky terrain. The rocks are colored by speed
with purple (stationary), blue (low-speed) and red (2 m/s).

Figure 2 shows the truck at different positions. It propels itself forward due to
traction forces applied by the tyres to the gravel. The rocks are colored based on
speed. As each wheel comes into contact with new parts of the terrain, the rocks
surrounding the wheel are mobilized. Interaction between the vehicle assembly,
tyres and ground result in wheel motion that is a combination of rolling (via trac-
tion forces) and plowing through the relatively loose surface layer of rocks. The
latter occurs because there is mild but observable sinkage of the wheels into the
terrain. The load distributed through each tyre-ground contact then exerts pressure
on the granular terrain in front of the moving vehicle. This is demonstrated by
rocks well in advance of the approaching wheels being forced into motion. The
plowing action of the wheels tends to push these rocks sideways and around the
wheel.
(a) (b)
Figure 3: a) Forward motion of the truck; b) Vertical position of the wheels.

Figure 3a shows the truck accelerating over four seconds. At 6 s, the vehicle
speed is about 1/5 the tangential speed of the wheels and the acceleration shows
considerable variability. This is due to low traction between the smooth tyre tread
and the ground, and to the shifting granular material under each wheel. Figure 3b
shows the vertical position of the wheels with time. The front wheels (objects 3
and 4) and back wheels (objects 5 and 6) move up and down together as pairs.
This is mainly due to the relatively flat nature of the terrain across the width of the
car. Occasionally the wheels either climb over larger rocks or sink into the surface
layer giving up to 200 mm variation in their positions. Each vertical impulse felt
by the tyre-ground contacts varies the torque about the centre of mass of the truck.

The variable torque causes the truck to pitch backwards and forwards and is the
reason for the differences observed in the front and back wheel heights.
CONCLUSION
Dynamical models of multi-component vehicles driving over granular terrain
are now possible. The model in this paper contains all the elements needed to pre-
dict both the nature of ground-vehicle engagement and the full resultant motion of
the vehicle. We observed destabilization of the surface layer of the ground due to
the wheels sinking and plowing through the loose rocks. Quantitative predictions
of vehicle motion demonstrated non-linear acceleration due to the vehicle “bounc-
ing around” on the uneven terrain. This resulted from the vehicle repeatedly pitch-
ing forwards and backwards, and experiencing variable traction on different
wheels.
The next steps for this model will be to add in realistic tyre-tread patterns and
tyre deformation. We will also explore more complex, terrain topography and ex-
tend the truck description to include full drive-train, steering and braking control.
REFERENCES
1. Yoshimoto, K. and Suetomi, T., J. Mech. Systems for Transport. and Logistics,
2 159-169, 2008.
2. Tordesillas, A. and Shi, J., Proc. 13th Int. Conf. of ISTVS, 1 201-208, 1999.
3. Nakashima, H., Takatsu, Y., Shinone, H., Matsukawa, H. and Kasetani, T., J.
Mech. Systems for Transport. and Logistics 2 55-65, 2009.
4. Horner, D.A., Peters, J.F. and Carrillo, A., J. Eng. Mech., 127 1027-32, 2001.
5. Cleary, P.W., “Granular Flows: Fundamentals and Applications”, Granular and
Complex Materials, ed. T. Aste, pub. World Sci., pp. 144-145, 2007.
6. Cleary, P.W. and Prakash, M., Phil. Transaction of Royal Society, 362 2003-
2030, 2004.

CHARGE AND STRUCTURE BEHAVIOUR IN A

TUMBLING MILL
Pär Jonsén1, Bertil I. Pålsson2, Kent Tano3, Andreas Berggren4

1
Division of Solid Mechanics, Luleå University of Technology, SE-97187 Luleå, Sweden
2
Division of Mineral Processing, Luleå University of Technology, SE-97187 Luleå, Sweden
3
Technology and Business Development, LKAB, SE-98381 Malmberget, Sweden
4
Technology and Development, Boliden Minerals, SE-93681 Boliden, Sweden
The grinding process in tumbling mills is complex and to include all

phenomena that occur in a single numerical model is today not possi-
ble. This paper presents the results of a study in which the deflection
of a lifter bar in a pilot ball mill is measured by an embedded strain
gauge sensor and compared to deflections predicted from finite ele-
ment (FE) simulations. The flexible rubber lifter and the lining in a
tumbling mill are modelled with the finite element method (FEM) and
the grinding medium modelled with the distinct element method
(DEM). The deflection profile obtained from DEM-FE simulation
shows a reasonably good correspondence to pilot mill measurements.
The approach presented here is a contribution to the validation of
DEM-FE simulations and an introduction to the description of a bend-
able rubber lifter implemented in a DEM-FEM mill model. It opens up
the possibility to predict contact forces for varying mill dimensions
and liner combinations. FEM is especially valuable in this case, since
there are readily available libraries with material models. This is a fol-
low-up work to previous preliminary result from a mono-size ball
charge interaction study.

INTRODUCTION
The behaviour of tumbling mills is complex and several parameters do signifi-
cantly influence the effectiveness of the grinding operation. Many of these pa-
rameters are either difficult or laborious to measure. Understanding of the charge
motion within the mill is of importance in mill optimisation. Both the breakage of
ore particles, deformation of the lining and the wear of liners/ball media are
closely linked to the charge motion. To study these phenomena in a physically
correct manner, suitable numerical models for the different parts of the mill
system is required.
To model charge motion and its interaction with the lining, the grinding balls
are modelled with DEM and the mill structure modelled with FEM. All parts of
the mill system will affect its response and a DEM-FEM model gives the
opportunity to study the influence of the mill structure. The validation of this task
is done by comparing numerical results with experimental measurements from
pilot grinding with an instrumented rubber lifter. This work is a continuation of
previous preliminary work on DEM-FEM modelling of tumbling milling
processes, see Jonsén et al. [1].
EXPERIMENTAL SETUP
The pilot mill is 1.414 m in diameter and 1.22 m in length. It is a grate-
discharge mill, equipped with 12 rubber lifters of square size 0.1 m and a face an-
gle of 45 degrees. Steel balls with a diameter ranging between 10-30 mm and a
density of 7800 kg/m3 were used in the experimental tests. The test material, a
hematite pellet feed with d50 around 35µm and a solids density of 5200 kg/m3,
was chosen to obtain stable grinding conditions with respect to feed size varia-
tions. Feed rate was kept constant at approx. 1.5 tonne/h.
One of the lifters has a strain gauge mounted on the leaf spring that converts
this deflection to an electric signal. The signal is then amplified, filtered and
transmitted to a computer. The sensor system is marketed by Metso Minerals un-
der the name Continuous Charge Measurement system (CCM). As the mill rotates
and the lifter with the sensor dips into the charge, the force acting on the lifter in-
creases, which in turn, causes a deflection.

MODELLING
The model of the mill is a combined DEM-FEM model; the charge consists of
DEM particles and the mill structure is modelled with FEM. The mill structure
consists of rubber lifter and liners and a mantel made of solid steel. For the elastic
behaviour of the rubber a Blatz-Ko hyper-elastic model is used [2]. Experimental
data for the rubber was provided by the supplier of the lining. The mill mantel is
modelled as a rigid material. The original length of the mill is 1.22 m, however, in
the model only a slice of 0.10 m is modelled. Fully integrated eight node solid
elements with a reduced integration of the pressure part are used to model the
structure.
a b
Fig 1. In a. the non-graded charge with 20 mm diameter particles and in b. the

graded charge with a diameter distribution of 15-25 mm.
A total number of 9401 solid elements are used to model the structure. The
friction coefficients of the particle-particle contact was set to µ= 0.5 and for the
particle-wall contact to µ= 0.9 based on values published by Rajamani [3]. For the
contact between the particles and the structure a “nodes to surface” contact is
used. A viscous damping of 60% is used in this contact based on the particle-
particle bouncing collision behaviour observed in the current calculations. Be-

cause of the two-dimensional representation, the balls in the numerical mill are
represented by rods rather than spheres. In this case, the disk thickness is set equal
to the length of the pilot mill model, 0.1 meters. In order to mimic the mass and
inertia of a string of charge balls rather than a solid rod, the density of the parti-
cles representing the charge was lowered accordingly. The consequence of this
two-dimensional representation is that any single particle impacting the lifter
within the two-dimensional model is actually simulating the simultaneous impact
of a string of charge balls running the full length of the mill. Such simultaneous
impact is not likely happening in reality and so this should be considered when
interpreting deflection results from the 2D model.
1 Diameter 15−25 mm
Diameter 20 mm
Measured
0.8
Normalised deflection [−]
0.6
0.4
0.2
0 50 100 150 200 250

Rotation angle [degrees]
Fig 2. Displacement of the lifter during passage through the charge. Compari-
son between graded charge, non-graded charge calculations and measurements.
RESULTS
In the numerical study the mill fill was J = 25% and speed 73% of critical rota-
tional speed. Two different ball charges are used in the study. One with a constant
particle diameter of 20 mm and one with a particle diameter distribution from 15
to 25 mm, one third of each type, also called graded charge. In Fig 1 the two cases
are shown at steady state. The graded charge tends, according to the simulation to

have more voids in the packing pattern of the particles and the toe angle is
smaller.
Fig 3. A snapshot of the von Mises’ stress field for a part of the mill model
during its passage through the charge.
To study the charge impact on the mill structure during the milling process one
can study forces and deflections of a lifter in the system. To observe the forces
that go through the system of liners a section plane is introduced in the liner be-
hind the studied lifter. The structure will respond to deformation upon the incom-
ing load from the charge. The deflection of the lifter is also analysed and com-
pared to the experimental measurements, see Fig 2. These deformations will give
rise to strains and stresses which are dependent on the material properties of the
structure. A snapshoot of the von Mises’ stress field for two lifters and the liner in
between and during their passage in the charge is shown in Fig 3. This is one of
the additional properties that are available from the FE-program.

CONCLUSION
The DEM-FEM model can predict the classical DEM results, but can also
predict responses from the mill structure like, e.g., stress and strain. The model
gives the opportunity to optimize the material selection of the mill structure. Criti-
cal response values for e.g. stress and strain can be identified during the milling
process. Forces and mechanical waves in the mill structure can be found. The ma-
jor difference between DEM only and DEM-FEM models is that the latter give a
direct coupling between force, stress and displacement for the whole mill system.
To include the mill structure response in the numerical models gives more in-
formation regarding charge motion and results in better correlation between ex-
perimental measurements and numerical models.
ACKNOWLEDGMENTS
For financial support of the project ModPulp project number P341106-1 within
Gruvforskningsprogrammet are Vinnova gratefully acknowledged.
REFERENCES
1. P. Jonsén, J. Alatalo, B. I. Pålsson, K. Tano, ‘Prediction of contact forces be-
tween a grinding charge and mill lifters’ 12th European Symposium on Com-
minution and Classification, 2009,
2. P. J.Blatz and W. L. Ko, ‘Application of Finite Element Theory to the Defor-
mation of rubber Materials’, Trans. Soc. of Rheology, 6, 223-251., 1962
3. R. K. Rajamani, Semi-Autogenous Mill Optimization with DEM Simulation
Software. In: Control 2000 – Mineral and Metallurgical Processing. Society
for Mining, Metallurgy, and Exploration, Inc., Littleton, CO., USA. pp. 209-
215, ISBN 0-87335-197-5, 2000

Page 1
ANAYSIS OF PARTICLE CHARING BEHAVIOUR

VIA ROTATING CHUTE OF BLAST FURNACE BY
USING DISCRETE ELEMENT METHOD
Hiroshi Mio1, Shinroku Matsuzaki1, Kazuya Kunitomo1, Jusuke Hidaka2

1. Nippon Steel Corporation, Environment & Process Technology Center
2. Doshisha University, Department of Chemical Engineering & Materials Science
The objective of this paper is to analyze the particle behavior at a

charging process of blast furnace by using Discrete Element Method.
The effect of chute inclination angle on the flowing behavior or the
particle segregation was discussed. The charging area of burden from
the rotating chute became closer with decreasing the chute angle, and
the charged layer was collapsed by the following charged particles.
The particles, which were charged at early time, were not able to
move toward the center of blast furnace, on the other hand, the others
moved.
INTRODUCTION
A blast furnace is a huge reactor having more than 5000 m3 in volume for pro-
ducing pig-iron from iron ore particles. Ore and coke particles are being stacked
alternately in layer, and the gas having high temperature is blown from tuyeres.
Iron ore particles are reduced during descending, and many physical changes and
chemical reactions between each phase are undergoing in there. It has been
thought that the burden distribution affects the operating efficiency strongly. Be-
cause the void fraction, the gas flow distribution and the burden descending be-

Page 2
havior depend on the burden distribution, and the particle segregation during
charging at the top of blast furnace affects it very much. Therefore, people want
to control the particles location in the blast furnace when particles are charged in
there. For this reason, the rotating chute is installed at the top of the blast furnace
for charging equipment. The burden may be charged at the suitable area by
changing the inclination angle of the chute.
However, it is quite difficult to confirm the results which people expected, be-
cause there is massive dust during charging. Therefore, the computational simula-
tion using Discrete Element Method (DEM) [1] is required for the estimation and
optimization of the charging processes. DEM is one of the most reliable simula-
tion methods for analyzing the solid particle behaviour. The authors have pro-
posed the method to solve the effect of particle shape on flowing behaviour; i.e.
each particle has its own rolling friction [2]. The particle behavior and the segre-
gation during flowing through the laboratory-scale chute analyzed by the simula-
tion agreed with the experimental results very well, and its agreement could be
seen under several different conditions (chute angle, installation of a damper or
particle mixture) [2]. Therefore, this simulation method has a high potential to
analyze the particle segregation during charging to the actual blast furnace. In this
paper, the particle charging behavior of sintered ore at the charging process of
blast furnace was simulated by using DEM. The inclination angle was changed
during charging like an actual operation, and the burden position for each inclina-
tion angle and the particle size segregation were discussed.
SIMULATION
The three-dimensional solid particle behavior in the charging process of blast
furnace was simulated by using DEM. This simulation method consists of an idea
of determining the kinematic force to each finite-sized particle. All forces acting
on each particle are modeled and calculated at every discrete-time step. The tra-
jectories of particles are updated by Newton’s low of motion, and The contact
model between two particles is given by Voigt model, which consists of a spring-
dashpot and a slider for the friction in a tangential component. The shape of par-
ticle in DEM is usually assumed to be spherical due to the ease of contact detec-
tion or calculation of contact force, although the shape of sintered ore particle is
completely irregular. Thus, its effect on the particle motion must be considered in

Page 3
DEM by setting the proper rolling friction for each particle, and it is given in fol-
lowing equation.
3 ω
M r , i = − α i b Fn i (1)
8 ωi
Where, b is a radius of contact area, and αi denotes the coefficient of rolling

friction. Every particle has different αi, because the shapes of sintered ore
particles are totally different from each other. Its distribution is related with the
rollability of particle [2]. The method having individual rolling friction had given
the significant agreement with the experimental results [2], therefore, this method
was applied to the present work.
The bell-less type (rotating chute) charging system of the blast furnace was
modeled in this work. The length of horseshoe-shaped chute was about 5.3 m,
and a damper was installed at the outlet of the chute with 26º. 31 tons (density, ρ:
3300 kg/m3, Young’s modulus, E: 35 GPa, Poisson’s ratio, ν: 0.25) of sintered ore
were positioned in the hopper randomly and packed. The coke layer was made at
the top of blast furnace, and its surface profile was in agreement with the actual
one. The number of coke particles was 112712 (ρ: 1050 kg/m3, E: 5.4 GPa, ν:
0.22). The particle size distributions of sintered ore and coke were related to
actual ones, however the particle sizes were adjusted relatively three times bigger
than actual ones to reduce the number of calculated particles (sintered ore:
22.5~97.5 mm; coke: 67.5~262.5 mm). Numbers of particles were 520231
(sintered ore) and 112712 (coke). The particles in the hopper were discharged on
the rotating chute, and they were charged into the blast furnace via the chute. The
chute incination angle against the virtical direction, θc, was changed from 43.1° to
36.9° every two or three rotations (see Table 1), and the total number of rotations
was 16. The rotational speed was 8.0 rpm. The coke layer in the blast furnace
didn’t descend during charging, and the effect of gas flow was not considered. ∆t
was 1.75 µs, and the total number of calculation steps was about 28 millions steps.
All calculations were executed by parallel computing using OpenMP.

Page 4
Chute inclination angle 43.1° 42.0° 40.9° 39.5° 38.2° 36.9°

Number of rotations 3 3 2 3 3 2
Table 1. Condition of chute inclination angle
RESULTS AND DISCUSSIONS

Figure 1 shows the snapshots of particle charging behavior under the different
chute inclination angles (1st, 4th, 7th, 9th, 12th and 15th charging). The particle
colors are changed as the chute angle, and all coke particles are black. The dis-
tance between the charging point and the center of blast furnace was also shown.
The charging point shifted to the center with decreasing the inclination angle.
The charged layer was compressed and widened by the following charging, and
the particles slid toward the center. When the chute inclination angle was de-
creased, the particles were charged at the slope of the heap. Then they were able
to move toward the center easily.
LR=4.10 m 3.76 m 3.66 m
(a) 43.1° (b) 42.0° (c) 40.9°
3.42 m 3.31 m 3.08 m
(d) 39.5° (e) 38.2° (f) 36.9°

Fig.1 Snapshots of charging behavior

Page 5
Figure 2 shows the changes of radial distribution of burdens which were

charged under 43.1° or 42.0° in chute angle (1st to 3rd rotation; red particles in
Fig.1 or 4th to 6th, cyan particles). Although the distribution shifted slightly to
the center, these particles hardly move toward the center under 43.1°. Because
they are pressed into the coke layer. On the other hand, the particles, which were
charged under 42.0°, moved toward the center continuously. Some of them
reached around the center.
Figure 3 shows the relation between the relative charged mass of sintered ore
particles and the radial distance. It is also found from this figure that most of
particles near the wall were the ones which were charged during 1st to 3rd
charging. The ones charged after 42.0° charging reached near the center of blast
furnace. Because they are charged on the sintered ore layer, they can slide easily
toward the center comparing to the case which the sintered ore particle were
charged on the surface of coke layer (they have larger size and smaller dinsity). It
is very important to know the traveling distance of particles under each inclination
angle condition and the simulation using DEM is very helpful for grasping the
phenomena.
100 100
After 3 charges After 6 charges
Specific charged mass [kg/m ]
Specific charged mass [kg/m ]

2
80 After 8 charges 80 After 11 charges

60 After 16 charges 60
40 40
20 20
0 0
0 1 2 3 4 5 0 1 2 3 4 5
Radial distance [m] Radial distance [m]
(a) 43.1° (b) 42.0°

Fig.2 Cahnges of radial distribution of burdens

Page 6
1.0 o
43.1
o
42.0
Relative charged mass [-]

0.8 40.9
o
o
39.5
o
0.6 38.2
o
36.9
0.4
0.2
0.0
0 1 2 3 4 5
Radial distance [m]
Fig.3 Relation between the relative charged mass and the radial distance
CONCLUSION
The particle behavior at the charging process of blast furnace was simulated by
using DEM. The effect of chute inclination angle on the flowing behavior was
discussed. The charging area of burden from the rotating chute became closer to
the center with decreasing the chute angle, and the charged layer was collapsed by
the following charged particles. The particle behavior during charging can be
simulated, and this process will be optimized by using DEM.
REFERENCES
1. P.A. Cundall, O.D.L. Strack, ‘A Discrete Numerical Model for Granular As-
semblies’, Geotechnique, 29, 47-65, 1979.
2. H. Mio, S. Komatsuki, M. Akashi, A. Shimosaka, Y. Shirakawa, J. Hidaka, M.
Kadowaki, S. Matsuzaki, K. Kunitomo, ‘Validation of Particle Size Segrega-
tion of Sintered Ore during Flowing through Laboratory-scale Chute by Dis-
crete Element Method’, ISIJ International, 48, 1696-1703, 2008.

STUDY ON CUTTING FORCES OF SIC MACHINING

PROCESS WITH PRE-STRESSED USING DEM
SIMULATION
Shengqiang Jiang1 Yuanqiang Tan1 Dongmin Yang1,2 Yong Sheng2

1
School of Mechanical Engineering, Xiangtan University, Hunan 411105, China
2
School of Civil Engineering, University of Leeds, Leeds, LS2 9JT, UK
Based on the theory of metal pre-stressed machining, the method of

pre-stressed suitable for ceramic materials machining was presented.
And using the cluster method and bonded particle model (BPM), the
discrete element method (DEM) model of ceramic materials of pre-
stressed machining was established. An orthogonal test was applied to
select the optimum machining parameters which included pre-stressed
size, tool rake angle, depth of cut, tool edge radius and cutting speed
in the SiC ceramic pre-stressed machining process. It was found that
the cutting force was basically changed a little under the action of pre-
stressed. Based on the orthogonal test results, the influences of five
elements on cutting force were discussed.
INTRODUCTION
Ceramic materials have been widely used as functional materials and structural
materials in many industrial fields. However, due to their inherent high hardness
and brittleness, grinding has become the main machining method. In the process
of machining ceramics material, the damage that caused by tools contact with the
ceramics surface is mainly reflected in the surface/sub-surface cracks and residual

stress [1]. Clearly, owing to the high machining cost of ceramic components and
the difficulty of controlling and measuring the damage of machined surface, its
wider application is limited. Therefore, the study of damage and control mecha-
nism has become an important research topic of ceramic machining and its appli-
cation.
Pre-stressed cutting method was first proposed by Zhou [2] in the machining of
metallic materials, which had been used in turning, grinding and milling. In this
paper, based on the theory of metal pre-stressed machining, the method of pre-
stressed suitable for ceramic materials machining was presented. The influence of
pre-stressed size, depth of cut, cutting speed, tool rake angle and tool edge radius
to the tangential force and normal force were investigated.
DEM SIMULATION
DEM model of SiC ceramic. Discrete element method (DEM) was introduced
by Cundall [3] for analysis of rock-mechanics problems and then applied to soils,
concrete and ceramics. Bonded particle model (BPM) was adopted to simulate a
kind of cemente with a finite size joining the two particles. In order to represent
the complex-shaped grains of SiC, we make a number of particles bonded into a
cluster (see Fig 1). Here we set the maximum allowable number of particles in a
cluster equal to 7.
In order to have confidence that a particular model is reproducing desired
physical behavior, it is necessary to calibration the micro-parameters of DEM
model. Here we used the uniaxial compressive test, the three-point bending test
and the fracture toughness test [4] to calibrate the particle properties in polycrys-
talline SiC DEM model.
DEM model of pre-stressed machining. In DEM cutting model, the work-
piece which represents for the polycrystalline SiC parts was set up by the packing
particles into a cell surrounded by three fixed walls, as shown in Fig 1. The di-
mension of the work piece was 0.85mm in length and 0.3mm in height, consisting
of 2219 clusters and 10909 balls. There is no friction between the work piece and
the walls. A two-double wall was used as a cutting tool which ploughed into the
work piece and moved with a cutting speed of v to simulate a real cutting process.
The stiffness of the walls was much bigger than that of the particles to avoid any

distortion of the cutting tool. The velocities of lateral walls are controlled by a
servo-mechanism that maintains a specified pre-stressed P. The method of or-
thogonal test was used to investigate the influence of machining parameters to
cutting force. Pre-stressed size, depth of cut, cutting speed, tool rake angle and
tool edge radius were selected as controllable independent variables of DEM
simulation named A, B, C, D and E, respectively, each at four levels and their test
levels were listed in Table 1.
γ
ρ v
α
P P
Fig 1 DEM model of the pre-stressed cutting process.

Elements and levels A (P/MPa) B (ap/µm) C (v/m/min) D (ρ/µm) E (γ/º)
1 100 10 5 5 5
2 300 20 10 10 -5
3 500 30 15 15 -15
4 700 15 20 20 -25
Table 1 The orthogonal test factors and levels

The test results were listed in Table 2-3. In the orthogonal test, K is the sum of
target function of every level and k = K/4. R is equal to max {ki} – min {ki}, and R
scales the effect of variables on the result. High R value of variable means this
variable has a strong effect on the result.
Fig 2 shows the effects of five elements on cutting force (tangential force and
normal force) according to the quantity of pre-stressed, tool rake angle, depth of
cut, tool edge radius and cutting speed by comparing level values in Table 1. It

can be seen from Fig 2(a) that the tangential force and normal force are basically
changed little under the action of pre-stressed. The tool rake angle is also effect
the outcomes, and the tangential force and normal force are decreased when the
rake angle changes form -25° to 5°. It is indicated that the small rake angle can
reduce the cutting force. The larger the depth of cut, the higher the tangential
force and the normal force are. The tangential force is smaller than the normal
force when the depth of cut is less than 12.5μm, but it becomes higher as the
quantity of depth of cut increases. The effect of tool edge radius to cutting force is
just as shown in Fig 2(d). As the increasing of tool edge radius, the tangential
force gradually decreases, but the normal force is increased. It may be attributed
to the increased cutting edge’s passivation and the increased contact surface areas
in y-direction between the tool and materials cause the cutting force changed from
the tangential force to normal force. The lower the cutting speed, the lower the
cutting force is.
Test no. A B C D E Fx(N) Fy( N)
1 1 1 1 1 1 18.7 22.8
2 1 2 2 2 2 92.0 58.9
3 1 3 3 3 3 136.8 99.3
4 1 4 4 4 4 87.8 96.6
5 2 1 2 3 4 93.5 71.7
6 2 2 1 4 3 26.9 64.8
7 2 3 4 1 2 82.5 53.2
8 2 4 3 2 1 137.3 91.4
9 3 1 3 4 2 103.5 80.1
10 3 2 4 3 1 37.0 50.4
11 3 3 1 2 4 73.6 63.4
12 3 4 2 1 3 140.5 78.1
13 4 1 4 2 3 61.4 46.4
14 4 2 3 1 4 149.9 80.3
15 4 3 2 4 1 53.4 74.5
16 4 4 1 3 2 60.4 68.7
Table 2 The orthogonal test results with L16 (4) 5

Tangential force Fx (N) Normal force Fy (N)

A B C D E A B C D E
K1 335.3 277.2 179.7 391.6 246.5 277.6 221.1 219.7 234.4 239.0
K2 340.3 305.8 379.4 364.3 338.4 281.1 254.4 283.2 260.1 260.9
K3 354.8 346.3 527.5 327.7 365.7 272.0 290.3 351.1 290.1 288.7
K4 325.0 426.1 268.8 271.7 404.9 269.9 334.8 246.6 316.0 312.0
k1 83.8 69.3 44.9 97.9 61.6 69.4 55.3 54.9 58.6 59.8
k2 85.1 76.5 94.9 91.1 84.6 70.3 63.6 70.8 65.0 65.2
k3 88.7 86.6 131.9 81.9 91.4 68.0 72.6 87.8 72.5 72.2
k4 81.3 106.5 67.2 67.9 101.2 67.5 83.7 61.7 79.0 78.0
R 7.4 37.2 87.0 30.0 39.6 2.8 28.4 32.9 20.4 12.8
Table 3 Analysis of L16 (4)5 test results
135 (c)
135 (a) 135 (b) Tangential force
Tangential force 120
120 Tangential force Normal force
Normal force 120
Normal force
105
Cutting force (N)
105 105
Cutting force (N)
Cutting force (N)
90
90 90
75
75
75
60 60 60
45 45 45
100 200 300 400 500 600 700 -25 -20 -15 -10 -5 0 5 10 15 20 25 30
Pre-stress size (MPa) Tool rake angle (degree) Depth of cut (μm)
135 135
(d) (e)
Tangential force Tangential force
120 Normal force 120 Normal force
105 105
Cutting force (N)
Cutting force (N)
90 90
75 75
60 60
45 45
4 6 8 10 12 14 16 18 20 22 4 6 8 10 12 14 16 18 20 22
Tool edge radius (μm) Cutting speed (m/min)
Fig 2 Effects of pre-stress size, tool rake angle, depth of cut, tool edge radius and
cutting speed on the cutting force (tangential force and normal force).

CONCLUSION
In this study, a DEM method was used to simulate the cutting process of SiC
ceramic under the action of pre-stressed. And using the orthogonal test method,
the effects of the five cutting parameters on cutting force were studied. From the
test results we can see that: using the method of pre-stressed machining, the cut-
ting force was fluctuated within a small range, so it didn’t bring the difficult to
cutting the materials. And under the action of pre-stressed, with the increasing of
depth of cut and the cutting speed, the cutting force increased lineally but it de-
creased with the tool rake angle changing from negative to positive. And also the
tangential force increased with the increasing of tool edge radius while the normal
force decreased.
ACKNOWLEDGMENTS
This research work was supported by the National Natural Science Foundation
of China (50875224).
REFERENCES
1. R. Hessert, B. Eigenmann, O. Vöhringer, D. Löhe. Fracture mechanical
evaluation of the effects of grinding residual stresses on bending strength of
ceramics. Materials Science and Engineering A, 234-236, 1126- 1129, 1997.
2. Z.H. Zhou, D.T. Guo. Pre-stressed machining: combined use of heuristics and
optimization methods, In Proceedings of IX the ICPR, Cincinnati, 1, 257-262,
1987.
3. P.A. Cundall, O.D.L. Strack. A discrete numerical model for granular assem-
blies, Geotechnique, 29(1), 47–65, 1979.
4. Y.Q. Tan, D.M. Yang, Y. Sheng. Study of polycrystalline Al2O3 machining
cracks using discrete element method, International Journal of Machine tool
& Manufacture, 48(9), 975-982, 2008.

DEM SIMULATION OF PARTICLE MOTION IN A

PADDLE MIXER
Ali Hassanpour1, Hongsing Tan2, Andrew Bayly2, Prasad Go-
palkrishnan3, Boonho Ng2 and Mojtaba Ghadiri1
1
Institute of Particle Science and Engineering, School of Process, Envi-
ronmental and Materials Engineering, University of Leeds, UK
2
P&G Technical Centre Ltd., Newcastle Upon Tyne, UK
3
P&G Co., ITC, Cincinnati, Ohio, USA
This article compares DEM simulations of particle flow in a paddle mixer to
experimental measurements using Positron Emission Particle Tracking
(PEPT). Good qualitative agreement is shown between DEM and PEPT in
terms of flow patterns inside the mixer. Quantitative comparison of PEPT
measurements and DEM analyses shows reasonable agreement in terms of
velocity distribution. In general, the DEM is shown to be capable of predict-
ing the dynamics of the particulate flow inside the mixer.
INTRODUCTION
Powder mixing is a fundamental process operation in many applications such
as detergent, cosmetic, food and pharmaceutical manufacturing. A wide variety of
mixers are used depending on specific application requirements, including batch
and continuous mixers with a wide range in shear rates, residence times and spe-
cific energy inputs. Understanding, control and optimisation of these processes are
of strategic importance, but at the same time very challenging. To optimise the
above process operations, a key step is to understand powder kinematics behav-
iour (flow fields, mixing patterns etc.) to enable efficient process design and con-
trol. There has been considerable work reported in the literature to characterise the
powder flow behaviour during mixing under various process conditions [e.g. 1-3].
However, it is difficult to obtain an insight to the internal flow field during mixing
processes and to address the kinematic behaviour of powders using experimental
approaches, particularly at large scales. Most experimental measurements of in-
ternal flow are based on Positron Emission Particle Tracking (PEPT) and are suit-
able for smaller scales [3-4]. The capability to model mixing process in mixers

using Distinct Element Method (DEM) has grown in relation to computing power.
It is now possible to simulate complex equipment geometries. However, computer
power is still inadequate for simulating full populations of particles (millions and
billions). A number of researchers have considered the use of larger particles in
the simulations to reduce the particle number and simulation time [5-7], whilst in
some cases the properties of large particles are being calibrated [6]. However re-
liability of this method in terms of predicting actual powder flow characteristics
remains a fundamental question and needs to addressed.
The work here is an effort to compare the powder flow predicted by from
DEM with experimental data. Here, a 6 litre Forberg commercial batch paddle
mixer is used as the subject of investigation. In a separate study the internal pow-
der flow dynamics of the mixer was investigated using PEPT measurements under
various process conditions, i.e. fill level, impeller rotation speed and particle den-
sity, at the University of Birmingham [8], providing experimental validation for
the model. The modelling results are compared with PEPT data in terms of the
particle velocity distribution.
MODEL SET-UP
For analysis of flow field of the mixer, the EDEM computer code developed by
DEM-solutions is used as it is capable of generating complex geometries, i.e. ves-
sel and the impellers. Particle motion in the mixer has been analysed to determine
the granular dynamics, and in particular to identify the velocity field. A 3-D rep-
resentation of the mixer is created using a CAD software and is imported into the
EDEM computer code (Fig. 1a). As it can be seen, the mixer consists of two inter-
sected semi-cylinders of the same span and two counter-rotating impellers, each
with 10 paddles positioned pair-wise along 5 axial positions. Particles are generat-
ed randomly at spatial locations above the impellers (the position shown in Fig.
1a). In the current simulations the Hertz-Mindlin model is considered for the nor-
mal and tangential interactions and at this stage the powders are considered to be
non-cohesive. This is because the experimental work using PEPT are carried out
on dry, free-flowing particles. The schematic flow field inside the paddle mixer
simulated by DEM is shown in Fig. 1b.

Position of
particle
generation
Impellers
(a) (b)
Fig 1. The imported geometry (a) and the schematic flow fields (b) inside the
paddle mixer simulated by DEM.
Due to the limitation of computer power, it is not possible at this stage to simu-
late the actual number of particles inside a 6l mixer (around 50 millions) within a
reasonable time. Therefore the simulation should be carried out with smaller
number of particles using larger particles. When larger particles are used in the
simulations, densities are adjusted to maintain a similar momentum exchange be-
tween particles in DEM [6]. However, it is essential to elaborate the effect of this
simplifying assumption on the dynamics of powders inside the mixer by simula-
tion. Therefore, initially the sensitivity of powder dynamics on the particle size
was analysed in this work. Systems with various particles sizes (Table 1) were
simulated and their particle dynamics in terms of time averaged particle velocity
were compared.
Simulation Case A B C D
Particle diameter (mm) 11.40 7.20 4.52 2.26
Total particle number 7,000 15,000 60,000 500,000
Net simulation time for 3 6 54 580

10s real time (hrs)
Table 1. The modelling parameters for the sensitivity analysis on particle size.

The steady state average velocity as a function of particle size is presented in

Fig. 2 and it can be seen that it slightly decreases as the particle size is reduced.
This analysis shows, although the average particle velocity is indeed sensitive to
the particle size, the effect is not very significant. For the simulation of powder
flow in the mixer under various process conditions (corresponding to the available
experimental results) it is essential to use a modelling system which produces re-
sults in a reasonable time (e.g. in this work to be less than a week). Therefore the
modelling conditions of case C, i.e. particles with 4.52mm size, were used in fur-
ther simulations.
Average Normalised Velocity (m/s)
0.5
0.45 Actual particle size
0.4
0.35
0.3
0.25 y = 0.0089x + 0.3591
0.2 R2 = 0.9875
0.15
0.1
0.05
0
0 2 4 6 8 10 12
Size (mm)
Fig 2. Steady state average particle velocity as a function of particle size.

For the simulation under various process conditions, spherical particles, were
generated in similar way to the previous section. After reaching a stable particle
bed the mixer model was activated, running at various process conditions, i.e. dif-
ferent impeller speed, fill level, and particle density. For confidentiality reasons
it is not possible to disclose the impeller rotational speeds. Real-time simulations
lasted for up to 10 seconds. Simulation conditions were selected based on availa-
ble experimental data from PEPT, with the exception that experiments were done
with finer particles (D50 < 1 mm) and a broader size distribution. For the experi-
mental work, the PEPT facility of the University of Birmingham (Birmingham,
UK) was used to track particle motion. The principles of the PEPT technique and
its capability can be found elsewhere [8].


The time-averaged normalised velocity distribution obtained from DEM analy-
sis is compared with that from PEPT measurements in Fig. 3a for a representative
process condition. The DEM model predicts a smooth distribution of particle ve-
locities while the PEPT data show more scatter or fluctuation in the frequency
plot. Potential explanations for this difference could be attributed to the particle
size discrepancies and/or the sampling method: PEPT analysis is based on data
from only one particle, i.e. the tracer, while the DEM results consider the velocity
profiles of the full population of particles in the simulation. Overall there is rea-
sonable agreement in the velocity distribution.
(a) (b)
Fig 3. Quantitative comparison between the DEM simulation and PEPT meas-
urements for particle velocity distribution (a) and average speed (b).
The results of the average particle velocity, normalised to paddle tip speed, are
compared across a range of process conditions including fill level, impeller speed,
and particle density in Fig. 3b; predicted speeds from DEM are plotted on the ab-
scissa and PEPT results on the ordinate. While there is consistent trend, the aver-
age speeds simulated in DEM are systematically lower than the PEPT measure-
ments over the range of conditions tested. However, this is in contradiction with
our expectations, as it can be seen from Fig. 2. A higher average velocity is ex-
pected for DEM simulations as the particle size in the simulation is significantly
larger that real particles size. A number of reasons could contribute to the sys-

tematic differences observed between the simulations and measurements. One

reason could be due to the single tracer particle sampling method used by PEPT,
discussed earlier. For a precise evaluation, the results of PEPT should be com-
pared to those of a single representative particle in DEM, for which longer simula-
tions (more than 15 minutes in real time) are required. Particle shape could be
another factor – the current simulations consider spherical particles while particles
in the experimental study have somewhat irregular shapes. Any and all of the
above are good topics for future work.
CONCLUSIONS
The powder flow inside a paddle mixer has been simulated using DEM. The
simulation results are compared with experimental measurements using PEPT.
Quantitative comparison of PEPT measurements and DEM analyses shows rea-
sonable agreement in terms of velocity distribution. There is a minor but system-
atic difference between DEM and PEPT time-averaged velocities compared over a
range of process and material conditions. DEM modelling provides useful infor-
mation in understanding of dynamics of particulate processes which is otherwise
difficult to measure by experiment, e.g. flow fields within mixing equipment.
REFERENCES
1. J. Bridgwater, Pow. Tech. 15, 215–236, 1976,
2. P.M. Portillo, M.G. Ierapetritou, F.J. Muzzio, Pow. Tech. 82, 368–378, 2007.
3. B.F.C. Laurent, J. Bridgwater, AIChE. Journal 46, 1723–1734, 2000.
4. B.H. Ng, C.C. Kwan, Y.L. Ding, M. Ghadiri, M., Z.F. Fan, Chem. Eng. Sci.
62(3), 756-765, 2007.
5. Y. Sato, H. Nakamura, S. Watano, Pow. Tech. 186, 130–136, 2008.
6. A. Hassanpour, C.C. Kwan, B.H. Ng, N. Rahmanian, N., Y.L. Ding, S.J. Anto-
ny, X.D. Jia, M. Ghadiri, Pow. Tech. 189, 304–312, 2009.
7. A. Sarkar, C.R. Wassgren, Chem. Eng. Sci. 64 (11), 2672-2682, 2009.
8. D.J. Parker, R.N. Forster, P. Fowles, P.S. Takhar, Nuclear Instruments and
Methods in Physics Research A 477, 540-545, 2002.

PREDICTION OF PARTICLE BREAKAGE IN

AGITATED DRYERS: A COMBINED DEM AND
EXPERIMENTAL APPROACH
Colin Hare, Mojtaba Ghadiri

Institute of Particle Science and Engineering, SPEME, University of Leeds, Leeds, LS2 9JT
During agitated drying of particles shear deformation leads to

breakage and results in a poor product. The rate and extent of this
breakage is difficult to predict and control. This paper aims to eluci-
date breakage rates in agitated contact dryers using a small quantity of
pharmaceutical material to analyse breakage rates in a shear cell. This
is combined with an estimation of the shearing conditions in the dryer
using DEM to predict breakage rates. The prediction compares well
with experimental results, and captures the independency of impeller
rotational speed. Further development of this work will include shear
cell and dryer operations under dryer conditions; elevated temperature
and humidity, and the presence of liquid in the bed.
INTRODUCTION
After crystallisation, pharmaceutical particles have to be dried. A common ap-
proach is to use an agitated vacuum filter dryer. Whilst the material is dried the
bed is rotated to prevent the occurrence of hot-spots. However this agitation leads
to shear deformation which subsequently leads to breakage. Breakage of particles
can lead to waste material, downstream processing problems or even dust explo-
sions. Prediction of the breakage that will occur under these drying conditions is
therefore desirable prior to industrial production.
In order to predict breakage occurring in the dryer the influence of the stress,
strain and strain rate needs to be considered. The method used here is to estimate
the stress and strain rate distributions in the dryer using the Distinct Element
Method (DEM). The extent of breakage occurring in a shear cell under a range of

stresses and strains is quantified and applied to the above results of the DEM sim-
ulations to predict attrition.
ATTRITION UNDER SHEAR DEFORMATION

The attrition shear cell, introduced by Paramanathan and Bridgwater [1] has
been used extensively to assess particle breakage under well-defined shearing
conditions [2-4]. After shearing the breakage of material is assessed. The most
commonly used method is that of Neil & Bridgwater [3].

   
W  K N     (1)
  ref  
Where σ is the stress applied to the bed, σref a reference stress, Г total strain and
KN, υ and β are fitting parameters. Paracetamol particles with a sieve size of 500
– 600 μm are placed into the cell up to a bed height of 3.5 mm to provide a bed
height of approximately 6 particles; roughly the width of a shear band [5]. The lid
is lowered onto the bed to provide a normal stress and the cell is rotated to achieve
a desired level of strain. After shearing the material is sieved through a (355 μm)
sieve that is two standard sieve sizes below the feed size to establish the percent-
age of broken material [6]. This procedure is carried out for shear stresses, τ, of
0.1, 0.6, 1.0, 1.6 and 2.1 kPa and strains of 30, 60, 120 and 240, shown in figure 1.
10
W = 0.53[(τD/σref).Г1.28]0.46
W
0.1
0.1 1 10 100
(τD/σref).Гυ
Figure 1. Attrition in the shear cell

The lower stresses are achieved by using a pulley-counterbalance system.

Equation 1 is applied to the attrition data, as shown in Figure 1. This is then cou-
pled with the estimation of stresses and strain rates in the dryer to provide a pre-
diction of attrition.
ATTRITION IN THE DRYER

A small-scale dryer is used here. The vessel is cylindrical with a diameter of 94
mm and contains a vertical-axis, double-bladed, retreat-curve impeller with a di-
ameter of 90 mm, a blade height of 18 mm and a blade angle of 60o to the hori-
zontal plane. The attrition of 500 – 600 μm Paracetamol particles in this vessel is
quantified using the same method as in the shear cell for a bed height of approxi-
mately 55 mm, impeller speeds of 20, 45 and 78 rpm, and rotational times of 5, 20
and 80 minutes, as shown in figure 3.
ESTIMATION OF STRESSES AND STRAIN RATES IN THE DRYER

DEM is used to obtain an estimate of the distribution of stresses and strain
rates in the small-scale dryer. The Paracetamol particles are represented by
spheres with a median size of 3 mm. The span of the size distribution is equal to
that of the experimental Paracetamol. The particles have sliding and rolling fric-
tion coefficients of 0.5 and 0.1 respectively. This results in an average torque that
corresponds well with the experimental value. After 2 impeller rotations the
torque is consistent and the bed has reached steady state. Throughout the follow-
ing complete rotation the stresses and strain rates in a sector of the bed (figures
2a-c) are analysed at approximately every 1o of rotation. This analysis sector is
divided both radially and vertically into measurement regions of equal volume
(figures 2i-iii).
The deviatoric stress, τD, in a given measurement region is calculated from the
normal stresses [7] (eq 2). The average velocities in the direction of impeller rota-
tion, vθ, on the vertical and radial boundaries of each measurement region are ob-
tained to give the strain rate [7] (eq 3). The deviatoric stresses and strain rates are
obtained for each measurement region at each measurement interval.

a) 25o b) 10o c) 40o
i) 3 x 3 ii) 4 x 4 iii) 2 x 2
Figure 2. Measurement sector (a-c) and region division (i-iii) arrangements
   yy    xx   zz    yy   zz 
2 2 2
D 
xx
(2)
6
2 2
1  v v   v 
      (3)
2  r r   z 
A sector size of 25o and region division of 3 by 3 measurement cells were cho-
sen primarily as this leads to 4 – 8 particle diameters in each dimension for the 3
mm particles, and so approximates the width of a shear band well.
ATTRITION PREDICTION
The stress and strain rate distributions throughout the dryer bed obtained from
DEM are combined with the analysis of attrition as a function of stress and strain
established in the shear cell to predict attrition,
  
c n    n   n 1   c
W1   K N   c ,n  c ,n     n      n     Ac (4)
  1   1
1 1   ref     1

where χ is the mass fraction in cell c at interval n,

tB
 c ,n    c ,n dt (5)
tA
 c ,n   c ,t   c ,t  / 2
A B
(6)
tA and tB are the time at the start and end of interval n respectively. As the
stresses and strain rates are considered not to deviate significantly for subsequent
impeller rotations, the predicted attrition after N rotations,
 c 
WN    Ac  N R  A  N R (7)
 1 
3.5 Exp - 20 rpm eq.7 - 25° 2x2

3 Exp - 45 rpm eq.7 - 25° 4x4
Exp - 78 rpm eq.7 - 40° 3x3
Attrition, W (%)
2.5 eq.7 - 25° 3x3 eq.7 - 10° 3x3

2
1.5
1
0.5
0
0 2000 4000 6000 8000
Number of Rotations, NR (-)
Figure 3. Experimental and predicted attrition in the small-scale dryer
The predicted attrition using this approach is shown in Figure 3. The initial
sector size and region division (a,i) provide a good estimate of the attrition occur-
ring in the dryer. Increasing the measurement cell division (a,ii) or reducing the
sector size (b,i) results in a smaller cell size, so the stresses and strain rates ob-
tained approach that of a single particle, the predicted attrition is therefore in-
creased. A reduction of the cell division (a,iii) or increase in the sector size (c,i)

leads to a reduced resolution of the bed, and so the influence of the higher stress
regions at the base of the dryer is not captured and the predicted attrition is re-
duced. In all cases the simulations were carried out at 20, 45 and 78 rpm, and the
attrition was predicted to be independent of impeller speed (not shown here)
which corresponds well with the experimental results in Figure 3.
CONCLUSION
A method for predicting attrition in agitated dryers by analysing the influence
of stress and strain on the attrition of Paracetamol in a shear cell, and combining
this with an estimation of the stresses and strain rates in the dryer using DEM, was
developed. The prediction compares well with the extent of attrition measured
experimentally at a range of rotational speeds and rotation times. The size of the
measurement region chosen in the DEM affects the predicted attrition; each region
should be comparable in size to a shear band in the shear cell. The extent of attri-
tion occurring in the dryer is independent of the impeller speed in the range tested.
ACKNOWLEDGMENTS
Financial support from the Engineering and Physical Sciences Research Coun-
cil and GlaxoSmithKline are greatfully acknowledged.
REFERENCES
1. Paramanathan, B.K., Bridgwater, J., 1983, Chemical Engineering Science, 38
(2), 197-206.
2. Ouwerkerk, C.E., 1991, Powder Technology, 65, 125-138.
3. Neil, A.U, Bridgwater, J., 1994, Powder Technology, 80, 207-219.
4. Ghadiri, M., Ning, Z., Kenter, S.J., Puik, E., 2000, Chemical Engineering Sci-
ence, 55, 5445-5456.
5. Neil, A.U., 1986, PhD thesis, University of Birmingham, UK.
6. Kwan, C.C., Chen, Y.Q, Ding, Y.L., Papadopoulous, D.G., Bentham, A.C.,
Ghadiri, M., 2004, European Journal of Pharmaceutical Science, 23, 327-336.
7. Luding, S., 2008, Particuology, 6, 501-505.

SIMULATION OF THE CONFINED COMPRESSION

TEST OF IRON ORE PELLETS USING RANDOM
DISTRIBUTED 3D MULTI PARTICLE FINITE
ELEMENTS
Gustaf Gustafsson, Hans-Åke Häggblad

Luleå University of Technology, Division of Solid Mechanics
In this work, a multi particle finite element model is used to simulate

the axial compression of iron ore pellets inside a steel cylinder. Each
individual pellet is discretized with a 3D finite element mesh. From
experiment, the load, displacement and the circumferential strain are
measured. Experimental results are compared with simulation results.
INTRODUCTION
Numerical simulation of the compaction of granular materials is an area of ac-
tive research. One approach is to use deformable discrete elements of the individ-
ual particles using a 2D finite element (FE) mesh, see e.g. [1] and [2]. In this
work, the axial compression of iron ore pellets inside a steel cylinder is studied
and the individual particles are discretized with a coarse FE mesh in 3D.
SINGLE IRON ORE PELLET

Iron ore pellets are fine particles of ore that has been rolled and sintered to cen-
timetre-sized irregular bodies. They are produced in two varieties: blast furnace
pellets and direct reduction pellets. This study handles the blast furnace pellets.

Experimental two point loading between parallel punches of 18 random cho-

sen pellets were performed in order to characterize the load displacement behav-
iour of the iron ore pellets. The pellets were gathered into four groups of different
sizes; see Table 1. Each pellet was loaded and unloaded in a cyclic increasing
manner until failure, see pictures from a test in Fig. 1.
Group/No of samples Mean mass [g] Ømean [mm] Distribution [%]
1/2 2.22 10.32 11.1

2/3 2.49 10.96 16.7
3/8 3.07 11.59 44.4
4/5 3.58 12.22 27.8
Table 1. Experimentally tested iron ore pellets.
Fig. 1. Iron ore pellet. Picture before and after failure.

FE mesh, configuration 1 and 2.
Finite element models corresponding to the four different experimental groups
were simulated. Each of the four models were meshed with 54 eight-node ele-
ments, see Fig. 1. As the pellet has an irregular shape, the load displacement re-
sponse is depending on the orientation of the pellet in the test rig. Therefore, the
finite element simulations were carried out on two extreme orientations of the
same pellet, configuration 1 and 2, see Fig. 1. An elastic plastic material model
with linear hardening was used. The Young’s modulus, E, the plastic hardening
modulus, H, and the yield stress, σy, were adjusted to fit the experiments. Pois-
son’s ratio, υ, is assumed similar to metals. Material parameters for the finite ele-
ment models are; density, ρ = 3700kg/m3, E = 94GPa, υ=0.29, H=3.5GPa and
σy=1MPa. In [3] a formula for deriving the tensile stress from the applied load in a

two point loading situation is derived, see equation 1, where St is the tensile stress,
F is the applied load, and r is the half distance between the loading points.
F
St = 1.4 (1)
2π r 2
Results from the experiments and simulations are shown in Fig. 2. The tensile
stress versus the displacement divided by the diameter of the specimens is plotted.
∅mean = 10.32 mm ∅mean = 10.96 mm
30 30
Specimen 1 Specimen 1
25 FE, config 1 Specimen 3
FE, config 2 25
FE, config 1
FE, config 2
20 20
St [MPa]
15 St [MPa] 15
10 10
5 5
0 0
0 0.01 0.02 0.03 0 0.01 0.02 0.03
δ/D δ/D
∅mean = 11.59 mm ∅mean = 12.22 mm

30 30
Specimen 3 25 Specimen 3
25
Specimen 5 FE, config 1
Specimen 6 20 FE, config 2
20
FE, config 1
FE, config 2
St [MPa]
St [MPa]
15 15
10 10
5 5
0 0
0 0.01 0.02 0.03 0 0.01 0.02 0.03
δ/D δ/D
Fig. 2. Tensile stress versus the displacement divided by the diameter.

Experiment and simulation. Group 1 to 4.

CONFINED COMPRESSION TEST

Experiment and simulation of iron ore in a confined compression test were per-
formed. The LS-DYNA non-linear finite element software [4] was used for the
simulations in this work.
The experiment consists of an upper and lower punch of thick circular steel
plates surrounded by a steel cylinder as seen in Fig. 3. Iron ore pellets are placed
inside the cylinder to a height of 305mm. Total mass of iron ore pellets is 46.0kg.
An axial load is applied to the lower punch. Measured data are the force and dis-
placement of the punch. Strain gauges are measuring the circumferential strain in
the middle of the steel membrane.
Fig. 3. Experimental setup.

The multi particle finite element model of the confined compression test is a
quarter-model of the real experimental setup. In this simulation, 4290 discretized
rigid particles of four sizes with the distribution given in Table 1 are representing
the iron ore pellets in the model. The filling height is 305mm and corresponding
mass 38.5kg. The rough surface of the iron ore pellet motivates a friction coeffi-
cient between the particles of 0.6. Friction coefficient between particles and wall
is set to 0.1. The punches are considered rigid in the simulation and the membrane
is modelled with thin elastic shell elements. The compression is simulated in two
steps. The iron ore pellets models are randomly sparse placed the cylinder. In the
first step, a gravity driven simulation is carried out where the pellets are arranged

in the cylinder. In the second step, the compression is simulated by a prescribed

displacement of the upper punch. The steps are seen in Fig. 4.
Fig. 4. Multi particle finite element simulation. From left: initial arrangement,
arrangement after gravity simulation and at maximum load.
Results from experiment and simulation are seen in Fig. 5, where the axial
forces versus the displacements are plotted.
140 140
120 120
100 100
80 80
Fax [kN]
Fax [kN]
60 60
40 40
20 20
0 0
0 5 10 15 20 25 0 5 10 15 20 25
δ [mm] δ [mm]
Fig. 5. Left: Experimental data, force versus displacement. Right: Simulation

data, force versus displacement.

DISCUSSION AND CONCLUSION

In this work, experiments and simulations of the mechanical behaviour on both
single and multiple iron ore pellets are performed. The single pellet simulations
show good agreement with experimental results. A multi particle finite element
model is used to simulate the axial compression of iron ore pellets inside a steel
cylinder. A comparison between experiment and simulation shows that the simu-
lated response is softer than the experimental. An explanation is the difference in
packing density (mass in experiment 46kg and simulation 38,5kg). A more de-
tailed measurement of the size distribution and larger number of sizes in the simu-
lation will increase the packing density resulting in a stiffer response. Another un-
certainty is the frictional behaviour between the particles.
ACKNOWLEDGMENTS
The financial support from Hjalmar Lundbohm Research Centre (HLRC),
LKAB and Luleå University of Technology are gratefully acknowledged.
REFERENCES
1. A.T. Procopio, A. Zavaliangos, ‘Simulation of multi-axial compaction of
granular media from loose to high relative densities’, Journal of the Mechan-
ics and Physics of Solids, 53, 1523-1551, 2005.
2. D.T. Gethin, R.S. Ransing, R.W. Lewis, M. Dutko, A.J.L. Crook, ‘Numerical
comparison of a deformable discrete element model and an equivalent contin-
uum analysis for the compaction of ductile porous material’, Computers and
Structures, 79, 1287-1294, 2001.
3. Livermore Software Technology Corporation, LS-DYNA keyword user´s man-
ual, version 971, Livermore Software Technology Corporation, 2007.
4. Y. Hiramatsu, Y. Oka, ‘Determination of the tensile strength of rock by a
compression test of an irregular test piece’, Int. J. Rock. Mech. Sci., 3, 89-99,
1966.

ON APPLICATION OF SYMBIOTIC CELLULAR

AUTOMATON METHOD FOR SIMULATION OF
THE MECHANICAL RESPONSE OF LIGNITE
A.V. Dimaki, E.V. Shilko, A.I. Dmitriev, *S. Zavsek, *J. Pezdic, S.G. Psakhie
ISPMS SB RAS, Akademicheskii av. 2/4, Tomsk, Russia, 634021
*Velenje Coal Mine, Partizanska 78, Velenje, Slovenia, 3320
The symbiotic cellular automaton method, which combines two dis-

crete approaches to describe the simulated media – the method of
classical cellular automaton and method of movable cellular automa-
ton, has been developed for numerical study of heterogeneous materi-
als. The paper describes the developed method and the results of its
application to study of the mechanical response of lignite specimens
under uniaxial loading in carbon dioxide atmosphere. It is shown that
the presence of adsorbed gas and the external gas pressure can pro-
duce a significant impact on the mechanical response of lignite.
INTRODUCTION
The presence of adsorbed gas in the coal seam produces a significant influ-
ence on the mechanical behavior of rock [1]. At the present time, the adsorbed gas
is often considered as the reason of instantaneous outbursts of large quantities of
crushed coal, which constitutes a danger to human life and complicates the tech-
nological process.
From the beginning of the production of Velenje lignite the protection from
instantaneous outbursts of coal is actual problem. One of the ways of solution of

this problem is the computer-aided simulation of the behavior of rock within the
natural conditions.
In the paper the symbiotic cellular automaton (SCA) method is described,
which represents the development of movable cellular automaton (MCA) method
by means of introducing an additional layer of classical cellular automata into the
computational region [2]. Within this layer the simulation of the behavior gas
phase is carried out. We use simple phenomenological relations for description as
the effect of the gas phase on the mechanical response of lignite as sorption proc-
esses. The simulation of the high-pressure sorption of gas in the lignite shows the
good compliance between simulation results and experiments. The possibilities of
SCA to study the behavior of the coal seam are demonstrated within the modeling
of uniaxial loading of lignite in the CO2 atmosphere.
DESCRIPTION OF SYMBIOTIC CELLULAR AUTOMATA METHOD

In the framework of SCA method the investigated medium is considered as
a superposition of two interrelated media (Fig. 1a). One of them is described by a
set of movable cellular automata [2,3] and another – by a mesh of classical cellu-
lar automata [4]. The step of calculation consists of two main substeps, first of
which is the step of the MCA model, called as ‘mechanical step’, and the second,
called as ‘net step’, is the step of CA model.
Description of the ‘mechanical step’. At this substep the mass transfer in
solid phase under the mechanical loading is considered by means of solving of the
motion equations of movable automata. MCA approach represents the simulated
medium as an ensemble of interacting finite size elements (automata) [2,3]. The
movable cellular automaton is an object of finite size, possessing translational and
rotational degrees of freedom. Within the MCA approach a many-body interaction
is used. Also, we introduce the new type of state – the state of a pair of automata,
that leads to a new parameter defining the criteria of switching of the inter-
automata interaction – the automata overlapping parameter: hij = rij − roij
(Fig. 1b). In the simplest case there are two states of the pair: linked (hij < hijmax)
and unlinked (hij > hijmax). The linked state is indicative of chemical bonds be-
tween elements and the unlinked state indicates that there is no chemical bond be-
tween them.

a) b)
Fig. 1. a) Layers of movable and classical cellular automata; b) Definition of
automata overlapping parameter.
We use the piecewise-linear approximation of real stress-strain dependence
in order to reproduce the mechanical response function of automata. In order to
take into the account the damage generation at a scale level lower than automaton
size, the degradation response function can be used [2,3].
Description of the ‘net step’. Next for the mechanical – ‘net’ substep is
performed on the CA mesh. At this substep the process of mass transfer of gas in
the pores and channels is considered, as well as the values of the forces acting on
the movable cellular automata from gas phase are calculated [5]. The configura-
tion of pores and channels through which gas propagates, is projected to the CA
mesh from the MCA layer. This symbiotic approach combines solutions for me-
chanical and gas dynamic problems and allows description of multiphase hetero-
geneous media. In the present model, the Mendeleev-Klapeyron equation is
applied to describe the state of gaseous phase. We use two types of sorption
processes on the boundaries between solid material and pores: adsorption and
desorption of gas, which are considered as independent and are described by the
following linear equations:
I ads = α ads (µ max min

solid − µ solid )µ env , I des = K des (µ solid − µ solid ) , (1)
where αads – adsorption intensity coefficient; Kdes – is the desorption coefficient,
µenv – is the concentration of the gas in the atmosphere; µsolid – is the concentration

of the gas in solid; µ min max

solid and µ solid – are the minimal and maximal possible
concentrations of gas in solid, consequently.
The presence of gas in the pores as well as in the bulk of material leads to
appearance of additional forces acting on the solid surface and causing the growth
of internal stresses in solid. The total force, acting from the surrounding gas, ap-
plied to the centre of mass of movable automaton, is written as follows [5]:
r gas N ksurf N agas r
Fk = S ∑ ∑ Pb nba , (2)
a =1 b =1
where the index a enumerates the ‘surface’ classical cellular automata, which ‘are
belonging’ to k-th movable automaton, the index b enumerates the CCA, which
are bordering with ‘surface’ a-th CCA, Pb – gas pressure in the volume of the
r
CCA b; nba – a vector of normal to the contact square between the b and a CCA;
S – ‘contact square’ of CCA.
SIMULATION OF FRACTURE OF DETRITUS IN THE ATMOSPHERE

OF CARBON DIOXIDE
To demonstrate the possibilities of SCA method and its applicability for de-
scribing the ‘porous solid – gas’ medium, a series of calculations were made. In
these calculations the effect of the CO2 on the mechanical response and the failure
of the samples of fine detritus was studied. Mechanical properties of fine detritus
and xylite inclusions were used as given in [5]. The response and failure of fine
detritus under uniaxial compression in the CO2 atmosphere and in vacuum were
simulated. Size of specimens was 5x10 cm, loading rate was V = 0.02 m/sec (Fig.
2a).
In the absence of CO2 atmosphere first crack appears in specimen when the
strain is 1.23% (Fig. 2b). In the presence of the gas the crack is formed within the
deformation 1.26%. Complete destruction of the sample in vacuum occurs by
means of the formation of a crosscutting diagonal macrocrack. This corresponds
to a deformation ~1.35%. The destruction of the sample under 1 bar pressure oc-
curs by the formation of vertical cracks within the deformation 1.45%.

Thus, little increasing of gas pressure from 0 to 1 bar resulted in an increase of

strength and ultimate strain of specimens at 2–5% without a noticeable change in
Young's modulus. Also, the fracture pattern of specimens has undergone some
changes. The results suggest that even relatively low atmospheric pressure is able
to exert some influence on the response and the fracture of materials having rela-
tively low strength.
a) b)
Fig. 2. a) Scheme of mechanical loading of simulated specimen; b) Response of
lignite specimen at the beginning of failure (inset shows total loading diagram).
CONCLUSION
The approach, describing the behavior of heterogeneous medium under exter-
nal influence, is proposed. This approach, called as SCA method, represents the
combination of methods of classical and movable cellular automaton. Verification
of this approach was carried out by means of comparison of the results of model-
ing of the samples of fine detritus in the CO2 atmosphere with the results of ex-
perimental studies performed by researchers of Velenje coal mine (Faculty of
natural sciences and engineering, University of Ljubljana, Slovenia) [6]. These
results demonstrate a qualitative agreement between model and experimental data.
It is clear that the proposed approach does not cover the entire spectrum of
processes in such a complex system like a multiphase heterogeneous medium. In
particular, the developed model does not take into account the filtration processes,

the phase transitions etc. In addition, since the simulated physical processes have
different (several orders of magnitude) characteristic times, within the framework
of the proposed approach it is possible to simulate only the first few seconds. Fur-
thermore, we use a 2D approximation of the simulated object. Of course, these
restrictions greatly influence the outcome of the numerical study. Nevertheless,
proper selection of model parameters allows obtaining good agreement with the
results of natural experiments. This demonstrates the correct formulation of the
problem and qualitatively correct description of the basic processes occurring in
the system.
Further directions of development of the method are the introduction to the
model the state equation for non-ideal gas, which is important for realistic CO2
simulation, as well as the possibility of modeling filtration processes in porous
solid body.
REFERENCES
1. B. Beamish, P. Crosdale, ‘Instantaneous outbursts in underground coal mines:
An overview and association with coal type’, International Journal of Coal
Geology, No.35, 27–55, 1998.
2. S.G. Psakhie etc., ‘Movable cellular automata method for simulating materials
with mesostructure’, Theor. and Appl. Fracture Mech., No.37, 311-334, 2001.
3. E.V. Shilko etc., ‘Development of Movable Cellular Automaton Method for
Simulation of Deformation and Fracture of Heterogeneous Elastoplastic Mate-
rials and Media’, Proc. Int. Conf. ‘Particles 2009’, 349-352, 2009.
4. S. Wolfram, Theory and applications of cellular automata, New Jersey: World
Scientific, 1986.
5. S. Zavsek etc, ‘Computer-aided investigation of response and fracture of lig-
nite using MCA method and symbiotic cellular automata approach’, Proc. Int.
Conf. on New Challenges in Mesomechanics, 19-25, 2002.
6. J. Pezdic etc., ‘Laboratory simulation of adsorption-desorption processes on
different lignite lithotypes from the Velenje lignite mine’, RMZ – Materials
and Geoenvironment, 46, No.3, 555-568, 1999.

MODELLING FLEXIBLE MATERIAL USING EDEM:

CALIBRATION AND INDUSTRIAL APPLICATIONS
Stephen Cole
DEM Solutions Ltd
The Discrete Element Method (DEM) is now being used for simula-
tion and analysis of a wide range of particulate solids handling and
processing operations in industry.
The commercial DEM software package, EDEM™, has a bonded par-

ticle model capability [1] that allows particles to be bonded together to
form solid shapes. These solids are created by connecting discrete
element spheres with flexible bonds. The bond strength and stiffness
can be varied will allows the user to model a variety of different parti-
cle processes where particles can deform and break up due to interac-
tion with surfaces or other particles. Bonds will resist normal and
tangential forces and moments up to a predefined breaking stress.
When load on a bond has exceeded predefined breaking stress the
bond will break.
Model predictions were validated against previously published data,

with a view to define a range of material behavior that encompasses
the observed material behavior during the experiments.

INTRODUCTION
The bulk behaviour of particulate materials is governed by the properties of the
particles and their local interactions. DEM simulates bulk particle behaviour by
modelling granular materials at the particle-scale, which enables linkage between
particle properties and process conditions. A DEM model can directly model par-
ticle size distribution and shape, mechanical properties and surface-to-surface
interactions. DEM accounts for momentum, mass and energy exchange between
particles, and between particles and surrounding fluid, electromagnetic field and
contacting equipment surfaces.
Using the EDEM capability to generate flexible bonds between spheres, flex-
ible material from small fibres and organic materials to steel beams and concrete
structures can be simulated.
Figure 1. Left, analysis of sieve efficiency using flexible stalks. Right, investi-
gation into structural collapse
Figure 1 shows examples of flexible material flowing through a sieve during

the harvesting process, the DEM simulation is used to analyst the efficiently of the
sieve. The same technique at a larger scale using different input parameters can
also be used to simulate failure and collapse of large structures.

PARTICLE DEFINITION
The commercial DEM software package, EDEM™ has the capability to gener-
ate multi-sphere particles; each sphere in a multi-sphere particle is rigidly joined
to the neighbouring sphere(s), as shown in Figure 2.
Figure 2. Three-sphere multi-sphere particle
The bonded particle model in EDEM™ can be used to generate flexible parti-
cles by bonding single sphere particles together with a flexible bond. This bonded
structure is defined as a meta-particle (Figure 3).
Figure 3. Two alternative meta-particles with flexible joints
In contrast to a multi-sphere particle, which is a rigid body, a meta-particle can

flex and break. The flexibility and strength of a meta-particle depends on the
number of spheres used and the strength of the bonds between the spheres.
CASE STUDY: SAFFLOWER STALK

Safflower seeds contain 30%-45% of oil; this oil can be used in biodiesel pro-
duction. Information of the mechanical properties of the stalk can improve har-
vesting conditions and improve yield [2].

The bending properties of the stalk detailed by O. Ozbek et al [2] were repro-
duced using the DEM Software EDEM™. A three point bending test (Figure 4)
was used to determine the bending stress and modulus of elasticity.
Figure 4. Experimental setup as detailed in O. Ozbek et al

The same bending model was setup in EDEM™ to gather force-deflection data
from the virtual stalk. Experimental results are shown in Figure 5.
Figure 5. Experimental results from O. Ozbek et al
The moisture content and stalk diameter are seen to change the mechanical be-
haviour of the material. The simulation parameters were setup to re-produce this
range of behaviour with bonded particle model. Figure 6 shows a slice through

the centre of a virtual stalk under compression. Red shows areas of compression
while blue colouring indicates areas of expansion.
Figure 6. A 6000 sphere stalk (half slice shown) during bending
Each sphere is joined to its neighbours with a flexible bond; the bulk behaviour
of the stalk is dependent on the individual properties of the bonds. The bond be-
haviour is controlled by 3 main parameters, the bond normal stiffness, bond shear
stiffness and bond disk radius, which is defined as the radius of cylindrical bond
between two spheres. A 27 simulation DOE was run with low medium and high
values for each setup.
The DOE produced a minimum Modulus of Elasticity of 10.5 MPa and a
maximum of 552 MPa, encompassing the range of 100 MPa to 450 MPa seen ex-
perimentally. The DOE also indicated that the primary value that affected the elas-
ticity was the bond disk radius, the second parameter was the normal stiffness and
the third most important was the combination of the two values, it was therefore
chosen to keep the shear stiffness constant (1e10 Pa) and vary the normal stiffness
and bond disk radius.
Figure 7 shows the response of the bending behaviour of a 7 sphere stalk with
the two parameters varied. The required modulus of elasticity detailed in the sur-
face plot can be mapped to the relevant moisture content of the stalk.

Figure 7. Surface plot of Modulus of Elasticity v Bond Disk Radius and Bond
Normal stiffness
CONCLUSION
The required mechanical response of the safflower stalk can be directly attrib-
uted to the input to the EDEM™ model, allowing the DEM software to predict the
behavior of the material under different operating conditions.
REFERENCES
1. Potyondy D. O. and Cundall P. A, A bonded-particle model for rock, Interna-
tional Journal of Rock Mechanics and Mining Sciences, 41, 1329-1364 (2004)
2. O. Özbek, A. Y. Seflek, K. Carman, ‘Some Mechanical Properties of Saf-
flower Stalk’, Applied Engineering in Agriculture. 25(4): 619-625, 2009

DEM ANALYSIS ON PARTICLE BEHAVIOR IN THE

COURSE OF SINTER MIXTURE CHARGING
Tsukasa Abe, Junya Kano, Masanori Nakano*)

Tohoku University, Graduate School of Environmental Studies
*)
Nippon Steel Corporation, Ironmaking R&D Lab
A DEM model for the charging apparatus of iron ore sintering has
been developed and applied to the estimation of vertical size segrega-
tion in the bed of raw mixture. The model generated several results
close to those obtained by the plant observation, which has proven the
potential of the method for wider use regarding the segregation phe-
nomena during charging.
INTRODUCTION
Steelmaking industry releases 11% of the total CO2 emission in Japan. Most of
the emission occurs in the course of ironmaking, the process of which consists of
blast furnaces, coke ovens and sintering plants. The sintering plants make sinter
from iron ores as the ferrous burden for blast furnace, burning cokes as fuel,
which are preliminarily mixed by 4% with iron ores. The CO2 emission from the
sintering machines amounts to 1% of the total.
Sintering proceeds with heat wave descending across sintering bed due to the
heat generation by coke combustion and heat transfer by down draft; the heat
wave causes iron ores particles to partially fuse over 1200°C and the consequent
melt bonds the particles to mass. As an intrinsic manner, the heat generated in up-

per layer is utilized in subsequent lower layer; therefore, coke content in the lower
layer is potentially slashed comparing that in the upper layer.
Charging process of the mixture into pallets of sintering machine helps to de-
velop vertical segregation of coke concentration in the sintering bed from the uni-
form mixture of cokes and iron ores. Some amount of size segregation occurs
naturally after the raw mixture flows over the sloping chute and settles on the
slope formed by preceding ones, which results in the segregation of coke
concentration, as the size of coke is smaller than that of ores.
The technical issue was to intensify the segregation by modifying the charging
appratus. Already many had been done with experimental apploaches1,2); however,
little3,4) seemed to challenge them theoretically. Therefore, we aimed to apply
DEM to the issue.
DEM MODELLING
The actual granules of sintering raw mixture are complicated in structure [Fig
1], consisting of a centered particle (>2mm) and adhering fine particles (<0.5mm)
on it by water with lone particles of intermediate size (0.5~2mm) ; consequently,
they are not rigid, but soft and sticky.
For the DEM simulation, we simplified such granules to be single rigid spheres
of three different sizes, neglecting their plasticity and reflecting their stickiness to
their friction coefficents. The model particle sizes were set larger than actual in
the interst of computing time.
The structure of actual granules also made it difficult to measure the material
parameters of particles used in the DEM simulation. Consequently, in this paper,
we apply rough values of iron ore to the particle parameters. The detailed
simulation condition of DEM is shown in Table 1.
The DEM model of charging apparatus is shown in Fig 2. The dimension of
charging apparatus followed that of Kimitsu No.1 sintering machine, Nippon Steel,
except the length of pallet width. The actual was 1.3m ; the model was reduced to
30cm for saving the number of particles to execute.

Fig 1. Schematic diagram of single granule.
Number of particle 12524, 52774, 69419 [-]

Diameter of particle 30, 18, 12 [mm]
Time step 1.0x10-4 [s]
Density of particle 3.2 [g/cm3]
Friction coefficient
Particle-particle 0.7 [-]
Particle-wall 0.7 [-]
Young’s modulus 0.0018 [GPa]
Poisson’s ratio 0.3 [-]
Rolling friction coefficient 1.885 [-]
Table 1. DEM simulation condition of DEM.

4000
Hopper
4000
Drum Feeder
Chute Drum feeder

3.79 [rpm]
f 1300 speed
Chute angle 45 [deg]
1400
300 (depth) Pallet speed 3.18 [cm/s]

Pallet Unit: mm
Direction of pallet movement
Fig 2. Schematic diagram of three-dimensional model of sinter charging.
SIMULATION RESULTS
The simulation runs for 60 sec. At 0 sec, all of the particles appeared in the
hopper, and then went over the drum feeder and the chute into the pallet. At 60
sec, the execution terminated to evaluate the vertical size segregation of the par-
ticles in the range shown in Fig 3.
Fig 4 shows a simulation result at 60sec. Fig 4a is for the particle movements
where the colors means their sizes, showing that the coarser particles (red) ran fur-
ther down than others over the slope of a sinter bed and lay at lower position in a
sinter bed; on the other hand, finer particles (blue) lay at upper position. Fig 4b
has numerically supported the observation, showing the distribution of the particle
sizes across the sinter bed. The coarser particles are fewer, and finer are lot at
higher position in the sinter bed.
In the actual operation of sintering machines, such trends in size segregation
are distinctively observed.

The agreement between the actual and the model results suggests a possibility
that the DEM simulation replicates the charging process of the mixture into pallets
in terms of the size segregation of the particles.
Sampling position
Direction of transfer
h
Fig 3. Schematic diagram of sampling position in a sinter bed.

Sampling position h /mm
500-600
400-500
300-400
200-300
100-200
0-100
0% 20% 40% 60% 80% 100%
30 mm 18 mm 12 mm
(a) (b)
Fig 4. Simulation results at 60sec: (a) a snapshot of the flow of the particles (red is
30mm, green is 18mm, and blue is 12mm), and (b) the distribution of the particle
sizes across the height of sinter bed.

CONCLUSION
We applied DEM to the charging process of the mixture into pallets, and
simulated well the particle behavior in terms of the vertical size segregation of the
particles. Therefore, it is possible that the DEM simulation replicates the particle
segregation phenomena in charging process qualitatively.
Further we will re-adjust the material parameters of particles to promote accu-
racy of simulation and examine the effect of the chute angle and the feeding rate
on the size segregation to seek the best machine condition under which the par-
ticle size segregation becomes most intensified.
REFERENCES
1. K. Fukami, H. Itaya, F. Aratani, K. Nakanishi and H. Nishimura, ‘A Study on
the Particle Size Segregation in the Sintering bed Formed at the Feeding Stage
of a Sinter Machine’, Tetsu-to-Hagane, 71, No.16, 1888-1894, 1985.
2. T. Inazumi, M. Fujimoto, S. Kasama and K. Sato, ‘Characteristic of Raw Mix
Packed Structure on Sinter Strand with Chute Type Feeder and Its Effect on
the Sintering Reaction’, Tetsu-to-Hagane, 77, No.1, 63-70, 1991.
3. M. Nakano, Y. Hori, S. Nakano, E. Ide, S. Yoshinaga and Y. Hosotani, ‘A
Model Design for the Charging Apparatus of a Sintering Machine’, Tetsu-to-
Hagane, 82, No.11, 887-892, 1996.
4. P.A. Moysey, M.H.I. Baird, ‘Size segregation of spherical nickel pellets in the
surface flow of a packed bed: Experiments and Discrete Element Method si-
mulations’, Powder Technology, 196, No.3, 298-308, 2009.

MODELING EMERGENCY EGRESS FROM A

PUBLIC FACILITY
Gary A. Geissinger, Member, IEEE

Graduate Student, Colorado School of Mines
This study uses discrete element techniques to investigate how the ur-
gency and politeness of patrons in a public facility can influence the
time required to evacuate the facility. Patron’s different weights, siz-
es, and maximum motion rates were modeled.
RELATED WORK
Pedestrian motion and evacuation has been studied using a variety of tech-
niques [1]. In the literature the approaches are generally considered to be either
macroscopic or microscopic [2]. Macroscopic models deal with the average be-
havior of the pedestrians and are often based on partial differential equation based
models. Microscopic models view pedestrians as independent agents. Discrete
element methods (DEM) have been applied to model pedestrian flow in the past.
Each pedestrian can be considered an independent agent [3] or the social forces
between patrons can strongly influence behavior and motion [4] in groups.
PATRON MODELING
In this 2-D study each patron in an entertainment venue is considered an inde-
pendent agent and is a member of a model patron set with member adults, senior
citizens, and children each with different velocity, mass, and acceleration values.
Without others in close proximity, maximum walking velocity of an adult patron

is limited to 55 inches per second and is based on the assumed minimum rate at
which a pedestrian can walk across an intersection [5]. The patron models include
the work of Imms and Edholm who showed that the elderly tend to move more
slowly with age [6]. The average walking rate for children 11 to 15 years of age
is 30 inches per second is also included in the model [7].
Dimensionally patrons are modeled as ellipses using width and depth values
from each patron model. It is assumed that a patron moves such that the depth is
aligned with the motion vector and the width is normal to it. Jacobs concluded
that people are more sensitive to maintaining personal space in front of them-
selves and with 6 to 8 square feet desired per person [8]. This spacing is assumed
at minimal urgency and is decreased as urgency is increased. Patron seating is
random within the venue and based on the population density for each model set.
As this investigation deals with patrons leaving an area during an emergency, it
was assumed that the force behind their motion should be proportional to their
weight (i.e., the patrons “throw their weight around” to get out of the theatre).
The more urgency a patron feels the more force is applied on a scale from 0 to 10
with corresponding forces from 0 to full body weight.
For the purposes of this study patrons are in an open seating area and proceed
to the closest exit; however, if a patron is located in a seating row, the direct vec-
tor is likely not the preferred motion orientation. Assuming that the patrons do
not hop over the seats and change rows, motion will be down a seating row until
an aisle is reached and then down the aisle to the exit.
CATCHMENT MODELS
The catchment models used for this study were taken from the 2000 Interna-
tional Building Code (IBC) 1008 and used as specified or as a point of departure
for parametric evaluation [9]. The IBC models provide a consistent framework
for study and are well documented. The current analysis allows for fixed seating,
regular and emergency exits, and a level floor and aisles.
The seating configuration of an airliner was also studied [10]. On commercial
aircraft, due to the curved cross section of the fuselage, the aisles are not located
down the sides of the seating areas. As a result a patron can become trapped if

motion is toward the closest exit without regard for seats and other obstacles in
the path.
SIMULATION DESIGN
The simulation is based on DEM techniques and was first modeled using Mat-
lab® and FORTAN. The final version was written in CodeGear™ C++ Builder
2007® in the Vista 64™ operating environment.
Each class of patrons is a different color: white for adults, magenta for child-
ren, and green for seniors. The exits are marked in red. A patron is removed from
further calculation if determined to sufficiently close to an exit.
TEST PARAMETERS
Spring constant patron to wall/row – A value of 10,000 for the spring constant
gives a deformation of approximately 10% which seems reasonable for patrons.
Spring constant patron to patron – Once again a value of 10,000 provided
about a 10% deformation in each patron.
Coefficient of restitution - When human beings collide with either objects or
other human beings the resultant velocity is very low; as a result the coefficient of
restitution was set to 0.1.
Velocity dependent drag - Once a patron is not motivated to move in a given
direction, the velocity quickly decreases to 0. This is modeled using a velocity
dependent drag coefficient that is set to 1.0.
Update interval - As always there is a trade-off between fidelity and reasonable
simulation execute rate. The recoil velocity is approximately 99% of the original
value with an update increment of 500 microseconds. As a result a one millise-
cond update interval was used.
SIMULATION RESULTS
A catchment specified as IBC 1008.5.1.2 was first used to study politeness
changes the time required to empty a catchment. Figure 1 shows that with just a
few patrons, any politeness has no effect on the time required to clear the area. As

the number of patrons increases, the impact of politeness becomes evident. In a

fully occupied seating area it takes excessive time if all patrons are fully polite.
The message there (not very politically correct) is that if one attends a packed
seating area and there is an emergency, then being polite could impact your lon-
gevity.
Evacuation Time as a Function of Politeness
252 Patrons
Evacuation Time (s)
100 75 Patrons
25 Patrons
10 Patrons
5 Patrons
10
0.0 2.0 4.0 6.0 8.0 10.0
Politeness
Fig 1. Evacuation time as a function of politeness

Another result shown in Table 1 flowed from comparing four emergency exit
configurations of an IBC 1008 catchment.
Exit Configuration Evacuation Time

Single Exit in Rear 64.4 s
Double Exit in Rear 34.7 s
Single Exit in Rear, Dual Front Exits 33.2 s
Double Exit in Rear, Dual Front Exits 20.3 s
Table 1. Evacuation time as a function of exit configuration.

In the case of deplaning after a flight, the usual scenario looks like Figure 2
with the passengers being polite waiting for the passengers ahead of them to dep-
lane. In this case the politeness is set at 10. Figure 3 shows what happens when
a flight arrives late; the passengers are not polite (set at 0) and jam the aisles.
Fig 2. Polite deplaning Fig 3. Late deplaning
CONCLUSION
Politeness impacts the rate at which patrons egress a facility. While children
and senior citizens have different maximum walking rates, the simulations indi-
cate that the actual rate at which patrons leave the facility is below that rate for all
but the most sparsely populated catchments. Additional emergency exits as well
as double wide exits can reduce the time to empty a facility by as much as a third.
While developing the models and simulation, a number of shortcomings and fu-
ture enhancements were identified. Patrons will likely move to an exit that will
allow the fastest egress, not the closest exit. Some patrons will hop over rows of
seats if the urgency is high enough. Some venues do not have clear sightlines to

all exits and also can trap patrons who attempt the shortest path without regard for
dead-ends. And finally there are cases where the obvious path may not be chosen
as a less direct path may be an advantage.
ACKNOWLEDGMENTS
I’d like to thank Dr. Graham Mustoe for his encouragement and guidance dur-
ing this investigation.
REFERENCES
1. E. Kuligowski and R. Peacock. A Review of Building Evacuation Models.
Technical Note 1471, National Institute of Standards and Technology, 2005.
2. P. Kachroo, S. Al-nasur, S. Wadoo, and A. Shende, Pedestrian Dynamics.
Springer, 2008.
3. M. J. Markowski. Modeling Behavior in Vehicular Pedestrian Traffic Flows.
Dissertation, University of Delaware, 2008.
4. T. Lakoba, D. J. Kaup, and N. Finkelstein. Modifications of the Helbing-
Molnar-Farkas-Vicsek social force model for Pedestrian Evolution. In Simu-
lation, Vol.81, 5:339-352, 2005.
5. Imms and Edholm. Studies of gait and mobility in the elderly. In Age Ageing,
10:147-156, 1981.
6. Prince F. et al, 1997. Gait in the elderly. In: Gait and Posture, Vol. 5, pp. 128-
135.
7. L. J. H. Schulze. Evaluation of cross-walk timing and the application of a
standard crossing light timing formula. Texas A&M University.
8. R. Sommer, Personal Space. Bosko, 2007.
9. International Building Council. International Building Code 2000. IBC,
2000.
10. From web URL:
http://www.united.com/page/middlepage/0,6823,1076,00.html

IMPLEMENTATION OF A ROCK JOINT SHEAR

STRENGTH CRITERION INSIDE A COMBINED
FINITE-DISCRETE ELEMENT METHOD
(FEM/DEM) CODE
Omid K. Mahabadi, Giovanni Grasselli

University of Toronto, Department of Civil Engineering
This manuscript discusses the implementation of the Giovanni Gras-

selli shear strength criterion for rock joints inside a combined finite-
discrete element method code. The relevance of the criterion, its for-
mulation, and application by a simple case study are presented.
INTRODUCTION
In the past decades, researchers have attempted at estimating the shear strength
of a rock joint or discontinuity. These efforts have led to the derivation of empiri-
cal rock joint shear strength criteria based on the joint wall strength parameters
and the geometry of the joint, i.e. its roughness. These criteria were however lim-
ited in correctly quantifying the peak shear strength mainly due to the issues re-
lated to obtaining the required material properties and roughness. State-of-the-art
digital and computational techniques have helped the researchers overcome many
of these limitations, namely those related to assessing the joint roughness. Nowa-
days, high-accuracy laser and optical devices can produce precise digital represen-
tation of joint walls, ultimately facilitating the acquisition of the required rough-
ness parameters.

By the application of high-resolution, three-dimensional, digitized rock joint

surfaces, (Grasselli and Egger 2003) proposed an empirical shear strength crite-
rion, denoted as the Giovanni Grasselli (GG) shear strength criterion in this text.
Based on the surface measurements, a number of statistical roughness parameters
characterizing the joint wall can be derived. The GG-shear strength criterion was
developed from the empirical correlation of these roughness parameters with labo-
ratory testing data.
This paper discusses the implementation of the updated GG-shear strength cri-
terion (Cottrell 2009) inside a combined finite-discrete element method code
(FEM/DEM), which is an expanded version of Y-code(Munjiza 2004).
FORMULATION
Peak shear strength. Cottrell (2009) recently modified the GG-shear strength
criterion originally developed by Grasselli and Egger (2003). According to this
new formulation, the shear strength of a rock joint or discontinuity (p) is given as
    
   
*
B
  
  max
*
max n

C 1  9 A  
 p  1  e  0 t   n tan  b  
 C  1   (1)
  
   
 
where max/[C+1] is the 3D roughness (average asperity angle) in the direction
of shearing; max is the maximum apparent dip angle of surface; C is a dimen-
sionless fitting parameter; σn and σt are the applied normal stress and tensile
strength of the intact material, respectively; b is the basic friction angle; and B is
a scale parameter which accounts for the measurement resolutions (i.e., 1.15 for
the 0.044 mm points spacing) (Tatone, et al. 2010). Further discussion on the
roughness parameters can be found in Tatone and Grasselli (2009).
Post-peak behaviour. Once the peak strength of the rock joint is reached dur-
ing progressive shearing, its strength decays hyperbolically towards a residual
shear strength (r), given as:
 r   n tan  r  (2)

where r is the residual friction angle, which is related to the basic friction an-
gle (b) as (Grasselli 2001)
 r   b   (3)
where  characterizes the contribution of roughness to the residual shear
strength and falls in the 15 – 24° range (Grasselli 2001).
The post-peak decay function for shear stress () is expressed as (when u  u p )
up
   r   p   r  (4)
u
where u is the horizontal shear displacement and up the horizontal peak shear
displacement. In his experiments, Grasselli (2001) found up = 0.27 ~ 0.65 mm.
Implementation. The GG-shear strength criterion and the post-peak regime
are implemented in the contact interaction algorithm of the FEM/DEM code. De-
tails on this interaction algorithm can be found in Munjiza (2004). Eqs. (1)-(4) are
used to modify the shear stress () calculated within the interaction code. Once the
edge breaks, i.e. an actual discontinuity is formed, the GG strength criterion and
the progressive frictional coefficient degradation is used.
CASE STUDY: RESULTS AND DISCUSSION

To demonstrate the application of the GG-shear strength criterion and to verify
it, a simple example is presented here. The model, as illustrated in Fig. 1a, is a
rectangle, 50 mm in width and 150 mm in height. The model was prepared using
Y-GUI (Mahabadi, et al. 2010). Note that the bottom of the model is fixed. The
model has a weak interface aligned at 45°. The parameters for the GG-shear
strength criterion are summarized in Table 1. As soon as the simulation begins,
the model fails along this weak interface solely due to the application of gravity
(Fig. 1b).

Fig 1. a) Schematic of the model showing the location of the weak interface; b)
model failed along the weak interface; c) normalized shear stress (  /  n ) versus
shear displacement (u)
Fig. 1c portrays the normalized shear stress (  /  n ) versus shear displacement
(u) for the fracture formed due to the existence of the weak interface. This figure
reveals that initially the shear stress of the joint is equal to its peak shear strength
(p). As more shear displacement occurs, the shear stress that the joint can bear
decreases hyperbolically until the residual shear strength (r) is reached.
A0 0.47 mm2
max 41.7°
C 3.21
b 30.0°
σt 1.45 MPa
Table 1. GG-shear strength criterion parameters for the case study

CONCLUDING REMARKS
The GG-shear strength criterion, together with its peak and post-peak regimes,
as well as its implementation inside our FEM/DEM code were discussed in this
paper. The application and verification of the implementation were demonstrated
in a simple case study. The example illustrated that as soon as a discontinuity sur-
face, i.e. a joint, is formed, the right behaviour is observed according to the shear
strength criterion. This new feature together with the inherent capabilities of the
hybrid continuum-discontinuum method, makes FEM/DEM to be a suitable simu-
lation tool for rock mechanics practices.
ACKNOWLEDGMENTS
The Y2D FEM/DEM code was originally developed by Antonio Munjiza from
Queen Mary, University of London. Also, Tomas Lukas’s assistance during cod-
ing is appreciated.
REFERENCES
1. G Grasselli, P Egger, Constitutive law for the shear strength of rock joints
based on three-dimensional surface parameters, International Journal of Rock
Mechanics and Mining Sciences, 40, No.1142, 25-40, 2003
2. B Cottrell, Updates to the GG-Shear Strength Criterion. M.Eng., University of

Toronto at Toronto, 2009.
3. A Munjiza, The combined finite-discrete element method. Chichester, West

Sussex, England ; Hoboken, NJ, John Wiley & Sons, 2004.
4. BSA Tatone, G Grasselli, B Cottrell, Accounting for the influence of

measurement resolution on discontinuity roughness estimates. EUROCK 2010
-- Rock Mechanics in Civil and Environmental Engineering. Lausanne,
Switzerland, A.A. Balkema, 2010
5. BSA Tatone, G Grasselli, A method to evaluate the three-dimensional

roughness of fracture surfaces in brittle geomaterials, Review of Scientific
Instruments, 80, No.Compendex, 2009

6. G Grasselli, Shear strength of rock joints based on quantified surface

description. PhD, Ecole Polytechnique Fédérale de Lausanne at Lausanne,
2001.
7. OK Mahabadi, G Grasselli, A Munjiza, Y-GUI: A graphical user interface and

pre-processor for the combined finite-discrete element code, Y2D,
incorporating material heterogeneity, Computers & Geosciences, 36, No.2,
241-252, 2010

ISBN: 978-0-9551179-8-5
This guide has been produced by

the Publications and Web Office Pub6469 for:
The School of Engineering and Materials Science

Queen Mary, University of London
Mile end Road
E1 4NS
Tel: +44 (0)20 7882 8732
Fax: +44 (0)20 7882 3390
email: [email protected]
www.sems.qmul.ac.uk

Discrete Element Methods PDF

Uploaded by

Copyright:

Available Formats

Discrete Element Methods PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Discrete Element Methods PDF

Uploaded by

Copyright:

Available Formats

DISCRETE ELEMENT

Theory and Applications

© 2010 by Queen Mary, University of London

1. ALGORITHMS AND SOLVERS

3. VERIFICATION AND VALIDATION

4. PARALLELISATION, GPU, OPTIMISATION

5. FRACTURE AND FRAGMENTATION, CRACKS

7. SOILS, ICE, ROCK, LANDSLIDES, EARTHQUAKES

Simulations of Discontinua – Theory and Applications 1

FROM NANOSCIENCE TO ASTROPHYSICS –

This is an opening lecture for the DEM5 International Conference on

FINITE ELEMENT METHODS

Simulations of Discontinua – Theory and Applications 2

DISCRETE ELEMENT METHODS

TOWARDS DISCRETE POPULATIONS METHODS

Simulations of Discontinua – Theory and Applications 3

Simulations of Discontinua – Theory and Applications 4

GENERALISATION OF THE CONCEPT OF EMERGENT PROPERTY

Simulations of Discontinua – Theory and Applications 5

an emergent qualitative property such as a new chemical compound or a new state

Figure 1: Droplets of liquid argon from a nanoscale virtual experiment

Figure 2: Planetesimal scale virtual experiment

Simulations of Discontinua – Theory and Applications 6

Simulations of Discontinua – Theory and Applications 7

GPU AS AN EXAMPLE OF WHAT COULD BE

Simulations of Discontinua – Theory and Applications 8

Simulations of Discontinua – Theory and Applications 9

Simulations of Discontinua – Theory and Applications 10

Multi-core Strategies For Particle Methods

John R. Williams, David Holmes and Peter Tilke*

Simulations of Discontinua – Theory and Applications 11

PARTICLE BASED NUMERICAL METHODS

Simulations of Discontinua – Theory and Applications 12

cles provide positions at which to enforce a partition of unity. By partitioning

SPATIAL HASHING IN PARTICLE METHODS

Simulations of Discontinua – Theory and Applications 13

Counter Intuitive Design In a typical SPH simulator, two operations must be

Simulations of Discontinua – Theory and Applications 14

Simulations of Discontinua – Theory and Applications 15

Fig 2. Rayleigh-Taylor instability example showing extremely complex fluid

Simulations of Discontinua – Theory and Applications 16

DEM5-2010, London, 25-26 August 2010 Page 7

5. A. Ferrari, M. Dumbser, E. F. Toro, A. Armanini, A new 3D parallel SPH

Simulations of Discontinua – Theory and Applications 17

Simulations of Discontinua – Theory and Applications 18

DEM VALIDATION – A SIMPLE PROBLEM

Colin Thornton1, Sharen Cummins2, Paul Cleary2

Using a variety of contact force models reported in the literature we

Simulations of Discontinua – Theory and Applications 19

Simulations of Discontinua – Theory and Applications 20

RESULTS AND DISCUSSION

Simulations of Discontinua – Theory and Applications 21

Simulations of Discontinua – Theory and Applications 22

Simulations of Discontinua – Theory and Applications 23

stress on particle interactions’, Chemical Engineering Science, 50, 967-87,

Simulations of Discontinua – Theory and Applications 24

COUPLED DISCRETE ELEMENT AND FLUID

J. E Hilton and P. W. Cleary

We present a coupled discrete element and fluid method in which the

1 3 αi2 αi4 αi6 αi8