CFD-DEM simulation of two-phase

flows in the flash smelting settler

Jani-Petteri Jylhä

School of Chemical Technology

Thesis submitted for examination for the degree of Master of

Science in Technology.

Espoo July 6, 2018

Supervisor

Prof. Ari Jokilaakso

Advisor

M.Sc. Nadir Khan

Copyright ⃝
c 2018 Jani-Petteri Jylhä
 Aalto University, P.O. BOX 11000, 00076 AALTO
www.aalto.fi
Abstract of the master’s thesis

Author Jani-Petteri Jylhä

Title CFD-DEM simulation of two-phase flows in the flash smelting settler

Degree programme Master’s Programme in Chemical, Biochemical and Materials

Engineering
Major Functional Materials Code of major CHEM3025

Supervisor Prof. Ari Jokilaakso

Advisor M.Sc. Nadir Khan

Date July 6, 2018 Number of pages 78 Language English

Abstract
Computer simulations have become a important tool in research. CFD-DEM
coupling has been used to simulate particle and fluid flows in pyrometallurgical
reactors, but no study could be found where it was used to simulate liquid droplets
in fluid. In this work viability of the coupling for flash smelting settler was studied
as the tool could help in studying copper loss mechanisms and optimizing the
process. First, experiment with water droplet in oil was made, and results from the
experiment were used to validate CFD-DEM model using two liquids. Validated
model was used to simulate slag layer in a flash smelting furnace. It was found
that CFD-DEM coupling can be used to simulate settling and collisions of matte
droplets. With limited number of droplets, the droplets formed mixed layers, which
started to separate due to differences in settling velocities. The settling velocities
corresponded well to theoretically calculated values. Also, two different collision
types were observed: high and low velocity collision in which the droplets bounced
apart or stayed in contact until the larger droplet had bypassed the smaller one,
respectively. However, phenomena such as coalescence should be researched and
taken into account for more sophisticated simulation.
Keywords modelling, simulating, copper, matte, particle, droplet, fluent, edem
 Aalto-yliopisto, PL 11000, 00076 AALTO
www.aalto.fi
Diplomityön tiivistelmä

Tekijä Jani-Petteri Jylhä

Työn nimi CFD-DEM simulaatio kaksifaasivirtauksista liekkisulatusuunin

alauunissa
Koulutusohjelma Kemian, biokemian ja materiaalitekniikan maisteriohjelma

Pääaine Toiminnalliset materiaalit Pääaineen koodi CHEM3025

Työn valvoja Prof. Ari Jokilaakso

Työn ohjaaja DI Nadir Khan

Päivämäärä July 6, 2018 Sivumäärä 78 Kieli Englanti

Tiivistelmä
Tietokonesimulaatiot ovat nykyään tärkeitä tutkimustyökaluja. CFD-DEM kyt-
köstä on käytetty tutkimaan partikkeli- ja fluidivirtauksia pyrometallurgisissa
reaktoreissa, mutta aiempaa tutkimusta, jossa tätä olisi käytetty nestepisaroiden
mallintamiseen, ei ollut löydettävissä. Tässä työssä tutkittiin kytköksen toimivuutta
liekkusulatusuunin alauunin mallinnuksessa. Tämä voisi helpottaa kuparihäviöme-
kanismien tutkimisessa ja prosessin optimoinnissa. Aluksi tutkittiin kokeellisesti
vesipisaraa öljyssä. Koetta käytettiin kahden nesteen CFD-DEM mallin validoi-
miseen. Validoitua mallia käytettiin liekkisulatusuunin kuonakerroksen mallinta-
miseen. CFD-DEM kytkös kykeni mallintamaan kivipisaroiden laskeutumista ja
törmäilyä. Rajallisella pisaramäärällä pisarat muodostivat sekoittuneen kerrok-
sen, joka alkoi jakautumaan erillisiksi kerroksiksi erilaisten laskeutumisnopeuksien
vuoksi. Laskeutumisnopeudet vastasivat hyvin teoreettisia laskettuja nopeuksia.
Törmäyksiä oli havaittavissa kahta tyyppiä: suuri ja pieni nopeuksisia. Suurinopeuk-
sisessa törmäyksessä pisarat kimposivat toisistaan erilleen ja pieninopeuksisessa
törmäyksessä pisarat pysyivät kontaktissa kunnes isompi pisara ohitti pienemmän.
Erilaisia ilmiöitä, kuten pisaroiden yhdistymistä ja reaktiota tulisi tutkia ja ottaa
huomioon monimutkaisemmissa malleissa.
Avainsanat mallinnus, kupari, partikkeli, pisara, fluent, edem
 5

Preface

This master’s thesis was conducted at Aalto University in Department of Chemical

and Metallurgical Engineering during 1.1.2018 - 6.7.2018.

I would like to thank Professor Ari Jokilaakso and my instructor M.Sc. Nadir Khan
for their guidance on this work and Juuso Elo-Rauta for assisting me in creating
tools for observing ongoing CFD-DEM simulation.

I would also like to thank my family and friends for their company and support
during my journey to this point.

Otaniemi, 29.6.2018

Jani-Petteri Jylhä
 6

Contents
Abstract 3

Abstract (in Finnish) 4

Preface 5

Contents 6

Symbols and abbreviations 7

1 Introduction 8

2 Background 10
2.1 Copper smelting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Numerical modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3 Studies using CFD-DEM coupling . . . . . . . . . . . . . . . . . . . . 15

3 Computational methods 30
3.1 CFD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.2 DEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.3 CFD-DEM coupling . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

4 Modelling setup 39
4.1 Validations and calculations . . . . . . . . . . . . . . . . . . . . . . . 39
4.2 Created models for CFD-DEM coupling . . . . . . . . . . . . . . . . 45

5 Results and discussion 49

6 Summary and conclusions 70

 7

Symbols and abbreviations

Symbols

dp particle diameter
E Young’s modulus
Ek kinetic energy
g acceleration due to gravity
k−ε k-epsilon turbulence model
k−ω k-omega turbulence model
K bulk modulus
r radius
u strain
vt terminal velocity
γ surface tension
θ contact angle
κ viscosity ratio of bubble
µ dynamic viscosity
ν kinematic viscosity
νp Poisson’s ratio
ρ density
σ stress

Abbreviations

API application programming interface

CAE computer aided engineering
CFD computational fluid dynamics
DEM discrete element method
DPM discrete particle method
FSF flash smelting furnace
GPGPU general-purpose computing on graphics processing units
MSDS material safety data sheet
PSUF particle scale-up factor
RANS Reynolds averaged Navier–Stokes -equation
VOF volume of fluid -method
1 Introduction

Simulating different phenomena in different fields of science has become more impor-
tant as computational capabilities of computers have grown and more sophisticated
simulation software have been developed. Some applications are challenging to study
experimentally, especially in industrial scale. Modelling has made studying these
phenomena possible as different properties can be observed safely from any point
of a system. However, some phenomena are difficult to study with existing tools
and, thus, new tools must be searched for to depict the phenomena with adequate
accuracy.

Some pyrometallurgical reactors have been studied using computational fluid dy-
namics (CFD) and discrete element method (DEM) in CFD-DEM coupling. For
example, cooling elements [1], gas flows [2] and slag flows [3] have been studied with
CFD. DEM has been used to simulate processes such as ore feed to flash smelting
furnace [2]. DEM has also been used in CFD-DEM simulations to study, for example,
gas and particle flows in blast furnace [4]. However, focus of the studies has been in
fluid-solid systems. In this work both phases are immiscible liquids which requires
approximation of liquid-like behaviour of solid particles.

During smelting process, some copper is lost in a slag. Several different mechanisms,
such as droplet entrapment by solid particles and gas bubbles, cause matte droplets
to not settle through a slag layer [5]. Unsettled droplets are removed from the
furnace with the slag. However, the slag is treated in slag cleaning furnace in
order to minimize copper losses. CFD-DEM simulation could be valuable tool in
studying copper loss mechanisms and improving the copper recovery of the smelting
process.

The aim of this work is to investigate if CFD-DEM coupling is a viable option in

studying matte settling through the slag layer. CFD-DEM coupling was chosen for
this study as CFD has been succesfully used to simulate fluid flows in many different
fields of research, but cannot simulate very small droplets with multiphase model [6]
and discrete particle method (DPM) cannot model collisions between particles [7].
DEM could simulate large amounts of very small particles [8]. However, in this work
particles are replaced with droplets and no study could be found, in which DEM
was used to simulate fluids. ANSYS and EDEM have created supported coupling
for the two methods [9]. If CFD-DEM coupling is capable to simulate droplets in
the slag, more sophisticated models, which take into account physical phenomena,
such as coalescence, and chemical phenomena, such as reaction kinetics, could be
created.

Viability of CFD-DEM coupling is investigated by first experimenting with water

droplet in oil and replicating droplet behaviour with soft elastic spherical particle
in the simulation. After validation of the model, the working model was applied to
 9

high temperature environment to study matte droplets in the slag. The evaluation
of viability is based on particle settling rate and realistic behaviour during collisions.
Reactions or coalescence are not taken into account in this work. The modelling
software are ANSYS FLUENT and EDEM for CFD- and DEM-parts of the coupling,
respectively.

First, copper smelting is introduced. Also, some previous studies with the simulation
methods are presented. In the second part, methods used in this study are presented
in more detail. Then, the simulations and experiments are presented in the third
part. These are followed by results, conclusions and summary.

This work is a part of Development of digitalized real time models of pyrometallurgical

processes for adjustable, eco- and cost-efficient production of metals or "DigiPyro"
project. The aim of "DigiPyro" is on creating tools for real-time simulation of py-
rometallurgical processes, including chemical reactions, and to transfer experimentally
obtained data of the pyrometallurgical processes to user-defined models.
 10

2 Background

2.1 Copper smelting

Flash smelting process was developed and started in 1940’s in Finland as an answer
to a energy shortage during World War II. The energy needed by flash smelting
process is produced by oxidizing sulfur and iron in copper concentrate. [10]

FSFs are used in copper and nickel production [10]. In Outokumpu FSF (known today
as OT) the concentrate is fed to a reaction shaft with slag formers with preheated
air. The oxidization reaction melts the feed in the reaction shaft forming gas and
liquid matte and slag. Oil burners can also be used if necessary. The matte and
the slag then descend to the settler of the furnace and produced gases exit through
uptake shaft. The matte descends to bottom of the settler, while less dense slag
floats on top of the matte. Temperatures of the matte and slag are usually around
1250 ◦ C [11]. The matte and the slag are tapped regularly as their surface levels rise.
Flash smelting furnace is presented in Figure 1 [12].

Figure 1: Schematic of flash smelting furnace. [12]

The matte is taken to a converting process such as a Peirce-Smith converter or a

flash converting furnace. A lot of SO2 is also produced in smelting and converting
processes. The SO2 gas must be collected and processed into asulfuric acid, which is
 11

the most often used method, but SO2 can also be compressed into liquid or reduced
to elemental sulfur. [13]

Peirce-Smith converter has been widely used, but it has few major challenges. First
challenge is batch processing. The other is SO2 emissions. Peirce-Smith converter has
open structure, which allows SO2 gases to escape especially during pouring making
capturing the gas difficult.

Some copper is lost with slag, which is taken to slag cleaning process. The amount
of produced slag is significantly greater than the amount of produced metal. 2.2
tonnes of slag is produced per tonne of metal. This slag may have as much as 2.1
% of copper. [14] The main reasons of copper losses to the slag are solid particles
behaving as obstacles for settling droplets, matte surface film rising with gas bubbles,
dispersion of precipitated matte caused by local differences in solubility of copper
and mixing caused by, for example, turbulence. Metallic copper can also penetrate
to refractory of the furnace causing copper losses. [5]

Also, Inco flash smelting is used in copper production. However, the Inco flash
smelting process uses industrial oxygen instead of air and does not use additional
fuel. The Inco flash smelting process produces all of the needed energy by oxidiza-
tion reaction. Inco FSFs are more compact than Outokumpu FSFs. Reactions in
the Inco flash smelting process are similar to those in Outokumpu flash smelting
process. However, the concentration of SO2 in offgases are much higher compared to
Outokumpu process. [13]

Also, different copper production methods exist such as reverberatory furnace and
Ausmelt TSL (top submerged lance). In the Ausmelt TSL process the feed is pelletized
mixture of concentrate, flux, coal and recycled products. The energy required to
smelt the concentrate is produced by coal in the pellets and oil injection. The oil is
injected into the melt through the lance with oxygen enriched air. Then the slag
and matte are taken to electric settling furnace from which the matte continues to
converter. The process is continuous. [15]

Reverberatory furnaces were used over a century and they could produce copper with
little copper losses as slags produces contained only 0.3 – 0.4 % copper. However,
reverberatory furnaces were energy inefficient and caused SO2 emissions as the energy
in the sulfide was not utilized. The concentration of SO2 in offgases were too low for
capturing the gas. [16]

The reverberatory furnace uses fossil fuel to produce energy needed to smelt the
concentrate. Slag and matte are tapped and matte is transported to converter.
[17]

However, low copper content of the slag required low grade matte as the copper
levels were related. Also, the matte grade of 70 % Cu in FSF was not unusual [18],
 12

but the matte grade in reverberatory furnace were typically significantly lower 30 –
40 % Cu. The slag from Peirce–Smith converter also needed to be returned to the
furnace in order to minimize copper losses from the converting. [19]

The slag used in copper production consists of Fe, SiO2 , S and Cu, but may also
have CaO, MgO, Al2 O3 , Co, Mn, Ni and Zn. The slag typically has density of
2800 – 3800 kg/m3 . Copper is recovered from the slag with pyrometallurgical or
hydrometallurgical process, or with mix of the processes. However, the slag contains
too low amounts of some other metals to economically recover them, and is therefore
used in other applications. As the mechanical properties of the slag are good, the
slag is used, for example, in abrasive tools, tiles and pavement. [14]

Slag chemistry requires compatible refractory material. For example, magnesia-

chrome bricks have been used in FSF refractories as the bricks have high resistance
to erosion caused by slag and thermal shock. [10] The settler of Outokumpu FSF
and the hearth of reverberatory furnace can be seen in Figure 2 [20, 21]. The Figure
shows refractory linings constructed from bricks. The campaign life of the FSF can
be increased with the right refractory material and cooling solution. The campaign
life of FSFs has increased from less than two months to ten years [18].

Figure 2: Settlers of the Outokumpu flash smelting (left) [20] and the reverberatory
(right) [21] furnaces.

Slag and matte layer thicknesses vary in the settler as the liquids are tapped with
regular intervals. The thicknesses of the layers can be in range of 0.15m – 0.8 m and
0.3 m – 0,7 m for slag and matte, respectively [22]. The slag and matte are tapped
from tapping holes, which can be located on the different sides or ends of the settler.
For example, the slag can be tapped from side as presented in Figure 1 [12] or from
the end of the settler as presented in Figure 3.
 13

Figure 3: Geometry of model depicting slag layer in flash smelting furnace.

2.2 Numerical modelling

Modelling has become important tool as computer technology has advanced. However,
history of simulation starts from 1777 with Buffon needle experiment, which was
used to calculate probability of dropped needles crossing line on a plane, on which
the needles were dropped on [23, 24]. Simulation evolved rapidly after construction
of computers in 1940s [24]. First large scale simulation done with computers was
simulating nuclear detonation during World War II [25]. During 1960s, several
programming languages were developed and were used to create simulations for
applications such as production plant operation, traffic control and plane ticket
reservations. Development of modelling and analysis tools occurred in 1970s and
1980s and moving images were developed to ease observation of changes in simulations.
[24]

Increasing computational capabilities have caused increase in the use of computer

aided engineering, including modelling flow dynamics and particle interactions. Dif-
ferent phenomena are studied by different modelling methods, but they have been
coupled to simulate, for example, particle interactions in flow fields or stresses caused
by particle loads. Modelling is one research method complementing experimental
and analytical methods [26]. Computational modelling is important tool in several
different applications, such as weather forecast, earthquake research and simulator
training [27]. Additionally, computer simulations are used by entertainment industry
in movies and video games to simulate, for example, particle trajectiories in special
effects [28].
 14

Several different methods of computer assisted engineering (CAE) can be used in

different fields of research. Such methods include computational fluid dynamics
(CFD), finite element analysis (FEA), multi-body dynamics (MBD) [29] and discrete
element method (DEM) [30]. Some methods can be coupled to produce more precise
results such as wind pressure and structural stresses on airplanes, particle movement
in fluid and bulk material causing forces on equipment [29, 31].

Some methods can be coupled in order to create more complex models, where one
software calculates one part of the simulation and other calculates other part. For
example, CFD can be coupled with DEM which allows studying interactions between
discrete particles and particle flow in surrounding fluid flow. Such applications are,
for example, solid-gas flows in blast furnaces and in separation and mixing systems,
such as cyclones and fluidized beds.

Like other methods, also modelling as a research method has its advantages but also
disadvantages. Some of these have been collected in Table 1.

Table 1: Some advantages and disadvantages of modelling.

Advantages Disadvantages

• Detailed data of the model can • Accuracy of the model must be

be obtained relatively cheaply and critically evaluated before results
quickly. [32] can be accepted. [32]
• The environment and conditions • Simplification of the model is of-
in a simulation are controlled. [32] ten necessary. [32]
• New designs can be tested and op- • Many models, including turbu-
timized before making any physi- lence models, are incomplete and
cal models or constructions. [32] are applicable only on certain type
• Some systems cannot be experi- of situations [32, 35].
mentally studied as they may be • Models may be relatively sensitive
too dangerous or the studied sys- to defined boundary conditions
tems are too large. [32] and errors during calculations.
• If necessary, extremely detailed • Large and detailed models may
models can be created. [32] be computationally very resource
• Users can implement their own intensive.
modifications on existing mod-
elling platforms with user defined
functions (UDF) or application
programming interface (API) [33,
34].
 15

2.3 Studies using CFD-DEM coupling

CFD-DEM coupling has previously been used to simulate solid-gas and solid-liquid
flows in applications such as blast furnaces [4, 36, 37], cyclones [38] and fluidized
beds [39–41]. Also, three-phase flows have been studied [42]. More complex systems
such as combustion reactions in blast furnaces [43] or particle melting [42, 44, 45],
can be taken into account in CFD-DEM models. However, studies of CFD-DEM
systems where particle is in liquid form has not yet been reported.

Particle wetting

In study investigating droplet wetting in a fluidized bed droplets were depicted by

particles. The droplets were sprayed on other particles from top of the bed. The
schematic of the model and some of the results from the study are presented in
Figures 4 and 5, respectively. [41]

Figure 4: Schematic and dimensions of the fluidized bed with top-spray. [41]
 16

Figure 5: The particle wetting in fluidized bed with top-spray. Gray particles are
dry, red particles have been wetted. [41]

A contact model of the droplets was modified to deposit the droplets on surface of
other particle or on the wall and then be deleted. Collision data of each particle was
stored to distinguish collisions between the particle and the wall, droplet or other
particle. [41] In this study, however, the droplets did not have more complex liquid
properties, such as surface tension and viscosity. Also, fluid related phenomena, such
as coalescence, was not taken into account.

Layer inversion

CFD-DEM model utilization in liquid-solid systems has been validated by a layer

inversion in liquid fluidized beds by using CFD-DEM model. Few studies were
selected to include conditions from a range of different hydrodynamic conditions. To
reduce computational requirements the simulations were conducted by using pseudo-
2D bed and two different kinds of particles in either water or water polyethylene
glycol mixture. More detailed information about the simulations and original studies
(cases A, B and C) are presented in Table 2. The simulation time was up to 300 s
before steady state was reached. [39]

The study included 12000, 37300 and 21400 particles in systems A, B and C,
respectively. Column height in system A could not be made high enough for expansion
with the highest liquid velocities. As a result, comparison between simulated and
experimental results of system A works mostly as a proof that CFD-DEM can
represent particle inversion.
 17

Table 2: Eperimental in simulation parameters used in the study. [39]

In the system B it was concluded that drag force model was the main factor in
formation of voidage and composition profile. Results from the system B are presented
in Figure 6. In the experiment, layer inversion conditions are at u = 12.5 mm/s.
Comparison of the experiment and simulation results show that profile of bed and
mixed layer interface height as a function of inlet velocity were matching well as
presented in Figure 7. [39]
 18

Figure 6: Fluidization at different inlet velocities in the system B. Black particles

are denser than grey particles. [39]

Figure 7: Bed and mixed layer interface height as a function of inlet velocity in 27
◦
C and 25 ◦ C. [39]

Figure 8: Bed and mixed layer interface height as a function of inlet velocity in 27
◦
C and 25 ◦ C. [39]
 19

Particle collision

A study was made on collision dynamics in fluidised bed granulator. The study used
Hertz-Mindlin based contact model and used older versions of EDEM and FLUENT.
In the study, 150 000 spherical particles with 2 mm diameter were simulated. The
simulation parameters used in the study are presented in Table 3. [46]

Table 3: The simulation parameters of the particle collision dynamics study. [46]
Parameter Value Unit
Fluidisation air flow rate 360 kg/h
Atomizer flow rate 5.7 kg/h
Gap distance below Wurster 30 mm
Normal coefficient of restitution 0.8 -
Sliding friction coefficient of particles 0.1 -
Rolling friction coefficient of particles 0.05 -
Poisson’s ratio of particles 0.25 -
Shear modulus of particles 108 Pa
DEM simulation time step 10−6 s
CFD simulation time step 10−4 s
Total simulation time 4 s
Number of particles 150 000 -
Number of grid cells (CFD) 9600 – 17 600 -

In the simulation, the particles were wetted in three different types of granulators:
top-spray, Wurster-coater and spouted bed. The granulators are presented in Figure
9. The nozzle in a granulator formed spray cone in which the particle and collision
dynamics were analysed. It was found that particle velocities were 0.47 m/s, 0.61
m/s and 0.60 m/s, but relative particle-particle collision velocities were only 0.13 m/s,
0.11 m/s and 0.08 m/s for the Wurster-coater, the spouted bed and the top-spray
granulators, respectively. The Wurster-coater showed interparticle collision frequency
of 357 1/s, which was significantly smaller compared to frequencies of 1149 1/s in
the spouted bed and 1300 1/s in the top-spray granulators. In addition of relative
collision velocity, the collision frequency was a important factor on agglomeration:
too high collision velocity would have caused the particles to rebound while lower
velocity collisions lead to particle agglomeration. Growth rate of the agglomerate
was affected by collision frequency and agglomerate breakage. [46]

A lab scale experiment was also made. The experiment was used as a comparison
when particle agglomeration was investigated. The experiment used dried dextrose
syrup and pure water. The study concluded from the simulation results that the
Wurster-coater had low agglomeration and breakage rates and medium wetting
intensity, but inferior agglomerate strength was expected. The top-spray had low
 20

adhesion rate, medium breakage rate and low wetting intensity which should cause
low growth rate. According to simulations, the spouted bed could produce dense
agglomerate, as it had high wetting, growth and breakage rates. These results were
verified with the experiments. The study concluded that the Wurster-coater was the
best for producing large and stable agglomerates, but the most stable, dense and
compact agglomerate was produced with the spouted bed granulator. [46]

Figure 9: The three granulator types used in the study, a top-spray, b Wurster-coater
and c spouted bed. [46]

Particles used in the study were more cohesive than droplets in this study. However,
this should not affect the simulations in other terms, except that rebound of the
droplets should occur at lower collision velocity. Also, droplets have lower friction
coefficients which together with lower cohesiveness decrease adhesion between them.
This should decrease the agglomeration significantly in this study. Properties of the
droplets are discussed in more detail further in this work.

Pyrometallurgy

CFD-DEM coupling has been used in studying other pyrometallurgical processes,

such as solid material flow and formation of cohesive zone in blast furnace. One
research team first studied if CFD-DEM coupling could be used to study 3D-model
of blast furnace. The study did not include chemical reactions or heat transfer. Also,
only gas velocity was taken into account as particle velocity was assumed to be small
enough compared to the gas velocity. The volume of the modelled blast furnace
was over 5000 m3 [37]. Due to large volume of the blast furnace and computational
intensity of DEM the particles needed to be enlarged. Also, the blast furnace was
modelled as semicircular which decreases the computational requirements of the
model. Model had predefined melting zone under which ore particles were eliminated
 21

as they supposedly did not influence the model. Coke was removed at raceways in
arbitrarily-specified intervals mimicking coke consumption. The calculation region of
the blast furnace is presented in Figure 10. [4]

Figure 10: Calculation region of three-dimensional blast furnace model. Layout of

the melting zone and raceways are also presented. [4]

In order to the model to work with different particle sizes in the CFD and DEM
models the void fraction of particles had to be calculated by using controlled volume
method. To transfer the void fraction of particles to the cells of the CFD-model both
averaging method and weighted mean method were compared. The weighted mean
method showed better results, and was selected to be used. [4]

The model produced fluidization of the ore and coke particles in the middle area of
top of the blast furnace when gas flow was taken into account. The fluidization was
caused by the increased gas velocity in the center of the throat of the blast furnace
during charging, which can be seen in Figure 11. However, the packed bed showed
no noticeable changes. Even though, the gas did not noticeably affect the particles it
did reduce normal stresses caused by particle load by 8 % and 12 % in lower and
upper parts of the blast furnace, respectively. [4]
 22

Figure 11: Charging the blast furnace and gas velocity change in the throat of blast
furnace. [4]

A study on gas injection from tuyeres and auxiliary tuyeres was conducted by the
same group. In the study, auxiliary tuyeres were located at 8 and 16 meters above
the tuyere level. The gas injection from the auxiliary tuyeres in the shaft of the blast
furnace counted to 30 % of the total gas flow to the furnace. The auxiliary tuyere
levels are presented in Figures 12 and 13. The void fraction was calculated in same
way as in their previous paper presented above. [36]

Figure 12: Velocity vectors in auxiliary tuyere levels. [36]

 23

Figure 13: Velocity vectors in the blast furnace. Auxiliary tuyere injection shown by
white arrows. [36]

Cases b and c correspond to upper and lower shaft injection, respectively. Inlet
volume at tuyere was 8240 Nm3 /min and inlet volumes at auxiliary tuyeres were
3516 Nm3 /min. In case b-1, 12 auxiliary tuyeres were blowing 39.3 m/s, and in
case b-2 6 auxiliary tuyeres were used with 78.5 m/s inlet velocity. In cases c-1 and
c-2, 20 and 10 auxiliary tuyeres were injecting gas at velocity of 19.6 and 39.2 m/s,
respectively. As can be seen from Figures 12 and 13, the effect of shaft injection
does not penetrate deep into the blast furnace, but instead turns quickly to rising
vertical flow. [36]

CFD-DEM coupling was used to solve gas drag forces caused by auxiliary tuyeres. It
was found that gas drag force distribution had no significant differences as can be
seen in Figure 14. Due to burden resistance, the amount of tuyeres or gas velocity
did not have significant effect on overall gas flow in the blast furnace. This can be
also seen when inspecting isobar planes as gas inject did not affect them above the
gas injection level. [36]
 24

Figure 14: Distribution of the gas drag force above shaft injection level. [36]

The same research group also conducted a study on how uneven gas flow from tuyeres
affects solids flow in the blast furnace. In this study, CFD-DEM model was used to
study gas and particle flow speeds and isobar planes. The study was conducted with
varying number of tuyeres blocked with varying spacing. Also, the gas inject velocity
from tuyeres was varied. The coke was not consumed in raceway if corresponding
tuyere was not active. [37]

Figure 15: The tyuere setup and particle velocities with uneven gas flow from the
tuyeres in different cases at (a) 4 m and (b) 2 m above (c) the tuyere level. [37]
 25

As can be seen from Figure 16, one clogged tuyere does not have significant effect on
the gas velocity or pressure distribution in the blast furnace. Also, even distribution
of clogged tuyeres, as in case B, has relatively small effect compared to cases A,
D and E, where clear difference can be seen especially in the isobar planes in the
lower part of the blast furnace. However, gas velocity and pressure distribution
becomes more even in the upper parts of the blast furnace. Unlike the gas velocity
and pressure distributions, particle velocity distribution has differences both in upper
and lower part of the blast furnace. Differences are, however, larger in the upper
part. [37]

Figure 16: Gas velocity vectors and isobar planes in the blast furnace. [37]

These previous three studies had time step size of 7.2 s and 10−4 s for CFD and
DEM, respectively. The gas flow was set to update every 600 steps in DEM model
and the particles consumption rate was accelerated 120 times to accelerate packed
bed descension rate to increase time range that could be analyzed with DEM. This
was done to keep simulation time reasonable despite of large particle number and
large model. [4, 36, 37] The time step values were several orders of magnitude larger
than in other studies presented in this work. This is very interesting, but the Figures
from studies seem to have stable model as the particle and gas velocities are not
unreasonable.

Another study was made, where CFD-DEM coupling was used to create LKAB
experimental blast furnace as a computer simulation. The volume of the blast
furnace was 8.2 m3 . The blast furnace included 13000 particles in the solid-gas
 26

flow, heat transfer and chemical reactions. However, only the key reactions were
included. To decrease computational requirements of the simulation, the study used
two-dimensional slot model. For the CFD model, gas composition at inlet was set
according to heat and mass balance. The meshing was examined, and it was found
that results did not improve significantly if finer mesh was used. The slot model of
the blast furnace is presented in Figure 17. [43]

Figure 17: Dimensions and the CFD mesh of two-dimensional slot model of the blast
furnace. [43]

Also in this study, coke particles in the raceway were removed at given rate to simulate
combustion. The iron ore particles were removed if their temperature reached 1400
◦
C. This occurred as the temperature field was calculated by using heat and mass
transfer by solid-gas flow and chemical reactions. [43]

The increasing temperature softened the iron ore which formed the cohesive zone
in the blast furnace. This behaviour was, however, based on another study [44].
The model proved capable to simulate reasonably well factors such as gas and solid
temperature distribution and the shape of the cohesive zone. Figure 18 presents
comparison of the predicted and measured cohesive zones as well as softening and
deformation of the ore particles. [43]
 27

Figure 18: Particles depicting iron ore softening and deforming on the left. On the
right comparison between predicted and measured cohesive zones is shown. [43]

The CFD and the DEM have also been used to model improved disperser for a
FSF. However, in the study about the improved disperser, the CFD and the DEM
were used separately, not as coupled system. The CFD was used to study flow of
oxygen enriched gas and combustion in reaction shaft of the furnace. The CFD
results are presented in Figure 19. The study found that the temperature was over
1800 ◦ C at the lower parts of the reaction shaft. The muzzle velocity of the gas
entering the reaction shaft was 100 m/s, but it decreased rapidly as the gas diffused
outwards. [2]
 28

Figure 19: Results from modelling the FSF reaction shaft with the CFD. [2]

The DEM was used to study concentrate flow to the furnace. According to the DEM
model, the existing disperser was not capable of delivering the concentrate with
uniform distribution. The new disperser with helical spirals was developed. The new
disperser produced significantly more uniform feed distribution. Both DEM results
are presented in Figure 20. [2]

The new design was installed in the furnace. When the slag composition was stabilized
and the data was compared to the data from the old design, it was found that the
oxygen efficiency had improved by 10 % and amounts of both oxygen and SO3 in
offgas was decreased. This translates to reduced oxygen consumption, increased
campaign life, more stable slag composition, reduced maintenance requirements and
improved autogenous combustion. The new disperser design was put to further
testing. [2]
 29

Figure 20: Particle flows with different disperser designs. [2]

 30

3 Computational methods

3.1 CFD

Computational fluid dynamics (CFD) was at first used in aerospace engineering,

but has been adopted by many different engineering fields, such as automotive,
chemical and civil engineering [26]. The CFD is used to model for example fluid flows,
reactions and heat transfer [6, 26, 47]. The CFD can also be used to solve multiphase
models with gas-liquid and liquid-liquid systems. However, multiphase model requires
droplets to be significantly larger than the computational cells. [6]

The CFD is based on the Navier–Stokes equations which are derived from Newton’s
second law. The equations (1) and (2) form the Navier–Stokes equations. [26] The
equations can be solved directly if flow in the system is laminar. However, flows
in practical applications are often turbulent. With turbulent flows several different
turbulence models can be used to solve the equations. However, they all have
advantages and disadvantages when compared to other turbulence models. The
turbulence models can be divided in two types: to Reynolds averaged Navier–Stokes
models (RANS) and direct numerical simulations (DNS). The DNS numerically
integrates Navier–Stokes equations and is limited to simple geometries, while the
RANS models add viscosity to the Navier–Stokes equations and can be used with
more complex models. [35, 48]. The user must be able to select right solver for the
model [6]. The k-epsilon (k − ε) model is often considered as good general-purpose
turbulent model, but also k-omega (k − ω) turbulence model is considered as good
general-purpose model [35].

∂u ∂u ∂u 1 ∂p ∂ 2u ∂ 2u
+u +ν =− +ν 2 +ν 2 (1)
∂t ∂x ∂y ρ ∂x ∂x ∂y

∂ν ∂ν ∂ν 1 ∂p ∂ 2ν ∂ 2ν
+u +ν =− +ν 2 +ν 2 (2)
∂t ∂x ∂y ρ ∂y ∂x ∂y

µ
ν= (3)
ρ

The CFD model requires a mesh to be generated to designate the cells for solving
the flow. Meshing also groups cells into a boundary zones which are needed to
apply boundary conditions in simulation. [49] Meshing is generated by algorithms
in software and often comes in commercial software, such as ANSYS Workbench,
but also independent mesh generation software are available. The CFD results are
sensitive to the quality of mesh, which depends on cell shape and size, but the mesh
 31

has to be fine enough to show all relevant geometries. Meshing also has effect on
converging rate. Often, hexahedral cells are favoured as they are computationally
effective and require less cells to achieve the same computational accuracy as with
tetrahedral cells. [49,50] However, hexahedral cells are not always achievable due to a
lack of mesh generation algorithm that is robust enough. Increase in computational
power of computers has decreased the significance of high cell counts in simulation,
which makes easier to use more tetrahedral cells without compromising mesh quality.
Also hybrid mesh consisting of hexahedral and tetrahedral cells can be utilized to
achieve high quality mesh. [49,50] The mesh quality consists of three different factors:
skewness and aspect ratio of individual cells and smoothness of the mesh. [49]

A sample model is depicting cooling block with a internal water flow. Roughly similar
plate coolers may be utilized in, for example, pyrometallurgical reactors [51, 52]. A
copper cooling systems with internal water circulation are considered to be the most
feasible cooling solutions available [53]. Meshing of the sample model is shown in
Figure 21. Heat transfer and internal water flow in the model are presented in Figures
22 and 23, respectively. The cooling block was modelled with ANSYS FLUENT and
was created only to illustrate some basic applications of CFD-modelling. The block
has dimensions of 200 mm x 500 mm x 1000 mm, channel is round and has diameter
of 50 mm. Water input was 1 kg/s and cooling block has front against hot wall at
350 K.

Figure 21: Unstructured meshing of the cooling block model split from the center
line.
 32

Figure 22: Heat transfer in the cooling block with internal water flow.

Figure 23: Internal water flow in the cooling block.

 33

The CFD can also be used to simulate several flowing fluids as presented in Figure
24. In the Figure, a water sphere with diameter of 20 mm is falling in box filled with
air. The box has width and depth of 50 mm and height of 150 mm. The starting
point of the droplet is at height of 120 mm. Falling of the droplet is simulated using
the volume of fluid (VOF) method.

The volume of fluid can be used to model interactions of separate fluid phases. The
interface of the surface is tracked in fixed mesh with Eulerian method. [54] The
eulerian mesh tracks cells, which contain material. The cell can also have several
materials, which forms interface between phases. [55]

Figure 24: Water sphere falling in air at times of a) 0 s, b) 0.05 s, c) 0.1 s, d) 0.15 s,
e) 0.2 s and f) 0.25 s.

Many commercial CFD programs can be modified with additional user defined func-
tion/subroutine (UDF/UDS) -files. A UDF-file can be used to give initial parameters
in a simulation and new functions, such as customizing boundary conditions, material
properties and models, adding model equations, adjusting functions and initializing
 34

model. [33, 56] Some programs, such as ANSYS FLUENT, use UDF-files written in
C [6, 33] but some software use Fortran [6].

Commercial software, such as ANSYS FLUENT, can utilize the discrete particle
method (DPM). The Discrete particle method can be used to study particle tracks in
a continuum. The particles can have different shapes. However, rather than creating
realistic geometry for the particle, a shape factor is used to calculate drag forces
focused on the particle.

3.2 DEM

Discrete element method (DEM) is used to model interactions between granular

material flows, such as ores and grains [57]. These particles can be in any shape
and size from spherical to complex and from large to small [8, 57]. However, more
complex shapes are computationally more resource intensive. DEM can also be
used to calculate electrostatics, particle bonding and heat transfer [28]. In Figure
25 a snapshot of particle flow in a hopper is presented with both spherical and
non-spherical particles compared.

Figure 25: Comparison of particle flows in hoppers with spherical (on left) and
unspherical particles. [58]

The DEM calculates contact and external forces for every particle individually, which
can be used to, for example, predict particle breakage in particle flow. However, it
 35

also makes DEM computationally intensive and makes efficient parallel computing
essential for practical applications [57]. The calculation time of DEM is affected by
material properties, such as bulk modulus, density and particle size which all affect
Rayleigh time-step. Also, hardware has large effect as more powerful processors
can parallelize the calculations to larger number of computing cores and perform
calculations faster. Also, the number of the particles affects the calculation times
almost linearly. [59]

The particles are allowed to overlap. The amount of overlap is calculated by using
Young’s modulus of the particle [60]. The overlapping distance is used with normal
and tangential velocities to calculate contact forces of particles [61]. Some methods
use Voigt model presented in Figure 26 to describe the contact forces between
particles [4, 36, 37], while others, such as EDEM, use Hertz-Mindlin contact model
presented in Figure 27 [38, 62]. However, several contact models may be available
within DEM software.

Figure 26: Concept of Voigt model. [4]

Figure 27: Concept of Hertz-Mindlin contact model. [62]

To avoid excess overlapping of particles, time steps size must be kept relatively
 36

small. Excessively large overlap would cause numerical instability and inaccuracy as
calculated forces would be too large. Rayleigh waves are used to estimate suitable
time step by calculating the Rayleigh time step. The time step used in simulation is
often in the range of 20 - 40 % of the Rayleigh time step. [63]

The particles in the DEM model can be spherical, some specific geometry or clustered
spheres depicting geometry of the particle. The shape of the particle has effect on
performance of the model. Well defined simulation will be computationally easiest
to solve with spherical particles, but realistic geometry may provide more accurate
solution. [64] However, some results have shown that simple realistic particles may
be easier to calculate than similar shape produced with clustered spheres [65].

If the used software is not capable to simulate some phenomena happening in model,
users may be able to modify their simulation by writing and adding their own code
in simulation [34, 66, 67]. Added code can add custom physics for example to add
moisture induced particle cohesion, fracturing of particles or magnetism [34]. Some
developers may even offer software customization as a service [68]. For example,
PFC and EDEM offer modification of simulation with written C++-code. In case
of EDEM the code is inserted in the simulation by using application programming
interface (API) [34].

3.3 CFD-DEM coupling

CFD and DEM can be coupled to get more precise simulations for models containing
discrete particles and fluids. [9, 69] The coupling is done with a UDF-file linking the
two methods. Some commercial DEM software, such as EDEM and Rocky, come
with official support for the coupling with commercial CFD software.

CFD and DEM can be coupled in two different ways: one way or two way coupling.
One way coupling works best on solutions where solids content is low. In the one way
coupling fluid flow is first solved in CFD and then imported in DEM, where flow is
used to solve particle flow. However, particles do not affect the fluid flow. In the two
way coupling fluid flow is first solved for one time step, then flow data is imported
to DEM solver. The particle flow is solved to the same time point as the CFD, and
particle data is imported back to the CFD solver and this is repeated until simulation
has run pre-defined amount of time. Depending on the simulation, results may vary
largely between these two coupling methods as can be seen in Figure 28. [28] The
flow sheet depicting calculation of the two way coupling of a gas-solid-simulation is
shown in Figure 29 [4, 37].
 37

Figure 28: Comparison of the one way and two way coupling results after simulating
sand falling through water. [28]

Figure 29: Calculation flow sheet of the two way CFD-DEM coupling in gas-solid-
simulation. [4]
 38

One challenge in the CFD-DEM coupling is the small time steps of the DEM model.
A DEM model can have time steps of, for example, 10−4 s, while CFD may have
time step size of several seconds. The DEM time step may be even smaller [41]. Also,
large DEM models may require larger particles to reduce simulation time, but CFD
models may be able to use real size particles. Larger particles may cause scaling
issues on the contact and drag forces. [4, 37, 70]

In the CFD-DEM model, DEM uses Lagrangian approach, where discrete elements
are moving in a domain. The Newton’s second law is used to describe movement of
the particle according to equation (4).

dvi V i βm
mi = Vi ∇p + (u − v) + mi g + Fc,i (4)
dt ϵp

where m is mass, v is gas velocity, t is time, V is volume, p is pressure, βm is

inter-phase momentum transfer coefficient, ϵp is particle volume fraction, u is particle
velocity, g gravitational constant and Fc is contact force. [41]

Figure 30: Basic principle of pseudo-three-phase CFD-DEM model where DEM is

simulating droplets and alumina spheres. [41]
 39

4 Modelling setup

4.1 Validations and calculations

An experiment was made in this work in order to validate the created droplet model.
In the experiment, water droplet was observed flowing through immiscible liquid.
Chosen liquid was wood oil as it was immiscible with water and less dense than water.
The oil was in a clear container and water droplets were dropped in the oil with
pipette. A slow motion video of the droplet was taken and then falling speed was
calculated from the video. As can be seen in Figure 31, the shape of the falling water
droplet is practically spherical. The droplet also holds its shape close to spherical
with missing hemisphere at the contact area.

Figure 31: Water droplet falling through oil (left) and in final shape at the bottom
of the container (center). Diameter of the droplet was roughly 4.5 - 5 mm.

The chosen wood oil was Teknos Woodex as Teknos provided both density and
kinematic viscosity in their MSDS [71]. The kinematic viscosity was then converted
to dynamic viscosity by using the equation (3) [26]. However, the kinematic viscosity
given by Teknos was upper limit for the viscosity. Thus, few different viscosities
were studied. The density of the oil was also measured to ensure realistic behaviour
of the simulations. The properties of water were taken from the ANSYS FLUENT
database.

The modelled droplets were created as cubes inside cuboid geometry as this allowed a
high quality mesh to be created as presented in Figure 32. The orthogonal quality of
the mesh was 1 and skewness was in range of 1.3057 · 10−10 – 1.3061 · 10−10 . For VOF
model, a 2-phase model was chosen to track different phases in the system. Realizable
k − ε turbulence model with scalable wall functions was used to simulate the fluid
flow. The interfacial tension between water and oil did cause the droplet to form a
sphere, but cubical beginning form did allow the creation of high quality mesh. The
values for the interfacial tension was taken from the literature [72]. The wall adhesion
 40

was also modelled, but the walls were given very high contact angles as they were
wetted by oil in the experiment. The VOF model was used to simulate the behaviour
of the droplet as the droplet impacted to the bottom of the oil container.

Figure 32: High quality mesh for VOF model.

The VOF model with smaller droplet was also made. The smaller droplet had edge
length of 3.3 mm, which has similar volume as the droplet in experiment. The
modelled volume was decreased to 19.8 mm deep and wide with height of 123.3 mm.
However, this model was not able to keep the water droplet intact and also the falling
velocity of the droplet was over 20 times slower than in the experiment. Decreasing
time step size and mesh size did not improve results of the model. The problems
were possibly caused by very small water volume fractions. Results from one of the
tests is presented in Figure 33.
 41

Figure 33: Volume fraction of small water droplet after simulating droplet descending
for 40 seconds. Starting position of the droplet is presented with a wire-frame.

In the DPM model no droplet geometry was made as in ANSYS FLUENT the droplet
is injected by using DPM settings. Single particle was injected from a specified point
in the middle of the upper part of the geometry. The trajectory of the particle was
tracked as it fell through the oil. From the trajectory with different time steps, the
falling speed was calculated and compared to results from the experiment.

The dimensions of geometries and measurements from the experiment are given in
Table 4 and all used variables are presented in Table 5 below.
 42

Table 4: Dimensions of oil container, water droplet and created geometries.

Droplet Drop
Height Width Depth
size height
(mm) (mm) (mm)
(mm) (mm)
4.5 - 5
Experiment 120 100 65 120
(diameter)
4.5 & 5
DPM 130 40 40 120
(diameter)
VOF single
130 19.8 40 7 (edge) 120
droplet

Table 5: Variables used in models. γ is oil-water interfacial tension.

ν (cm2 /s) µ (kg/ms) ρ (kg/m3 ) γ (mN/m) θ (◦ )
0.01845 [71]
900 [71]
Oil <0.205 0.015 49 [72] -
837.5
0.010
Water - 0.001003 998.2 49 [72] 170

To get the DEM model to produce realistic behaviour, the droplets were approximated
as soft, elastic spheres. This required estimating and iterating values for solid material
properties, such as the Young’s modulus and Poisson’s ratio.

Hooke’s law shows relation between bulk modulus K, Young’s modulus E and
Poisson’s ratio νp as shown in equation (5) [73]. The isothermal bulk modulus of
fresh water is 2180 MPa [74] and as water is almost incompressible, Poisson’s ratio
of water is close to 1/2. A study was made, that estimated Poisson’s ratio of salt
water to be 0.4999 [75] and the same value was chosen for this study. With these
values, Young’s modulus of E = 1.3 MPa was calculated.

E
K= (5)
3(1 − 2νp )

Similar values were chosen for depicting a matte droplet as the matte droplet is also
practically incompressible. Slightly higher Young’s modulus of 1.5 MPa was chosen
as water has lower viscosity (µwater = 0.001003 kg/ms) than matte (µmatte = 0.01
kg/ms [76]).

The properties for the slag and matte droplets were taken from literature and are
presented in Table 6. Coefficients of friction and restitution were estimated and
 43

validated with the water droplet experiment as they are solid material properties.

Table 6: Used properties for slag and matte.

Slag Matte
µ (kg/ms) 0.45 -
ρ (kg/m3 ) 3150 [3] 4600 [76]
νp - 0.4999 [75]
E (MPa) - 1.5
Coefficient of Restitution - 0.5
Coefficient of Static Friction - 0.01
Coefficient of Rolling Friction - 0.01

Validations were also made with larger particles by using particle scale-up factor
(PSUF) of the EDEM-FLUENT coupling. However, this required also changes to
the particle density in order to achieve results similar to the true size simulations.
As the radius of the particle was changed, the density was divided by change to the
power of 3 according to equations (6) and (7). This was done because duplication
the radius of the sphere increases its volume eightfold as can be seen in equation (8).
The calculated values are presented in Table 7.

ρp
ρscaled = (6)
xs 3

rscaled
xs = (7)
rp

4
V2 π(xs r1 )3 xs 3 r1 3
3
= 4 3
= 3
= xs 3 (8)
V1 3
πr1 r1

where ρscaled is the particle density after scaling, ρp is original particle density, xs is
scaling multiplier, rscaled is the radius of the particle after scaling, rp is the radius of
the original particle, V1 and V2 are volumes of spheres 1 and 2, respectively, and r1
is the radius of the sphere 1.
 44

Table 7: Calculated densities for droplets when using different particle scale-up
factors.
PSUF ρwater (kg/m3 ) ρmatte (kg/m3 ) Density scaling
1 998.2 4600 1
2 124.8 575 1/8
4 15.60 71.88 1/64
8 1.950 8.984 1/512

This data may not be entirely correct as it was revealed that PSUF in EDEM-
FLUENT coupling did not return scaled values to EDEM. This was corrected in later
update of the coupling. [77] However, the effect of this must be later evaluated.

As the stable time step size in CFD seemed to depend on the particle size in the drag
calculations, significant difference in calculations speed could not be observed. New
approach was studied where the droplets have the same diameter but the densities
were changed to get the terminal velocity of the droplets to match original droplets
with different diameters. Original particle diameters were 100, 250 and 500 µm, the
enlarged diameter was 2 mm. However, the particle depicting 100 µm droplet was
also enlarged to 1 mm diameter. First, settling rate was calculated with Stokes’ law
equation (9) [78]. Then the new particle density was solved with equation (10) by
using new diameter and original terminal velocity. Calculated values are presented
in Table 8.

2 ρp − ρf 2
vt = gr (9)
9 µ

9 µvt
ρp = + ρf (10)
2 gr2

where vt is terminal velocity, ρp is particle density, ρf is density of fluid medium and

µ is viscosity of the fluid g is acceleration due to gravity, r is particle radius. This is
only applicable if crition vt r/νp is small. [78] In this case, the largest criterion value
was for a particle depicting 500 µm droplet, which was roughly 1.76 · 10−6 . The
particles were tested with CFD time step sizes of 10−4 s, 5 · 10−5 s and 10−5 s. Used
densities were approximations as EDEM rounds density to four digits.
 45

Table 8: Calculated terminal velocities and densities for the original particles and
enlarged particles.
ddepicted (µm) dreal (µm) ρ (kg/m3 ) vt (m/s)
100 100 4600 1.76 · 10−5
250 250 4600 1.10 · 10−4
500 500 4600 4.39 · 10−4
100 1000 3165 1.94 · 10−5
100 2000 3154 1.82 · 10−5
250 2000 3173 1.11 · 10−4
500 2000 3241 4.41 · 10−4

4.2 Created models for CFD-DEM coupling

The first created CFD-DEM model was the validation model. A single particle was
injected in a cuboid container with measurement presented in Table 4. Then the
same container was used to inject two particles with interval of 0.25 seconds. The
models were used to select properties for the particle depicting a droplet and to select
the right drag law. The variables for the oil were ν = 0.015 kg/ms and ρ = 837.5
kg/m3 . Properties of the water droplet were set as calculated previously in this
study.

A enlarged cuboid container was used to validate the PSUF models. The dimensions
of the container were 170 mm x 100 mm x 100 mm. The distance between the bottom
of the container and the lowest point of the particle was 120 mm at the starting
point. Realizable k − ε with scalable wall functions was used and the fluid was set
without flows. Original diameter of the water droplet and the matte droplet were 4.5
mm and 2 mm, respectively.

The k − ω shear stress transport (SST) turbulence model was used in a cube model
as it produced more stable scaled residual curves than the k − ε model. However,
the k − ε was used in larger models. All tests, except the water droplet validations,
were made with PSUF set to 4 in EDEM-FLUENT coupling. CFD time step size
was 10−4 as with a higher time step size the smallest particles started to bounce
upwards out of domain. This was most likely caused by too large movement causing
unrealistically large drag force. Time step size in EDEM was controlled by the
software. Particle injection velocity was set to 7 m/s as previous study has simulated
similar gas velocities in the reaction shaft as presented in Figure 34 [79].

The one way and two way coupling were compared due to the size of required time
step. However, the CFD time step size could not be increased by selecting one way
coupling as smaller particles were not stable. This could be observed by too small
 46

particles disappearing at every CFD time step.

Figure 34: 7 m/s gas velocity iso-surface. [79]

The second model created was a cube depicting part of the slag layer. The edge of
the cube was 500 mm. 2000 particles with diameter of 500 µm was injected into the
cube from half of the top face. Slag flow was calculated to roughly correspond to the
value of 70 t/h in the slag layer of the whole FSF settler. The value of the mass flow
was from selected from literature [3], but only one third of the full slag layer was
taken into account. The cube is presented in Figure 35. In the next test, the particle
injection was modified to have particles with diameter of 50, 100, 250 and 500 µm
with equal particle size distribution. However, the particles were created four times
larger in EDEM to keep the time step length feasible.

Figure 35: Cube model used to test different particle flows in a smaller system.
 47

As the particles affected the slag flow in the cube unrealistically much a larger test
model was created. The test model was a cuboid with dimensions of 6000 mm x
6000 mm x 500 mm. A circular area with the diameter of 4500 mm was on the top
surface. The dimensions were again selected from literature [3]. The circular area
depicted the area directly under the reaction shaft and was used as a mass flow inlet
and as a particle injection. One side surface of the geometry was set as a outflow.
The geometry with slag stream is presented in Figure 36. Tests with the larger model
were also performed first with particle size of only 500 µm. After this particles with
diameter of 50, 100, 250 and 500 µm were tested.

Figure 36: Streamline of the slag flow in the geometry depicting section of the slag
layer under the reaction shaft.

A smaller partial furnace slag layer was created for the enlarged particles. Dimensions
of the model were 3000 mm x 3000 mm x 500 mm. The reaction shaft of the model had
the diameter of 1000 mm and it was centred on top of the layer. The tapping hole was
110 mm in diameter and was on side of the slag layer. The tapping hole was 50 mm
 48

above the slag-matte interface. Due to stability issues of the particles corresponding
to the smallest size, only particles depicting 500 µm and 250 µm particle sizes were
used. 150 000 particles were injected with the rate of 4000 particles per second
with equal distribution between particle sizes. Two simulations were made with this
model: one without flow and the other with 0.96 kg/s mass flow from the shaft to
the tapping hole. Used CFD time step was 5 · 10−4 s. The model is presented in
Figure 37.

Figure 37: Geometry of the small partial slag layer model.

 49

5 Results and discussion

Validation

The validation model with one droplet produced very good results as they were very
similar to the experimental test. The initial trajectory of the droplet was not directly
downwards, but the droplet quickly lost lateral velocity due to drag. This was caused
by too tight definition of the injection point in EDEM, which was corrected in later
models. However, the droplet hit the bottom of the container in 1.76 s and in 2.02
s with diameters of 4.5 mm and 5 mm, respectively. This corresponds well to 1.87
seconds measured from the video as the droplet size is estimated to be between 4.5
and 5 mm, and also, as in the model does not have to first break surface tension of
the oil but instead is submerged in the oil from the beginning.

The two droplet validation did not produce as good results. The falling time was
similar to the one droplet model, but unlike in the experimental part, the second
droplet did not manage to catch the first droplet during the descending. However,
with spherical drag law the second droplet did descend faster than the first droplet.
When coefficient of static friction in EDEM was reduced to 0.01 from default value
of 0.5, the droplet slid or rolled from top of the other droplet next to it, as can be
seen happening in the experimental video.

As a results from CFD-DEM coupled simulation correspond well to those from the
VOF model and experiment, CFD-DEM can be used to realistically simulate a droplet
settling. The two droplet test demonstrated that the drag flow caused by the first
droplet was not high enough as the second droplet did not catch the first droplet.
However, this is partly caused by measuring accuracy from the video. Also, initial
submersion of the droplets may cause some error. In the experiment the drag flow
was already formed in the area where the second droplet entered the oil. In the
CFD-DEM model the second droplet took longer time to get to the starting point.
Also, droplets in the experiment slightly deformed, which increased drag, but droplets
in the CFD-DEM model remained spherical due to limitations of EDEM.

Particle scale-up factor provided reliable results with the water droplet tests. However,
with large PSUF, the error could grow too large as the 8 times scaled droplet had
almost 5 % error in the settling time. With the matte droplets the PSUF caused
systematic error of roughly 21 %. Descension times and their corresponding error-%
are presented in Table 9 for both water and matte droplets. The error is calculated
from unscaled (PSUF = 1) model.

High error with the matte droplets suggests that the validation test should be done if
the particle scaling is needed. However, the scaling cannot be decrease the simulation
time in the CFD as the drag forces are calculated with the true sized particles.
 50

Table 9: Results and errors from PSUF validation models.

Water Matte
PSUF t (s) error-% t (s) error-%
1 2.02 - 18.00 -
2 1.99 1.49 14.26 20.78
4 2.02 0.00 14.24 20.89
8 1.92 4.95 14.11 21.61

A single enlarged droplet produced falling times similar to values calculated with
Stokes’ law, as can be seen in Figure 38. However, the limits of calculational accuracy
can be seen in the droplets depicting 100 µm droplet. The velocity curves of the
droplets are not smooth but rather form steps. The 1 mm approximation of the 100
µm droplet was also unstable with the larger CFD time step. The enlarged droplets
may also trap other droplets more easily as with the scaled-up droplets.

Figure 38: Velocities of the enlarged droplets as a function of time with the different
CFD time steps. Named sizes are droplet size depicted by the enlarged droplets.
Droplets have diameter of 2 mm, except 100 µm small which has diameter of 1 mm.
 51

Cubic model

The cube model did not produce very realistic results. As can be seen in Figure 39,
without the droplets the slag stream was even, but when droplets were introduced to
the system, a vortex formed in the slag. The first 500 µm droplets flowed through the
slag in 3.4 seconds having average velocity of roughly 147 mm/s. This is unrealistic
as the vertical flow was roughly 10.5 mm/s and according to Stokes’ law the terminal
velocity of 500 µm droplet in the slag is only 0.44 mm/s.

However, the unrealistic particle and slag flows could be caused by mass-imbalance
as volume fraction taken by the droplets grows quickly in the small model. It could
also be caused by faulty coupling algorithm as the particle properties were scaled in
FLUENT but were not scaled back properly when returning the values to EDEM [77].
Too short droplet settling times indicates that FLUENT scales the droplets down
and calculates the created drag, but the values would not be scaled back up. This
would create escalating error which could create too high turbulence in FLUENT.
This in turn could possibly cause the whirl visible in Figure 39.

Figure 39: Streamlines of slag flow in the cube model. Left side stream is without
droplets and right side is after the droplets have bypassed through the slag layer.

Partial slag layer

Despite too fast settling velocities the partial slag layer with 4 different droplet sizes
was simulated with PSUF = 4. This was done as significant difference in droplet
size could affect droplet collisions. As can be seen in Figure 40, which is taken from
edge of the reaction shaft area, larger droplets descend through the slag relatively
 52

quickly and descend relatively straight downwards. The smaller droplets are, however,
significantly affected by turbulence caused by the larger droplets. Especially the
smallest droplets have relatively large change in their velocity vectors.

Figure 40: Velocity vectors of the scaled-up droplets in slag. The turbulence caused
by the drag of the larger droplets has an significant effect on direction of the smaller
droplets. Black (the shortest arrow), red (medium length arrow), blue (the longest
arrow) and green particles correspond to 50 µm, 100 µm, 250 µm and 500 µm
droplets, respectively.
 53

The effect of the turbulence is most likely even more significant as droplets in the
simulation had too high settling velocity. Also, in operational FSF the droplets
could coalesce which would increase their settling velocity and drag caused by them
further increasing the effect of the turbulence on smaller droplets. As the droplet
size decreases, effect of the turbulence increases. This also was supported by known
copper loss mechanisms of the smallest droplets getting trapped in the slag. Droplet
size in the slag can be much smaller than 100 µm studied in this simulation, as
presented by other study in Figure 41 [5].

In addition to turbulence moving the smaller droplets they were also pushed aside
as they collide with faster descending larger droplets, as can be seen in Figures 42
and 43. The collisions between droplets may lead to coalescence. However, in this
work coalescence was not taken into account as only the viability of CFD-DEM in
liquid-liquid modelling in FSF settler simulations was studied without developing
more sophisticated simulation.

Also, as can be seen in Figures 42 and 43, collision produced realistic behaviour
despite the particles having very low bulk modulus. In Figure 42, the droplets were
moving at their settling velocities. As the larger droplet contacted the smaller one,
the smaller moved significantly more in horizontal direction than the larger. Similar
observations can be made in Figure 43 where the larger droplet was moving faster
than its settling velocity as the contact point was close to the surface of the slag
layer, and the velocity of the larger droplet had not yet decreased to settling velocity.
High velocity collision had enough energy to push the droplets apart even in the slag,
which has high viscosity presented in Table 6. Experimental water droplet collisions
used as a validation and comparison are presented in Figure 44.

Figure 41: Trapped copper droplets in slag. [5]

 54

Figure 42: Slow velocity collision of 250 µm droplet and 100 µm droplets. Tip of the
contact vector is shown with red. Droplets at their settling velocity.

Figure 43: High velocity collision between 500 µm and 250 µm droplets. Velocity of
the 500 µm droplet has not yet decreased to settling velocity.

The low and high velocity collisions were also in good agreement with the literature [46].
As the impact velocity in the low velocity collision was below critical, the droplets
stuck together until the larger one had bypassed the smaller droplet and they were
 55

separated. In the high velocity impact, the impact velocity was significantly higher
and the droplets bounced apart.

Figure 44: Water droplets descending in oil and colliding to each other and with
other droplets on the bottom of the container without coalescing. The orange and
blue arrows show the first and the second droplet descending, respectively.

The number of collisions grew as new particles entered the slag layer. However, the
growth turned relatively linearly as more particles enter the slag. As different sized
droplets would spread through the slag the growth would most likely stop and fairly
even number of collisions would occur on any given time. Frequency distribution as
a function of time and cumulative distribution as a function of particles in the slag
are presented in Figure 45.
 56

Figure 45: Frequency and cumulative distributions of collisions between droplets.

 57

Small partial slag layer with enlarged droplets

The enlarged droplets were simulated in the smaller partial furnace. In the model
without the slag flow, the settling rate was close to values calculated with Stokes’ law.
This was in agreement with previous study, where settling in FSF was simulated [3].
However, the highest peak in the velocity distribution curves was 7 · 10−7 m/s lower
than the calculational values. This is practically insignificant difference, but still
shows that the majority of the droplets were slower than the calculated velocities.
However, this can be explained by particles colliding with each other due to the drag
and initial velocity differences. Collisions in the lower parts of the droplet layer could
cause chain of collisions, which would decrease velocity of the colliding particles. The
velocity distributions are presented in Figures 46 and 47.

From Figure 46 large difference between velocities of different droplet sizes can
be observed. This can be explained by smaller and, thus, lighter droplets having
significantly lower settling velocity compared to the larger ones. Velocities had spread
of 2.9·10−5 m/s and 2.3·10−4 m/s for 500 µm and 250 µm droplets, respectively.

Figure 46: Droplet velocity distribution in the model without the slag flow.
 58

Figure 47: Droplet velocity distributions in the slag without flow for droplets sepa-
rately. Velocities calculated with Stokes’ law are shown with arrows.

500 µm droplets had faster settling rate which caused different sizes to separate
into layers descending with different velocities. However, some of the smaller ones
were trapped by the larger droplets and vice versa, which created mixed interface
between the layers. In operational FSF this interface would not exist, however, as
droplets would enter the system continuously and in the system they would have
 59

much larger size distribution. This would create one thick layer which would consist
of droplets of all sizes. The velocity distribution presented in Figure 46 shows two
distinct curves with no overlapping. This is in good agreement with the interface
presented in Figure 48 as at 60 seconds the mixing of the interface is very low and
the droplet sizes have almost completely separated.

Figure 48: Interface between two droplet layers at times of 15, 30, 45 and 60 seconds
without the slag flow.

Number of collisions grew quickly as the droplets were injected in the slag layer.
When the injection was stopped collision frequency continued to increase for few
seconds. After that, the collision frequency started to decrease. This delay was
caused by the last larger droplets colliding frequently with the smaller droplets until
most of the smaller ones were bypassed. The highest collision frequency was at 42
seconds, as can be seen in Figure 49. Stable collision frequency would most likely be
 60

reached in continuous process.

As can be seen in Figure 50, the particles either collided and bounced apart, or
pushed each other aside as the larger droplets bypassed by the smaller ones. The
collisions were similar to previous simulations. This is also in good agreement with
the literature [46], similarly to the cube model collisions.

Figure 49: Frequency and cumulative distributions of collisions as a function of time

with slag flow.
 61

Figure 50: High and slow speed collisions in the model without the slag flow. In
high speed collision the second droplet was injected into the model, which caused the
collision. The second droplet was not in the model in previous time step and, thus,
is not presented in the Figure. The arrows point direction of movement experienced
by the particle, red arrows point direction of the 250 µm droplets and blue arrows
show direction of 500 µm droplets. Observed droplets have been numbered.

The enlarged particle model with two particles depicting different droplets was also
simulated with 0.96 kg/s slag mass flow. Droplets behaved very similarly to the
model without the slag flow, when particle layering was considered. The particles
were mixed during injection period and started to separate after injection period as
particles depicting 500 µm droplets descended faster. The interface between layers is
presented in Figure 51.
 62

Figure 51: Interface between the two droplet layers at times of 15, 30, 45 and 60
seconds with slag the flow.

As can be seen in Figure 52, the interfaces were similar. However, the interface
between the layers in the model with the slag flow had descended more and spread
more than in the model without slag flow. Faster descending was purely caused by
the slag flow moving both droplet layers. The higher spread was caused by the higher
velocity distribution, which is discussed later in this work. Also, the interfaces of the
layers were significantly more mixed in the model with the slag flow.
 63

Figure 52: Comparison of the droplet layer interface at 30 and 60 seconds. The
model without the slag flow on the left and with the slag flow on the right.

In velocity distribution curves presented in Figures 53 and 54, main peaks can be
seen that are in similar places compared to the model without flow. However, much
smaller secondary peaks can also be seen on both sides of the main peaks. The
secondary peaks were most likely caused by droplet collisions. The main peaks are
significantly taller than those in the model without flow. The peaks in the model
with the slag flow have tens of thousands of droplets, while the model without the
slag flow had only thousands in the highest peaks. As both models have practically
identical number of droplets, it can be concluded that the model with the slag flow
 64

has the majority of the droplets with much narrower velocity distribution compared
to the model without the slag flow. Due to this narrower distribution, the secondary
peaks can be seen only in the model with the slag flow.

The velocity distribution was much larger when the slag was flowing. The velocities
of the droplets had the spread of 2.4 · 10−4 m/s and 2.3 · 10−4 m/s for 500 µm and
250 µm droplets, respectively. However, the spread also takes the mixed interface
into account, which could be seen as lines connecting curves of the two droplet sizes
in Figure 53. The connecting lines showed droplet velocities having relatively smooth
distribution in the mixed interface. When only the distinguishable peaks were taken
into account, the different sizes had 2.2 · 10−4 m/s and 2.1 · 10−4 m/s for 500 µm and
250 µm droplets, respectively. The peaks were considered distinguishable when over
10 droplets were found within velocity spread of 5.4 · 10−6 m/s. The velocity spread
of the distinguishable peaks were lower than the velocity ranges of the droplets when
the slag was not flowing through the model.

The flow velocity of the slag in the shaft area was 3.9 · 10−4 m/s. Velocities of
8.3 · 10−4 m/s in 500 µm and 5.0 · 10−4 m/s in 250 µm curves are shown with arrows
in the Figure 53. These are combined values of the settling velocities calculated with
Stokes’ law and the slag flow velocity in the shaft area. As in the model without
the slag flow, the majority of the droplets were moving slower than the calculated
velocity. However, the difference was slightly larger at 1.8 · 10−6 m/s and 1.6 · 10−6
m/s for the larger and the smaller droplet, respectively. This difference was caused
by collisions between droplets similarly to the model without the slag flow.

Figure 53: Droplet velocity distribution in the simulation with the slag flow.
 65

Figure 54: Droplet velocity distributions in the model with the slag flow for droplets
separately. The arrows are pointing to the calculated values.

As the droplets were injected in the slag layer, number of collisions grew quickly. The
number of collisions started to decrease after the droplet injection stopped. However,
the collision frequency started to decrease with a delay as the last of the larger
droplets were colliding frequently with the smaller droplets. The highest collision
frequency was at 43 seconds, as can be seen in Figure 55. In continuous process
the collisions would at some point reach a stable rate. This is very similar to model
without the slag flow. Comparing to the model without the slag flow, number of the
 66

collisions decreased at slower rate. This was caused by larger spread of the layers, as
the larger droplets had longer distance in which the collisions could occur.

Figure 55: Frequency and cumulative distributions of collisions as a function of time

with slag flow.

As can be seen in Figure 56, multiple collisions can occur in the small area and
relatively short time period. In this work the droplets push each other when colliding,
but taking coalescence into account could change the situation significantly as merging
droplets would form larger droplet which would have even higher settling velocity.
As in the model without the slag flow and the cube model, the low velocity collisions
behaved similarly as reported in the literature [46].
 67

Figure 56: Collision of several droplets observed directly above the system. Arrows
point direction of movement experienced by the particle, red arrows point direction
of the 250 µm droplets and blue arrows show direction of 500 µm droplets. Observed
droplets have been numbered.
 68

The collisions were very similar in every model, but also in the validation experiment.
High velocity collisions caused droplets to move sideways quickly which caused the
droplets to bounce in opposing directions. In slow speed collisions, the particles
pushed each other aside until the larger particle could bypass the smaller and slower
one. When observed directly above the droplets, they seemed to stay in tangential
contact unless another collision occurred.

Computational performance

Small domains were used in the single droplet simulations. This combined with
only one particle and short times to be simulated meant that the simulations were
relatively fast. Calculation times for these were in range of 2 – 8 hours. Simulations
with larger systems and particle counts were significantly slower. Also, adding more
complexity, such as flow, caused even further increase in required calculation times.
These models had calculation times of several days. For example, smaller partial slag
layer model required roughly 20 days when simulated with flow and slightly under 5
days without flow with 4 core Intel Xeon with hyper-threading at 3.4 GHz. Some
programs may work better without hyper-threading but according to company’s own
performance tests, EDEM will benefit of hyper-threading.

The time steps used in CFD ranged in 10−4 s – 10−3 s, while the time steps in
DEM were in range of 10−6 s – 10−5 s. These are similar to those used in other
studies [38–40, 44]. However, one team managed to use as high as 7.2 s time step in
CFD, while time step in DEM was 10−4 s [4,36,37]. This could be probably explained
by relatively stationary particles. The diameter of the coke and ore particles used
in the simulation were 3.5 cm and 2 cm in CFD, DEM had corresponding particle
sizes of 15 cm and 7.5 cm. As the gas was injected to the blast furnace from tuyeres,
where particles are held in place with large mass above them, they cannot be moved
significantly by the gas flow. This could allow significantly larger time step size to
be used as EDEM advices to reduce CFD time step size if the change in the DEM
simulation is large during one CFD time step [80].

In validation models the residuals in FLUENT started converging quickly, in a

few iterations, the majority of the calculation time was consumed by saving and
transferring data between EDEM and FLUENT as the actual calculations in both
software seemed instantaneous for naked eye. However, with larger systems the
majority of the calculation time was consumed by FLUENT as a large number of
iterations was needed to calculate each time step. However, the calculation speed
increased when all particles were created as the residuals started converging faster
until only a couple of iterations were needed to achieve the convergence.

In future studies, the cube model should be further studied as a mass-imbalance

may have been the reason for a large change in the slag flow after the droplets were
 69

injected. Also, effects of new, corrected EDEM-FLUENT coupling must be studied

as it may also have an effect on the slag and particle flow as forces returned to EDEM
were not correct. Too fast moving particles could cause the slag to move too fast
in the model due to unrealistically large movement. This effect may multiply as
FLUENT calculated with downscaled values, but the values were not scaled back up
for EDEM.

Future studies should consider adding a larger particle injection area with realistic
particle size gradient with the areas as gas flows in FSF move the smallest particle
closer to tapping hole. This could make significant difference to copper losses during
slag tapping. Also, particle amount should be adjusted to better correspond to the
mass-flow in a real FSF.

Methods to increase the CFD time step should be considered in the future studies.
This would make longer simulations more viable and allow more detailed models
to be simulated. One possible method would be utilizing GPGPU in calculations.
Parallelizing the calculations should be considered as one method of increasing
calculation speed. Also, memory optimization should be taken into account as the
software need to repeatedly read and write data to memory. The software also needs
to write data to a drive, which is significantly slower. This should be minimized as it
cannot be eliminated. The software save data to the drive and the data is used to
later observe the simulation at time step specific to the saved file. If all writing were
to be eliminated, only the end result could be observed.

Droplet coalescence in the simulations should also be studied as the droplet size has
significant effect on the settling velocity and, thus, on the copper loss mechanisms.
The coalescence studies should be the next subject due to a large effect on the matte
flow. With API, some factor could be used that would merge colliding droplets if
right conditions are met. This could have a large effect on settling of especially
smaller matte droplets as they may not settle through the slag, but may start to
settle if the droplets coalesce. Coalescence could reduce copper losses to the slag as
it could decrease the number of the smallest droplets in the slag and increase settling
velocity of the droplets.
 70

6 Summary and conclusions

Computer simulations have become an important tool in many different fields of

science. For example, CFD and CFD-DEM coupling have been used in the studies
of pyrometallurgical processes and phenomena. However, found CFD-DEM studies
were investigating solid-fluid flows, but the focus of this work was in liquid-liquid
flows. Thus, the viability of the CFD-DEM coupling in simulating the matte flow
through the slag layer was investigated.

The CFD-DEM coupling produced reasonable matte droplet settling velocities with
and without the slag flow. The settling velocities of the majority of droplets were
either close to calculated values or differences could be explained by collisions or
drag. However, in too small model, the particle and slag flow were unrealistically fast
which could be due to incorrect values returned to EDEM by used EDEM-FLUENT
coupling.

Also, collisions of the particles produced realistic results despite very low bulk
modulus. The high velocity collisions caused the droplets to bounce in opposing
directions, while in the low velocity collisions the droplets remained close to each
other. The collision behaviour was in good agreement with the literature. When
different kinds of droplets were injected in the simulation, the droplets separated to
different layers due to differing settling velocities. However, interface of the layers
was mixed due to velocity distribution of the droplets. As the droplets collided and
created turbulence in the slag, part of the droplets had increased velocities, while
other ones had decreased velocities.

As forces and movement were calculated for every individual droplet, the slag flow
had to be calculated and the data had to be transferred between the software, the
simulations were slow. The simulation speed was also affected by very small time
steps, for example 10−4 s for CFD and 10−6 s for DEM, was needed to keep the
simulation stable. Similar time step sizes have also been used in previous studies. On
relatively common processor configuration of 4 hyper-threaded cores, the simulations
took up to few weeks. However, the calculation times should be improvable by
paralellizing the calculation process, decreasing the particle count and acquiring more
powerful computer hardware. Also, the memory utilization should be optimized and
data writing to drive should be kept at minimum. However, writing to drive cannot
be completely eliminated as frequently saved data is needed in order to get data
during the simulation, not just the final results.

All things considered, the CFD-DEM coupling shows promising results as a tool for
simulating droplets settling in the flash smelting settler. Adding a variety of droplet
sizes shows also other interesting phenomena, such as droplets changing directions
due to collisions and drag flow caused by larger droplets passing by.
 71

In addition of getting reliable results, visualization of the results is also important.

Both EDEM and FLUENT have various methods to visualize them. The ocation
of droplets or slag streamline can be easily imaged every saved time step. Both are
also capable of producing video from which changes, such as collisions, are very easy
to observe. Both software can also be used to extract statistical information, such as
velocities at any saved time step.

Much more sophisticated model could be created with additional user-defined func-
tions depicting physical and other phenomena, such as coalescence and reactions.
Adding these functions to the simulation should be considered and investigated in
future studies. Coalescence affects settling velocities of the droplets and reactions
affect composition and, thus, density of the matte.

Computer based process simulation has been growing in importance parallel with the
increase in computing power. Demand for realistic and real-time simulations of the
complicated processes or phenomena does not diminish. In the future CFD-DEM
could be used to quickly model conditions in the flash furnace. However, in addition of
more sophisticated simulation, that would also require advancements in the available
computing power. With a realistic model considering also reactions and coalescence,
CFD-DEM could be used to simulate copper losses. This would be useful in the slag
cleaning optimization.
 72

References

[1] M. Kennedy, A. MacRae, R. Jones, L. Kolbeinsen, P. Nos, and A. Filzwieser,

“Some considerations for safer furnace cooling,” Com 2015, vol. 53, no. 9, pp.
1689–1699, 2013.

[2] M. White, R. J. Haywood, W. Taylor, S. Chen, J. Ranasinghe, and R. West,

“Improved Copper Flash Smelting at Olympic Dam,” no. April, 2015.

[3] J. Xia, T. Ahokainen, T. Kankaanpää, J. Järvi, and P. Taskinen, “Numerical

modelling of copper droplet setting behavior in the settler of a flash smelting
furnace,” Proceedings - European Metallurgical Conference, EMC 2005, vol. 3,
no. January, 2005.

[4] S. Natsui, H. Nogami, S. Ueda, J. Kano, R. Inoue, and T. Ariyama, “Simultane-

ous three-dimensional analysis of gas-solid flow in blast furnace by combining
discrete element method and computational fluid dynamics,” ISIJ International,
vol. 51, no. 1, pp. 41–50, 2011.

[5] E. De Wilde, I. Bellemans, M. Campforts, M. Guo, B. Blanpain, N. Moelans,

and K. Verbeken, “Investigation of High-Temperature Slag/Copper/Spinel In-
teractions,” Metallurgical and Materials Transactions B: Process Metallurgy
and Materials Processing Science, vol. 47, no. 6, pp. 3421–3434, 2016.

[6] B. Andersson, R. Andersson, L. Hakansson, M. Mortensen, R. Sudiyo,

and B. van Wachem, Computational Fluid Dynamics for En-
gineers. Cambridge University Press, 2012. [Online]. Available:
http://ingegneriaterni.altervista.org/wp-content/uploads/2016/06/BOOK-
Bengt-Andersson-et-al-Computational-fluid-dynamics-for-engineers-2012.pdf

[7] C. Kloss, C. Goniva, G. Aichinger, and S. Pirker, “Comprehensive

DEM-DPM-CFD Simulations-Model Synthesis, Experimental Validation and
Scalability,” Seventh International Conference on CFD in the Minerals
and Process Industries, no. December, pp. 1–7, 2009. [Online]. Available:
http://www.cfd.com.au/cfd{_}conf09/PDFs/090KLO.pdf

[8] “Frequently Asked Questions.” [Online]. Available: https://www.edemsimulation.

com/resources-learning/faqs/

[9] “EDEM - Software Partner | ANSYS.” [Online]. Available: http:

//www.ansys.com/About-ANSYS/partner-ecosystem/software-partners/edem

[10] I. V. Kojo, A. Jokilaakso, and P. Hanniala, “Flash smelting and converting

furnaces: a 50 year retrospect,” JOM, vol. 52, no. 2, pp. 57–61, 2000.
 73

[11] L. Bergh, I. Cornejo, F. Romero, and C. Sulzer, “Improving copper matte

grade control in a concentrate flash furnace,” IFAC Proceedings Volumes (IFAC-
PapersOnline), pp. 13–18, 2012.

[12] M. G. King, “The evolution of technology for extractive metallurgy over the
last 50 years—Is the best yet to come?” JOM, vol. 59, no. 2, pp. 21–27,
2007. [Online]. Available: http://www.tms.org/pubs/journals/jom/0702/king-
0702.html{%}5Cnhttp://link.springer.com/10.1007/s11837-007-0019-2

[13] “Background report AP-42 section 12.3 primary copper smelting,” Tech. Rep.
[Online]. Available: https://www3.epa.gov/ttnchie1/ap42/ch12/bgdocs/b12s03.
pdf

[14] B. Gorai, R. K. Jana, and Premchand, “Characteristics and utilisation of copper

slag - A review,” Resources, Conservation and Recycling, vol. 39, no. 4, pp.
299–313, 2003.

[15] R. W. Matusewicz, P. M. A. Reuter, and S. P. Hughes, “Large Scale Copper

Smelting Using Ausmelt TSL Technology at the Tongling Jinchang Smelter,”
no. May 1953, pp. 961–970, 2010.

[16] W. Davenport, “Copper extraction from the 60’s into the 21st century,” Copper
99-Cobre 99, vol. I, pp. 55–79, 1999.

[17] Willie Scott, “Smelting Copper from Ore,” 2011. [Online]. Avail-
able: https://www.brighthubengineering.com/manufacturing-technology/68414-
smelting-copper-from-ore/

[18] I. V. Kojo and H. Storch, “Copper Production With Outokumpu Flash Smelting:
An Update,” Symposium A Quarterly Journal In Modern Foreign Literatures,
vol. 8, no. January 2006, pp. 225–238, 2006.

[19] P. Coursol, N. C. Valencia, P. Mackey, S. Bell, and B. Davis, “Minimization of

copper losses in copper smelting slag during electric furnace treatment,” JOM,
vol. 64, no. 11, pp. 1305–1313, 2012.

[20] “fsfsettler.” [Online]. Available: https://www.outotec.com/products/smelting-

and-converting/flash-furnace/

[21] “reversettler.” [Online]. Available: www.coppercountryexplorer.com/2009/11/

smelter-tech-the-reverberatory-furnace/

[22] W. G. Davenport and E. H. Partelpoeg, Flash Smelting: Analysis,

Control and Optimization. Elsevier Science, 2015. [Online]. Available:
https://books.google.fi/books?id=cMcgBQAAQBAJ
 74

[23] D. Goldsman, R. E. Nance, and J. R. Wilson, “A brief history of simulation,”

in Winter Simulation Conference, 2009.

[24] “History of simulation.” [Online]. Available: http://www.landersimulation.com/

eng/training-with-simulation/the-world-in-motion/history-of-simulation/

[25] “Computer simulation.” [Online]. Available: https://en.wikipedia.org/wiki/

Computer{_}simulation{#}History

[26] J. Tu, G. H. Yeoh, and C. Liu, Computational Fluid Dynamics:

A Practical Approach. Elsevier Science, 2007. [Online]. Available:
https://books.google.fi/books?id=BJwuMQOhM0sC

[27] National Institute of Biomedical Imaging and Bioengineer-

ing, “Computational Modeling,” no. September, 2016. [On-
line]. Available: https://www.nibib.nih.gov/sites/default/files/
Computational{_}Modeling{_}Fact{_}Sheet{_}ENGLISH{_}v4{_}508.
pdf

[28] R. Wood, “Modeling Fluid-Particle Systems using EDEM with CFD,” 2017.
[Online]. Available: https://www.edemsimulation.com/webinar/modeling-fluid-
particle-systems-using-edem-cfd/

[29] “CAE Integration.” [Online]. Available: https://www.edemsimulation.com/

software/cae-integration/

[30] “World first in Computer Aided Engineering - New tool from EDEM
brings Discrete Element Modelling to every engineer,” 2017. [Online].
Available: https://www.edemsimulation.com/news/world-first-computer-aided-
engineering-new-tool-edem-brings-discrete-element-modelling-every-engineer/

[31] E. Mestres, “Coupling simulations: A simple way to increase results accuracy,”

2012. [Online]. Available: http://insider.altairhyperworks.com/coupling-
simulations/

[32] M. Patel, “4 Advantages and Disadvantages of CFD,” 2013. [Online]. Avail-

able: https://cfdmodelingservices.wordpress.com/2013/06/12/4-advantages-
and-disadvantages-of-cfd/

[33] “Lecture 8 - User Defined Functions (UDFs),” Introduction to ANSYS FLUENT,

no. December, pp. 1–23, 2010.

[34] “Customization / API - EDEM Software.” [Online]. Available: https:

//www.edemsimulation.com/software/customization-api/
 75

[35] S. Wasserman, “Turbulence Models Offered by CFD Simulation Vendors.”

[Online]. Available: https://www.engineering.com/ResourceMain.aspx?resid=
484

[36] S. Natsui, S. Ueda, H. Nogami, J. Kano, R. Inoue, and T. Ariyama, “Dynamic

Analysis of Gas and Solid Flows in Blast Furnace with Shaft Gas Injection by
Hybrid Model of DEM-CFD,” ISIJ International, vol. 51, no. 9, pp. 51–58,
2011.

[37] ——, “Analysis on non-uniform gas flow in blast furnace based on DEM-CFD
combined model,” Steel Research International, vol. 82, no. 8, pp. 964–971,
2011.

[38] J. Wei, H. Zhang, Y. Wang, Z. Wen, B. Yao, and J. Dong, “The gas-solid flow
characteristics of cyclones,” Powder Technology, vol. 308, pp. 178–192, 2017.

[39] A. Di Renzo, F. Cello, and F. P. Di Maio, “Simulation of the layer inversion

phenomenon in binary liquid–fluidized beds by DEM-CFD with a drag law for
polydisperse systems,” Chemical Engineering Science, vol. 66, no. 13, pp. 2945–
2958, 2011. [Online]. Available: http://dx.doi.org/10.1016/j.ces.2011.03.035

[40] A. Di Renzo and F. P. Di Maio, “Homogeneous and bubbling fluidization

regimes in DEM-CFD simulations: Hydrodynamic stability of gas and liquid
fluidized beds,” Chemical Engineering Science, vol. 62, no. 1-2, pp. 116–130,
2007.

[41] M. Börner, A. Bück, and E. Tsotsas, “DEM-CFD investigation of particle

residence time distribution in top-spray fluidised bed granulation,” Chemical
Engineering Science, vol. 161, pp. 187–197, 2017.

[42] “CFD-DEM model of particle melting.” [Online]. Available: https://www.dcs-

computing.com/cfd-dem-model-particle-melting

[43] Q. Hou, E. Dianyu, S. Kuang, Z. Li, and A. B. Yu, “DEM-based virtual

experimental blast furnace: A quasi-steady state model,” Powder Technology,
vol. 314, pp. 557–566, 2017.

[44] W. J. Yang, Z. Y. Zhou, A. B. Yu, and D. Pinson, “Particle scale simulation

of softening-melting behaviour of multiple layers of particles in a blast furnace
cohesive zone,” Powder Technology, vol. 279, pp. 134–145, 2015.

[45] M. Baniasadi, M. Baniasadi, and B. Peters, “Coupled CFD-DEM with heat

and mass transfer to investigate the melting of a granular packed bed,”
Chemical Engineering Science, vol. 178, pp. 136–145, 2018. [Online]. Available:
http://linkinghub.elsevier.com/retrieve/pii/S0009250917307819
 76

[46] L. Fries, S. Antonyuk, S. Heinrich, D. Dopfer, and S. Palzer, “Collision dynamics

in fluidised bed granulators: A DEM-CFD study,” Chemical Engineering Science,
vol. 86, pp. 108–123, 2013.

[47] “ANSYS Fluent Software: CFD Simulation.” [Online]. Available: http:

//www.ansys.com/products/fluids/ansys-fluent

[48] B. Ray, R. Bhaskaran, and L. R. Collins, “Introduction to

CFD Basics,” Cornell Online Course, 2012. [Online]. Avail-
able: https://www.academia.edu/attachments/45468154/download{_}file?st=
MTUyODY5Nzk4OSwxMzAuMjMzLjEwOC4xMzQsODI4MTYwODM{%}3D{&}s=
swp-toolbar{&}ct=MTUyODY5ODA1MywxNTI4Njk4MTAwLDgyODE2MDgz

[49] A. Bakker, “Lecture 7 - Meshing,” Presentation, no. 2002, pp. 1–35, 2006.

[50] J. Kortelainen, “Meshing Tools for Open Source CFD – A Practical Point of
View,” Report, p. 25, 2009.

[51] K. Verscheure, A. K. Kyllo, A. Filzwieser, B. Blanpain, and P. Wollants, “Fur-

nace Cooling Technology in Pyrometallurgical Processes,” Sohn International
Symposium Advanced processing of metals and materials volume 4 - New, im-
proved and existing technologies: Non-ferrous materials extraction and processing,
vol. 4, no. January, pp. 139–154, 2006.

[52] “Plate Coolers for Blast Furnace.” [Online]. Avail-

able: http://fundicaoeusinagem.com/high-conductivity-copper-cooled-
components-for-steel-industry/plate-coolers-for-blast-furnace/

[53] J. Jansson, L. Pesonen, and I. Vaajamo, “Ensuring the furnace integrity with
intelligent operations, efficient cooling and refined design,” Com 2015, pp. 1–12,
2015.

[54] A. Bakker, “Lecture 16 - Free Surface Flows Applied Computational Fluid

Dynamics Example : spinning bowl,” Fluid Dynamics, no. 2002, pp. 1–27, 2006.

[55] “Lagrangian and Eulerian Reference Frames,” 2015. [Online]. Avail-

able: https://www.sharcnet.ca/Software/Ansys/16.2.3/en-us/help/wb{_}sim/
exp{_}dyn{_}theory{_}lag{_}eul{_}360.html

[56] “AcuSolve⃝R Computational Fluid Dynamics.” [Online]. Avail-

able: http://www.altairhyperworks.com/ResourceLibrary.aspx?keywords=
2017{&}altair{_}products=AcuSolve{&}partner{_}products={&}category=
Brochures

[57] “Rocky Discrete Element Method Package.” [Online]. Available: http://rocky-

 77

dem.com/images/pdf/Rocky{_}DEMP.pdf

[58] “DEM simulation of hopper/silo flow - non-sphercial particles.” [Online].

Available: https://www.dcs-computing.com/dem-simulation-hoppersilo-flow-
non-sphercial-particles

[59] S. Cole, “It’s about time,” 2015. [Online]. Available: https://www.

edemsimulation.com/blog/its-about-time/

[60] B. Peng, “Discrete Element Method (DEM) Contact Models Applied

to Pavement Simulation,” Master Thesis, 2014. [Online]. Available:
https://vtechworks.lib.vt.edu/handle/10919/50399

[61] P. Cleary and M. L. Sawley, “DEM modelling of industrial granular flows:3D case
studies and the effect of particle shape on hopper discharge,” Appl.Math.Model.,
vol. 26, pp. 89–111, 2002.

[62] C. B. Padros, “Making contact – Using EDEM API to write custom contact
models,” 2015. [Online]. Available: https://www.edemsimulation.com/blog/
making-contact-using-edem-api-to-write-custom-contact-models/

[63] M. Sousani, “What is DEM – an Introduction to the Discrete Element Method,”

2017. [Online]. Available: https://www.edemsimulation.com/webinar/dem-
introduction-discrete-element-method/

[64] “Realistic particle shapes.” [Online]. Available: http://rocky-dem.com/index.

php?pg=particle{_}shapes

[65] D. Mann, “STAR-CCM+ v12.04 preview: Model realistic par-

ticle shapes with polyhedral DEM particles.” [Online]. Avail-
able: https://mdx2.plm.automation.siemens.com/blog/david-mann/star-ccm-
v1204-preview-model-realistic-particle-shapes-polyhedral-dem-particles

[66] “Itasca Consulting Group.” [Online]. Available: https://www.itascacg.com/

software/pfc

[67] “CFDEM⃝workbench
R for LIGGGHTS⃝
R DEM CFDEM⃝,”
R no. May, pp. 0–4,
2017.

[68] “Why we are unique.” [Online]. Available: https://www.dcs-computing.com/

why-we-are-unique

[69] “CFDEM⃝coupling
R Open Source CFD-DEM Framework.” [Online]. Available:
https://www.cfdem.com/cfdemrcoupling-open-source-cfd-dem-framework
 78

[70] C. Goniva, C. Kloss, A. Hager, and S. Pirker, “An open source CFD-DEM
perspective,” Proceedings of OpenFOAM workshop, pp. 1–10, 2010.
[Online]. Available: ftp://ftp.heanet.ie/disk1/sourceforge/o/op/openfoam-
extend/OpenFOAM{_}Workshops/OFW5{_}2010{_}Gothenburg/Papers/
ChristophGonivaPaperOFWS5.pdf

[71] “KÄYTTÖTURVALLISUUSTIEDOTE WOODEX WOOD OIL,” pp. 1–16,

2017. [Online]. Available: https://www.teknos.com/globalassets/inriver/
resources/woodex-wood-oil{_}9471{_}fi{_}fi{_}17363.pdf

[72] H. Kim and D. J. Burgess, “Prediction of interfacial tension between oil mixtures
and water,” Journal of Colloid and Interface Science, vol. 241, no. 2, pp. 509–
513, 2001.

[73] W. L. Hosch, “Hooke’s Law,” Encylopeadia Britanica, 2006. [Online]. Available:

http://www.cns.gatech.edu/PHYS-4421/lautrup/2.8/elasticity.pdf{%}0Ahttp:
//www.cns.gatech.edu/PHYS-4421/lautrup/2.8/elasticity.pdf{%}5Cnhttps:
//www.britannica.com/science/Hookes-law

[74] “Bulk modulus of elasticity of liquids.” [Online]. Available: https:

//neutrium.net/properties/bulk-modulus-of-elasticity-of-liquids/

[75] Y. J. Yoon and S. C. Cowin, “The elastic moduli estimation of the solid-water
mixture,” International Journal of Solids and Structures, vol. 46, no. 3-4, pp. 527–
533, 2009. [Online]. Available: http://dx.doi.org/10.1016/j.ijsolstr.2008.09.010

[76] D. K. Chibwe, G. Akdogan, C. Aldrich, and R. H. Eric, “CFD Modelling of

Global Mixing Parameters in a Peirce-Smith Converter with Comparison to
Physical Modelling,” Chemical Product and Process Modeling, vol. 6, no. 1, 2011.
[Online]. Available: https://www.degruyter.com/view/j/cppm.2011.6.issue-
1/cppm.2011.6.1.1584/cppm.2011.6.1.1584.xml

[77] R. Wood, “Correction to scale factor implementation,” 2018. [Online]. Available:

https://www.edemsimulation.com/forum/topic/scale-factor-implementation/

[78] H. Lamb, “Hydrodynamics,” pp. 532–533, 1916. [Online]. Avail-

able: https://archive.org/stream/hydrodynamics00horarich{#}page/n555/
mode/2up/search/terminal

[79] C. B. Solnordal, F. R. Jorgensen, P. T. Koh, and A. Hunt, “CFD modelling of

the flow and reactions in the Olympic Dam flash furnace smelter reaction shaft,”
Applied Mathematical Modelling, vol. 30, no. 11, pp. 1310–1325, 2006.

[80] “Using the EDEM-Fluent Coupling Fluent Case,” pp. 1–17, 2016.