This action might not be possible to undo. Are you sure you want to continue?
GRANULAR DYNAMICS OF GASSOLID
TWOPHASE FLOWS
PROEFSCHRIFT
ter verkrijging van
de graad van doctor aan de Universiteit Twente,
op gezag van de rector magnificus,
prof.dr. F.A. van Vught,
volgens besluit van het College voor Promoties
in het openbaar te verdedigen
op vrijdag 21 januari 2000, te 16.45 uur.
door
Bob Petrus Bernardus Hoomans
geboren op 2 augustus 1971
te Oldenzaal.
ii
Dit proefschrift is goedgekeurd door de promotoren
prof.dr.ir. J.A.M. Kuipers
prof.dr. W.J. Briels
prof.dr.ir. W.P.M. van Swaaij
iii
iv
Samenstelling promotiecommissie:
prof.dr. W.E. van der Linden, voorzitter Universiteit Twente
prof.dr.ir. J.A.M. Kuipers, promotor Universiteit Twente
prof.dr. WJ. Briels, promotor Universiteit Twente
prof.dr.ir. W.P.M. van Swaaij, promotor Universiteit Twente
prof.dr. J.P.K. Seville University of Birmingham, UK
prof.dr.ing. B.H. Hjertager Aalborg University Esbjerg, DK
prof.ir. C.M. van den Bleek Technische Universiteit Delft
prof.dr.ir M.M.C.G. Warmoeskerken Universiteit Twente
prof.dr.ir. G.F. Versteeg Universiteit Twente
Cover: snapshots of a simulation with a ternary density distribution (Chapter 6)
© december 1999 Bob Hoomans, Maastricht, Nederland. All rights reserved.
Second impression august 2001
ISBN 903651401
Contents
______________________________________________________________________________________
v
Contents
Summary 1
Samenvatting 5
1. GENERAL INTRODUCTION
Abstract 9
1. Fluidisation 11
2. Hydrodynamic modelling 12
2.1 Multiscale modelling 13
3. Discrete particle modelling 16
4. Outline of this thesis 17
References 19
2. GRANULAR DYNAMICS
Abstract 23
1. Introduction 25
1.1 Hardparticle approaches 26
1.2 Softparticle approaches 28
1.3 Monte Carlo techniques 31
2. Hardsphere approach 32
2.1 Collision model 32
2.2 Key parameters of the collision model 38
2.3 Sequence of collisions 40
2.4 Optimisation 43
3. Softsphere approach 47
3.1 The linearspring/dashpot model 48
3.2 Model parameters 53
4. Hardsphere vs. Softsphere 56
4.1 Static situations 57
4.2 Spring stiffness 57
4.3 Energy considerations 59
4.4 Multiple particle interactions 60
5. Measurement of collision parameters 62
6. External forces 66
Notation 68
References 70
Contents
______________________________________________________________________________________
vi
3. GAS PHASE HYDRODYNAMICS
Abstract 79
1. Introduction 81
1.1 Direct solution of the NavierStokes equations 81
1.2 Lattice Boltzmann simulations 82
1.3 Dissipative Particle Dynamics 83
2. Governing equations 84
3. Constitutive equations 85
3.1 Gas phase density 85
3.2 Gas phase stress tensor 85
4. Numerical solution 86
5. Boundary conditions 87
6. Two way coupling 88
6.1 Void fraction 89
6.1.1 Calculation of the void fraction in 2D 89
6.1.2 Calculation of the void fraction in 3D 91
6.2 Momentum transfer 92
Notation 93
References 95
4. THE EFFECT OF PARTICLE PROPERTIES ON THE HYDRODYNAMICS
OF GASFLUIDISED BEDS WITH HOMOGENEOUS INFLOW
CONDITIONS
Abstract 99
1. Introduction 101
2. Model 104
2.1 Granular dynamics 104
2.2 External forces 105
2.3 Gas phase hydrodynamics 106
3. Effects of collision parameters 107
4. Energy conservation 113
5. Influence of a particle size distribution 116
6. 3D simulations 118
6.1 Influence of collision parameters 119
6.2 2D versus 3D 121
7. Conclusions 123
Notation 124
References 125
Contents
______________________________________________________________________________________
vii
5. GRANULAR DYNAMICS SIMULATION OF BUBBLE FORMATION IN A
GASFLUIDISED BED: HARDSPHERE VS. SOFTSPHERE APPROACH
Abstract 129
1. Introduction 131
2. Models 132
2.1 Hardsphere granular dynamics 132
2.2 Softsphere granular dynamics 133
2.3 External forces 134
2.4 Gas phase hydrodynamics 136
3. Preliminary simulations 137
4. Experimental validation 142
4.1 Experimental 142
4.2 Influence of a particle size distribution 144
4.3 Hardsphere vs. softsphere 146
4.4 Effects of collision parameters 149
5. Conclusions 151
Notation 153
References 154
6. GRANULAR DYNAMICS SIMULATION OF SEGREGATION
PHENOMENA IN BUBBLING GASFLUIDISED BEDS
Abstract 157
1. Introduction 159
2. Models 161
2.1 Hardsphere granular dynamics 161
2.2 Softsphere granular dynamics 162
2.3 External forces 164
2.4 Gas phase hydrodynamics 165
3. Ternary density distribution 166
4. Binary size distribution 169
4.1 Base case 169
4.2 Statistical analysis of segregation 172
4.3 Effects of collision parameters 174
5. Experimental validation 175
5.1 Experimental 175
5.2 Results 177
6. Conclusions 183
Notation 184
References 186
Contents
______________________________________________________________________________________
viii
7. THE INFLUENCE OF COLLISION PROPERTIES ON THE FLOW
STRUCTURE IN A RISER
Abstract 189
1. Introduction 191
2. Model 192
2.1 Granular dynamics 192
2.2 External forces 193
2.3 Inlet and outlet conditions 196
2.4 Gas phase hydrodynamics 196
3. Results 198
3.1 Effect of collision parameters 198
3.2 Axial effects 202
3.3 Influence of lift forces 206
4. Conclusions 207
Notation 208
References 209
8. EXPERIMENTAL VALIDATION OF GRANULAR DYNAMICS
SIMULATIONS OF GASFLUIDISED BEDS WITH HOMOGENEOUS
INFLOW CONDITIONS USING POSITRON EMISSION PARTICLE
TRACKING
Abstract 213
1. Introduction 215
2. Model 217
2.1 Granular dynamics 217
2.2 External forces 217
2.3 Gas phase hydrodynamics 218
3. Positron Emission Particle Tracking 219
4. Comparison between PEPT data and simulation 221
5. Results 222
6. Conclusions 230
Notation 231
References 232
Publications 235
Dankwoord (Acknowledgements) 237
Levensloop 241
Summary
________________________________________________________________________
1
Summary
Gassolid twophase flows are encountered in a wide variety of industrial applications.
The complex hydrodynamics of these systems is still not fully understood which renders
the scaleup of these units difficult. Therefore the development and validation of
fundamental hydrodynamic models is of utmost importance to gain more insight into the
complex hydrodynamics.
The study reported in this thesis is concerned with the granular dynamics of gassolid
twophase flows. In granular dynamics simulations the Newtonian equations of motion
are solved for each individual particle in the system while taking into account the mutual
interaction between particles and between particles and walls. The gasphase
hydrodynamics is described by the volume averaged NavierStokes equations for two
phase flow. The gasphase flow is resolved on a length scale that is larger than the
particle size. Two types of discrete particle models have been developed to be
incorporated into the granular dynamics simulations. The first is a (2D and 3D) hard
sphere model where the particles are assumed to interact through instantaneous, binary
collisions. A sequence of collisions is processed one collision at a time. The second is a
(2D) softsphere linear spring/dashpot model where contact forces between the particles
are calculated from the overlap between the particles. This softsphere model was chosen
since it is the most frequently used model in the literature and hence it is best suited for a
comparison with the hardsphere model. The key collision parameters in both models are
the coefficient of restitution (1 ≥ e ≥ 0), the coefficient of friction (µ ≥ 0) and the
coefficient of tangential restitution (1 ≥ β
0
≥ 0). The linear spring/dashpot model also
requires a spring stiffness to describe the particle interaction. The softsphere model is
capable of handling multiple particle interactions and can handle static situations in
contrast to the hardsphere model.
The effect of the collision parameters on the bed dynamics in a gas fluidised bed with
homogeneous inflow conditions was investigated. The collision parameters (except for
Summary
________________________________________________________________________
2
β
0
) turned out to have a profound influence on the fluidisation behaviour. When the
collisions were assumed to be fully elastic and perfectly smooth (e = 1, µ = 0) no bubbles
were observed and pressure fluctuations inside the bed were rather small. When more
realistic values for the collision parameters were used (e < 1, µ > 0) bubbles did appear
and the pressure fluctuations were larger. These trends were observed in 2D simulations
as well as in 3D simulations. The Root Mean Square (RMS) of the pressure fluctuations
inside the bed showed an almost linear dependency on the energy dissipation rate by
collisions during the simulation at low values of the energy dissipation rate. The
influence of a (lognormal) particle size distribution on the bed dynamics was less
pronounced than the influence of the collision parameters. When a wider size distribution
was taken into account the pressure peaks inside the bed were smaller.
The case of bubble formation at a single central orifice was chosen for a comparison
between the (2D) hardsphere model and the (2D) softsphere model. Preliminary soft
sphere simulations revealed that a minimum value of the spring stiffness was required to
ensure stable simulations. For higher values of the spring stiffness the influence of the
value of the spring stiffness on the simulation results was negligible but the required CPU
time increased significantly. Incorporation of a (log normal) particle size distribution
improved the agreement between simulation and experiment. Hardly any difference could
be observed between the results of the hardsphere simulations and the softsphere
simulations indicating that the assumption of binary collisions in the hardsphere model is
not limiting. With both types of models it was found that the assumption of fully elastic
and perfectly smooth collisions resulted in much worse agreement between simulation
and experiment: hardly any bubble could be observed.
Segregation was successfully simulated with the (2D) hardsphere model for systems
consisting of particles of equal size but different density as well as for systems consisting
of particles of equal density but different size. In the latter case a clear steady state was
not reached since the larger particles were continuously transported to the upper layers of
the bed in bubble wakes. A statistical analysis showed a rather wide spread in segregation
profiles indicating the limited value of a single frame analysis. A simulation assuming
Summary
________________________________________________________________________
3
fully elastic and perfectly smooth collisions showed a rapid and almost perfect
segregation due to the absence of bubbles in this system. Preliminary experimental
validation showed rather poor agreement between simulation and experiment. The
simulation predicted segregation at a lower gas velocity than used in the experiment.
The results of simulations of the riser section of a circulating fluidised bed proved to be
very sensitive with respect to the collision parameters. In the case of fully elastic and
perfectly smooth collisions hardly any clustering of particles could be observed as
opposed to the case where these collision parameters were assigned realistic values.
Particlewall collisions turned out to have very little influence on the flow structure. A
strong effect of the collision properties on the axial solids profile was found where a
pronounced buildup of solids was observed in the simulation with realistic values for the
collision parameters. In the simulation assuming fully elastic and perfectly smooth
collisions no buildup of solids was observed. This result is supported by experimental
findings reported in the open literature. Lift forces acting on the suspended particles
turned out to have a slightly redispersive effect on the flow structure which made the
radial segregation of the solids a little less pronounced.
The (2D) hardsphere model was experimentally validated using the Positron Emission
Particle Tracking (PEPT) facility at the University of Birmingham. A quasi 2D, gas
fluidised bed with homogeneous inflow conditions was used for the validation. In the
experiment the motion of a single tracer particle was tracked during one hour. The PEPT
data was timeaveraged to allow for a comparison with the results of a simulation where
15,000 particles were tracked during 45 seconds. The collision parameters required for
the simulation were obtained from independent measurements at the Open University at
Milton Keynes. The results showed good agreement between experiment and simulation
when the measured values for the collision parameters were used. When the collisions
were assumed to be fully elastic and perfectly smooth the agreement was much worse.
Summary
________________________________________________________________________
4
Samenvatting
________________________________________________________________________
5
Samenvatting
Gasvast tweefasenstromingen vormen een belangrijk onderdeel van een grote
verscheidenheid aan industriële processen. Het opschalen van dergelijke processen wordt
bemoeilijkt door de complexiteit van de hydrodynamica van dergelijke systemen. Het is
daarom van het grootste belang meer inzicht te krijgen in deze complexe hydrodynamica
door het ontwikkelen en valideren van fundamentele hydrodynamische modellen.
Het onderzoek dat ten grondslag ligt aan dit proefschrift houdt zich bezig met de
granulaire dynamica van gasvast tweefasenstromingen. In granulaire dynamica
simulaties worden de Newtonse bewegingsvergelijkingen opgelost voor elk individueel
granulair deeltje in het systeem waarbij de interactie tussen deeltjes onderling alsmede de
interactie tussen deeltjes en systeemwanden wordt verdisconteerd. De hydrodynamica
van de gasfase wordt beschreven door de volume gemiddelde NavierStokes
vergelijkingen voor tweefasenstroming. De stroming van de gasfase wordt opgelost op
een lengteschaal die groter is dan de grootte van een afzonderlijk vaste stof deeltje. Voor
de granulaire dynamica simulaties zijn twee soorten discrete deeltjes modellen
ontwikkeld. De eerste is een (2D en 3D) harde bollen model waar de interactie tussen
de deeltjes wordt verondersteld te verlopen via instantane, binaire botsingen. De
opeenvolgende botsingen worden hierbij één voor één afgehandeld in chronologische
volgorde. Het tweede model is een (2D) zachte bollen lineaireveer/smoorpot model
waarbij de contactkrachten tussen de deeltjes worden berekend uit hun onderlinge
overlap. Dit zachte bollen model is gekozen omdat het het meest gebruikt is in de
literatuur en daardoor het best geschikt is voor een vergelijk met het harde bollen model.
De belangrijkste botsingsparameters in beide modellen zijn de restitutiecoëfficiënt (1 ≥ e
≥ 0), de frictiecoëfficiënt (µ ≥ 0) en de tangentiele restitutiecoëfficiënt (1 ≥ β
0
≥ 0). In het
lineaireveer/smoorpotmodel is behalve deze parameters ook nog een veerkonstante
vereist om de deeltjesinteractie te beschrijven. In tegenstelling tot het harde bollen model
Samenvatting
________________________________________________________________________
6
is het zachte bollen model in staat om meervoudige deeltjesinteractie te verdisconteren en
tevens is het geschikt om statische situaties te simuleren.
De invloed van de botsingsparameters op het stromingsgedrag van een gasgefluidiseerd
bed met homogene instroomcondities is onderzocht. De botsingsparameters (met
uitzondering van β
0
) bleken een grote invloed te hebben op het fluïdisatiegedrag. Als
werd aangenomen dat de botsingen volledig elastisch en perfect glad verliepen (e = 1, µ =
0) werden geen bellen waargenomen en waren de drukfluctuaties in het bed klein. Met
realistische waarden voor de botsingsparameters (e < 1, µ > 0) werden wel bellen
waargenomen en waren de drukfluctuaties in het bed aanzienlijk groter. Deze trend werd
zowel in 2D als in 3D simulaties waargenomen. De gemiddelde kwadratische waarde
van de drukfluctuaties vertoonde een zo goed als lineaire afhankelijkheid van de
energiedissipatiesnelheid bij lage waarden van deze laatste. The invloed van een (log
normale) deeltjesgrootteverdeling op het fluïdisatiegedrag was minder groot dan de
invloed van de botsingsparameters. Met een bredere deeltjesgrootteverdeling werden de
drukpieken in het bed iets lager.
Een vergelijk tussen het (2D) zachte bollen en het (2D) harde bollen model werd
gemaakt aan de hand van simulaties van belvorming aan een centraal inspuitpunt.
Verkennende berekeningen met het zachte bollen model lieten zien dat een minimum
waarde van de veerkonstante vereist was om verzekerd te zijn van een stabiele simulatie.
Voor hogere waarden van de veerkonstante werd de invloed ervan op de resultaten van de
simulaties verwaarloosbaar waarbij echter wel de benodigde rekentijd drastisch toenam.
Wanneer een (log normale) deeltjesgrootteverdeling werd verdisconteerd verbeterde de
overeenkomst tussen simulatie en experiment aanzienlijk. Er kon echter nauwelijks enig
verschil worden waargenomen tussen de resultaten van het harde bollen model en die van
het zachte bollen model. Dit gaf aan dat de aanname van binaire botsingen in het harde
bollen model niet beperkend is. Met beide typen modellen werd de overeenkomst tussen
simulatie en experiment veel slechter wanneer een simulatie werd uitgevoerd onder de
aanname van volledig elastische en perfect gladde botsingen.
Samenvatting
________________________________________________________________________
7
Het (2D) harde bollen model werd met succes ingezet bij het simuleren van segregatie in
systemen die bestaan uit deeltjes van gelijke grootte maar verschillende dichtheid
alsmede systemen bestaande uit deeltjes van gelijke dichtheid maar verschillende grootte.
In dit laatste geval werd geen stationaire toestand bereikt omdat de deeltjes continu
omhoog werden getransporteerd in het zog van bellen. Een statistische analyse liet een
grote spreiding zien in het segregatieprofiel wat aangeeft dat een analyse op basis van een
tijdsopname beperkte waarde heeft. Een simulatie waarbij de botsingen volledig elastisch
en perfect glad werden verondersteld liet een zeer snelle en bijna volledige segregatie
zien wat toegeschreven kon worden aan de afwezigheid van bellen in dit systeem. Een
eerste experimentele validatie liet een matige overeenkomst zien tussen simulatie en
experiment. De simulatie voorspelde segregatie bij lagere gassnelheden.
In simulaties van de risersectie van een circulerend gefluïdiseerd bed werd gevonden dat
het stromingsgedrag sterk afhankelijk is van de botsingsparameters. Clustervorming werd
niet of nauwelijks waargenomen als de botsingen volledig elastisch en perfect glad
werden verondersteld. Met realistische waarden voor deze parameters kon clustervorming
wel degelijk worden waargenomen. De botsingsparameters voor deeltjeswand botsingen
bleken nauwelijks invloed op het stromingsgedrag te hebben. In de simulatie met
realistische waarden voor de botsingsparameters werd een duidelijke opbouw van een
axiaal vaste stof profiel geconstateerd in tegenstelling tot de simulatie met ideale
botsingsparameters. Dit is in overeenstemming met experimenteel waargenomen trends
gerapporteerd in de open literatuur. Liftkrachten bleken slechts een klein dispersief effect
te hebben wat er toe bij droeg dat de radiële segregatie van de vaste stof wat minder groot
was.
Het (2D) harde bollen model werd experimenteel gevalideerd met behulp van Positron
Emission Particle Tracking (PEPT) experimenten uitgevoerd aan de universiteit van
Birmingham. Een quasi 2D gas gefluïdiseerd bed met homogene instroomcondities werd
gebruikt voor de validatie. In het PEPT experiment werd de beweging van een tracer
deeltje gevolgd gedurende een uur. De data van dit experiment werd tijdgemiddeld om
een vergelijk mogelijk te maken met een simulatie waarin 15000 deeltjes werden gevolgd
Samenvatting
________________________________________________________________________
8
gedurende 45 seconden. De botsingsparameters benodigd voor de simulatie werden
onafhankelijk gemeten aan de Open University in Milton Keynes. De resultaten van een
simulatie waarin deze waarden werden gebruikt vertoonde goede overeenstemming met
het experiment. De overeenstemming tussen het experiment en een simulatie waarin de
botsingen volledig elastisch en perfect glad werden verondersteld was beduidend minder
goed.
General Introduction
______________________________________________________________________________________
9
Chapter 1.
GENERAL INTRODUCTION
Abstract:
In this chapter a brief introduction to fluidisation is presented. The hydrodynamics of
gasfluidised beds is very complex and still not very well understood which renders the
scaleup of these units difficult. Therefore fundamental hydrodynamic models are
required to gain more insight into the hydrodynamics of gasfluidised beds. These
fundamental models can be classified into three categories based on the level of
microscopy featured in the model. In this thesis the focus is on discrete particle models
which form the intermediate category in the concept of multiscale modelling. The
position of discrete particle models within the multiscale modelling concept is explained
and the objective of the work presented in this thesis is formulated. Finally the outline of
this thesis is presented.
Chapter 1
______________________________________________________________________________________
10
General Introduction
______________________________________________________________________________________
11
1. Fluidisation
Gas fluidised beds are widely applied in the chemical process industry (Kunii and
Levenspiel, 1991) because of several advantageous properties including isothermal
conditions throughout the bed, excellent heat and mass transfer properties and the
possibility of continuous operation. Typical applications cover a wide variety of physical
and chemical processes such as fluidised bed combustion, catalytic cracking of oil, gas
phase polymerisation of olefins and fluidised bed granulation (detergents, fertilisers) to
name a few.
In gas fluidised beds the gravity force acting on the solid particles is compensated by the
drag forces exerted on the particles by the upward flowing gas. The minimum fluidisation
velocity (u
mf
) is defined as the superficial gas velocity at which the gravity force acting on
the particles is just counterbalanced by the drag forces exerted on the particles by the gas
phase. When operated at gas velocities above u
mf
several regimes are encountered. The
three regimes that are featured in this thesis are presented in Figure 1.1. In this figure the
gas velocity increases from left to right.
Figure 1.1. The three fluidisation regimes featured in this thesis. From left to right:
the bubbling regime, the slug flow regime and the fast fluidisation regime.
Chapter 1
______________________________________________________________________________________
12
At velocities exceeding u
mf
usually gas bubbles are present in the bed. These bubbles
have a decisive influence on the hydrodynamics of a bubbling fluidised bed and hence on
its performance as a chemical reactor and/or a heat exchange unit. When the gas velocity
is increased further also the size of the bubbles increases. The bubble size may approach
the bed diameter in which case the slug flow regime prevails. When the gas velocity is
increased beyond the terminal fall velocity of the particles the fast fluidisation regime is
encountered. In this regime the particles are entrained with the gas flow and transported
upward through the riser. At the top of the riser the particles are separated from the gas in
a cyclone and fed back to the bottom of the riser. Such systems are commonly referred to
as Circulating Fluidised Beds (CFB’s). In between the slug flow and the fast fluidisation
regime the turbulent fluidisation regime can be distinguished and when the gas velocity is
further increased starting in the fast fluidisation regime the pneumatic transport regime is
encountered. In this thesis however, the focus is on the bubbling, slug flow, and fast
fluidisation regime.
Although gas fluidised beds are widely applied, the scaleup of these systems is still very
complicated which is mainly due to the complex hydrodynamics of these systems.
Therefore it is of crucial importance to develop a thorough understanding of the
hydrodynamics of gas fluidised beds. Together with the development of dedicated
experimental techniques the development of fundamental hydrodynamic models is of
utmost importance to achieve a better understanding of fluidisation. Eventually this will
lead to the improvement of existing processes, improved scale up and the design of more
efficient future processes. In the work presented in this thesis the development of such a
fundamental hydrodynamic model is described and the results obtained with it are
presented.
2. Hydrodynamic modelling
During the last few decades Computational Fluid Dynamics (CFD) has become a very
powerful and versatile tool for the numerical analysis of transport phenomena (Kuipers
General Introduction
______________________________________________________________________________________
13
and van Swaaij, 1998). With continuously increasing computer power combined with the
development of improved physical models CFD has become a very useful tool for
chemical engineers. CFD modelling of gasfluidised beds has proven to be successful and
new developments in this area are promising. The majority of studies on modelling of
fluidised beds is concerned with the hydrodynamics only. Although attempts have been
reported where the hydrodynamics were modelled combined with mass transfer and
chemical reaction (Samuelsberg and Hjertager (1996), Gao et al., 1999) the results of
such attempts depend strongly on how well the hydrodynamics is modelled. Kuipers et al.
(1998) demonstrated that the predicted performance of a riser reactor in terms of
chemical conversion depends strongly on the prevailing flow structure in the riser. If the
flow structure is not well captured by the hydrodynamic model a sensible prediction of
the reactor performance is impossible. Therefore the development of reliable
hydrodynamic models is of utmost importance in order to arrive ultimately at models that
are capable of predicting the performance of fluidised beds reactors. Hence, the focus of
the present study is on the hydrodynamics only.
2.1 Multiscale modelling
Due to the complexity of the hydrodynamics of multiphase flows it has become accepted
that a single generalised CFD model cannot cover the wide variety of phenomena
encountered in multiphase flows (Delnoij et al. 1997, Kuipers and van Swaaij, 1997).
Instead, specific models have to be developed that are tailor made to capture the relevant
phenomena occurring at the length scale to which they are applied. By incorporating
microscopic information from sub scale models and passing on information to super scale
models a multi scale modelling concept can be established.
Three different classes of fundamental hydrodynamic models (learning models according
to van Swaaij, 1985) of gas fluidised beds can be distinguished. These models can be
combined together in a multi scale concept for fundamental hydrodynamic models of
gas fluidised beds as is schematically represented in Figure 1.2.
Chapter 1
______________________________________________________________________________________
14
In order to model a large (industrial) scale fluidised bed a continuum model, where the
gas phase and the solids phase are regarded as interpenetrating continuous media, is the
appropriate choice. This EulerianEulerian type of model have been developed and
successfully applied over the last two decades (Kuipers et al., 1992, Gidaspow, 1994
among many others). These models require closure relations for the solids phase stress
tensor and the fluidparticle drag where commonly empirical relations are used in the
absence of more accurate closures. Improved closure relations for the solids phase stress
tensor can be obtained by using the kinetic theory of granular flow (Sinclair and Jackson,
1989, Nieuwland et al., 1996, among many others).
Figure 1.2. Multiscale modelling concept for fundament al hydrodynamic models of
gasfluidised beds.
In discrete particle models the Newtonian equations of motion are solved for each
individual solid particle in the system. In this EulerianLagrangian type of model a
closure relation for the solids phase rheology is no longer required since the motion of the
individual particles is solved directly. However, the number of particles that can be taken
into account in this technique is limited (< 10
6
). Therefore it is not yet possible, even
with modern day super computers, to simulate a large (industrial) scale system. However,
this type of model can be used to arrive at improved closure equations for continuum
models by employing techniques from statistical mechanics. Also assumptions made
Continuum models
Large (industrial) scale
simulations
Discrete particle models
Particleparticle interaction
closure laws
Lattice Boltzmann models
Fluid particle interaction
closure laws
General Introduction
______________________________________________________________________________________
15
within the framework of the kinetic theory of granular flow, as incorporated in advanced
continuum models, can be verified. Since discrete particle models are very well suited to
study the influence of particle properties on the hydrodynamics of gasfluidised beds this
makes them very useful models within the multiscale modelling concept. However,
discrete particle models still require closure relations for the fluidparticle drag since the
gas flow is resolved on a length scale larger than the particle size. In the absence of better
closures empirical relations for the fluidparticle drag have to be used.
When the gas flow is resolved on a length scale smaller than the particle size these
closure relations for fluidparticle drag are no longer required. Instead they can actually
be obtained from the simulations. The Lattice Boltzmann technique seems to be best
suited for such simulations because it is very flexible in dealing with complex flow
geometries. In Chapter 3 some additional techniques besides Lattice Boltzmann
simulations are presented that can be used for the same purpose. It is important to realise
that such simulations are limited to systems consisting of a number of particles that is
significantly smaller (<10
3
) than the number of particles that can be taken into account
using discrete particle models (<10
6
).
In short the multiscale concept as presented in Figure 1.2 consists of three classes of
models where more detail of the twophase flow is resolved going from continuum
models to discrete particle models to Lattice Boltzmann models. This goes at the cost of
increased computational requirements which necessitates a size reduction of the
simulated system. The model capable of simulating a larger system is fed with a closure
relation obtained from a more microscopic simulation. In return the results of these
simulations can be used to pass on information to models capable of simulating the flow
on a larger scale. Before such a connection between separate scales can be established the
individual simulation techniques must be well developed, verified and experimentally
validated. In the work presented in this thesis the focus is on the development and the use
of discrete particle models.
Chapter 1
______________________________________________________________________________________
16
3. Discrete particle modelling
In granular dynamics simulations of gas fluidised beds the Newtonian equations of
motion are solved for each individual solid particle by using a discrete particle model. For
the fluidparticle interaction empirical relations have to be used since the hydrodynamics
of the gasphase is resolved on a length scale larger than the particle size.
The discrete particle approach for gasfluidised beds was pioneered by Tsuji et al. (1993)
who developed a twodimensional softsphere discrete particle model of a gas fluidised
bed based on the work of Cundall and Strack (1977). Kawaguchi et al. (1998) extended
this model to three dimensions as far as the motion of the particles is concerned. Hoomans
et al. (1996) presented a hardsphere approach in their twodimensional discrete particle
model of a gasfluidised bed. Ouyang and Li (1998) developed a slightly different version
of this model. Xu and Yu (1997) presented a hybrid simulation technique that features
elements from both hardsphere and softsphere techniques. Mikami et al. (1998)
extended the model originally developed by Tsuji et al. (1993) to include cohesive forces
between the particles. Recent developments in this area include the (2D) simulation of
fluidised bed with internals (Rong et al., 1999) and the (2D) simulation of gasphase
olefin polymerisation (Kaneko et al., 1999) where energy balances and chemical reaction
rates were taken into account.
As far as particle interaction is concerned a multiscale modelling concept can be
distinguished as is schematically presented in Figure 1.3. As mentioned in the previous
section continuum models require closure relations for the solids phase rheology (i.e.
viscosity, solids phase pressure). These relations can be obtained from discrete particle
models by employing techniques from statistical mechanics. Discrete particle models can
also be used to verify assumptions made in the kinetic theory of granular flow which is
used in most of the recently developed continuum models.
General Introduction
______________________________________________________________________________________
17
Figure 1.3. Multiscale modelling concept for models involving particle interaction.
The particle interaction parameters required for the discrete particle models can be
obtained from experiments as was the case in the work presented in this thesis. However,
it is important to realise that these parameters can also be obtained from microscopic
particle interaction models. By using the appropriate contact theory the particle
interaction parameters can be obtained using only material properties as input. In this way
a multiscale concept for particleinteraction models arises. Transport properties for the
solidsphase in a continuum model are obtained from discrete particle models and the
particle interaction parameters required in discrete particle models are obtained from
contact theory based on material properties.
4. Outline of this thesis
The objective of the work presented in this thesis is to study the influence of particle
properties on the dynamics of gassolid twophase flows using discrete particle models.
In chapters 2 and 3 the theoretical framework of the granular dynamics simulations of
gas fluidised beds will be presented and in the chapters 4, 5, 6, 7 and 8 several
applications and experiment al validation of these models will be discussed.
Solids phase rheology Continuum models
Particle interaction
parameters
Discrete particle models
Material properties Contact theory
Chapter 1
______________________________________________________________________________________
18
Chapter 2 deals with granular dynamics. The two types of discrete particle models used in
this work, the hardsphere and the softsphere model, are presented in detail. The hard
sphere model has been developed both in 2D and in 3D whereas the softsphere linear
spring/dashpot model has been developed in 2D only. The two types of models are
thoroughly discussed and a comparison will be presented. Furthermore an experimental
technique to measure the collision parameters required in both types of models is
presented. Finally the external forces acting on a single (Lagrangian) particle in a gas
fluidised bed are presented.
In Chapter 3 the gas phase dynamics is discussed for which the Eulerian approach was
adopted. Alternative techniques will be discussed briefly. The volume averaged
continuity and momentum conservation equations are presented together with the closure
equations. Special attention is paid to the twoway coupling between the motion of the
solid particles and the motion of the continuous gasphase.
In Chapter 4 the hardsphere model is applied to gasfluidised beds with homogenous
inflow conditions. The dependency of the bed dynamics on the collision parameters will
be investigated. Furthermore the influence of the incorporation of a (log normal) particle
size distribution on the bed dynamics is studied. Finally the results obtained with the 2D
model are compared with the results obtained with the 3D model.
In Chapter 5 the hardsphere model is compared with the softsphere model. The
formation of a single bubble at a central orifice will be used as a test case for the
comparison. The dependency of the results of the softsphere model on the value of the
spring stiffness is investigated and the influence of the incorporation of a (lognormal)
particle size distribution will be studied. The results of the hardsphere and softsphere
model will be compared with each other and with experimental data. Finally the influence
of the collision parameters on the bubble formation process will be investigated.
General Introduction
______________________________________________________________________________________
19
Chapter 6 deals with segregation phenomena in bubbling gas fluidised beds. Simulations
of systems consisting of particles of equal size but different density and systems
consisting of particles of equal density but different size will be presented. Both the hard
sphere (2D) and the softsphere models are used. Segregation will be demonstrated for
each of these systems and a statistical analysis of the segregation dynamics is presented
for the system consisting of particles of equal density but different size. For the latter
system the influence of the collision parameters on the segregation is investigated. Finally
an experimental validation is performed using a specially designed set up.
In Chapter 7 the hardsphere model is applied to riser flow. The model is modified to
allow for a continuous transport of particles throughout the riser duct. The influence of
the collision parameters on the flow structure in the riser will be investigated. The
formation of clusters is studied and radial and axial solids profiles are obtained after time
averaging of the data. The results will be compared with experimental data available in
the open literature.
Finally in Chapter 8 experimental validation of granular dynamics simulations of a gas
fluidised bed with homogeneous inflow conditions is performed using the Positron
Emission Particle Tracking (PEPT) facility at the University of Birmingham. In the PEPT
experiment the motion of a single radioactive tracer particle is tracked during one hour.
The time averaged data obtained from the experiment will be compared with the results
of a simulation where 15,000 particles are tracked during 45 seconds. The results will be
compared on the basis of velocity maps, occupancy plots and speed histograms. The
results of a simulation where independently measured collision parameters are used will
be compared with the results of a simulation assuming fully elastic, perfectly smooth
collisions and the results of the PEPT experiment.
References
Delnoij, E., Kuipers, J.A.M. and van Swaaij W.P.M., (1997). Computational fluid
dynamics applied to gas liquid contactors, Chem. Engng Sci. 52, 3623.
Chapter 1
______________________________________________________________________________________
20
Gao, J., Xu, C., Lin, S., Tang, G. and Guo, Y., (1999). Advanced model for turbulent gas
solid flow and reaction in FCC riser reactors, AIChE J., 45, 1095.
Hoomans, B.P.B., Kuipers, J.A.M., Briels, W.J. and van Swaaij, W.P.M., (1996).
Discrete particle simulation of bubble and slug formation in a twodimensional gas
fluidised bed: a hardsphere approach, Chem. Engng Sci., 51, 99.
Hoomans, B.P.B., Kuipers, J.A.M., Briels, W.J. and van Swaaij, W.P.M., (1998).
Comments on the paper “Numerical simulation of the gassolid flow in a fluidized bed by
combining discrete particle method with computational fluid dynamics” by B.H. Xu and
A.B. Yu, Chem. Engng Sci., 53, 2645.
Kaneko, Y., Shiojima, T. and Horio, M., (1999). DEM simulation of fluidized beds for
gasphase olefin polymerization, Chem. Engng Sci., 54, 5809.
Kawaguchi, T., Tanaka, T. and Tsuji, Y., (1998). Numerical simulation of two
dimensional fluidized beds using the discrete element method (comparison between the
two and threedimensional models), Powder Technol. 96, 129.
Kuipers, J.A.M., van Duin K.J., van Beckum, F.P.H. and van Swaaij, W.P.M., (1992). A
numerical model of gasfluidized beds, Chem. Engng Sci., 47, 1913.
Kuipers, J.A.M. and van Swaaij, W.P.M., (1997). Application of computational fluid
dynamics to chemical reaction engineering, Rev. Chem Eng., 13, No 3, 1.
Kuipers, J.A.M. and van Swaaij, W.P.M., (1998). Computational fluid dynamics applied
to chemical reaction engineering, Adv. Chem Eng., 24, 227.
Kunii, D. and Levenspiel, O., (1991). Fluidization engineering, 2
nd
edition, Butterworth
Heinemann, Boston, USA.
General Introduction
______________________________________________________________________________________
21
Mikami, T., Kamiya, H. and Horio, M., (1998). Numerical simulation of cohesive powder
behavior in a fluidized bed, Chem. Engng Sci. 53, 1927.
Ouyang, J. and Li, J. (1999). Particle motionresolved discrete model for simulating gas
solid fluidization, Chem. Engng Sci., 54, 2077.
Rong, D., Mikami, T. and Horio, M., (1999). Particle and bubble movements around
tubes immersed in fluidized beds – a numerical study, Chem. Engng Sci., 54, 5737.
Samuelsberg, A.E. and Hjertager, B.H., (1996). Computational fluid dynamic simulation
of an oxychlorination reaction in a fullscale fluidized bed reactor, in Proceedings of the
5
th
Int. Conf. Circulating Fluidized beds, Beijing (China), May 28 June 1, 1996.
van Swaaij, W.P.M., (1985)., Chemical reactors, in Fluidisation, 2
nd
edition p. 595,
Davidson, J.F., Clift, R. and Harrison, D. (Eds), Academic Press, London, UK.
Tsuji, Y., Kawaguchi, T. and Tanaka, T., (1993). Discrete particle simulation of two
dimensional fluidized bed, Powder Technol. 77, 79.
Xu, B.H. and Yu, A.B., (1997). Numerical simulation of the gassolid flow in a fluidized
bed by combining discrete particle method with computational fluid dynamics, Chem.
Engng Sci. 52, 2785.
Chapter 1
______________________________________________________________________________________
22
Granular Dynamics
______________________________________________________________________________________
23
Chapter 2.
GRANULAR DYNAMICS
Abstract:
In this chapter a review is presented of the various granular dynamics simulation
techniques available in the literature. The approaches can be divided into three groups:
hardparticle techniques, softparticle techniques and Monte Carlo techniques. These
three types of simulation are discussed with emphasis on their application to simulation
of gassolid twophase flow. In this work two types of models were developed. The first is
a hardsphere model where the particles are assumed to interact through instantaneous,
binary collisions. A sequence of collisions is processed one collision at a time. The
second is a softsphere linear spring/dashpot model where contact forces between the
particles are calculated from the overlap between the particles. Both models are
described in detail. The key parameters in these models to describe a collision are the
coefficient of restitution (e), the coefficient of friction (µ) and the coefficient of tangential
restitution (β
0
). The effect of these parameters on a single collision is demonstrated. A
comparison between the hardsphere and the softsphere model is presented. The soft
sphere model is capable of simulating static situations unlike the hardsphere technique.
However in the softsphere model special care must be taken in the choice of time step
and the spring stiffness required for the calculation of the repulsive force. It is shown that
the spring stiffness of the tangential spring should not be taken equal to the stiffness of
the normal spring in order to avoid unrealistic behaviour. Experiments are described that
enable measurement of the three collision parameters by careful observation of single
impacts. Finally the external forces acting on the particles used in the simulations of gas
fluidised beds in this work are presented.
Chapter 2
______________________________________________________________________________________
24
Parts of this chapter are based on the paper:
B.P.B. Hoomans, J.A.M. Kuipers, W.J. Briels and W.P.M. van Swaaij, (1996). Discrete particle simulation
of bubble and slug formation in twodimensional gasfluidised beds: a hardsphere approach, Chem. Engng
Sci., 51, 99.
Granular Dynamics
______________________________________________________________________________________
25
1. Introduction
Measured by tons granular matter is after water the most manipulated material in the
world (de Gennes, 1999). However, roughly 40% of the capacity of industrial plants is
wasted due to problems related to transport of granular matter (Knowlton et al., 1994).
Therefore it is not surprising that Granular Dynamics (GD) has attracted the interest of a
large number and a wide variety of researchers over the last decades. The interest not
only originates from industrial needs but there is also an increasing interest in granular
media from a more fundamental perspective. Phenomena like heap formation and arching
provide both theoreticians and experimentalists with challenging problems concerning the
statics of granular matter (de Gennes, 1999). However the greatest challenge is provided
by the dynamics of this material which unveils a wealth of phenomena, such as standing
waves in vibrated beds, segregation, clustering and inelastic collapse, that we are just
beginning to understand (Jaeger et al., 1996). In this work the focus will be on the
granular dynamics of gassolid twophase flow which is a field of research that has
gained an increasing amount of attention over the last decade.
The systems considered in this work are all operated in the grain inertia regime according
to Bagnold (1954) which implies that particleparticle and particlewall interactions are
dominated by inertia rather than viscous forces. This holds for gassolid flow with rather
coarse particles which is the subject of study in this work. In the case of liquidsolid
fluidised beds lubrication forces have to be taken into account (Schwarzer, 1995) and for
liquidsolid systems with smaller particles (< 100 µm) direct particle interaction does not
even occur. Such systems can be studied by means of Stokesian dynamics simulations
(Ichiki and Hayakawa, 1995). The modelling approaches adopted in Granular Dynamics
can be roughly divided into two groups: soft particle and hard particle approaches. Before
the two approaches adopted in this work (hardsphere and softsphere linearspring/dash
pot) will be described in detail in the preceding paragraphs a short review of the different
approaches that are available in the literature will be presented. These approaches can be
divided into three types of simulations: hardparticle approaches, softparticle approaches
Chapter 2
______________________________________________________________________________________
26
and MonteCarlo approaches. Hardparticle simulations can be typified as event driven
because the interaction times are small compared to free flight times. In event driven
simulations the progression depends on the number of collisions that occurs. Softparticle
simulations can be typified as time driven because the interaction times are large
compared to free flight times. In time driven simulations a constant time step is used to
progress through the dynamics of the system.
1.1 HardParticle Approaches
The hardsphere simulation technique was first presented by Alder and Wainwright
(1957) in order to study phase transitions in molecular systems by means of numerical
simulations. In a later paper (Alder and Wainwright, 1959) the technique was presented
in more detail including a way to deal with a square well interaction potential apart from
merely hardsphere interaction. In hardsphere simulations the particles are assumed to
interact through instantaneous, binary collisions. A sequence of collisions is processed
one collision at a time in order of occurrence. For this purpose a list of future collisions is
compiled and updated when necessary. For a comprehensive introduction to this type of
simulation the reader is referred to Allen and Tildesley (1990). A lot of effort has been
put into the further optimisation and development of this event driven type of simulation
technique (Marin et al., 1993). Over two decades after the publications by Alder and
Wainwright the hardsphere approach was discovered as a useful tool for granular
dynamics simulations (Campbell, 1985). The dissipative particle interaction in granular
media makes these systems significantly different from molecular systems where energy
is always conserved. Hence energy has to be continuously supplied to a granular system
in order to keep the particles in motion. This can for instance be achieved by applying a
shear rate through a proper choice of boundary conditions (Campbell, 1985). In granular
dynamics simulations of gassolid twophase flow there is a constant stream of energy
supplied to the particles through gravity and the drag force exerted on the particles by the
gasphase. For the remainder of this paragraph the focus will be on the various techniques
used in granular dynamics simulations of gassolid twophase flow.
Granular Dynamics
______________________________________________________________________________________
27
Tsuji et al. (1987) presented a hardsphere granular dynamics model for dilute gassolid
flow in a horizontal channel. They neglected the particleparticle interaction but instead
focussed on particlewall interaction where an irregular bouncing model was used. This
can be justified since the solids fraction in their system was very low (ε
s
< 0.01).
Sommerfeld (1990) and Frank et al. (1992) presented similar approaches where also
attention was paid to experimental validation.
For denser flows the particleparticle interaction can no longer be neglected. The problem
in these simulations is not so much the description of the particleparticle interaction
itself since accurate collision models are available. The main problem is the large number
of particles and hence the large number of collisions that have to be detected and
processed. A technique that takes particleparticle interaction into account without
detecting and processing every single collision that occurs in the system is the Direct
Simulation Monte Carlo (DSMC) technique that was originally developed by Bird (1976)
for molecular simulations. This technique was employed by Yonemura et al. (1993) in
their simulations of gassolid flow in a vertical channel. A more detailed description of
this technique was presented by Tsuji et al. (1998). The particles tracked in this
simulation technique are not actual particles but ‘sample’ particles that represent several
‘real’ particles. The DSMC method can therefore be regarded as a coarse grained model
of the actual dynamics. The occurrence of a collision as well as the geometry of a
collision (i.e. the collision coordinate system) are determined by a random number
generator where the probability of a collision depends on the local solids fraction (the
greater the solid fraction the greater the probability of a collision). However one should
be careful when applying this technique since the modified Nanbu method used by Tsuji
et al. (1998) does not guarantee exact conservation of energy in the absence of dissipative
terms (Frezzoti, 1997). Oesterlé and Petitjean (1993) presented a technique that resembles
the DSMC technique strongly but differs in the sense that the particles in their
simulations are ‘real’ particles instead of ‘sample’ particles.
Chapter 2
______________________________________________________________________________________
28
Hopkins et al. (1991) developed a type of hardparticle simulation technique where
collisions were detected at constant time intervals. Although they did not apply it directly
to gassolid flow the technique is worthwhile mentioning here since Ge and Li (1997) and
Ouyang and Li (1999) adopted this approach in their simulations of gas fluidised beds. In
this computational strategy the particles are assumed to interact as hardspheres but the
motion update is performed at constant time steps. After each time step the whole system
is scanned for overlap between particles and when this is detected the particles involved
are assumed to have collided during that time step and subsequently their new velocities
are calculated. This technique works fine for relatively dilute systems when the time step
is chosen relatively small. For denser systems overlap may be found between more than
two particles which gives rise to incorrect dynamics. In the case of larger time steps it
may be possible that collisions are overlooked. Hogue and Newland (1994) used a similar
strategy in their simulations where a technique was used that was especially designed to
handle nonspherical (i.e. polygonal) particles. In addition to the model of Hogue and
Newland, Müller and Liebling (1995) presented a triangulation technique that enables a
sequence of collisions between polygonal particles to be correctly processed.
Hoomans et al. (1996) presented a true hardsphere type of simulation technique for gas
fluidised beds. This was the first time that this technique was developed and applied with
success to such a dense system as a bubbling gas fluidised bed. In this twodimensional
model a sequence of collisions is processed proceeding from one collision to the next by
using a collision list that is compiled and updated in a highly efficient manner. In the
following chapter this technique will be explained in detail. Lun and Liu (1997) presented
a threedimensional hardsphere model for a more dilute gassolid flow where they used a
granular dynamics model presented earlier by Lun (1996).
1.2 SoftParticle Approaches
The Distinct Element Method (DEM) developed by Cundall and Strack (1979) was the
first granular dynamics simulation technique published in the open literature. They used a
Granular Dynamics
______________________________________________________________________________________
29
twodimensional soft particle model where the particles were allowed to overlap slightly.
The contact forces were subsequently calculated from the deformation history of the
contact using a linear spring/dashpot model. This method allowed for multiple particle
overlap although the net contact force was obtained from pairwise interactions. Soft
particle approaches differ in the choice of force scheme used to calculate the interparticle
forces. A review of various popular schemes for repulsive interparticle forces is
presented by Schäfer et al. (1996). Before the various softparticle approaches used in
fluidised bed simulations will be discussed, two alternatives to the Cundall and Strack
model are briefly outlined here.
Walton and Braun (1986) presented a model that uses two different spring constants to
model energy dissipation instead of a dashpot. The spring constant for the compression
phase (loading) of the constant is taken to be lower than the constant used for the
restitution phase (unloading). The coefficient of restitution can be related to the ratio of
the two spring constants. This model was used by McCarthy and Ottino (1998) and
Wightman et al. (1998) in their studies of granular mixing in a rotating container.
Langston et al. (1994) presented a force scheme that was based on a continuous potential
of an exponential form containing two unknown parameters, the stiffness of the
interaction and an interaction constant. In a later paper (Langston et al., 1995) they
presented a threedimensional version of this model. The repulsive force was obtained by
taking the gradient of this potential while the stiffness parameter was chosen in such a
way that particle overlap could not become too high without requiring too small a time
step. The interaction constant was chosen in such a way that the net force would be zero
when a particle would rest on top of another particle in a gravity field. One should be
careful when applying this model since it features rather unrealistic behaviour when using
a cutoff distance greater than the particle diameter. When two particles that do not touch
each other are positioned at the same height within the cutoff distance of the potential
they will still experience a repulsive force. This is not physically correct behaviour for
dry granular material.
Chapter 2
______________________________________________________________________________________
30
The softsphere approach was first applied to gas fluidised beds by Tsuji et al. (1993). In
their twodimensional model the particle interaction was calculated using a softsphere
interaction model similar to the one presented by Cundall and Strack (1979). Previously
Tsuji et al. (1992) presented a threedimensional model (as far as the particles were
concerned) for gassolid flow in a horizontal pipe. However, in this model the gasphase
dynamics was represented by a onedimensional model with hardly any dynamic features.
A threedimensional version of their fluidised bed model was presented by Kawaguchi et
al. (1998). In that model the particle motion was resolved in full 3D whereas the gas
phase dynamics was still calculated in 2D. This can be justified because the system used
in their simulations was a rather flat (quasi 2D) fluidised bed. In such a system the
motion of the gasphase in the third dimension can be neglected.
Schwarzer (1995) presented a twodimensional model of a liquidfluidised bed. Apart
from the interparticle forces, for which the Cundall and Strack model was used, also
lubrication forces were taken into account. These lubrication forces can be neglected for
gas fluidised beds but play an important role in liquidfluidised beds. Also a Gaussian
particle size distribution was taken into account in these simulations although the effect
of the size distribution on the flow behaviour was not investigated.
Xu and Yu (1997) presented a twodimensional model of a gas fluidised bed that was
based on the model developed by Tsuji et al. (1993). However in their simulations a
collision detection algorithm that is normally found in hardsphere simulations was used
to determine the first instant of contact precisely. The spring constants used in their
simulations were much higher than the ones used by Tsuji et al. (1993). Unfortunately no
results were reported that could show the importance of the detection algorithm for the
overall simulation results.
Mikami et al. (1998) extended the model developed by Tsuji et al. (1993) by
incorporating liquid bridge forces to simulate cohesive particle fluidisation. The particles
used in their simulations are still Geldart D particles but due to the liquid bridge forces
the fluidisation behaviour resembles the behaviour of Geldart C particles strongly.
Granular Dynamics
______________________________________________________________________________________
31
Mikami (1998) was also the first to present full 3D simulations (also with respect to the
gasphase) where no less than 500,000 particles were used.
1.3 Monte Carlo techniques
Another popular method to study many particle systems is the Monte Carlo technique
(Frenkel and Smit 1996). This method has been applied to granular systems by Rosato et
al. (1986). In their Monte Carlo simulations a new overlap free particle configuration is
generated at each step. The change in the system energy is then calculated and if this
change is negative the new configuration is accepted. If the system energy has increased
the new configuration is accepted with a probability obtained from Boltzmann
distribution based on the change in energy. Using this method Rosato et al. were able to
simulate segregation phenomena in shaken or vibrated systems where effects due to an
interstitial fluid were neglected. It is important to realise that time is not a variable in
Monte Carlo simulations. A Monte Carlo step can only be linked to an actual time step by
means of calibration but this is not a straightforward task. Therefore this method is not
capable of simulating the dynamics of a granular system without the input of apriory
knowledge.
The Monte Carlo technique is capable of predicting steady state (i.e. equilibrium)
conditions and for that purpose it has certain advantages over dynamic simulations.
Seibert and Burns (1998) were able to predict segregation phenomena in liquidfluidised
beds using an extended version of their previously developed model (Seibert and Burns,
1996). In their model they used a net force (calculated from gravity and fluid drag) to
calculate the change of energy involved with a particle movement. Although the results
compared rather well with experimental data no statement was made about the time scale
over which the phenomena occurred. This is of course to be expected for this type of
simulation technique.
Chapter 2
______________________________________________________________________________________
32
2. HardSphere approach
In the hard sphere model used in this work the particles are assumed to interact through
binary, quasi instantaneous collisions where contact occurs at a point. The particles are
perfect, homogeneous spheres and the interaction forces are impulsive. In between
collisions the particles are in free flight. First the collision model will be presented and
then the computational strategy and some optimisation techniques will be described.
Since the hardsphere model was developed in 3D as well as in 2D the collision model
will be presented in vector notation. For the 2D version the zcomponent of the position
and velocity vectors are zero and only rotation about the zaxis is considered.
2.1 Collision model
In the collision model it is assumed that the interaction forces are impulsive and therefore
all other finite forces are negligible during collision. The original 2D collision model
was mapped after the model presented by Wang and Mason (1992). In this work however
we will mainly adopt the notation used by Foerster et al. (1994) since that is more widely
accepted (see for example Lun, 1997 and Tsuji et al. 1998). The coordinate systems used
in our model are defined in Figure 2.1.
Consider the two colliding spheres a and b in Figure 2.1 with position vectors r
a
and r
b
.
The normal unit vector can now be defined:
b a
b a
r r
r r
n
−
−
· . (2.1)
Hence the normal unit vector points in the direction from the centre of particle b to the
centre of particle a. The point of origin is the contact point. Prior to collision, the spheres
with radii R
a
and R
b
and masses m
a
and m
b
have translation velocity vectors v
a
and v
b
and
Granular Dynamics
______________________________________________________________________________________
33
rotational velocity vectors ω
a
and ω
b
(clockwise rotation is negative by definition).
Velocities priortocollision are indicated by the subscript
0
.
y
x
z
n
t
b
a
Figure 2.1. Definition of the coordinate systems.
For a binary collision of these spheres the following equations can be derived by applying
Newton’s second and third laws:
J v v · − ) (
0 a, a a
m (2.2)
J v v − · − ) (
0 b, b b
m (2.3)
( ) J n × − · ω − ω
a a, a a
R I ) (
0
(2.4)
( ) J n − × · ω − ω
b b, b b
R I ) (
0
(2.5)
J v v v v · − − · − ) ( ) (
0 0 b, b b a, a a
m m (2.6)
J n× − · ω − ω · ω − ω ) ( ) (
0 0 b, b
b
b
a, a
a
a
R
I
R
I
(2.7)
2
5
2
mR I · . (2.8)
The impulse vector J is defined as follows:
Chapter 2
______________________________________________________________________________________
34
∫
·
·
·
c
t t
t
ab
dt
0
F J , (2.9)
where t
c
stands for the contact time (i.e. the duration of the contact).
From equations 2.6 and 2.7 it is clear that the postcollision velocities of both particles
can be calculated when the impulse vector J is known. If the force F
ab
in equation 2.9
were known as a function of all the parameters involved, the impulse J could be
calculated directly. Thornton (1997) demonstrated that based on a simplified theoretical
model for the normal interaction between elasticperfectly plastic spheres an analytical
solution could be obtained for the rebound velocity. Walton (1992) used two types of
finite element codes (DYNA2D and NIKE2D) to simulate the collision process in detail
on a subparticle level. The only input parameters necessary in these calculations are
material properties although assumptions have to be made about the deformation
behaviour (elastic/plastic) of the material. In simulations of gasfluidised beds a large
number of collisions (typically 10
6
10
9
) have to be processed and therefore the actual
physics of a binary collision has to be simplified to some extent and constitutive relations
have to be introduced.
Before these constitutive relations will be introduced first the relative velocity at the
contact point (v
ab
) has to be defined:
) (
, , c b c a ab
v v v − ≡ . (2.10)
( ) ( ) n v n v v
b b b a a a ab
R R × ω + − × ω − · . (2.11)
n v v v × ω + ω − − · ) ( ) (
b b a a b a ab
R R . (2.12)
From this relative velocity, the tangential unit vector can be obtained since the normal
unit vector is already defined in equation 2.1:
Granular Dynamics
______________________________________________________________________________________
35
n) n(v v
n) n(v v
t
⋅ −
⋅ −
·
0 , 0 ,
0 , 0 ,
ab ab
ab ab
. (2.13)
Equations 2.6 and 2.7 can now be rearranged using ( ) n J n J n J n ⋅ − · × × ) ( and equation
2.12 to obtain:
( ) ) (
2 1 1 0 ,
n J n J v v ⋅ − − · − B B B
ab ab
, (2.14)
where
,
`
.

+ ·
b a
m m
B
1 1
2
7
1
(2.15)
and
b a
m m
B
1 1
2
+ · . (2.16)
At this point constitutive relations are required to close the set of equations. Through
these constitutive relations three parameters enter the model. The first parameter is the
coefficient of (normal) restitution, ( ) 1 0 ≤ ≤ e :
( ) n v n v ⋅ − · ⋅
0 , ab ab
e . (2.17)
For nonspherical particles this definition can lead to energy inconsistencies (Stronge,
1990) but for spherical particles this definition holds. The second parameter is the
coefficient of (dynamic) friction, ( ) 0 ≥ µ :
( ) J n J n ⋅ − · × µ . (2.18)
The third parameter is the coefficient of tangential restitution, ( ) 1 0
0
≤ ≤ β :
Chapter 2
______________________________________________________________________________________
36
( )
0 , 0 ab ab
v n v n × − · × β . (2.19)
Notice that this relation does not affect the components parallel to n and that the
components orthogonal to n are related by a factor –β
0
. Although it is accepted that these
coefficients depend on particle size and impact velocity this is not taken into account in
this model. The only exception is made for the coefficient of normal restitution where
collisions occurring at a normal impact velocity less than a threshold value ‘MINC0’
(typically 10
4
m/s) are assumed to be perfectly elastic (e = 1.0).
Combining equations 2.14 and 2.17 yields the following expression for the normal
component of the impulse vector:
( )
2
0
1
B
e J
ab,
n
n v ⋅
+ − · (2.20)
For the tangential component two types of collisions can be distinguished that are called
sticking and sliding. If the tangential component of the relative velocity is sufficiently
high in comparison to the coefficients of friction and tangential restitution that gross
sliding occurs throughout the whole duration of the contact, the collision is of the sliding
type. The nonsliding collisions are of the sticking type. When β
0
is equal to zero the
tangential component of the relative velocity becomes zero during a sticking collision.
When β
0
is greater than zero in such a collision, reversal of the tangential component of
the relative velocity will occur. The criterion to determine the type of collision is as
follows:
( )
1
0 0
1
B J
n
ab,
t v ⋅ +
<
β
µ sliding (2.21)
( )
1
0 0
1
B J
n
ab,
t v ⋅ +
≥
β
µ sticking (2.22)
Granular Dynamics
______________________________________________________________________________________
37
For collisions of the sticking type, the tangential impulse is given by:
( ) ( )
1
0
0
1
0
0
1 1
B B
J
ab, ab,
t
t v v n ⋅
+ − ·
×
+ − · β β . (2.23)
For collisions of the sliding type, the tangential impulse is given by:
n t
J J µ − · . (2.24)
The total impulse vector is then simply obtained by addition:
t n J
t n
J J + · . (2.25)
The postcollision velocities can now be calculated from equations 2.6 and 2.7.
In particlewall collisions the mass of particle b (i.e. the wall) is infinitely large which
makes all terms 1/m
b
equal to zero. It is possible to implement a moving/rotating wall
through the velocity vectors v
b
and ω
b
but in the simulations performed for this work
these velocities are all set equal to zero.
The energy dissipated during a collision can be obtained by solving the following integral
over the duration of the collision:
∫ ∫
+ ·
t t ab n n ab tot dsp
dJ v dJ v E
, , ,
. (2.26)
The energy dissipated by the normal component in a collision is:
( )
2
2
2
0
1
2
e
B
v
E
ab,n,
dsp,n
− · . (2.27)
Chapter 2
______________________________________________________________________________________
38
For the energy dissipated by the tangential component the two types of collision have to
be distinguished again. If the collision is of the sticking kind, the dissipated energy is:
( )
2
0
1
2
0
1
2
β − ·
B
v
E
ab,t,
dsp,t
, (2.28)
and if the collision is of the sliding type, the dissipated energy is:
,
`
.

− ⋅ − ·
n ab, n dsp,t
B E J
2
1
J
1 0
µ µ t v . (2.29)
The total amount of energy dissipated in a collision is then obtained by adding the
tangential and normal contributions:
dsp,t dsp,n dsp,to
E E E + ·
t
. (2.30)
2.2 Key parameters of the collision model
Since the three key parameters of the collision model are of crucial importance for the
remainder of this work the effect of each of the three will now be highlighted. The system
considered here is a particle that collides with a flat wall under the influence of gravity
(g). For all these examples the particles do not experience any friction of the gas phase. In
Figure 2.2 the effect of the coefficient of restitution is illustrated.
When a particle collides perfectly elastically, without any energy dissipation (e = 1) with
a horizontal flat wall it will bounce back to the same height as it was initially released
from. No energy is dissipated in this process and the particle will eternally continue to
bounce. When e < 1, which is always the case for granular material, energy is dissipated
in the collision and the particle will not bounce back to the same height as it was initially
released from. Eventually the particle will come to rest on the wall.
Granular Dynamics
______________________________________________________________________________________
39
Figure 2.2. The effect of the coefficient of restitution (e).
The effect of the coefficient of friction (µ) is demonstrated in Figure 2.3. Consider the
same case as before but now the particle initially rotates as well (β
0
= 0). In the case of a
perfectly smooth particle (µ = 0) this rotation does not affect the translation motion of the
particle after collision. No energy is dissipated in this case. When the particle is not
perfectly smooth (µ > 0), which is always the case for granular material, the rotation does
affect both the translation and the rotation after collision as illustrated. In this case energy
is dissipated during the collision. For sticking collisions this effect is much more
pronounced than for sliding collisions.
Figure 2.3. The effect of the coefficient of friction (µ).
g
e = 1 e < 1
µ = 0
µ > 0
g
Chapter 2
______________________________________________________________________________________
40
The effect of the coefficient of tangential restitution is demonstrated in Figure 2.4. The
case is identical to the one above for the coefficient of friction except that now the
coefficient of friction is high enough to ensure that the collision is of the sticking type.
Figure 2.4. The effect of the coefficient of tangential restitution (β
0
).
In the case of β
0
= 0 the result is identical to the case described above provided the
collision is of the sticking type. The tangential component of the relative velocity at the
contact point is equal to zero at the end of the collision and energy is dissipated.
However, when β
0
= 1 the tangential component of the relative velocity at the contact
point reverses and the particle bounces back in the opposite direction. No energy is
dissipated in this case.
In section 5 experiments are described that enable determination of these collision
parameters from careful observation of single impacts.
2.3 Sequence of collisions
In the hardsphere model a constant time step DT is used to take the external forces acting
on the particles into account. Within this time step DT the velocities are assumed to
change only due to collisions and a sequence of collisions is processed one collision at a
time like in a regular hardsphere simulation in Molecular Dynamics. So in fact a separate
MD hardsphere simulation is performed within each time step. To do so it is necessary to
β
0
= 0
β
0
= 1
g
Granular Dynamics
______________________________________________________________________________________
41
determine what pair of particles will collide first which requires the determination of the
collision times of all relevant collision pairs. The collision time t
ab
of a pair of particles
(a,b) is defined as the time remaining until these particles will collide. It can be calculated
from the initial positions and velocities of both particles.
Figure 2.5. Determination of the collision time t
ab
.
When particles a and b collide the distance between the two centres of mass is equal to
the sum of the two radii as is shown in Figure 2.5. This yields a quadratic equation in t
ab
the smallest solution of which corresponds to collision (Allen and Tildesley, 1990):
( ) ( ) ( )
2
2 2 2 2
ab
b a ab ab ab ab ab ab
ab
v
R R r v
t
+ − − ⋅ − ⋅ −
·
v r v r
, (2.31)
where
b a ab
r r r − ≡ and
b a ab
v v v − ≡ (in this definition of v
ab
the particle rotation is not
taken into account unlike in equation 2.12). Note that if 0 > ⋅
ab ab
v r the particles are
moving away from each other and will not collide.
a
b
v
a
t
ab v
b
t
ab
R
a
+ R
b
r
ab
Chapter 2
______________________________________________________________________________________
42
In case of a collision with a wall the collision time follows simply from the distance to
the wall and the normal velocity component toward that wall which leads for a vertical
wall to the following expression:
a x
a x a wall
wall a
v
R x
t
,
,
,
) ( r − +
· . (2.32)
The algorithm used to process a sequence of collisions within a constant time step DT is
presented in Figure 2.6.
Figure 2.6. Computational strategy of a hardsphere simulation within a time step DT.
set up collision list
locate minimum collision time t
ab
increment acctim by t
ab
move (t
ab
)
reset collision lists
collision dynamics
locate minimum collision time t
ab
increment acctim by t
ab
move (DT(acctimt
ab
))
acctim < DT ?
yes
no
Granular Dynamics
______________________________________________________________________________________
43
First the collision lists are initialised in which for each particle a collision partner and a
corresponding collision time are stored. For each particle the smallest collision time is
determined by scanning all relevant collision partners. The variable acctim (accumulated
time) keeps track of the time spent since the beginning of the time step. In the routine
move(t
ab
) the collision times of all particles are reduced with t
ab
and the particle positions
are updated using a first order explicit integration:
ab a a ab a
t t t t v r r + · + ) ( ) ( . (2.33)
The calculation of the collision dynamics involves the collision model presented in the
previous paragraph. Subsequently the routine reset collision lists is entered where new
collision times and partners have to be found all the particles involved in the collision.
This does not only effect the particles a and b but also the particles that were about to
collide with either a or b. Finally a new collision pair has to be detected and acctim can
be incremented with the new collision time t
ab
. As soon as a minimum collision time is
found that after addition to acctim is greater than the time step DT, the loop is finished.
After the loop is finished the particles have to be moved forward until acctim equals DT.
During this motion no collision occurs.
2.4 Optimisation
To perform simulations of relatively large systems for relatively long times it is essential
to optimise the hardsphere computational strategy. The first step to achieve this is to
minimise the number of particles that have to be scanned for a possible collision by
employing a neighbour list. In the neighbour list of particle a all the particles that are
found within a square of size D
nblist
with particle a located at the centre, are stored. When
looking for a collision partner for particle a only the particles in the neighbour list need to
be scanned. In Figure 2.7 particle a is coloured black and its neighbour particles are
shaded. The neighbour lists are updated at each time step dt
nblist
.
Chapter 2
______________________________________________________________________________________
44
Figure 2.7. The neighbour list principle: all shaded particles are stored in the
neighbour list of the black particle.
When updating the neighbour list it is still not necessary to scan all particles. The solution
of the gas flow field (Chapter 3) requires the computational domain to be divided into
cells; for each cell the particles whose centre can be found in that cell are stored in a list.
When updating the neighbour list, only the cell where the particle’s centre is found and
the three nearest adjacent cells are scanned for possible neighbours.
With the implementation of the neighbour lists substantial speed up has been achieved
and simulations with over 1,000 particles are easily possible. The search for collision
partners (set up collision lists and reset collision lists in Figure 2.6) is no longer the most
CPU time consuming routine in the algorithm. Instead the update of the particle motion
(move in Figure 2.6) takes up 50 % of the CPU time in a typical fluidised bed simulation.
This can be optimised by applying a more efficient motion update strategy. In the original
version of the code all particles were moved to their new positions before each collision.
This implies N
particles
*N
collisions
motion updates per time step which means that particles
that do not collide are moved over a straight line in far too many steps as illustrated in
Figure 2.8.
D
nblist
Granular Dynamics
______________________________________________________________________________________
45
Figure 2.8. Efficient motion update.
Since the particles in Figure 2.8 collide only once, two motion updates will suffice for
both particles. In the previous version the particles were moved to their new position in a
total of seven steps for this specific example, that is five motion updates too many. This
does not only slow down the algorithm but also gives rise to numerical errors. In typical
fluidised bed simulations the total number of collisions per time step is of the same order
as the total number of particles (i.e. 10
4
10
5
) indicating that the total number of
unnecessary motion updates is of the same order as the total number of particles. In the
new algorithm special care must be taken when looking for new collision partners for
particles that just collided, since the positions stored in memory for the particles not
involved in that collision are not their actual positions. This causes some overhead for the
new algorithm but nonetheless the speed gain is substantial since the routine move(t
ab
) in
Figure 2.6 went down from 50% to less than 1% of the total amount of CPU time.
After optimising the motion update strategy the search for the smallest collision time
(locate minimum collision time in Figure 2.6) is the main CPU time consumer in the hard
sphere routine. This search has to be performed after each collision and in the old version
all particles were scanned and the smallest collision time was stored. In the new strategy
advantage is taken of the fact that the computational domain is already divided into cells
as illustrated in Figure 2.9. This figure is rather idealised for clarity but on average a
typical grid cell can contain up to 100 particles.
Chapter 2
______________________________________________________________________________________
46
Figure 2.9. Efficient search for smallest collision times using grid cells.
When for each grid cell the smallest collision time is stored it is not necessary to scan all
particles after each collision. Instead all the grid cells are scanned and since the total
number of grid cells is at least one order of magnitude smaller than the total number of
particles (i.e. 2,340 cells vs. 40,000 particles for the bubble formation simulations in
Chapter 5) this is much faster. Of course a new smallest collision time has to be found in
the grid cells containing particles that were involved in the last collision but this causes
negligible overhead.
After implementation of these optimised routines the main CPU time consumer is again
the search for possible collisions even though this search is performed only within the
neighbour list! A suitable choice of the size of the neighbour list and the time step for
updating the neighbour list is now critical. These choices however depend on the sort of
system that is simulated. For instance in bubbling beds (Chapters 4 and 5) a relatively
small neighbour list (D
nblist
= 3 D
p
) can be used that is updated every second time step.
For risers (Chapter 7) however a larger neighbour list has to be used (D
nblist
= 8 D
p
) since
the velocity differences between the particles are larger as well. If a neighbour list is
chosen to be too small it is possible that a collision is not detected and overlap between
Granular Dynamics
______________________________________________________________________________________
47
particles can occur. This cannot be tolerated in hardsphere simulations and if such an
overlap is detected in our code the simulation is stopped immediately. On the other hand
if the neighbour list is chosen to be rather large all collisions will be detected but this will
go at cost of the computational speed.
A hardsphere simulation is an event driven simulation which implies that the amount of
collisions to be processed per time step depends on the dynamics of the system. The
number of collisions to be processed can be considerably higher in a dense region of the
bed than in a dilute region. Hence the CPU time required to progress a time step can vary
significantly. It is therefore not straightforward to benefit from parallel computing with a
highly optimised event driven code.
3. SoftSphere approach
Although the hardsphere model was used for the majority of the simulations in this work
also a softsphere model was implemented. It was not the objective to select the best soft
sphere model available but the aim was to compare the results obtained with the most
popular softsphere model for fluidisation simulations with the results obtained with the
hardsphere model. The linear spring/dashpot model (Cundall and Strack, 1979) is the
most popular softsphere granular dynamics model since it was used by Tsuji et al.
(1993), Schwarzer (1995), Xu and Yu (1997), Kawaguchi et al. (1998) and Mikami et al.
(1998). For a review of various contact force models used in softsphere simulations the
reader is referred to Walton (1992) or Schäfer et al. (1996).
The softsphere model was implemented in 2D. In the vector notation employed in the
preceding paragraphs the zcomponent of the position and velocity vectors is equal to
zero and only rotation about the zaxis is considered. The notation and some definitions
can differ somewhat from the ones used for the hardsphere model. Unfortunately there is
no standard notation in Granular Dynamics (yet) and therefore it was attempted here to
stay close to the notations used in the references above.
Chapter 2
______________________________________________________________________________________
48
3.1 The linear spring/dashpot model
In softsphere models the following equations of motion are used:
external contact
dt
d
m F F
r
+ ·
2
2
, (2.34)
T ·
ω
dt
d
I , (2.35)
where r is the position vector of the centre of the particle, m is the mass of the particle,
F
contact
is the contact force acting on the particle, F
external
is the external force acting on
the particle, ω is the rotation velocity, T is the torque acting on the particle and I is the
moment of inertia of the particle as defined in equation 2.8. In this section the focus will
be on the contact forces between the particles, the external forces will be discussed in
section 6.
The particle velocities are updated using the accelerations from equations 2.34 and 2.35
by means of a first order explicit integration:
DT
0 0
v v v & + · , (2.36)
DT
0 0
ω ω ω & + · . (2.37)
The new particle positions are subsequently also obtained from a first order explicit
integration:
DT v r r + ·
0
, (2.38)
where the subscript
0
denotes the value at the previous time step.
Granular Dynamics
______________________________________________________________________________________
49
The interaction forces between particles in contact are modelled with a spring, a dashpot
and a friction slider, as shown in Figure 2.10.
spring
friction slider
dashpot
Figure 2.10. The linear spring/dashpot model.
The contact forces are evaluated from the overlap between the particles and their relative
velocities. Two particles a and b are in contact (i.e. have mutual overlap) if the distance
between their centres is less than the sum of their radii:
b p a p a b
R R
, ,
+ < −r r (2.39)
The repulsive force acting on the particles is divided into a normal (F
n
) and a tangential
(F
t
) component. The procedure of evaluating these contact forces starts with defining a
normal unit vector, pointing from the mass centre of particle a to the mass centre of
particle b as is presented in Figure 2.11.
Chapter 2
______________________________________________________________________________________
50
ω
b
ω
a
r
b
r
a

R
p,a
R
p,b
v
a
v
b
n
ab
t
ab
x
y
Figure 2.11. Definition of the coordinate system used in the softsphere model
Note that this definition differs from equation 2.1 used in the hardsphere model:
a b
a b
ab
r r
r r
n
−
−
· . (2.40)
The relative velocity of particle a with respect to b is given by:
ab b b a a b a ab
R R n v v v × ω + ω + − · ) ( ) ( . (2.41)
The normal component of the relative velocity is given by:
( )
ab ab ab ab n
n n v v ⋅ ·
,
. (2.42)
The tangential relative velocity, or slip velocity of particle a with respect to b can now be
obtained as follows:
Granular Dynamics
______________________________________________________________________________________
51
ab n ab t,ab ,
v v v − · (2.43)
The tangential unit vector is defined as follows:
ab t
ab t
ab
,
,
v
v
t · . (2.44)
By defining the tangential unit vector this way the vector is always pointing in the
direction of the slip velocity.
The overlap in the normal direction can immediately be calculated as the difference
between the sum of the particle radii and the distance between the particles:
( )
b a b a n
R R r r − − + · ξ (2.45)
The tangential displacement that has been established since the beginning of the contact
is obtained by integrating the relative velocities with respect to time:
( )
∫
·
t
t
ab t t
dt t
0
,
v ξ (2.46)
The contact forces are now given by:
ab n n ab n n n,ab
k
,
v n F η ξ − − · (2.47)
t,ab t t t t,ab
k v ? F η − − · (2.48)
If however the following relation is satisfied:
Chapter 2
______________________________________________________________________________________
52
ab n t,ab ,
F F µ > , (2.49)
then sliding occurs and the tangential force is given by:
ab n,ab t,ab
t F F µ − · . (2.50)
For contacts between particles and walls, the walls are assumed to be non moving and of
infinite mass just like in the hardsphere model.
The energy dissipated during contact between particles a and b can be calculated by
solving the following integral over the duration of the contact:
∫
⋅ − · dt E
ab ab ab dsp
v F
,
. (2.51)
Particle a can be in contact with several particles at the same time. Therefore the resulting
force and torque acting on particle a are obtained by summation of the forces with respect
to b:
( )
∑
+ ·
b
t,ab n,ab a contact
F F F
,
, (2.52)
( )
∑
× ·
b
t,ab ab a a
R F n T . (2.53)
Since the contact forces are in general at least an order of magnitude larger than the
external forces a separation of time scales was introduced. At each time step DT the
external forces were taken into account while at 0.1DT the equations of motion were
solved by taking only the contact forces into account.
It should be stressed here once again that the contact force model presented above was
used because it is the most popular model for fluidised bed simulations. Therefore it is
Granular Dynamics
______________________________________________________________________________________
53
the most useful model for a comparison with the hardsphere model. However, it is not
necessarily the best contact force available. In fact, the expression for the tangential
contact force (equation 2.48) can lead to very unrealistic behaviour if no special measures
are taken as will be demonstrated in the following paragraph.
3.2 Model parameters
The linear spring/dashpot model features three key parameters that will be discussed
here. The first is the spring stiffness (k) for both the linear normal and the linear
tangential spring. Measurements by Mullier et al. (1991) for 6 mm diameter cellulose
acetate spheres showed a linear dependency of the normal load upon the normal
displacement apart from a short initial stage (Walton, 1992). Therefore the use of the
linear spring can be justified for this type of material although it is one of the simplest
models available. For a more elaborate discussion on contact theory the reader is referred
to Johnson (1985). In most simulations the value of the spring stiffness is chosen rather
low for computational convenience. Tsuji et al. (1993), Kawaguchi et al. (1998) and
Mikami et al. (1998) all used a spring stiffness of 800 N/m for aluminum particles
whereas Xu and Yu (1997) used a much higher value of 50,000 N/m in their simulations.
The measurements by Mullier et al. (1991) suggest a value of about 400,000 N/m for the
spring stiffness of the 6 mm diameter cellulose acetate spheres. It is likely that the values
for the spring stiffness of 4 mm diameter aluminum spheres used in the work mentioned
above are even higher. At higher values for the spring stiffness the duration of the contact
becomes smaller and this requires a smaller time step to ensure numerical stability. From
equation 2.47 (harmonic oscillator with (weak) damping) a solution for the displacement
and relative velocity as a function of time can be obtained using the initial conditions (ξ
n
= 0 at t = 0, and v
n
= v
n0
at t = 0):
( )
( ) ( ) t B t B k B
B k B
v
t
n
n
n
n 2
2
2 2
2
2 2
0
5 . 0 exp 5 . 0 sin
5 . 0
) ( η η
η
ξ −
,
`
.

−
−
· , (2.54)
Chapter 2
______________________________________________________________________________________
54
From this solution the (normal) contact time between two particles can be found using ξ
n
= 0 at t = t
contact, n
, and v
n
= e v
n0
at t = t
contact, n
:
( )
n
n contact
k B
e
t
2
2
2
,
ln +
·
π
, (2.55)
where B
2
is defined in equation 2.16 and e is the normal coefficient of restitution defined
in equation 2.17. In a similar way a relation for the normal damping coefficient η
n
can be
obtained:
n contact
n
t B
e
, 2
ln 2 −
· η . (2.56)
Note that this result differs from the result presented by Mikami et al. (1998) which is
only valid for a particlewall contact.
In the case of (e = 0) the system is a critically damped harmonic oscillator and the contact
time will be infinite. In that case the following relation should be used to obtain the
damping coefficient η
n
:
2
2
B
k
n
n
· η . (2.57)
Xu and Yu (1997) presented a trial and error method to obtain the value of the damping
coefficient that is far less elegant. They simulated a particle wall collision similar to the
case presented in Figure 2.2 and adjusted the damping coefficient until the rebound
height of the particle matched the expected value for that particular coefficient of
restitution.
Granular Dynamics
______________________________________________________________________________________
55
Another important issue is the difference between the normal and the tangential spring
stiffness. In all softsphere simulations of fluidised beds reported in the literature the
tangential spring stiffness is chosen to be equal to the normal spring stiffness. This can
lead to energy inconsistencies as will be demonstrated here.
Figure 2.12. Binary collision between rough spheres.
Consider the case of a binary collision between rough particles (no sliding occurs) where
there is a significant tangential component of the relative velocity as depicted in Figure
2.12 (see also Figure 2.13a). When applying the same analysis as for the normal contact
time starting from equation 2.48 the following expression is found for the tangential
contact time:
( )
t
t contact
k B
t
1
2
0
2
,
ln β π +
· , (2.58)
where B
1
is defined in equation 2.15 and β
0
is the coefficient of tangential restitution
defined in equation 2.19. Since B
1
and B
2
are not equal and e and β
0
are not necessarily
equal as well, the following relation must be satisfied to ensure that the normal and
tangential contact times are equal:
( )
( )
n t
k
e
k
,
`
.

+
+
·
2
2
2
0
2
ln
ln
7
2
π
β π
. (2.59)
If this relation is not satisfied the tangential spring will still be loaded at the instant of
detachment. This implies that energy is stored in the spring that causes a defect in the
a b
Chapter 2
______________________________________________________________________________________
56
energy balance over the contact period. In the following section this will be
demonstrated.
For the tangential damping coefficient the following result is obtained:
t contact
t
t B
, 1
0
ln 2 β
η
−
· . (2.60)
Note that also the tangential damping coefficient differs from the normal damping
coefficient. This has never been addressed before in any of the softsphere simulations of
gas fluidised beds presented in the literature.
In the case of (β
0
= 0) the system is a critically damped harmonic oscillator and the
contact time will be infinite. In that case the following relation should be used to obtain
the tangential damping coefficient η
t
:
1
7
4
B
k
t
t
· η . (2.61)
The coefficient of friction (µ) plays the same role in the linear spring/dashpot model as it
does in the hardsphere model. It can be obtained from experiments as will be explained
in section 5.
4. HardSphere vs. SoftSphere
For the majority of the simulations presented in this work the hardsphere model is used.
However, there are certain cases where a softsphere model is to be preferred. In the
following paragraphs the key differences between the two approaches will be highlighted.
Granular Dynamics
______________________________________________________________________________________
57
4.1 Static situations
For dynamic systems in general a hardsphere simulation is computationally faster than a
softsphere simulation as long as there is sufficient motion in the system and the void
fraction does not become too low. Severe problems are encountered when static situations
like for example de fluidisation occur. The particles become very closely packed with
very low relative velocities. The number of collisions then increases exponentially whilst
the impulse concerned with a single collision becomes negligible. Despite of the double
precision accuracy used in the simulations the particles eventually overlap (although this
overlap is far smaller than the atomic length scale) due to number loss and the algorithm
breaks down. A softsphere model has no problems in dealing with static situations and is
therefore to be preferred for simulations where such situations can occur. This includes
systems where attractive forces between particles (like for example the liquid bridge
forces in the simulations by Mikami et al. 1998) play an important role.
4.2 Spring stiffness
In the limit of an infinitely high value of the spring stiffness the softsphere model meets
the hardsphere model. However the softsphere model cannot be run at very high values
of the spring stiffness due to computational limitations. The assumption of instantaneous
collisions in the hardsphere model does not capture the whole physics of the collision
process but neither does the assumption of using an artificially low value for the spring
stiffness in the softsphere model. It is important to realise that where a softsphere model
in principle is capable of predicting the contact time between particles (which is of great
importance for studying heat and mass transfer phenomena in gas fluidised beds) these
predictions have no value when an artificially low spring stiffness is used.
Furthermore there is the issue of the tangential spring stiffness versus the normal spring
stiffness. In the vast majority of the linear spring/dashpot softsphere models presented
in the literature the tangential spring stiffness is chosen to be equal to the normal spring
stiffness. This can lead to energy inconsistencies as will be demonstrated here. Consider
the case presented in Figure 2.12. where two particles collide perfectly elastically and
Chapter 2
______________________________________________________________________________________
58
perfectly rough (e = 1, β
0
= 1 and µ is large enough to ensure a sticking collision).
Particle a moves at a velocity of 0.01 m/s and a rotation velocity of 1.0 1/s toward
particle b which does not rotate and moves at a velocity of –0.01 m/s. In Figure 2.13a the
rotation velocity of particle a is plotted as a function of time for three cases: a hardsphere
simulation, a softsphere simulation with a tangential spring stiffness (k
t
) at 2/7 of the
normal spring stiffness (k
n
) and a softsphere simulation with k
t
= k
n
.
Figure 2.13. a) Rotation velocity of particle a as a function of time. b) Tangential
contact forces as a function of time.
It is clear that in the case where k
t
= k
n
the rotation velocity after collision is not predicted
correctly. From Figure 2.13b) it can be seen that the tangential contact force is nonzero
at the end of collision implying that not all the energy that is put into the spring is
recovered. Hence energy inconsistencies can arise when k
t
is chosen to be equal to k
n
. It
should be noted however that these inconsistencies are rather small since in general the
rotation energy is two orders of magnitude smaller than the kinetic energy.
The particles in these simulations were 4 mm diameter aluminium spheres (ρ = 2700
kg/m
3
) as previously used by Tsuji et al. (1993) and a value of 1000 kg/m was used for
the normal spring stiffness. A time step of 10
7
s was used for the numerical integration.
Granular Dynamics
______________________________________________________________________________________
59
4.3 Energy considerations
The main advantage of the hardsphere model is that there is an analytical solution
available for the collision model. Given the velocities of the particles prior to collision
together with the particle properties the postcollision velocities can be calculated
exactly. Hence numerical errors cannot cause any problems concerning energy
conservation. This is not guaranteed for softsphere models where the accuracy of the
solution depends on the time step used for the numerical integration. The time step should
always be chosen carefully in such simulations to avoid energy inconsistencies. Consider
the same case as in the previous paragraph (see Figure 2.12) but now none of the particles
do rotate. A spring stiffness of 800 kg/m was chosen in order to match the parameter
settings used by Tsuji et al. (1993) and Kawaguchi et al. (1998). For this binary headon
collision with the coefficient of (normal) restitution (e) set equal to 0.9 a contact time of
7.47 10
4
s can be calculated using equation 2.55. Tsuji et al. (1993) and Kawaguchi et al.
(1998) used a time step of 2 10
4
s in their simulations with these particles. Mikami et al.
(1998) performed simulations of a different system but used the same ratio between
contact time and time step of 3.7. In Figure 2.14 the effect of the time step on the
resolution of the normal contact force as a function of time is presented.
Figure 2.14. The normal contact force as a function of time for a binary headon
collision at two different time steps (e = 0.9).
Chapter 2
______________________________________________________________________________________
60
It can be seen that the curve for the case with biggest time step (2 10
4
s) is by far not as
smooth as the curve for the case with the smallest time step (1 10
5
s). However this
results in a relative error in the energy balance that is surprisingly low (1.64 % for the 2
10
4
s case and 0.05% for the 1 10
5
s case). Nonetheless one should keep in mind that in a
typical fluidised bed simulation a lot of collisions occur and this error can accumulate
quite rapidly. Another typical feature of the linear spring/dashpot model can be observed
in this figure as well: in the final stage the contact force becomes attractive instead of
repulsive (DT = 10
5
s). This is due to the damping term in equation 2.47: at the end of
the contact, when the overlap approaches zero, the relative velocity increases and hence
the damping term causes the contact force to become negative (i.e. attractive).
Another error arises due to the determination of contact between particles at a constant
time step. In Figure 2.14 this can be observed very well since the time at which the
particles first come into contact is not the same for both time steps. The strategy used by
Xu and Yu (1997) who determine the instant of first contact precisely by employing a
collision search algorithm from hardsphere simulations would prevent this. However the
additional search for first contact goes at a cost of the computational speed whereas the
actual integration is still performed at a rather large time step. Improvements can be
achieved by employing higher order numerical schemes.
4.4 Multiple particle interactions
A drawback of the hardsphere model is the assumption that the particles interact via
instantaneous binary collisions. Consider the case presented in Figure 2.15 where a
particle approaches two other particles at constant velocity in the absence of an external
force. The particles are positioned in such a way that a three body collision should occur.
The collision times for both pairs of particles are identical. In the hardsphere routine one
of the two possible collisions is detected as the first collision to occur and this binary
collision (in this case a perfectly elastic, perfectly smooth collision) is processed first.
The second collision is processed immediately afterwards where the postcollision
Granular Dynamics
______________________________________________________________________________________
61
velocity of the first collision is used as the initial velocity for the second collision. This
renders the rebound pattern shown in Figure 2.15. which is physically not correct.
Figure 2.15. Threebody collision simulated with a hardsphere model.
When the same case is repeated with the softsphere model the result presented in Figure
2.16 is obtained.
Figure 2.16. Threebody collision simulated with a softsphere model.
Because the softsphere model is capable of handling multiple particle interactions the
rebound pattern is predicted correctly here. It should be noted that this specific test case
Chapter 2
______________________________________________________________________________________
62
was carefully selected to demonstrate the effect of the assumption of instantaneous,
binary collisions.
5. Measurement of collision parameters
In the hardsphere model used in this work three parameters are required to describe to
describe a collision: the coefficient of restitution (e), the coefficient of tangential
restitution (β
0
) and the coefficient of friction (µ). These three parameters are also required
for the linear spring/dashpot model besides the spring stiffness. Unfortunately
experimental data for these collision parameters are scarce in the open literature. This is
rather surprising since rigidbody impact was already a research subject in the days of
Galilei (1638) and the concept of the coefficient of restitution was introduced by Newton
(1686) accompanied by the first experimental data. The book by Goldsmith (1960)
provides a good review of both theory and experiments concerning impacts.
Measurements of the coefficient of restitution are reported there as well including the
dependency of this parameter upon impact velocity and particle size. In more recent years
the attention was focussed on the impact of particles on flat plates (Maw et al., 1981 and
Sondergaard et al., 1990) where more detailed data was obtained.
It was not until the last decade that also particleparticle collisions became a subject of
experimental investigations. Foerster et al. (1994) presented a method to obtain the three
collision parameters (e, µ and β
0
) by careful experimentation and image processing. The
difficulty in these experiments lies in the control of the geometry of the impact. Lorenz et
al. (1997) presented a refinement of this technique together with additional data for more
types of particles then the cellulose acetate and soda lime glass spheres used by Foerster
et al. 1994). Bernasconi et al. (1997) presented an advanced image processing technique
employing three synchronised CCD cameras that enables accurate measurements of the
particle position as well as the orientation of the particles. Labous et al. (1997) presented
an experimental technique to measure collisional properties of spheres using highspeed
video analysis. They studied the dependency of the coefficient of restitution on impact
Granular Dynamics
______________________________________________________________________________________
63
velocity and particle mass. Despite the fact that quite a number of research groups are
capable of measuring collision properties there is still a lack of collision data in the
literature.
The Impact Research Group of the Open University at Milton Keynes has developed an
accurate technique to measure collision properties. The collision properties of the glass
particles used in simulations reported in Chapters 6 and 8 were measured at this facility.
The technique to measure particle impact parameters accurately was first developed for
particle impacts on flat plates and is described in detail by Kharaz et al. (1999). This set
up was used to measure the particlewall impact parameters used for a simulation
reported in Chapter 6. To enable measurement of particleparticle impact parameters
some adjustments to the set up were required that are reported by Gorham and Kharaz
(1999) and will be briefly outlined here. The setup used for the measurements is
schematically represented in Figure 2.17.
a) b)
Figure 2.17. Schematic representation of the experimental set up used for the
measurement of particleparticle collision properties: a) before collision
b) after collision.
pivot
Falling particle
Stationary particle
Lever
arm
Solenoid
Triggering device
CCD camera
Chapter 2
______________________________________________________________________________________
64
A particle is released from a vacuum nozzle with zero rotation. It then falls through an
optical fibre triggering device which generates a sequence of pulses to control the CCD
camera, the strobe light and the solenoid (Figure 2.17a). The leverarm is then quickly
moved away at 0.3 ms before collision leaving the stationary particle entirely free and in
the correct position for a collision (Figure 2.17b). By following this procedure the particle
can be assumed to be non moving, non rotating and still at the same initial position. The
collision is then recorded on a single frame of the CCD camera using strobe flashes that
are timed in such a way that particle images do not overlap. This is required for accurate
determination of the particle positions from the recording by digital image processing
which finally yields the particle velocities before and after collision.
The coefficient of restitution (e) can easily be obtained from the normal components of
the relative velocity before and after impact. Since the particles do not rotate before
collision it is not necessary to know the particle rotation after impact since the velocity
after impact (v
ab
), which includes a rotation contribution, can be calculated using
equations 2.6 and 2.14. The two other collision parameters can now be obtained from a
plot of Ψ
after
as a function of Ψ
before
that are defined as follows:
( )
( ) n v
t v
⋅
⋅ −
· Ψ
0 , ab
ab
after
, (2.62)
( )
( ) n v
t v
⋅
⋅ −
· Ψ
0 ,
0 ,
ab
ab
before
. (2.63)
Such a plot is schematically represented in Figure 2.18. Each data point represents one
collision measurement where the impact velocity was about 1.0 m/s.
Two straight lines are fitted through the data points: one for the sticking regime and one
for the sliding regime. The coefficient of tangential restitution can be obtained from the
slope of the sticking line:
after before
Ψ − · Ψ
0
β , (2.64)
Granular Dynamics
______________________________________________________________________________________
65
and the coefficient of friction can be obtained from the intercept of the sliding line:
( )µ e
after before
+ − Ψ · Ψ 1
2
7
. (2.65)
Figure 2.18. Schematic representation of a plot of Ψ
after
vs. Ψ
before
for a typical result
of a measurement.
Each individual measurement yields either a value for e and β
0,
or e and µ depending on
the collision regime. Nevertheless a series of measurements is required to obtain the two
linear fits that determine the two regimes. Typical results obtained with this technique for
glass particles show the following values for the collision parameters: e = 0.97 t0.01, µ =
0.1 t0.01 and β
0
= 0.33 t0.05. For nearly headon collisions (low values of Ψ
before
) the
measurements show positive values of Ψ
after
. This phenomenon is referred to as micro
slip and cannot be explained by our hardsphere model. The model developed by Maw et
al. (1976) however is capable of predicting this phenomenon. Nevertheless it is believed
Ψ
after
Ψ
before
sticking
sliding
Chapter 2
______________________________________________________________________________________
66
that the threeparameter hardsphere collision model contains sufficient physics to be
capable of reliable predictions.
6. External Forces
The incorporation of external forces differs somewhat from the approach followed by
Hoomans et al. (1996). In this work the external forces are used in accordance with those
implemented in the two fluid model presented by Kuipers et al. (1992) where, of course,
the forces now act on a single particle:
( )
( ) m
d
dt
m
V
V p
p
p
p
p
p p
v
g u v · +
−
− − ∇
β
ε 1
(2.66)
where m
p
represents the mass of a particle, v
p
its velocity, u the local gas velocity and V
p
the volume of a particle. A similar equation of motion was used by Kawaguchi et al.
(1998). In equation 2.66 the first term on the right hand side is due to gravity. The second
term is due to the drag force where β represents an interphase momentum exchange
coefficient as it usually appears in twofluid models. For low void fractions (ε < 0.80) β is
obtained from the wellknown Ergun equation:
( )
( )
p
p
g
p
g
D D
v u − − +
−
·
ρ
ε
µ
ε
ε
β 1 75 . 1
1
150
2
2
(2.67)
where D
p
represents the particle diameter, µ
g
the viscosity of the gas and ρ
g
the density of
the gas. For high void fractions (ε ≥ 0.80) the following expression for the interphase
momentum transfer coefficient has been used which is basically the correlation presented
by Wen and Yu (1966) who extended the work of Richardson and Zaki (1954):
Granular Dynamics
______________________________________________________________________________________
67
( )
65 . 2
1
4
3
−
−
−
· ε ρ
ε ε
β
p g
p
d
D
C v u (2.68)
The drag coefficient C
d
is a function of the particle Reynolds number:
( )
¹
¹
¹
'
¹
≥
<
·
1000 Re 44 . 0
1000 Re Re 0.15 + 1
Re
24
0.687
p
p p
p
d
C
(2.69)
where the particle Reynolds number in this case is defined as:
g
p p g
p
D
µ
ερ v u −
· Re . (2.70)
The pressure gradient in the third term on the right hand side of equation 2.66 is
calculated by using a first order approximation. The local value was obtained from an
area weighted averaging technique using the values of the pressure gradients at the four
surrounding grid nodes. This technique is also used to obtain local gas velocities and
local void fractions at the position of the centre of the particle. The area weighted
averaging technique used to obtain the local averaged value Q of a quantity Q(i,j) from
the four surrounding computational nodes is shown in Figure 2.19. The local averaged
value is calculated as follows:
Q
A Q A Q A Q A Q
DXDY
i j i j ii j ii j ii jj ii jj i jj i jj
·
+ + +
, , . , , , , ,
(2.71)
where:
( )
( )
( )
( )
A DX DY
A DY
A
A DX
i j x y
ii j x y
ii jj x y
i jj x y
,
,
,
,
· − −
· −
·
· −
δ δ
δ δ
δ δ
δ δ
(2.72)
Chapter 2
______________________________________________________________________________________
68
The distances δ
x
and δ
y
, necessary in this averaging technique, are calculated from the
position of the particle in the staggered grid.
Figure 2.19. Area weighted averaging.
For the integration of equation 2.66 an explicit first order scheme is used to update the
velocities and the positions of the particles. Other external forces than the ones included
in equation 2.66 can be taken into account as well. In Chapter 7 simulations of more
dilute flows in the riser section of a circulating fluidised bed are presented where lift
forces were also taken into account.
Notation
B
1
, B
2
collision constants, 1/kg
C
d
drag coefficient, []
D
nblist
diameter of neighbour square
DX
A
i,j
A
i,jj
A
ii,jj
A
ii,j
δ
x
δ
y
DY
Q
ii,j
Q
i,j
Q
ii,jj
Q
i,jj
Q
position for the
evaluation of
A Area
Granular Dynamics
______________________________________________________________________________________
69
D
p
particle diameter, m
dt
nblist
time step for neighbour list update, s
DT time step, s
e coefficient of restitution, []
E energy, J
F force, N
g gravitational acceleration, m/s
2
I moment of inertia, kgm
2
J impulse vector, kgm/s
k spring stiffness, N/m
m particle mass, kg
n normal unit vector, []
n
tot
total number of particles, []
n
coll
total number of collisions, []
R
p
particle radius, m
r position, m
T torque, Nm
t tangential unit vector, []
t time, s
t
ab
collision time, s
u gas phase velocity, m/s
v velocity, m/s
Greek symbols
β volumetric interphase momentum transfer coefficient, kg/(m
3
s)
β
0
coefficient of tangential restitution, []
δ distance, m
ε void fraction, []
µ
g
gas shear viscosity, kg/(ms)
µ friction coefficient, []
η damping coefficient, Ns/m
Chapter 2
______________________________________________________________________________________
70
ρ density, kg/m
3
ω angular velocity, 1/s
ξ displacement, m
Ψ defined in equations 2.62 and 2.63, []
Subscripts
0 initial condition
a,b particle indices
av average
cp contact point
dsp dissipated
gyr gyration
nblist neighbour list
p particle
w wall
References
Alder, B.J. and Wainwright, T.E., (1957). Phase Transition for a HardSphere System, J.
Chem. Phys. 27, 1208.
Alder, B.J. and Wainwright, T.E., (1959). Studies in Molecular Dynamics. I. General
Method, J. Chem. Phys. 31, 459.
Allen, M.P. and Tildesley, D.J., (1990). Computer simulations of liquids, Oxford Science
Publications, Oxford, UK.
Bagnold, R.A., (1954). Experiments on a gravity free dispersion of large solid particles in
a Newtonian fluid under shear, Proc. R. Soc. Lond., A225, 49.
Granular Dynamics
______________________________________________________________________________________
71
Bernasconi, D.J., Fandrich, M.E. and Hogue, C., (1997). Experimental 3D Impact
Analysis. Multibody System Dynamics, 1, 361.
Bird, G.A., (1976). Molecular Gas Dynamics and Direct Simulation of Gas Flows,
Oxford University Press, Oxford, UK.
Campbell, C.S. and Brennen, C.E., (1985). Computer simulations of granular shear flows,
J.Fluid. Mech. 151, 167.
Cundall, P.A. and Strack, O.D.L., (1979). A discrete numerical model for granular
assemblies. Géotechnique, 29, 47.
de Gennes, P.G., (1999). Granular matter: a tentative view, Rev. Mod. Phys., 71, S374.
Foerster, S.F., Louge, M.Y., Chang, H. and Allia, K., (1994). Measurements of the
collision properties of small spheres. Phys. Fluids, 6 (3), 1108.
Frank, Th., Schade, K.P. and Petrak, D., (1992). Numerical simulation and experimental
investigation of a gassolid twophase flow in a horizontal channel, Int. J. Multiphase
Flow, 19, 187.
Frenkel, D. and Smit, B., (1996). Understanding Molecular Simulation, from Algorithms
to Applications, Academic Press, Inc., San Diego, USA.
Frezzotti, A., (1997). A particle scheme for the numerical solution of the Enskog
equation, Phys. Fluids, 9 (5), 1329.
Galilei, G., (1638). Unterredungen und Demonstrationen über zwei neue Wissenszweige,
Oswald’s Klassiker 25, 38, Leipzig, W. Engelmann, 189091.
Chapter 2
______________________________________________________________________________________
72
Ge, W. and Li, J., (1997). PseudoParticle Approach to Hydrodynamics of Gas/Solid
TwoPhase Flow, In Proceedings of the 5
th
Int. Conf. Circulating Fluidized beds, Beijing
(China), May 28 June 1, 1996, DT8, 1.
Goldsmith, W., (1960). Impact, the theory and physical behaviour of colliding solids, E.
Arnold Pub., London.
Gorham, D.A. and Kharaz, A.H., (1999). Results of particle impact tests, Impact
Research Group Report IRG 13, The Open University, Milton Keynes, UK.
Hogue, C. and Newland, D., (1994). Efficient computer simulation of moving granular
particles, Powder Technol. 78, 51.
Hoomans, B.P.B., Kuipers, J.A.M., Briels, W.J. and van Swaaij, W.P.M., (1996).
Discrete particle simulation of bubble and slug formation in a twodimensional gas
fluidised bed: a hardsphere approach, Chem. Engng Sci., 51, 99.
Hopkins, M.A. and Louge, M.Y., (1991). Inelastic microstructure in rapid granular flows
of smooth disks, Phys. Fluids A 3(1), 47.
Ichiki, K. and Hayakawa, H., (1995). Dynamical simulation of fluidized beds:
hydrodynamically interacting granular particles, Phys. Rev. E. 52, 658.
Jaeger, H.M., Nagel, S.R. and Behringer, R.P., (1996). Granular solids, liquids and gases,
Rev. Mod. Phys., 68, 1259.
Johnson, K.L., (1985). Contact Mechanics, Cambridge University Press, Cambridge, UK.
Kawaguchi, T., Tanaka, T. and Tsuji, Y., (1998). Numerical simulation of two
dimensional fluidized beds using the discrete element method (comparison between the
two and threedimensional models), Powder Technol. 96, 129.
Granular Dynamics
______________________________________________________________________________________
73
Kharaz, A.H., Gorham, D.A., and Salman, A.D., (1999). Accurate measurement of
particle impact parameters, Measurement Science and Technology 10, 31.
Knowlton, T.M., Carson, J.W., Klinzing, G.E. and Yang, W.C., (1994). The importance
of storage, transfer and collection, Chem. Eng. Progress. 90 (4), 44.
Kuipers, J.A.M., van Duin K.J., van Beckum, F.P.H. and van Swaaij, W.P.M., (1992). A
numerical model of gasfluidized beds, Chem. Engng Sci., 47, 1913
Labous, L., Rosato, A.D. and Dave, R.N., (1997). Measurements of collisional properties
of spheres using highspeed video analysis, Phys. Rev. E. 56, 5717.
Langston, P.A., Tüzün U. and Heyes, D.M., (1994). Continuous potential discrete particle
simulations of stress and velocity fields in hoppers transition from fluid to granular flow,
Chem. Engng Sci. 49, 1259.
Langston, P.A., Tüzün, U. and Heyes, D.M., (1995). Discrete element simulation of
granular flow in 2D and 3D hoppers: dependence of discharge rate and wall stress on
particle interactions. Chem. Engng Sci. 50, 967.
Lorenz, A., Tuozzolo, C. and Louge, M.Y., (1997). Measurement of impact properties of
small, nearly spherical particles. Exp. Mech. 37, 292.
Lun, C.C.K., (1996). Granular dynamics of slightly inelastic spheres in Couette flow,
Phys. Fluids 8, 2868.
Lun, C.C.K., and Liu, H.S., (1997). Numerical simulation of dilute turbulent gassolid
flows in horizontal channels, Int. J. Multiphase Flow 23, 575.
Chapter 2
______________________________________________________________________________________
74
Marin, M., Risso, D. and Cordero, P., (1993). Efficient algorithms for manybody hard
particle molecular dynamics, J. Comput. Phys. 109, 306.
Maw, N., Barber, J.R. and Fawcett, J.N., (1976). The oblique impact of elastic spheres,
Wear 38, 101.
Maw, N., Barber, J.R. and Fawcett, J.N., (1981). The role of elastic tangential compliance
in oblique impact, Trans. ASME: J. Lubrication Tech. 103, 74.
McCarthy, J.J., and Ottino, J.M., (1998). Particle dynamics simulation: a hybrid
technique applied to granular mixing, Powder Technol. 97, 91.
Mikami, T., Kamiya, H. and Horio, M., (1998). Numerical simulation of cohesive powder
behavior in a fluidized bed, Chem. Engng Sci. 53, 1927.
Mikami, T., (1998). Agglomeration fluidization of liquid/solid bridging particles an its
control, PhD thesis, Tokyo University of Agriculture and Technology, Department of
Chemical Engineering, Japan.
Müller, D. and Liebling, Th.M., (1995). Detection of collisions of polygons by using a
triangulation, In Contact Mechanics (M. Raous, M. Jean and J.J. Moreau eds.),
proceedings of the 2
nd
Contact Mechanics International Symposium, September 1923,
Carry le –Rouet, France, p. 369.
Mullier, M. Tüzün, U. and Walton, O.R., (1991). A singleparticle friction cell for
measuring contact frictional properties of granular materials, Powder Technol. 65, 61.
Newton. I., (1686). Philosophiae naturalis principia mathematica, (Translated into
English by A. Motte). New York, D. Adee, 1848.
Granular Dynamics
______________________________________________________________________________________
75
Oesterlé, B. and Petitjean, A., (1993). Simulation of particletoparticle interactions in
gassolid flows, Int. J. Multiphase flow 19, 191.
Ouyang, J. and Li, J. (1999). Particle motionresolved discrete model for simulating gas
solid fluidization, Chem. Engng Sci. 54, 2077.
Richardson, J.F. and Zaki, W.N., (1954). Sedimentation and fluidization: part I. Trans.
Inst. Chem. Eng. 32, 35.
Rosato, A., Prinz, F., Standburg, K.J. and Swendsen, R., (1986). Monte Carlo Simulation
of particulate matter segregation, Powder Technol. 49, 59.
Schäfer, J., Dippel, S. and Wolf, D.E., (1996). Force schemes in simulations of granular
materials, J. Phys. I France 6, 5.
Schwarzer, S., (1995). Sedimentation and flow through porous media: Simulating
dynamically coupled discrete and continuum phase, Phys. Rev. E. 52, 6461.
Seibert, K.D. and Burns, M.A., (1996). Simulation of fluidized beds and other fluid
particle systems using statistical mechanics, AIChE J. 42, 660.
Seibert, K.D. and Burns, M.A., (1998). Simulation of structural phenomena in mixed
particle fluidized beds, AIChE J. 44, 528.
Sommerfeld, M., (1990). Numerical simulation of the particle dispersion in turbulent
flow: the importance of particle lift forces ad particle/wall collision models, In Numerical
Methods for Multiphase Flows, Vol. 91. ASME, New York.
Sondergaard, R., Chaney, K. and Brennen, C.E., (1990). Measurements of solid spheres
bouncing off flat plates, J. Appl. Mech. 57, 694.
Chapter 2
______________________________________________________________________________________
76
Stronge, W.J., (1990). Rigid body collisions with friction, Proc. R. Soc. Lond. A. 431,
169.
Thornton, C., (1997). Coefficient of restitution for collinear collisions of elasticperfectly
plastic spheres, J. Appl. Mech. 64, 383.
Tsuji, Y., Morikawa, Y., Tanaka. T., Nakatsukasa, N. and Nakatani, N., (1987).
Numerical simulation of gassolid twophase flow in a twodimensional horizontal
channel, Int. J. Multiphase Flow 13, 671.
Tsuji, Y., Tanaka, T. and Ishida, T., (1992). Lagrangian numerical simulation of plug
flow of cohesionless particles in a horizontal pipe, Powder Technol. 71, 239.
Tsuji, Y., Kawaguchi, T. and Tanaka, T., (1993). Discrete particle simulation of two
dimensional fluidized bed, Powder Technol. 77, 79.
Tsuji, Y., Tanaka, T. and Yonemura, S., (1998). Cluster patterns in circulating fluidized
beds predicted by numerical simulation (discrete particle model versus two fluid model),
Powder Technol. 95, pp. 254264.
Walton, O.R. and Braun, R.L., (1986). Viscosity and temperature calculations for
assemblies of inelastic frictional disks, Journal of Rheology 30, 949.
Walton, O.R., (1992). Numerical simulation of inelastic, frictional particleparticle
interactions, in Particulate TwoPhase Flow, Part I, edited by M.C. Roco, Butterworth
Heinemann, Boston, 884.
Wang, Y. and Mason, M.T., (1992). Twodimensional rigidbody collisions with friction.
J. Appl. Mech. 59, 635.
Granular Dynamics
______________________________________________________________________________________
77
Wen, C.Y. and Yu, Y.H., (1966). Mechanics of fluidization. Chem. Eng. Prog. Symp. Ser.
62 (62), 100.
Wightman, C., Moakher, M., Muzzio, F.J. and Walton, O., (1998). Simulation of flow
and mixing of particles in a rotating and rocking cylinder, AIChE J. 44, 1266.
Xu, B.H. and Yu, A.B., (1997). Numerical simulation of the gassolid flow in a fluidized
bed by combining discrete particle method with computational fluid dynamics. Chem.
Engng Sci. 52, 2785.
Yonemura, S., Tanaka, T. and Tsuji, Y., (1993). Cluster formation in gassolid flow
predicted by the DSMC method, In Proc. ASME/FED Symp. No. 166, ASME, New York,
USA, 303.
Chapter 2
______________________________________________________________________________________
78
Gas Phase Hydrodynamics
______________________________________________________________________________________
79
Chapter 3.
GAS PHASE HYDRODYNAMICS
Abstract:
In this chapter the model for the gas phase hydrodynamics is presented. The two
dimensional model is based on the volume averaged NavierStokes equations for two
phase flow and is solved on a scale larger than the particle size. This requires the drag
force exerted on the particles by the gas phase to be incorporated in the model through
empirical relations. A short review of techniques that allow the flow around each
individual particle to be calculated, and thus do not require any empirical input
concerning fluidparticle drag, is presented as well. The numerical solution technique
and the applied boundary conditions are briefly outlined. Twoway coupling between the
motion of the particles and the motion of the gasphase is established via the void
fraction and a source term in the momentum conservation equation for the gas phase. A
detailed description of the calculation of the void fraction from both the 2D and the 3D
granular dynamics model is presented. The incorporation of the momentum source term
in the gas phase hydrodynamics model is explained as well.
Chapter 3
______________________________________________________________________________________
80
Parts of this chapter are based on the papers:
B.P.B. Hoomans, J.A.M. Kuipers, W.J. Briels and W.P.M. van Swaaij, (1996). Discrete particle simulation
of bubble and slug formation in twodimensional gasfluidised beds: a hardsphere approach, Chem. Engng
Sci., 51, 99.
B.P.B. Hoomans, J.A.M. Kuipers, W.J. Briels and W.P.M. van Swaaij, (1998). Comments on the paper
‘Numerical simulation of the gassolid flow in a fluidized bed by combining discrete particle method with
computational fluid dynamics’ by B.H. Xu and A.B. Yu, Chem. Engng Sci., 54, 2645.
Gas Phase Hydrodynamics
______________________________________________________________________________________
81
1. Introduction
The drag force exerted on a particle by the gas phase requires the velocity of the gas
phase at the position of the particle. In the approach used in this work the gas phase
hydrodynamics is resolved on a length scale that is larger than the particle size. Hence
empirical relations are required to determine the drag force exerted on the particles by the
gas phase. This implies that special care has to taken to achieve correct twoway
coupling. The technique adopted in this work is reminiscent of the ParticleSourceIn Cell
method developed by Crowe et al. (1977). An alternative for this method was presented
by Pan and Banerjee (1996) where the particle effects on the fluid motion were fed back
by calculating the velocity disturbance caused by the particles assuming that the flow
around them is locally Stokesian. This technique is however rather expensive in terms of
CPU time and is limited to rather dilute flows of particles that are slightly heavier than
fluid.
The approach followed in this work is described in the preceding sections. However, it is
worthwhile to discuss some techniques here that are capable of solving the flow field
around each individual particle and therefore enable the calculation of the force exerted
on the particle by the fluid without any empirical input. This can provide improved
closure relations for the drag force that can be used in Eulerian Lagrangian simulations
where the flow field is resolved on a scale larger than the particle diameter.
1.1 Direct solution of the NavierStokes equations
When the flow around a particle is solved using a noslip boundary condition on the
particle surface, the drag force exerted on the particle can be obtained by integration of
the stress tensor over the particle surface. There are several methods available to solve the
NavierStokes equations in such a complex geometry. Guj and de Matteis (1986)
presented a technique to solve the twodimensional flow field around an arrangement of
particles using boundary fitted coordinates. Hu (1996) used a finite element technique
Chapter 3
______________________________________________________________________________________
82
based on moving unstructured grids to simulate the twodimensional motion of solid
particles in a liquid. Kalthoff et al. (1997) developed a technique where a finite difference
NavierStokes solver is used to solve the twodimensional liquid motion. Since the grid
size is smaller than the particle diameter a forcing technique was used to ensure noslip
boundary conditions on the particle surface. The drag force exerted on a particle was then
obtained by numerical integration of the stress tensor over the particle surface using an
analytical expansion of the fluid field.
Although these techniques provide a lot of detail without requiring any empirical input,
the number of particles that can be used in these simulations is still limited (10
2
10
3
).
Furthermore these are still twodimensional simulations and hence threedimensional
models would put even higher demands on computing power.
1.2 Lattice Boltzmann simulations
A flow simulation technique that has become very popular in recent years is the lattice
Boltzmann technique pioneered by McNamara and Zanetti (1988). In this technique a
oneparticle velocity distribution function rather than ‘real particles’ (as in lattice gas
cellular automata as described by Frisch et al., 1986) is transported on a carefully
selected lattice. A time step in a lattice Boltzmann simulation consists of two phases: a
collision phase and a propagation phase. The transport properties of the system (for
example the viscosity) are defined by the eigenvalues of the collision operator used in the
collision phase. This technique is capable of dealing with complex geometries which is
important for the majority of flows encountered in industry. Eggels and Somers (1995)
extended the technique by including a turbulent stress tensor as well as a scalar transport
equation. Eggels (1996) used this model to perform largeeddy simulations of baffled
stirred tank reactors in full 3D.
The lattice Boltzmann technique is also capable of performing simulations of multiphase
systems. Ladd (1994) used this technique to simulate colloidal systems. In these 3D
simulations special attention was paid to ensure noslip boundary conditions on the
Gas Phase Hydrodynamics
______________________________________________________________________________________
83
particle fluid boundaries. In principle this method can also be applied to gassolid
systems as encountered in gas fluidised beds. However the system should be of limited
size (several solid particles instead of thousands) due to computational limitations. The
method is therefore very well suited to obtain improved closure relations for the gas
particle drag force on the length scale at which it is applied in granular dynamics
simulations of gas fluidised beds as described in this work.
1.3 Dissipative Particle Dynamics
A somewhat similar approach to flow simulations is the dissipative particle dynamics
technique originally developed by Koelman and Hoogerbrugge (1992). Hoogerbrugge
and Koelman (1993) applied this technique to perform 3D simulations of hardsphere
suspensions under steady shear. This technique is based on the dynamics of particles that
should be regarded as momentum carriers rather than physical particles. Unlike in lattice
Boltzmann simulations these particles are not restricted to lattice sites. The dynamics of
these particles consists of a collision phase, where the interaction between the particles is
taken into account, and a propagation phase, where the particles are moved according to
an equation of motion that contains contributions of a random and a dissipative force.
Español and Warren (1995) have shown that the dissipative force as used by
Hoogerbrugge and Koelman (1993) needs to be modified in order to obey the fluctuation
dissipation theorem. Simulations of suspensions are possible by local freezing of particles
that are found within the volume defined by the suspended particle. By summation of the
forces experienced by all the particles that constitute the suspended particle, the net force
exerted on the suspended particle is obtained. The DPD technique is capable of handling
complex geometries and has been used to study colloidal systems. It seems however that
the DPD technique is limited to low Reynolds number flow and therefore its use for
inertia dominated systems, like gasfluidised beds, is doubtful.
Ge and Li (1997) presented a technique that resembles DPD in the sense that the gas
phase is represented by discrete particles rather than the NavierStokes equations. Despite
the fact that this method is rather original it leaves quite some questions unanswered. For
Chapter 3
______________________________________________________________________________________
84
example it is not clear how physical properties of the gas are incorporated into the model.
Furthermore it is not clear what equation of motion is used for the gas particles. It would
be helpful if it could be demonstrated that this model is capable of predicting key
fluidisation phenomena such as the minimum fluidisation velocity and the terminal
velocity of a particle falling under the influence of gravity in a quiescent gas.
2. Governing Equations
The calculation of the gasphase hydrodynamics mainly follows the lines presented by
Kuipers et al. (1992). It is based on the numerical solution of the following set of partial
differential equations that can be seen as a generalised form of the NavierStokes
equations for a gas interacting with a solid phase as originally derived by Anderson and
Jackson (1967).
Continuity equation gas phase:
( )
( )
∂ ερ
∂
ερ
g
g
t
+ ∇⋅ · u 0. (3.1)
Momentum equation gas phase:
( )
( ) ( )
∂ ερ
∂
ερ ε ε ερ
g
g p g g
t
p
u
uu S g + ∇ ⋅ · − ∇ − − ∇ ⋅ + τ . (3.2)
In this work isothermal, twodimensional motion is considered which implies that three
basic variables have to be specified. The three basic variables in the model are the
pressure (p) and the two velocity components of the gasphase (u
x
and u
y
). The void
fraction (ε) and the momentum exchange source term (S
p
) are obtained from the discrete
Gas Phase Hydrodynamics
______________________________________________________________________________________
85
particle model as will be explained in section 6. All remaining variables have to be
specified in terms of the three basic variables and/or the variables obtained from the
discrete particle model through constitutive equations.
3. Constitutive Equations
3.1 Gas phase density
The gas phase density (ρ
g
) is related to the pressure (p) and the gas phase temperature (T)
by the ideal gas law:
p
RT
M
?
g
g
· , (3.3)
where R is the gas constant (8.314 J/(mol K)). The average molecular weight of air (M
g
=
28.8 10
3
kg/mol) was used and the temperature was set to a constant value of T = 293 K.
3.2 Gas phase stress tensor
The viscous stress tensor τ
g
is assumed to depend only on the gas motion. The general
form for a Newtonian fluid (Bird et al., 1960) has been implemented:
( ) ( ) ( ) ( )
]
]
]
]
∇ + ∇ + ⋅ ∇
,
`
.

− − ·
T
g g g g
u u I u µ µ λ
3
2
τ . (3.4)
In the simulations the bulk viscosity of the gas phase λ
g
was set equal to zero which is
allowed for gases (Bird et al., 1960) whereas for the gas phase shear viscosity a constant
value of µ
g
= 1.8 10
5
kg/ms was used. I denotes the unit tensor.
Note that no turbulence modelling was taken into account. For bubbling beds this can be
justified since the turbulence is damped out in the bed due to the very high solids fraction.
Chapter 3
______________________________________________________________________________________
86
4. Numerical Solution
The numerical solution follows the lines of Kuipers et al. (1993) and will not be
discussed in detail here. A finite difference technique, employing a staggered grid to
ensure numerical stability, is used to solve the gasphase conservation equations 3.1 and
3.2. This implies that the scalar variables (p and ε) are defined at the cell centre and that
the velocity components are defined at the cell faces as is shown in Figure 3.1.
Figure 3.1. Lay out of the staggered grid.
A pressure correction technique was employed to solve the set of partial differential
equations. The model is capable of performing transient twodimensional calculations in
a Cartesian or an axisymmetrical geometry. In the simulations reported in this work only
the Cartesian option was used.
DY
DX
(i,j)
(i,j+0.5)
(i+0.5,j)
yvelocity component
scalar variable
xvelocity component
Gas Phase Hydrodynamics
______________________________________________________________________________________
87
5. Boundary Conditions
For the incorporation of the boundary conditions a flag matrix is used which allows
boundary conditions to be specified for each single cell. A variety of boundary conditions
can be applied by specification of the value of the cell flag fl(i,j) which is associated with
the relevant boundary condition for that cell(i,j). The typical set of boundary conditions
used in the simulations performed in this study is shown in Figure 3.2.
Figure 3.2. Cell flags for the boundary conditions for the hydrodynamic model
corner cells
1
1
5
1
1
4
1 3
1 3
7 5
1
1 3
3
7
4
interior cells
fictitious cells
NY+1
NY
j
1
0
0
1
i NX
NX+1
3
3
7
3
3
7
1
1
5
1
1
4
Chapter 3
______________________________________________________________________________________
88
The cell flags and the corresponding boundary conditions that are featured in the present
code are listed in Table 3.1.
Table 3.1. Cell flags and corresponding cell types.
fl(i,j) Cell type of cell(i,j)
1 Interior cell, no boundary conditions have to be specified
2 Impermeable wall, free slip boundaries
3 Impermeable wall, no slip boundaries
4 Influx cell, velocities have to be specified
5 Prescribed pressure cell, free slip boundaries
6 Continuous outflow cell, free slip boundaries
7 Corner cell, no boundary conditions have to be specified
To mimic a distributor plate the void fraction in the influx cells (fl(i,j) = 4) was set to a
constant value of 0.4.
6. Twoway coupling
An important issue in granular dynamics simulations of twophase flow is the twoway
coupling. The calculation of the drag force is based on empirical relations and therefore
not exact. Hence it is required that twoway coupling is achieved in such a way that the
model is capable to predict typical fluidisation phenomena such as the pressure drop over
the bed at minimum fluidisation conditions. This pressure drop, multiplied by the cross
sectional area, should balance the force exerted by gravity on the particles. Xu and Yu
(1997) used a technique similar to the one used by Schwarzer (1995) where the total drag
force exerted on a particle was fed back to the gas phase with a minus sign. With this
technique the pressure drop over the bed at minimum fluidisation conditions in the
simulations of Xu and Yu (1997) was over predicted by a factor 1.5 (Hoomans et al.,
1998). Therefore special care has to be taken in order to select a twoway coupling
technique that ensures the key fluidisation features to be incorporated correctly.
Gas Phase Hydrodynamics
______________________________________________________________________________________
89
Twoway coupling is achieved via the calculation of the void fraction and the
incorporation of an interaction term in the momentum conservation equation for the gas
phase (equation 3.2). In the following paragraphs it is explained how twoway coupling is
achieved in the model used in this work.
6.1 Void fraction
6.1.1 Calculation of the void fraction in 2D
The solution of equations 3.1 and 3.2 requires specification of the void fraction (ε) which
can be obtained from the discrete particle model. Since the particle positions are known
the void fraction ε(i,j) can be calculated based on the area occupied by the particles in
that cell i,j. Since the void fraction is an important parameter which considerably
influences the motion of the gas phase, a detailed check for overlap was performed in
which multiple cell overlap was taken into account as illustrated in Figure 3.3. In this
figure a case is presented where a particle overlaps four different grid cells. The distances
from the particle centre to the nearest boundaries of the grid cells are indicated by δ
1
and
δ
2
.
Figure 3.3. Multiple particlecell overlap.
δ
2
δ
1
A
i,jj
A
ii,j
i,jj
ii,jj
i,j ii,j
A
ii,jj
Chapter 3
______________________________________________________________________________________
90
The smallest overlapping area of the particle A
ii,jj
can be calculated as follows:
]
]
]
]
]
,
`
.

,
`
.

−
,
`
.

−
,
`
.

− +
,
`
.

− − ·
p p
p
p p
p jj ii
R R
R
R R
R A
2 1
2
2
2
2
1
1 2 1 ,
arcsin arccos 1 1
2
1 δ δ δ
δ
δ
δ δ δ
. (3.5)
For the area A
jj
overlapping with the upper two cells in Figure 3.3 the following relation
can be obtained:
,
`
.

,
`
.

− −
,
`
.

· + ·
2
2
2
2
, ,
1 arccos
p p
p p jj ii jj i jj
R R
R R A A A
δ
δ
δ
. (3.6)
Keeping in mind that:
2
, , , , p j ii jj ii jj i j i
R A A A A π · + + + , (3.7)
all four areas can now be calculated using basic subtractions. In the case where a particle
overlaps with only two cells the area of overlap can simply be obtained from equation
3.6.
However, the void fraction calculated in this way is based on a twodimensional analysis
that is inconsistent with the applied empiricism in the calculation of the drag force
exerted on a particle (Chapter 2, section 6). To correct for this inconsistency the void
fraction calculated on the basis of area (ε
2D
) is transformed into a threedimensional void
fraction (ε
3D
) using the following equation:
( ) ε
π
ε
3 2
3
2
1
2
3
1
D D
· − − . (3.8)
Gas Phase Hydrodynamics
______________________________________________________________________________________
91
This equation has been derived on the basis of a comparison between a twodimensional
hexagonal lattice and a threedimensional FCC unit cube assuming equal interparticle
distances. It ensures that the closest packing in the 2D hexagonal lattice is transformed
into the closest packing in the 3D FCC case. Ouyang and Li (1999) used a
transformation similar to equation 3.8 but used the square root of 2 instead of 2 in the
second term on the right hand side. This renders void fractions that in general are too
high.
Tsuji et al. (1993), Xu and Yu (1997), Kawaguchi et al. (1998) and Mikami et al. (1998)
all used a different approach. They calculated the void fraction on the basis of volumes
assuming that the system consists of one particle layer. Hence the third dimension is
equal to the particle diameter. This method yields void fractions that are in general too
high as well. Using this method the closest packing can never be obtained and therefore it
is not a correct representation for a threedimensional system.
6.1.2 Calculation of the void fraction in 3D
In the 3D model the problem with the conversion of a 2D void fraction to a 3D void
fraction no longer exists. Since the gas phase hydrodynamics is still resolved in 2D each
particle can still have overlap with four cells at maximum. Consider Figure 3.3 but now
read volumes instead of areas. The main difficulty in the 3D void fraction calculation is
to obtain an expression for the volume V
ii,jj
. Since there is no analytical expression
available for this volume an approximation has to be used. When a particle overlaps with
two cells the volume of a spherical cap can be calculated exactly by:
( )
( )
1
2
1
2
3
δ
δ π
+
−
·
p
p
cap
R
R
V , (3.9)
with δ
1
being the distance from the centre of the particle to the cell boundary. V
ii,jj
can
now be approximated by:
Chapter 3
______________________________________________________________________________________
92
p
p
ii,j ii,jj
p
i,jj ii,jj
ii,jj
V
V
V V
V
V V
V
,
`
.
 +
,
`
.
 +
· . (3.10)
Since V
ii,jj
+V
i.jj
and V
ii,jj
+V
ii,j
are sphere caps, these volumes can be calculated using
equation 3.9. Each of the 4 volumes can subsequently be obtained using basic
subtractions.
6.2 Momentum transfer
The method used in this work is based on Newton’s third law and was also used by
Delnoij et al. (1997). It differs from the technique used by Hoomans et al. (1996). The
reaction force to the drag force exerted on a particle per unit volume is included in the
momentum conservation equation 3.2 via a source term S
p
that has the dimension N/m
3
:
( )
( ) ( )
∫
∑
− −
−
− ·
·
dV
V
V
a a
Npart
a
p
p
r r v u S δ
ε
β
0
1
1
(3.11)
The δfunction ensures that the reaction force acts as a point force at the position of the
particle in the system. In the numerical implementation this forcepervolume term is
distributed to the four nearest grid nodes using the area weighted averaging technique
described in Chapter 2, section 6, where also the expression for the volumetric
momentum exchange coefficient β can be found. A mixed explicit implicit numerical
treatment was applied similar to the one used by Kuipers et al. (1993) to solve the gas
phase conservation equations.
Since the source term S
p
has the dimension of force per unit volume the force exerted on
the particles has to de divided by the volume of a grid cell. In the 3D model this is
straightforward since the third dimension is determined by the depth of the bed. In the 2
D model a virtual third dimension has to be introduced. This virtual third dimension is
Gas Phase Hydrodynamics
______________________________________________________________________________________
93
estimated based on the same calculation as the conversion from ε
2D
to ε
3D
. In the 2D
simulations the volume of a computational cell was chosen to be:
p cell
D DZ DR V
75 . 0
3 2
−
· , (3.12)
in which the third dimension is slightly less than the particle diameter. Note that this
volume depends on the particle size and hence is not a constant in simulations where a
particle size distribution is taken into account.
Tsuji et al. (1993), Kawaguchi et al. (1998) and Mikami et al. (1998) use a technique
where the coupling was achieved in a similar way as in twofluid models. This technique
was also employed by Hoomans et al. (1996). A two fluid model requires averaged
values of particle velocities and these can be obtained by averaging over all the particles
in a computational cell. However, this technique is limited to systems consisting of
particle of uniform size and therefore the technique used in this work was preferred.
Notation
A area, m
2
D
p
particle diameter, m
DX horizontal computational cell dimension, m
DY vertical computational cell dimension, m
fl(i,j) cell flag [], defined in Table 1
g gravitational acceleration, m/s
2
I unity tensor, []
i,j cell indices, []
M molecular weight, kg/mol
NX number of computational cells in xdirection, []
NY number of computational cells in ydirection, []
p pressure, Pa
Chapter 3
______________________________________________________________________________________
94
r particle position vector, m
R
p
particle radius, m
R gas constant, J/mol K
S
p
source term defined in equation 3.11
t time, s
T temperature, K
u gas phase velocity, m/s
V volume, m
3
v velocity, m/s
Greek symbols
β volumetric interphase momentum transfer coefficient, kg/(m
3
s)
δ distance, m
ε void fraction, []
λ
g
gas bulk viscosity, kg/(m.s)
µ
g
gas shear viscosity, kg/(m.s)
ρ density, kg/m
3
τ gas phase stress tensor, kg/ms
2
Subscripts
g gas phase
i,j cell indices
p particle
x xcomponent
y ycomponent
Superscripts
T transposed
Gas Phase Hydrodynamics
______________________________________________________________________________________
95
References
Anderson, T.B. and Jackson, R., (1967). A fluid mechanical description of fluidized beds
(equations of motion), Ind. Eng. Chem., Fundam., 6, 527.
Bird, R.B., Stewart, W.E. and Lightfood, E.N., (1960). Transport phenomena, John
Wiley & Sons, New York.
Crowe, C.T., Sharma, M.P. and Stock, D.E., (1977). The ParticleSourceIn Cell (PSI
CELL) model for gasdroplet flows, ASME Trans. J. Fluids Eng., 325.
Delnoij, E., Lammers, F.A., Kuipers, J.A.M. and van Swaaij W.P.M., (1997). Dynamic
simulation of dispersed gas liquid twophase flow using a discrete bubble model, Chem.
Engng Sci., 52, 1429.
Eggels, J.G.M., and Somers, J.A., (1995). Numerical simulation of free convective flow
using the latticeBoltzmann scheme, Int. J. Heat and Fluid Flow, 16, 357.
Eggels, J.G.M., (1996). Direct and largeeddy simulation of turbulent and fluid flow
using the latticeBoltzmann scheme, Int. J. Heat and Fluid Flow, 17, 307.
Español, P. and Warren, P., (1995). Statistical mechanics of dissipative particle
dynamics, Europhys. Lett., 30, 191.
Ge, W. and Li, J., (1997). PseudoParticle Approach to Hydrodynamics of Gas/Solid
TwoPhase Flow, In Proceedings of the 5
th
Int. Conf. Circulating Fluidized beds, Beijing
(China), May 28 June 1, 1996, DT8, 1.
Guj, G. and de Matteis, G., (1986). Fluidparticles interaction in particulate fluidized
beds: numerical and experimental analysis, Phys. Chem. Hydrodynamics, 7, 145.
Chapter 3
______________________________________________________________________________________
96
Hoogerbrugge, P.J. and Koelman, J.M.V.A., (1992). Simulating microscopic
hydrodynamic phenomena with dissipative particle dynamics, Europhys. Lett., 19, 155.
Hoomans, B.P.B., Kuipers, J.A.M., Briels, W.J. and van Swaaij, W.P.M., (1996).
Discrete particle simulation of bubble and slug formation in a twodimensional gas
fluidised bed: a hardsphere approach, Chem. Engng Sci., 51, 99.
Hoomans, B.P.B., Kuipers, J.A.M., Briels, W.J. and van Swaaij, W.P.M., (1998).
Comments on the paper “Numerical simulation of the gassolid flow in a fluidized bed by
combining discrete particle method with computational fluid dynamics” by B.H. Xu and
A.B. Yu, Chem. Engng Sci., 53, 2645.
Hu, H.H., (1996). Durect simulation of flows of solidliquid mixtures, Int. J. Multiphase
Flow, 22, 335.
Kalthoff, W., Schwarzer, S. and Hermann, H.J., (1997). Algorithn for the simulation of
particle suspensions with inertia effects, Phys. Rev. E, 56, 2234.
Kawaguchi, T., Tanaka, T. and Tsuji, Y., (1998). Numerical simulation of two
dimensional fluidized beds using the discrete element method (comparison between the
two and threedimensional models), Powder Technol., 96, 129.
Koelman, J.M.V.A. and Hoogerbrugge, P.J., (1993). Dynamics simulations of hard
sphere suspensions under steady shear, Europhys. Lett., 21, 363.
Kuipers, J.A.M., van Duin K.J., van Beckum, F.P.H. and van Swaaij, W.P.M., (1992). A
numerical model of gasfluidized beds, Chem. Engng Sci., 47, 1913.
Kuipers, J.A.M., van Duin K.J., van Beckum, F.P.H. and van Swaaij, W.P.M., (1993).
Computer simulation of the hydrodynamics of a twodimensional gas fluidized bed,
Comput. Chem. Engng, 17, 839.
Gas Phase Hydrodynamics
______________________________________________________________________________________
97
Ladd, A.J.C., (1994). Numerical simulations of particulate suspensions via a discretized
Boltzmann equation. Part 1. Theoretical foundation, J. Fluid Mech., 271, 185.
McNamara, G.R. and Zanetti, G., (1988). Use of the Boltzmann equation to simulate
laticegas automata, Phys. Rev. Lett., 61, 2332.
Mikami, T., Kamiya, H. and Horio, M., (1998). Numerical simulation of cohesive powder
behavior in a fluidized bed, Chem. Engng Sci., 53, 1927.
Ouyang, J. and Li, J. (1999). Particle motionresolved discrete model for simulating gas
solid fluidization, Chem. Engng Sci., 54, 2077.
Pan, Y. and Banerjee, S., (1995). Numerical simulation of particle interactions with wall
turbulence, Phys. Fluids, 8, 2733.
Schwarzer, S., (1995). Sedimentation and flow through porous media: Simulating
dynamically coupled discrete and continuum phase, Phys. Rev. E., 52, 6461.
Tsuji, Y., Kawaguchi, T. and Tanaka, T., (1993). Discrete particle simulation of two
dimensional fluidized bed, Powder Technol., 77, 79.
Xu, B.H. and Yu, A.B., (1997). Numerical simulation of the gassolid flow in a fluidized
bed by combining discrete particle method with computational fluid dynamics, Chem.
Engng Sci., 52, 2785.
Chapter 3
______________________________________________________________________________________
98
Effect of Particle Properties…
___________________________________________________________________________________
99
Chapter 4.
THE EFFECT OF PARTICLE PROPERTIES ON THE
HYDRODYNAMICS OF GASFLUIDISED BEDS WITH
HOMOGENOUS INFLOW CONDITIONS
Abstract:
A hardsphere Granular Dynamics model was used to study the influence of particle
properties on the hydrodynamics of gasfluidised beds with homogeneous inflow
conditions. In the model the Newtonian equations of motion are solved for each solid
particle while taking into account the particle particle and particlewall collisions.
The gas phase hydrodynamics is described by the spatially averaged NavierStokes
equations for twophase flow. The hydrodynamics of a homogeneously fluidised bed
(0.5 m height, 0.15 m width, 2400 particles of 4 mm diameter, ρ = 2700 kg/m
3
, u
gas
=
1.5 u
mf
) was strongly affected by the collision parameters (i.e. the coefficients of
restitution and friction): the more energy is dissipated in collisions the more bubbling
and slugging could be observed. The Root Mean Square (RMS) of the pressure
fluctuations inside the bed showed an almost linear dependency on the energy
dissipation rate by collisions during the simulation for low dissipation rates. The
effect of a particle size distribution on the bed hydrodynamics was studied as well.
Pressure peaks inside the bed decreased with increasing geometric standard deviation
of the lognormal size distribution. Simulations with a threedimensional model (as
far as the particles are concerned) required a minimum depth of the bed of five
particle diameters for stable simulations. The results of the threedimensional
simulations confirmed the trends observed in the twodimensional simulations
concerning the effect of the collision parameters on the bed hydrodynamics. Time
series of the pressure inside the bed showed a lower frequency and a lower RMS
value in the 3D simulations. The particlewall interaction had a more pronounced
influence on the hydrodynamics in the 3D simulations than in the 2D simulations.
Chapter 4
___________________________________________________________________________________
100
Parts of this chapter are based on the papers:
B.P.B. Hoomans, J.A.M. Kuipers, W.J. Briels and W.P.M. van Swaaij, (1996). Discrete particle
simulation of bubble and slug formation in twodimensional gasfluidised beds: a hard sphere
approach, Chem. Engng Sci., 51, 99.
B.P.B. Hoomans, J.A.M. Kuipers and W.P.M. van Swaaij, (1998). The influence of a particle size
distribution on the granular dynamics of dense gas fluidized beds: a computer simulation study, AIChE
Symp. Ser. No. 318, Vol. 94, 15.
B.P.B. Hoomans, J.A.M. Kuipers and W.P.M. van Swaaij, (1998). The influence of particle properties
on pressure signals in dense gasfluidised beds: a computer simulation study, World Congress on
Particle Technology 3, July 79, Brighton, UK.
Effect of Particle Properties…
___________________________________________________________________________________
101
1. Introduction
With increasing computer power Granular Dynamics simulations have become a very
useful and versatile research tool to study the dynamics of dense gasparticle flows. In
these models the Newtonian equations of motion are solved for each individual
granular particle in the system. The mutual interactions between particles and the
interaction between particles and walls are taken into account directly. Applications of
this type of models are very diverse ranging from agglomerate dynamics (Thornton et
al., 1996) to hopper flow (Langston et al., 1995) and gasfluidised beds (Hoomans et
al., 1996).
In fluidisation simulations the Granular Dynamics models have a clear advantage over
continuum models since the particle interactions are taken into account directly and
consequently the necessity to specify closure relations for the solidsphase stress
tensor, as encountered in continuum models (Kuipers et al., 1992 among many others),
is circumvented. Furthermore a discrete particle model is very well suited to take an
arbitrary particle size distribution into account which is far more complex, if not
impossible, in case a continuum modelling approach is adopted.
Tsuji et al. (1993) developed a twodimensional softsphere discrete particle model of
a gas fluidised bed based on the work of Cundall and Strack (1977). In this approach
the particles are allowed to overlap slightly and this overlap is subsequently used to
calculate the contact forces. Kawaguchi et al. (1998) extended the model developed by
Tsuji et al. (1993) to threedimensions as far as the motion of the particles is
concerned. Hoomans et al. (1996) used a hardsphere approach in their two
dimensional discrete particle model of a gasfluidised bed which implies that the
particles interact through binary, instantaneous, inelastic collisions with friction. Xu
and Yu (1997) used a twodimensional softsphere model similar to the one used by
Tsuji et al. (1993) in their fluidised bed simulations but they also employed techniques
from hardsphere simulations in order to determine the instant when the particles first
come into contact precisely. Mikami et al. (1998) presented a twodimensional soft
Chapter 4
___________________________________________________________________________________
102
sphere model of a gasfluidised bed where cohesive forces between the particles were
taken into account. A more detailed discussion of the models mentioned above is
presented in Chapter 2.
In Granular Dynamics simulations particleparticle and particlewall interactions are
taken into account directly which is a clear advantage over continuum models.
Collision parameters like the coefficient of restitution can be included in continuum
models by employing the kinetic theory of granular flow to arrive at improved closure
relations for the solidsphase stress tensor (Sinclair and Jackson (1989), Gidaspow
(1994), Nieuwland et al. (1996)). However, these improved closures rely on a one
parameter collision model, do not take particle rotation into account and more
importantly show an extreme and nonrealistic sensitivity of the overall bed
hydrodynamics to very small deviations from one for the coefficient of restitution
(Hrenya and Sinclair, 1997). Within the framework of a discrete particle model the
collision parameters are included in a more natural way and therefore this type of
model is to be preferred when investigating the influence of particle parameters on the
bed hydrodynamics.
Hoomans et al. (1996) have shown that the dynamics of a fluidised bed depends
strongly on the collision parameters ( i.e. coefficients of restitution (e) and friction (µ)).
In the case where no energy is dissipated (i.e. e = 1 and µ = 0) no bubble or slug
formation was observed and the bed showed a rather homogeneous fluidisation
behaviour. In the case where the impact parameters were given more realistic values
(i.e. e = 0.9 and µ = 0.3) and hence energy is dissipated in collisions, bubble formation
was observed and the bed behaviour appeared to be in much better agreement with that
observed in reality. Van den Bleek and Schouten (1993) already showed that the
coefficient of restitution has an effect on the dynamics of a fluidised bed. They used a
onedimensional fiveparticlearray model to generate time series of particle position
and voidage which is described in more detail by Schouten and van den Bleek (1992).
These time series were subsequently analysed by applying techniques from
deterministic chaos theory. They found that for increasing values of the coefficient of
restitution (up to 0.9) the bed dynamics became less chaotic. The coefficient of
Effect of Particle Properties…
___________________________________________________________________________________
103
restitution was varied over a range from 0.0 to 0.9, which is unfortunately not the most
interesting range. Typical values for this parameter of granular material generally used
in fluidisation are found within 0.8 and 1.0. However, it is clear that despite the fact
that this fiveparticlearray model is a rather crude representation of the dynamics of a
fluidised bed, it does capture the essential feature that the collision parameters have a
pronounced effect on the bed hydrodynamics.
Experimental studies on the influence of particle interaction parameters on the
dynamics of gasfluidised beds are unfortunately rather scarce. For example the
Geldart (1974) classification of the fluidisation behaviour of powders does not include
any particle interaction parameters. Thiel and Potter (1977) reported an influence of
the particleparticle interaction on the dynamics of slugging fluidised beds. They
characterised this interaction through the angle of internal friction (Zenz and Othmer,
1960). This is a bulk property rather than a particle property but in general it can be
assumed that the rougher the particles the lower the angle of internal friction. Chen et
al. (1997a and 1997b) investigated the influence of solid frictional interactions on
pressure fluctuations in fluidised beds. They also characterised the particle interaction
through the angle of internal friction. For relatively shallow beds they found larger
root mean square (RMS) values for the pressure fluctuations in beds consisting of
particles with lower angles of internal friction. They also reported a negligible
influence of the wall roughness on the RMS values of the measured pressure
fluctuations. Chang and Louge (1992) studied the hydrodynamics in a riser where they
applied a coating to glass particles in order to make them smoother and hence obtain a
lower coefficient of friction. Although this system is operated in a completely different
regime compared to those prevailing in the experiments mentioned above and the
simulations presented in this chapter, the experimental results obtained by Chang and
Louge (1992) are worthwhile mentioning here. For the coated particles a completely
different axial pressure profile was measured than for the regular glass particles. In the
case of the coated particles (i.e. lower coefficient of friction) the solids hold up in the
riser section (and thus the axial pressure gradient) was a lot smaller. Since the only
difference between the two experiments was the use of the coated particles instead of
the regular glass particles their experiments indicate that particleparticle interaction
has a significant effect on the flow structure in the riser.
Chapter 4
___________________________________________________________________________________
104
In this chapter the dependency of the hydrodynamics of a gas fluidised bed with
homogeneous inflow conditions on the collision parameters is investigated in more
detail using the hardsphere model. Furthermore the model was used to study the
influence of a particle size distribution on the hydrodynamics of gas fluidised beds
with homogeneous inflow conditions. The majority of the simulations was performed
with the twodimensional model, but also simulations with the threedimensional
model were performed.
2. Model
Since a detailed description of the model is presented in Chapter 2 and Chapter 3, the
key features will be summarised briefly here.
2.1 Granular dynamics
The collision model as originally developed by Wang and Mason (1992) is used to
describe a binary, instantaneous, inelastic collision with friction. The key parameters
of the model are the coefficient of restitution (0 ≤ e ≤ 1) and the coefficient of friction
(µ ≥ 0). Foerster et al. (1994) have shown that also the coefficient of tangential
restitution (0 ≤ β0
≤ 1) should be used in order to describe the collision dynamics more
accurately. These three collision parameters are all included in the model.
In the hardsphere approach a sequence of binary collisions is processed. This implies
that a collision list is compiled in which for each particle a collision partner and a
corresponding collision time is stored. A constant time step is used to take the external
forces into account and within this time step the prevailing collisions are processed
sequentially. In order to reduce the required CPU time neighbour lists are used. For
each particle a list of neighbouring particles is stored and only for the particles in this
list a check for possible collisions is performed.
Effect of Particle Properties…
___________________________________________________________________________________
105
2.2 External forces
The incorporation of external forces differs somewhat from the approach followed by
Hoomans et al. (1996). In this work the external forces are used in accordance with
those implemented in the two fluid model presented by Kuipers et al. (1992) where, of
course, the forces now act on a single particle:
( )
( ) m
d
dt
m
V
V p
p
p
p
p
p p
v
g u v · +
−
− − ∇
β
ε 1
, (4.1)
where m
p
represents the mass of a particle, vp
its velocity, u the local gas velocity and
V
p
the volume of a particle. In equation (4.1) the first term is due to gravity and the
third term is the force due to the pressure gradient. The second term is due to the drag
force where β represents an interphase momentum exchange coefficient as it usually
appears in twofluid models. For low void fractions (
ε
< 0.80)
β
is obtained from the
wellknown Ergun equation:
( )
( )
p
p
g
p
g
D D
v u − − +
−
·
ρ
ε
µ
ε
ε
β 1 75 . 1
1
150
2
2
, (4.2)
where D
p
represents the particle diameter, µg
the viscosity of the gas and ρg
the
density of the gas. For high void fractions (ε ≥ 0.80) the following expression for the
interphase momentum transfer coefficient has been used which is basically the
correlation presented by Wen and Yu (1966) who extended the work of Richardson
and Zaki (1954):
( )
65 . 2
1
4
3
−
−
−
· ε ρ
ε ε
β
p g
p
d
D
C v u
. (4.3)
The drag coefficient C
d
is a function of the particle Reynolds number and given by:
Chapter 4
___________________________________________________________________________________
106
( )
¹
¹
¹
'
¹
≥
<
·
1000 Re 44 . 0
1000 Re Re 0.15 + 1
Re
24
0.687
p
p p
p
d
C ,
(4.4)
where the particle Reynolds number (Re
p
) in this case is defined as follows:
g
p p g
p
D
µ
ερ v u −
· Re . (4.5)
For the integration of equation (4.1) a simple explicit first order scheme was used to
update the velocities and positions of the particles.
2.3 Gas phase hydrodynamics
The calculation of the gas phase hydrodynamics mainly follows the lines presented by
Kuipers et al. (1992). It is based on the numerical solution of the following set of
partial differential equations that can be seen as a generalised form of the Navier
Stokes equations for a gas interacting with a solid phase as originally derived by
Anderson and Jackson (1967).
Continuity equation gas phase:
( )
( )
∂ ερ
∂
ερ
g
g
t
+ ∇⋅ · u 0 . (4.6)
Momentum equation gas phase:
( )
( ) ( )
∂ ερ
∂
ερ ε ε ερ
g
g p g g
t
p
u
uu S g + ∇ ⋅ · − ∇ − − ∇ ⋅ + τ . (4.7)
Effect of Particle Properties…
___________________________________________________________________________________
107
In this work transient, twodimensional, isothermal (T = 293 K) flow of air at
atmospheric conditions is considered. The constitutive equations can be found in
Chapter 3. There is one important modification with respect to the model presented by
Hoomans et al. (1996) and that deals with the way in which the twoway coupling
between the gasphase and the dispersed particles is established. In the present model
the reaction force to the drag force exerted on a particle per unit of volume is fed back
to the gas phase through the source term S
p
which has the dimension of force per unit
of volume N/m
3
.
3. Effects of collision parameters
A large number of simulations was performed in order to investigate the influence of
the collision properties on the bed dynamics. As a test system the same system as used
before by Tsuji et al. (1993), Hoomans et al. (1996) and Xu and Yu (1997) was
chosen. The most important system parameters are summarised in table 4.1.
Table 4.1. Parameter settings for all simulations
Particles Bed
Shape Spherical Width 0.15 m
Diameter, D
p
4.0 mm Height 0.50 m
Material Aluminium Number of xcells 15
Density, ρ 2700 kg/m
3
Number of ycells 25
Number, N
part
2400
In all simulations the minimum fluidisation conditions were used as initial conditions
while the values for the coefficients of restitution (e) and friction (
µ
) were varied. A
time step of 10
4
seconds was used in all simulations and the gas inflow at the bottom
was set equal to 1.5 u
mf
for the whole width of the bed (u
mf
= 1.78 m/s). The
simulations were continued for 8 seconds real time and during the simulations the
pressure in the bed 0.2 m above the centre of the bottom plate was monitored.
Chapter 4
___________________________________________________________________________________
108
Figure 4.1. Snapshots of particle configurations for three simulations with u
g
= 1.5
u
mf
. Upper: e = 1, µ = 0, middle: e = 0.99, µ = 0, bottom: e = 0.9, µ =
0.3.
Effect of Particle Properties…
___________________________________________________________________________________
109
A typical (no n ideal) run takes about 6 minutes of CPUtime per simulated second on
a Silicon Graphics Origin200 server (R10000 processor, 180 MHz). Snapshots of the
particle configurations for three different simulations are presented in Figure 4.1. The
two extreme cases (e = 1.0, µ = 0.0 (ideal) and e = 0.9, µ = 0.3 (nonideal)) are shown
there together with a simulation where e = 0.99 and µ = 0.0. A number of particle
layers was marked by colour at t = 0 s in order to visualise t he induced particle mixing
during the fluidisation process. It can be seen that in the ideal case barely any bubbles
or slugs are being formed which enables the initial structure to remain intact for quite
some time. In the non ideal case the formation of bubbles and slugs can be observed
clearly and due to these instabilities the particle mixing in the bed is much better.
After 2 seconds the initial structure can still be observed in the ideal case whereas it
has completely disappeared in the nonideal case. The simulation with e = 0.99
resembles the ideal case very much although the particle mixing appears to be slightly
better. This is a much more natural dependency than the one reported for instance by
Hrenya and Sinclair (1997). They reported simulations of riser flow using a twofluid
model that incorporates the kinetic theory for granular flow a change in e from 1.0 to
0.99 completely changed the hydrodynamic behaviour.
Figure 4.2. Pressure at 0.2 m above the centre of the distributor plate as a function
of time.
e µ
e µ
e µ
Chapter 4
___________________________________________________________________________________
110
However, it should be noted that these simulations were concerned with a different
hydrodynamic regime than encountered in the simulations reported here and therefore
a direct comparison may not be justified.
The computed pressure signals at 0.2 m above the centre of the bottom plate for the
three simulations are presented in Figure 4.2. It can be observed that the pressure
fluctuations in the non ideal case are much stronger than in the ideal case. In fact the
pressure fluctuations turn out to depend strongly on the collision parameters. In Figure
4.3 the root mean square (RMS) of the pressure signals is presented as a function of
the coefficient of restitution (e).
Figure 4.3. Root Mean Square (RMS) of the pressure fluctuations as a function of
the coefficient of restitution.
It can clearly be observed that the RMS increases with decreasing value of e. The fact
that, for instance, the RMS value for the simulation with e = 0.96 is higher than that
for the simulation with e = 0.95 is due to the different dynamics. Despite of the lower
value for the coefficient of restitution a higher amount of energy is dissipated in
collisions during the simulation. Because the total number of collisions in these
simulations is about the same (8.3 million) the average impact velocity was apparently
higher in the case of e = 0.96. With a slightly different initial condition this might
change which is due the complex (i.e. chaotic) nature of the system’s dynamics. In all
the simulations that were performed in order to obtain the RMS values presented in
e
Effect of Particle Properties…
___________________________________________________________________________________
111
Figure 4.3, the coefficient of friction was set equal to zero. In the calculation of the
RMS the initial stages of the simulation (approximately 2 s) were not taken into
account in order to avoid startup effects from influencing the results.
In Figure 4.4 the RMS is presented as a function of the value of the coefficient of
friction (µ). For all the simulations that were performed in order to obtain the RMS
values presented in Figure 4.4 the coefficient of restitution (e) was set equal to 1.0.
Figure 4.4. Root Mean Square (RMS) of pressure fluctuations as a function of the
coefficient of friction.
Here it can be observed that the RMS increases with increasing value for the
coefficient of friction. In Figure 4.4 the value for the RMS is lower for the simulation
with µ = 0.25 than for the simulations with µ = 0.2 which is not expected. However
the amount of energy dissipated in collisions for both simulations is not the same; in
fact more energy is dissipated in the simulation with µ = 0.2 than in the simulation
with µ = 0.25. In the latter case fewer collisions occurred and apparently the collisions
that occurred had less impact.
Also the influence of the coefficient of tangential restitution (β
0
) was investigated.
However, this parameter turned out to have little influence on the overall bed
hydrodynamics. This is probably due to the low (but realistic) values used for the
coefficient of friction (µ). When the coefficient of friction was increased (µ = 1000)
µ
Chapter 4
___________________________________________________________________________________
112
to ensure that all collisions were of the sticking type the influence of β0
was much
more pronounced. With e =1.0, µ = 1000 and β0
= 1.0 similar results were obtained as
for the case with e = 1.0, µ = 0.0 and β0
= 0.0. However, the influence of the
coefficient of tangential restitution for realistic values of the coefficient of friction (µ
< 1.0) was negligible.
In all simulations mentioned above the values of the collision parameters for particle
wall collisions were taken to be equal to the values for particleparticle collisions. The
influence of the particlewall interaction on the overall bed hydrodynamics was
investigated as well. This influence was found to be negligible. The particleparticle
interaction is very much dominant over the particlewall interaction in the 2D model
used here. A simulation with ideal particleparticle interaction (e = 1.0, µ = 0.0) but
nonideal particlewall interaction (e
w
= 0.9 , µ
w
= 0.3) showed very similar results as
in the fully ideal case.
In Figure 4.5 the RMS of the pressure is plotted as a function of the energy dissipation
rate by collisions during the simulation. All simulations performed in this study are
included in this figure.
Figure 4.5. Root Mean Square (RMS) of pressure fluctuations as a function of the
energy dissipation rate by collisions during the simulation.
Effect of Particle Properties…
___________________________________________________________________________________
113
It can be seen that the RMS increases almost linearly with the energy dissipation rate
by collisions in the system. For higher values of the dissipation rate (i.e. lower
coefficients of restitution and/or higher coefficients of friction) the data points show
some scatter. This is due to the richer dynamics of the system (i.e. more vigorous
bubbling and slugging which causes a more irregular pressure signal) at higher
dissipation rates which requires longer simulation times than the 8 seconds employed
here in order to obtain better statistics.
4. Energy conservation
Since the amount of energy dissipated in collisions turns out to have a profound
influence on the bed hydrodynamics the energy housekeeping during a simulation was
investigated in more detail. The energies concerned with the particle positions and
velocities (i.e. kinetic energy, potential energy and rotation energy) can be calculated
(instantaneously) as follows:
E m v
kin k k
k
Npart
·
∑
1
2
2
, (4.8)
∑
·
Npart
k
k y k pot
gr m E
,
, (4.9)
∑
·
Npart
k
k k rot
I E
2
2
1
ω . (4.10)
The energy dissipated in collisions is calculated using the equations presented in
Chapter 2. This dissipated energy is taken to be positive by definition. The energy that
the particles receive from the interaction with the gas phase can be obtained as
follows:
( )
( )
∑
∫
⋅
,
`
.

∇ − −
−
·
Npart
k
k k p k
k p
drag
dt p V
V
E v v u
,
,
1 ε
β
. (4.11)
All the five different energies mentioned above are calculated independently. Since
the dissipated energy is taken to be positive by definition the sum of the potential,
Chapter 4
___________________________________________________________________________________
114
kinetic, rotation and dissipated energy should equal the amount of energy that the
particles receive from the interaction with the gas phase. The total amount of energy
in the system can now be obtained as follows:
drag disp rot kin pot tot
E E E E E E − + + + · , (4.12)
and this amount should remain constant during a simulation.
In Figure 4.6 the energy housekeeping during a simulation with nonideal collision
parameters (e = 0.9, µ = 0.3) is presented. It can be observed that the potential energy
is larger than the kinetic energy. The rotation energy is about 2 orders of magnitude
smaller than the potential energy and therefore appears to be zero in this figure. The
energy dissipated in collisions increases almost linearly with time and its behaviour is
closely followed by the energy supplied by the drag force. In Figure 4.7 the energy
housekeeping during a simulation with ideal collision parameters (e = 1.0, µ = 0.0) is
presented. It should be stressed that the only differences between the system in Figure
4.6 and 4.7 are the values of the collision parameters.
Figure 4.6. Energy housekeeping for a simulation with nonideal collision
parameter settings (e = 0.9, µ = 0.3) with a homogenous gasinflow
velocity u
gas
= 1.5 u
mf
.
Effect of Particle Properties…
___________________________________________________________________________________
115
Figure 4.7. Energy housekeeping for a simulation with ideal collision parameter
settings (e = 1.0, µ = 0.0) with a homogenous gasinflow velocity u
g
=
1.5 u
mf
.
In Figure 4.7 it can be observed that the amount of energy dissipated in collisions
remains zero throughout the whole simulation. The energy put into the system by drag
force remains constant after a short initial stage (2 seconds). This is the amount of
energy required for the expansion of the bed. Note that in both the non ideal and the
ideal cases the total energy remains constant throughout the whole simulation as
required. It can also be observed that the potential energy in the nonideal case (Figure
4.6) fluctuates more strongly than in the ideal case (Figure 4.7) which is obviously
due to the bubbling and slugging in the non ideal case. Due to this vigorous bubbling
and slugging the pressure fluctuations in the bed are stronger than in the ideal case as
reported in the previous section. The energy required for these stronger fluctuations is
supplied by the drag force which puts a lot more energy into the system in the non
ideal case (Figure 4.6) than in the ideal case (Figure 4.7).
Chapter 4
___________________________________________________________________________________
116
5. Influence of a particle size distribution
Hardly any distribution of particle size encountered in fluidisation studies is
symmetrical, most of them are skewed to larger diameters (Seville et al. (1997)). A
symmetrical size distribution like a normal or a Gaussian distribution is therefore not
representative for particles used in laboratory or industrial practice. In this work the
particle diameters are obtained from a lognormal distribution which is asymmetrical
and can be represented in mathematical terms as follows:
( )
( )
p
g
CM p p
g p
D
D D
D
f d
ln 2
ln ln
exp
ln 2
1
d
2
2
,
]
]
]
]
−
− ·
σ σ π
, (4.13)
where df is the fraction of particles having diameters whose logarithms lie between ln
D
p
and ln D
p
+ d(ln D
p
). In equation 4.13 D
p,CM
is the count median diameter and σg
is
the geometric standard deviation. When creating the particle size distribution all the
diameters which are smaller than D
p,CM
σ
g
or larger than
D
p,CM
σ
g
are rejected to
mimic the effects of sieving. An example of a particle size distribution generated with
this method is represented in a (discrete) frequency histogram in Figure 4.8.
Figure 4.8. Frequency histogram of a lognormal particle size distribution with a
count median diameter of 4.0 mm and a geometric standard deviation
(GSD or σg
)
of 0.1 mm.
Effect of Particle Properties…
___________________________________________________________________________________
117
The influence of the distribution width on the bed hydrodynamics at homogeneous
inflow conditions was studied. Simulations were performed using the parameter
settings presented in table 4.1 where the particle diameter of 4 mm is now the mean
particle diameter. A time step of 10
4
s was used and all simulations were run for 10 s.
The coefficient of restitution (e) was set equal to 0.9 and the coefficient of friction (µ)
was set equal to 0.3 for both particleparticle and particlewall collisions. The initial
configurations were obtained by placing the particles in the system and allowing them
to fall under the influence of gravity while the homogeneous gas inflow was set equal
to u
mf
. Simulations were performed for a system consisting of particles of uniform size
and systems consisting of particles with a lognormal size distribution with a
geometric standard deviation
σg
= 0.1, 0.5 and 1.0 mm respectively. A homogeneous
gas inflow at 1.5 u
mf
was specified at the bottom of the system.
In Figure 4.9 the pressure fluctuations inside the bed at 0.2 m above the centre of the
bottom plate are presented for all the four cases.
Figure 4.9. Pressure at 0.2 m above the centre of the distributor plate as a function
of time for the four cases, GSD = Geometric Standard Deviation (σ
g
).
It can be observed that the pressure peaks decrease with increasing geometric standard
deviation of the size distribution. This can be explained by the lower void fraction in
the uniform case due to the closer packing which results in a higher force acting on
Chapter 4
___________________________________________________________________________________
118
the particles which in turn causes the higher pressure peaks. After the first 1.5 s the
differences become less pronounced which indicates that especially in situations
where close packing can occur, such as at minimum fluidisation conditions, it is
important to take polydispersity into account.
6. 3D simulations
A threedimensional version of the hardsphere model was developed. Similar to the
model presented by Kawaguchi et al. (1998), the gas phase hydrodynamics is still
resolved in two dimensions. This can be justified since the simulated system is a flat
fluidised bed with a depth of 22 mm where the motion of the gas phase in the third
dimension can be neglected. The 3D bed is schematically represented in Figure 4.10.
Figure 4.10. Schematic representation of the 3D bed.
All other parameter settings are the same as used for the twodimensional system
presented in table 4.1. In preliminary tests it became clear that a minimum depth of 5
particle diameters was required to ensure stable simulations. At smaller depths the
particles get stuck between the front and back walls and the hardsphere algorithm
breaks down. Eventually the simulation crashes due to particle overlap. With a depth
22 mm
0.5 m
0.15 m
Effect of Particle Properties…
___________________________________________________________________________________
119
of 22 mm (5.5 particle diameters) such problems did not occur and therefore this
depth was used in the simulations. The number of particles used in the 3D
simulations was 14,000 in order to match initial bed height in the 2D simulations.
6.1 Influence of collision parameters
The collision parameters turned out to have a key influence in the twodimensional
simulations and therefore their influence was also investigated using the three
dimensional model. The setup of the 3D simulations was the same as for the 2D
simulations. The bed was homogeneously fluidised with a gas inflow of 1.5 u
mf
.
In Figure 4.11 snapshots obtained from three different 3D simulations are presented.
A simulation with nonideal collision parameters (e = e
w
= 0.9, µ = µ
w
= 0.3) was
performed as well as a simulation with ideal collision parameters (e = e
w
= 1.0, µ = µ
w
= 0.0). A simulation with ideal particleparticle interaction but nonideal particlewall
interaction (e = 1.0, e
w
= 0.9, µ = 0.0, µw
= 0.3) was also carried out. The snapshots
presented in this figure are front views similar to the snapshots of the 2D simulations
in Figure 4.1. A number of particle layers was marked by colour based on the position
at t = 0 s in order to visualise the induced particle mixing during the fluidisation
process. For the 3D case it should be noted that the bed is 5.5 particle diameters deep
and this affects the visualisation: the particles near the front wall dominate the image.
The trends observed in the 2D simulations are clearly confirmed in the 3D
simulations. In the ideal case no bubbles are observed and after a vigorous expansion
the bed remains rather homogeneously fluidised. In the non ideal case bubbles are
observed and the bed height is continuously fluctuating. In the snapshots this may not
be very clear due to the visualisation technique employed here, but this fluctuating
motion could clearly be observed in animations. Also the potential energy of the
system showed these fluctuations in the non ideal case.
Chapter 4
___________________________________________________________________________________
120
Figure 4.11. Snapshots of particle configurations for three 3D simulations with u
g
= 1.5 u
mf
. Upper: e = e
w
= 1.0, µ = µw
= 0.0, middle: e = 1.0, e
w
= 0.9,
µ = 0.0, µ
w
= 0.3 bottom: e = e
w
= 0.9, µ = µ
w
= 0.3.
Effect of Particle Properties…
___________________________________________________________________________________
121
The particlewall interaction has a significant influence on the hydrodynamics. In the
snapshots this can be observed especially during the initial stages of the simulation.
Also the potential energy of this system shows stronger fluctuations than the ideal
system.
6.2 2D versus 3D
The snapshots of the 3D nonideal simulation in Figure 4.11 show less bubbling and
slugging than the snapshots in the 2D equivalent in Figure 4.1. This can be partially
due to the visualisation technique employed here but also the pressure fluctuations
inside the bed are less strong and have a lower frequency (2.3 Hz in 3D vs. 3.3 Hz in
2D for the nonideal case).
In Figure 4.12 the pressure 0.2 m above the centre of the distributor plate is plotted as
a function of time for the 2D and 3D simulation with ideal values for the collision
parameters.
Figure 4.12. Pressure 0.2 m above the centre of the distributor plate as a function of
time (e = e
w
= 1.0, µ = µw
= 0.0): 2D simulations compared to 3D
simulations
Although in both the 2D and the 3D simulations the fluctuations are not very strong,
the fluctuations in the 2D case are more pronounced compared to the 3D case. The
Chapter 4
___________________________________________________________________________________
122
RMS value is about 70 Pa higher for the 2D simulation. The same difference was
found for the non ideal case. These differences may be explained in terms of
improved statistics since the number of particles in a computational cell is much
higher in the 3D case. Also the absence of a degree of freedom of motion in the 2D
case can cause the system to show a more constrained behaviour and hence cause
stronger fluctuations.
In Figure 4.13 the pressure fluctuations in the 2D and 3D simulations are compared
for the case with ideal particle particle interaction and nonideal particlewall
interaction (e = 1.0, e
w
= 0.9, µ = 0.0, µ
w
= 0.3).
Figure 4.13. Pressure 0.2 m above the centre of the distributor plate as a function of
time (e
w
= 0.9, µ = 0.0, µw
= 0.3): 2D simulations compared to 3D
simulations.
For this case the fluctuations are stronger in the 3D simulation. The RMS value of
the fluctuations was 60 Pa higher in the 3D case. Baring in mind that in the other
cases the fluctuations in the 2D simulations were stronger than in the 3D simulations
it can be concluded that the particlewall interaction has a more pronounced effect on
the hydrodynamics in 3D simulations than in 2D simulations. This is obviously due
to the presence of the front and the back wall in the 3D simulations of the ‘flat’ bed
Effect of Particle Properties…
___________________________________________________________________________________
123
which renders a much higher number of particlewall collisions than in the 2D
simulations.
The 3D simulations are much more computational intensive than the 2D
simulations. It takes 6 minutes of CPU time per simulated second on a Silicon
Graphics Origin200 server (R10000 processor, 180 MHz) with the 2D model (2400
particles) whereas it takes 220 minutes per second to perform the same (nonideal)
simulation with the 3D model (14,000 particles).
7. Conclusions
A hardsphere Granular Dynamics model was used to study the effect of the particle
properties on the hydrodynamics of gasfluidised beds with homogeneous inflow
conditions. The bed dynamics is strongly affected by the values of the collision
parameters (i.e. the coefficients of restitution (e) and friction (µ)). When ideal particle
interaction was assumed (e = 1.0, µ = 0.0, i.e. no energy is dissipated in collisions)
nonrealistic bed dynamics is observed: no bubble or slug formation occurs and
particle mixing is poor. When realistic particle interaction was assumed (e = 0.9, µ =
0.3) bubble and slug formation are observed and the fluidisation behaviour is in much
better accordance with that inferred from experimental observations. The coefficient
of tangential restitution (β0
) has little influence on the overall behaviour. The particle
particle interaction parameters turned out to be much more important than the
particlewall interaction parameters in the twodimensional simulations. Pressure
fluctuations inside the bed are much stronger in the nonideal case than in the ideal
case. With intermediate values for the collision parameters the behaviour also changes
gradually with pressure fluctuations becoming less strong when the coefficients are
changed toward their ideal values. It turned out to be more convenient to determine
the root mean square (RMS) of the pressure fluctuations as a function of the energy
dissipation rate since this is affected both by the collision parameters and the system
dynamics. The RMS values of the fluctuations show an almost linear increase with the
energy dissipation rate during a simulation. The higher the energy dissipation rate the
more energy was put into the system by the drag force.
Chapter 4
___________________________________________________________________________________
124
Twodimensional simulations have been performed with particles which diameters
have been obtained from a lognormal distribution. The main influence of the poly
dispersity could be observed when the particles were in a rather close packing.
Pressure peaks inside a slugging fluidised bed decreased with increasing geometric
standard deviation of the lognormal size distribution during the initial stages of a
simulation starting from minimum fluidisation conditions. This indicates that it is
especially important to take a particle size distribution into account in systems where
locally very dense particle configurations can occur.
Also a threedimensional model (as far as the particles are concerned) was developed.
A minimum depth of the bed of 5 particle diameters was required to ensure stable
simulations. The 3D model confirmed the trends observed in the 2D model with
respect to the dependency of the overall bed hydrodynamics on the collision
parameters. Time series of the pressure inside the bed showed a lower frequency and a
lower RMS value in the 3D simulations than in the 2D case. The particlewall
interaction had a more pronounced influence on the hydrodynamics in the 3D
simulations than in the 2D simulations. As anticipated the 3D simulations required
considerably more CPU time than the 2D simulations.
Notation
C
d
drag coefficient, []
e coefficient of restitution, []
D
p
particle diameter, m
g gravitational acceleration, m/s
2
m
p
particle mass, kg
p pressure, Pa
r position vector, m
S
p
momentum source term N/m
3
T temperature, K
t time, s
Effect of Particle Properties…
___________________________________________________________________________________
125
u gas velocity vector, m/s
vp
particle velocity vector, m/s
V
p
particle volume, m
3
Greek symbols
β volumetric interphase momentum transfer coefficient, kg/(m
3
s)
β
0
coefficient of tangential restitution, []
ε void fraction, []
µ coefficient of friction, []
µg
gas viscosity, kg/ms
τ gas phase stress tensor, kg/ms
2
ρ density, kg/m
3
σ
g
Geometric Standard Deviation (GSD), m
Subscripts
CM Count Median
dsp dissipated
drg drag
GSD Geometric Standard Deviation
g gas phase
kin kinetic
mf minimum fluidisation
p particle
pot potential
rot rotation
tot total
w wall
References
Anderson, T.B. and Jackson, R., (1967). A fluid mechanical description of fluidized
beds (equations of motion), Ind. Eng. Chem., Fundam., 6, 527.
Chapter 4
___________________________________________________________________________________
126
Bleek, van den, C.M., and Schouten, J.C., (1993). Can deterministic chaos create
order in fluidizedbeds scaleup ?, Chem. Engng Sci., 48, 2367.
Chang, H. and Louge, M.Y., (1992), Fluid dynamic similarity of circulating fluidized
beds, Powder Technol., 70, 259.
Chen, Z., Gibilaro, L.G., and Foscolo, P.U., (1997a). Fluid pressure loss in slugging
fluidised beds, Chem. Engng Sci., 52, 55.
Chen, Z., Di Felice, R., Gibilaro, L.G., and Foscolo, P.U., (1997b). Effects of solid
frictional interactions on the applicability of the fluiddynamic scaling rules for
fluidisation, Chem. Eng. Comm., 157, 205.
Cundall, P.A. and Strack, O.D.L., (1979). A discrete numerical model for granular
assemblies. Géotechnique, 29, 47.
Foerster S.F., Louge, M.Y., Chang,H. and Allia, K. (1994). Measurements of the
collision properties of small spheres. Phys. Fluids,, 6, 1108.
Geldart, D., (1973). Types of gas fluidization, Powder Technol., 7, 285.
Gidaspow, D. (1994). Multiphase Flow and Fluidization, Academic Press, Boston,
USA.
Hoomans, B.P.B., Kuipers, J.A.M., Briels, W.J. and van Swaaij, W.P.M., (1996).
Discrete particle simulation of bubble and slug formation in a twodimensional gas
fluidised bed: a hardsphere approach, Chem. Engng Sci., 51, 99.
Hrenya, C.M. and Sinclair, J.L. (1997). Effects of particlephase turbule nce in gas
solid flows, AIChE J., 43, 853.
Kuipers, J.A.M., van Duin K.J., van Beckum, F.P.H. and van Swaaij, W.P.M., (1992).
A numerical model of gasfluidized beds, Chem. Engng Sci., 47, 1913
Effect of Particle Properties…
___________________________________________________________________________________
127
Langston, P.A., Tüzün, U. and Heyes, D.M., (1995). Discrete element simulation of
granular flow in 2D and 3D hoppers: dependence of discharge rate and wall stress on
particle interactions. Chem. Engng Sci., 50, 967.
Kawaguchi, T., Tanaka, T. and Tsuji, Y., (1998). Numerical simulation of two
dimensional fluidized beds using the discrete element method (comparison between
the two and threedimensional models), Powder Technol., 96, 129.
Mikami, T., Kamiya, H. and Horio, M., (1998). Numerical simulation of cohesive
powder behavior in a fluidized bed, Chem. Engng Sci., 53, 1927.
Nieuwland, J.J., van Sint Annaland, M., Kuipers, J.A.M. and van Swaaij, W.P.M.,
(1996), Hydrodynamic modelling of gasparticle flows in riser reactors, AIChE J., 37,
1009.
Richardson, J.F. and Zaki, W.N., (1954). Sedimentation and fluidization: part I.
Trans. Inst. Chem. Eng., 32, 35.
Schouten, J.C. and Bleek, van den, C.M., (1992). Chaotic hydrodynamics of
fluidization: consequences for scaling and modeling of fluid bed reactors, AIChE
Symp. Ser., No. 289 Vol. 88, 70.
Seville, J.P.K., U. Tüzün, and R. Clift, (1997). Processing of Particulate Solids,
Blackie Academic & Professional.
Sinclair, J.L. and Jackson, R. (1989). Gasparticle flow in a vertical pipe with particle
particle interactions, AIChE J., 35, 1473.
Thiel, W.J. and Potter, O.E., (1977). Slugging in Fluidized beds, Ind. Eng. Chem,,
Fundam. 16, 242.
Thornton, C., K. K. Yin, and M. J. Adams, 1996, Numerical simulation of the impact
fracture and fragmentation of agglomerates, J, Phys. D: Appl. Phys., 29, 424.
Chapter 4
___________________________________________________________________________________
128
Tsuji, Y., Kawaguchi, T. and Tanaka, T., (1993). Discrete particle simulation of two
dimensional fluidized bed, Powder Technol. 77, 79.
Wang, Y. and Mason, M.T., (1992). Twodimensional rigidbody collisions with
friction. J. Appl. Mech., 59, 635.
Wen, C.Y. and Yu, Y.H., (1966). Mechanics of fluidization. Chem. Eng. Prog. Symp.
Ser., 62 (62), 100.
Xu, B.H. and Yu, A.B., (1997). Numerical simulation of the gassolid flow in a
fluidized bed by combining discrete particle method with computational fluid
dynamics. Chem. Engng Sci. 52, 2785.
Zenz, F.A. and Othmer D.F., (1960). Fluidisation and fluidparticle systems, Reinhold
Publishing Corp. New York, USA.
HardSphere vs. Soft Sphere
______________________________________________________________________________________
129
Chapter 5.
GRANULAR DYNAMICS SIMULATION OF BUBBLE
FORMATION IN A GAS FLUIDISED BED: HARDSPHERE
VS. SOFTSPHERE APPROACH
Abstract:
Two types of Granular Dynamics models have been developed to study bubble formation
in a gasfluidised bed. For each individual particle the Newtonian equations of motion
are solved taking into account the external forces (i.e. gravity and drag). In the hard
sphere simulations a sequence of binary, instantaneous collisions is processed one
collision at a time while a collision model is used to describe the inelastic collisions with
friction. In the softsphere model the particleparticle and particlewall interactions are
taken into account through a linear spring/dashpot contact force model. The gas phase
hydrodynamics is described by the spatially averaged NavierStokes equations for two
phase flow. Bubble formation at a central jet in a bed of 0.2 m width and 0.3 m height
with 40,000 glass ballotini particles of 850 µm diameter was chosen as a test case for the
simulations. Preliminary simulations with the softsphere model revealed that the results
are rather insensitive to the time step used for the integration of the contact forces as well
as the value of the spring stiffness. The agreement between simulation and experiment
was improved by incorporation of a lognormal particle size distribution. Hardly any
difference could be observed between the results of the hardsphere and the softsphere
simulations that both compared well with the experiment. The computational efficiency of
the hardsphere simulations depends strongly on the dynamics of the system in contrast to
the computational efficiency of the softsphere simulations. The collision parameters
turned out to have a key influence in both types of simulations. When the collisions were
assumed to be fully elastic and perfectly smooth the agreement with the experiment was
much worse.
Chapter 5
______________________________________________________________________________________
130
Parts of this chapter are based on the papers:
B.P.B. Hoomans, J.A.M. Kuipers, W.J. Briels and W.P.M. van Swaaij, (1996). Discrete particle simulation
of bubble and slug formation in twodimensional gasfluidised beds: a hardsphere approach, Chem. Engng
Sci., 51, 99.
J.A.M. Kuipers, B.P.B. Hoomans and W.P.M. van Swaaij, (1998). Hydrodynamic models of gasfluidized
beds and their role for design and operation of fluidized bed chemical reactors, in Fluidization IX, L. S. Fan
and T.M. Knowlton (eds), 15.
B.P.B. Hoomans, J.A.M. Kuipers and W.P.M. van Swaaij, (1998). The influence of a particle size
distribution on the granular dynamics of dense gas fluidized beds: a computer simulation study, AIChE
Symp. Ser. No. 318, Vol. 94, 15.
HardSphere vs. Soft Sphere
______________________________________________________________________________________
131
1. Introduction
Due to increasing computer power Granular Dynamics simulations have become a very
useful and versatile research tool to study the hydrodynamics of gasfluidised beds. In
these simulations the Newtonian equations of motion for each individual particle in the
system are solved. Particleparticle and particlewall interactions are taken into account
directly which is a clear advantage over two fluid models which require closure relations
for the solidsphase stress tensor (Gidaspow (1994), Sinclair and Jackson (1989) and
Kuipers et al. (1992) among others). Furthermore a discrete particle model is very well
suited to take an arbitrary particle size distribution into account which is far more
complex, if not impossible, in case a continuum modelling approach is adopted.
Tsuji et al. (1993) developed a softsphere discrete particle model of a gasfluidised bed
based on the work of Cundall and Strack (1977). In their approach the particles are
allowed to overlap slightly and from this overlap the contact forces are subsequently
calculated. Kawaguchi et al. (1998) extended the model developed by Tsuji et al. (1993)
to threedimensions as far as the motion of the particles is concerned. Hoomans et al.
(1996) used a hardsphere approach in their discrete particle model of a gas fluidised bed
which implies that the particles interact through binary, instantaneous, inelastic collisions
with friction. Xu and Yu (1997) used a twodimensional softsphere model similar to the
one used by Tsuji et al. (1993) in their fluidised bed simulations but they also employed
techniques from hardsphere simulations in order to determine the instant when the
particles first come into contact precisely. Mikami et al. (1998) presented a two
dimensional softsphere model of a gasfluidised bed where the model developed by Tsuji
et al. (1993) was extended in order to take cohesive forces between the particles into
account.
In all the softsphere models used for fluidised bed simulations a linearspring/dashpot
model is used to describe the normal contact forces. In these simulations the value of the
Chapter 5
______________________________________________________________________________________
132
spring stiffness is chosen artificially low for computational convenience. It is assumed
that this is allowed but it has never been demonstrated that the simulation results are not
affected by the value of the spring stiffness. In order to investigate the effect of the spring
stiffness on the simulation results and to allow for a comparison with hardsphere
simulations, a softsphere linear spring/dashpot model was developed. Since for both
types of models the gas phase hydrodynamics and the twoway coupling was incorporated
in exactly the same manner it is possible to make a direct comparison that highlights only
the effects due to the choice of model. The formation of a single bubble at a central orifice
in a gas fluidised bed was chosen as a test case because of the availability of a well
defined experiment (Hoomans et al., 1996) that allows for validation.
2. Models
Since a detailed description of the models is presented in Chapter 2 and Chapter 3, the
key features will be summarised only briefly here.
2.1 Hardsphere granular dynamics
The collision model as originally developed by Wang and Mason (1992) is used to
describe a binary, instantaneous, inelastic collision with friction. The key parameters of
the model are the coefficient of restitution (0 ≤ e ≤ 1) and the coefficient of friction (µ ≥
0). Foerster et al. (1994) have shown that also the coefficient of tangential restitution (0 ≤
β
0
≤ 1) should be used in order to describe the collision dynamics more accurately. These
three collision parameters are all included in the model.
In the hardsphere approach a sequence of binary collisions is processed. This implies that
a collision list is compiled in which for each particle a collision partner and a
corresponding collision time is stored. A constant time step is used to take the external
forces into account and within this time step the prevailing collisions are processed
sequentially. In order to reduce the required CPU time neighbour lists are used. For each
HardSphere vs. Soft Sphere
______________________________________________________________________________________
133
particle a list of neighbouring particles is stored and only for the particles in this list a
check for possible collisions is performed.
2.2 Softsphere granular dynamics
The linear spring/dashpot model (Cundall and Strack, 1979) is the most popular soft
sphere granular dynamics model since it was used by Tsuji et al. (1993), Schwarzer
(1995), Xu and Yu (1997), Kawaguchi et al. (1998) and Mikami et al. (1998). For a
review of various contact force models used in softsphere simulations the reader is
referred to Walton (1992) or Schäfer et al. (1996).
In softsphere models the following equations of motion are used:
external contact
dt
d
m F F
r
+ ·
2
2
, (5.1)
T ·
ω
dt
d
I , (5.2)
where r is the position vector of the centre of the particle, m is the mass of the particle,
F
contact
is the contact force acting on the particle, F
external
is the net external force acting
on the particle, ω is the rotation velocity, T is the torque acting on the particle and I is the
moment of inertia of the particle. In this paragraph the focus will be on the contact forces
between the particles, the external forces will be discussed in the following paragraph.
The contact forces are given by:
ab n n ab n n n,ab
k
,
v n F η ξ − − · (5.3)
t,ab t t t t,ab
k v ? F η − − · (5.4)
If however the following relation is satisfied:
Chapter 5
______________________________________________________________________________________
134
ab n t,ab ,
F F µ > , (5.5)
then sliding occurs and the tangential force is given by:
ab n,ab t,ab
t F F µ − · . (5.6)
For contacts between particles and walls, the walls are assumed to be non moving and of
infinite mass just like in the hardsphere model.
The resulting force and torque for particle a are now simply obtained by adding the pair
wise contributions of all the particles that are in contact with a:
( )
∑
+ ·
b
t,ab n,ab a contact
F F F
,
, (5.7)
( )
∑
× ·
b
t,ab ab a a
R F n T . (5.8)
Since the contact forces are in general at least an order of magnitude larger than the
external forces a separation of time scales was introduced. By default at each time step
DT the external forces were taken into account while at 0.1DT the equations of motion
were solved by taking only the contact forces into account. Hence the external forces
remain constant during a time step DT.
2.3 External forces
The incorporation of external forces differs somewhat from the approach followed by
Hoomans et al. (1996). In this work the external forces are used in accordance with those
implemented in the two fluid model presented by Kuipers et al. (1992) where, of course,
the forces now act on a single particle:
HardSphere vs. Soft Sphere
______________________________________________________________________________________
135
( )
( ) m
d
dt
m
V
V p
p
p
p
p
p p
v
g u v · +
−
− − ∇
β
ε 1
, (5.9)
where m
p
represents the mass of a particle, v
p
its velocity, u the local gas velocity and V
p
the volume of a particle. In equation (5.9) the first term is due to gravity and the third term
is the force due to the pressure gradient. The second term is due to the drag force where β
represents an interphase momentum exchange coefficient as it usually appears in two
fluid models. For low void fractions (ε < 0.80) β is obtained from the wellknown Ergun
equation:
( )
( )
p
p
g
p
g
D D
v u − − +
−
·
ρ
ε
µ
ε
ε
β 1 75 . 1
1
150
2
2
, (5.10)
where D
p
represents the particle diameter, µ
g
the viscosity of the gas and ρ
g
the density of
the gas. For high void fractions (ε ≥ 0.80) the following expression for the interphase
momentum transfer coefficient has been used which is basically the correlation presented
by Wen and Yu (1966) who extended the work of Richardson and Zaki (1954):
( )
65 . 2
1
4
3
−
−
−
· ε ρ
ε ε
β
p g
p
d
D
C v u . (5.11)
The drag coefficient C
d
is a function of the particle Reynolds number and given by:
( )
¹
¹
¹
'
¹
≥
<
·
1000 Re 44 . 0
1000 Re Re 0.15 + 1
Re
24
0.687
p
p p
p
d
C
,
(5.12)
where the particle Reynolds number (Re
p
) in this case is defined as follows:
Chapter 5
______________________________________________________________________________________
136
g
p p g
p
D
µ
ερ v u −
· Re . (5.13)
For the integration of equation (5.9) a simple explicit first order scheme was used to
update the velocities and positions of the particles.
2.4 Gas phase hydrodynamics
The calculation of the gas phase hydrodynamics mainly follows the lines presented by
Kuipers et al. (1992). It is based on the numerical solution of the following set of partial
differential equations that can be seen as a generalised form of the NavierStokes
equations for a gas interacting with a solid phase as originally derived by Anderson and
Jackson (1967).
Continuity equation gas phase:
( )
( )
∂ ερ
∂
ερ
g
g
t
+ ∇⋅ · u 0. (5.14)
Momentum equation gas phase:
( )
( ) ( )
∂ ερ
∂
ερ ε ε ερ
g
g p g g
t
p
u
uu S g + ∇ ⋅ · − ∇ − − ∇ ⋅ + τ . (5.15)
In this work transient, twodimensional, isothermal (T = 293 K) flow of air at atmospheric
conditions is considered. The constitutive equations can be found in Chapter 3. There is
one important modification with respect to the model presented by Hoomans et al. (1996)
and that deals with the way in which the twoway coupling between the gas phase and the
dispersed particles is established. In the present model the reaction force to the drag force
HardSphere vs. Soft Sphere
______________________________________________________________________________________
137
exerted on a particle per unit of volume is fed back to the gas phase through the source
term S
p
which has the dimension of force per unit of volume N/m
3
.
3. Preliminary simulations
At first several preliminary simulations were performed using the softsphere model to
investigate the influence of the spring stiffness on the simulations results. The formation
of a bubble at a central orifice was chosen as a test case. The parameter settings for these
simulations are presented in table 5.1.
Table 5.1. Parameter settings bubble formation simulations
Particles: Bed:
Shape spherical width 195 mm
diameter, D
p
850 µm height 300 mm
Material ballotini glass orifice diameter, d
o
15 mm
density, ρ
p
2930 kg/m
3
cell width, DX 5 mm
e 0.96 cell height, DY 5 mm
e
w
0.86 number xcells, NX 39
µ 0.15 number ycells, NY 60
µ
w
0.15
Number, N
p
40,000 time step, DT 10
4
s
The values of the collision parameters are taken from Hoomans et al. (1996). The
coefficient of tangential restitution (β
0
) is assumed to be equal to zero. Unless indicated
otherwise a time step DT = 10
4
s is used while DT
contact
= 0.1 DT. At the side walls no
slip boundary conditions were applied and at the upper boundary cells a prescribed
pressure was chosen. At the bottom row of cells the gas inflow velocity was set equal to
the minimum fluidisation velocity of the glass particles (u
mf
= 0.5 m/s). In the three
central cells the gas velocity was set equal to 5 u
mf
during the first 0.2 s of the simulation
to generate a single bubble.
Chapter 5
______________________________________________________________________________________
138
The effect of the spring stiffness was studied for a system consisting of particles of
uniform size. The contact time for such particles within the framework of the linear
spring/dashpot model can be calculated as follows:
( )
n
p
contact
k
e m
t
2
ln
2 2
+
·
π
. (5.16)
The values of the spring stiffness were varied from 1 to 10,000 N/m. The contact times
for these values of the spring stiffness are presented in table 5.2. Unless indicated
otherwise the value of the stiffness of the tangential spring was chosen as 2/7 times the
normal spring stiffness (k
t
= 2/7 k
n
). Note that the contact times in the linear spring/dash
pot model do not depend on the impact velocity.
Table 5.2. Contact times and results of simulations
*
for various values of the normal
spring stiffness. In all cases DT = 10
4
s.
k
n
N/m
t
contact
s
DT
contact
.
10
4
s
DT
contact
.
10
5
s
DT
contact
.
10
6
s
1 3.05 10
3
No No No
10 9.64 10
4
Yes Yes Yes
100 3.05 10
4
No Yes Yes
1000 9.64 10
5
No Yes Yes
10000 3.05 10
5
No No Yes
*
No: simulation not completed successfully, Yes: simulation completed successfully
With a normal spring stiffness of 1 N/m no successful simulations could be performed.
The repulsive force is simply not strong enough compared to the external forces (gravity
and drag) to prevent the particles from excessive overlap. During the simulation this leads
to extremely low local void fractions (ε < 0.2) which eventually causes the simulation to
crash. With a spring stiffness of 10 N/m or higher such problems were not encountered.
HardSphere vs. Soft Sphere
______________________________________________________________________________________
139
However, at higher values of the spring stiffness the time step for the numerical
integration of the contact forces should be chosen sufficiently small in comparison with
the contact time in order to perform stable simulations. It turned out that in the case of k
n
= 100 N/m (t
contact
= 3.05 10
4
s) a time step of 10
4
s was not sufficiently small to perform
a stable simulation. However, when a time step of 5 10
5
s was used, the simulation was
successfully completed. Hence the ratio of the contact time and the time step should at
least be greater than 3 (t
contact
/DT
contact
> 3) to ensure the stability of the simulation. In
Figure 5.1 the sum of the potential energy of all the particles in the system is plotted as a
function of time for several time steps (DT
contact
).
Figure 5.1. The potential energy of the system as a function of time for several time
steps. In all cases k
n
= 100 N/m.
From Figure 5.1 it can be observed that there is hardly any influence of the time step used
for the integration of the contact forces on the overall results of the simulation which is
rather surprising. Snapshots of the simulations do not reveal a significant difference
either. There is however a big difference in CPU time because a simulation requires
nearly 10 times as much CPU time when the time step is reduced by a factor 10. In the
simulations presented by Tsuji et al. (1993), Kawaguchi et al. (1998) and Mikami et al.
(1998) a time step was used that was 3.7 times smaller than the contact time for the
particles used in their systems. In the simulations presented by Xu and Yu (1997) the
Chapter 5
______________________________________________________________________________________
140
ratio of the contact time and the time step was 6.2. Although a ratio of 3.7 appears to be
rather close to the stability limit the results presented in Figure 5.1 suggest that it is not
required to choose a very small time step. This result is somewhat surprising since in
Chapter 2, Figure 2.14, it was demonstrated that the effect of the time step on the time
resolution of the contact force was rather large. However, the error in the energy balance
over the collision was rather low in that particular case. This could explain the minor
influence of DT
contact
on the simulation results. The results presented in Figure 5.1 also
indicate that the technique employed by Xu and Yu (1997) to determine the precise
instant of first contact within a time step does not improve the results of a simulation.
Additional simulations were performed using the hardsphere model where the overall
time step (DT) for the integration of the external forces was varied (DT =10
3
s, DT = 10
4
s, DT = 10
5
s). In all these cases DT
contact
= 10
5
s. For all three time steps very similar
results were obtained. The simulation with DT = 10
5
s was significantly slower in terms
of CPU time than the simulation with DT = 10
4
s but the difference between the
simulations with DT = 10
3
s and DT = 10
4
s was remarkably small. This is due to the fact
that in the simulation with DT = 10
3
s a larger neighbour list was required since this list
could only be updated at every time step.
In Figure 5.2 snapshots at t = 0.2 s are presented for simulations with various values of
the spring stiffness: k
n
= 10 N/m, k
n
= 100 N/m and k
n
= 10000 N/m. A simulation with k
n
= 1000 N/m was not included in the figure but showed very similar results. By default the
value for the tangential spring stiffness is set equal to 2/7 times the value of normal
spring stiffness (k
t
= 2/7 k
n
) for reasons discussed in Chapter 2. In Figure 5.2 a snapshot at
t = 0.2 s is included of a simulation where the value for the tangential spring stiffness is
set equal to the value of normal spring stiffness (k
t
= k
n
= 100 N/m).
All snapshots in Figure 5.2 appear to be surprisingly identical. Apparently the value of
the spring stiffness has a negligible effect on the simulation results as long as it is high
enough compared to the net external force to prevent the particles from overlapping too
much (more than roughly 10% of the particle diameter). However, the choice of spring
HardSphere vs. Soft Sphere
______________________________________________________________________________________
141
stiffness does affect the CPU time required for a simulation significantly since the time
step used for the integration of the contact forces should at least be three times smaller
than the contact time corresponding to the spring stiffness used in the simulation.
a) b)
c) d)
Figure 5.2. Snapshots of particle configurations at t = 0.2 s for various values of the
spring stiffness a) k
n
= 10 N/m b) k
n
= 10000 N/m c) k
n
= 100 N/m d) k
n
=
k
t
= 100 N/m.
Chapter 5
______________________________________________________________________________________
142
In Chapter 2 it was demonstrated that the value of the tangential spring stiffness should
be set equal to 2/7 times the value of normal spring stiffness (k
t
= 2/7 k
n
) in order to
ensure equal normal and tangential contact times. From the snapshot in Figure 5.2d (k
t
=
k
n
= 100 N/m) hardly any difference can be observed with the snapshot in Figure 5.2c (k
n
= 100 N/m, k
t
= 2/7 k
n
). This is however not surprising because of the low value of the
coefficient of friction (µ) that causes a considerable amount of sliding contacts. For the
same reason the coefficient of tangential restitution (β
0
) had a negligible influence on the
simulation results (compared to the coefficients of normal restitution (e) and friction)
presented in Chapter 4. Furthermore the effect is most pronounced for sticking contacts
with a coefficient of tangential restitution equal to one (β
0
= 1.0) while in the simulation
presented in Figure 5.2 the coefficient of tangential restitution is set equal to zero (β
0
=
0.0). However, an additional simulation with β
0
= 1.0 showed very similar results and
hence the effect due to the low coefficient of friction is dominant.
4. Experimental validation
The hardsphere and the softsphere codes were compared with each other and validated
using data of a bubble formation experiment obtained from the experimental setup
described below.
4.1 Experimental
The set up used for the experiment is shown schematically in Figure 5.3 (Nieuwland,
1994). The primary fluidising gas was introduced into the bed through two equally sized
distribution chambers. From these distribution chambers the primary fluidisation gas was
introduced into the fluidised bed section through a porous sintered stainless steel plate
with a mean pore diameter of 10 µm. This porous plate, which served as the main
distributor of the primary fluidising gas, was equipped with a central rectangular pipe
(internal dimensions: 15.0 x 15.0 mm) to inject secondary fluidising gas. This rectangular
HardSphere vs. Soft Sphere
______________________________________________________________________________________
143
jet was covered with a piece of stainless steel gauze to prevent the particles from entering
it.
Figure 5.3. Schematic representation of the experimental setup.
During the experiments three different gas flow rates were adjusted independently by
means of calibrated thermal mass flow controllers: a gas flow through the distributor
section to keep the bed at incipient fluidisation conditions, a gas flow through the
rectangular orifice to fluidise the particles above the orifice and a gas flow for the bubble
formation at the orifice. Prior to bubble injection, the secondary gas was purged, while
the primary gas flowed through the gas distributor and the orifice to keep the bed at
minimum fluidisation conditions. Rapidly responding, computer controlled, magnetic
valves were used to inject secondary gas through the orifice into the bed to form a bubble
and to purge the primary gas during the period of injection. In the experiment the primary
gas was introduced at a superficial velocity of 0.5 m/s (u
mf
) and during a period of 0.2 s
secondary gas was injected through the central orifice at a superficial velocity of 2.5 m/s
(5 u
mf
). An SVHS video camera was used to observe the bed during the whole process
Chapter 5
______________________________________________________________________________________
144
of bubble formation. A sieve fraction of glass ballotini particles (ρ
p
= 2930 kg/m
3
) with
diameters between 800 and 900 µm was used.
4.2 Influence of a particle size distribution
In order to test whether the simulation results were improved by taking a particle size
distribution into account, a simulation with particles of uniform size was compared with a
simulation where a log normal particle size distribution was taken into account (Hoomans
et al., 1998). The results of these simulations were validated using data of a bubble
formation experiment obtained from the experimental set up described above.
The main parameter settings for the simulations are presented in Table 5.1. Apart form a
simulation with particles of a uniform diameter of 850 µm a simulation was performed
where the particle diameters were obtained from a lognormal distribution with a count
median average of 850 µm and a geometric standard deviation of 50 µm. The particle
size distribution employed in the simulation is presented in Figure 5.4.
Figure 5.4. Particle size distribution for the bubble formation simulations
HardSphere vs. Soft Sphere
______________________________________________________________________________________
145
Snapshots at t = 0.2 s of the two simulations and corresponding video images of the
experiment are presented in Figure 5.5. The simulations presented here were performed
using the hardsphere code.
a) b) c)
Figure 5.5. Snapshots at t = 0.2 s of: a) experiment, b) simulation with uniform
particles, c) simulation with lognormal particle size distribution.
It can be observed that in the system with particles of uniform size small ‘satellite’
bubbles appear above and alongside the main bubble. This is not observed in the
experiment nor in the simulation with the polydisperse particle assembly. These
‘satellite’ bubbles are probably due to locally low void fraction due to close packing. In a
system consisting of particles of uniform size such locally low void fractions can occur
since it is possible for an assembly of particles to gather in a lattice like structure. When
locally low void fractions prevail the drag force (which depends strongly on the void
fraction) can become strong enough to generate a bubble. The size of the main bubble
agrees reasonably well with the experiment for both of the (2D) simulations which is
rather encouraging especially since the model parameters were obtained on beforehand
and independently.
Chapter 5
______________________________________________________________________________________
146
4.3 Hardsphere vs. softsphere
The simulations were performed using the parameter settings presented in Table 5.1. For
the softsphere simulations a normal spring stiffness (k
n
) of 100 N/m was used and the
time step for the integration of the contact forces (DT
contact
) was set equal to 10
5
s. In
both simulations the particle size distribution presented in Figure 5.4 is taken into
account. In Figure 5.6 snapshots of the hardsphere (middle) and the softsphere (right)
simulations are presented together with the corresponding pictures of the experiment
(left). The initial conditions were identical in both simulations and hence there was no
overlap between particles at t = 0 s. which is required for the hardsphere simulation.
In the snapshots presented in Figure 5.6 it can be observed that the bubble is being
formed at the central orifice during the first 0.1 s which is in good agreement with the
experiment. After 0.1 s the bubble detaches from the distributor plate and starts to rise in
the bed. Since the jet remains active during the first 0.2 s the formation of a second
bubble is initiated but the jet does not remain active long enough for this second bubble to
grow to full size. After 0.1 s the shape of the bubble predicted in the simulations starts to
deviate from the bubble shape observed in the experiment although the main bubble size
as well as the position of the bubble in the bed is fairly well predicted. The bubble shape
observed in the experiment is more round and the gassolid boundary appears to be
sharper than in the simulations. This can be very well due to the twodimensional nature
of the simulations. Although the experimental setup was quasi twodimensional the bed
was still more than 17 particle diameters deep (16 mm) which is a significant difference
with the simulations where the particle motion is restricted to two dimensions. In addition
a higher order numerical scheme for the convective fluxes in the gas phase momentum
conservation equations is known to yield results with sharper bubble boundaries in
Eulerian simulations. The use of such a higher order scheme (like for instance the Barton
scheme) may also improve the bubble definition in (already 2D) Lagrangian simulations.
HardSphere vs. Soft Sphere
______________________________________________________________________________________
147
a) t = 0.05 s
b) t = 0.10 s
Figure 5.6. Snapshots of bubble formation simulations compared with an experiment.
From left to right: experiment, hardsphere simulation and softsphere
simulation, a) t = 0.05 s b) t = 0.10 s…
Chapter 5
______________________________________________________________________________________
148
c) t = 0.15 s
d) t = 0.20 s
Figure 5.6. (…continued) Snapshots of bubble formation simulations compared with
an experiment. From left to right: experiment, hardsphere, softsphere c) t
= 0.15 s d) t = 0.20 s.
HardSphere vs. Soft Sphere
______________________________________________________________________________________
149
It is remarkable however that hardly any difference can be observed between the results
of the hardsphere and the softsphere simulations. The snapshots appear to be almost
identical. In the hardsphere all collisions are assumed to be binary whereas the soft
sphere code is able to handle multiple contacts. Since the results show hardly any
difference it can be concluded that the assumption in the hardsphere code that all
collisions are binary is not limiting. There was however a difference in computational
efficiency between the two simulations. The hardsphere simulation required about 220
minutes of CPU time on a Silicon Graphics Origin200 server with a R10000 processor
(180 MHz, 1Mb cache memory) for a simulation of 0.3 s whereas the softsphere
simulation required 250 minutes of CPU time on the same machine. The main difference
between the two simulations is that the softsphere simulation progresses in time at a
rather constant speed whereas this is certainly not the case in the hardsphere simulation.
In the hardsphere simulation the first 0.2 s (with an active jet) only require 30% of the
total amount of CPU time. This is due to the strong increase in the number of collisions
after 0.2 s when the jet is no longer active. From t = 0.0 s to t = 0.2 s a total number of
37,8 million collisions was processed whereas from t = 0.2 s to t = 0.3 s the respectable
number of 110 million collisions was processed. Hence the computational efficiency of a
hardsphere simulation depends strongly on the dynamics of the system.
4.4 Effect of collision parameters
Kuipers et al. (1998) demonstrated the effect of the collision parameters on bubble
formation at a single, central orifice using a hardsphere simulation. When the collisions
were assumed to be fully elastic and perfectly smooth (e = 1.0, µ = 0.0, also referred to as
ideal collision parameters) no bubble could be observed after 0.2 s. The difference with
the simulation where realistic values for the collision parameters were specified (e = 0.96,
µ = 0.15, also referred to as nonideal collision parameters) was tremendous. This
simulation was repeated here using the softsphere code with, apart from the collision
parameters, the same parameter settings as used for the simulation in the previous
paragraph. Snapshots of both the hardsphere and the softsphere simulations are
presented in Figure 5.7 together with video images of the experiment.
Chapter 5
______________________________________________________________________________________
150
a) t = 0.1 s
b) t = 0.2 s
Figure 5.7. Snapshots of bubble formation simulations with ideal collision parameter
settings (e = 1.0, µ = 0.0, β
0
= 0.0) compared with an experiment. From
left to right: experiment, hardsphere, softsphere a) t = 0.1 s b) t = 0.2 s.
HardSphere vs. Soft Sphere
______________________________________________________________________________________
151
In Figure 5.7 it can clearly be observed that the result presented earlier by Kuipers et al.
(1998) obtained with a hardsphere code is reproduced here with a softsphere code. The
snapshot at t = 0.1 s is still quite similar to the snapshot of the simulation with the non
ideal collision parameters presented in Figure 5.6 as well as the experiment. But in the
snapshot at t = 0.2 s in Figure 5.7 the bubble seems to have disappeared which is not in
agreement with the snapshot at t = 0.2 s in Figure 5.6 and more importantly the
experiment. This stresses the importance of the collision parameters in granular dynamics
simulations of gasfluidised beds. When nonrealistic values for the collision parameters
are specified nonrealistic behaviour is predicted. This is found in both the hardsphere
and softsphere granular dynamics simulations.
5. Conclusions
A hardsphere and a softsphere granular dynamics model of a gas fluidised bed have
been developed. These twodimensional models were compared with each other and
validated experimentally using the formation of a single bubble at a central orifice as a
test case. Preliminary simulations with the softsphere code revealed that the time step
used for the integration of the contact forces should be more than 3 times smaller than the
contact time. The results of the simulations did not change significantly when the time
step for the integration of the contact forces was decreased. The results of the simulations
were also rather insensitive to the value of the normal spring stiffness. The value of the
normal spring stiffness should be high enough compared to the net external force in order
to prevent the particles from overlapping too much (more than roughly 10% of their
diameter). Although the use of very high values for the spring stiffness does not change
the results significantly it requires more CPU time since smaller time steps have to be
used. A simulation were the value of the tangential spring stiffness was set equal to the
value of the normal spring stiffness (k
t
= k
n
) showed hardly any difference with the
default simulations where k
t
= 2/7 k
n
.
Chapter 5
______________________________________________________________________________________
152
The incorporation of a particle size distribution turned out to improve the agreement
between simulation and experiment. In a bubble formation simulation with particles of
uniform size small ‘satellite’ bubbles appeared above and alongside the main bubble.
This could not be observed in the experiment nor in the simulation where a log normal
particle size distribution was taken into account.
A hardsphere and a softsphere simulation where a log normal particle size distribution
was taken into account were compared with each other and with an experiment. Both
simulations agreed rather well with the experiment especially during the initial stages.
The main bubble size and the position of the bubble in the bed were found to be in good
agreement with the experiment. The shape of the bubble observed in the experiment was
more round than the one observed in the simulations. Also the gassolid boundary
appeared to be sharper in the experiment than in the simulations. This is probably due to
the twodimensional nature of the simulations. Further improvement can be achieved by
extending the model to three dimensions because although the bed used in the experiment
was quasi twodimensional it was still 16 mm deep. However, a full 3D system
simulation of this system would require about 750,000 particles which is extremely
challenging even for state of the art computing resources of today. Applying periodic
boundary conditions in the zdirection would reduce the total number of particles required
in the simulation but in that case the influence of the front and back walls would be
eliminated as well. Another improvement can be achieved by employing a higher order
numerical scheme for the convective fluxes in the gas phase momentum equations that
suffer less from numerical diffusion. Results of Eulerian simulations with such higher
order schemes show sharper bubble boundaries and that could also be the case for
Lagrangian simulations. The results of both the hardsphere and the softsphere
simulation were nearly identical. Therefore it can be concluded that the assumption in the
hardsphere code that all collisions are binary is not limiting. However, there was a
difference in computational efficiency. Where the softsphere simulation progresses
through the simulation at a rather constant speed the computational efficiency of the hard
sphere simulation depends strongly on the dynamics of the system. The more collisions
HardSphere vs. Soft Sphere
______________________________________________________________________________________
153
occur the slower the progress. The overall CPU time requirements of the two codes were
very similar for a simulation of 0.3 s.
The collision parameters (i.e. coefficients of restitution and friction) turned out to have a
key influence on the bubble formation process. When the collisions were assumed to be
fully elastic and perfectly smooth (e = 1.0, µ = 0.0) no bubble could be observed after 0.2
s. This result was obtained with both the hardsphere and the softsphere model.
Notation
C
d
drag coefficient, []
D
o
orifice width, m
D
p
particle diameter, m
DT time step, s
DX horizontal computational cell dimension, m
DY vertical computational cell dimension, m
e coefficient of restitution, []
F force, N
g gravitational acceleration, m/s
2
I moment of inertia, kgm
2
k spring stiffness, N/m
m particle mass, kg
N
p
number of particles, []
NX number of computational cells in xdirection, []
NY number of computational cells in ydirection, []
n normal unit vector, []
p pressure, Pa
R
p
particle radius, m
r position vector, m
S
p
momentum source term, N/m
3
Chapter 5
______________________________________________________________________________________
154
T torque, Nm
T temperature, K
t time, s
u gas velocity vector, m/s
v particle velocity vector, m/s
V
p
particle volume, m
3
Greek symbols
β volumetric interphase momentum transfer coefficient, kg/(m
3
s)
β
0
coefficient of tangential restitution, []
ε void fraction, []
µ coefficient of friction, []
µ
g
gas viscosity, kg/ms
η damping coefficient, Ns/m
τ gas phase stress tensor, kg/ms
2
ρ density, kg/m
3
ξ displacement, m
Subscripts
g gas phase
mf minimum fluidisation
p particle
w wall
References
Anderson, T.B. and Jackson, R., (1967). A fluid mechanical description of fluidized beds
(equations of motion), Ind. Eng. Chem., Fundam., 6, 527.
HardSphere vs. Soft Sphere
______________________________________________________________________________________
155
Cundall, P.A. and Strack, O.D.L., (1979). A discrete numerical model for granular
assemblies. Géotechnique, 29, 47.
Foerster S.F., Louge, M.Y., Chang,H. and Allia, K. (1994). Measurements of the
collision properties of small spheres. Phys. Fluids,, 6, 1108.
Gidaspow, D. (1994). Multiphase Flow and Fluidization, Academic Press, Boston, USA.
Hoomans, B.P.B., Kuipers, J.A.M., Briels, W.J. and van Swaaij, W.P.M., (1996). Discrete
particle simulation of bubble and slug formation in a twodimensional gas fluidised bed: a
hardsphere approach, Chem. Engng Sci., 51, 99.
Hoomans, B.P.B., Kuipers, J.A.M. and van Swaaij, W.P.M., (1998). The Influence of a
Particle Size Distribution on the Granular Dynamics of Dense GasFluidized beds: a
Computer Simulation Study”, AIChE Symp. Ser. No. 318, 94, 15.
Kuipers, J.A.M., van Duin K.J., van Beckum, F.P.H. and van Swaaij, W.P.M., (1992). A
numerical model of gasfluidized beds, Chem. Engng Sci., 47, 1913
Kuipers, J.A.M., Hoomans, B.P.B. and van Swaaij, W.P.M., (1998). Hydrodynamic
Models of GasFluidized Beds and their Role for Design and Operation of Fluidized Bed
Chemical Reactors”, in Fluidization IX, L.S. Fan and T.M. Knowlton eds., 15.
Kawaguchi, T., Tanaka, T. and Tsuji, Y., (1998). Numerical simulation of two
dimensional fluidized beds using the discrete element method (comparison between the
two and threedimensional models), Powder Technol., 96, 129.
Mikami, T., Kamiya, H. and Horio, M., (1998). Numerical simulation of cohesive powder
behavior in a fluidized bed, Chem. Engng Sci., 53, 1927.
Chapter 5
______________________________________________________________________________________
156
Nieuwland, J.J., (1994). Hydrodynamic modelling of gassolid twophase flows, PhD
Thesis, University of Twente, The Netherlands.
Richardson, J.F. and Zaki, W.N., (1954). Sedimentation and fluidization: part I. Trans.
Inst. Chem. Eng., 32, 35.
Schäfer, J., Dippel, S. and Wolf, D.E., (1996). Force schemes in simulations of granular
materials, J. Phys. I France 6, 5.
Schwarzer, S., (1995). Sedimentation and flow through porous media: Simulating
dynamically coupled discrete and continuum phase, Phys. Rev. E. 52, 6461.
Sinclair, J.L. and Jackson, R. (1989). Gasparticle flow in a vertical pipe with particle
particle interactions, AIChE J., 35, 1473.
Tsuji, Y., Kawaguchi, T. and Tanaka, T., (1993). Discrete particle simulation of two
dimensional fluidized bed, Powder Technol. 77, 79.
Walton, O.R., (1992). Numerical simulation of inelastic, frictional particleparticle
interactions, in Particulate TwoPhase Flow, Part I, edited by M.C. Roco, Butterworth
Heinemann, Boston, 884.
Wang, Y. and Mason, M.T., (1992). Twodimensional rigidbody collisions with friction.
J. Appl. Mech., 59, 635.
Wen, C.Y. and Yu, Y.H., (1966). Mechanics of fluidization. Chem. Eng. Prog. Symp.
Ser., 62 (62), 100.
Xu, B.H. and Yu, A.B., (1997). Numerical simulation of the gassolid flow in a fluidized
bed by combining discrete particle method with computational fluid dynamics. Chem.
Engng Sci. 52, 2785.
Segregation Phenomena
______________________________________________________________________________________
157
Chapter 6.
GRANULAR DYNAMICS SIMULATION OF
SEGREGATION PHENOMENA IN BUBBLING GAS
FLUIDISED BEDS
Abstract:
Granular dynamics simulations of gasfluidised beds were performed in order to simulate
segregation phenomena in systems consisting of particles of different size and density. In
the model the gasphase hydrodynamics is described by the spatially averaged Navier
Stokes equations for twophase flow. For each solid particle the Newtonian equations of
motion are solved taking into account the particleparticle and particlewall interactions.
The (2D) model was applied to a system consisting of particles of three different densities
but equal size and to systems consisting of particles of two different sizes but equal
density. In the different density system segregation was observed over a time scale of
several seconds. In the different size systems segregation was also observed but a clear
steady state was not reached since the larger particles were continuously transported to
the upper regions of the bed in bubble wakes. A statistical analysis of the segregation in
terms of mass fraction distributions is presented. When the particleparticle and particle
wall interactions were assumed to be fully elastic and perfectly smooth, segregation
occurred very fast and was almost complete due to the absence of bubbles. Preliminary
experimental validation showed rather poor agreement between simulation and
experiment. The simulation predicts segregation at a lower gas velocity than used in the
experiment.
Chapter 6
______________________________________________________________________________________
158
Parts of this chapter are based on the following papers:
B.P.B. Hoomans, J.A.M. Kuipers and W.P.M. van Swaaij, (1998). Discrete Particle
Simulation of Segregation Phenomena in Dense GasFluidized Beds, in Fluidization IX,
L.S. Fan and T.M. Knowlton (eds),485.
B.P.B. Hoomans, J.A.M. Kuipers and W.P.M. van Swaaij, (1999). Granular Dynamics
Simulation of Segregation Phenomena in Dense GasFluidised beds, Powder Technology
(in press).
Segregation Phenomena
______________________________________________________________________________________
159
1. Introduction
Segregation phenomena can play an important role in fluidised systems consisting of
particles of different size and/or density. Typical examples of such processes are fluidised
bed polymerisation and fluidised bed granulation among others. In order to improve the
performance of these processes detailed knowledge about the distribution of the different
solids species throughout the bed in different operating conditions is required. In a system
consisting of particles of equal density but different size the bigger (heavier) particles tend
to reside at the bottom of the bed if the fluidisation velocity does not exceed the minimum
fluidisation velocity (u
mf
)
of the big particles. The big particles are in this case referred to
as jetsam. The smaller (lighter) particles show the tendency to float and reside at the bed
surface. These particles are referred to as flotsam. At gas velocities much higher than the
u
mf
of the jetsam better mixing is normally achieved. In case of a system consisting of
particles of equal size but different density the denser particles are referred to as jetsam
whereas the lighter particles are referred to as flotsam.
Segregation phenomena in gasfluidised beds have been the subject of several
experimental studies reported in the literature. Nienow et al. (1987), Hoffmann et al.
(1993) and Wu and Baeyens (1998) studied systems consisting of particles of equal
density but different size. Nienow and Naimer (1980) performed experiments on systems
consisting of particles of equal size but different density. Rowe et al. (1972) presented
experimental results for both types of systems. In general the different density systems
tend to segregate more easily than the polydisperse systems.
Due to increasing computer power discrete particle models have become a very useful and
versatile research tool in order to study the hydrodynamics of gasfluidised beds. In these
models the Newtonian equations of motion for each individual particle are solved.
Particleparticle and particlewall interactions are taken into account directly which is a
clear advantage over two fluid models that require closure relations for the solidsphase
Chapter 6
______________________________________________________________________________________
160
stress tensor (Kuipers et al. (1992) and Gidaspow (1994) among many others). When
simulating gas fluidised beds with particles of different size and/or particles of different
density multi fluid models can be used (Gidaspow et al., 1990) but several difficulties are
encountered due to the fact that large sets of continuum equations have to be solved. In
addition, and more fundamentally, significant problems arise when closure laws for the
mutual interaction of particles belonging to different classes have to be formulated.
Although the Kinetic Theory of Granular Flow (KTGF) offers a theoretical framework to
overcome this problem this approach is not very flexible and the computational cost
increases dramatically with each additional solid phase. Kumaran and Koch (1993) and
Manger (1996) applied this theory to binary mixtures and recently Mathiesen et al.
(1998) extended the work of Manger (1996) in order to handle three solid phases. A
discrete particle approach offers a more natural way to overcome these problems since
each individual particle in the simulation is tracked. Hence discrete particle models are
very useful in order to study segregation phenomena in gas fluidised beds consisting of
particles of different size and/or density in detail. Moreover they can be used to generate
data which can subsequently be used to develop closure models for continuum models.
However the number of particles that can be taken into account in a simulation is limited
(typically < 10
6
) which implies that currently the method can only be applied to rather
small systems consisting rather coarse particles.
The discrete particle approach for gasfluidised beds was pioneered by Tsuji et al. (1993)
who developed a softsphere discrete particle model based on the work of Cundall and
Strack (1979). In this approach the particles are allowed to overlap slightly and this
overlap is subsequently used to calculate the contact forces. Recently Kawaguchi et al.
(1998) presented results for a threedimensional version of this model. Schwarzer (1995)
used a model similar to that of Tsuji et al. (1993) to simulate liquid fluidised beds in
which lubrication forces were also taken into account. Hoomans et al. (1996) were the
first to present a hardsphere granular dynamics model of a gas fluidised bed. In this
discrete particle model the particles interact through binary, instantaneous, inelastic
collisions with friction. Xu and Yu (1997) presented a hybrid technique where they used a
contact force model in order to determine the interparticle forces and a collision detection
Segregation Phenomena
______________________________________________________________________________________
161
algorithm in order to determine the precise instant at which the particles first come into
contact. Mikami et al. (1998) recently extended the model of Tsuji et al. (1993) in order to
include cohesive forces between the particles. Seibert and Burns (1998) were able to
predict segregation phenomena in liquid fluidised beds using a Monte Carlo simulation
technique. However, the Monte Carlo technique is only capable of predicting a certain
steady state and is not suitable to simulate the dynamics of segregation.
In this chapter twodimensional granular dynamics simulations will be used to study
segregation phenomena in bubbling gasfluidised beds. The majority of the simulations is
concerned with systems consisting of particles of equal density but different size but also
a system consisting of particles of uniform size but different density is included.
2. Models
Since a detailed description of the models is presented in Chapter 2 and Chapter 3, the
key features will be summarised only briefly here.
2.1 Hardsphere granular dynamics
The collision model as originally developed by Wang and Mason (1992) is used to
describe a binary, instantaneous, inelastic collision with friction. The key parameters of
the model are the coefficient of restitution (0 ≤ e ≤ 1) and the coefficient of friction (µ ≥
0). Foerster et al. (1994) have shown that also the coefficient of tangential restitution (0 ≤
β
0
≤ 1) should be used in order to describe the collision dynamics more accurately. These
three collision parameters are all included in the model.
In the hardsphere approach a sequence of binary collisions is processed. This implies that
a collision list is compiled in which for each particle a collision partner and a
corresponding collision time is stored. A constant time step is used to take the external
Chapter 6
______________________________________________________________________________________
162
forces into account and within this time step the prevailing collisions are processed
sequentially. In order to reduce the required CPU time neighbour lists are used. For each
particle a list of neighbouring particles is stored and only for the particles in this list a
check for possible collisions is performed. When simulating a binary system of particles
of different size two cut off distances are used as schematically shown in Figure 6.1.
Figure 6.1. The neighbour list principle using two cut off distances. The shaded
particles are stored in the neighbour list of the black one.
The neighbour list consists of all the small particles which are found within the small
square (coloured grey) and the big particles which are found within the big square
(coloured grey). By using this approach the number of particles in the neighbour list will
never become too high which significantly reduces both CPU time and memory space
requirements.
2.2 Softsphere granular dynamics
Although the majority of the simulations presented in this chapter were performed with
the hardsphere code also a simulation with the softsphere code was performed and
therefore that model is described briefly here as well.
In softsphere models the following equations of motion are used:
D
nblist,small
D
nblist ,big
Segregation Phenomena
______________________________________________________________________________________
163
external contact
dt
d
m F F
r
+ ·
2
2
, (6.1)
T ·
ω
dt
d
I , (6.2)
where r is the position vector of the centre of the particle, m is the mass of the particle,
F
contact
is the contact force acting on the particle, F
external
is the net external force acting
on the particle, ω is the rotation velocity, T is the torque acting on the particle and I is the
moment of inertia of the particle. In this paragraph the focus will be on the contact forces
between the particles, the external forces will be discussed in the following paragraph.
The contact forces are given by:
ab n n ab n n n,ab
k
,
v n F η ξ − − · (6.3)
t,ab t t t t,ab
k v ? F η − − · (6.4)
If however the following relation is satisfied:
ab n t,ab ,
F F µ > , (6.5)
then sliding occurs and the tangential force is given by:
ab n,ab t,ab
t F F µ − · . (6.6)
For contacts between particles and walls, the walls were assumed to be non moving and
of infinite mass just like in the hardsphere model.
The resulting force and torque for particle a are now simply obtained by adding the pair
wise contributions of all the particles that are in contact with a:
Chapter 6
______________________________________________________________________________________
164
( )
∑
+ ·
b
t,ab n,ab a contact
F F F
,
, (6.7)
( )
∑
× ·
b
t,ab ab a a
R F n T . (6.8)
Since the contact forces are in general at least an order of magnitude larger than the
external forces a separation of time scales was introduced. By default at each time step
DT the external forces were taken into account while at 0.1DT the equations of motion
were solved by taking only the contact forces into account.
2.3 External forces
The incorporation of external forces differs somewhat from the approach followed by
Hoomans et al. (1996). In this work the external forces are used in accordance with those
implemented in the two fluid model presented by Kuipers et al. (1992) where, of course,
the forces now act on a single particle:
( )
( ) m
d
dt
m
V
V p
p
p
p
p
p p
v
g u v · +
−
− − ∇
β
ε 1
, (6.9)
where m
p
represents the mass of a particle, v
p
its velocity, u the local gas velocity and V
p
the volume of a particle. In equation (6.9) the first term is due to gravity and the third term
is the force due to the pressure gradient. The second term is due to the drag force where β
represents an interphase momentum exchange coefficient as it usually appears in two
fluid models. For low void fractions (ε < 0.80) β is obtained from the wellknown Ergun
equation:
( )
( )
p
p
g
p
g
D D
v u − − +
−
·
ρ
ε
µ
ε
ε
β 1 75 . 1
1
150
2
2
, (6.10)
Segregation Phenomena
______________________________________________________________________________________
165
where D
p
represents the particle diameter, µ
g
the viscosity of the gas and ρ
g
the density of
the gas. For high void fractions (ε ≥ 0.80) the following expression for the interphase
momentum transfer coefficient has been used which is basically the correlation presented
by Wen and Yu (1966) who extended the work of Richardson and Zaki (1954):
( )
65 . 2
1
4
3
−
−
−
· ε ρ
ε ε
β
p g
p
d
D
C v u . (6.11)
The drag coefficient C
d
is a function of the particle Reynolds number and given by:
( )
¹
¹
¹
'
¹
≥
<
·
1000 Re 44 . 0
1000 Re Re 0.15 + 1
Re
24
0.687
p
p p
p
d
C
,
(6.12)
where the particle Reynolds number (Re
p
) in this case is defined as follows:
g
p p g
p
D
µ
ερ v u −
· Re . (6.13)
For the integration of equation (6.9) a simple explicit first order scheme was used to
update the velocities and positions of the particles.
2.4 Gas phase hydrodynamics
The calculation of the gas phase hydrodynamics mainly follows the lines presented by
Kuipers et al. (1992). It is based on the numerical solution of the following set of partial
differential equations that can be seen as a generalised form of the NavierStokes
equations for a gas interacting with a solid phase as originally derived by Anderson and
Jackson (1967).
Chapter 6
______________________________________________________________________________________
166
Continuity equation gas phase:
( )
( )
∂ ερ
∂
ερ
g
g
t
+ ∇⋅ · u 0. (6.14)
Momentum equation gas phase:
( )
( ) ( )
∂ ερ
∂
ερ ε ε ερ
g
g p g g
t
p
u
uu S g + ∇ ⋅ · − ∇ − − ∇ ⋅ + τ . (6.15)
In this work transient, twodimensional, isothermal (T = 293 K) flow of air at atmospheric
conditions is considered. The constitutive equations can be found in Chapter 3. There is
one important modification with respect to the model presented by Hoomans et al. (1996)
and that deals with the way in which the twoway coupling between the gas phase and the
dispersed particles is established. In the present model the reaction force to the drag force
exerted on a particle per unit of volume is fed back to the gas phase through the source
term S
p
which has the dimension of force per unit of volume N/m
3
. For the calculation of
the void fraction (ε) the technique described in Chapter 3 was used. No special
modifications were made to take the effect of the particle size distribution into account.
3. Ternary density distribution
A simulation was performed of a system that consisted of three types of particles of
uniform size but of different density using the hardsphere code. The system was filled
with 800 particles with a density of 2700 kg/m
3
(u
mf
= 1.79 m/s), 800 particles with a
density of 1800 kg/m
3
(u
mf
= 1.42 m/s) and 800 particles with a density of 900 kg/m
3
(u
mf
= 0.98 m/s). The main parameter settings for this simulation can be found in table 6.1. The
Segregation Phenomena
______________________________________________________________________________________
167
gas inflow velocity was set equal to the minimum fluidisation velocity of the densest (ρ
p
=
2700 kg/m
3
) particles (u
mf
= 1.79 m/s).
Table 6.1. Parameter settings for the simulation with a different density mixture
Particles: Bed:
shape spherical width 0.15 m.
diameter 4.0 mm height 0.50 m
ρ
p
= 2700 kg/m
3
N
p
= 800 number xcells, NX 15
ρ
p
= 1800 kg/m
3
N
p
= 800 number ycells, NY 25
ρ
p
= 900 kg/m
3
N
p
= 800 cell width, DX 10 mm
e = e
w
0.9 cell height, DY 10 mm
µ = µ
w
0.3
N
p,total
2400 time step, DT 10
4
s
The coefficients of tangential restitution were assumed to be equal to zero. The bed was
fluidised at the minimum fluidisation velocity of the densest particles (u
g
= 1.79 m/s) for
15 s starting from perfectly mixed initial conditions. In Figure 6.2 snapshots of the
particle configurations are presented.
Figure 6.2. Snapshots of the particle configurations in a system with a ternary density
distribution.
Chapter 6
______________________________________________________________________________________
168
In Figure 6.2 the segregation effect can clearly be observed. After only 15 seconds three
distinct layers can be identified. In Figure 6.3 the mass fractions are plotted as a function
of the bed height at t = 15 s. The mass fractions are obtained by averaging over the full
width of a bed section of 0.05 m height. In this figure it can be observed that the mass
fraction of the jetsam particles (ρ
p
= 2700 kg/m
3
) is nearly one in the bottom region where
the mass fraction of the flotsam particles (ρ
p
= 900 kg/m
3
) is zero. In the upper regions of
the bed the mass fraction jetsam is zero whereas the mass fraction flotsam approaches
one. The middle section of the bed clearly shows a peak in the mass fraction of the
particles with ρ
p
= 1800 kg/m
3
.
Figure 6.3. Mass fractions as a function of the bed height at t = 15 s for the simulation
with a ternary density distribution.
It is remarkable that even though the two less dense types of particles are both fluidised
well above their u
mf
there is a clear segregation effect between these two layers as well.
Even though the fluidisation velocity is 1.3u
mf
of the particles with ρ
p
= 1800 kg/m
3
, the
segregation between the upper two layers is rather clear. However, this is in agreement
with the experimental observations by Rowe et al. (1972) who reported that different
Segregation Phenomena
______________________________________________________________________________________
169
density systems still show segregation above the u
mf
of the jetsam particles. Since the
particle size is rather large in this relatively small system wake effects due to bubbles are
less pronounced. A dispersive mechanism was therefore not dominantly present and this
may have enhanced the segregation.
4. Binary size distribution
4.1 Base case
A simulation was performed of a system that consisted of two types of particles of
different size but equal density using the hardsphere code. A homogeneous mixture of
250 particles of 4.0 mm diameter and 4750 particles of 1.5 mm diameter (50/50 wt.) was
fluidised at the minimum fluidisation velocity of the bigger particles (u
mf,big
= 1.7 m/s).
The simulation was run for 50 seconds using the parameter settings presented in table 6.2.
The collision parameters for the glass particles were obtained from Hoomans et al. (1996).
Table 6.2. Parameter settings for the simulations with a binary size distribution
Particles: Bed:
shape spherical width 0.15 m.
density, ρ
p
2480 kg/m
3
height 0.25 m
D
p
= 1.5 mm N
p
= 4750 number xcells, NX 15
D
p
= 4.0 mm N
p
= 250 number ycells, NY 25
e 0.96 cell width, DX 10 mm
e
w
0.86 cell height, DY 10 mm
µ = µ
w
0.15
N
p,total
5000 time step, DT 10
–4
s
The coefficients of tangential restitution were assumed to be equal to zero. Snapshots of
particle configurations are shown in Figure 6.4. In this figure it can be observed that
segregation does occur with the bigger particles accumulating at the bottom of the bed.
Chapter 6
______________________________________________________________________________________
170
However, this is a very dynamic situation as could be observed in animations where it
became clear that the bigger particles are continuously transported to the upper regions of
the bed in the wake of a bubble and descend again in the denser regions. This process
continued throughout the simulation.
Figure 6.4. Snapshots of the particle configurations for a binary system of particles of
different size.
Segregation Phenomena
______________________________________________________________________________________
171
A typical simulation took about 25 minutes of CPU time per simulated second on a
Silicon Graphics Power Challenge with a 196 MHz R10000 processor. In Figure 6.5 the
mass fraction jetsam at t = 20 s is plotted as a function of the height in the bed. Since all
particle positions are known at each instant during a simulation such a plot can be
calculated instantaneously. In experiments normally the gas supply is first shut off after
which the bed is sectioned and each section is analysed separately. In the figure an
additional line at the average mass fraction of 0.5 is included to guide the eye.
Figure 6.5. Mass fraction jetsam as a function of the bed height at t = 20 s for the
simulations with a binary size distribution at three different gas velocities
(u_mf refers to the u
mf
of the big particles)
It can be observed that the mass fraction jetsam is indeed higher in the lower part of the
bed although the fraction does not become one. This is not surprising since it could
already be observed from the snapshots in Figure 6.4 that there is no pure layer of jetsam
particles being formed on the bottom of the bed. The common knowledge that different
density systems tend to segregate more easily than different size systems (Rowe et al.,
1972) is confirmed here when Figure 6.5 is compared with Figure 6.3. In Figure 6.5 also
results for higher gas velocities are included from which it can be observed that the solids
mixing is better at increased gas velocities.
Chapter 6
______________________________________________________________________________________
172
4.2 Statistical analysis of segregation
Although segregation could be observed in the base case simulation presented in the
previous paragraph the system remained very dynamic. Therefore an analysis based on a
single frame, as was presented for the base case simulation, does not provide a complete
understanding of the segregation phenomenon. A similar problem is encountered in
segregation experiments where segregation profiles are determined after the gas supply is
abruptly shut off and the bed is subsequently divided in sections that are separately
analysed (Hoffmann et al. (1993), Wu and Baeyens (1998)). By repeating the experiments
the reliability of the results can be improved which is however rather time consuming. In
granular dynamics simulations the positions of all the particles in the system are known at
each instant. This enables the determination of segregation profiles by averaging over a
large number of particle configurations, also referred to as frames. For the results
presented in Figure 6.6 an average segregation profile was obtained using the particle
configurations at each 0.01 seconds for a total duration of 50 seconds. Hence a reliable
average and a corresponding variance could be obtained from 5000 frames. A line
indicating the average mass fraction jetsam is included to guide the eye.
Figure 6.6. Segregation profiles obtained from both single and multiple frame
analysis
Segregation Phenomena
______________________________________________________________________________________
173
In this figure a segregation profile obtained from a single frame analysis (t = 20.0 s) is
included as well. Since this profile differs significantly from the average profile it clearly
shows that one has to be careful with the application of single frame analysis. It only gives
an impression of an instantaneous situation and therefore does not provide a complete
understanding of the segregated state of the system.
In Figure 6.7 the probability distribution of the mass fraction jetsam at a bed height of
0.075 m obtained over the full duration of the simulation is presented.
Figure 6.7. Probability distribution of the jetsam mass fraction at 0.075 m above the
distributor plate obtained over 50 seconds
The wide spread in this distribution again emphasises that a single frame analysis may
lead to an incomplete understanding of the segregated state of the system. It is important
to note that the segregation profiles obtained from the simulation are based on the actual
positions of the particles during fluidisation while in experiments the gas supply is always
shut off before the collapsed bed is sectioned and analysed. The segregation profiles
obtained in such experiments are therefore based on the particle positions in a collapsed
bed rather than a bed in a fluidised state.
Chapter 6
______________________________________________________________________________________
174
4.3 Effects of collision parameters
Since the collision parameters turned out to have a significant influence on the bed
hydrodynamics in all the simulations presented in the previous chapters an additional
simulation was performed assuming fully elastic (e = 1) and perfectly smooth (µ = 0)
collisions. All other parameters were set to the same values as presented in table 6.2.
Snapshots of this simulation are presented in Figure 6.8.
Figure 6.8. Snapshots of particle configurations for the simulation with ideal collision
parameters (e = 1.0, µ = 0.0)
The result is striking, a segregated state is reached after a few seconds and this situation
does not change significantly anymore. Due to the absence of bubbles there is no
mechanism present to transport the bigger particles to the upper regions of the bed, which
stresses the crucial role of bubbles for solids mixing in fluidised beds. The simulation was
Segregation Phenomena
______________________________________________________________________________________
175
continued for 20 seconds and an average segregation profile over 2000 frames was
obtained. This average segregation profile is presented in Figure 6.9.
Figure 6.9. Segregation profiles obtained from both single and multiple frame
analysis for the simulation with ideal collision parameters (e = 1.0, µ =
0.0)
In this figure it can be observed that the segregation is far more pronounced than in the
base case simulation (Figure 6.6) with non ideal collision parameters (e = 0.96, µ = 0.15)
and that the variance is a lot less as well.
5. Experimental validation
5.1 Experimental
For the purpose of experimental validation of the segregation simulations a setup was
designed and constructed that is schematically represented in Figure 6.10.
Chapter 6
______________________________________________________________________________________
176
width 150 mm
height 400 mm depth 15 mm
porous plate
pore size ~10 micron
3mm glass beads
gas
gas (jet)
height 100 mm
PI
PI
PI
PI
PI
PI
PI
PI
PI
PI
steam
steam
Figure 6.10. Schematic representation of the experimental setup
In this quasi twodimensional set up the particles can be observed with a video camera
during the fluidisation process since glass was used as the material for the construction of
the bed. Coloured glass particles obtained from Worf Glaskugeln GmbH in Mainz,
Germany were used that can easily be distinguished by visual observation. The biggest
particles were coloured red and had a diameter of 2.5 mm, the smallest were coloured
yellow and had a diameter of 1.5 mm. Since homogeneous gas inflow conditions were
used during the segregation experiments the option to invoke a jet was not used. Instead,
the gas velocity in the jet was set equal to the uniform inflow velocity. A 50/50 wt.
particle mixture was used for the validation with an initial bed height of 0.15 m. The
fluidising gas (air at ambient conditions) was humidified in order to minimise the
influence of electrostatic forces. In preliminary measurements the minimum fluidisation
velocities were determined separately for both types of particles. For the 1.5 mm diameter
particles a minimum fluidisation velocity (u
mf
) of 0.92 m/s was found whereas for the 2.5
mm diameter particles a u
mf
of 1.28 m/s was found. For a 50/50 wt.% mixture the
Segregation Phenomena
______________________________________________________________________________________
177
segregation turned out to be most pronounced when the gas inflow velocity was chosen
equal to 1.3 m/s.
5.2 Results
The parameter settings used for the simulations that were to be compared with the
experiment are presented in table 6.3.
Table 6.3. Parameter settings for the simulation duplicating the experiment
Particles: Bed:
Shape Spherical Width 0.15 m.
density, ρ
p
2525 kg/m
3
Height 0.40 m
D
p,small
= 1.5 mm N
p
= 6481 number xcells, NX 15
D
p,big
= 2.5 mm N
p
= 1400 number ycells, NY 40
cell width, DX 10 mm
k
n
1000 N/m cell height, DY 10 mm
N
p,total
7881 time step, DT 10
–4
s
The collision parameters for the particles used in the experiment were measured at the
Open University at Milton Keynes using the facility of the Impact Research group that is
described in Chapter 2. In the simulations presented in the previous paragraphs the
collision parameters for particleparticle collisions were identical: no distinction was
made between collisions of particles of different types. For the simulation presented here
the code was extended to take the various collision parameters into account. The values of
the collision parameters that were obtained from the measurements at the Open University
are presented in table 6.4 (Gorham and Kharaz, 1999).
Chapter 6
______________________________________________________________________________________
178
Table 6.4. Collision parameters obtained by Gorham and Kharaz (1999)
bigbig bigsmall smallsmall bigwall smallwall
e 0.97 0.97 0.97 0.97 0.97
µ 0.11 0.10 0.15 0.09 0.10
β
0
0.33 0.33 0.33 0.33 0.33
At first a simulation was run with a homogeneous gas inflow velocity of 1.3 m/s using the
hardsphere code. However, after 60 seconds still no segregation could be observed which
is not in agreement with the results of the experiment. At lower gas velocities segregation
was observed but for these simulations the hardsphere code could not be used due to
stability problems. The hardsphere approach is not able to handle static zones in the bed
and at gas velocities below u
mf
such static zones cannot be avoided. Therefore the soft
sphere code was used to perform a simulation with a homogeneous gas inflow velocity of
0.9 m/s. For this simulation the softsphere code was extended to take the various
collision parameters into account. The normal spring stiffness was set equal to 1000 N/m
for all sorts of contacts and a time step of 10
5
s was used for the integration of the contact
forces.
Video images of the experiment and snapshots of the simulation are presented together in
Figure 6.11. It can clearly be observed that the segregation is more pronounced and faster
in the experiment than in the simulation. In the video images of the experiment bubbling
appears to be far more pronounced than in the simulation. In the simulation with a gas
inflow velocity of 1.3 m/s the bubbling was also much more pronounced but in that
particular simulation segregation did not occur.
Segregation Phenomena
______________________________________________________________________________________
179
t = 0 s t = 20 st = 40 s
Figure 6.11. Snapshots of the simulation compared with video images of the experiment
Chapter 6
______________________________________________________________________________________
180
t = 60 st = 80 st = 100 s
Figure 6.11. (continued) Snapshots of the simulation compared with video images of the
experiment
Segregation Phenomena
______________________________________________________________________________________
181
At t = 100 s the segregation was almost complete in the experiment whereas this was by
far not the case in the simulation. Therefore the simulation was continued in order to
check whether the segregation would become more complete. In Figure 6.12 snapshots of
the simulation at t = 150 s and t = 200 s are presented.
Figure 6.12. Snapshots of the simulation at t = 150 s and t = 200 s
Indeed the segregation is more pronounced than it was at t = 100 s but still not as
complete as it was in the experiment. Also the bigger particles tend to accumulate in the
centre of the bed whereas in the experiment there is a sharper horizontal division between
the two particle layers. Since the simulation was very time consuming (280 minutes of
CPU time per simulated second on a Silicon Graphics Origin200 server with a R10000
processor, 225 MHz, 2 Mb cache) the simulation was not repeated at different gas
velocities.
Chapter 6
______________________________________________________________________________________
182
Although the softsphere model is capable of predicting segregation for this system it is
clear that the agreement with the experiment is rather poor. This is most likely due to the
fact that the motion of the particles is restricted to two dimensions in the simulation. Also
the particle sizes used in the experiment do not differ so much (1.5 vs. 2.5 mm).
Segregation is expected to be more pronounced in the case of a larger size difference as
was used in the simulation in the previous paragraph (1.5 vs. 4.0 mm). However the use of
4.0 mm diameter particles in a bed of 15 mm depth is questionable. On the other hand
particles of 1.0 mm diameter could be used instead of 1.5 mm diameter but this would
increase the number of particles in the simulation and hence the computational cost.
It is important to realise that in the present code the hydrodynamics of the gas phase is
still resolved on a length scale that is larger than the particle size. This affects the twoway
coupling between the particle dynamics and the gas phase hydrodynamics. In Figure 6.13
two computational cells are schematically represented with each a different particle
configuration but identical void fraction.
Figure 6.13. Schematic representation of two computational cells with identical void
fraction but different solids distribution
In the present code these two configurations are treated in exactly the same manner as far
as the twoway coupling is concerned, assuming that the particle velocities are identical in
both cases. It is however to be expected that when the gas flow is solved on a length scale
Segregation Phenomena
______________________________________________________________________________________
183
smaller than the particle size these two configurations will show different results in terms
of twoway coupling. In Chapter 3 some techniques are discussed that are capable of
solving the gas phase hydrodynamics on this subparticle scale. Using for instance the
Lattice Boltzmann technique the effects of the subgrid particle configuration on the flow
behaviour can be studied on a length scale smaller than the particle size. Such microscopic
simulations can subsequently be used to improve the twoway coupling in the Eulerian
Lagrangian simulations by for instance an improved correlation for the drag force acting
on a particle. This can eventually lead to improved agreement between simulation and
experiment for the case presented in this section.
6. Conclusions
Gas fluidised beds consisting of particles of different size and density were simulated
using a hardsphere discrete particle model. Segregation was observed in the case of a
binary system of particles of different size as well as in the case of a ternary system
consisting of particles of different density. In the latter case the segregation effect was
more pronounced which is in agreement with experimental findings reported in the
literature. Using realistic values for the collision parameters (e = 0.96, µ = 0.15)
segregation was observed on a time scale of several seconds although the system never
reached a clear steady state. The bigger (jetsam) particles were continuously transported
to the upper regions of the bed by the bubbles and then moved down again in the denser
regions. An average segregation profile was determined using 5000 frames (every 0.01
second for a total duration of 50 seconds). The results showed a rather large variance. An
analysis based on a single frame was found to lead to an incomplete understanding of the
segregated state. When the collisions were assumed to be perfectly elastic and perfectly
smooth (e = 1.0, µ = 0.0) a completely different behaviour was observed. Due to the
absence of bubbles segregation became almost complete after only a few seconds and this
situation did not change significantly anymore during the remainder of the simulation.
This stresses the important role of bubbles in the mixing of solids in fluidised beds.
Chapter 6
______________________________________________________________________________________
184
Finally an experimental setup was designed and constructed for the purpose of
experimental validation. At a gas inflow velocity of 1.3 m/s segregation was observed in
the experiments over a time scale of less than a minute but in the simulations no
segregation could be observed over the same time scale. At a lower gas inflow velocity
(0.9 m/s) segregation was observed in the simulations. For these simulations the soft
sphere code had to be used due to the presence of static zones. However the time scale
over which the segregation occurred and the final segregation pattern were in poor
agreement with the experimental results. This is most likely due to the fact that the motion
of the particles in the simulation is restricted to two dimensions in contrast with the
experiment. Furthermore the difference in the particle sizes used for the validation is
rather small (1.5 mm vs. 2.5 mm diameter) which renders a tough test for the model.
Notation
C
d
drag coefficient, []
D
nblist
width of neighbour list square, m
D
p
particle diameter, m
DT time step, s
DX horizontal computational cell dimension, m
DY vertical computational cell dimension, m
e coefficient of restitution, []
F force, N
g gravitational acceleration, m/s
2
I moment of inertia, kgm
2
k spring stiffness, N/m
m particle mass, kg
N
p
Number of particles, []
NX number of computational cells in xdirection, []
NY number of computational cells in ydirection, []
Segregation Phenomena
______________________________________________________________________________________
185
n normal unit vector, []
p pressure, Pa
R
p
particle radius, m
r position vector, m
S
p
momentum source term, N/m
3
T torque, Nm
T temperature, K
t tangential unit vector, []
t time, s
u gas velocity vector, m/s
v particle velocity vector, m/s
V
p
particle volume, m
3
Greek symbols
β defined in Eqs () and (), kg/m
3
s
β
0
coefficient of tangential restitution, []
ε void fraction, []
µ coefficient of friction, []
µ
g
gas viscosity, kg/ms
η damping coefficient, Ns/m
τ gasphase stress tensor, kg/ms
2
ρ density, kg/m
3
ω angular velocity, s
1
ξ displacement, m
Subscripts
g gas phase
mf minimum fluidisation
n normal
Chapter 6
______________________________________________________________________________________
186
nblist neighbour list
p particle
t tangential
w wall
References
Anderson, T.B. and Jackson, R., (1967). A fluid mechanical description of fluidized beds
(equations of motion), Ind. Eng. Chem., Fundam., 6, 527.
Cundall, P.A. and Strack, O.D.L., (1979). A discrete numerical model for granular
assemblies. Géotechnique, 29, 47.
Foerster S.F., Louge, M.Y., Chang, H. and Allia, K. (1994). Measurements of the
collision properties of small spheres. Phys. Fluids,, 6, 1108.
Gidaspow, D., Ding, J. and Jayaswal, U.K. (1990). Multiphase NavierStokes equation
solver. Numerical Methods for Multiphase Flows, FED 91, 47.
Gidaspow, D. (1994). Multiphase Flow and Fluidization, Academic Press, Boston, USA.
Gorham, D.A. and Kharaz, A.H., (1999). Results of particle impact tests, Impact Research
Group Report IRG 13, The Open University, Milton Keynes, UK.
Hoffmann, A.C., Janssen, L.P.B.M. and Prins, J. (1993). Particle segregation in fluidised
binary mixtures, Chem. Engng Sci., 48, 1583.
Hoomans, B.P.B., Kuipers, J.A.M., Briels, W.J. and van Swaaij, W.P.M., (1996). Discrete
particle simulation of bubble and slug formation in a twodimensional gas fluidised bed: a
hardsphere approach, Chem. Engng Sci., 51, 99.
Segregation Phenomena
______________________________________________________________________________________
187
Kawaguchi, T., Tanaka, T. and Tsuji, Y., (1998). Numerical simulation of two
dimensional fluidized beds using the discrete element method (comparison between the
two and threedimensional models), Powder Technol., 96, 129.
Kuipers, J.A.M., van Duin, K.J. van Beckum F.P.H. and van Swaaij, W.P.M. (1992). A
numerical model of gasfluidized beds, Chem. Engng Sci., 47, 1913.
Kumaran, V. and Koch, D.L. (1993). Properties of a bidisperse particle gas suspension
Part 1. Collision time small compared with viscous relaxation time, J. Fluid Mech., 247,
623.
Manger, E. (1996). Modelling and simulation of gas/solids flow in curvilinear
coordinates, Ph.D. Thesis, Telemark Institute of Technology, Porsgrunn, Norway.
Mathiesen, V., Solberg T. and Hjertager, B.H., (1998). A computational study of
multiphase flow behavior in a circulating fluidized bed with realistic particle size
distribution, Proc. 3
rd
Int. Conf. On Multiphase Flow, 812 June 1998, Lyon, France
Mikami, T., Kamiya, H. and Horio, M., (1998). Numerical simulation of cohesive powder
behavior in a fluidized bed, Chem. Engng Sci., 53, 1927.
Nienow, A.W. and Naimer, N.S. (1980). Continuous mixing of two particulate species of
different density in a gas fluidised bed, Trans IChemE., 58, 181.
Nienow, A.W., Naimer, N.S. and Chiba, T. (1987). Studies of segregation/mixing in
fluidised beds of different size particles, Chem. Eng. Comm., 62, 52.
Richardson, J.F. and Zaki, W.N., (1954). Sedimentation and fluidization: part I. Trans.
Inst. Chem. Eng., 32, 35.
Chapter 6
______________________________________________________________________________________
188
Rowe, P.N., Nienow, A.W. and Agbim, A.J. (1972). The mechanism by which particles
segregate in gas fluidised beds: binary systems of nearspherical particles, Trans. Instn
Chem. Engrs, 50, 310.
Schwarzer, S., (1995). Sedimentation and flow through porous media: Simulating
dynamically coupled discrete and continuum phase, Phys. Rev. E. 52, 6461.
Seibert, K.D. and Burns, M.A. (1998). Simulation of structural phenomena in mixed
particle fluidized beds, AIChE J. 44, 528.
Tsuji, Y., Kawaguchi, T. and Tanaka, T., (1993). Discrete particle simulation of two
dimensional fluidized bed, Powder Technol. 77, 79.
Wang, Y. and Mason, M.T., (1992). Twodimensional rigidbody collisions with friction.
J. Appl. Mech., 59, 635.
Wen, C.Y. and Yu, Y.H., (1966). Mechanics of fluidization. Chem. Eng. Prog. Symp.
Ser., 62 (62), 100.
Wu, S.Y. and Baeyens, J. (1998). Segregation by size difference in gas fluidized beds,
Powder Technol. 98, 139.
Xu, B.H. and Yu, A.B., (1997). Numerical simulation of the gassolid flow in a fluidized
bed by combining discrete particle method with computational fluid dynamics. Chem.
Engng Sci. 52, 2785.
Riser Flow
______________________________________________________________________________________
189
Chapter 7.
THE INFLUENCE OF COLLISION PROPERTIES ON THE
FLOW STRUCTURE IN A RISER
Abstract:
A twodimensional hardsphere discrete particle model was used in order to simulate the
gassolid flow in the riser section of a circulating fluidised bed (CFB). The Newtonian
equations of motion are solved for each individual particle in the system while taking into
account the particleparticle and particlewall collisions. The gas phase hydrodynamics
is described by the spatially averaged NavierStokes equations for twophase flow. The
results proved to be very sensitive with respect to the collision parameters (i.e. coefficient
of restitution (e) and coefficient of friction (µ)). In the case of fully elastic and perfectly
smooth collisions (e = 1.0, µ = 0.0) hardly any clustering of particles could be observed
as opposed to the case where these collision parameters were assigned realistic values (e
= 0.94, µ = 0.28). Particlewall collisions turned out to have very little influence on the
flow structure. A strong effect of the collision properties on the axial solids profile was
found where a pronounced buildup of solids was observed in the simulation with
realistic values for the collision parameters. In the simulation assuming fully elastic and
perfectly smooth collisions no buildup of solids was observed. This result is supported by
the experimental findings of Chang and Louge (1992) where a smaller pressure drop
over the riser was found when glass particles were used to which a coating was applied
in order to decrease the coefficient of friction. Lift forces acting on the suspended
particles turned out to have a slightly redispersive effect on the flow structure which
made the radial segregation of the solids a little less pronounced.
Chapter 7
______________________________________________________________________________________
190
Parts of this chapter are based on the papers:
B.P.B. Hoomans, J.A.M. Kuipers and W.P.M. van Swaaij, (1998). Granular dynamics simulation of cluster
formation in dense riser flow, Third International Conference on Multiphase Flows 1998, June 812, Lyon,
France.
B.P.B. Hoomans, J.A.M. Kuipers and W.P.M. van Swaaij, (1999). Discrete particle simulation of cluster
formation in dense riser flow, Circulating Fluidized Bed Technology VI, J. Werther (Ed.), p. 266.
Riser Flow
______________________________________________________________________________________
191
1. Introduction
Cluster formation is a characteristic feature of dense gas solid flow in the riser section of
a Circulating Fluidized Bed (CFB). Although a clear definition is lacking (Chen, 1995)
clusters can be typified as regions of locally higher solid fraction. Since the existence of
such clusters has a profound influence on the performance of a CFB unit as a chemical
reactor they have been the subject of a number of studies. Kuipers et al. (1998)
demonstrated that the radial solids segregation in a riser (i.e. coreannulus flow) has a
negative effect on the performance of the riser as a chemical reactor. Radial solids
segregation in a riser is largely due to the existence of clusters that tend to accumulate
near the wall. Horio and Kuroki (1994) experimentally observed clusters in dilute riser
flow using a laser sheet technique. They found that the clusters had a characteristic
parabolic shape.
Tsuo and Gidaspow (1990) used a TwoFluid approach with constant solids viscosity in
order to simulate the riser section of a CFB. They reported the formation of cluster like
structures even though the spatial resolution in their simulation was rather low
(computational cells of 7.62 mm width and 76.2 mm height) compared to typical cluster
sizes. After time averaging the typical coreannulus flow structure where the average
solids concentration is considerably higher near the wall was obtained. Other TwoFluid
approaches incorporating the kinetic theory of granular flow (Sinclair and Jackson
(1989), Nieuwland et al. (1996)) did not focus particularly on cluster formation but more
on the radial segregation of solids. This type of model possesses a peculiar dependency
on the magnitude of the coefficient of restitution where a small deviation from unity
causes the flow structure to change completely while the agreement with experimental
findings deteriorates. This was (among others) pointed out by Hrenya and Sinclair (1997)
who also reported that when particle phase turbulence was included in the model the
aforementioned dependency became far less pronounced.
Chapter 7
______________________________________________________________________________________
192
Tanaka et al. (1996) performed simulations of gassolid flow in a vertical duct using the
Lagrangian approach for the solid particles. They employed the Direct Simulation Monte
Carlo (DSMC) method to describe the particle dynamics. In a later paper Tsuji et al.
(1998) made a comparison between the work of Tsuo and Gidaspow (1990) and Tanaka
et al. (1996). In the DSMC method the simulated particles actually represent a certain
number of ‘real‘ particles. A random number generator is invoked to determine the
collision partners and the geometry of the collision. Hence this method does not account
for actual particleparticle and particlewall interaction in a direct way. Moreover the
modified Nanbu method used in their work does not guarantee exact conservation of
energy (Frezzotti, 1997). This can be an important drawback especially since the collision
parameters, that determine how much energy is dissipated in collisions, turned out to be
of key importance in their simulations. Recently Helland et al. (1999) and Ouyang and Li
(1999) presented discrete particle approaches similar to the one used in this work but in
their simulations collisions are detected at constant time intervals allowing the particles to
slightly overlap. In this chapter a EulerianLagrangian simulation technique for dense
gassolid twophase flow in a riser is presented which features direct incorporation of the
particleparticle and particlewall interaction.
2. Model
Since a detailed description of the model is presented in Chapter 2 and Chapter 3, the key
features will be summarised briefly here.
2.1 Granular dynamics
The collision model as originally developed by Wang and Mason (1992) is used to
describe a binary, instantaneous, inelastic collision with friction. The key parameters of
the model are the coefficient of restitution (0 ≤ e ≤ 1) and the coefficient of friction (µ ≥
0). Foerster et al. (1994) have shown that also the coefficient of tangential restitution (0 ≤
β
0
≤ 1) should be used in order to describe the collision dynamics more accurately. These
Riser Flow
______________________________________________________________________________________
193
three collision parameters are all included in the model. In all simulations presented in this
chapter, the coefficient of tangential restitution is assumed to be equal to zero.
In the hardsphere approach a sequence of binary collisions is processed. This implies that
a collision list is compiled in which for each particle a collision partner and a
corresponding collision time is stored. A constant time step is used to take the external
forces into account and within this time step the prevailing collisions are processed
sequentially. In order to reduce the required CPU time neighbour lists are used. For each
particle a list of neighbouring particles is stored and only for the particles in this list a
check for possible collisions is performed. Since in riser flow the velocity differences
between the particles is larger than in bubbling fluidised beds the neighbour list is chosen
rather large (D
nblist
= 8 D
p
) to ensure that all collisions are properly detected.
2.2 External forces
The equation of motion used for the simulations of riser flow is:
LM LG drag p
p
p
m
dt
d
m F F F g
v
+ + + · (7.1)
where m
p
represents the mass of a particle and v
p
its velocity. In equation 7.1 the first
term on the right hand side is due to gravity and the second term is due to drag:
( )
7 . 2 2
8
1
−
− − · ε ρ π
p p g d p drag
C D v u v u F (7.2)
In equation 7.2 ρ
g
represents the density of the gas phase and u represents its velocity. It
should be noted that the drag correction due to the presence of other particles has been
accounted for via the wellknown Richardson and Zaki (1954) correction factor. The
drag coefficient for an isolated particle depends on the particle Reynolds number as
given by Rowe and Henwood (1961):
Chapter 7
______________________________________________________________________________________
194
( ) ( )
¹
¹
¹
'
¹
≥
< +
·
1000 Re 44 . 0
1000 Re Re 15 . 0 1
Re
24
p
p
687 . 0
p
p
d
C
(7.3)
where the particle Reynolds number is defined as:
g
p p g
p
D
µ
ερ v u−
· Re . (7.4)
Gravity and drag force are implemented by default in all simulations shown in this work.
The lift forces are not used by default but since an additional simulation was performed in
order to test their influence the lift forces will be discussed in more detail here.
The second term on the right hand side of equation 7.1 is the lift force due to a velocity
gradient. Following the lines of Tanaka et al. (1996) only the gradient in the xdirection is
considered:
( ) χ ρ π sign C D F
p g LG p x LG
2
2
,
8
1
v u − · (7.5)
where χ denotes the velocity gradient ∂(u
y
v
p,y
)/∂x.
The empirical relations used for the lift coefficient C
LG
are summarised in table 7.1.
The third term on the right hand side in equation 7.1 is the lift force due to particle
rotation, also known as the Magnus lift force:
( )
r
r p
p g LM p LM
C D
ω
ω
ρ π
× −
− ·
v u
v u F
2
8
1
. (7.6)
Riser Flow
______________________________________________________________________________________
195
where ω
r
is defined as follows:
u × ∇ − ·
2
1
ω ω
r
(7.7)
The empirical relations used for the lift coefficient C
LM
can be found in table 7.1.
Table 7.1. Empirical relations for the lift coefficients
( )
40 Re Re 0741 . 0
40 Re 3314 . 0
10
Re
exp 3314 . 0 1
2 / 1
,
2 / 1 2 / 1
,
> ·
≤ +
,
`
.

− − ·
p
Sa LG
LG
p
p
Sa LG
LG
C
C
C
C
χ
ξ ξ
(Mei, 1992)
with:
Re
Re
225 . 8
2 / 1
,
χ
·
Sa LG
C
(Saffman, 1965),
p
p
D
v u −
·
2
χ
ξ
and
g
p g
D
µ
χ ρ
χ
4
Re
2
·
γ 2 ·
LM
C Re
p
< 10 with:
p
r p
D
v u−
·
2
ω
γ
2 / 1
Re 8 . 3
−
·
p LM
C γ 10 < Re
p
< 500 (Oesterlé, 1994)
γ 40 . 0 ·
LM
C 500 < Re
p
(Tsuji et al., 1985)
( ) C
m
· +
− −
1 1
2
1
π
α β
ω ω
Re Re
with:
g
p r g
D
µ
ω ρ
ω
4
Re
2
·
α = 5.32; β = 37.2 10< Re
ω
≤ 20
α = 6.44; β = 32.2 20< Re
ω
≤ 50 (Dennis et al. 1980)
α = 6.45; β = 32.1 Re
ω
≥ 50
The rotational velocity of particles is not only affected by collisions with the system walls
or with other particles, but also by the torque exerted on them by the fluid:
Chapter 7
______________________________________________________________________________________
196
r r g m p
C D
dt
d
I ω ω ρ π
ω
5
64
1
− · . (7.8)
The empirical relations used for the rotational drag coefficient C
m
are summarised in
table 7.1.
For the integration of equation 7.1 as well as equation 7.8 a simple explicit first order
scheme was used to update the velocities and positions of the particles.
2.3 Inlet and outlet conditions
In the riser section of a CFB there is a continuous transport of particles throughout the
system that has to be taken into account in the simulation. Instead of simulating the entire
riser section, including inlet and exit configurations, only the middle section was
simulated. At the inlet of the simulated system particles were placed at random in the
bottom row of cells ensuring that there is no overlap with other particles or walls. An
initial axial velocity of 0.4 m/s was assigned to each particle fed to the system. At each
time step a specific number of particles was fed to the riser duct in accordance with a
specified solids mass flux (G
s
). The cross sectional area in this (twodimensional) case
was assumed to be equal to the width of the system times the particle diameter. At the
outlet particles which had crossed the upper boundary were simply removed from the
system at each time step. The particles were only allowed to leave the system at the upper
boundary.
2.4 Gas phase hydrodynamics
The calculation of the gas phase hydrodynamics mainly follows the lines presented by
Kuipers et al. (1992). It is based on the numerical solution of the following set of partial
differential equations that can be seen as a generalised form of the NavierStokes
equations for a gas interacting with a solid phase as originally derived by Anderson and
Jackson (1967).
Riser Flow
______________________________________________________________________________________
197
Continuity equation gas phase:
( )
( )
∂ ερ
∂
ερ
g
g
t
+ ∇⋅ · u 0 . (7.9)
Momentum equation gas phase:
( )
( ) ( ) g S uu
u
g g p g
g
p
t
ερ ε ε ερ
∂
ερ ∂
+ ⋅ ∇ − − ∇ − · ⋅ ∇ + τ . (7.10)
In this work transient, twodimensional, isothermal (T = 293 K) flow of air at slightly
elevated pressure (p = 1.2 Bar) is considered. The constitutive equations and the boundary
conditions used can be found in Chapter 3. Note that no turbulence modelling is included
in the present description. The reasons for this can be summarised as follows. Firstly the
main goal of this work is to study the influence of particle properties on the flow structure.
Secondly the particles used in the simulations in this chapter are rather coarse (500 µm)
and are therefore not much influenced by turbulent eddies as for instance is the case for
FCC particles (70 µm). And last but not least there is no generally accepted and
experimentally validated turbulence model for the type of twophase flow that is the
subject of study here. Helland et al. (1999) did include an additional term in the
momentum equation of the gas phase in order to account for sub grid scale effects of
turbulent eddies. However this term was taken directly from a singlephase flow model
without any modification for the presence of the particles. It cannot be denied that the
assumption of laminar flow is questionable for the type of flow encountered in the
simulations in this chapter but there is no doubt that the application of a singlephase flow
turbulence model in this situation is questionable as well. The void fraction (ε) is
calculated from the particle positions in the bed. There is one important modification with
respect to the model of Hoomans et al. (1996) and that deals with the way in which the
Chapter 7
______________________________________________________________________________________
198
twoway coupling between the gasphase and the particle motion is established. In the
present model the reaction forces to the forces exerted on a particle by the gasphase per
unit of volume is fed back through the source term S
p
which has the dimension Nm
3
.
3. Results
As a first test case for the model the same system was chosen as was used by Tsuo and
Gidaspow (1990) and Tanaka et al. (1996) in their simulations. The general parameter
settings are summarised in Table 7.2. In all simulations a time step of 10
4
s was used
together with a computational grid of 20 (NX) by 100 (NY) cells.
Table 7.2. General parameter settings
particles system gas
diameter, D
p
500 µm width 0.08 m temperature, T 293 K
density, ρ 2620 kg/m
3
height 2.00 m pressure, p 1.2 Bar
solids mass flux, G
s
25 kg/m
2
s
velocity, u
g
5.0 m/s
3.1 Effect of collision parameters
At first the influence of the collision parameters was examined. Tanaka et al. (1996)
reported a strong dependency of the flow structure with respect to the coefficients of
restitution (e) and friction (µ). In the work of Hoomans et al. (1996) on bubbling and
slugging fluidised beds these parameters also turned out to have a decisive influence on
the flow behaviour. Therefore three different simulations were performed where the
collision parameters were set according to the values presented in Table 7.3.
Riser Flow
______________________________________________________________________________________
199
Table 7.3. Settings for the collision parameters
e µ e
w
µ
w
Ideal 1.0 0.0 1.0 0.0
non ideal 0.94 0.28 0.94 0.28
Ideal particles/ nonideal walls 1.0 0.0 0.94 0.28
At first it was checked if each simulation had reached a steady state where the outlet
solids mass flux was equal to the specified solids mass flux of 25 kg/m
2
s. For the
simulation with ideal collision parameters this steady state was reached within 10 s but
for the simulation with non ideal collision parameters it took more than 30 s before the
outlet solids mass flux was equal to inlet solids mass flux. Once the steady state was
reached the simulations were run for 10 s in order to obtain statistically reliable time
averages. In Figure 7.1 the outlet solids mass flux is presented as a function of time for
the simulation with the non ideal collision parameters.
Figure 7.1. Outlet solids flux as a function of time for the simulation with nonideal
collision parameters (e = 0.94, µ = 0.28).
Chapter 7
______________________________________________________________________________________
200
Note that t = 0 s in this figure corresponds to the initial condition for the actual simulation
of the system at steady state conditions. The solids flux during the time required to reach
this steady state is not shown here. In Figure 7.1 it can be observed that the outlet solids
mass flux has a very peaky nature but the time average value is equal to the specified
mass flux (G
s
= 25 kg/m
2
s). It should be noted that in the simulations presented by Tsuo
and Gidaspow (1990) and Tsuji et al. (1998) the time averaged outlet solids mass flux
was by far not equal to 25 kg/m
2
s and therefore their results do not apply to fully
developed riser flow.Two snapshots of simulations with ideal (e = 1.0, µ = 0.0) and non
ideal (e = 0.94 and µ = 0.28) collision parameters are shown in Figure 7.2. These are
snapshots of the middle section of the riser (1.0 to 1.2 m height).
In Figure 7.2 a typical cluster can clearly be observed in the case of nonideal collision
parameters. The snapshot of the simulation with the ideal collision parameters shows a
much more homogeneous distribution of the particles.
Figure 7.2. Snapshots of a small section of the simulated riser (width = 0.08 m). Left:
ideal collision parameters, e = 1.0, µ = 0.0, Right: nonideal collision
parameters: e = 0.94, µ = 0.28.
Riser Flow
______________________________________________________________________________________
201
From animations of the simulations it could be observed that in the simulation with the
non ideal collision parameters the clusters moved downward preferably along the walls
and had a streaky, semiparabolic shape. These clusters are rather dynamic in nature: they
continuously form, grow and break up. Typical downward velocities for these clusters are
in the order of magnitude of 1.0 m/s which is in agreement with the experimental findings
of Tsuo and Gidaspow (1990). In the case of ideal collision parameters the flow was
much more homogeneous and although locally slightly denser regions can be
distinguished, these can hardly be identified as clusters. Similar results were obtained in a
simulation with ideal collision parameters for particleparticle collisions but non ideal
parameters for particlewall collisions.In Figure 7.3 the time averaged (10 s) radial solids
profile at 1.0 m height is presented. For the case with non ideal collision parameters the
typical coreannulus flow structure can be observed where the solids fraction is relatively
high near the wall and relatively low in the centre of the riser duct. In the other two cases
the radial profile is flat. This result is a strong indication that the presence of clusters,
which is clearly observed in the non ideal case, causes the typical coreannulus structure
since no cluster formation, and hence no coreannulus structure, could be observed in the
other two cases where no energy was dissipated in particleparticle collisions.
Figure 7.3. Radial profile of the time averaged solids volume fraction at 1.0 m above
the inlet of the simulated riser.
0.00
0.02
0.04
0.06
0.08
0.10
0.00 0.02 0.04 0.06 0.08
x, m
s
o
l
i
d
s
f
r
a
c
t
i
o
n
,
[

]
ideal particles and walls
nonideal particles and walls
ideal particles, nonideal walls
Chapter 7
______________________________________________________________________________________
202
It is also remarkable that the average solids hold up is higher in the simulation with the
non ideal collision parameters than in the simulations with ideal particleparticle
collisions. This suggests that there is an effect of the collision parameters on the axial
solids profile in the riser.
3.2 Axial effects
The reason why it took the simulation with the non ideal collision parameters much
longer to reach a steady state turned out to be the build up of a solids profile in the riser.
By the time the steady state was reached a total number of about 60,000 particles was
found inside the system whereas in the simulations with ideal particleparticle collisions
this number was about 25,000. This had also a pronounced effect on the CPU time. The
simulation with non ideal collision parameters required over 27 hours of CPU time per
simulated second on a Silicon Graphics Origin200 server with a R10000 processor (180
MHz, 1 Mb cache).
Figure 7.4. Snapshots of the bottom section of the simulated riser (width = 0.08 m)
from 0.0 to 0.2 m. Left: ideal collision parameters, e =1.0, µ = 0.0, Right:
nonideal collision parameters: e = 0.94, µ = 0.28.
Riser Flow
______________________________________________________________________________________
203
The simulations with ideal particleparticle collisions required 2.5 hrs of CPU time per
simulated second on the same machine. The reason for this huge difference is the total
number of collisions that had to be processed in the simulations: 1 million collisions per
second for the simulation with ideal collision parameters, 65 million collisions per second
for the simulation with non ideal collision parameters.Since clusters of particles move
downward in the riser and the particles cannot leave the system at the bottom it can be
understood that an axial solids profile will event ually be established. In Figure 7.4 a
snapshot of the bottom section of the riser is presented for the simulations with ideal and
non ideal collision parameters. In this figure the same difference between ideal and non
ideal collision parameters can be observed as in Figure 7.2 but in the bottom section the
difference is even more pronounced. Large strands of particles move down alongside the
walls and then break up again after impinging on the bottom wall. In Figure 7.5 the time
averaged (10 s) radial solids profile at 0.2 m above the bottom is presented for the three
different cases.
Figure 7.5. Radial profile of the time averaged solids volume fraction at 0.2 m above
the inlet of the simulated riser.
0.00
0.02
0.04
0.06
0.08
0.10
0.00 0.02 0.04 0.06 0.08
x, m
s
o
l
i
d
s
f
r
a
c
t
i
o
n
,
[

]
ideal particles and walls
nonideal particles and walls
ideal particles, nonideal walls
Chapter 7
______________________________________________________________________________________
204
Similar to the radial solids profiles presented in Figure 7.3 the coreannulus flow
structure can be observed in the simulation with non ideal collision parameters but not in
the simulations with ideal particleparticle collisions. Again, hardly any difference could
be observed between the simulation with both ideal particleparticle and particlewall
collisions and the simulation with ideal particleparticle but non ideal particlewall
collisions indicating that particleparticle collisions dominate the flow behaviour. It is
also clear that the average solids hold up for the simulation with the non ideal collisions
at 0.2 m in the riser is higher than at 1.0 m (Figure 7.3). This suggests the existence of a
nonuniform axial solids profile.
In Figure 7.6 the time averaged (10 s) axial distribution of the solids volume fraction in
the riser are presented for the two cases mentioned above. For the non ideal case clearly a
build up of solids in the riser can be observed which is in agreement with the different
total number of particles in the simulations reported earlier (app. 60,000 vs. 25,000). In
this figure the results of the simulation with ideal particleparticle but non ideal particle
wall collisions are not included for clarity reasons. However, these results are very
similar to the results of the simulation with ideal collision parameters.
Figure 7.6. Axial profile of the timeaveraged solids volume fraction (nondeal: e =
0.94, µ = 0.28, ideal: e = 1.0, µ = 0.0)
Riser Flow
______________________________________________________________________________________
205
In Figure 7.6 it can clearly be observed that a build up of solids exists in the case of the
simulation with non ideal particle collisions whereas in the case of ideal collisions the
axial solids distribution is nearly uniform. Chang and Louge (1992) reported results from
experiments using a somewhat different system but operated at a rather similar mass
loading and Froude number. They found a significant change in axial pressure profile in
the riser when they applied a coating to glass particles in order to make them smoother.
For the coated glass particles (i.e. lower coefficient of friction) the pressure drop over the
riser was much less than for the noncoated glass particles indicating a lower solids hold
up in the system. The results of the simulations presented above support the idea that the
difference in the axial pressure profile found in these experiments is caused by the
difference in particleparticle interaction in both cases. In the case where less energy is
dissipated in collisions (i.e. the more ideal system: lower values for the coefficient of
friction, higher values for the coefficient of restitution) less clusters are being formed and
hence there is less internal solids circulation within the riser. This results in a lower solids
hold up in the riser and hence a lower pressure drop.
Chang and Louge (1992) pointed out that these effects become less pronounced at higher
solids fluxes. A simulation was performed with ideal collision parameters at a solids flux
of 50 kg/m
2
s in order to investigate this. Although the system appeared less homogeneous
than at a solids flux of 25 kg/m
2
s still the formation of clusters was far less pronounced
than in the nonideal case with a solids flux of 25 kg/m
2
s. Obviously there is still a large
difference between the simulated system with ideal collision parameters and the coated
glass particles used in the experiments. Nevertheless the simulation results clearly show
the important effect of the collision parameters on the flow structure in a riser and this
result is supported by the experiments of Chang and Louge (1992). It is therefore of great
importance that the collision parameters (i.e. coefficient of restitution and friction) are
taken into account in the scale up as well as the design of riser reactors.
Chapter 7
______________________________________________________________________________________
206
3.3 Influence of lift forces
By default the two lift forces in equation 7.1 and the effect of the gas flow on the rotation
of the particles (equation 7.8) were not included in the simulations. However, an
additional simulation was performed including all these effects in order to investigate
their influence on the flow structure. The case with the non ideal collision parameters and
a superficial gas velocity of 5.0 m/s was chosen as a test case. The time averaged (10 s)
radial distributions of the solids volume fraction at 0.2 m above the riser inlet are
presented in Figure 7.7 for the two simulations.
Figure 7.7. Influence of the lift forces on the timeaveraged radial distribution of the
solids volume fraction at 0.2 m above the inlet of the simulated riser.
From Figure 7.7 it is apparent that the lift forces did not have a significant influence on
the flow structure. The radial segregation is slightly less pronounced when the lift forces
are included in the equations of motion. This can be understood since the lift forces tend
to move the particles toward the centre of the riser and therefore have a redispersive
effect.
0.00
0.02
0.04
0.06
0.08
0.10
0.00 0.02 0.04 0.06 0.08
x, m
s
o
l
i
d
s
f
r
a
c
t
i
o
n
,
[

]
without lift forces
with lift forces
Riser Flow
______________________________________________________________________________________
207
4. Conclusions
Granular Dynamics simulations of riser flow have been performed using a two
dimensional hardsphere discrete particle model. Particles of 500 µm diameter and a
density of 2620 kg/m
3
were introduced in a riser of 0.08 m width and 2.00 m height at a
solids mass flux of 25 kg/m
2
s. By default the superficial gas velocity was set equal to 5.0
m/s. The simulations turned out to be rather sensitive to the collision parameters, i.e. the
coefficients of restitution (e) and friction (µ). When these parameters were assigned non
ideal values (e = 0.94, µ = 0.28) cluster formation was observed. When the collision
parameters were set to ideal values (e = 1, µ = 0) no cluster formation was observed. This
was also the case in a simulation where only particlewall collisions were considered non
ideal whereas particleparticle collisions were still taken to be ideal. The time averaged
radial distribution of the solids volume fraction clearly revealed the typical coreannulus
structure for the simulation with the non ideal collision parameters. For the simulations
with ideal particleparticle collisions the time averaged radial distribution of the solids
volume fraction was rather flat. Particlewall collisions turned out to have very little
influence on the flow structure.
In the non ideal case the solids holdup in the riser was much higher than in the ideal case
as could be observed from axial solids volume fraction profiles. These results are
supported by the experimental findings of Chang and Louge (1992) where a large
difference in axial pressure profile was found when a coating was applied to glass
particles in order to change the surface roughness (i.e. decrease the coefficient of
friction). In those experiments a smaller pressure drop (i.e. smaller solids hold up) was
found for the coated (smoother) particles which is in agreement with the trend observed
in the simulations.
Finally the influence of the lift forces was investigated. These forces turned out to have a
slightly redispersive effect on the flow structure which made the radial segregation of the
Chapter 7
______________________________________________________________________________________
208
solids a little less pronounced. Since this was only a minor effect it could be regarded as a
justification for the fact that the lift forces were neglected in the default simulations.
Notation
C
d
drag coefficient, []
C
m
rotational drag coefficient, []
e coefficient of restitution, []
D
nblist
width of neighbour list square, m
D
p
particle diameter, m
I moment of inertia, kgm
2
F force, N
G
s
solids mass flux, kg/m
2
s
g gravitational acceleration, m/s
2
m
p
particle mass, kg
NX number of computational cells in xdirection, []
NY number of computational cells in ydirection, []
p pressure, Pa
S
p
momentum source term N/m
3
T temperature, K
t time, s
u gas velocity vector, m/s
v
p
particle velocity vector, m/s
V
p
particle volume, m
3
x x coordinate, m
y y coordinate, m
Greek symbols
α coefficient in expression Magnus lift force (table 7.1), []
β coefficient in expression Magnus lift force (table 7.1), []
Riser Flow
______________________________________________________________________________________
209
β
0
coefficient of tangential restitution, []
χ velocity gradient, 1/s
ε void fraction, []
µ coefficient of friction, []
µ
g
gas viscosity, kg/ms
τ gas phase stress tensor, kg/ms
2
ρ density, kg/m
3
ξ dimensionless velocity gradient, []
ω
p
particle rotation velocity, 1/s
ω
r
relative rotation velocity gas/particle, 1/s
Subscripts
g gas phase
LG lift force due to velocity gradient
LM Magnus lift force
p particle
w wall
References
Anderson, T.B. and Jackson, R., (1967). A fluid mechanical description of fluidized beds
(equations of motion), Ind. Eng. Chem., Fundam., 6, 527.
Chang, H. and Louge, M.Y., (1992). Fluid dynamic similarity of circulating fluidized
beds, Powder Technol., 70, 259.
Chen, J.C., (1996). Clusters (The Thomas Baron Plenary Lecture 1995), AIChE Symp.
Ser. 92 (No. 313), 1.
Chapter 7
______________________________________________________________________________________
210
Dennis, S.C.R., Singh, S.N., Ingham, D.B., (1980). The steady flow due to a rotating
particle at low and moderate Reynolds numbers, J. Fluid Mech., 101 (2), 257.
Foerster S.F., Louge, M.Y., Chang,H. and Allia, K., (1994). Measurements of the
collision properties of small spheres, Phys. Fluids, 6, 1108.
Frezzotti, A., (1997). A particle scheme for the numerical solution of the Enskog
equation, Phys. Fluids, 9, 1329.
Helland, E., van den Moortel, T., Occelli, R., Tadrist, L., Hjertager, B.H., Solberg, T. and
Mathiesen, V., (1999). Numerical/experimental study of gasparticle flow behaviour in a
circulating fluidized bed, in Circulating Fluidized Bed Technology VI, J. Werther (Ed.),
p. 261.
Hoomans, B.P.B., Kuipers, J.A.M., Briels, W.J. and van Swaaij, W.P.M. (1996). Discrete
particle simulation of bubble and slug formation in a twodimensional gas fluidised bed:
A hardsphere approach, Chem. Engng Sci., 51, 99.
Horio, M., Kuroki, H., (1994). Threedimensional flow visualization of dilutely dispersed
solids in bubbling and circulating fluidized beds, Chem. Engng Sci., 49, 2413.
Hrenya, C.M. and Sinclair, J., (1997). Effects of particle phase turbulence in gassolid
flows, AIChE J., 43, 853.
Kuipers, J.A.M., van Duin, K.J., van Beckum, F.P.H. and van Swaaij, W.P.M., (1992). A
numerical model of gasfluidized beds, Chem. Engng Sci., 47, 1913.
Kuipers, J.A.M., Hoomans, B.P.B. and van Swaaij, W.P.M., (1998). Hydrodynamic
models of gasfluidized beds and their role for design and operation of fluidized bed
chemical reactors, in Fluidization IX, L.S. Fan and T.M. Knowlton eds., p. 15
Riser Flow
______________________________________________________________________________________
211
Nieuwland, J.J., van Sint Annaland, M., Kuipers, J.A.M. and van Swaaij, W.P.M.,
(1996). Hydrodynamic modelling of gasparticle flows in riser reactors, AIChE J., 37,
1009.
Mei, R., (1992). An approximate expression for the shear lift force on a spherical particle
at finite Reynolds number, Int. J. Multiphase flow, 18, 145.
Oesterlé, B., (1994). Une étude de l’influence des forces transversales agissant sur les
particules dans les écoulements gazsolide, Powder Technol., 79, 81.
Ouyang, J. and Li, J., (1999). Discrete simulations of heterogeneous structure and
dynamic behavior in gassolid fluidization, Chem. Engng Sci., 54, 5427.
Richardson, J.F. and Zaki, W.N., (1954). Sedimentation and Fluidisation: Part I, Trans.
Inst. Chem. Engng, 32, 35.
Rowe, P.N. and Henwood, G.A., (1961). Drag forces in a hydraulic model of a fluidised
bedpart I, Trans. Instn Chem. Engrs, 39, 43.
Saffman, P.G., (1965). The lift on a small sphere in a slow shear flow, J. Fluid Mech.,
22(2), 385.
Sinclair, J. and Jackson, R., (1989). Gasparticle flow in a vertical pipe with particle
particle interactions, AIChE J., 35, 1473.
Tanaka, T., Yonemura, S., Kiribayashi, K., Tsuji, Y., (1996). Cluster formation and
particle induced instability in gassolid flows predicted by the DSMC method, JSME Int.
J. Series B, 39, 239.
Tsuji. Y., Morikawa, Y. and Mizuno, O., (1985). Experimental measurement of the
Magnus lift force on a rotating sphere at low Reynolds numbers, J. Fluids Eng., 107, 484.
Chapter 7
______________________________________________________________________________________
212
Tsuji, Y., Tanaka, T. and Yonemura, S., (1998). Cluster patterns in circulating fluidized
beds predicted by numerical simulation (discrete particle model versus two fluid model),
Powder Technol., 95, 254.
Tsuo, Y.P. and Gidaspow, D., (1990). Computation of flow patterns in circulating
fluidized beds, AIChE J., 36, 885.
Wang, Y. and Mason, M.T., (1992). Twodimensional rigidbody collisions with friction.
J. Appl. Mech., 59, 635.
Experimental Validation using PEPT
___________________________________________________________________________________
213
Chapter 8.
EXPERIMENTAL VALIDATION OF GRANULAR
DYNAMICS SIMULATIONS OF GASFLUIDISED BEDS
WITH HOMOGENOUS INFLOW CONDITIONS USING
POSITRON EMISSION PARTICLE TRACKING
Abstract:
A hardsphere Granular Dynamics model of a twodimensional gasfluidised bed was
experimentally validated using Positron Emission Particle Tracking (PEPT). In the
model the Newtonian equations of motion are solved for each solid particle while
taking into account the particleparticle and particlewall collisions. The gas phase
hydrodynamics is described by the spatially averaged NavierStokes equations for
twophase flow. A quasi twodimensional bed of 0.185 m width and 0.4 m height with
homogenous inflow conditions at 1.5 u
mf
was chosen as a test case. Glass particles (ρp
= 2435 kg/m
3
) with diameters ranging from 1.25 mm to 1.5 mm were used as the bed
material. The collision parameters required in the simulation were obtained from
separate, independent measurements. In the PEPT experiment the motion of a single
tracer particle in the bed was tracked during the period of one hour. In the simulation
the motion of 15,000 particles was tracked during a period of 45 s. The simulation
data was time averaged over 45 s for each particle and subsequently ensemble
averaged over all the particles in the simulation. The comparison was made on the
basis of averaged velocity maps, occupancy plots and speed histograms. The results
showed good agreement between experiment and simulation when the measured
values for the collision parameters were used. When collisions were assumed to be
fully elastic and perfectly smooth the agreement was much worse.
Chapter 8
___________________________________________________________________________________
214
Experimental Validation using PEPT
___________________________________________________________________________________
215
1. Introduction
Experimental validation of flow models for dense gasfluidised beds, that form the
basis for (amongst others) reactor models, is of cr ucial importance to arrive at models
that can predict the performance of fluidised beds with confidence. The trend in both
experiments and models is that systems can be studied in more and more detail with
continuously increasing accuracy. It is not the objective to present a complete review
of available experimental techniques in gasfluidised beds here. A comprehensive
review on measuring techniques in fluidised beds was presented by Werther (1999)
whereas an extensive overview of existing techniques for the measurement of solids
concentration and velocity was presented by Nieuwland et al. (1996).
With increasing computer power Granular Dynamics simulations have become a very
useful and versatile research tool to study the dynamics of dense gasparticle flows. In
these models the Newtonian equations of motion are solved for each individual
granular particle in the system. The mutual interactions between particles and the
interaction between particles and walls are taken into account directly. The discrete
particle approach was pioneered by Tsuji et al. (1993) who used a soft sphere model
to describe the interaction between the particles. A three dimensional version of this
model was later presented by Kawaguchi et al. (1998) whereas Mikami et al. (1998)
extended the (2D) model in order to include cohesive forces between the particles.
Hoomans et al. (1996) presented a hardsphere approach where collisions are assumed
to be binary and instantaneous. Xu and Yu (1997) developed a hybrid technique
combining elements of both softsphere and hardsphere techniques. In the studies on
discrete particle simulation of gas fluidised beds previously reported in the literature
experimental validation has received little attention. This was partly due to the fact
that proper experimental techniques were not widely available.
In Chapter 5 and Chapter 6 of this thesis experimental validation was performed by
direct observation of transient phenomena that occurred in a transparent system by
using a video camera. This is a very basic technique that nonetheless can provide
useful information on for instance the formation of a single bubble at an orifice or the
dynamics of segregation. In the latter case it is required that the segregating species
Chapter 8
___________________________________________________________________________________
216
can be visually identified. However, the direct observation technique is limited to
systems that allow for direct observation (i.e. transparent quasi twodimensional set
ups) in contrast to for example Xray techniques (Yates et al., 1994). Furthermore the
information that is obtained using direct observation covers macroscopic behaviour:
Important quantities such as local solids volume fraction, local solids velocities etc.
cannot be obtained using this technique.
Particle tracking techniques are noninvasive measurement techniques where the
motion of a single tracer particle is tracked for a period of time. Such techniques are
ideally suited to validate a discrete particle model since the motion of a single particle
in the bed can be tracked and therefore it allows for relatively direct comparison
between measured data and simulation results. Lin et al. (1985) presented a technique
where a radioactive tracer particle was used that was made of scandium46 (Sc46). A
total of 12 scintillation detectors were positioned around the fluid ised bed to enable
determination of the location of the particle. More recently Larachi et al. (1996)
applied the tracking technique using Sc46 tracers to gas liquidsolid flows and
Mostoufi and Chaouki (1999) applied the technique to liquidfluidised beds.
In this work a cooperation between the University of Twente and the University of
Birmingham was initiated in order to use the Positron Emission Particle Tracking
(PEPT) technique developed at Birmingham to validate the granular dynamics model
of a gas fluidised bed developed at Twente. PEPT has been developed at Birmingham
since 1987 and was successfully applied to a large number of systems for solids
processing including mixers and gas fluidised beds (Seville et al. 1995). A
comprehensive introduction to PEPT can be found in Stein et al. (1997). Recently a
PEPT system was also developed by Stellema et al. (1998). PEPT differs from other
tracking techniques in the sense that it uses positron emitting radioisotopes which
have the unique feature that their decay leads to simultaneous emission of a pair of
backtoback γrays. In a following section a more elaborate discussion of the PEPT
technique will be presented.
Experimental Validation using PEPT
___________________________________________________________________________________
217
2. Model
Since a detailed description of the model is presented in Chapter 2 and Chapter 3, the
key features will be summarised briefly here.
2.1 Granular dynamics
The collision model as originally developed by Wang and Mason (1992) is used to
describe a binary, instantaneous, inelastic collision with friction. The key parameters
of the model are the coefficient of restitution (0 ≤ e ≤ 1) and the coefficient of friction
(µ ≥ 0). Foerster et al. (1994) have shown that also the coefficient of tangential
restitution (0 ≤ β
0
≤ 1) should be used in order to describe the collision dynamics more
accurately. These three collision parameters are all included in the model.
In the hardsphere approach a sequence of binary collisions is processed. This implies
that a collision list is compiled in which for each particle a collision partner and a
corresponding collision time is stored. A constant time step is used to take the external
forces into account and within this time step the prevailing collisions are processed
sequentially. In order to reduce the required CPU time neighbour lists are used. For
each particle a list of neighbouring particles is stored and only for the particles in this
list a check for possible collisions is performed.
2.2 External forces
The incorporation of external forces differs somewhat from the approach followed by
Hoomans et al. (1996). In this work the external forces are used in accordance with
those implemented in the two fluid model presented by Kuipers et al. (1992) where, of
course, the forces now act on a single particle:
( )
( ) m
d
dt
m
V
V p
p
p
p
p
p p
v
g u v · +
−
− − ∇
β
ε 1
, (8.1)
Chapter 8
___________________________________________________________________________________
218
where m
p
represents the mass of a particle, vp
its velocity, u the local gas velocity and
V
p
the volume of a particle. In equation (8.1) the first term is due to gravity and the
third term is the force due to the pressure gradient. The second term is due to the drag
force where β represents an interphase momentum exchange coefficient as it usually
appears in twofluid models. For low void fractions (ε < 0.80) β is obtained from the
wellknown Ergun equation:
( )
( )
p
p
g
p
g
D D
v u − − +
−
·
ρ
ε
µ
ε
ε
β 1 75 . 1
1
150
2
2
, (8.2)
where D
p
represents the particle diameter, µg
the viscosity of the gas and ρg
the
density of the gas. For high void fractions (ε ≥ 0.80) the following expression for the
interphase momentum transfer coefficient has been used which is basically the
correlation presented by Wen and Yu (1966) who extended the work of Richardson
and Zaki (1954):
( )
65 . 2
1
4
3
−
−
−
· ε ρ
ε ε
β
p g
p
d
D
C v u . (8.3)
The drag coefficient C
d
is a function of the particle Reynolds number and given by:
( )
¹
¹
¹
'
¹
≥
<
·
1000 Re 44 . 0
1000 Re Re 0.15 + 1
Re
24
0.687
p
p p
p
d
C ,
(8.4)
where the particle Reynolds number (Re
p
) in this case is defined as follows:
g
p p g
p
D
µ
ερ v u −
· Re . (8.5)
For the integration of equation (8.1) a simple explicit first order scheme was used to
update the velocities and positions of the particles.
Experimental Validation using PEPT
___________________________________________________________________________________
219
2.3 Gas phase hydrodynamics
The calculation of the gas phase hydrodynamics mainly follows the lines presented by
Kuipers et al. (1992). It is based on the numerical solution of the following set of
partial differential equations that can be seen as a generalised form of the Navier
Stokes equations for a gas interacting with a solid phase as originally derived by
Anderson and Jackson (1967).
Continuity equation gas phase:
( )
( )
∂ ερ
∂
ερ
g
g
t
+ ∇⋅ · u 0 . (8.6)
Momentum equation gas phase:
( )
( ) ( )
∂ ερ
∂
ερ ε ε ερ
g
g p g g
t
p
u
uu S g + ∇ ⋅ · − ∇ − − ∇ ⋅ + τ . (8.7)
In this work transient, twodimensional, isothermal (T = 293 K) flow of air at
atmospheric conditions is considered. The constitutive equations can be found in
Chapter 3. There is one important modification with respect to the model presented by
Hoomans et al. (1996) and that deals with the way in which the twoway coupling
between the gasphase and the dispersed particles is established. In the present model
the reaction force to the drag force exerted on a particle per unit of volume is fed back
to the gas phase through the source term Sp
which has the dimension of force per unit
of volume N/m
3
.
3. Positron Emission Particle Tracking
Positron Emission Particle Tracking (PEPT) is a technique that allows noninvasive
observation of the motion of a single radioactive tracer particle. The PEPT technique
is schematically represented in Figure 8.1.
Chapter 8
___________________________________________________________________________________
220
a) detection
b) reconstruction
c) particle location
Figure 8.1. The principle of the PEPT technique
In the experiment reported here a glass particle taken from the sample of particles used
in the fluidisation experiment was activated by direct irradiation in a cyclotron beam.
detector detector
Experimental Validation using PEPT
___________________________________________________________________________________
221
The glass particle was irradiated with the
3
He beam from a cyclotron to produce the
positron emitter
18
F from reactions involving the oxygen on the glass. The decay of the
18
F isotope features the conversion of a proton to a neutron with the emission of a
positron, the antiparticle of the electron. The positron then annihilates with an
electron to produce a pair of back to back γrays. The γrays are detected by the
positron cameras which consists of two positionsensitive γray detectors that have an
active area of 0.3 by 0.6 m. By using a reconstruction algorithm the position of the
particle can be obtained as the intersection point of successive annihilation vectors.
The algorithm employs an iterative scheme to discard corrupt annihilation vectors that
can be caused by γray scattering or random coincidences. When the tracer particle
does not move the more annihilation vectors are used the more accurate the particle
position can be determined. However, when the particle is moving the set of
annihilation vectors should be large enough to locate the particle accurately but not so
large that it has moved significantly during the time period over which the set was
measured.
The technique was not optimised for the use of a quasi twodimensional system and
for the post processing the standard software was used. The instantaneous particle
velocity is inferred from the difference between successive locations.
4. Comparison between PEPT data and simulation
The PEPT experiment as performed in this work renders the trajectory of a single
particle during one hour. This time scale cannot be reached by means of simulation on
modern day computers. Instead the motion of 15,000 particles is tracked during a
shorter period of time (45 s). In fact considerably more data is generated in the
simulation since 15,000 times 45 seconds is a far larger number than 1 times 3600
seconds. And since the 15,000 simulated particles cover the whole of the fluidised
system, the simulation data does not suffer from poorly sampled regions in contrast
with the experimental data.
Chapter 8
___________________________________________________________________________________
222
It is assumed that it is justified to compare the simulation results with the time
averaged experimental data by first time averaging the simulation data for each of the
15,000 particles and subsequently taking an ensemble average over all the 15,000
particles. By doing this there is a risk that rare events, which occur over a time scale
close to or greater than the duration of a simulation, are poorly sampled. On the
contrary, the chance that such a rare event is experienced by the tracer particle in the
experiment is rather low as well.
The actual comparison is made on the basis of occupancy plots, velocity maps and
speed histograms which are standard outputs of the PEPT software (Tin et al., 1997)
that was supplied by the University of Birmingham. To obtain an occupancy plot from
the PEPT data the system is first divided into cells (10 mm width/height). In the
occupancy plot the fraction of the total time that the tracer particle has spend in that
particular cell is displayed using a colour code that is explained in the legend
accompanying the plot. For the simulations a similar procedure was followed after all
the individual particle trajectories were added together. In order to obtain the velocity
maps the system was again divided into cells. The velocity vectors in the figure are
averaged velocity for that particular cell. How the averages were obtained is explained
above. Finally the PEPT data was compared with the results of the simulation using
speed histograms. For the PEPT experiment the speed of the tracer particle was
calculated on the basis of the 3 velocity components (x,y,z) at each instant. The
histogram shows the speed distribution based on the data set covering the entire
duration of the experiment of one hour. The speed histograms obtained from the
simulation results are based on the two velocity components (x,y) taken into account
in the simulations.
5. Results
As a test system for the experimental validation a gasfluidised bed (0.185 m width, 20
mm depth) with homogeneous inflow conditions (u
g
= 1.5 u
mf
, u
mf
= 0.9 m/s) was
chosen. The parameter settings for the simulation are summarised in table 8.1. The bed
was filled with the glass particles in such a way the static bed height was about 0.17 m.
Experimental Validation using PEPT
___________________________________________________________________________________
223
In the simulation a total number of 15,000 particles was used to match the bed height
encountered in the experiment. The bed used in the experiment was assembled using
glass as the construction material. It was fitted with a porous plate distributor with a
sufficiently high pressure drop to ensure homogeneous inflow conditions. In
preliminary experiments a malfunctioning distributor plate was used which caused a
strongly nonsymmetrical solids flow pattern in the bed indicating a nonuniform gas
distribution. The actual PEPT experiment used for the experimental validation was run
for one hour in order to collect sufficient data to calculate statistically reliable
averages.
Table 8.1. Parameter settings for the PEPT simulation
Particles: Bed:
Shape spherical Width 0.185 m.
density, ρp
2418 kg/m
3
Height 0.40 m
particle diameter, D
p
1.251.50 mm Number xcells, NX 37
Number ycells, NY 80
e = e
w
0.97 cell width, DX 5 mm
µ = µ
w
0.10 cell height, DY 5 mm
β0
= β0,w
0.33
N
p
15,000 time step, DT 10
–4
s
The particleparticle collision parameters presented in table 8.1 were independently
measured by Gorham and Kharaz (1999) using the facility at the Open University at
Milton Keynes described in Chapter 2. The particlewall collision parameters were
assumed to be equal to the particleparticle collision parameters. This is justified since
earlier simulations (Chapter 4) showed that the influence of the particle wall collision
parameters were negligible compared to the influence of the particleparticle collision
parameters. A lognormal particle size distribution about an average diameter of 1.375
mm was taken into account in the simulation. Particle diameters lower than 1.25 mm
and greater than 1.50 mm were rejected in order to mimic the effects of sieving and
hence match the particle size distribution encountered in the experiment.
Chapter 8
___________________________________________________________________________________
224
Figure 8.2. Snapshots of particle configurations of the simulation.
An initial simulation was performed first to ensure that the actual simulation would not
suffer from any startup effects. During this initial simulation the system was operated
at a homogeneous gas inflow of 1.5 u
mf
(u
mf
= 0.9 m/s) with the parameter settings
Experimental Validation using PEPT
___________________________________________________________________________________
225
presented in table 8.1. After this initial simulation the actual simulation (duration 45 s)
was started and the required data (positions and velocities) was stored at time intervals
correspond ing to those used in the experiment. Snapshots of the simulation are
presented in Figure 8.2. In this figure it can be observed that the bed is bubbling quite
vigorously and this behaviour could also be visually observed during the PEPT
experiment. In Figure 8.3 an example of the trajectory of one single (randomly chosen)
particle in the simulation is presented. This trajectory covers the whole duration of the
simulation of 45 s.
Figure 8.3. Example of a trajectory of a single particle during the simulation.
Although the figure above gives some insight into the motion of a particle in a
fluidised bed it does not provide a solid basis for a comparison. Therefore averaging
techniques were applied as discussed in the previous section which yield data that
permit a comparison between the PEPT data and the simulation results.
In Figure 8.4 the velocity map obtained from the PEPT data is presented together with
the velocity maps obtained from two simulations. In the centre the velocity map of the
simulation using the measured values for the collision parameters (table 8.1) is
presented and on the right the velocity map of a simulation assuming fully elastic,
perfectly smooth collisions (e = 1, µ = 0, also referred to as ideal collisions) is
presented. A reference vector indicating the magnitude of the velocity is included in
all of the three velocity maps. In velocity map of the PEPT data a circulation pattern
Chapter 8
___________________________________________________________________________________
226
can be observed where particles rise in the centre of the bed and descend near the
walls. Two circulation cells can be distinguished which together form a rather
symmetric picture indicating that the gas inflow was indeed homogeneous.
Figure 8.4. Velocity map obtained from the PEPT data (left) compared with the
velocity maps of the simulation using the measured collision
parameters (centre) and the simulation assuming fully elastic, perfectly
smooth collision (right)
From the simulation with the measured collision parameters a very similar velocity
map was obtained. Although not perfectly symmetric two circulation cells can be
distinguished which is in good agreement with the PEPT data. The velocity map of the
simulation assuming fully elastic, perfectly smooth collisions shows a somewhat
different behaviour. The velocity vectors are smaller indicating lower speeds and also
two additional circulation cells on top of the two main cells can be observed. These
additional circulation cells rotate in the opposite direction and were not present in the
PEPT experiment.
In Figure 8.5 the occupancy plots obtained from the PEPT data and the two
simulations are presented together. In this figure it can be observed that in the PEPT
data the occupancy was higher near the walls which was also the case in the simulation
with the measured collision parameters. This indicates that the particles spend
relatively more time near the walls. In the simulation assuming fully elastic, perfectly
Experimental Validation using PEPT
___________________________________________________________________________________
227
smooth collisions the occupancy is almost the same at each position in the system. In
other words the residence time of the particles is evenly distributed throughout the
whole system. This is due to the absence of bubbles in the latter case which causes a
very homogeneous type of fluidisation.
Figure 8.5. Occupancy plots obtained from the PEPT data (above), the simulation
using the measured collision parameters (left) and the simulation
assuming fully elastic, perfectly smooth collisions (right)
It should be noted that the occupancy plot obtained from the PEPT data is not as
smooth as the ones obtained from the simulations. This is obviously due to the fact
that in the experiment one single particle was tracked for one hour. This does not
guarantee that the whole system is covered and therefore it can be understood that
Chapter 8
___________________________________________________________________________________
228
some regions of the bed suffer from poor statistics. For the simulation results all the
particles in the system were taken into account which renders a far smoother picture
since the whole of the bed was automatically sampled.
Figure 8.6. Speed histogram obtained from the PEPT experiment (top) compared
with the speed histograms obtained from the simulation using the
measured collision parameters (left) and the simulation assuming
perfectly smooth, fully elastic collisions (right)
Experimental Validation using PEPT
___________________________________________________________________________________
229
In Figure 8.6 the speed histogram obtained from the PEPT experiment is presented
together with the speed histograms obtained from the two simulations. Where the
velocity map and the occupancy plot provide a basis for a more qualitative comparison
between simulation and experiment, the speed histogram offers the possibility for a
quantitative comparison.
In Figure 8.6 it can be observed that the results of the simulation using the measured
collision parameters compare rather well with results of the PEPT experiment.
Although the simulation results show a somewhat higher average speed than the PEPT
data (about 0.25 m/s in the simulation and 0.15 m/s in the experiment) the shape of the
distribution is rather similar. The higher average speed is most likely due to the two
dimensional nature of the simulation. The absence of the front and back walls in the
simulation implies that the particles in the simulation do experience a lower amount of
wall friction that effectively results in a higher average speed. It is important to note
that in both the experiment and the simulation particle speeds above 0.4 m/s are
observed. Since these higher speeds are closely related to the bubbling behaviour in
the bed it is important that good agreement between simulation and experiment is
achieved on this matter. Although the speeds observed in the simulation are somewhat
higher than the ones observed in the experiment for reasons discussed above, the
agreement is encouraging.
The agreement between the results of the PEPT experiment and the simulation
assuming fully elastic, perfectly smooth collisions is much worse in this respect. In the
latter simulation the average speed (about 0.1 m/s) is actually closer to the average
speed observed in the experiment but more importantly however, the distribution of
speeds in the histogram is far more narrow: no speeds above 0.4 m/s were observed.
The reason for this is the absence of bubbles in the simulation with ideal collision
parameters since the particles attain their maximum speeds when they are accelerated
into the wake of a bubble. This once again emphasises the profound influence that
these collision parameters have on the dynamics of a gas fluidised bed.
Chapter 8
___________________________________________________________________________________
230
6. Conclusions
Granular dynamics simulations of gasfluidised beds with homogeneous inflow
conditions were experimentally validated using the Positron Emission Particle
Tracking facility at the University of Birmingham. A quasi twodimensional bed of
0.185 m width and 0.4 m height with homogenous inflow conditions at 1.5 u
mf
was
chosen as a test case. Glass particles (ρ
p
= 2435 kg/m
3
) with diameters ranging from
1.25 mm to 1.5 mm were used as the bed material. In the PEPT experiment the motion
of a single tracer particle in the bed was tracked during the period of one hour. In the
simulation the motion of 15,000 particles was tracked during a period of 45 s. The
simulation data was time averaged over 45 s for each particle and subsequently
ensemble averaged over all the particles in the simulation. The results showed good
agreement between experiment and simulation when measured values for the collision
parameters were used. The particle speeds observed in the simulation were somewhat
higher than those observed in the experiment which is most likely due to the absence
of the front and back wall in the simulation. When collisions were assumed to be fully
elastic and perfectly smooth the agreement was worse. No speeds higher than 0.4 m/s
were observed whereas in the experiment as well as in the simulation with the
measured collision parameters speeds up to 0.8 m/s were found. This demonstrates the
profound influence of the collision parameters on the bed hydrodynamics since these
higher speeds are closely related to the presence of bubbles in the bed. From a direct
comparison with an experiment it was therefore shown that the assumption of fully
elastic, perfectly smooth collisions is not valid for fluidised bed simulations.
From this first comparison between granular dynamics simulations and a PEPT
experiment it can be concluded that PEPT is a powerful tool for the experimental
validation of these simulations. In future work it should not only be attempted to
perform the validation on a more detailed level but also the simulation data can be
used to verify the PEPT software. Since the simulation yields particle positions and
corresponding velocities that are exact, the PEPT software should be able reproduce
these results.
Experimental Validation using PEPT
___________________________________________________________________________________
231
Notation
C
d
drag coefficient, []
e coefficient of restitution, []
D
p
particle diameter, m
DT time step, s
DX horizontal computational cell dimension, m
DY vertical computational cell dimension, m
g gravitational acceleration, m/s
2
m
p
particle mass, kg
N
p
number of particles, []
NX number of computational cells in xdirection, []
NY number of computational cells in ydirection, []
p pressure, Pa
r position vector, m
S
p
momentum source term N/m
3
T temperature, K
t time, s
u gas velocity vector, m/s
v
p
particle velocity vector, m/s
V
p
particle volume, m
3
Greek symbols
β volumetric interphase momentum transfer coefficient, kg/(m
3
s)
β0
coefficient of tangential restitution, []
ε void fraction, []
µ coefficient of friction, []
µ
g
gas viscosity, kg/ms
τ gas phase stress tensor, kg/ms
2
ρ density, kg/m
3
Subscripts
g gas phase
Chapter 8
___________________________________________________________________________________
232
mf minimum fluidisation
p particle
w wall
References
Anderson, T.B. and Jackson, R., (1967). A fluid mechanical description of fluidized
beds (equations of motion), Ind. Eng. Chem., Fundam., 6, 527.
Foerster S.F., Louge, M.Y., Chang,H. and Allia, K. (1994). Measurements of the
collision properties of small spheres. Phys. Fluids,, 6, 1108.
Gorham, D.A. and Kharaz, A.H., (1999). Results of particle impact tests, Impact
Research Group Report IRG 13, The Open University, Milton Keynes, UK.
Hoomans, B.P.B., Kuipers, J.A.M., Briels, W.J. and van Swaaij, W.P.M., (1996).
Discrete particle simulation of bubble and slug formation in a twodimensional gas
fluidised bed: a hardsphere approach, Chem. Engng Sci., 51, 99.
Kuipers, J.A.M., van Duin K.J., van Beckum, F.P.H. and van Swaaij, W.P.M., (1992).
A numerical model of gasfluidized beds, Chem. Engng Sci., 47, 1913
Larachi, F., Cassanello, M., Chaouki, J. and Guy, C., (1996). Flow structure of the
solids in a 3D gasliquidsolid fluidized bed, AIChE J., 42, 2439.
Lin, J.S., Chen, M.M. and Chao, B.T., (1985). A novel radioactive particle tracking
facility for measurement of solids motion in gas fluidized beds, AIChE J., 31, 465.
Kawaguchi, T., Tanaka, T. and Tsuji, Y., (1998). Numerical simulation of two
dimensional fluidized beds using the discrete element method (comparison between
the two and threedimensional models), Powder Technol., 96, 129.
Experimental Validation using PEPT
___________________________________________________________________________________
233
Mikami, T., Kamiya, H. and Horio, M., (1998). Numerical simulation of cohesive
powder behavior in a fluidized bed, Chem. Engng Sci., 53, 1927.
Mostoufi, N. and Chaouki, J., (1999). Prediction of effective drag coefficient in
fluidized beds, Chem. Engng Sci., 54, 851.
Nieuwland, J.J., Meijer, R., Kuipers, J.A.M. and van Swaaij, W.P.M., (1996),
Measurements of solids concentration and axial solids velocity in gassolid twophase
flows, Powder Technol., 87, 127.
Richardson, J.F. and Zaki, W.N., (1954). Sedimentation and fluidization: part I.
Trans. Inst. Chem. Eng., 32, 35.
Seville, J.P.K., Simons, S.J.R., Broadbent, C.J., Martin, T.W., Parker, D.J. and
Beynon, T.D., (1995). Particle velocities in gasfluidised beds, Fluidization VIII,
Engineering Foundation, Preprints Tours, 319.
Stein. M., Martin, T.W., Seville, J.P.K., McNeil, P.A. and Parker, D.J., (1997).
Positron emission particle tracking: particle velocities in gas fluidised beds, mixers
and other applications, in Noninvasive monitoring of multiphase flows, Chaouki, J.,
Larachi, F. and Dududkovic, M.P. (eds), Elsevier Science, Amsterdam, the
Netherlands.
Stellema, C.S., Vlek, J., Mudde, R.F., de Goeij, J.J.M. and van den Bleek, C.M.,
(1998). Development of an improved positron emission particle tracking system,
Nucl. Instr. And Meth. In Phys. Res. A, 404, 334.
Tan, M., Parker, D.J. and Dee, P.R., (1997). PEPT data presentation software, manual,
Birmingham University, UK.
Tsuji, Y., Kawaguchi, T. and Tanaka, T., (1993). Discrete particle simulation of two
dimensional fluidized bed, Powder Technol. 77, 79.
Chapter 8
___________________________________________________________________________________
234
Wang, Y. and Mason, M.T., (1992). Twodimensional rigidbody collisions with
friction. J. Appl. Mech., 59, 635.
Werther, J., (1999). Measurement techniques in fluidized beds, Powder Technol., 102,
15.
Wen, C.Y. and Yu, Y.H., (1966). Mechanics of fluidization. Chem. Eng. Prog. Symp.
Ser., 62 (62), 100.
Xu, B.H. and Yu, A.B., (1997). Numerical simulation of the gassolid flow in a
fluidized bed by combining discrete particle method with computational fluid
dynamics. Chem. Engng Sci. 52, 2785.
Yates, J.G., Cheesman, D.J. and Sergeev, Y.A., (1994). Experimental observations of
voidage distribution around bubbles in a fluidized bed, Chem. Engng Sci. 49, 1885.
Publications
________________________________________________________________________
235
Publications
1. B.P.B. Hoomans, J.A.M. Kuipers, W.J. Briels and W.P.M. van Swaaij, (1996).
Discrete particle simulation of bubble and slug formation in twodimensional gas
fluidised beds: a hardsphere approach, Chem. Engng Sci., 51, 99.
2. B.P.B. Hoomans, J.A.M. Kuipers and W.P.M. van Swaaij, (1998). Discrete particle
simulation of segregation phenomena in dense gas fluidized beds, in Fluidization IX,
L.S. Fan and T.M. Knowlton (eds), 485.
3. J.A.M. Kuipers, B.P.B. Hoomans and W.P.M. van Swaaij, (1998). Hydrodynamic
models of gas fluidized beds and their role for design and operation of fluidized bed
chemical reactors, in Fluidization IX, L.S. Fan and T.M. Knowlton (eds), 15.
4. B.P.B. Hoomans, J.A.M. Kuipers and W.P.M. van Swaaij, (1998). Granular
Dynamics Simulation of Cluster Formation in Dense Riser Flow, Third International
Conference on Multiphase Flows 1998, June 812, Lyon, France.
5. B.P.B. Hoomans, J.A.M. Kuipers and W.P.M. van Swaaij, (1998). The influence of
particle properties on pressure signals in dense gasfluidised beds: a computer
simulation study, World Congress on Particle Technology 3, July 79, Brighton, UK.
6. B.P.B. Hoomans, J.A.M. Kuipers, W.J. Briels and W.P.M. van Swaaij, (1998).
Comments on the paper ‘Numerical simulation of the gassolid flow in a fluidized bed
by combining discrete particle method with computational fluid dynamics’ by B.H.
Xu and A.B. Yu, Chem. Engng Sci., 54, 2645.
7. B.P.B. Hoomans, J.A.M. Kuipers and W.P.M. van Swaaij, (1998). The influence of a
particle size distribution on the granular dynamics of dense gasfluidized beds: a
computer simulation study, AIChE Symp. Ser. No. 318, Vol. 94, 15.
Publications
________________________________________________________________________
236
8. B.P.B. Hoomans, J.A.M. Kuipers and W.P.M. van Swaaij, (1999). Discrete particle
simulation of cluster formation in dense riser flow, Circulating Fluidized Bed
Technology VI, J. Werther (Ed.), .
9. B.P.B. Hoomans, J.A.M. Kuipers and W.P.M. van Swaaij, (2000). Granular dynamics
simulation of segregation phenomena in dense gasfluidised beds, Powder
Technology (in press).
*
vrij naar Marco Stam, afstudeerverslag Universiteit Twente, 1996. 237
Dankwoord (Ackowledgements)
En zo komt er een einde aan een promotie met ballen
*
. Een promotie die zeker niet het
werk is van slechts één persoon. De mensen die een bijdrage hebben geleverd aan de
totstandkoming van dit proefschrift wil ik op deze plek dan ook graag bedanken.
Allereerst gaat mijn dank uit naar mijn drie promotoren. Modellering van de
hydrodynamica van meerfasenstromingen is een onderzoekslijn waarmee Hans Kuipers
inmiddels wereldwijd een grote naam heeft opgebouwd. Hans, het was me een waar
genoegen om deel uit te maken van jouw team. Ik ben je dankbaar voor het grote
vertrouwen dat je in me had. Ook zonder Wim Briels was dit promotieproject niet van de
grond gekomen. Zijn vermogen om formules te doorgronden en die vervolgens te
manipuleren is ook de kwaliteit van dit proefschrift ten goede gekomen. Wim, ik denk
met veel plezier terug aan de tijd die ik in jouw groep heb doorgebracht. Wim van Swaaij
volgde het onderzoek weliswaar van wat grotere afstand maar wist op beslissende
momenten een zet in de goede richting te geven. Wim, ik zal altijd onthouden dat je me
een sleepkabel aanreikte toen ik met pech langs de weg stond.
Een vijftal afstudeerders heeft een zeer gewaardeerde bijdrage geleverd aan de diverse
onderwerpen die in dit proefschrift aan bod komen. Robert Heijnen (segregatie, het idee
voor het plaatje op de kaft van dit proefschrift is aan zijn brein ontsproten), Marco Stam
(zachte ballen), Peter Schinkelshoek (riserstroming), Reneke van SoestSeij (3D harde
ballen) en Jan Gerard Schellekens (segregatie): allemaal ontzettend bedankt !
Veel steun heb ik gehad aan mijn collega’s Erik Delnoij, Mathijs Goldschmidt, Michiel
Gunsing en Jie Li van de multiphase flow crew van Hans Kuipers waar we dankbaar van
elkaars kennis en vaardigheden gebruik konden maken.
De experimenten beschreven in hoofdstuk 5 waren niet mogelijk geweest zonder de hulp
van Jelle Nieuwland, Patrick Huttenhuis en Olaf Veehof. Ik wil Bart Lammers bedanken
voor het ontwikkelingswerk aan de visualisatiesoftware dat hij heeft uitgevoerd tijdens
Dankwoord
________________________________________________________________________
238
zijn afstudeeropdracht bij Erik. De videoanimaties van mijn simulaties zijn inmiddels in
alle werelddelen te zien geweest. Mathijs heeft een cruciale rol gespeeld bij de
totstandkoming en het bedrijven van de opstelling die gebruikt is voor de experimenten
beschreven in hoofdstuk 6. Hierbij werd hij technisch ondersteund door Gerrit
Schorfhaar. PaulGuillaume Schmitt voerde in het kader van zijn stage een deel van de
metingen met deze opstelling uit (merci beaucoup !) en dankzij Siebren Mellema konden
de videobeelden digitaal verwerkt worden. I wish Jie Li and Pranay Darda all the best
with their continuation of the discrete particle work.
Professor Jonathan Seville and Dr. David Parker of the University of Birmingham are
gratefully acknowledged for their support of the PEPT validation project reported in
Chapter 8. I want to thank Matthias Stein, Mohamad Amran Mohd Salleh, Dennis Allen,
and Aidan McCormack for their valuable contribution to this BirminghamTwente
project. The collision parameters required for the validation simulations in Chapter 6 and
Chapter 8 were experimentally obtained by Dr. David Gorham and Ahmad Kharaz of the
Open University at Milton Keynes.
Deze metingen werden financiëel ondersteund door het Unilever Research Laboratorium
Vlaardingen. Ik wil Joop Olieman, Frank van de Scheur en AtzeJan van der Goot
bedanken voor de mogelijkheid die ze Mathijs en mij geboden hebben om te participeren
in een groots opgezet internationaal Unilever onderzoeksprogamma. Dit heb ik als zeer
leerzaam ervaren.
Met veel plezier denk ik terug aan de contacten met Silicon Graphics (tegenwoordig
SGI). Met name de cursussen in het kader van de Powergroup en de directe interactie met
experts op het gebied van supercomputing heb ik erg op prijs gesteld. Ruud van der Pas,
Ronald van Pelt, Peter Michielse en Bart Mellenberg: bedankt !
Secretariële ondersteuning werd altijd vriendelijk en correct verzord door (in
chronologische volgorde): Greet van der VoortKamminga, Bartie Bruggink, Gery
Stratingh, Yvonne Bruggert, Nicole Haitjema, Annemiek Vos, Ria Stegehuisde Vegte en
Dankwoord
________________________________________________________________________
239
Brigitte Sanderink. Rik Akse zorgde voor een keurige afhandeling van de financiële
zaken.
Ik wil Ben Betlem, Niels Kruyt, Ties Bos, Martin van der Hoef en Peter Roos hier
nogmaals bedanken voor de bijdrage die ze geleverd hebben in de diverse
afstudeercommissies binnen het promotieproject.
Ook buiten kantooruren om heb ik me altijd uitstekend vermaakt met mijn collega AIO’s
in Enschede. Met groot plezier denk ik terug aan de diverse borrels, kroegtochten en
etentjes, het zaalvoetballen met Proceskunde en vooral ook het karten om de Briels trofee
en De Dhont Cup met Computational Chemistry (nog altijd ongeslagen binnen CT !).
Reinier, Harald, Johan, Frank, Roelof, Nico, Wouter, Hein, Wim, Margie en José,
bedankt !
Een speciaal woord van dank is op zijn plaats voor mijn goede vriend en paranimf
Martijn Oversteegen. Tinus heeft in de beginperiode een zeer belangrijke rol gespeeld
door mij te attenderen op de simulatietechnieken uit de Moleculaire Dynamica. De basis
voor dit proefschrift is in die periode gelegd. Verder is hij de afgelopen jaren altijd goed
geweest voor veel gezelligheid, een luisterend oor en een kritische noot op zijn tijd.
Tinus, het was me een groot genoegen !
Tot slot wil ik mijn ouders, Peter en Ine, en mijn broers (paranimf) Tom & Rick
bedanken die samen een solide basis vormen waar ik altijd op kon terugvallen. Allemaal
ontzettend bedankt !
Bob.
Dankwoord
________________________________________________________________________
240
241
Levensloop
Bob Hoomans werd geboren op 2 augustus 1971 in Oldenzaal waar hij ook de lagere
school bezocht. In 1989 behaalde bij het VWO diploma aan het Thijcollege te Oldenzaal.
In september 1989 werd begonnen met de studie Chemische Technologie aan de
Universiteit Twente in Enschede. In het kader van deze opleiding liep hij van september
1993 tot maart 1994 stage bij het Joint Research Centre (JRC) in Ispra, Italië. In januari
1995 studeerde hij af bij de vakgroep Proceskunde op het onderwerp discrete
deeltjessimulatie van gasgefluïdiseerde bedden.
In juni 1995 trad hij in dienst als promovendus (AIO) bij de vakgroepen Chemische
Fysica (later Computational Chemistry) en Proceskunde waar hij onderzoek verrichte op
het gebied van granulaire dynamica van gasvast tweefasenstromingen. De resultaten van
dit onderzoek vormen de basis voor dit proefschrift.
In oktober 1999 trad hij in dienst als CFD engineer bij DSM Research in Geleen.
242
Dit proefschrift is goedgekeurd door de promotoren
prof.dr.ir. J.A.M. Kuipers prof.dr. W.J. Briels prof.dr.ir. W.P.M. van Swaaij
ii
iii
Samenstelling promotiecommissie:
prof.dr. W.E. van der Linden, voorzitter prof.dr.ir. J.A.M. Kuipers, promotor prof.dr. WJ. Briels, promotor prof.dr.ir. W.P.M. van Swaaij, promotor prof.dr. J.P.K. Seville prof.dr.ing. B.H. Hjertager prof.ir. C.M. van den Bleek prof.dr.ir M.M.C.G. Warmoeskerken prof.dr.ir. G.F. Versteeg
Universiteit Twente Universiteit Twente Universiteit Twente Universiteit Twente University of Birmingham, UK Aalborg University Esbjerg, DK Technische Universiteit Delft Universiteit Twente Universiteit Twente
Cover: snapshots of a simulation with a ternary density distribution (Chapter 6)
© december 1999 Bob Hoomans, Maastricht, Nederland. All rights reserved. Second impression august 2001 ISBN 903651401
iv
Contents ______________________________________________________________________________________
Contents
Summary Samenvatting 1 5
1.
GENERAL INTRODUCTION Abstract 1. Fluidisation 2. Hydrodynamic modelling 2.1 Multiscale modelling 3. Discrete particle modelling 4. Outline of this thesis References 9 11 12 13 16 17 19
2.
GRANULAR DYNAMICS Abstract 1. Introduction 1.1 Hardparticle approaches 1.2 Softparticle approaches 1.3 Monte Carlo techniques 2. Hardsphere approach 2.1 Collision model 2.2 Key parameters of the collision model 2.3 Sequence of collisions 2.4 Optimisation 3. Softsphere approach 3.1 The linearspring/dashpot model 3.2 Model parameters 4. Hardsphere vs. Softsphere 4.1 Static situations 4.2 Spring stiffness 4.3 Energy considerations 4.4 Multiple particle interactions 5. Measurement of collision parameters 6. External forces Notation References 23 25 26 28 31 32 32 38 40 43 47 48 53 56 57 57 59 60 62 66 68 70
v
2 Momentum transfer Notation References 79 81 81 82 83 84 85 85 85 86 87 88 89 89 91 92 93 95 4.2 2D versus 3D 7.1 Void fraction 6.1 Granular dynamics 2.1.Contents ______________________________________________________________________________________ 3. Introduction 2.1 Gas phase density 3.3 Dissipative Particle Dynamics 2.2 External forces 2. Conclusions Notation References 99 101 104 104 105 106 107 113 116 118 119 121 123 124 125 vi . GAS PHASE HYDRODYNAMICS Abstract 1. Boundary conditions 6. Model 2. Two way coupling 6. Governing equations 3.1.2 Calculation of the void fraction in 3D 6. Numerical solution 5. Energy conservation 5.2 Gas phase stress tensor 4.1 Direct solution of the NavierStokes equations 1.2 Lattice Boltzmann simulations 1. THE EFFECT OF PARTICLE PROPERTIES ON THE HYDRODYNAMICS OF GASFLUIDISED BEDS WITH HOMOGENEOUS INFLOW CONDITIONS Abstract 1.3 Gas phase hydrodynamics 3. Constitutive equations 3. Effects of collision parameters 4.1 Calculation of the void fraction in 2D 6.1 Influence of collision parameters 6. 3D simulations 6. Influence of a particle size distribution 6. Introduction 1.
Binary size distribution 4.2 Statistical analysis of segregation 4.3 External forces 2. GRANULAR DYNAMICS SIMULATION OF BUBBLE FORMATION IN A GASFLUIDISED BED: HARDSPHERE VS.1 Experimental 5. SOFTSPHERE APPROACH Abstract 1. Conclusions Notation References 129 131 132 132 133 134 136 137 142 142 144 146 149 151 153 154 6. Introduction 2.2 Influence of a particle size distribution 4.1 Experimental 4.4 Gas phase hydrodynamics 3.4 Gas phase hydrodynamics 3. Ternary density distribution 4. Conclusions Notation References 157 159 161 161 162 164 165 166 169 169 172 174 175 175 177 183 184 186 vii . Experimental validation 4. softsphere 4.4 Effects of collision parameters 5.2 Results 6. Introductio n 2. Experimental validation 5. GRANULAR DYNAMICS SIMULATION OF SEGREGATION PHENOMENA IN BUBBLING GASFLUIDISED BEDS Abstract 1.1 Hardsphere granular dynamics 2.1 Hardsphere granular dynamics 2.3 External forces 2.2 Softsphere granular dynamics 2.3 Hardsphere vs. Models 2.3 Effects of collision parameters 5.1 Base case 4.Contents ______________________________________________________________________________________ 5.2 Softsphere granular dynamics 2. Preliminary simulations 4. Models 2.
EXPERIMENTAL VALIDATION OF GRANULAR DYNAMICS SIMULATIONS OF GASFLUIDISED BEDS WITH HOMOGENEOUS INFLOW CONDITIONS US ING POSITRON EMISSION PARTICLE TRACKING Abstract 1. Introduction 2.4 Gas phase hydrodynamics 3. Introduction 2.1 Effect of collision parameters 3.2 External forces 2. Positron Emission Particle Tracking 4.1 Granular dynamics 2. Results 3.3 Inlet and outlet conditions 2. Model 2.3 Influence of lift forces 4. THE INFLUENCE OF COLLISION PROPERTIES ON THE FLOW STRUCTURE IN A RISER Abstract 1. Model 2. Conclusions Notation References 189 191 192 192 193 196 196 198 198 202 206 207 208 209 8. Conclusions Notation References 213 215 217 217 217 218 219 221 222 230 231 232 Publications Dankwoord (Acknowledgements) Levensloop 235 237 241 viii .3 Gas phase hydrodynamics 3.2 Axial effects 3. Comparison between PEPT data and simulation 5.1 Granular dynamics 2.Contents ______________________________________________________________________________________ 7. Results 6.2 External forces 2.
The study reported in this thesis is concerned with the granular dynamics of gassolid twophase flows. The linear spring/dashpot model also requires a spring stiffness to describe the particle interaction. Therefore the development and va lidation of fundamental hydrodynamic models is of utmost importance to gain more insight into the complex hydrodynamics. Two types of discrete particle models have been developed to be incorporated into the granular dynamics simulations. The effect of the collision parameters on the bed dynamics in a g fluidised bed with ashomogeneous inflow conditions was investigated. The gasphase hydrodynamics is described by the volume averaged NavierStokes equations for twophase flow. The gasphase flow is resolved on a length scale that is larger than the particle size. the coefficient of friction (µ ≥ 0) and the coefficient of tangential restitution (1 ≥ β 0 ≥ 0).Summary ________________________________________________________________________ Summary Gassolid twophase flows are encountered in a wide variety of industrial applications. In granular dynamics simulations the Newtonian equations of motion are solved for each individual particle in the system while taking into account the mutual interaction between particles and between particles and walls. binary collisions. The collision parameters (except for 1 . The complex hydrodynamics of these systems is still not fully understood which renders the scaleup of these units difficult. A sequence of collisions is processed one collision at a time. The first is a (2D and 3D) hardsphere model where the particles are assumed to interact through instantaneous. The second is a (2D) softsphere linear spring/dashpot model where contact forces between the particles are calculated from the overlap between the particles. The key collision parameters in both models are the coefficient of restitution (1 ≥ e ≥ 0). This softsphere model was chosen since it is the most frequently used model in the literature and hence it is best suited for a comparison with the hardsphere model. The softsphere model is capable of handling multiple particle interactions and can handle static situations in contrast to the hardsphere model.
µ > 0) bubbles did appear and the pressure fluctuations were larger. The case of bubble formation at a single central orifice was chosen for a comparison between the (2D) hardsphere model and the (2D) softsphere model. µ = 0) no bubbles were observed and pressure fluctuations inside the bed were rather small. The Root Mean Square (RMS) of the pressure fluctuations inside the bed showed an almost linear dependency on the energy dissipation rate by collisions during the simulation at low values of the energy dissipation rate. For higher values of the spring stiffness the influence of the value of the spring stiffness on the simulation results was negligible but the required CPU time increased significantly.Summary ________________________________________________________________________ β0 ) turned out to have a profound influence on the fluidisation behaviour. Incorporation of a (log. When the collisions were assumed to be fully elastic and perfectly smooth (e = 1. Hardly any difference could be observed between the results of the hardsphere simulations and the softsphere simulations indicating that the assumption of binary collisions in the hardsphere model is not limiting. A statistical analysis showed a rather wide spread in segregation profiles indicating the limited value of a single frame analysis. These trends were observed in 2D simulations as well as in 3D simulations. Preliminary softsphere simulations revealed that a minimum value of the spring stiffness was required to ensure stable simulations. With both types of models it was found that the assumption of fully elastic and perfectly smooth collisions resulted in much worse agreement between simulation and experiment: hardly any bubble could be observed. When mo re realistic values for the collision parameters were used (e < 1. When a wider size distribution was taken into account the pressure peaks inside the bed were smaller. A simulation assuming 2 .normal) particle size distribution improved the agreement between simulation and experiment. Segregation was successfully simulated with the (2D) hardsphere model for systems consisting of particles of equal size but different density as well as for systems consisting of particles of equal density but different size. In the latter case a clear steady state was not reached since the larger particles were continuously transported to the upper layers of the bed in bubble wakes. The influence of a (lognormal) particle size distribution on the bed dynamics was less pronounced than the influence of the collision parameters.
The (2D) hardsphere model was experimentally validated using the Positron Emission Particle Tracking (PEPT) facility at the University of Birmingham. A strong effect of the collision properties on the axial solids profile was found where a pronounced buildup of solids was observed in the simulation with realistic values for the collision parameters. The collision parameters required for the simulation were obtained from independent measurements at the Open University at Milton Keynes. When the collisions were assumed to be fully elastic and perfectly smooth the agreement was much worse. 3 .Summary ________________________________________________________________________ fully elastic and perfectly smooth collisions showed a rapid and almost perfect segregation due to the absence of bubbles in this system. Lift forces acting on the suspended particles turned out to have a slightly redispersive effect on the flow structure which made the radial segregation of the solids a little less pronounced. The results showed good agreement between experiment and simulation when the measured values for the collision parameters were used. Preliminary experimental validation showed rather poor agreement between simulation and experiment.000 particles were tracked during 45 seconds. In the case of fully elastic and perfectly smooth collisions hardly any clustering of particles could be observed as opposed to the case where these collision parameters were assigned realistic values. A quasi 2D. The simulation predicted segregation at a lower gas velocity than used in the experiment. Particlewall collisions turned out to have very little influence on the flow structure. In the simulation assuming fully elastic and perfectly smooth collisions no buildup of solids was observed. In the experiment the motion of a single tracer particle was tracked during one hour. The results of simulations of the riser section of a circulating fluidised bed proved to be very sensitive with respect to the collision parameters. This result is supported by experimental findings reported in the open literature. The PEPT data was timeaveraged to allow for a comparison with the results of a simulation where 15. gasfluidised bed with homogeneous inflow conditions was used for the validation.
Summary ________________________________________________________________________ 4 .
In het lineaireveer/smoorpotmodel is behalve deze parameters ook nog een veerkonstante vereist om de deeltjesinteractie te beschrijven. Dit zachte bollen model is gekozen omdat het het meest gebruikt is in de literatuur en daardoor het best geschikt is voor een vergelijk met het harde bollen model. de frictiecoëfficiënt (µ ≥ 0) en de tangentiele restitutiecoëfficiënt (1 ≥ β 0 ≥ 0). Het tweede model is een (2D) zachte bollen lineaireveer/smoorpot model waarbij de contactkrachten tussen de deeltjes worden berekend uit hun onderlinge overlap. De stroming van de gasfase wordt opgelost op een lengteschaal die groter is dan de grootte van een afzonderlijk vaste stof deeltje. De eerste is een (2D en 3D) harde bollen model waar de interactie tussen de deeltjes wordt ve rondersteld te verlopen via instantane. binaire botsingen. Het opschalen van dergelijke processen wordt bemoeilijkt door de complexiteit van de hydrodynamica van dergelijke systemen. Het onderzoek dat ten grondslag ligt aan dit proefschrift houdt zich bezig met de granulaire dynamica van gasvast tweefasenstromingen. De opeenvolgende botsingen worden hierbij één voor één afgehandeld in chronologische volgorde. Het is daarom van het grootste belang meer inzicht te krijgen in deze complexe hydrodynamica door het ontwikkelen en valideren van fundamentele hydrodynamische modellen.gemiddelde NavierStokes vergelijkingen voor tweefasenstroming. In granulaire dynamica simulaties worden de Newtonse bewegingsvergelijkingen opgelost voor elk individueel granulair deeltje in het systeem waarbij de interactie tussen deeltjes onderling alsmede de interactie tussen deeltjes en systeemwanden wordt verdisconteerd. Voor de granulaire dynamica simulaties zijn twee soorten discrete deeltjes modellen ontwikkeld. In tegenstelling tot het harde bollen model 5 . De hydrodynamica van de gasfase wordt beschreven door de volume. De belangrijkste botsingsparame ters in beide modellen zijn de restitutiecoëfficiënt (1 ≥ e ≥ 0).Samenvatting ________________________________________________________________________ Samenvatting Gasvast tweefasenstromingen vormen een belangrijk onderdeel van een grote verscheidenheid aan industriële processen.
Dit gaf aan dat de aanname van binaire botsingen in het harde bollen model niet beperkend is. µ = 0) werden geen bellen waargenomen en waren de drukfluctuaties in het bed klein. Met een bredere deeltjesgrootteverdeling werden de drukpieken in het bed iets lager. µ > 0) werden wel bellen waargenomen en waren de drukfluctuaties in het bed aanzienlijk groter. De botsingsparameters (met uitzondering van β 0 ) bleken een grote invloed te hebben op het fluïdisatiegedrag. Een vergelijk tussen het (2D) zachte bollen en het (2D) harde bollen model werd gemaakt aan de hand van simulaties van belvorming aan een centraal inspuitpunt.normale) deeltjesgrootteverdeling werd verdisconteerd verbeterde de overeenkomst tussen simulatie en experiment aanzienlijk. Als werd aangenomen dat de botsingen volledig elastisch en perfect glad verliepen (e = 1. Wanneer een (log. Deze trend werd zowel in 2D als in 3D simulaties waargenomen. 6 . The invloed van een (lognormale) deeltjesgrootteverdeling op het fluïdisatiegedrag was minder groot dan de invloed van de botsingsparameters. Met realistische waarden voor de botsingsparameters (e < 1. Er kon echter nauwelijks enig verschil worden waargenomen tussen de resultaten van het harde bollen model en die van het zachte bollen model. De gemiddelde kwadratische waarde van de drukfluctuaties vertoonde een zo goed als lineaire afhankelijkheid van de energiedissipatiesnelheid bij lage waarden van deze laatste. De invloed van de botsingsparameters op het stromingsgedrag van een gasgefluidiseerd bed met homogene instroomcondities is onderzocht. Verkennende berekeningen met het zachte bollen model lieten zien dat een minimum waarde van de veerkonstante vereist was om verzekerd te zijn van een stabiele simulatie.Samenvatting ________________________________________________________________________ is het zachte bollen model in staat om meervoudige deeltjesinteractie te verdisconteren en tevens is het geschikt om statische situaties te simuleren. Met beide typen modellen werd de overeenkoms t tussen simulatie en experiment veel slechter wanneer een simulatie werd uitgevoerd onder de aanname van volledig elastische en perfect gladde botsingen. Voor hogere waarden van de veerkonstante werd de invloed ervan op de resultaten van de simulaties verwaarloosbaar waarbij echter wel de benodigde rekentijd drastisch toenam.
gefluïdiseerd bed met homogene instroomcondities werd gebruikt voor de validatie. Het (2D) harde bollen model werd experimenteel gevalideerd met behulp van Positron Emission Particle Tracking (PEPT) experimenten uitgevoerd aan de universiteit van Birmingham. Een quasi 2D gas. Een simulatie waarbij de botsingen volledig elastisch en perfect glad werden verondersteld liet een zeer snelle en bijna volledige segregatie zien wat toegeschreven kon worden aan de afwezigheid van bellen in dit systeem. De simulatie voorspelde segregatie bij lagere gassnelheden. Clustervorming werd niet of nauwelijks waargenomen als de botsingen volledig elastisch en perfect glad werden verondersteld. In dit laatste geval werd geen stationaire toestand bereikt omdat de deeltjes continu omhoog w erden getransporteerd in het zog van bellen. Dit is in overeenstemming met experimenteel waargenomen trends gerapporteerd in de open literatuur. In de simulatie met realistische waarden voor de botsingsparameters werd een duidelijke opbouw van een axiaal vaste stof profiel geconstateerd in tegenstelling tot de simulatie met ideale botsingsparameters.Samenvatting ________________________________________________________________________ Het (2D) harde bollen model werd met succes ingezet bij het simuleren van segregatie in systemen die bestaan uit deeltjes van gelijke grootte maar verschillende dichtheid alsmede systemen bestaande uit deeltjes van gelijke dichtheid maar verschillende grootte. Met realistische waarden voor deze parameters kon clustervorming wel degelijk worden waargenomen. In het PEPT experiment werd de beweging van een tracerdeeltje gevolgd gedurende een uur. In simulaties van de risersectie van een circulerend gefluïdiseerd bed werd gevonden dat het stromingsgedrag sterk afhankelijk is van de botsingsparameters. De data van dit experiment werd tijdgemiddeld om een vergelijk mogelijk te maken met een simulatie waarin 15000 deeltjes werden gevolgd 7 . Een eerste experimentele validatie liet een matige overeenkomst zien tussen simulatie en experiment. De botsingsparameters voor deeltjeswand botsingen bleken nauwelijks invloed op het stromingsgedrag te hebben. Een statistische analyse liet een grote spreiding zien in het segregatieprofiel wat aangeeft dat een analyse op basis van een tijdsopname beperkte waarde heeft. Liftkrachten bleken slechts een klein dispersief effect te hebben wat er toe bij droeg dat de radiële segregatie van de vaste stof wat minder groot was.
De overeenstemming tussen het experiment en een simulatie waarin de botsingen volledig elastisch en perfect glad werden verondersteld was beduidend minder goed. 8 . De botsingsparameters benodigd voor de simulatie werden onafhankelijk gemeten aan de Open University in Milton Keynes. De resultaten van een simulatie waarin deze waarden werden gebruikt vertoonde goede overeenstemming met het experiment.Samenvatting ________________________________________________________________________ gedurende 45 seconden.
The hydrodynamics of gasfluidised beds is very complex and still not very well understood which renders the scaleup of these units difficult.General Introduction ______________________________________________________________________________________ Chapter 1. Finally the outline of this thesis is presented. These fundamental models can be classified into three categories based on the level of microscopy featured in the model. In this thesis the focus is on discrete particle models which form the intermediate category in the concept of multiscale modelling. GENERAL INTRODUCTION Abstract: In this chapter a brief introduction to fluidisation is presented. The position of discrete particle models within the multiscale modelling concept is explained and the objective of the work presented in this thesis is formulated. Therefore fundamental hydrodynamic models are required to gain more insight into the hydrodynamics of gasfluidised beds. 9 .
Chapter 1 ______________________________________________________________________________________ 10 .
General Introduction ______________________________________________________________________________________ 1. Figure 1.1. gasphase polymerisation of olefins and fluidised bed granulation (detergents. From left to right: the bubbling regime. the slug flow regime and the fast fluidisation regime. The three regimes that are featured in this thesis are presented in Figure 1. catalytic cracking of oil. In this figure the gas velocity increases from left to right. fertilisers) to name a few. excellent heat and mass transfer properties and the possibility of continuous operation.1.fluidised beds the gravity force acting on the solid particles is compensated by the drag forces exerted on the particles by the upward flowing gas. The minimum fluidisation velocity (umf) is defined as the superficial gas velocity at which the gravity force acting on the particles is just counterbalanced by the drag forces exerted on the particles by the gas phase. 11 . Fluidisation Gas. The three fluidisation regimes featured in this thesis. In gas. 1991) because of several advantageous properties including isothermal conditions throughout the bed. Typical applications cover a wide variety of physical and chemical processes such as fluidised bed combustion.fluidised beds are widely applied in the chemical process industry (Kunii and Levenspiel. When operated at gas velocities above umf several regimes are encountered.
When the gas velocity is increased further also the size of the bubbles increases. improved scale. and fast fluidisation regime. slug flow. the focus is on the bubbling.Chapter 1 ______________________________________________________________________________________ At velocities exceeding umf usually gas bubbles are present in the bed. The bubble size may approach the bed diameter in which case the slug flow regime prevails. Although gas.fluidised beds are widely applied. Such systems are commonly referred to as Circulating Fluidised Beds (CFB’s). the scaleup of these systems is still very complicated which is mainly due to the complex hydrodynamics of these systems. 2. Together with the development of dedicated experimental techniques the development of fundamental hydrodynamic models is of utmost importance to achieve a better understanding of fluidisation. When the gas velocity is increased beyond the terminal fall velocity of the particles the fast fluidisation regime is encountered. Eventually this will lead to the improvement of existing processes. Therefore it is of crucial importance to develop a thorough understanding of the hydrodynamics of gas. These bubbles have a decisive influence on the hydrodynamics of a bubbling fluidised bed and hence on its performance as a chemical reactor and/or a heat exchange unit.up and the design of more efficient future processes. In this regime the particles are entrained with the gas flow and transported upward through the riser. In the work presented in this thesis the development of such a fundamental hydrodynamic model is described and the results obtained with it are presented. In between the slug flow and the fast fluidisation regime the turbulent fluidisation regime can be distinguished and when the gas velocity is further increased starting in the fast fluidisation regime the pneumatic transport regime is encountered. Hydrodynamic modelling During the last few decades Computational Fluid Dynamics (CFD) has become a very powerful and versatile tool for the numerical analysis of transport phenomena (Kuipers 12 . At the top of the riser the particles are separated from the gas in a cyclone and fed back to the bottom of the riser. In this thesis however.fluidised beds.
. By incorporating microscopic information from sub scale models and passing on information to super scale models a multi. Three different classes of fundamental hydrodynamic models (learning models according to van Swaaij. the focus of the present study is on the hydrodynamics only. 1997.2. 2. Although attempts have been reported where the hydrodynamics were modelled combined with mass transfer and chemical reaction (Samuelsberg and Hjertager (1996).scale concept for fundamental hydrodynamic models of gas. CFD modelling of gasfluidised beds has proven to be successful and new developments in this area are promising.General Introduction ______________________________________________________________________________________ and van Swaaij. 1998).1 Multiscale modelling Due to the complexity of the hydrodynamics of multiphase flows it has become accepted that a single generalised CFD model cannot cover the wide variety of phenomena encountered in multiphase flows (Delnoij et al. 1997).fluidised beds as is schematically represented in Figure 1. With continuously increasing computer power combined with the development of improved physical models CFD has become a very useful tool for chemical engineers. specific models have to be developed that are tailor made to capture the relevant phenomena occurring at the length scale to which they are applied.scale modelling concept can be established.fluidised beds can be distinguished. 1985) of gas. Gao et al. Kuipers et al. The majority of studies on modelling of fluidised beds is concerned with the hydrodynamics only. 1999) the results of such attempts depend strongly on how well the hydrodynamics is modelled. Hence. These models can be combined together in a multi. Instead. Therefore the development of reliable hydrodynamic models is of utmost importance in order to arrive ultimately at models that are capable of predicting the performance of fluidised beds reactors. 13 . Kuipers and van Swaaij. If the flow structure is not well captured by the hydrodynamic model a sensible prediction of the reactor performance is impossible. (1998) demonstrated that the predicted performance of a riser reactor in terms of chemical conversion depends strongly on the prevailing flow structure in the riser.
However. 1996. the number of particles that can be taken into account in this technique is limited (< 106 ).Chapter 1 ______________________________________________________________________________________ In order to model a large (industrial) scale fluidised bed a continuum model. where the gas phase and the solids phase are regarded as interpenetrating continuous media. Therefore it is not yet possible. among many others). Multiscale modelling concept for fundament al hydrodynamic models of gasfluidised beds. These models require closure relations for the solids phase stress tensor and the fluidparticle drag where commonly empirical relations are used in the absence of more accurate closures.. even with modern day super computers. Also assumptions made 14 . In this EulerianLagrangian type of model a closure relation for the solids phase rheology is no longer required since the motion of the individual particles is solved directly. This EulerianEulerian type of model have been developed and successfully applied over the last two decades (Kuipers et al. In discrete particle models the Newtonian equations of motion are solved for each individual solid particle in the system. is the appropriate choice. 1992. 1994 among many others). Gidaspow. Improved closure relations for the solids phase stress tensor can be obtained by using the kinetic theory of granular flow (Sinclair and Jackson. 1989.. Nieuwland et al. to simulate a large (industrial) scale system. Continuum models Large (industrial) scale simulations Discrete particle models Particleparticle interaction closure laws Lattice Boltzmann models Fluid particle interaction closure laws Figure 1. this type of model can be used to arrive at improved closure equations for continuum models by employing techniques from statistical mechanics. However.2.
2 consists of three classes of models where more detail of the twophase flow is resolved going from continuum models to discrete particle models to Lattice Boltzmann models. discrete particle models still require closure relations for the fluidparticle drag since the gas flow is resolved on a length scale larger than the particle size. 15 . In the work presented in this thesis the focus is on the development and the use of discrete particle models. In Chapter 3 some additional techniques besides Lattice Boltzmann simulations are presented that can be used for the same purpose. This goes at the cost of increased computational requirements which necessitates a size reduction of the simulated system. as incorporated in advanced continuum models. Instead they can actually be obtained from the simulations. Since discrete particle models are very well suited to study the influence of particle properties on the hydrodynamics of gasfluidised beds this makes the m very useful models within the multiscale modelling concept. When the gas flow is resolved on a length scale smaller than the particle size these closure relations for fluidparticle drag are no longer required. In the absence of better closures empirical relations for the fluidparticle drag have to be used. It is important to realise that such simulations are limited to systems consisting of a number of particles that is significantly smaller (<103 ) than the number of particles that can be taken into account using discrete particle models (<106 ). verified and experimentally validated. In return the results of these simulations can be used to pass on information to models capable of simulating the flow on a larger scale. In short the multiscale concept as presented in Figure 1.General Introduction ______________________________________________________________________________________ within the framework of the kinetic theory of granular flow. The Lattice Boltzmann technique seems to be best suited for such simulations because it is very flexible in dealing with complex flow geometries. The model capable of simulating a larger system is fed with a closure relation obtained from a more microscopic simulation. However. can be verified. Before such a connection between separate scales can be established the individual simulation techniques must be well developed.
(1996) presented a hardsphere approach in their twodimensiona l discrete particle model of a gasfluidised bed. Ouyang and Li (1998) developed a slightly different version of this model. For the fluidparticle interaction empirical relations have to be used since the hydrodynamics of the gasphase is resolved on a length scale larger than the particle size.3. 16 . Hoomans et al. Kawaguchi et al. Discrete particle modelling In granular dynamics simulations of gas.Chapter 1 ______________________________________________________________________________________ 3. (1993) who developed a twodimensional softsphere discrete particle model of a gas. Mikami et al. (1998) extended this model to three dimensions as far as the motion of the particles is concerned. 1999) where energy balances and chemical reaction rates were taken into account. (1993) to include cohesive forces between the particles. The discrete particle approach for gasfluidised beds was pioneered by Tsuji et al. As far as particle interaction is concerned a multiscale modelling concept can be distinguished as is sche matically presented in Figure 1.fluidised bed based on the work of Cundall and Strack (1977).fluidised beds the Newtonian equations of motion are solved for each individual solid particle by using a discrete particle model.e. solids phase pressure). i viscosity.. (1998) extended the model originally developed by Tsuji et al. Recent developments in this area include the (2D) simulation of fluidised bed with internals (Rong et al. Xu and Yu (1997) presented a hybrid simulation technique that features elements from both hardsphere and softsphere techniques. Discrete particle models can also be used to verify assumptions made in the kinetic theory of granular flow which is used in most of the recently developed continuum models. As mentioned in the previous section continuum models require closure relations for the solids phase rheology (. These relations can be obtained from discrete particle models by employing techniques from statistical mechanics. 1999) and the (2D) simulation of gasphase olefin polymerisation (Kaneko et al..
In chapters 2 and 3 the theoretical framework of the granular dynamics simulations of gas. it is important to realise that these parameters can also be obtained from microscopic particle interaction models. The particle interaction parameters required for the discrete particle models can be obtained from experiments as was the case in the work presented in this thesis.3. However.General Introduction ______________________________________________________________________________________ Solids phase rheology Continuum models Particle interaction parameters Discrete particle models Material properties Contact theory Figure 1. In this way a multiscale concept for particleinteraction models arises. Multiscale modelling concept for models involving particle interaction. 4. 6.fluidised beds will be presented and in the chapters 4. 5. 17 . Outline of this thesis The objective of the work presented in this thesis is to study the influence of particle properties on the dynamics of gassolid twophase flows using discrete particle models. Transport properties for the solidsphase in a continuum model are obtained from discrete particle models and the particle interaction parameters required in discrete particle models are obtained from contact theory based on material properties. 7 and 8 several applications and experiment al validation of these models will be discussed. By using the appropriate contact theory the particle interaction parameters can be obtained using only material properties as input.
Alternative techniques will be discussed briefly. In Chapter 3 the gas phase dynamics is discussed for which the Eulerian approach was adopted. In Chapter 5 the hardsphere model is compared with the softsphere model. Furthermore the influence of the incorporation of a (log. Finally the results obtained with the 2D model are compared with the results obtained with the 3D model. The dependency of the bed dynamics on the collision parameters will be investigated. are presented in detail. Finally the external forces acting on a single (Lagrangian) particle in a gasfluidised bed are presented. Furthermore an experimental technique to measure the collision parameters required in both types of models is presented. The dependency of the results of the softsphere model on the value of the spring stiffness is investigated and the influence of the incorporation of a (lognormal) particle size distribution will be studied. Special attention is paid to the twoway coupling between the motion of the solid particles and the motion of the continuous gasphase. The formation of a single bubble at a central orifice will be used as a test case for the comparison. the hardsphere and the softsphere model. The volume averaged continuity and momentum conservation equations are presented together with the closure equations. The results of the hardsphere and softsphere model will be compared with each other and with experimental data. The two types of discrete particle models used in this work. In Chapter 4 the hardsphere model is applied to gasfluidised beds with homogenous inflow conditions. The hardsphere model has been developed both in 2D and in 3D whereas the softsphere linear spring/dashpot model has been developed in 2D only. Finally the influence of the collision parameters on the bubble formation process will be investigated.normal) particle size distribution on the bed dynamics is studied.Chapter 1 ______________________________________________________________________________________ Chapter 2 deals with granular dynamics. The two types of models are thoroughly discussed and a comparison will be presented. 18 .
perfectly smooth collisions and the results of the PEPT experiment. The model is modified to allow for a continuous transport of particles throughout the riser duct.000 particles are tracked during 45 seconds. 19 . For the latter system the influence of the collision parameters on the segregation is investigated. In Chapter 7 the hardsphere model is applied to riser flow. The influence of the collision parameters on the flow structure in the riser will be investigated. In the PEPT experiment the motion of a single radioactive tracer particle is tracked during one hour. Segregation will be demonstrated for each of these systems and a statistical analysis of the segregation dynamics is presented for the system consisting of particles of equal density but different size.. occupanc y plots and speed histograms.fluidised beds. 52. Simulations of systems consisting of particles of equal size but different density and systems consisting of particles of equal density but different size will be presented. The results will be compared with experimental data available in the open literature. The results of a simulation where independently measured collision parameters are used will be compared with the results of a simulation assuming fully elastic. Computational fluid dynamics applied to gas. E..General Introduction ______________________________________________________________________________________ Chapter 6 deals with segregation phenomena in bubbling gas. Finally an experimental validation is performed using a specially designed set. Both the hardsphere (2D) and the softsphere models are used. and van Swaaij W.up. 3623.M. The formation of clusters is studied and radial and axial solids profiles are obtained after time averaging of the data. Engng Sci. The time averaged data obtained from the experiment will be compared with the results of a simulation where 15.M. J. References Delnoij. Kuipers.A. The results will be compared on the basis of velocity maps.liquid contactors.P. (1997). Finally in Chapter 8 experimental validation of granular dynamics simulations of a gasfluidised bed with homogeneous inflow conditions is performed using the Positron Emission Particle Tracking (PEPT) facility at the University of Birmingham. Chem.
. C. Powder Technol.M. W. Briels. Shiojima.B.M. Xu. B. W. Adv. 47. and Horio. 1913. Kuipers. (1998).P.. 2nd edition. Lin. and van Swaaij.M. and van Swaaij. W. Boston. J.. Y. and van Swaaij. Tanaka. Hoomans.. Discrete particle simulation of bubble and slug formation in a twodimensional gasfluidised bed: a hardsphere approach.M. 45. Chem. 1095.. and Tsuji..P.A. D.. Comments on the paper “Numerical simulation of the gassolid flow in a fluidized bed by combining discrete particle method with computational fluid dynamics” by B. Chem. 20 . (1996).M. Chem. USA. O. S. Numerical simulation of twodimensional fluidized beds using the discrete element method (comparison between the two. Hoomans.A..M.M. 51. and Guo.. Engng Sci. Chem Eng. Kuipers. and van Swaaij. Advanced model for turbulent gassolid flow and reaction in FCC riser reactors.. B.P.. (1998).and threedimensional models). G. Yu.M.. 96. Kaneko. and van Swaaij. Chem Eng. No 3. 24.. Rev.B. Fluidization engineering. (1992). J. M. 54. 1. 5809. Tang.P. Kuipers. T.Chapter 1 ______________________________________________________________________________________ Gao. 13..M. J. (1999).H. W. Y. (1999)..A... Xu and A. Engng Sci. DEM simulation of fluidized beds for gasphase olefin polymerization. van Beckum. Engng Sci. Y. van Duin K. 227. AIChE J. 129.B.. Chem. Kuipers. W.P. Computational fluid dynamics applied to chemical reaction engineering.P.. T... (1998).P. 99.... F. 53. A numerical model of gasfluidized beds.P... and Levenspiel. Engng Sci. J.. T.A. J. (1997). Application of computational fluid dynamics to chemical reaction engineering.A.J.M. Kunii. W. W.J. ButterworthHeinemann. Kawaguchi. Kuipers.H. Briels. (1991). J.J. 2645.
Particle and bubble movements around tubes immersed in fluidized beds – a numerical study. 2785.M. 2077. in Proceedings of the 5th Int. Chem. Samuelsberg..motionresolved discrete model for simulating gassolid fluidization.. Engng Sci. Chemical reactors.. Engng Sci. Engng Sci. 53. Numerical simulation of the gassolid flow in a fluidized bed by combining discrete particle method with computational fluid dynamics.. and Harrison. Kawaguchi. Ouyang. Chem. May 28. D. Chem. and Li. T. and Hjertager. in Fluidisation. Mikami. Conf. Engng Sci.P. (1998). Y. H. 595. and Horio. (1985). (1999). (Eds). 52. A.. Davidson. 21 . J...June 1.. Rong. and Horio. Computational fluid dynamic simulation of an oxychlorination reaction in a fullscale fluidized bed reactor.. 1996. A. Xu. T. 77.. and Yu. T. W. R. T. J. Academic Press. Chem. (1999).F.H.. UK. Clift. D.E. B.. Particle. Powder Technol. M. Kamiya. van Swaaij.H. (1993). Numerical simulation of cohesive powder behavior in a fluidized bed. 54. and Tanaka.. (1996). Tsuji.B. Beijing (China). J. Circulating Fluidized beds. 54.General Introduction ______________________________________________________________________________________ Mikami. 1927. 79. M. 2nd edition p. 5737. London. Discrete particle simulation of two dimensional fluidized bed. (1997). B.
Chapter 1 ______________________________________________________________________________________ 22 .
The first is a hardsphere model where the particles are assumed to interact through instantaneous. binary collisions. The key parameters in these models to describe a collision are the coefficient of restitution (e). These three types of simulation are discussed with emphasis on their application to simulation of gassolid twophase flow. The approaches can be divided into three groups: hardparticle techniques. The softsphere model is capable of simulating static situations unlike the hardsphere technique. Finally the external forces acting on the particles used in the simulations of gasfluidised beds in this work are presented. GRANULAR DYNAMICS Abstract: In this chapter a review is presented of the various granular dynamics simulation techniques available in the literature. It is shown that the spring stiffness of the tangential spring should not be taken equal to the stiffness of the normal spring in order to avoid unrealistic behaviour. Both models are described in detail. A comparison between the hardsphere and the softsphere model is presented. The second is a softsphere linear spring/dashpot model where contact forces between the particles are calculated from the overlap between the particles. the coefficient of friction (µ) and the coefficient of tangential restitution (β0 ). Experiments are described that enable measurement of the three collision parameters by careful observation of single impacts. softparticle techniques and Monte Carlo techniques. The effect of these parameters on a single collision is demonstrated. A sequence of collisions is processed one collision at a time. 23 . However in the softsphere model special care must be taken in the choice of time step and the spring stiffness required for the calculation of the repulsive force. In this work two types of models were developed.Granular Dynamics ______________________________________________________________________________________ Chapter 2.
24 . Engng Sci. Briels and W.M. van Swaaij.J. 51. Chem. 99.Chapter 2 ______________________________________________________________________________________ Parts of this chapter are based on the paper: B. Kuipers.. (1996).P. Hoomans.A. Discrete particle simulation of bubble and slug formation in twodimensional gasfluidised beds: a hardsphere approach.B.M. J. W.P.
Before the two approaches adopted in this work (hardsphere and softsphere linearspring/dashpot) will be described in detail in the preceding paragraphs a short review of the different approaches that are available in the literature will be presented. In this work the focus will be on the granular dynamics of gassolid twophase flow which is a field of research that has gained an increasing amount of attention over the last decade. that we are just beginning to understand (Jaeger et al. These approaches can be divided into three types of simulations: hardparticle approaches.Granular Dynamics ______________________________________________________________________________________ 1. 1999). Introduction Measured by tons granular matter is after water the most manipulated material in the world (de Gennes. 1999). 1994). Such systems can be studied by means of Stokesian dynamics simulations (Ichiki and Hayakawa. 1995). 1996). However. segregation. Phenomena like heap fo rmation and arching provide both theoreticians and experimentalists with challenging problems concerning the statics of granular matter (de Gennes. softparticle approaches 25 . Therefore it is not surprising that Granular Dynamics (GD) has attracted the interest of a large number and a wide variety of researchers over the last decades.. This holds for gassolid flow with rather coarse particles which is the subject of study in this work. The modelling approaches adopted in Granular Dynamics can be roughly divided into two groups: soft particle and hard particle approaches. The interest not only originates from industrial needs but there is also an increasing interest in granular media from a more fundamental perspective. In the case of liquidsolid fluidised beds lubrication forces have to be taken into account (Schwarzer. clustering and inelastic collapse.. roughly 40% of the capacity of industrial plants is wasted due to problems related to transport of granular matter (Knowlton et al. However the greatest challenge is provided by the dynamics of this material which unveils a wealth of phenomena. The systems considered in this work are all operated in the grain inertia regime according to Bagnold (1954) which implies that particleparticle and particlewall interactions are dominated by inertia rather than viscous forces. 1995) and for liquidsolid systems with smaller particles (< 100 µm) direct particle interaction does not even occur. such as standing waves in vibrated beds.
In granular dynamics simulations of gassolid twophase flow there is a constant stream of energy supplied to the particles through gravity and the drag force exerted on the particles by the gasphase. 1959) the technique was presented in more detail including a way to deal with a square well interaction potential apart from merely hardsphere interaction.Chapter 2 ______________________________________________________________________________________ and MonteCarlo approaches. In time driven simulations a constant time step is used to progress through the dynamics of the system. For this purpose a list of future collisions is compiled and updated when necessary. In event driven simulations the progression depends on the number of collisions that occurs. This can for instance be achieved by applying a shear rate through a proper choice of boundary conditions (Campbell. In a later paper (Alder and Wainwright. 1993). binary collisions. A sequence of collisions is processed one collision at a time in order of occurrence. 26 . Softparticle simulations can be typified as time driven because the interaction times are large compared to free flight times.1 HardParticle Approaches The hardsphere simulation technique was first presented by Alder and Wainwright (1957) in order to study phase transitions in molecular systems by means of numerical simulations. For the remainder of this paragraph the focus will be on the various techniques used in granular dynamics simulations of gassolid twophase flow. In hardsphere simulations the particles are assumed to interact through instantaneous. 1985). Hence energy has to be continuously supplied to a granular system in order to keep the particles in motion. 1. For a comprehensive introduction to this type of simulation the reader is referred to Allen and Tildesley (1990). The dissipative particle interaction in granular media makes these systems significantly different from molecular systems where energy is always conserved. A lot of effort has been put into the further optimisation and development of this event driven type of simulation technique (Marin et al. 1985). Over two decades after the publications by Alder and Wainwright the hardsphere approach was discovered as a useful tool for granula r dynamics simulations (Campbell.. Hardparticle simulations can be typified as event driven because the interaction times are small compared to free flight times.
(1993) in their simulations of gassolid flow in a vertical channel. However one should be careful when applying this technique since the modified Nanbu method used by Tsuji et al. The occurrence of a collision as well as the geometry of a collision ( i. the collision coordinate system) are determined by a random number generator where the probability of a collision depends on the local solids fraction (the greater the solid fraction the greater the probability of a collision). ε Sommerfeld (1990) and Frank et al. The main problem is the large number of particles and hence the large number of collisions that have to be detected and processed. For denser flows the particleparticle interaction can no longer be neglected. A more detailed description of this technique was presented by Tsuji et al. They neglected the particleparticle interaction but instead focussed on particlewall interaction where an irregular bouncing model was used. (1998). The particles tracked in this simulation technique are not actual particles but ‘sample’ particles that represent several ‘real’ particles. 1997). This technique was employed by Yonemura et al. A technique that takes particleparticle interaction into account without detecting and processing every single collision that occurs in the system is the Direct Simulation Monte Carlo (DSMC) technique that was originally developed by Bird (1976) for molecular simulations. (1992) presented similar approaches where also attention was paid to experimental validation. The DSMC method can therefore be regarded as a coarse grained model of the actual dynamics. (1987) presented a hardsphere granular dynamics model for dilute gassolid flow in a horizontal channel. The problem in these simulations is not so much the description of the particleparticle interaction itself since accurate collision models are available. 27 .01).Granular Dynamics ______________________________________________________________________________________ Tsuji et al. Oesterlé and Petitjean (1993) presented a technique that resembles the DSMC technique strongly but differs in the sense that the particles in their simulations are ‘real’ particles instead of ‘sample’ particles.e. (1998) does not guarantee exact conservation of energy in the absence of dissipative terms (Frezzoti. This can be justified since the solids fraction in their system was very low ( s < 0.
Hoomans et al.e. Although they did not apply it directly to gassolid flow the technique is worthwhile mentioning here since Ge and Li (1997) and Ouyang and Li (1999) adopted this approach in their simulations of gas. polygonal) particles. This was the first time that this technique was developed and applied with success to such a dense system as a bubbling gas. After each time step the whole system is scanned for overlap between particles and when this is detected the particles involved are assumed to have collided during that time step and subsequently their new velocities are calculated. 1. They used a 28 . Müller and Liebling (1995) presented a triangulation technique that enables a sequence of collisions between polygonal particles to be correctly processed. In the following chapter this technique will be explained in detail. Lun and Liu (1997) presented a threedimensional hardsphere model for a more dilute gassolid flow where they used a granular dynamics model presented earlier by Lun (1996). In this computational strategy the particles are assumed to interact as hardspheres but the motion update is performed at constant time steps.fluidised bed. In the case of larger time steps it may be possible that collisions are overlooked.Chapter 2 ______________________________________________________________________________________ Hopkins et al. (1996) presented a true hardsphere type of simulation technique for gasfluidised beds. In this twodimensional model a sequence of collisions is processed proceeding from one collision to the next by using a collision list that is compiled and updated in a highly efficient manner.fluidised beds. In addition to the model of Hogue and Newland. Hogue and Newland (1994) used a similar strategy in their simulations where a technique was used that was especially designed to handle nonspherical (i.2 SoftParticle Approaches The Distinct Element Method (DEM) developed by Cundall and Strack (1979) was the first granular dynamics simulation technique published in the open literature. This technique works fine for relatively dilute systems when the time step is chosen relatively small. For denser systems overlap may be found between more than two particles which gives rise to incorrect dynamics. (1991) developed a type of hardparticle simulation technique where collisions were detected at constant time intervals.
two alternatives to the Cundall and Strack model are briefly outlined here. 1995) they presented a threedimensional version of this model. Softparticle approaches differ in the choice of force scheme used to calculate the interparticle forces. The spring constant for the compression phase (loading) of the constant is taken to be lower than the constant used for the restitution phase (unloading). When two particles that do not touch each other are positioned at the same height within the cutoff distance of the potential they will still experience a repulsive force. 29 . Walton and Braun (1986) presented a model that uses two different spring constants to model energy dissipation instead of a dashpot. This method allowed for multiple particle overlap although the net contact force was obtained from pairwise interactions. (1994) presented a force scheme that was based on a continuous potential of an exponential form containing two unknown parameters. This is not physically correct behaviour for dry granular material. A review of various popular schemes for repulsive interparticle forces is presented by Schäfer et al. Langston et al. One should be careful when applying this model since it features rather unrealistic behaviour when using a cutoff distance greater than the particle diameter. The interaction constant was chosen in such a way that the net force would be zero when a particle would rest on top of another particle in a gravity field.Granular Dynamics ______________________________________________________________________________________ twodimensiona l soft particle model where the particles were allowed to overlap slightly. In a later paper (Langston et al. The repulsive force was obtained by taking the gradient of this potential while the stiffness parameter was chosen in such a way that particle overlap could not become too high without requiring too small a time step. Before the various softparticle approaches used in fluidised bed simulations will be discussed.. The coefficie nt of restitution can be related to the ratio of the two spring constants. The contact forces were subsequently calculated from the deformation history of the contact using a linear spring/dashpot model. (1998) in their studies of granular mixing in a rotating container. (1996). This model was used by McCarthy and Ottino (1998) and Wightman et al. the stiffness of the interaction and an interaction constant.
Xu and Yu (1997) presented a twodimensional model of a gas. Also a Gaussian particle size distribution was taken into account in these simulations although the effect of the size distribution on the flow behaviour was not investigated. In such a system the motion of the gasphase in the third dimension can be neglected. (1993). The spring constants used in their simulations were much higher than the ones used by Tsuji et al. (1993). in this model the gasphase dynamics was represented by a onedimensional model with hardly any dynamic features. (1993). (1992) presented a threedimensional model (as far as the particles were concerned) for gassolid flow in a horizontal pipe. Apart from the interparticle forces. In that model the particle motion was resolved in full 3D whereas the gasphase dynamics was still calculated in 2D. The particles used in their simulations are still Geldart D particles but due to the liquid bridge forces the fluidisation behaviour resembles the behaviour of Geldart C particles strongly. for which the Cundall and Strack model was used. A threedimensional version of their fluidised bed model was presented by Kawaguchi et al. This can be justified because the system used in their simulations was a rather flat (quasi 2D) fluidised bed. Previously Tsuji et al. Unfortunately no results were reported that could show the importance of the detection algorithm for the overall simulation results. However. (1998). (1998) extended the model developed by Tsuji et al.fluidised beds by Tsuji et al.Chapter 2 ______________________________________________________________________________________ The softsphere approach was first applied to gas.fluidised beds but play an important role in liquidfluidised beds. 30 . In their twodimensional model the particle interaction was calculated using a softsphere interaction model similar to the one presented by Cundall and Strack (1979). However in their simulations a collision detection algorithm that is normally found in hardsphere simulations was used to determine the first instant of contact precisely. Mikami et al. Schwarzer (1995) presented a twodimensional model of a liquidfluidised bed. (1993) by incorporating liquid bridge forces to simulate cohesive particle fluidisation.fluidised bed that was based on the model developed by Tsuji et al. also lubrication forces were taken into account. These lubrication forces can be neglected for gas.
equilibrium) conditions and for that purpose it has certain advantages over dynamic simulations. If the system energy has increased the new configuration is accepted with a probability obtained from Boltzmann distribution based on the change in energy.000 particles were used. (1986). This method has been applied to granular systems by Rosato et al. Although the results compared rather well with experimental data no statement was made about the time scale over which the phenomena occurred. Seibert and Burns (1998) were able to predict segregation phenomena in liquidfluidised beds using an extended version of their previously developed model (Seibert and Burns. In their Monte Carlo simulations a new overlap. 31 .Granular Dynamics ______________________________________________________________________________________ Mikami (1998) was also the first to present full 3D simulations (also with respect to the gasphase) where no less than 500. In their model they used a net force (calculated from gravity and fluid drag) to calculate the change of energy involved with a particle movement. 1996).3 Monte Carlo techniques Another popular method to study many particle systems is the Monte Carlo technique (Frenkel and Smit 1996). 1. Therefore this method is not capable of simulating the dynamics of a granular system without the input of apriory knowledge. Using this method Rosato et al.e. It is important to realise that time is not a variable in Monte Carlo simulations.free particle configuration is generated at each step. The Monte Carlo technique is capable of predicting steady state (i. were able to simulate segregation phenomena in shaken or vibrated systems where effects due to an interstitial fluid were neglected. This is of course to be expected for this type of simulation technique. The change in the system energy is then calculated and if this change is negative the new configuration is accepted. A Monte Carlo step can only be linked to an actual time step by means of calibration but this is not a straightforward task.
The original 2D collision model was mapped after the model presented by Wang and Mason (1992). homogeneous spheres and the interaction forces are impulsive. The point of origin is the contact point. Since the hardsphere model was developed in 3D as well as in 2D the collision model will be presented in vector notation. In between collisions the particles are in free flight.Chapter 2 ______________________________________________________________________________________ 2. the spheres with radii Ra and Rb and masses ma and mb have translation velocity vectors va and vb and 32 . The coordinate systems used in our model are defined in Figure 2. Consider the two colliding spheres a and b in Figure 2.instantaneous collisions where contact occurs at a point. ra − rb n= (2. The particles are perfect. (1994) since that is more widely accepted (see for example Lun. For the 2D version the zcomponent of the position and velocity vectors are zero and only rotation about the zaxis is considered. In this work however we will mainly adopt the notation used by Foerster et al.1 Collision model In the collision model it is assumed that the interaction forces are impulsive and therefore all other finite forces are negligible during collision. HardSphere approach In the hard sphere model used in this work the particles are assumed to interact through binary. The normal unit vector can now be defined: ra − rb . quasi.1) Hence the normal unit vector points in the direction from the centre of particle b to the centre of particle a.1. 1997 and Tsuji et al. 2. First the collision model will be presented and then the computational strategy and some optimisation techniques will be described.1 with position vectors ra and rb . 1998). Prior to collision.
0 ) = − n × J Ra Rb I= 2 mR 2 . 5 (2.0 ) = Rb n × (− J ) m a ( v a − v a.0 ) = −(R a n ) × J I b ( ωb − ωb.Granular Dynamics ______________________________________________________________________________________ rotational velocity vectors ωa and ωb (clockwise rotation is negative by definition).6) (2.0 ) = − J I a ( ω a − ωa.1.4) (2. Velocities priortocollision are indicated by the subscript 0 .5) (2.2) (2. For a binary collision of these spheres the following equations can be derived by applying Newton’s second and third laws: m a ( v a − v a. y n a b t x z Figure 2.3) (2. Definition of the coordinate systems.8) The impulse vector J is defined as follows: 33 .0 ) = −m b ( v b − v b.7) Ia I ( ωa − ω a.0 ) = J (2.0 ) = J m b ( v b − v b.0 ) = b ( ωb − ωb.
The only input parameters necessary in these calculations are material properties although assumptions have to be made about the deformation behaviour (elastic/plastic) of the material. v ab = ( v a − v b ) − ( R a ω a + Rb ω b ) × n .12) v ab = (v a − ω a × Ra n) − (v b + ω b × Rb n ) . From this relative velocity. From equations 2. Walton (1992) used two types of finite element codes (DYNA2D and NIKE2D) to simulate the collision process in detail on a subparticle level. Thornton (1997) demonstrated that based on a simplified theoretical model for the normal interaction between elasticperfectly plastic spheres an analytical solution could be obtained for the rebound velocity.9) where t c stands for the contact time (i. the duration of the contact). If the force Fab in equation 2.6 and 2. c − v b. Before the se constitutive relations will be introduced first the relative velocity at the contact point (vab ) has to be defined: v ab ≡ ( v a.e. In simulations of gasfluidised beds a large number of collisions (typically 106 109 ) have to be processed and therefore the actual physics of a binary collision has to be simplified to some extent and constitutive relations have to be introduced.9 were known as a function of all the parameters involved.Chapter 2 ______________________________________________________________________________________ t= tc J= t =0 ∫ Fab dt .10) (2. the impulse J could be calculated directly. (2.c ) .7 it is clear that the postcollision velocities of both particles can be calculated when the impulse vector J is known. the tangential unit vector can be obtained since the normal unit vector is already defined in equation 2.11) (2.1: 34 . (2.
0 = B1 J − (B1 − B 2 )n( J ⋅ n ) . (µ ≥ 0 ) : n × J = −µ (n ⋅ J ) .18) 35 .13) Equations 2.12 to obtain: v ab − v ab. 0 ⋅ n) v ab.7 can now be rearranged using ( n × J ) × n = J − n(J ⋅ n ) and equation 2. The first parameter i the s coefficient of (normal) restitution. (0 ≤ e ≤ 1) : v ab ⋅ n = −e (v ab.17) For nonspherical particles this definition can lead to energy inconsistencies (Stronge.6 and 2. (2. ma mb (2.15) 1 1 + . (2. 0 − n(v ab. The second parameter is the coefficient of (dynamic) friction.16) At this point constitutive relations are required to close the set of equations. 0 ⋅ n) . Through these constitutive relations three parameters enter the model. (0 ≤ β 0 ≤ 1) : (2. The third parameter is the coefficient of tangential restitution. 0 ⋅ n) .Granular Dynamics ______________________________________________________________________________________ t= v ab. (2. 0 − n(v ab. 1990) but for spherical particles this definition holds.14) where B1 = and B2 = 7 1 1 m + m 2 a b (2.
20) For the tangential component two types of collisions can be distinguished that are called sticking and sliding. (2.Chapter 2 ______________________________________________________________________________________ n × v ab = − β 0 (n × v ab. The only exception is made for the coefficient of normal restitution where collisions occurring at a normal impact velocity less than a threshold value ‘MINC0’ (typically 104 m/s) are assumed to be perfectly elastic (e = 1. Although it is accepted that these coefficients depend on particle size and impact velocity this is not taken into account in this model. Combining equations 2.21) (1 + β 0 )v ab.0 ) .17 yields the following expression for the normal component of the impulse vector: v ab.22) 36 . If the tangential component of the relative velocity is sufficiently high in comparison to the coefficients of friction and tangential restitution that gross sliding occurs throughout the whole duration of the contact. When β 0 is equal to zero the tangential component of the relative velocity becomes zero during a sticking collision.14 and 2.0 ⋅ t J n B1 sticking (2.0). reversal of the tangential component of the relative velocity will occur.19) Notice that this relation does not affect the components parallel to n and that the components orthogonal to n are related by a factor –β 0 .0 ⋅ t J n B1 sliding (2. The criterion to determine the type of collision is as follows: µ< µ≥ (1 + β 0 )v ab. the collision is of the sliding type.0 ⋅ n B2 J n = −(1 + e ) (2. When β 0 is greater than zero in such a collision. The nonsliding collisions are of the sticking type.
It is possible to implement a moving/rotating wall through the velocity vectors vb and ωb but in the simulations performed for this work these velocities are all set equal to zero.26) E dsp.27) 37 .0 ⋅ t B1 . (2.6 and 2.Granular Dynamics ______________________________________________________________________________________ For collisions of the sticking type.0 B1 = −(1 + β 0 ) v ab. t dJ t .n = 2 v ab. (2.7. the tangential impulse is given by: J t = −(1 + β 0 ) n × v ab.0 2B2 (1 − e ) . the tangential impulse is given by: J t = −µ J n . 2 (2. The energy dissipated by the normal component in a collision is: (2.25) The postcollision velocities can now be calculated from equations 2.n. tot = ∫ v ab. n dJ n + ∫ v ab.e. the wall) is infinitely large which makes all terms 1/mb equal to zero.24) The total impulse vector is then simply obtained by addition: J = J nn+ Jtt . The energy dissipated during a collision can be obtained by solving the following integral over the duration of the collision: E dsp. (2. In particlewall collisions the mass of particle b (i.23) For collisions of the sliding type.
2 the effect of the coefficient of restitution is illustrated.2 Key parameters of the collision model Since the three key parameters of the collision model are of crucial importance for the remainder of this work the effect of each of the three will now be highlighted.30) 2.0 ⋅ t − µB1J n .28) and if the collision is of the sliding type. If the collision is of the sticking kind. (2.t.t = − µJ n v ab.t . the dissipated energy is: 1 E dsp. 38 . When a particle collides perfectly elastically.0 2 B1 (1 − β ) .29) The total amount of energy dissipated in a collision is then obtained by adding the tangential and normal contributions: E dsp. which is always the case for granular material. 2 0 (2. In Figure 2. For all these examples the particles do not experience any friction of the gas phase.t = 2 v ab.Chapter 2 ______________________________________________________________________________________ For the energy dissipated by the tangential component the two types of collision have to be distinguished aga in.tot = E dsp. the dissipated energy is: E dsp. No energy is dissipated in this process and the particle will eternally continue to bounce.n + E dsp. energy is dissipated in the collision and the particle will not bounce back to the same height as it was initially released from. without any energy dissipation (e = 1) with a horizontal flat wall it will bounce back to the same height as it was initially released from. Eventually the particle will come to rest on the wall. When e < 1. 2 (2. The system considered here is a particle that collides with a flat wall under the influence of gravity (g).
No energy is dissipated in this case. In the case of a perfectly smooth particle (µ = 0) this rotation does not affect the translation motion of the particle after collision.Granular Dynamics ______________________________________________________________________________________ g e=1 e<1 Figure 2.2.3. the rotation does affect both the translation and the rotation after collision as illustrated. For sticking collisions this effect is much more pronounced than for sliding collisions. The effect of the coefficient of friction (µ). g µ=0 µ>0 Figure 2. Consider the same case as before but now the particle initially rotates as well (β 0 = 0). The effect of the coefficient of friction (µ) is demonstrated in Figure 2. In this case energy is dissipated during the collision. The effect of the coefficient of restitution (e). 39 . which is always the case for granular material.3. When the particle is not perfectly smooth (µ > 0).
when β 0 = 1 the tangential component of the relative velocity at the contact point reverses and the particle bounces back in the opposite direction. No energy is dissipated in this case. Within this time step DT the velocities are assumed to change only due to collisions and a sequence of collisions is processed one collision at a time like in a regular hardsphere simulation in Molecular Dynamics. In the case of β 0 = 0 the result is identical to the case described above provided the collision is of the sticking type. However.Chapter 2 ______________________________________________________________________________________ The effect of the coefficient of tangential restitution is demonstrated in Figure 2. 2.4. The effect of the coefficient of tangential restitution (β0 ). The tangential component of the relative velocity at the contact point is equal to zero at the end of the collision and energy is dissipated. To do so it is necessary to 40 . The case is identical to the one above for the coefficient of friction except that now the coefficient of friction is high enough to ensure that the collision is of the sticking type.4. In section 5 experiments are described that enable determination of these collision parameters from careful observation of single impacts. So in fact a separate MD hardsphere simulation is performed within each time step.3 Sequence of collisions In the hardsphere model a constant time step DT is used to take the external forces acting on the particles into account. g β0 = 1 β0 = 0 Figure 2.
12). The collision time t ab of a pair of particles (a. 41 . This yields a quadratic equation in t ab the smallest solution of which corresponds to collision (Allen and Tildesley. When particles a and b collide the distance between the two centres of mass is equal to the sum of the two radii as is shown in Figure 2.5. (2. Ra + Rb va tab vb tab a rab b Figure 2. It can be calculated from the initial positions and velocities of both particles.31) where rab ≡ ra − rb and v ab ≡ v a − v b (in this definition of vab the particle rotation is not taken into account unlike in equation 2. 1990): t ab = − rab ⋅ v ab − 2 2 (rab ⋅ v ab )2 − v ab (rab − ( Ra + R b ) 2 ) 2 v ab .Granular Dynamics ______________________________________________________________________________________ determine what pair of particles will collide first which requires the determination of the collision times of all relevant collision pairs. Note that if rab ⋅ v ab > 0 the particles are moving away from each other and will not collide. Determination of the collision time t ab .b) is defined as the time remaining until these particles will collide.5.
wall = ( x wall + Ra ) − r x .Chapter 2 ______________________________________________________________________________________ In case of a collision with a wall the collision time follows simply from the distance to the wall and the normal velocity component toward that wall which leads for a vertical wall to the following expression: t a. set up collision list locate minimum collision time tab increment acctim by t ab no acctim < DT ? yes move (tab ) collision dynamics reset collision lists locate minimum collision time tab increment acctim by t ab move (DT(acctimtab )) Figure 2. 42 .6. (2.6. Computational strategy of a hardsphere simulation within a time step DT.32) The algorithm used to process a sequence of collisions within a constant time step DT is presented in Figure 2. a v x.a .
In Figure 2. Subsequently the routine reset collision lists is entered where new collision times and partners have to be found all the particles involved in the collision. In the routine move(tab ) the collision times of all particles are reduced with t ab and the particle positions are updated using a first order explicit integration: ra (t + t ab ) = ra ( t ) + v a t ab . The neighbour lists are updated at each time step dt nblist. During this motion no collision occurs.Granular Dynamics ______________________________________________________________________________________ First the collision lists are initialised in which for each particle a collision partner and a corresponding collision time are stored. Finally a new collision pair has to be detected and acctim can be incremented with the new collision time t ab . The first step to achieve this is to minimise the number of particles that have to be scanned for a possible collision by employing a neighbour list. The variable acctim (accumulated time) keeps track of the time spent since the beginning of the time step.7 particle a is coloured black and its neighbour particles are shaded. 2. This does not only effect the particles a and b but also the particles that were about to collide with either a or b. After the loop is finished the particles have to be moved forward until acctim equals DT. the loop is finished.4 Optimisation To perform simulations of relatively large systems for relatively long times it is essential to optimise the hardsphere computational strategy.33) The calculation of the collision dynamics involves the collision model presented in the previous paragraph. As soon as a minimum collision time is found that after addition to acctim is greater than the time step DT. In the neighbour list of particle a all the particles that are found within a square of size Dnblist with particle a located at the centre. For each particle the smallest collision time is determined by scanning all relevant collision partners. are stored. (2. When looking for a collision partner for particle a only the particles in the neighbour list need to be scanned. 43 .
for each cell the particles whose centre can be found in that cell are stored in a list. In the original version of the code all particles were moved to their new positions before each collision. With the implementation of the neighbour lists substantial speed up has been achieved and simulations with over 1. 44 .6) is no longer the most CPU time consuming routine in the algorithm.6) takes up 50 % of the CPU time in a typ ical fluidised bed simulation. This can be optimised by applying a more efficient motion update strategy. Instead the update of the particle motion (move in Figure 2. The solution of the gas flow field (Chapter 3) requires the computational domain to be divided into cells. This implies Nparticles*Ncollisions motion updates per time step which means that particles that do not collide are moved over a straight line in far too many steps as illustrated in Figure 2.Chapter 2 ______________________________________________________________________________________ Dnblist Figure 2. only the cell where the particle’s centre is found and the three nearest adjacent cells are scanned for possible neighbours. When updating the neighbour list it is still not necessary to scan all particles. The search for collision partners (set up collision lists and reset collision lists in Figure 2. When updating the neighbour list.8.000 particles are easily possible. The neighbour list principle: all shaded particles are stored in the neighbour list of the black particle.7.
Since the particles in Figure 2.6) is the main CPU time consumer in the hardsphere routine. In the new algorithm special care must be taken when looking for new collision partners for particles that just collided. In the previous version the particles were moved to their new position in a total of seven steps for this specific example. This search has to be performed after each collision and in the old version all particles were scanned and the smallest collision time was stored. Efficient motion update.8.Granular Dynamics ______________________________________________________________________________________ Figure 2.9. This does not only slow down the algorithm but also gives rise to numerical errors. 45 .e. since the positions stored in memory for the particles not involved in that collision are not their actual positions. After optimising the motion update strategy the search for the smallest collision time (locate minimum collision time in Figure 2.6 went down from 50% to less than 1% of the total amount of CPU time.8 collide only once. that is five motion updates too many. two motion updates will suffice for both particles. In typical fluidised bed simulations the total number of collisions per time step is of the same order as the total number of particles (i. This figure is rather idealised for clarity but on average a typical grid cell can contain up to 100 particles. In the new strategy advantage is taken of the fact that the computational domain is already divided into cells as illustrated in Figure 2. 104 105 ) indicating that the total number of unnecessary motion updates is of the same order as the total number of particles. This causes some overhead for the new algorithm but nonetheless the speed gain is substantial since the routine move(t ab ) in Figure 2.
340 cells vs. Efficient search for smallest collision times using grid cells. For instance in bubbling beds (Chapters 4 and 5) a relatively small neighbour list (Dnblist = 3 Dp ) can be used that is updated every second time step.Chapter 2 ______________________________________________________________________________________ Figure 2. 2.9. 40. After implementation of these optimised routines the main CPU time consumer is again the search for possible collisions even though this search is performed only within the neighbour list! A suitable choice of the size of the neighbour list and the time step for updating the neighbour list is now critical. Of course a new smallest collision time has to be found in the grid cells containing particles that were involved in the last collision but this causes negligible overhead. When for each grid cell the smallest collision time is stored it is not necessary to scan all particles after each collision. If a neighbour list is chosen to be too small it is possible that a collision is not detected and overlap between 46 .e. These choices however depend on the sort of system that is simulated. Instead all the grid cells are scanned and since the total number of grid cells is at least one order of magnitude smaller than the total number of particles ( . For risers (Chapter 7) however a larger neighbour list has to be used (Dnblist = 8 Dp ) since the velocity differences between the particles are larger as well.000 particles for the bubble formation simulations in i Chapter 5) this is much faster.
This cannot be tolerated in hardsphere simulations and if such an overlap is detected in our code the simulation is stopped immediately. The softsphere model was implemented in 2D. (1998) and Mikami et al. Xu and Yu (1997). In the vector notation employed in the preceding paragraphs the z component of the position and velocity vectors is equal to zero and only rotation about the zaxis is considered. It is therefore not straightforward to benefit from parallel computing with a highly optimised event driven code. 1979) is the most popular softsphere granular dynamics model since it was used by Tsuji et al. On the other hand if the neighbour list is chosen to be rather large all collisions will be detected but this will go at cost of the computational speed. 47 . (1993). Unfortunately there is no standard notation in Granular Dynamics (yet) and therefore it was attempted here to stay close to the notations used in the references above. For a review of various contact force models used in softsphere simulations the reader is referred to Walton (1992) or Schäfer et al. The notation and some definitions can differ somewhat from the ones used for the hardsphere model.Granular Dynamics ______________________________________________________________________________________ particles can occur. The linear spring/dashpot model (Cundall and Strack. (1998). Hence the CPU time required to progress a time step can vary significantly. The number of collisions to be processed can be considerably higher in a dense region of the bed than in a dilute region. A hardsphere simulation is an event driven simulation which implies that the amount of collisions to be processed per time step depends on the dynamics of the system. Kawaguchi et al. It was not the objective to select the best softsphere model available but the aim was to compare the results obtained with the most popular softsphere model for fluidisation simulations with the results obtained with the hardsphere model. Schwarzer (1995). (1996). SoftSphere approach Although the hardsphere model was used for the majority of the simulations in this work also a softsphere model was implemented. 3.
The particle velocities are updated using the accelerations from equations 2. the external forces will be discussed in section 6.37) The new particle positions are subsequently also obtained from a first order explicit integration: r = r0 + vDT . Fcontact is the contact force acting on the particle. T is the torque acting on the particle and I is the moment of inertia of the particle as defined in equation 2. (2.36) (2.34) (2. dt 2 dω = T.38) where the subscript 0 denotes the value at the previous time step. & ω = ω 0 + ω 0 DT . Fexternal is the external force acting on the particle. (2.8. ω is the rotation velocity.35) where r is the position vector of the centre of the particle. m is the mass of the particle.1 The linear spring/dashpot model In softsphere models the following equations of motion are used: m I d 2r = Fcontact + Fexternal . In this section the focus will be on the contact forces between the particles.Chapter 2 ______________________________________________________________________________________ 3.35 by means of a first order explicit integration: & v = v 0 + v 0 DT . dt (2.34 and 2. 48 .
The linear spring/dashpot model. Two particles a and b are in contact (i.11.10.e.39) The repulsive force acting on the particles is divided into a normal (Fn ) and a tangential (Ft) component. as shown in Figure 2.10. a + R p. The procedure of evaluating these contact forces starts with defining a normal unit vector.Granular Dynamics ______________________________________________________________________________________ The interaction forces between particles in contact are modelled with a spring. a dashpot and a friction slider. spring friction slider dashpot Figure 2. The contact forces are evaluated from the overlap between the particles and their relative velocities. pointing from the mass centre of particle a to the mass centre of particle b as is presented in Figure 2. b (2. 49 . have mutual overlap) if the distance between their centres is less than the sum of their radii: rb − ra < R p.
Definition of the coordinate system used in the softsphere model Note that this definition differs from equation 2. (2. rb −ra n ab = (2.40) The relative velocity of particle a with respect to b is given by: v ab = ( v a − v b ) + ( Ra ω a + Rb ωb ) × n ab .1 used in the hardsphere model: rb −ra .Chapter 2 ______________________________________________________________________________________ r bra Rp. ab = (v ab ⋅ n ab )n ab .a nab ωa tab y x Figure 2. or slip velocity of particle a with respect to b can now be obtained as follows: 50 .41) The normal component of the relative velocity is given by: v n.b ωb vb va Rp.42) The tangential relative velocity. (2.11.
ab (2.46) The contact forces are now given by: Fn.44) By defining the tangential unit vector this way the vector is always pointing in the direction of the slip velocity.47) Ft.ab = − k nξ n n ab − ηn v n .48) If however the following relation is satisfied: 51 .45) The tangential displacement that has been established since the beginning of the contact is obtained by integrating the relative velocities with respect to time: ξ t (t ) = ∫ v t.ab (2. (2. ab (2.ab = −k t ? t − η t v t.43) The tangential unit vector is defined as follows: t ab = v t. ab .ab = v ab − v n. ab dt t0 t (2. ab v t. The overlap in the normal direction can immediately be calculated as the difference between the sum of the particle radii and the distance between the particles: ξ n = (R a + Rb ) − ra − rb (2.Granular Dynamics ______________________________________________________________________________________ v t.
It should be stressed here once again that the contact force model presented above was used because it is the most popular model for fluidised bed simulations.ab ) .ab + Ft. (2. b (2.1DT the equations of motion were solved by taking only the contact forces into account.ab > µ Fn. The energy dissipated during contact between particles a and b can be calculated by solving the following integral over the duration of the contact: E dsp.53) Since the contact forces are in general at least an order of magnitude larger than the external forces a separation of time scales was introduced. b (2.moving and of infinite mass just like in the hardsphere model. Therefore it is 52 .52) Ta = ∑ (Ra n ab × Ft.ab ) .50) For contacts between particles and walls. Therefore the resulting force and torque acting on particle a are obtained by summation of the forces with respect to b: Fcontact . (2.ab = − µ Fn. (2.51) Particle a can be in contact with several particles at the same time. ab = − ∫ Fab ⋅ v ab dt .ab t ab . ab .49) then sliding occurs and the tangential force is given by: Ft. a = ∑ (Fn. the walls are assumed to be non. At each time step DT the external forces were taken into account while at 0.Chapter 2 ______________________________________________________________________________________ Ft.
54) 53 .000 N/m in their simulations. In fact.48) can lead to very unrealistic behaviour if no special measures are taken as will be demonstrated in the following paragraph. 3. However. The measurements by Mullier et al. Kawaguchi et al. Tsuji et al. In most simulations the value of the spring stiffness is chosen rather low for computational convenience. (2. From equation 2.5ηB 2 ) t exp (− 0.Granular Dynamics ______________________________________________________________________________________ the most useful model for a comparison with the hardsphere model. (1998) all used a spring stiffness of 800 N/m for aluminum particles whereas Xu and Yu (1997) used a much higher value of 50. the expression for the tangential o contact force (equation 2. At higher values for the spring stiffness the duration of the contact becomes smaller and this requires a smaller time step to ensure numerical stability. 1992). (1991) suggest a value of about 400. (1991) for 6 mm diameter cellulose acetate spheres showed a linear dependency of the normal load upon the normal displacement apart from a short initial stage (Walton. (1993). Therefore the use of the linear spring can be justified for this type of material although it is one of the simplest models available. The first is the spring stiffness (k) for both the linear normal and the linear tangential spring. For a more elaborate discussion on contact theory the reader is referred to Johnson (1985).000 N/m for the spring stiffness of the 6 mm diameter cellulose acetate spheres. and v n = vn0 at t = 0): ξ n (t) = v n0 B 2 k n − (0.5ηB 2 t ) .5ηB 2 ) 2 2 sin B 2 k n − (0. Measurements by Mullier et al. it is not necessarily the best contact f rce available. It is likely that the values for the spring stiffness of 4 mm diameter aluminum spheres used in the work mentioned above are even higher.2 Model parameters The linear spring/dashpot model features three key parameters that will be discussed here.47 (harmonic oscillator with (weak) damping) a solution for the displacement and relative velocity as a function of time can be obtained using the initial conditions (ξn = 0 at t = 0. (1998) and Mikami et al.
54 .56) Note that this result differs from the result presented by Mikami et al. B2 (2. In a similar way a relation for the normal damping coefficient ηn can be obtained: ηn = − 2 ln e . n = π 2 + (ln e ) B2 k n 2 . n (2. n. and vn = e vn0 at t = t contact. In the case of (e = 0) the system is a critically damped harmonic oscillator and the contact time will be infinite. n : t contact.2 and adjusted the damping coefficient until the rebound height of the particle matched the expected value for that particular coefficient of restitution.57) Xu and Yu (1997) presented a trial and error method to obtain the value of the damping coefficient that is far less elegant. (2.17.55) where B2 is defined in equation 2.16 and e is the normal coefficient of restitution defined in equation 2. They simulated a particle wall collision similar to the case presented in Figure 2.Chapter 2 ______________________________________________________________________________________ From this solution the (normal) contact time between two particles can be found using ξ n = 0 at t = t contact. (1998) which is only valid for a particlewall contact. B 2 t contact. In that case the following relation should be used to obtain the damping coefficient η n : ηn = 2 kn .
48 the following expression is found for the tangential contact time: t contact.59) If this relation is not satis fied the tangential spring will still be loaded at the instant of detachment. Binary collision between rough spheres. This implies that energy is stored in the spring that causes a defect in the 55 . t π 2 + (ln β 0 ) = .15 and β 0 is the coefficient of tangential restitution defined in equation 2. This can lead to energy inconsistencies as will be demonstrated here.13a). When applying the same analysis as for the normal contact time starting from equation 2. (2.Granular Dynamics ______________________________________________________________________________________ Another important issue is the difference between the normal and the tangential spring stiffness. Since B1 and B2 are not equal and e and β 0 are not necessarily equal as well. a b Figure 2. the following relation must be satisfied to ensure that the normal and tangential contact times are equal: 2 ( )2 2 π + ln β 0 kt = 7 π 2 + (ln e) 2 k n . B1 k t 2 (2. In all softsphere simulations of fluidised beds reported in the literature the tangential spring stiffness is chosen to be equal to the normal spring stiffness.58) where B1 is defined in equation 2.12 (see also Figure 2. Consider the case of a binary collision between rough particles (no sliding occurs) where there is a significant tangential component of the relative velocity as depicted in Figure 2.12.19.
SoftSphere For the majority of the simulations presented in this work the hardsphere model is used. there are certain cases where a softsphere model is to be preferred.fluidised beds presented in the literature. 4. For the tangential damping coefficient the following result is obtained: − 2 ln β 0 . In the following paragraphs the key differences between the two approaches will be highlighted. t ηt = (2. This has never been addressed befo re in any of the softsphere simulations of gas. However. In the following section this will be demonstrated. In that case the following relation should be used to obtain the tangential damping coefficient η t : ηt = 4 kt .60) Note that also the tangential damping coefficient differs from the normal damping coefficient. HardSphere vs. 56 . B1t contact. 7 B1 (2.61) The coefficient of friction (µ) plays the same role in the linear spring/dashpot model as it does in the hardsphere model. In the case of (β 0 = 0) the system is a critically damped harmonic oscillator and the contact time will be infinite.Chapter 2 ______________________________________________________________________________________ energy balance over the contact period. It can be obtained from experiments as will be explained in section 5.
Despite of the double precision accuracy used in the simulations the particles eventually overlap (although this overlap is far smaller than the atomic length scale) due to number loss and the algorithm breaks down. The number of collisions then increases exponentially whilst the impulse concerned with a single collision becomes negligible. 1998) play an important role.2 Spring stiffness In the limit of an infinitely high value of the spring stiffness the softsphere model meets the hardsphere model. This can lead to energy inconsistencies as will be demonstrated here.1 Static situations For dynamic systems in general a hardsphere simulation is computationally faster than a softsphere simulation as long as there is sufficient motion in the system and the void fraction does not become too low. where two particles collide perfectly elastically and 57 .Granular Dynamics ______________________________________________________________________________________ 4.fluidised beds) these predictions have no value when an artificially low spring stiffness is used. In the vast majority of the linear spring/dashpot softsphere models presented in the literature the tangential spring stiffness is chosen to be equal to the normal spring stiffness. 4. The particles become very closely packed with very low relative velocities. It is important to realise that where a softsphere model in principle is capable of predicting the contact time between particles (which is of great importance for studying heat and mass transfer phenomena in gas.12. A softsphere model has no problems in dealing with static situations and is therefore to be preferred for simulations where such situations can occur. However the softsphere model cannot be run at very high values of the spring stiffness due to computational limitations. This includes systems where attractive forces between particles (like for example the liquid bridge forces in the simulations by Mikami et al. Furthermore there is the issue of the tangential spring stiffness versus the normal spring stiffness. Consider the case presented in Figure 2.fluidisation occur. The assumption of instantaneous collisions in the hardsphere model does not capture the whole physics of the collision process but neither does the assumption of using an artificially low value for the spring stiffness in the softsphere model. Severe problems are encountered when static situations like for example de.
01 m/s and a rotation velocity of 1. 58 . A time step of 107 s was used for the numerical integration.Chapter 2 ______________________________________________________________________________________ perfectly rough ( = 1. From Figure 2. b) Tangential contact forces as a function of time.13a the rotation velocity of particle a is plotted as a function of time for three cases: a hardsphere simulation. (1993) and a value of 1000 kg/m was used for the normal spring stiffness. Figure 2. It is clear that in the case where k t = k n the rotation velocity after collision is not predicted correctly.13.13b) it can be seen that the tangential contact force is nonzero at the end of collision implying that not all the energy that is put into the spring is recovered.01 m/s. β 0 = 1 and µ is large enough to ensure a sticking collision). In Figure 2. a softsphere simulation with a tangential spring stiffness (k t ) at 2/7 of the normal spring stiffness (k n ) and a softsphere simulation with k t = kn .0 1/s toward particle b which does not rotate and moves at a velocity of –0. Hence energy inconsistencies can arise when k t is chosen to be equal to k n . It should be noted however that these inconsistencies are rather small since in general the rotation energy is two orders of magnitude smaller than the kinetic energy. a) Rotation velocity of particle a as a function of time. e Particle a moves at a velocity of 0. The particles in these simulations were 4 mm diameter aluminium spheres (ρ = 2700 kg/m3 ) as previously used by Tsuji et al.
47 104 s can be calculated using equation 2. Tsuji et al.14 the effect of the time step on the resolution of the normal contact force as a function of time is presented. Consider the same case as in the previous paragraph (see Figure 2. (1998) performed simulations of a different system but used the same ratio between contact time and time step of 3. In Figure 2. 59 . (1993) and Kawaguchi et al.55. The time step should always be chosen carefully in such simulatio ns to avoid energy inconsistencies. Hence numerical errors cannot cause any problems concerning energy conservation. Figure 2.14. A spring stiffness of 800 kg/m was chosen in order to match the parameter settings used by Tsuji et al.Granular Dynamics ______________________________________________________________________________________ 4.9). For this binary headon collision with the coefficient of (normal) restitution (e) set equal to 0. (1998).7.3 Energy considerations The main advantage of the hardsphere model is that there is an analytical solution available for the collision model.12) but now none of the particles do rotate.9 a contact time of 7. This is not guaranteed for softsphere models where the accuracy of the solution depends on the time step used for the numerical integration. The normal contact force as a function of time for a binary headon collision at two different time steps (e = 0. Mikami et al. (1998) used a time step of 2 104 s in their simulations with these particles. Given the velocities of the particles prior to collision together with the particle properties the postcollision velocities can be calculated exactly. (1993) and Kawaguchi et al.
The second collision is processed immediately afterwards where the postcollision 60 . However the additional search for first contact goes at a cost of the computational speed whereas the actual integration is still performed at a rather large time step. Consider the case presented in Figure 2. Another typical feature of the linear spring/dashpot model can be observed in this figure as well: in the final stage the contact force becomes attractive instead of repulsive (DT = 105 s). The particles are positioned in such a way that a three body collision should occur.15 where a particle approaches two other particles at constant velocity in the absence of an external force.64 % for the 2 104 s case and 0.e. In Figure 2. In the hardsphere routine one of the two possible collisions is detected as the first collision to occur and this binary collision (in this case a perfectly elastic. The collision times for both pairs of particles are identical. attractive). This is due to the damping term in equation 2. when the overlap approaches zero.05% for the 1 105 s case). perfectly smooth collision) is processed first. Nonetheless one should keep in mind that in a typical fluidised bed simulation a lot of collisions occur and this error can accumulate quite rapidly.14 this can be observed very well since the time at which the particles first come into contact is not the same for both time steps.Chapter 2 ______________________________________________________________________________________ It can be seen that the curve for the case with biggest time step (2 104 s) is by far not as smooth as the curve for the case with the smallest time step (1 105 s). The strategy used by Xu and Yu (1997) who determine the instant of first contact precisely by employing a collision search algorithm from hardsphere simulations would prevent this.4 Multiple particle interactions A drawback of the hardsphere model is the assumption that the particles interact via instantaneous binary collisions. Another error arises due to the determination of contact between particles at a constant time step. 4. Improvements can be achieved by employing higher order numerical schemes. However this results in a relative error in the energy balance that is surprisingly low (1. the relative velocity increases and hence the damping term causes the contact force to become negative (i.47: at the end of the contact.
16 is obtained. Figure 2.15.15. which is physically not correct. Threebody collision simulated with a hardsphere model.Granular Dynamics ______________________________________________________________________________________ velocity of the first collision is used as the initial velocity for the second collision. It should be noted that this specific test case 61 .16. This renders the rebound pattern shown in Figure 2. Threebody collision simulated with a softsphere model. Because the softsphere model is capable of handling multiple particle interactions the rebound pattern is predicted correctly here. When the same case is repeated with the softsphere model the result presented in Figure 2. Figure 2.
. They studied the dependency of the coefficient of restitution on impact 62 . Measurement of collision parameters In the hardsphere model used in this work three parameters are required to describe to describe a collision: the coefficient of restitution (e). 1994). In more recent years the attention was focussed on the impact of particles on flat plates (Maw et al. 5.Chapter 2 ______________________________________________________________________________________ was carefully selected to demonstrate the effect of the assumption of instantaneous. 1990) where more detailed data was obtained. (1997) presented an advanced image processing technique employing three synchronised CCD cameras that enables accurate measurements of the particle position as well as the orientation of the particles. the coefficient of tangential restitution (β 0 ) and the coefficient of friction (µ). Foerster et al. The book by Goldsmith (1960) provides a good review of both theory and experiments concerning impacts. 1981 and Sondergaard et al. This is rather surprising since rigidbody impact was already a research subject in the days of Galilei (1638) and the concept of the coefficient of restitution was introduced by Newton (1686) accompanied by the first experimental data. It was not until the last decade that also particleparticle collisions became a subject of experimental investigations. µ and β 0 ) by careful experimentation and image processing. Labous et al. (1997) presented an experimental technique to measure collisional properties of spheres using highspeed video analysis. These three parameters are also required for the linear spring/dashpot model besides the spring stiffness.. binary collisions. Lorenz et al. (1994) presented a method to obtain the three collision parameters (e. Bernasconi et al. Unfortunately experimental data for these collision parameters are scarce in the open literature. (1997) presented a refinement of this technique together with additional data for more types of particles then the cellulose acetate and soda lime glass spheres used by Foerster et al. The difficulty in these experiments lies in the control of the geometry of the impact. Measurements of the coefficient of restitution are reported there as well including the dependency of this parameter upon impact velocity and particle size.
Despite the fact that quite a number of research groups are capable of measuring collision properties there is still a lack of collision data in the literature. 63 .17. The collision properties of the glass particles used in simulations reported in Chapters 6 and 8 were measured at this facility. The technique to measure particle impact parameters accurately was first developed for particle impacts on flat plates and is described in detail by Kharaz et al. This setup was used to measure the particlewall impact parameters used for a simulation reported in Chapter 6.up were required that are reported by Gorham and Kharaz (1999) and will be briefly outlined here. Schematic representation of the experimental set up used for the measurement of particleparticle collision properties: a) before collision b) after collision.17.Granular Dynamics ______________________________________________________________________________________ velocity and particle mass. Falling particle Triggering device CCD camera Stationary particle Lever arm Solenoid a) b) pivot Figure 2. The setup used for the measurements is schematically represented in Figure 2. The Impact Research Group of the Open University at Milton Keynes has developed an accurate technique to measure collision properties. (1999). To enable measurement of particleparticle impact parameters some adjustments to the set.
The two other collision parameters can now be obtained from a plot of Ψafter as a function of Ψbefore that are defined as follows: − ( v ab ⋅ t ) .17a). 0 ⋅ n) . Two straight lines are fitted through the data points: one for the sticking regime and one for the sliding regime. can be calculated using equations 2. The leverarm is then quickly moved away at 0.moving. (2.64) 64 . Since the particles do not rotate before collision it is not necessary to know the particle rotation after impact since the velocity after impact (vab ). The coefficient of restitution (e) can easily be obtained from the normal components of the relative velocity before and after impact. It then falls through an optical. The coefficient of tangential restitution can be obtained from the slope of the sticking line: Ψbefore = − β 0 Ψafter .14. By following this procedure the particle can be assumed to be non. (v ab.0 ⋅ t ) Ψafter = Ψbefore = (2. the strobe light and the solenoid (Figure 2.0 m/s. non rotating and still at the same initial position. Each data point represents one collision measurement where the impact velocity was about 1.6 and 2.3 ms before collision leaving the stationary particle entirely free and in the correct position for a collision (Figure 2. which includes a rotation contribution.Chapter 2 ______________________________________________________________________________________ A particle is released from a vacuum nozzle with zero rotation.fibre triggering device which generates a sequence of pulses to control the CCD camera.0 ⋅ n ) − (v ab.18.62) (v ab .17b). This is required for accurate determination of the particle positions from the recording by digital image processing which finally yields the particle velocities before and after collision. The collision is then recorded on a single frame of the CCD camera using strobe flashes that are timed in such a way that particle images do not overlap. (2.63) Such a plot is schematically represented in Figure 2.
Typical results obtained with this technique for glass particles show the following values for the collision parameters: e = 0. The model developed by Maw et al. Each individual measurement yields either a value for e and β 0. For nearly headon collisions (low values of Ψbefore) the measurements show positive values of Ψafter.18. 2 (2. (1976) however is capable of predicting this phenomenon.01 and β 0 = 0. or e and µ depending on the collision regime. Ψbefore for a typical result of a measurement.01.97 ±0.65) Ψafter sliding Ψbefore sticking Figure 2.1 ±0. Nevertheless it is believed 65 .33 ±0. Schematic representation of a plot of Ψafter vs.Granular Dynamics ______________________________________________________________________________________ and the coefficient of friction can be obtained from the intercept of the sliding line: Ψbefore = Ψafter − 7 (1 + e )µ .05. µ = 0. This phenomenon is referred to as microslip and cannot be explained by our hardsphere model. Nevertheless a series of measurements is required to obtain the two linear fits that determine the two regimes.
A similar equation of motion was used by Kawaguchi et al. The second term is due to the drag force where β represents an interphase momentum exchange coefficient as it usually appears in twofluid models. External Forces The incorporation of external forces differs somewhat from the approach followed by Hoomans et al. In equation 2. For high void fractions (ε ≥ 0. u the local gas velocity and Vp the volume of a particle.75(1 − ε ) ρg Dp u −v p (2. the forces now act on a single particle: m dv p p dt = mg+ p ( u − v ) − V ∇p (1 − ε ) p p p Vβ (2. µg the viscosity of the gas and ρg the density of the gas. (1996). vp its velocity. 6.80) the following expression for the interphase momentum transfer coefficient has been used which is basically the correlation presented by Wen and Yu (1966) who extended the work of Richardson and Zaki (1954): 66 . (1998). of course.Chapter 2 ______________________________________________________________________________________ that the threeparameter hardsphere collision model contains sufficient physics to be capable of reliable predictions. (1992) where.66) where mp represents the mass of a particle.67) where Dp represents the particle diameter.80) β is obtained from the wellknown Ergun equation: (1 − ε )2 β = 150 ε µg D 2 p + 1.fluid model presented by Kuipers et al.66 the first term on the right hand side is due to gravity. In this work the external forces are used in accordance with those implemented in the two. For low void fractions (ε < 0.
68) The drag coefficient Cd is a function of the particle Reynolds number: 24 0.44 Re p < 1000 Re p ≥ 1000 (2. j + Aii .Granular Dynamics ______________________________________________________________________________________ β = 3 ε (1 − ε ) Cd ρ g u − v p ε − 2. The local value was obtained from an area weighted averaging technique using the values of the pressure gradients at the four surrounding grid nodes. jj DXDY Q = (2.15Re p ) Cd = p 0. The local averaged value is calculated as follows: Ai . j = δ x DY − δ y Aii . j = ( DX − δ x ) DY − δ y Aii .66 is calculated by using a first order approximation.70) The pressure gradient in the third term on the right hand side of equation 2. jj + Ai . (2. j Qii. jj = δ x δ y Ai. j + Aii . The area weighted averaging technique used to obtain the local averaged value Q of a quantity Q(i.687 Re (1 + 0.19. j Qi . This technique is also used to obtain local gas velocities and local void fractions at the position of the centre of the particle.65 4 Dp (2. jj Qii .j) from the four surrounding computational nodes is shown in Figure 2. jjQi.71) where: Ai. jj = ( DX − δ x )δ y ( ( ) ) (2.72) 67 .69) where the particle Reynolds number in this case is defined as: Re p = ερ g u − v p D p µg .
jj Qii. Notation B1 . are calculated from the position of the particle in the staggered grid.66 can be taken into account as well. In Chapter 7 simulations of more dilute flows in the riser section of a circulating fluidised bed are presented where lift forces were also taken into account.66 an explicit first order scheme is used to update the velocities and the positions of the particles.jj Qi.j Figure 2.j Ai. necessary in this averaging technique. B2 Cd Dnblist 68 collision constants.jj Ai. Area weighted averaging.j position for the evaluation of Q DY A Area δx δy Aii. Other external forces than the ones included in equation 2.j DX Qii.Chapter 2 ______________________________________________________________________________________ The distances δ x and δ y. For the integration of equation 2.19.jj Aii. Qi. 1/kg drag coefficient. [] diameter of neighbour square .
Granular Dynamics ______________________________________________________________________________________ Dp dt nblist DT e E F g I J k m n ntot ncoll Rp r T t t t ab u v particle diameter. N gravitational acceleration. J force. kgm2 impulse vector. [] total number of collisions. s collision time. [] gas shear viscosity. m torque. s gas phase velocity. kgm/s spring stiffness. s time step. m void fraction. s coefficient of restitution. Ns/m 69 . [] distance. m/s2 moment of inertia. [] particle radius. [] damping coefficient. kg normal unit vector. m time step for neighbour list update. m/s Greek symbols β β0 δ ε µg µ η volumetric interphase momentum transfer coefficient. m position. [] total number of particles. N/m particle mass. [] time. Nm tangential unit vector. kg/(ms) friction coefficient. [] energy. kg/(m3 s) coefficient of tangential restitution. m/s velocity.
General Method.P. T.J. Chem. 49. (1990). Computer simulations of liquids.. M. Soc. J. 31. 1/s displacement.. (1954). (1959).A.63. A225.Chapter 2 ______________________________________________________________________________________ ρ ω ξ Ψ density. B. 459. 27. kg/m3 angular velocity. Phys.. 70 .free dispersion of large solid particles in a Newtonian fluid under shear. and Wainwright. D.E. Experiments on a gravity. J. UK. Allen. Oxford Science Publications. T. Bagnold. R. 1208.J. m defined in equations 2. and Tildesley. (1957). Alder. Phase Transition for a HardSphere System.b av cp dsp gyr nblist p w initial condition particle indices average contact point dissipated gyration neighbour list particle wall References Alder.. Lond.62 and 2. and Wainwright..E. Proc. R. Oxford.J. I. Chem. Phys. B. Studies in Molecular Dynamics. [] Subscripts 0 a.
C. 9 (5).G. Fandrich. Oswald’s Klassiker 25. K. and Brennen..D. C. Géotechnique. (1985). (1638). 6 (3).L... and Smit. 19. 167. 29. Mech. Measurements of the collision properties of small spheres. 71. Inc.. (1996). H.Fluid.. G.. Understanding Molecular Simulation.J. O.E. M. from Algorithms to Applications. M.. Mod.. 1108.. (1976). Oxford. Experimental 3D Impact Analysis. UK. D. 47. 361. Engelmann. W. C.Y. Phys. San Diego. Granular matter: a tentative view. Schade. 71 .F. Fluids. (1997). A particle scheme for the numerical solution of the Enskog equation. Th. 38. (1992). B. A discrete numerical model for granular assemblies.. Oxford University Press. Louge.Granular Dynamics ______________________________________________________________________________________ Bernasconi. (1999). and Hogue. P. and Strack. 151.. Int. Academic Press. Frank. S... Unterredungen und Demonstrationen über zwei neue Wissenszweige. S374.. J. Numerical simulation and experimental investigation of a gassolid twophase flow in a horizontal channel. (1994). and Allia. Multibody System Dynamics. Campbell. Frenkel. Leipzig. Fluids. Computer simulations of granular shear flows. 187.A. Chang. Bird. A.A. de Gennes.E.S. Phys. USA. Multiphase Flow. 1. Galilei.P. G. Phys. J. and Petrak. (1997).. P. 1329.. 189091. Foerster. D. D. Cundall. K. Frezzotti. Molecular Gas Dynamics and Direct Simulation of Gas Flows. (1979). Rev.
and van Swaaij.. R. Rev. Gorham. J. Impact. Powder Technol. DT8. 1996.M. Engng Sci. Beijing (China). (1996).Chapter 2 ______________________________________________________________________________________ Ge. H. Nagel. J.. C. 68. S. and Hayakawa.. Cambridge. and Behringer.. UK.A. Tanaka. Powder Technol.. (1960). Discrete particle simulation of bubble and slug formation in a twodimensional gasfluidised bed: a hardsphere approach. Cambridge University Press.L. W. Contact Mechanics. (1999).. Phys. Circulating Fluidized beds. and Louge. Inelastic microstructure in rapid granular flows of smooth disks.P. A. 658.. W.A. H. The Open University. Milton Keynes.M. and Kharaz. London.. Ichiki. E. 72 . 99.R. Phys. and Newland.B. (1994).. 78. Efficient computer simulation of moving granular particles. Jaeger. D.A. 1259. Numerical simulation of twodimensional fluidized beds using the discrete element method (comparison between the two. May 28. Johnson. Hopkins. Results of particle impact tests.. Fluids A 3(1). W. E.. Goldsmith..and threedimensional models). liquids and gases. B. the theory and physical behaviour of colliding solids. K. Hoomans. (1995).Y. UK. Y.M. Chem. 51. M. M. In Proceedings of the 5th Int. Conf. T. 47.P. (1997).. and Li. (1998). (1996). Kuipers. Hogue.P.. 96.. Arnold Pub. 129. T. 51. (1985). Impact Research Group Report IRG 13.. Rev. Dynamical simulation of fluidized beds: hydrodynamically interacting granular particles. W.June 1.H. (1991). D. Mod..J. Briels. PseudoParticle Approach to Hydrodynamics of Gas/Solid TwoPhase Flow. Granular solids. Phys. and Tsuji. 1. K. 52. Kawaguchi.
(1997). van Duin K..N. Langston. Discrete element simulation of granular flow in 2D and 3D hoppers: dependence of discharge rate and wall stress on particle interactions.. and Yang. transfer and collection. and van Swaaij. Exp. Tüzün U.. G. A. W. Continuous potential discrete particle simulations of stress and velocity fields in hoppers transition from fluid to granular flow. Chem. Klinzing. Eng. 1259. and Louge. H..J..Granular Dynamics ______________________________________________________________________________________ Kharaz.K. Chem.. and Salman. D.. nearly spherical particle s.. 44.. (1994). (1996)..M. D. Numerical simulation of dilute turbulent gassolid flows in horizontal channels. Chem. Mech.M. 90 (4).C. Granular dynamics of slightly inelastic spheres in Couette flow.M. W. Engng Sci. Rev. Measurement Science and Technology 10. 292. F. Lun.S.A. Tuozzolo. Kuipers.P. 1913 Labous.. M.. E. Measurements of collisional properties of spheres using highspeed video analysis.. D. A numerical model of gasfluidized beds.. 47. Carson. T. 31. 967. Lorenz. Phys. Phys.C. 50. A..M. Knowlton.K. Engng Sci. Int. (1999). (1995). A.H. R. Tüzün. Accurate measurement of particle impact parameters.A. P.A.P. 49.. 37. (1997). and Liu.H.Y. L. Gorham. Chem. Progress. J. Fluids 8. van Beckum. 5717. Langston. Rosato.A. 575. C.. Measurement of impact properties of small. Lun.. Multiphase Flow 23. U. and Heyes. and Dave.M. (1992). and Heyes. 73 . (1994).E. C. 2868. C. A..D.. The importance of storage. J. 56. Engng Sci.D. J. (1997).W. P.C..
Carry. and Liebling. (1998). (1976). D. Efficient algorithms for manybody hard particle molecular dynamics. Moreau eds. 97.le –Rouet. Tüzün. Detection of collisions of polygons by using a triangulation. 91. and Ottino. J. J. France. (1991).J.N. 74. Department of Chemical Engineering. McCarthy. Th. (1998). (1981). M. The oblique impact of elastic spheres.R. M... Kamiya..R. J. 103. 74 .M. 306. P. Barber. September 1923. N... T. Maw. Japan. (1993). p. Adee.. H. Numerical simulation of cohesive powder behavior in a fluidized bed.. T.J. A singleparticle friction cell for measuring contact frictional properties of granular materials. Newton. 1927. Maw. Müller. J. In Contact Mechanics (M. Mikami. M. PhD thesis. (1686). Lubrication Tech. Mullier.. 65.. Powder Technol. Comput. Chem.M. 61.N. I. ASME: J. J. 369. and Fawcett. Jean and J. N. 1848. and Fawcett.. (1998). Motte). 109. Engng Sci. D. U. (Translated into English by A. Particle dynamics simulation: a hybrid technique applied to granular mixing. J. 53.Chapter 2 ______________________________________________________________________________________ Marin. Tokyo University of Agriculture and Technology. 101. M.. and Walton.R. J. Barber. Risso. and Horio.. Powder Technol.).. Phys. (1995). Mikami. and Cordero. O. Trans. D. Philosophiae naturalis principia mathematica. Wear 38.. The role of elastic tangential compliance in oblique impact. New York. Agglomeration fluidization of liquid/solid bridging particles an its control. Raous. proceedings of the 2nd Contact Mechanics International Symposium.
Granular Dynamics ______________________________________________________________________________________
Oesterlé, B. and Petitjean, A., (1993). Simulation of particletoparticle interactions in gassolid flows, Int. J. Multiphase flow 19, 191.
Ouyang, J. and Li, J. (1999). Particle motionresolved discrete model for simulating gassolid fluidization, Chem. Engng Sci. 54, 2077.
Richardson, J.F. and Zaki, W.N., (1954). Sedimentation and fluidization: part I. Trans. Inst. Chem. Eng. 32, 35.
Rosato, A., Prinz, F., Standburg, K.J. and Swendsen, R., (1986). Monte Carlo Simulation of particulate matter segregation, Powder Technol. 49, 59.
Schäfer, J., Dippel, S. and Wolf, D.E., (1996). Force schemes in simulations of granular materials, J. Phys. I France 6, 5.
Schwarzer, S., (1995). Sedimentation and flow through porous media: Simulating dynamically coupled discrete and continuum phase, Phys. Rev. E. 52, 6461.
Seibert, K.D. and Burns, M.A., (1996). Simulation of fluidized beds and other fluidparticle systems using statistical mechanics, AIChE J. 42, 660.
Seibert, K.D. and Burns, M.A., (1998). Simulation of structural phenomena in mixedparticle fluidized beds, AIChE J. 44, 528.
Sommerfeld, M., (1990). Numerical simulation of the particle dispersion in turbulent flow: the importance of particle lift forces ad particle/wall collision models, In Numerical Methods for Multiphase Flows, Vol. 91. ASME, New York.
Sondergaard, R., Chaney, K. and Brennen, C.E., (1990). Measurements of solid spheres bouncing off flat plates, J. Appl. Mech. 57, 694.
75
Chapter 2 ______________________________________________________________________________________
Stronge, W.J., (1990). Rigid body collisions with friction, Proc. R. Soc. Lond. A. 431, 169.
Thornton, C., (1997). Coefficient of restitution for collinear collisions of elasticperfectly plastic spheres, J. Appl. Mech. 64, 383.
Tsuji, Y., Morikawa, Y., Tanaka. T., Nakatsukasa, N. and Nakatani, N., (1987). Numerical simulation of gassolid twophase flow in a twodimensional horizontal channel, Int. J. Multiphase Flow 13, 671.
Tsuji, Y., Tanaka, T. and Ishida, T., (1992). Lagrangian numerical simulation of plug flow of cohesionless particles in a horizontal pipe, Powder Technol. 71, 239.
Tsuji, Y., Kawaguchi, T. and Tanaka, T., (1993). Discrete particle simulation of two dimensional fluidized bed, Powder Technol. 77, 79.
Tsuji, Y., Tanaka, T. and Yonemura, S., (1998). Cluster patterns in circulating fluidized beds predicted by numerical simulation (discrete particle model versus two fluid model), Powder Technol. 95, pp. 254264.
Walton, O.R. and Braun, R.L., (1986). Viscosity and temperature calculations for assemblies of inelastic frictional disks, Journal of Rheology 30, 949.
Walton, O.R., (1992). Numerical simulation of inelastic, frictional particleparticle interactions, in Particulate TwoPhase Flow, Part I, edited by M.C. Roco, ButterworthHeinemann, Boston, 884.
Wang, Y. and Mason, M.T., (1992). Twodimensional rigidbody collisions with friction. J. Appl. Mech. 59, 635.
76
Granular Dynamics ______________________________________________________________________________________
Wen, C.Y. and Yu, Y.H., (1966). Mechanics of fluidization. Chem. Eng. Prog. Symp. Ser. 62 (62), 100.
Wightman, C., Moakher, M., Muzzio, F.J. and Walton, O., (1998). Simulation of flow and mixing of particles in a rotating and rocking cylinder, AIChE J. 44, 1266.
Xu, B.H. and Yu, A.B., (1997). Numerical simulation of the gassolid flow in a fluidized bed by combining discrete particle method with computational fluid dynamics. Chem. Engng Sci. 52, 2785.
Yonemura, S., Tanaka, T and Tsuji, Y., (1993). Cluster formation in gassolid flow . predicted by the DSMC method, In Proc. ASME/FED Symp. No. 166, ASME, New York, USA, 303.
77
Chapter 2 ______________________________________________________________________________________
78
Gas Phase Hydrodynamics ______________________________________________________________________________________
Chapter 3.
GAS PHASE HYDRODYNAMICS
Abstract:
In this chapter the model for the gas phase hydrodynamics is presented. The twodimensional model is based on the volume averaged NavierStokes equations for twophase flow and is solved on a scale larger than the particle size. This requires the drag force exerted on the particles by the gas phase to be incorporated in the model through empirical relations. A short review of techniques that allow the flow around each individual particle to be calculated, and thus do not require any empirical input concerning fluidparticle drag, is presented as well. The numerical solution technique and the applied boundary conditions are briefly outlined. Twoway coupling between the motion of the particles and the motion o the gasphase is established via the void f fraction and a source term in the momentum conservation equation for the gas phase. A detailed description of the calculation of the void fraction from both the 2D and the 3D granular dynamics model is presented. The incorporation of the momentum source term in the gas phase hydrodynamics model is explained as well.
79
Kuipers. B..B.B. 54.A. 51. Engng Sci.M. J.P. W.J. Briels and W.P. Hoomans. 99. (1998). Discrete particle simulation of bubble and slug formation in twodimensional gasfluidised beds: a hardsphere approach.M.H. Yu. Chem. Briels and W.J.A.P.B.. J. W. Engng Sci.Chapter 3 ______________________________________________________________________________________ Parts of this chapter are based on the papers: B.M. Hoomans.M.P. 2645. 80 . Chem. van Swaaij. (1996). Xu and A. Kuipers. van Swaaij. Comments on the paper ‘Numerical simulation of the gassolid flow in a fluidized bed by combining discrete particle method with computational fluid dynamics’ by B.
Gas Phase Hydrodynamics ______________________________________________________________________________________ 1. This implies that special care has to taken to achieve correct twoway coupling. Hu (1996) used a finite element technique 81 . In the approach used in this work the gas phase hydrodynamics is resolved on a length scale that is larger than the particle size. Introduction The drag force exerted on a particle by the gas phase requires the velocity of the gas phase at the position of the particle. An alternative for this method was presented by Pan and Banerjee (1996) where the particle effects on the fluid motion were fed back by calculating the velocity disturbance caused by the particles assuming that the flow around them is locally Stokesian. (1977). 1. The approach followed in this work is described in the preceding sections. The technique adopted in this work is reminiscent of the ParticleSourceIn Cell method developed by Crowe et al. However. This can provide improved closure relations for the drag force that can be used in Eulerian Lagrangian simulations where the flow field is resolved on a scale larger than the particle diameter. There are several methods available to solve the NavierStokes equations in such a comple x geometry. Guj and de Matteis (1986) presented a technique to solve the twodimensional flow field around an arrangement of particles using boundary fitted coordinates. Hence empirical relations are required to determine the drag force exerted on the particles by the gas phase. the drag force exerted on the particle can be obtained by integration of the stress tensor over the particle surface. This technique is however rather expensive in terms of CPU time and is limited to rather dilute flows of particles that are slightly heavier than fluid. it is worthwhile to discuss some techniques here that are capable of solving the flow field around each individual particle and therefore enable the calculation of the force exerted on the particle by the fluid without any empirical input.1 Direct solution of the NavierStokes equations When the flow around a particle is solved using a noslip boundary condition on the particle surface.
the number of particles that can be used in these simulations is still limited (102 103 ). Eggels and Somers (1995) extended the technique by including a turbulent stress tensor as well as a scalar transport equation. In these 3D simulations special attention was paid to ensure noslip boundary conditions on the 82 . 1. Furthermore these are still twodimensional simulations and hence threedimensional models would put even higher demands on computing power. The transport properties of the system (for example the viscosity) are defined by the eigenvalues of the collision operator used in the collision phase. A time step in a lattice Bo ltzmann simulation consists of two phases: a collision phase and a propagation phase. Ladd (1994) used this technique to simulate colloidal systems. The lattice Boltzmann technique is also capable of performing simulations of multiphase systems. Kalthoff et al. In this technique a oneparticle velocity distribution function rather than ‘real particles’ (as in lattice gas cellular automata as described by Frisch et al. 1986) is transported on a carefully selected lattice. Although these techniques provide a lot of detail without requiring any empirical input.. Since the grid size is smaller than the particle diameter a forcing technique was used to ensure noslip boundary conditions on the particle surface.2 Lattice Boltzmann simulations A flow simulation technique that has become very popular in recent years is the lattice Boltzmann technique pioneered by McNamara and Za netti (1988). This technique is capable of dealing with complex geometries which is important for the majority of flows encountered in industry. Eggels (1996) used this model to perform largeeddy simulations of baffled stirred tank reactors in full 3D. (1997) developed a technique where a finite difference NavierStokes solver is used to solve the twodimensional liquid motion. The drag force exerted on a particle was then obtained by numerical integration of the stress tensor over the particle surface using an analytical expansio n of the fluid field.Chapter 3 ______________________________________________________________________________________ based on moving unstructured grids to simulate the twodimensional motion of solid particles in a liquid.
is doubtful. However the system should be of limited size (several solid particles instead of thousands) due to computational limitations. The DPD technique is capable of handling complex geometries and has been used to study colloidal systems.fluid boundaries. Simulations of suspensions are possible by local freezing of particles that are found within the volume defined by the suspended particle.fluidised beds as described in this work.Gas Phase Hydrodynamics ______________________________________________________________________________________ particle. The method is therefore very well suited to obtain improved closure relations for the gasparticle drag force on the length scale at which it is applied in granular dynamics simulations of gas.3 Dissipative Particle Dynamics A somewhat similar approach to flow simulations is the dissipative particle dynamics technique originally developed by Koelman and Hoogerbrugge (1992).fluidised beds. The dynamics of these particles consists of a collision phase. For 83 . the net force exerted on the suspended particle is obtained. like gasfluidised beds. Despite the fact that this method is rather original it leaves quite some questions unanswered. Unlike in lattice Boltzmann simulations these particles are not restricted to lattice sites. and a propagation phase. where the interaction between the particles is taken into account. By summation of the forces experienced by all the particles that constitute the suspended particle. 1. Español and Warren (1995) have shown that the dissipative force as used by Hoogerbrugge and Koelman (1993) needs to be modified in order to obey the fluctuationdissipation theorem. It seems however that the DPD technique is limited to low Reynolds number flow and therefore its use for inertia dominated systems. This technique is based on the dynamics of particles that should be regarded as momentum carriers rather than physical particles. Hoogerbrugge and Koelman (1993) applied this technique to perform 3D simulations of hardsphere suspensions under steady shear. In principle this method can also be applied to gassolid systems as encountered in gas. Ge and Li (1997) presented a technique that resembles DPD in the sense that the gasphase is represented by discrete particles rather than the NavierStokes equations. where the particles are moved according to an equation of motion that contains contributions of a random and a dissipative force.
1) Momentum equation gas phase: ∂ ερg u ∂t ( ) + (∇ ⋅ ερ uu ) = − ε∇p − S − (∇ ⋅ ε τ ) + ερ g . (1992). g p g g (3. 2. twodimensional motion is considered which implies that three basic variables have to be specified.Chapter 3 ______________________________________________________________________________________ example it is not clear how physical properties of the gas are incorporated into the model. Continuity equation gas phase: ∂ ερg ( ) + ∇ ⋅ ερ u = 0 . It is based on the numerical solution of the following set of partial differential equations that can be seen as a generalised form of the NavierStokes equations for a gas interacting with a solid phase as originally derived by Anderson and Jackson (1967). The three basic variables in the model are the pressure ( ) and the two velocity components of the gasphase (ux and uy). Governing Equations The calculation of the gasphase hydrodynamics mainly follows the lines presented by Kuipers et al. Furthermore it is not clear what equation of motion is used for the gas particles.2) In this work isothermal. It would be helpful if it could be demonstrated that this model is capable of predicting key fluidisation phenomena such as the minimum fluidisation velocity and the terminal velocity of a particle falling under the influence of gravity in a quiescent gas. ( ) ∂t g (3. The void p fraction (ε) and the momentum exchange source term (Sp ) are obtained from the discrete 84 .
1 Gas phase density The gas phase density (ρg ) is related to the pressure (p) and the gas phase temperature (T) by the ideal gas law: ?g = Mg RT p .Gas Phase Hydrodynamics ______________________________________________________________________________________ particle model as will be explained in section 6. All remaining variables have to be specified in terms of the three basic variables and/or the variables obtained from the discrete particle model through constitutive equations.. Constitutive Equations 3. (3. 85 . The average molecular weight of air (Mg = 28..3) where R is the gas constant (8.2 Gas phase stress tensor The viscous stress tensor τ g is assumed to depend only on the gas motion.314 J/(mol K)). The general form for a Newtonian fluid (Bird et al. I denotes the unit tensor. 3. 3.8 103 kg/mol) was used and the temperature was set to a constant value of T = 293 K. 1960) has been implemented: 2 T τ g = − λg − µ g (∇ ⋅ u )I + µ g (∇u ) + (∇u ) 3 ( ) . 1960) whereas for the gas phase shear viscosity a constant value of µg = 1. For bubbling beds this can be justified since the turbulence is damped out in the bed due to the very high solids fraction.4) In the simulations the bulk viscosity of the gas phase λg was set equal to zero which is allowed for gases (Bird et al.8 105 kg/ms was used. Note that no turbulence modelling was taken into account. (3.
5. Lay out of the staggered grid.5) scalar variable DY (i. employing a staggered grid to ensure numerical stability. 86 . In the simulations reported in this work only the Cartesian option was used. A pressure correction technique was employed to solve the set of partial differential equations.1. The model is capable of performing transient twodimensiona l calculations in a Cartesian or an axisymmetrical geometry.Chapter 3 ______________________________________________________________________________________ 4. (i.j) xvelocity component yvelocity component DX Figure 3.j+0.1 and 3. (1993) and will not be discussed in detail here.2. This implies that the scalar variables (p and ε) are defined at the cell centre and that the velocity components are defined at the cell faces as is shown in Figure 3. Numerical Solution The numerical solution follows the lines of Kuipers et al. A finite difference technique. is used to solve the gasphase conservation equations 3.1.j) (i+0.
2.2.j) which is associated with the relevant boundary condition for that cell(i.j).Gas Phase Hydrodynamics ______________________________________________________________________________________ 5. Boundary Conditions For the incorporation of the boundary conditions a flag matrix is used which allows boundary conditions to be specified for each single cell. The typical set of boundary conditions used in the simulations performed in this study is shown in Figure 3. Cell flags for the boundary conditions for the hydrodynamic model 87 . A variety of boundary conditions can be applied by specification of the value of the cell flag fl(i. NY+1 7 5 5 5 7 NY 3 1 1 1 3 interior cells 3 1 1 1 3 fictitious cells j 3 1 1 1 3 1 3 1 1 1 corner cells 3 0 7 4 4 4 7 0 1 i NX NX+1 Figure 3.
The calculation of the drag force is based on empirical relations and therefore not exact..1. should balance the force exerted by gravity on the particles. Xu and Yu (1997) used a technique similar to the one used by Schwarzer (1995) where the total drag force exerted on a particle was fed back to the gas phase with a minus sign. free slip boundaries Continuous outflow cell. Therefore special care has to be taken in order to select a twoway coupling technique that ensures the key fluidisation features to be incorporated correctly. velocities have to be specified Prescribed pressure cell. Hence it is required that twoway coupling is achieved in such a way that the model is capable to predict typical fluidisation phenomena such as the pressure drop over the bed at minimum fluidisation conditions. fl(i. free slip boundaries Impermeable wall.1.j) = 4) was set to a constant value of 0. 1998). free slip boundaries Corner cell.4. 6. Cell flags and corresponding cell types. With this technique the pressure drop over the bed at minimum fluidisation conditions in the simulations of Xu and Yu (1997) was over predicted by a factor 1.5 (Hoomans et al. This pressure drop. Twoway coupling An important issue in granular dynamics simulations of twophase flow is the twoway coupling. no boundary conditions have to be specified To mimic a distributor plate the void fraction in the influx cells (fl(i.Chapter 3 ______________________________________________________________________________________ The cell flags and the corresponding boundary conditions that are featured in the present code are listed in Table 3.j) 1 2 3 4 5 6 7 Cell type of cell(i. no slip boundaries Influx cell. multiplied by the cross sectional area. Table 3.j) Interior cell. 88 . no boundary conditions have to be specified Impermeable wall.
Since the void fraction is an important parameter which considerably influences the motion of the gas phase. Since the particle positions are known the void fraction ε(i.j Figure 3.jj Ai. a detailed check for overlap was performed in which multiple cell overlap was taken into account as illustrated in Figure 3.1.j ii. 89 .Gas Phase Hydrodynamics ______________________________________________________________________________________ Twoway coupling is achieved via the calculation of the void fraction and the incorporation of an interaction term in the momentum conservation equation for the gasphase (equation 3.1 6. 6.3.1 Void fraction Calculation of the void fraction in 2D The solution of equations 3.1 and 3. i. In this figure a case is presented where a particle overlaps four different grid cells.3.jj Aii.jj δ1 δ2 Aii. Multiple particlecell overlap.2).jj ii.j) can be calculated based on the area occupied by the particles in that cell i.j i.2 requires specification of the void fraction (ε) which can be obtained from the discrete particle model.j. In the following paragraphs it is explained how twoway coupling is achieved in the model used in this work. The distances from the particle centre to the nearest boundaries of the grid cells are indicated by δ1 and δ2 .
(3. section 6). p (3.6) Keeping in mind that: Ai.5) For the area Ajj overlapping with the upper two cells in Figure 3. jj δ = R p R p arccos 2 R p 2 δ2 −δ 2 1− . R p (3. To correct for this inconsistency the void fraction calculated on the basis of area (ε 2D) is transformed into a threedimensional void fraction (ε 3D) using the following equation: ε 3D = 1 − 2 π 3 (1 − ε ) 2D 3 2 . However.Chapter 3 ______________________________________________________________________________________ The smallest overlapping area of the particle Aii. In the case where a particle overlaps with only two cells the area of overlap can simply be obtained from equation 3.7) all four areas can now be calculated using basic subtractions.8) 90 .3 the following relation can be obtained: A jj = Ai. jj 2 2 δ1 δ2 δ 1 +δ 2 1− − R p arccos 1 = δ 1δ 2 − R p δ 1 1 − R R R 2 p p p δ − arcsin 2 R p . j + Ai . (3. jj + Aii.6. jj + Aii. j = πR 2 . the void fraction calculated in this way is based on a twodimensional analysis that is inconsistent with the applied empiricism in the calculation of the drag force exerted on a particle (Chapter 2. jj + Aii.jj can be calculated as follows: Aii .
8 but used the square root of 2 instead of 2 in the second term on the right hand side. They calculated the void fraction on the basis of volumes assuming that the system consists of one particle layer. (1993). Ouyang and Li (1999) used a transformation similar to equation 3. Tsuji et al.jj can now be approximated by: 91 . 6. (1998) all used a different approach.jj. Since the gas phase hydrodynamics is still resolved in 2D each particle can still have overlap with four cells at maximum.Gas Phase Hydrodynamics ______________________________________________________________________________________ This equation has been derived on the basis of a comparison between a twodimensional hexagonal lattice and a threedimensional FCC unit cube assuming equal interparticle distances. The main difficulty in the 3D void fraction calculation is to obtain an expression for the volume Vii.1. When a particle overlaps with two cells the volume of a spherical cap can be calculated exactly by: π (R p − δ 1 )2 3 Vcap = (2R p + δ 1 ). Kawaguchi et al. (1998) and Mikami et al.9) with δ 1 being the distance from the centre of the particle to the cell boundary. Using this method the closest packing can never be obtained and therefore it is not a correct representation for a threedimensional system. Hence the third dimension is equal to the particle diameter. Consider Figure 3. Vii. This renders void fractions that in general are too high.3 but now read volumes instead of areas. Xu and Yu (1997). (3. It ensures that the closest packing in the 2D hexagonal lattice is transformed into the closest packing in the 3D FCC case. Since there is no analytical expression available for this volume an approximation has to be used.2 Calculation of the void fraction in 3D In the 3D model the problem with the conversion of a 2D void fraction to a 3D void fraction no longer exists. This method yields void fractions that are in general too high as well.
11) The δfunction ensures that the reaction force acts as a point force at the position of the particle in the system. Each of the 4 volumes can subseque ntly be obtained using basic subtractions.2 Momentum transfer The method used in this work is based on Newton’s third law and was also used by Delnoij et al. This virtual third dimension is 92 . A mixed explicit implicit numerical treatment was applied similar to the one used by Kuipers et al.j Vii. 6. these volumes can be calculated using equation 3.jj Vii.jj + V i.jj + V ii.jj+Vi. In the 2D model a virtual third dimension has to be introduced.9.Chapter 3 ______________________________________________________________________________________ Vii. section 6. (1993) to solve the gas phase conservation equations. (3. (1997).j are sphere caps. In the 3D model this is straightforward since the third dimension is determined by the depth of the bed. (1996). where also the expression for the volumetric momentum exchange coefficient β can be found. Since the source term Sp has the dimension of force per unit volume the force exerted on the particles has to de divided by the volume of a grid cell. In the numerical implementation this forcepervolume term is distributed to the four nearest grid nodes using the area weighted averaging technique described in Chapter 2.jj = V V p p V p . The reaction force to the drag force exerted on a particle per unit volume is included in the momentum conservation equation 3.jj+Vii. It differs from the technique used by Hoomans et al.10) Since Vii.jj and Vii.2 via a source term Sp that has the dimension N/m3 : Vp β Sp = − 1 V ∫ ∑ (1 − ε ) (u − v )δ (r − r )dV a a a= 0 Npart (3.
75 D p .fluid model requires averaged values of particle velocities and these can be obtained by averaging over all the particles in a computational cell. Kawaguchi et al. [] cell indices. [] number of computational cells in ydirection. (1998) and Mikami et al. Pa 93 . However. This technique was also employed by Hoomans et al. m/s2 unity tensor. Notation A Dp DX DY fl(i. Tsuji et al.j M NX NY p area. (1998) use a technique where the coupling was achieved in a similar way as in twofluid models. defined in Table 1 gravitational acceleration. (3. (1996). A two. Note that this volume depends on the particle size and hence is not a constant in simulations where a particle size distribution is taken into account. (1993).Gas Phase Hydrodynamics ______________________________________________________________________________________ estimated based on the same calculation as the conversion from ε 2D to ε 3D. m2 particle diameter. [] molecular weight. m cell flag [].12) in which the third dimension is slightly less than the particle diameter. m horizontal computational cell dimension. this technique is limited to systems consisting of particle of uniform size and therefore the technique used in this work was preferred. In the 2D simulations the volume of a computational cell was chosen to be: Vcell = 2 DR DZ 3 −0. m vertical computational cell dimension. [] pressure. kg/mol number of computational cells in xdirection.j) g I i.
Chapter 3 ______________________________________________________________________________________ r Rp R Sp t T u V v particle position vector. kg/(m. kg/m3 gas phase stress tensor. J/mol K source term defined in equation 3. m3 velocity. K gas phase velocity.s) gas shear viscosity. [] gas bulk viscosity. m/s volume. kg/ms2 Subscripts g i.j p x y gas phase cell indices particle xcomponent ycomponent Superscripts T transposed 94 . m particle radius. kg/(m. kg/(m3 s) distance. m void fraction.11 time.s) density. m gas constant. s temperature. m/s Greek symbols β δ ε λg µg ρ τ volumetric interphase momentum transfer coefficient.
1. Eggels. E. Heat and Fluid Flow. ASME Trans. Sharma.. P.. (1995).A.. D. Eggels. Kuipers. J.P. P.T. and Jackson.. May 28. (1996).Gas Phase Hydrodynamics ______________________________________________________________________________________ References Anderson.N. Stewart. (1977).B. 17. 7. The ParticleSourceIn Cell (PSICELL) model for gasdroplet flows. 52. 1429.G. Español. (1967). 95 . and Somers. and Stock.M..P. 6. Engng Sci.. Lett. 527. Chem.. Circulating Fluidized beds. Transport phenomena. and Lightfood. Delnoij. J. Fluids Eng. F. T. and Li. In Proceedings of the 5th Int.. DT8. Chem. J. Fundam. Lammers.. Direct and largeeddy simulation of turbulent and fluid flow using the latticeBoltzmann scheme. Crowe.E. Eng.. G. 1996. PseudoParticle Approach to Hydrodynamics of Gas/Solid TwoPhase Flow. J... 16. A fluid mechanical description of fluidized beds (equations of motion). Guj. R. W. (1995). Hydrodynamics. Dynamic simulation of dispersed gas. W.A. 307. Ind. J.M. New York. 191. 30. 357. John Wiley & Sons. Phys.A.E. and de Matteis.. Chem.liquid twophase flow using a discrete bubble model...M.. R.G.June 1. G.. Ge. Heat and Fluid Flow. Int. Conf. Statistical mechanics of dissipative particle dynamics. Europhys. Int. E. C. and Warren. (1997).. (1960). 325. J. (1986). (1997). and van Swaaij W. Numerical simulation of free convective flow using the latticeBoltzmann scheme. J. Bird. M. Beijing (China). Fluidparticles interaction in particulate fluidized beds: numerical and experimental analysis.. J.B. 145.M.
(1993). Int.. (1998).. Briels.M.fluidized bed. Lett.. Engng Sci. Kuipers. 363.B.P. van Duin K. W. Dynamics simulations of hardsphere suspensions under steady shear. (1992).P.P.. (1996). F. B. Engng Sci. W. Chem. J.A.A. (1992). 19. Algorithn for the simulation of particle suspensions with inertia effects.B. 2234. and Hermann.. Chem. B.A.J. Simulating microscopic hydrodynamic phenomena with dissipative particle dynamics.M. and Koelman. S. Durect simulation of flows of solidliquid mixtures. J. Kawaguchi. W.A.H.and threedimensional models)..P.P.A.J. Comments on the paper “Numerical simulation of the gassolid flow in a fluidized bed by combining discrete particle method with computational fluid dynamics” by B.P. A numerical model of gasfluidized beds. Phys. Kalthoff.J. Schwarzer. Hoomans. W. 21. 17. J. 839. Koelman. 155.M. W. P.M.J...A. Engng.H... J... and van Swaaij. 2645. van Beckum. W. Comput.. and van Swaaij.. Powder Technol. Kuipers. J. Chem. T. 53.V. Tanaka. P.. Engng Sci.J. J.. 1913.. Lett.M. Y. van Beckum. J. Kuipers. (1997). Europhys. 96. and Hoogerbrugge. Hu. E...J. 56. (1998). 96 . F. 129.H.Chapter 3 ______________________________________________________________________________________ Hoogerbrugge. Chem. Rev. van Duin K.B. and van Swaaij. 99. Numerical simulation of twodimensional fluidized beds using the discrete element method (comparison between the two.V. T. Hoomans.H. Yu. Xu and A.M. Europhys.. Kuipers.. Multiphase Flow. 51.M. 47. 335.M..P. Computer simulation of the hydrodynamics of a twodimensional gas..P. H. and Tsuji.. and van Swaaij.M. Briels.M.J. (1996). (1993). H. Discrete particle simulation of bubble and slug formation in a twodimensional gasfluidised bed: a hardsphere approach. W. 22..
Phys. 52. Phys.R.. Mikami. T. Fluid Mech. Y..C. A. (1995). Powder Technol.Gas Phase Hydrodynamics ______________________________________________________________________________________ Ladd.J. (1998). G. and Banerjee. Ouyang. and Horio.. Engng Sci. H. Kamiya. Tsuji. Kawaguchi. Numerical simulation of the gassolid flow in a fluidized bed by combining discrete particle method with computational fluid dynamics. T. and Zanetti. Rev. T. McNamara. Pan. J. Phys. and Yu. Sedimentation and flow through porous media: Simulating dynamically coupled discrete and continuum phase... G. Part 1. and Li.. Fluids. (1993). 79. Chem. Numerical simulations of particulate suspensions via a discretized Boltzmann equation. Lett. 1927.. (1995). Y. 185. E. 2332.. 8. 2785.H. Theoretical foundation. (1997).. Chem. Use of the Boltzmann equation to simulate laticegas automata. S. Numerical simulation of cohesive powder behavior in a fluidized bed.. Engng Sci. 53. (1988). Engng Sci... 77.motionresolved discrete model for simulating gassolid fluidization. J. (1994). 54. Xu. S. 2077. and Tanaka. 52. J. Schwarzer. 271. 61. Chem.. B. M. 97 . 2733. Particle. Discrete particle simulation of two dimensional fluidized bed. (1999).. Rev.. 6461..B. A. Numerical simulation of particle interactions with wall turbulence.
Chapter 3 ______________________________________________________________________________________ 98 .
The Root Mean Square (RMS) of the pressure fluctuations inside the bed showed an almost linear dependency on the energy dissipation rate by collisions during the simulation for low dissipation rates. The results of the threedimensional simulations confirmed the trends observed in the twodimensional simulations concerning the effect of the collision parameters on the bed hydrodynamics. 99 . ρ = 2700 kg/m 3. Pressure peaks inside the bed decreased with increasing geometric standard deviation of the lognormal size distribution. 0.15 m width. THE EFFECT OF PARTICLE PROPERTIES ON THE HYDRODYNAMICS OF GASFLUIDISED BEDS WITH HOMOGENOUS INFLOW CONDITIONS Abstract: A hardsphere Granular Dynamics model was used to study the influence of particle properties on the hydrodynamics of gasfluidised beds with homogeneous inflow conditions. The effect of a particle size distribution on the bed hydrodynamics was studied as well. The gas phase hydrodynamics is described by the spatially averaged NavierStokes equations for twophase flow. The particlewall interaction had a more pronounced influence on the hydrodynamics in the 3D simulations than in the 2D simulations.Effect of Particle Properties… ___________________________________________________________________________________ Chapter 4. ugas = 1. 2400 particles of 4 mm diameter. the coefficients of mf restitution and friction): the more energy is dissipated in collisions the more bubbling and slugging could be observed. In the model the Newtonian equations of motion are solved for each solid particle while taking into account the particle particle and particle wall collisions. Simulations with a threedimensional model (as far as the particles are concerned) required a minimum depth of the bed of five particle diameters for stable simulations. The hydrodynamics of a homogeneously fluidised bed (0.5 u ) was strongly affected by the collision parameters (i.5 m height.e. Time series of the pressure inside the bed showed a lower frequency and a lower RMS value in the 3D simulations.
The influence of a particle size distribution on the granular dynamics of dense gas fluidized beds: a computer simulation study.J.Chapter 4 ___________________________________________________________________________________ Parts of this chapter are based on the papers: B.A. 94. Kuipers and W. 318. No. Ser. Engng Sci. van Swaaij. Chem. Hoomans. Kuipers. UK. J.P. (1996). 99.M.B. Hoomans.M. Briels and W. J. B.B.M. Vol. Discrete particle simulation of bubble and slug formation in twodimensional gasfluidised beds: a hard sphere approach. Hoomans.M.P.M. (1998). July 79.P. Kuipers and W. 100 .P. 51.P.B. The influence of particle properties on pressure signals in dense gasfluidised beds: a computer simulation study. 15. J. B. Brighton.P.A. World Congress on Particle Technology 3. van Swaaij. (1998).A.. van Swaaij. AIChE Symp. W.M.
. 1995) and gasfluidised beds (Hoomans et al. (1993) to threedimensions as far as the motion of the particles is concerned. is circumvented. (1993) in their fluidised bed simulations but they also employed techniques from hardsphere simulations in order to determine the instant when the particles first come into contact precisely. In these models the Newtonian equations of motion are solved for each individual granular particle in the system. Introduction With increasing computer power Granular Dynamics simulations have become a very useful and versatile research tool to study the dynamics of dense gasparticle flows. inelastic collisions with friction. In this approach the particles are allowed to overlap slightly and this overlap is subsequently used to calculate the contact forces. The mutual interactions between particles and the interaction betwee n particles and walls are taken into account directly. if not impossible. Mikami et al. as encountered in continuum models (Kuipers et al. Furthermore a discrete particle model is very well suited to take an arbitrary particle size distribution into account which is far more complex. 1996) to hopper flow (Langston et al. (1996) used a hardsphere approach in their twodimensional discrete particle model of a gasfluidised bed which implies that the particles interact through binary. (1998) presented a twodimensional soft 101 .fluidised bed based on the work of Cundall and Strack (1977). In fluidisation simulations the Granular Dynamics models have a clear advantage over continuum models since the particle interactions are taken into account directly and consequently the necessity to specify closure relations for the solidsphase stress tensor. Xu and Yu (1997) used a twodimensional softsphere model similar to the one used by Tsuji et al.. Applications of this type of models are very diverse ranging from agglomerate dynamics (Thornton et al. Kawaguchi et al. Tsuji et al. instantaneous.Effect of Particle Properties… ___________________________________________________________________________________ 1. (1993) developed a twodimensional softsphere discrete particle model of a gas. in case a continuum modelling approach is adopted. Hoomans et al. (1998) extended the model developed by Tsuji et al.. 1996).. 1992 among many others).
1997). In Granular Dynamics simulations particleparticle and particlewall interactions are taken into account directly which is a clear advantage over continuum models. They used a onedimensional fiveparticlearray model to generate time series of particle position and voidage which is described in more detail by Schouten and van den Bleek (1992).e. Hoomans et al. Nieuwland et al. e = 0.9 and µ = 0. (1996)). Collision parameters like the coefficient of restitution can be included in continuum models by employing the kinetic theory of granular flow to arrive at improved closure relations for the solidsphase stress tensor (Sinclair and Jackson (1989).9) the bed dynamics became less chaotic.Chapter 4 ___________________________________________________________________________________ sphere model of a gasfluidised bed where cohesive forces between the particles were taken into account. In the case where no energy is dissipated (i. A more detailed discussion of the models mentioned above is presented in Chapter 2. do not take particle rotation into account and more importantly show an extreme and nonrealistic sensitivity of the overall bed hydrodynamics to very small deviations from one for the coefficient of restitution (Hrenya and Sinclair.e. Gidaspow (1994). these improved closures rely on a one parameter collision model. (1996) have shown that the dynamics of a fluidised bed depends strongly on the collision parameters ( i. They found that for in creasing values of the coefficient of restitution (up to 0. However.e. Van den Bleek and Schouten (1993) already showed that the coefficient of restitution has an effect on the dynamics of a fluidised bed. These time series were subsequently analysed by applying techniques from deterministic chaos theory. e = 1 and µ = 0) no bubble or slug formation was observed and the bed showed a rather homogeneous fluidisation behaviour. coefficients of restitution (e) and friction ( µ)). bubble formation was observed and the bed behaviour appeared to be in much better agreement with that observed in reality.3) and hence energy is dissipated in collisions. Within the framework of a discrete particle model the collision parameters are included in a more natural way and therefore this type of model is to be preferred when investigating the influence of particle parameters on the bed hydrodynamics. The coefficient of 102 . In the case where the impact parameters were given more realistic values (i.
Effect of Particle Properties… ___________________________________________________________________________________ restitution was varied over a range from 0.8 and 1.e. For the coated particles a completely different axial pressure profile was measured than for the regular glass particles.0 to 0. Although this system is operated in a completely different regime compared to those prevailing in the experiments mentioned above and the simulations presented in this chapter. For relatively shallow beds they found larger root mean square (RMS) values for the pressure fluctuations in beds consisting of particles with lower angles of internal friction. They also characterised the particle interaction through the angle of internal friction. This is a bulk property rather than a particle property but in general it can be assumed that the rougher the particles the lower the angle of internal friction. the experimental results obtained by Chang and Louge (1992) are worthwhile mentioning here.up in the riser section (and thus the axial pressure gradient) was a lot smaller. (1997a and 1997b) investigated the influence of solid frictional interactions on pressure fluctuations in fluidised beds. Chang and Louge (1992) studied the hydrodynamics in a riser where they applied a coating to glass particles in order to make them smoother and hence obtain a lower coefficient of friction. Experimental studies on the influence of particle interaction parameters on the dynamics of gasfluidised beds are unfortunately rather scarce. 1960). They characterised this interaction through the angle of internal friction (Zenz and Othmer. lower coefficient of friction) the solids hold. In the case of the coated particles (i. Chen et al. which is unfortunately not the most interesting range.0. However.9. Si ce the only n difference between the two experiments was the use of the coated particles instead of the regular glass particles their experiments indicate that particleparticle interaction has a significant effect on the flow structure in the riser. They also reported a negligible influence of the wall roughness on the RMS values of the measured pressure fluctuations. For example the Geldart (1974) classification of the fluidisation behaviour of powders does not include any particle interaction parameters. it does capture the essential feature that the collision parameters have a pronounced effect on the bed hydrodynamics. Typical values for this parameter of granular material generally used in fluidisation are found within 0. it is clear that despite the fact that this fiveparticle array model is a rather crude representation of the dynamics of a fluidised bed. Thiel and Potter (1977) reported an influence of the particleparticle interaction on the dynamics of slugging fluidised beds. 103 .
2. In the hardsphere approach a sequence of binary collisions is processed. These three collision parameters are all included in the model. (1994) have shown that also the coefficient of tangential restitution (0 ≤ β 0 ≤ 1) should be used in order to describe the collision dynamics more accurately. The majority of the simulations was performed with the twodimensional model. instantaneous. the key features will be summarised briefly here. Furthermore the model was used to study the influence of a particle size distribution on the hydrodynamics of gas.Chapter 4 ___________________________________________________________________________________ In this chapter the dependency of the hydrodynamics of a gas. Foerster et al. but also simulations with the threedimensional model were performed.fluidised bed with homogeneous inflow conditions on the collision parameters is investigated in more detail using the hardsphere model. For each particle a list of neighbouring particles is stored and only for the particles in this list a check for possible collisions is performed. 2. Model Since a detailed description of the model is presented in Chapter 2 and Chapter 3. A constant time step is used to take the external forces into account and within this time step the prevailing collisions are processed sequentially.fluidised beds with homogeneous inflow conditions. 104 . In order to reduce the required CPU time neighbour lists are used. This implies that a collision list is compiled in which for each particle a collision partner and a corresponding collision time is stored.1 Granular dynamics The collision model as originally developed by Wang and Mason (1992) is used to describe a binary. inelastic collision with friction. The key parameters of the model are the coefficient of restitution (0 ≤ e ≤ 1) and the coefficient of friction (µ ≥ 0).
fluid model presented by Kuipers et al. µg the viscosity of the gas and ρ g the density of the gas.75(1 − ε ) ρg Dp u − vp . of course. For high void fractions (ε ≥ 0.1) the first term is due to gravity and the third term is the force due to the pressure gradient.3) The drag coefficient Cd is a function of the particle Reynolds number and given by: 105 . 4 Dp β= (4.65 . the forces now act on a single particle: m p dv Vβ = mg+ (u − v dt (1 − ε ) p p p p ) − V ∇p p .2) where Dp represents the particle diameter.1) where m p represents the mass of a particle. (1992) where. In equation (4. u the local gas velocity and Vp the volume of a particle. The second term is due to the drag force where β represents an interphase momentum exchange coefficient as it usually appears in twofluid models.80) the following expression for the interphase momentum transfer coefficient has been used which is basically the correlation presented by Wen and Yu (1966) who extended the work of Richardson and Zaki (1954): 3 ε (1 − ε ) Cd ρ g u − v p ε − 2 .Effect of Particle Properties… ___________________________________________________________________________________ 2. (1996). In this work the external forces are used in accordance with those implemented in the two.2 External forces The incorporation of external forces differs somewhat from the approach followed by Hoomans et al. (4.80) β is obtained from the wellknown Ergun equation: β = 150 (1 − ε )2 ε µg D2 p + 1 . v p its velocity. (4. For low void fractions (ε < 0.
(4.15Re p ) Cd = p 0.6) Momentum equation gas phase: ∂ ερg u ∂t ( ) + (∇ ⋅ ερ uu) = − ε∇p − S − ( ∇ ⋅ ετ ) + ερ g .7) 106 . (1992).3 Gas phase hydrodynamics The calculation of the gas phase hydrodynamics mainly follows the lines presented by Kuipers et al.5) For the integration of equation (4. ∂t g (4. Continuity equation gas phase: ∂ ερ g ( ) + (∇ ⋅ ερ u) = 0 . It is based on the numerical solution of the following set of partial differential equations that can be seen as a generalised form of the NavierStokes equations for a gas interacting with a solid phase as originally derived by Anderson and Jackson (1967).44 Re p < 1000 . 2.1) a simple explicit first order scheme was used to update the velocities and positions of the particles. (4.687 Re (1 + 0.4) Re p ≥ 1000 where the particle Reynolds number (Re p ) in this case is defined as follows: ε ρg u − v p D p µg Re p = . g p g g (4.Chapter 4 ___________________________________________________________________________________ 24 0.
2 m above the centre of the bottom plate was monitored. As a test system the same system as used before by Tsuji et al. The constitutive equations can be found in Chapter 3. isothermal (T = 293 K) flow of air at atmospheric conditions is considered. (1996) and Xu and Yu (1997) was chosen. Hoomans et al. Table 4. Dp Material Density. The simulations were continued for 8 seconds real time and during the simulations the pressure in the bed 0.1. (1996) and that deals with the way in which the twoway coupling between the gasphase and the dispersed particles is established. (1993). Effects of collision parameters A large number of simulations was performed in order to investigate the influence of the collision properties on the bed dynamics. In the present model the reaction force to the drag force exerted on a particle per unit of volume is fed back to the gas phase through the source term S p which has the dimension of force per unit 3 of volume N/m . 107 . The most important system parameters are summarised in table 4.Effect of Particle Properties… ___________________________________________________________________________________ In this work transient.78 m/s).1. ρ Number.15 m 0. There is one important modification with respect to the model presented by Hoomans et al. N part Parameter settings for all simulations Bed Spherical 4.5 um f for the whole width of the bed (umf = 1. Particles Shape Diameter. 3.50 m 15 25 In all simulations the minimum fluidisation conditions were used as initial conditions while the values for the coefficients of restitution (e) and friction (µ) were varied.0 mm Aluminium 2700 kg/m3 2400 Width Height Number of xcells Number of ycells 0. twodimensional. A time step of 10 4 seconds was used in all simulations and the gas inflow at the bottom was set equal to 1.
Chapter 4 ___________________________________________________________________________________
Figure 4.1 . Snapshots of particle configurations for three simulations with ug = 1.5 um f. Upper: e = 1, µ = 0, middle: e = 0.99, µ = 0, bottom: e = 0.9, µ = 0.3.
108
Effect of Particle Properties… ___________________________________________________________________________________
A typical (no n ideal) run takes about 6 minutes of CPUtime per simulated second on a Silicon Graphics Origin200 server (R10000 processor, 180 MHz). Snapshots of the particle configurations for three different simulations are presented in Figure 4.1. The two extreme cases (e = 1.0, µ = 0.0 (ideal) and e = 0.9, µ = 0.3 (nonideal)) are shown there together with a simulation where e = 0.99 and µ = 0.0. A number of particle layers was marked by colour at t = 0 s in order to visualise t he induced particle mixing during the fluidisation process. It can be seen that in the ideal case barely any bubbles or slugs are being formed which enables the initial structure to remain intact for quite some time. In the non ideal case the formation of bubbles and slugs can be observed clearly and due to these instabilities the particle mixing in the bed is much better. After 2 seconds the initial structure can still be observed in the ideal case whereas it has completely disappeared in the nonideal case. The simulation with e = 0.99 resembles the ideal case very much although the particle mixing appears to be slightly better. This is a much more natural dependency than the one reported for instance by Hrenya and Sinclair (1997). They reported simulation of riser flow using a twofluid s model that incorporates the kinetic theory for granular flow a change in e from 1.0 to 0.99 completely changed the hydrodynamic behaviour.
e e e
µ µ µ
Figure 4.2 .
Pressure at 0.2 m above the centre of the distributor plate as a function of time.
109
Chapter 4 ___________________________________________________________________________________
However, it should be noted that these simulations were concerned with a different hydrodynamic regime than encountered in the simulations reported here and therefore a direct comparison may not be justified.
The computed pressure signals at 0.2 m above the centre of the bottom plate for the three simulations are presented in Figure 4.2. It can be observed that the pressure fluctuations in the non ideal case are much stronger than in the ideal case. In fact the pressure fluctuations turn out to depend strongly on the collision parameters. In Figure 4.3 the root mean square (RMS) of the pressure signals is presented as a function of the coefficient of restitution (e).
e
Figure 4.3 .
Root Mean Square (RMS) of the pressure fluctuations as a function of the coefficient of restitution.
It can clearly be observed that the RMS increases with decreasing value of e. The fact that, for instance, the RMS value for the simulation with e = 0.96 is higher than that for the simulation with e = 0.95 is due to the different dynamics. Despite of the lower value for the coefficient of restitution a higher amount of energy is dissipated in collisions during the simulation. Because the total number of collisions in these simulations is about the same (8.3 million) the average impact velocity was apparently higher in the case of e = 0.96. With a slightly different initial condition this might change which is due the complex (i.e. chaotic) nature of the system’s dynamics. In all the simulations that were performed in order to obtain the RMS values presented in
110
Effect of Particle Properties… ___________________________________________________________________________________
Figure 4.3, the coefficient of friction was set equal to zero. In the calculation of the RMS the initial stages of the simulation (approximately 2 s) were not taken into account in order to avoid startup effects from influencing the results.
In Figure 4.4 the R MS is presented as a function of the value of the coefficient of friction (µ ). For all the simulations that were performed in order to obtain the RMS values presented in Figure 4.4 the coefficient of restitution (e) was set equal to 1.0.
µ
Figure 4.4 .
Root Mean Square (RMS) of pressure fluctuations as a function of the coefficient of friction.
Here it can be observed that the RMS increases with increasing value for the coefficient of friction. In Figure 4.4 the value for the RMS is lower for the simulation with µ = 0.25 than for the simulations with µ = 0.2 which is not expected. However the amount of energy dissipated in collisions for both simulations is no t the same; in fact more energy is dissipated in the simulation with µ = 0.2 than in the simulation with µ = 0.25. In the latter case fewer collisions occurred and apparently the collisions that occurred had less impact. Also the influence of the coefficient of tangential restitution (β 0 ) was investigated. However, this parameter turned out to have little influence on the overall bed hydrodynamics. This is probably due to the low (but realistic) values used for the coefficient of friction (µ). When the coefficient of friction was increased (µ = 1000)
111
Chapter 4 ___________________________________________________________________________________
to ensure that all collisions were of the sticking type the influence of β 0 was much more pronounced. With e =1.0, µ = 1000 and β 0 = 1.0 similar results were obtained as for the case with e = 1.0, µ = 0.0 and β0 = 0.0. However, the influence of the coefficient of tangential restitution for realistic values of the coefficient of friction (µ < 1.0) was negligible.
In all simulations mentioned above the values of the collision parameters for particle wall collisions were taken to be equal to the values for particle particle collisions. The influence of the particle wall interaction on the overall bed hydrodynamics was investigated as well. This influence was found to be negligible. The particle particle interaction is very much dominant over the particlewall interaction in the 2D model used here. A simulation with ideal particleparticle interaction (e = 1.0, µ = 0.0) but nonideal particle wall interaction (ew = 0.9 , µw = 0.3) showed very similar results as in the fully ideal case.
In Figure 4.5 the RMS of the pressure is plotted as a function of the energy dissipation rate by collisions during the simulation. All simulations performed in this study are included in this figure.
Figure 4.5 .
Root Mean Square (RMS) of pressure fluctuations as a function of the energy dissipation rate by collisions during the simulation.
112
11) All the five different energies mentioned above are calculated independently.k ∇p ⋅ v k dt . The energy dissipated in collisions is calculated using the equations presented in Chapter 2. kinetic energy. k .e. 4.Effect of Particle Properties… ___________________________________________________________________________________ It can be seen that the RMS increases almost linearly with the energy dissipation rate by collisions in the system. lower coefficients of restitution and/or higher coefficients of friction) the data points show some scatter. potential energy and rotation energy) can be calculated (instantaneously) as follows: E kin 1 Npart 2 = ∑ mk vk .10) E pot = ∑m k k gry . E rot 1 = 2 Npart ∑I k k ω k2 . The energies concerned with the particle positions and velocities (i. more vigorous bubbling and slugging which causes a more irregular pressure signal) at higher dissipation rates which requires longer simulation times than the 8 seconds employed here in order to obtain better statistics. Energy conservation Since the amount of energy dissipated in collisions turns out to have a profound influence on the bed hydrodynamics the energy housekeeping during a simulation was investigated in more detail. 113 . (4.e.9) (4. This dissipated energy is taken to be positive by definition. For higher values of the dissipation rate (i. Since the dissipated energy is taken to be positive by definition the sum of the potential. k k k p . This is due to the richer dynamics of the system (i.e. 2 k Npart (4.8) (4. The energy that the particles receive from the interaction with the gas phase can be obtained as follows: Edrag = Npart V β ∑ ∫ (1 − ε ) (u − v ) − V p.
7 are the values of the collision parameters. and this amount should remain constant during a simulation.7 the energy housekeeping during a simulation with ideal collision parameters (e = 1.0) is presented.3) is presented. 114 . Figure 4.6. µ = 0.9.6 and 4.9.0. The total amount of energy in the system can now be obtained as follows: E tot = E pot + E kin + E rot + E disp − E drag . Energy housekeeping for a simulation with nonideal collision parameter settings (e = 0. rotation and dissipated energy should equal the amount of energy that the particles receive from the interaction with the gas phase.3) with a homogenous gasinflow velocity ugas = 1. µ = 0. The energy dissipated in collisions increases almost linearly with time and its behaviour is closely followed by the energy supplied by the drag force.12) In Figure 4.6 the energy housekeeping during a simulation with nonideal collision parameters (e = 0.5 um f. In Figure 4.Chapter 4 ___________________________________________________________________________________ kinetic. (4. It should be stressed that the only differences between the system in Figure 4. The rotation energy is about 2 orders of magnitude smaller than the potential energy and therefore appears to be zero in this figure. It can be observed that the potential energy is larger than the kinetic energy. µ = 0.
The energy required for these stronger fluctuations is supplied by the drag force which puts a lot more energy into the system in the nonideal case (Figure 4.Effect of Particle Properties… ___________________________________________________________________________________ Figure 4.5 um f.7.0.6) than in the ideal case (Figure 4.ideal and the ideal cases the total energy remains constant throughout the whole simulation as required. µ = 0. The energy put into the system by drag force remains constant after a short initial stage (2 seconds).6) fluctuates more strongly than in the ideal case (Figure 4. Due to this vigorous bubbling and slugging the pressure fluctuations in the bed are stronger than in the ideal case as reported in the previous section.7 it can be observed that the amount of energy dissipated in collisions remains zero throughout the whole simulation. This is the amount of energy required for the expansion of the bed.7). It can also be observed that the potential energy in the nonideal case (Figure 4.7) which is obviously due to the bubbling and slugging in the non ideal case.0) with a homogenous gasinflow velocity ug = 1. Energy housekeeping for a simulation with ideal collision parameter settings (e = 1. 115 . In Figure 4. Note that in both the non.
(1997)).CM )2 df = exp − d D p . When creating the particle size distribution all the diameters which are smaller than Dp. In equation 4.13) where df is the fraction of particles having diameters whose logarithms lie between ln Dp and ln Dp + d(ln Dp ).0 mm and a geometric standard deviation (GSD or σ g) of 0. 116 .CM is the count median diameter and σg is the geometric standard deviation. most of them are skewed to larger diameters (Seville et al.σg or larger than Dp. Figure 4.8. In this work the particle diameters are obtained from a lognormal distribution which is asymmetrical and can be represented in mathematical terms as follows: (ln D p − ln D p . 2 2π D p ln σ g 2 (ln σ g ) 1 (4.13 Dp.1 mm. Frequency histogram of a lognormal particle size distribution with a count median diameter of 4.CM . Influence of a particle size distribution Hardly any distribution of particle size encountered in fluidisation studies is symmetrical.σg are rejected to mimic the effects of sieving.8 . An example of a particle size distribution generated with this method is represented in a (discrete) frequency histogram in Figure 4.Chapter 4 ___________________________________________________________________________________ 5.CM . A symmetrical size distribution like a normal or a Gaussian distribution is therefore not representative for particles used in laboratory or industrial practice.
This can be explained by the lower void fraction in the uniform case due to the closer packing which results in a higher force acting on 117 . 0. GSD = Geometric Standard Deviation (σg ).2 m above the centre of the bottom plate are presented for all the four cases.2 m above the centre of the distributor plate as a function of time for the four cases. A time step of 10 4 s was used and all simulations were run for 10 s. Simulations were performed for a system consisting of particles of uniform size and systems consisting of particles with a log normal size distribution with a geometric standard devia tion σg = 0.9 . Pressure at 0. A homogeneous gas inflow at 1.1 where the particle diameter of 4 mm is now the mean particle diameter.5 u mf was specified at the bottom of the system. Simulations were performed using the parameter settings presented in table 4. The coefficient of restitution (e) was set equal to 0. In Figure 4.3 for both particle particle and particlewall collisions. It can be observed that the pressure peaks decrease with increasing geometric standard deviation of the size distribution. The initial configurations were obtained by placing the particles in the system and allowing them to fall under the influence of gravity while the homogeneous gas inflow was set equal to u mf.5 and 1.0 mm respectively.9 the pressure fluctuations inside the bed at 0.9 and the coefficient of friction (µ) was set equal to 0. Figure 4.Effect of Particle Properties… ___________________________________________________________________________________ The influence of the distribution width on the bed hydrodynamics at homogeneous inflow conditions was studied.1.
After the first 1.Chapter 4 ___________________________________________________________________________________ the particles which in turn causes the higher pressure peaks. 3D simulations A threedimensional version of the hardsphere model was developed. such as at minimum fluidisation conditions. Schematic representation of the 3D bed. (1998). 6. it is important to take polydispersity into account.10. All other parameter settings are the same as used for the twodimensional system presented in table 4. the gas phase hydrodynamics is still resolved in two dimensions. This can be justified since the simulated system is a flat fluidised bed with a depth of 22 mm where the motion of the gas phase in the third dimension can be neglected.15 m Figure 4. The 3D bed is schematically represented in Figure 4. Eventually the simulation crashes due to particle overlap. Similar to the model presented by Kawaguchi et al. In preliminary tests it became clear that a minimum depth of 5 particle diameters was required to ensure stable simulations. With a depth 118 .1.5 m 0. 22 mm 0.10. At smaller depths the particles get stuck between the front and back walls and the hardsphere algorithm breaks down.5 s the differences become less pronounced which indicates that especially in situations where close packing can occur.
 119 . In the ideal case no bubbles are observed and after a vigorous expansion the bed remains rather homogeneously fluid ised.5 um f. A simulation with nonideal collision parameters (e = ew = 0. In Figure 4. µ = µw = 0.ideal case bubbles are observed and the bed height is continuously fluctuating. but this fluctuating motion could clearly be observed in animatio ns.0. µw = 0. µ = µw = 0.0).1 Influence of collision parameters The collision parameters turned out to have a key influence in the twodimensional simulations and therefore their influence was also investigated using the threedimensional model. In the non.3) was performed as well as a simulation with ideal collision parameters (e = ew = 1. 6. The trends observed in the 2D simulations are clearly confirmed in the 3D simulations. The bed was homogeneously fluidised with a gas inflow of 1. In the snapshots this may not be very clear due to the visualisation technique employed here. µ = 0.11 snapshots obtained from three different 3D simulations are presented.1.000 in order to match initial bed height in the 2D simulations.9. A simulation with ideal particleparticle interaction but nonideal particlewall interaction (e = 1.5 particle diameters) such problems did not occur and therefore this depth was used in the simulations. The setup of the 3D simulations was the same as for the 2 D simulations. For the 3D case it should be noted that the bed is 5. A number of particle layers was marked by colour based on the position at t = 0 s in order to visualise the induced particle mixing during the fluidisation process.3) was also carried out.0. Also the potential energy of the system showed these fluctuations in the non ideal case.5 particle diameters deep and this affects the visualisation: the particles near the front wall dominate the image.0.Effect of Particle Properties… ___________________________________________________________________________________ of 22 mm (5. ew = 0.9. The number of particles used in the 3D simulations was 14. The snapshots presented in this figure are front views similar to the snapshots of the 2D simulations in Figure 4.
ew = 0.Chapter 4 ___________________________________________________________________________________ Figure 4. middle: e = 1.0. Upper: e = ew = 1. µw = 0.0.11.0. 120 .0.9.3.3 bottom: e = ew = 0. µ = µw = 0. µ = µw = 0.9. µ = 0.5 um f. Snapshots of particle configurations for three 3D simulations with ug = 1.
2 2D versus 3D The snapshots of the 3D nonideal simulation in Figure 4.1.0.11 show less bubbling and slugging than the snapshots in the 2D equivalent in Figure 4.Effect of Particle Properties… ___________________________________________________________________________________ The particlewall interaction has a significant influence on the hydrodynamics. This can be partially due to the visualisation technique employed here but also the pressure fluctuations inside the bed are less strong and have a lower frequency (2. In Figure 4. 6. the fluctuations in the 2D case are more pronounced compared to the 3D case.12 the pressure 0. Figure 4. µ = µw = 0. 3.3 Hz in 2D for the nonideal case). Pressure 0.2 m above the centre of the distributor plate is plotted as a function of time for the 2D and 3D simulation with ideal values for the collision parameters. In the snapshots this can be observed especially during the initial stages of the simulation. The 121 .3 Hz in 3D vs.12.2 m above the centre of the distributor plate as a function of time (e = ew = 1.0): 2D simulations compared to 3D simulations Although in both the 2D and the 3D simulations the fluctuations are not very strong. Also the potential energy of this system shows stronger fluctuations than the ideal system.
Also the absence of a degree of freedom of motion in the 2D case can cause the system to show a more constrained behaviour and hence cause stronger fluctuations.ideal case.0. µ = 0.2 m above the centre of the distributor plate as a function of time (ew = 0.9. µw = 0.0. Figure 4.Chapter 4 ___________________________________________________________________________________ RMS value is about 70 Pa higher for the 2 simulation. µw = 0. In Figure 4. The RMS value of D the fluctuations was 60 Pa higher in the 3D case.3). µ = 0. Pressure 0.13 the pressure fluctuations in the 2D and 3D simulations are compared for the case with ideal particle particle interaction and non ideal particlewall interaction (e = 1.9. For this case the fluctuations are stronger in the 3 simulation.3): 2D simulations compared to 3D simulations. ew = 0. These differences may be explained in terms of improved statistics since the number of particles in a computational cell is much higher in the 3D case. The same difference was D found for the non.0. Baring in mind that in the other cases the fluctuations in the 2D simulations were stronger than in the 3D simulations it can be concluded that the particle wall interaction has a more pronounced effect on the hydrodynamics in 3D simulations than in 2D simulations.13. This is obviously due to the presence of the front and the back wall in the 3D simulations of the ‘flat’ bed 122 .
The higher the energy dissipation rate the more energy was put into the system by the drag force.Effect of Particle Properties… ___________________________________________________________________________________ which renders a much higher number of particlewall collisions than in the 2D simulations. It takes 6 minutes of CPU time per simulated second on a Silicon Graphics Origin200 server (R10000 processor. Conclusions A hardsphere Granular Dynamics model was used to study the effect of the particle properties on the hydrodynamics of gasfluidised beds with homogeneous inflow conditions. 7. When realistic particle interaction was assumed (e = 0. The 3D simulations are much more computational intensive than the 2D simulations.0. The particle particle interaction parameters turned out to be much more important than the particlewall interaction parameters in the twodimensional simulations.e. With intermediate values for the collision parameters the behaviour also changes gradually with pressure fluctuations becoming less strong when the coefficients are changed toward their ideal values.000 particles). 180 MHz) with the 2D model (2400 particles) whereas it takes 220 minutes per second to perform the same (nonideal) simulation with the 3D model (14. It turned out to be more convenient to determine the root mean square (RMS) of the pressure fluctuations as a function of the energy dissipation rate since this is affected both by the collision parameters and the system dynamics. no energy is dissipated in collisions) e nonrealistic bed dynamics is observed: no bubble or slug formation occurs and particle mixing is poor.3) bubble and slug formation are observed and the fluidisation behaviour is in much better accordance with that inferred from experimental observations.0. 123 . The bed dynamics is strongly affected by the values of the collision parameters (i.9. The RMS values of the fluctuations show an almost linear increase with the energy dissipation rate during a simulation. µ = 0. Pressure fluctuations inside the bed are much stronger in the nonideal case than in the ideal case. i.e. The coefficient of tangential restitution (β 0 ) has little influence on the overall behaviour. When ideal particle interaction was assumed ( = 1. the coefficients of restitution (e) and friction ( µ)). µ = 0.
The main influence of the polydispersity could be observed when the particles were in a rather close packing. s 124 . kg pressure. This indicates that it is especially important to take a particle size distribution into account in sy stems where locally very dense particle configurations can occur. The 3 model confirmed the trends observed in the 2 model with D D respect to the dependency of the overall bed hydrodynamics on the collision parameters. m/s2 particle mass. m momentum source term N/m3 temperature. Pressure peaks inside a slugging fluidised bed decreased with increasing geometric standard deviation of the lognormal size distribution during the initial stages of a simulation starting from minimum fluidisation conditions.Chapter 4 ___________________________________________________________________________________ Twodimensional simulations have been performed with particles which diameters have been obtained from a log normal distribution. Pa position vector. [] particle diameter. m gravitational acceleration. [] coefficient of restitution. As anticipated the 3D simulations required considerably more CPU time than the 2D simulations. Also a threedimensional model (as far as the particles are concerned) was developed. Notation Cd e Dp g mp p r Sp T t drag coefficient. Time series of the pressure inside the bed showed a lower frequency and a lower RMS value in the 3D simulations than in the 2 case. A minimum depth of the bed of 5 particle diameters was required to ensure stable simulations. K time. The particlewall D interaction had a more pronounced influence on the hydrodynamics in the 3D simulations than in the 2D simulations.
T. Eng. kg/ms gas phase stress tensor. Fundam. R... 125 . [] void fraction. and Jackson.Effect of Particle Properties… ___________________________________________________________________________________ u vp Vp gas velocity vector. m Greek symbols β β0 ε µ µg τ ρ σg 3 volumetric interphase momentum transfer coefficient. m/s 3 particle volume. kg/ms2 density. kg/(m s) coefficient of tangential restitution. kg/m3 Geometric Standard Deviation (GSD). 6. m/s particle velocity vector. Chem. A fluid mechanical description of fluidized beds (equations of motion). (1967). Ind. [] gas viscosity. m Subscripts CM dsp drg GSD g kin mf p pot rot tot w Count Median dissipated drag Geometric Standard Deviation gas phase kinetic minimum fluidisation partic le potential rotation total wall References Anderson. [] coefficient of friction.. 527.B.
Chem. Chem. J. 55. Effects of particle phase turbule nce in gassolid flows..U. Engng Sci. Chem. 853.. (1994).H. 259.. 7 .. Types of gas fluidization. 285.D. Louge. H.L. Boston. A numerical model of gasfluidized beds. Di Felice. W. Engng Sci. Z. A discrete numerical model for granular assemblies. (1997b).. and Allia. Discrete particle simulation of bubble and slug formation in a twodimensional gasfluidised bed: a hardsphere approach. (1979).A.M. C.Chapter 4 ___________________________________________________________________________________ Bleek. K. Effects of solid frictional interactions on the applicability of the fluid dynamic scaling rules for fluidisation. 6 . Chang. M. Engng Sci.. 1108. (1994). Chen. 52. J.M.Y. and van Swaaij. 70. Géotechnique. Gibilaro.. W. Briels. (1993).P.M.. D. 48 .. P. Measurements of the collision properties of small spheres.. B.F. Z. (1992).. 99..P. Can deterministic chaos create order in fluidizedbeds scaleup ?. AIChE J.M. J..L.U. (1992).G.. O.Y.M..G.... (1997)... Powder Technol. P.J.B. Eng. D. Chem.P. Hoomans. Engng Sci. Multiphase Flow and Fluidization. Kuipers. (1997a).P. 43 . Geldart.A. Hrenya.A. 47.. USA.J. P. Cundall. R. Comm.H. and van Swaaij. and Foscolo.C. Fluid pressure loss in slugging fluidised beds. Chen.. (1996). 2367. and Louge. F. C. Kuipers. Gidaspow.M. J. 51 . and Schouten. Phys... Powder Technol. Chang.. 29. (1973). van den.. van Beckum. W. 1913 126 . Fluids.. Foerster S. M. 47.. and Foscolo. Gibilaro. L. 205. 157. Academic Press. L. and Sinclair. Chem. Fluid dynamic similarity of circulating fluidized beds. and Strack.. van Duin K.
1009. Slugging in Fluidized beds. Processing of Particulate Solids. Phys. van den. and Zaki.E.. 53 . AIChE J. Gasparticle flow in a vertical pipe with particleparticle interactions. Numerical simulation of the impact fracture and fragmentation of agglomerates. 88 ... Kamiya. Ind. 1473. J.M.M. and R. Schouten. 424.Effect of Particle Properties… ___________________________________________________________________________________ Langston.. Mikami. Chem. J. Eng. Powder Technol. Tüzün. W. and Bleek... 37 . Eng. Hydrodynamic modelling of gasparticle flows in riser reactors. K. U.P. No. (1954). and M. 967. and Potter. W. J. Chaotic hydrodynamics of fluidization: consequences for scaling and modeling of fluid bed reactors. 35. (1997). Kawaguchi. Tanaka. Fundam. Discrete element simulation of granular flow in 2D and 3D hoppers: dependence of discharge rate and wall stress on particle interactions.. Clift. Richardson.. (1998).. M. 1996. 29. Sedimentation and fluidization: part I. Engng Sci. D: Appl. (1996). Inst. C. C.. 16. J.. and Tsuji. Phys. 129. Adams. and van Swaaij. J. J. Nieuwland. T. Y. K.. and Heyes.and threedimensional models).. T.. 70.M. Blackie Academic & Professional.P. 35... and Jackson.. 50 .. AIChE J. 127 . W..A.N. Seville. R. AIChE Symp. M. (1977). Numerical simulation of twodimensional fluidized beds using the discrete element method (comparison between the two. Thornton. 289 Vol. J.L. Ser. (1989). Kuipers.. 242. Chem.K.J. O. P. Chem.. 1927.. D.M. J. Chem. van Sint Annaland. Yin. 96. (1992). Engng Sci.. Numerical simulation of cohesive powder behavior in a fluidized bed.A. (1998). Tüzün. Trans. 32 .F. Thiel.J. and Horio. U. H. (1995). Sinclair. T.C.
A. Symp. 635. F. Twodimensional rigidbody collisions with friction.Y. Zenz. Fluidisation and fluidparticle systems. 77. Appl... T. J. 52. (1992). Y.F.. 128 . 79. Wen.H. M. Y. 2785..T. Prog.. (1960). Eng. USA. and Yu. Reinhold Publishing Corp. Ser. 100.. Kawaguchi. and Yu. Mech. T. (1993). 59. Xu. (1997).Chapter 4 ___________________________________________________________________________________ Tsuji. and Mason. B. Y. Numerical simulation of the gassolid flow in a fluidized bed by combining discrete particle method with computational fluid dynamics. (1966). and Tanaka.H. Discrete particle simulation of two dimensional fluidized bed..B. Powder Technol. 62 (62).A. Chem. Wang. and Othmer D. Chem. Mechanics of fluidization. C. New York.. Engng Sci.
Hardly any difference could be observed between the results of the hardsphere and the softsphere simulations that both compared well with the experiment. Bubble formation at a central jet in a bed of 0. 129 . The agreement between simulation and experiment was improved by incorporation of a lognormal particle size distribution. SoftSphere ______________________________________________________________________________________ Chapter 5. When the collisions were assumed to be fully elastic and perfectly smooth the agreement with the experiment was much worse.HardSphere vs.2 m width and 0. In the hardsphere simulations a sequence of binary. The gas phase hydrodynamics is described by the spatially averaged NavierStokes equations for twophase flow. The collision parameters turned out to have a key influence in both types of simulations. SOFTSPHERE APPROACH Abstract: Two types of Granular Dynamics models have been developed to study bubble formation in a gasfluidised bed. Preliminary simulations with the softsphere model revealed that the results are rather insensitive to the time step used for the integration of the contact forces as well as the value of the spring stiffness. GRANULAR DYNAMICS SIMULATION OF BUBBLE FORMATION IN A GAS FLUIDISED BED: HARDSPHERE VS. gravity and drag). The computational efficiency of the hardsphere simulations depends strongly on the dynamics of the system in contrast to the computational efficiency of the softsphere simulations. In the softsphere model the particleparticle and particlewall interactions are taken into account through a linear spring/dashpot contact force model.3 m height with 40.e. instantaneous collisions is processed one collision at a time while a collision model is used to describe the inelastic collisions with friction. For each individual particle the Newtonian equations of motion are solved taking into account the external forces (i.000 glass ballotini particles of 850 µm diameter was chosen as a test case for the simulations.
Ser.S. Engng Sci.P. Hoomans. Discrete particle simulation of bubble and slug formation in twodimensional gasfluidised beds: a hardsphere approach. 130 . J. 51.M. Kuipers. AIChE Symp. Vol.M. J. No. Fan and T. L. van Swaaij. B. Hydrodynamic models of gasfluidized beds and their role for design and operation of fluidized bed chemical reactors. Chem. Kuipers.A. Hoomans and W.M.A. 94.M. 318.Chapter 5 ______________________________________________________________________________________ Parts of this chapter are based on the papers: B.B. 15.B.P. Hoomans. Kuipers and W.P.P.P.M.B. The influence of a particle size distribution on the granular dynamics of dense gas fluidized beds: a computer simulation study.P. van Swaaij. 15. J.. 99. (1996).M. (1998). in Fluidization IX. W. B. Knowlton (eds).J.A. (1998). Briels and W. van Swaaij.M.
HardSphere vs. (1993) to threedimensions as far as the motion of the particles is concerned. SoftSphere ______________________________________________________________________________________ 1. (1993) in their fluidised bed simulations but they also employed techniques from hardsphere simulations in order to determine the instant when the particles first come into contact precisely.fluid models which require closure relations for the solidsphase stress tensor (Gidaspow (1994). (1992) among others). Tsuji et al. In these simulations the value of the 131 . if not impossible. inelastic collisions with friction. Xu and Yu (1997) used a twodimensional softsphere model similar to the one used by Tsuji et al.fluidised bed which implies that the particles interact through binary. (1993) was extended in order to take cohesive forces between the particles into account. Introduction Due to increasing computer power Granular Dynamics simulations have become a very useful and versatile research tool to study the hydrodynamics of gasfluidised beds. In these simulations the Newtonian equations of motion for each individual particle in the system are solved. in case a continuum modelling approach is adopted. Furthermore a discrete particle model is very well suited to take an arbitrary particle size distribution into account which is far more complex. In their approach the particles are allowed to overlap slightly and from this overlap the contact forces are subsequently calculated. Sinclair and Jackson (1989) and Kuipers et al. (1998) extended the model developed by Tsuji et al. Mikami et al. Kawaguchi et al. Particleparticle and particlewall interactions are taken into account directly which is a clear advantage over two. (1996) used a hardsphere approach in their discrete particle model of a gas. Hoomans et al. instantaneous. (1993) developed a softsphere discrete particle model of a gasfluidised bed based on the work of Cundall and Strack (1977). In all the softsphere models used for fluidised bed simulations a linearspring/dashpot model is used to describe the normal contact forces. (1998) presented a twodimensional softsphere model of a gasfluidised bed where the model developed by Tsuji et al.
For each 132 .. the key features will be summarised only briefly here. instantaneous. The key parameters of the model are the coefficient of restitution (0 ≤ e ≤ 1) and the coefficient of friction (µ ≥ 0). a softsphere linear spring/dashpot model was developed. 2. inelastic collision with friction. Foerster et al. A constant time step is used to take the external forces into account and within this time step the prevailing collisions are processed sequentially. This implies that a collision list is compiled in which for each particle a collision partner and a corresponding collision time is stored. In order to reduce the required CPU time neighbour lists are used. 1996) that allows for validation. These three collision parameters are all included in the model.1 Hardsphere granular dynamics The collision model as originally developed by Wang and Mason (1992) is used to describe a binary.fluidised bed was chosen as a test case because of the availability of a well defined experiment (Hoomans et al. In the hardsphere approach a sequence of binary collisions is processed. 2. The formation of a single bubble at a central orifice in a gas. (1994) have shown that also the coefficient of tangential restitution (0 ≤ β0 ≤ 1) should be used in order to describe the collision dynamics more accurately. Models Since a detailed description of the models is presented in Chapter 2 and Chapter 3. It is assumed that this is allowed but it has never been demonstrated that the simulation results are not affected by the value of the spring stiffness.Chapter 5 ______________________________________________________________________________________ spring stiffness is chosen artificially low for computational convenience. In order to investigate the effect of the spring stiffness on the simulation results and to allow for a comparison with hardsphere simulations. Since for both types of models the gas phase hydrodynamics and the twoway coupling was incorporated in exactly the same manner it is possible to make a direct comparison that highlights only the effects due to the choice of model.
(1998) and Mikami et al.4) If however the following relation is satisfied: 133 . dt 2 dω = T.2) where r is the position vector of the centre of the particle. Xu and Yu (1997). Kawaguchi et al. SoftSphere ______________________________________________________________________________________ particle a list of neighbouring particles is stored and only for the particles in this list a check for possible collisions is performed. Fcontact is the contact force acting on the particle. In softsphere models the following equations of motion are used: m I d 2r = Fcontact + Fexternal .HardSphere vs.3) Ft. T is the torque acting on the particle and I is the moment of inertia of the particle. In this paragraph the focus will be on the contact forces between the particles.1) (5. (1993).2 Softsphere granular dynamics The linear spring/dashpot model (Cundall and Strack. (1996). ω is the rotation velocity. (1998). For a review of various contact force models used in softsphere simulations the reader is referred to Walton (1992) or Schäfer et al. Fexternal is the net external force acting on the particle.ab = − k nξ n n ab − ηn v n . the external forces will be discussed in the following paragraph.ab = −k t ? t − η t v t. 1979) is the most popular softsphere granular dynamics model since it was used by Tsuji et al. dt (5. m is the mass of the particle. The contact forces are given by: Fn.ab (5. Schwarzer (1995). 2.ab (5.
fluid model presented by Kuipers et al.6) For contacts between particles and walls.3 External forces The incorporation of external forces differs somewhat from the approach followed by Hoomans et al. (5.ab > µ Fn. In this work the external forces are used in accordance with those implemented in the two. b (5. b (5. Hence the external forces remain constant during a time step DT.7) Ta = ∑ (Ra n ab × Ft.ab ) .moving and of infinite mass just like in the hardsphere model. ab . the forces now act on a single particle: 134 .1DT the equations of motion were solved by taking only the contact forces into account.8) Since the contact forces are in general at least an order of magnitude larger than the external forces a separation of time scales was introduced.5) then sliding occurs and the tangential force is given by: Ft. the walls are assumed to be non.ab = − µ Fn. of course. By default at each time step DT the external forces were taken into account while at 0.ab ) .ab t ab . (1992) where. 2.ab + Ft. a = ∑ (Fn. (5.Chapter 5 ______________________________________________________________________________________ Ft. The resulting force and torque for particle a are now simply obtained by adding the pair wise contributions of all the particles that are in contact with a: Fcontact . (1996).
The second term is due to the drag force where β represents an interphase momentum exchange coefficient as it usually appears in twofluid models.9) where mp represents the mass of a particle.80) β is obtained from the wellknown Ergun equation: (1 − ε ) 2 β = 150 ε µg D2 p + 1. SoftSphere ______________________________________________________________________________________ m dv p p dt = mg+ p ( u − v ) − V ∇p (1 − ε ) p p p Vβ .9) the first term is due to gravity and the third term is the force due to the pressure gradient. vp its velocity.44 Re p < 1000 . (5.11) The drag coefficient Cd is a function of the particle Reynolds number and given by: 24 0.75(1 − ε ) ρg Dp u − vp .15Re p ) Cd = p 0.12) Re p ≥ 1000 where the particle Reynolds number (Re p ) in this case is defined as follows: 135 .65 .10) where Dp represents the particle diameter.687 1 Re ( + 0.80) the following expression for the interphase momentum transfer coefficient has been used which is basically the correlation presented by Wen and Yu (1966) who extended the work of Richardson and Zaki (1954): 3 ε (1 − ε ) Cd ρ g u − v p ε − 2. (5. u the local gas velocity and Vp the volume of a particle. For high void fractions (ε ≥ 0. µg the viscosity of the gas and ρg the density of the gas. 4 Dp β= (5. (5. In equation (5.HardSphere vs. For low void fractions (ε < 0.
In the present model the reaction force to the drag force 136 . The constitutive equations can be found in Chapter 3.15) In this work transient.14) Momentum equation gas phase: ∂ ερg u ∂t ( ) + (∇ ⋅ ερ uu ) = − ε∇p − S − (∇ ⋅ ε τ ) + ερ g . ( ) ∂t g (5. There is one important modification with respect to the model presented by Hoomans et al. (1992). g p g g (5. (5.13) For the integration of equation (5.4 Gas phase hydrodynamics The calculation of the gas phase hydrodynamics mainly follows the lines presented by Kuipers et al. It is based on the numerical solution of the following set of partial differential equations that can be seen as a generalised form of the NavierStokes equations for a gas interacting with a solid phase as originally derived by Anderson and Jackson (1967). Continuity equation gas phase: ∂ ερg ( ) + ∇ ⋅ ερ u = 0 .9) a simple explicit first order scheme was used to update the velocities and positions of the particles.Chapter 5 ______________________________________________________________________________________ Re p = ερg u − v p D p µg . isothermal (T = 293 K) flow of air at atmospheric conditions is considered. (1996) and that deals with the way in which the twoway coupling between the gas phase and the dispersed particles is established. 2. twodimens ional.
SoftSphere ______________________________________________________________________________________ exerted on a particle per unit of volume is fed back to the gas phase through the source term Sp which has the dimension of force per unit of volume N/m3 . At the bottom row of cells the gas inflow velocity was set equal to the minimum fluidisation velocity of the glass particles (umf = 0.96 0.15 40. DT 195 mm 300 mm 15 mm 5 mm 5 mm 39 60 104 s The values of the collision parameters are taken from Hoomans et al. 137 . DY number xcells.1 DT. (1996). DX cell height. Unless indicated otherwise a time step DT = 104 s is used while DTcontact = 0. The parameter settings for these simulations are presented in table 5.86 0. Table 5. Dp Material density. 3. NX number ycells.000 Bed: width height orifice diameter.1. At the side walls noslip boundary conditions were applied and at the upper boundary cells a prescribed pressure was chosen.5 m/s). Np spherical 850 µm ballotini glass 2930 kg/m3 0.2 s of the simulation to generate a single bubble. In the three central cells the gas velocity was set equal to 5 umf during the first 0. The formation of a bubble at a central orifice was chosen as a test case.15 0. The coefficient of tangential restitution (β 0 ) is assumed to be equal to zero. ρp e ew µ µw Number. Parameter settings bubble formation simulations Particles: Shape diameter.1. Preliminary simulations At first several preliminary simulations were performed using the softsphere model to investigate the influence of the spring stiffness on the simulations results. do cell width. NY time step.HardSphere vs.
Unless indicated otherwise the value of the stiffness of the tangential spring was chosen as 2/7 times the normal spring stiffness (k t = 2/7 kn ).05 103 9.000 N/m.2.2. 106 s No Yes Yes Yes Yes No: simulation not completed successfully.64 104 3.16) The values of the spring stiffness were varied from 1 to 10. (5. Note that the contact times in the linear spring/dashpot model do not depend on the impact velocity. Contact times and results of simulations* for various values of the normal spring stiffness.05 105 DTcontact. Table 5. The contact time for such particles within the framework of the linear spring/dashpot model can be calculated as follows: m p (π 2 + ln 2 e ) 2k n t contact = .64 105 3. The repulsive force is simply not strong enough compared to the external forces (gravity and drag) to prevent the particles from excessive overlap. 105 s No Yes Yes Yes No DTcontact. kn N/m 1 10 100 1000 10000 * t contact s 3.05 104 9. During the simulation this leads to extremely low local void fractions (ε < 0.2) which eventually causes the simulation to crash. 138 . In all cases DT = 104 s. 104 s No Yes No No No DTcontact. The contact times for these values of the spring stiffness are presented in table 5. Yes: simulation completed successfully With a normal spring stiffness of 1 N/m no successful simulations could be performed. With a spring stiffness of 10 N/m or higher such problems were not encountered.Chapter 5 ______________________________________________________________________________________ The effect of the spring stiffness was studied for a system consisting of particles of uniform size.
In the simulations presented by Xu and Yu (1997) the 139 . In the simulations presented by Tsuji et al.1 it can be observed that there is hardly any influence of the time step used for the integration of the contact forces on the overall results of the simulation which is rather surprising.05 104 s) a time step of 104 s was not sufficiently small to perform a stable simulation.HardSphere vs. Hence the ratio of the contact time and the time step should at least be greater than 3 (t contact/DTcontact > 3) to ensure the stability of the simulation. SoftSphere ______________________________________________________________________________________ However. In all cases k n = 100 N/m. The potential energy of the system as a function of time for several time steps. (1993). However. It turned out that in the case of kn = 100 N/m (t contact = 3. Figure 5.1 the sum of the potential energy of all the particles in the system is plotted as a function of time for several time steps (DTcontact). From Figure 5. There is however a big difference in CPU time because a simulation requires nearly 10 times as much CPU time when the time step is reduced by a factor 10. when a time step of 5 105 s was used. the simulation was successfully completed. Snapshots of the simulations do not reveal a significant difference either. In Figure 5. at higher values of the spring stiffness the time step for the numerical integration of the contact forces should be chosen sufficiently small in comparison with the contact time in order to perform stable simulations. (1998) a time step was used that was 3.7 times smaller than the contact time for the particles used in their systems. Kawaguchi et al. (1998) and Mikami et al.1.
Apparently the value of the spring stiffness has a negligible effect on the simulation results as long as it is high enough compared to the net external force to prevent the particles from overlapping too much (more than roughly 10% of the particle diameter). By default the value for the tangential spring stiffness is set equal to 2/7 times the value of normal spring stiffness (k t = 2/7 k n ) for reasons discussed in Chapter 2.2 snapshots at t = 0.7 appears to be rather close to the stability limit the results presented in Figure 5.2. k n = 100 N/m and k n = 10000 N/m. However. This could explain the minor influence of DTcontact on the simulation results. the choice of spring 140 .14. All snapshots in Figure 5. it was demonstrated that the effect of the time step on the time resolution of the contact force was rather large.2 appear to be surprisingly identical. The results presented in Figure 5. For all three time steps very similar results were obtained. However. This result is somewhat surprising since in Chapter 2.2 s is included of a simulation where the value for the tangential spring stiffness is set equal to the value of normal spring stiffness (k t = k n = 100 N/m). Although a ratio of 3. the error in the energy balance over the collision was rather low in that particular case. In all these cases DTcontact = 105 s.1 also indicate that the technique employed by Xu and Yu (1997) to determine the precise instant of first contact within a time step does not improve the results of a simulation.1 suggest that it is not required to choose a very small time step. In Figure 5. This is due to the fact that in the simulation with DT = 103 s a larger neighbour list was required since this list could only be updated at every time step.Chapter 5 ______________________________________________________________________________________ ratio of the contact time and the time step was 6. DT = 104 s. A simulation with k n = 1000 N/m was not included in the figure but showed very similar results. DT = 105 s). The simulation with DT = 105 s was significantly slower in terms of CPU time than the simulation with DT = 104 s but the difference between the simulations with DT = 103 s and DT = 104 s was remarkably small. In Figure 5. Additional simulations were performed using the hardsphere model where the overall time step (DT) for the integration of the external forces was varied (DT =103 s.2 s are presented for simulations with various values of the spring stiffness: k n = 10 N/m.2 a snapshot at t = 0. Figure 2.
141 .2.2 s for various values of the spring stiffness a) k n = 10 N/m b) k n = 10000 N/m c) k n = 100 N/m d) k n = k t = 100 N/m. Snapshots of particle configurations at t = 0. SoftSphere ______________________________________________________________________________________ stiffness does affect the CPU time required for a simulation significantly since the time step used for the integration of the contact forces should at least be three times smaller than the contact time corresponding to the spring stiffness used in the simulation. a) b) c) d) Figure 5.HardSphere vs.
From these distribution chambers the primary fluidisation gas was introduced into the fluidised bed section through a porous sintered stainless steel plate with a mean pore diameter of 10 µm. For the same reason the coefficient of tangential restitution (β 0 ) had a negligible influence on the simulation results (compared to the coefficients of normal restitution (e) and friction) presented in Chapter 4. k t = 2/7 k n ). Furthermore the effect is most pronounced for sticking contacts with a coefficient of tangential restitution equal to one (β 0 = 1.0 mm) to inject secondary fluidising gas. 4.2 the coefficient of tangential restitution is set equal to zero (β 0 = 0.2c (k n = 100 N/m. However. 4. 1994). which served as the main distributor of the primary fluidising gas.3 (Nieuwland. This porous plate.2d (k t = kn = 100 N/m) hardly any difference can be observed with the snapshot in Figure 5. The primary fluidising gas was introduced into the bed through two equally sized distribution chambers.up used for the experiment is shown schematically in Figure 5.Chapter 5 ______________________________________________________________________________________ In Chapter 2 it was demonstrated that the value of the tangential spring stiffness should be set equal to 2/7 times the value of normal spring stiffness (k t = 2/7 k n ) in order to ensure equal normal and tangential contact times.1 Experimental The set.0 showed very similar results and hence the effect due to the low coefficient of friction is dominant. was equipped with a central rectangular pipe (internal dimensions: 15. an additional simulation with β 0 = 1. From the snapshot in Figure 5.0) while in the simulation presented in Figure 5. This rectangular 142 .0).0 x 15. This is however not surprising because of the low value of the coefficient of friction (µ) that causes a considerable amount of sliding contacts. Experimental validation The hardsphere and the softsphere codes were compared with each other and validated using data of a bubble formation experiment obtained from the experimental setup described below.
2 s secondary gas was injected through the central orifice at a superficial velocity of 2.5 m/s (5 umf). a gas flow through the rectangular orifice to fluidise the particles above the orifice and a gas flow for the bubble formation at the orifice. Figure 5. computer controlled.3. SoftSphere ______________________________________________________________________________________ jet was covered with a piece of stainless steel gauze to prevent the particles from entering it. the secondary gas was purged. while the primary gas flowed through the gas distributor and the orifice to keep the bed at minimum fluidisation conditions.HardSphere vs. During the experiments three different gas flow rates were adjusted independently by means of calibrated thermal mass flow controllers: a gas flow through the distributor section to keep the bed at incipient fluidisation conditions. magnetic valves were used to inject secondary gas through the orifice into the bed to form a bubble and to purge the primary gas during the period of injection. Prior to bubble injection. Rapidly responding.5 m/s (umf) and during a period of 0. An SVHS video camera was used to observe the bed during the whole process 143 . Schematic representation of the experimental setup. In the experiment the primary gas was introduced at a superficial velocity of 0.
The particle size distribution employed in the simulation is presented in Figure 5. 1998). Particle size distribution for the bubble formation simulations 144 .Chapter 5 ______________________________________________________________________________________ of bubble formation.up described above. Figure 5.4.1. Apart form a simulation with particles of a uniform diameter of 850 µm a simulation was performed where the particle diameters were obtained from a lognormal distribution with a count median average of 850 µm and a geometric standard deviation of 50 µm. 4.normal particle size distribution was taken into account (Hoomans et al. A sieve fraction of glass ballotini particles (ρp = 2930 kg/m3 ) with diameters between 800 and 900 µm was used. The results of these simulations were validated using data of a bubble formation experiment obtained from the experimental set. The main parameter settings for the simulations are presented in Table 5.4. a simulation with particles of uniform size was compared with a simulation where a log.2 Influence of a particle size distribution In order to test whether the simulation results were improved by taking a particle size distribution into account..
This is not observed in the experiment nor in the simulation with the polydisperse particle assembly. These ‘satellite’ bubbles are probably due to locally low void fraction due to close packing. It can be observed that in the system with particles of uniform size small ‘satellite’ bubbles appear above and alongside the main bubble. a) Figure 5. b) c) Snapshots at t = 0. The size of the main bubble agrees reasonably well with the experiment for both of the (2D) simulations which is rather encouraging especially since the model parameters were obtained on beforehand and independently.5. c) simulation with lognormal particle size distribution.like structure. When locally low void fractions prevail the drag force (which depends strongly on the void fraction) can become strong enough to generate a bubble. b) simulation with uniform particles. The simulations presented here were performed using the hardsphere code. 145 . SoftSphere ______________________________________________________________________________________ Snapshots at t = 0.HardSphere vs.5.2 s of: a) experiment.2 s of the two simulations and corresponding video images of the experiment are presented in Figure 5. In a system consisting of particles of uniform size such locally low void fractions can occur since it is possible for an assembly of particles to gather in a lattice.
In the snapshots presented in Figure 5. After 0.Chapter 5 ______________________________________________________________________________________ 4. In Figure 5. The initial conditions were identical in both simulations and hence there was no overlap between particles at t = 0 s.1 s the bubble detaches from the distributor plate and starts to rise in the bed. In both simulations the particle size distribution presented in Figure 5.1 s the shape of the bubble predicted in the simulations starts to deviate from the bubble shape observed in the experiment although the main bubble size as well as the position of the bubble in the bed is fairly well predicted.2 s the formation of a second bubble is initiated but the jet does not remain active long enough for this second bubble to grow to full size.6 snapshots of the hardsphere (middle) and the softsphere (right) simulations are presented together with the corresponding pictures of the experiment (left).4 is taken into account. Although the experimental setup was quasi twodimensional the bed was still more than 17 particle diameters deep (16 mm) which is a significant difference with the simulations where the particle motion is restricted to two dimensions.3 Hardsphere vs. which is required for the hardsphere simulation. The bubble shape observed in the experiment is more round and the gassolid boundary appears to be sharper than in the simulations. The use of such a higher order scheme (like for instance the Barton scheme) may also improve the bubble definition in (already 2D) Lagrangian simulations. For the softsphere simulations a normal spring stiffness (k n ) of 100 N/m was used and the time step for the integration of the contact forces (DTcontact) was set equal to 105 s. In addition a higher order numerical scheme for the convective fluxes in the gas phase momentum conservation equations is known to yield results with sharper bubble boundaries in Eulerian simulations. After 0. 146 .1. This can be very well due to the twodimensional nature of the simulations.6 it can be observed that the bubble is being formed at the central orifice during the first 0.1 s which is in good agreement with the experiment. softsphere The simulations were performed using the parameter settings presented in Table 5. Since the jet remains active during the first 0.
05 s b) t = 0. Snapshots of bubble formation simulations compared with an experiment.05 s b) t = 0.6. SoftSphere ______________________________________________________________________________________ a) t = 0.10 s Figure 5. a) t = 0.HardSphere vs.10 s… 147 . From left to right: experiment. hardsphere simulation and softsphere simulation.
softsphere c) t = 0.15 s d) t = 0. From left to right: experiment.20 s. (…continued) Snapshots of bubble formation simulations compared with an experiment. hardsphere. 148 .6.15 s d) t = 0.Chapter 5 ______________________________________________________________________________________ c) t = 0.20 s Figure 5.
0. When the collisions were assumed to be fully elastic and perfectly smooth (e = 1.15.2 s when the jet is no longer active.0 s to t = 0. From t = 0. The snapshots appear to be almost identical. SoftSphere ______________________________________________________________________________________ It is remarkable however that hardly any difference can be observed between the results of the hardsphere and the softsphere simulations.8 million collisions was processed whereas from t = 0.2 s a total number of 37.0. central orifice using a hardsphere simulation.2 s to t = 0. µ = 0. also referred to as nonideal collision parameters) was tremendous. also referred to as ideal collision parameters) no bubble could be observed after 0. There was however a difference in computational efficiency between the two simulations. 1Mb cache memory) for a simulation of 0.96. Since the results show hardly any difference it can be concluded that the assumption in the hardsphere code that all collisions are binary is not limiting. In the hardsphere simulation the first 0.3 s the respectable number of 110 million collisions was processed. 4. µ = 0. 149 . Snapshots of both the hardsphere and the softsphere simulations are presented in Figure 5.3 s whereas the softsphere simulation required 250 minutes of CPU time on the same machine. apart from the collision parameters. (1998) demonstrated the effect of the collision parameters on bubble formation at a single. Hence the computational efficiency of a hardsphere simulation depends strongly on the dynamics of the system.7 together with video images of the experiment.2 s (with an active jet) only require 30% of the total amount of CPU time. This simulation was repeated here using the softsphere code with.2 s.HardSphere vs. The main difference between the two simulations is that the softsphere simulation progresses in time at a rather constant speed whereas this is certainly not the case in the hardsphere simulation. This is due to the strong increase in the number of collisions after 0.4 Effect of collision parameters Kuipers et al. The difference with the simulation where realistic values for the collision parameters were specified (e = 0. The hardsphere simulation required about 220 minutes of CPU time on a Silicon Graphics Origin200 server with a R10000 processor (180 MHz. In the hardsphere all collisions are assumed to be binary whereas the softsphere code is able to handle multiple contacts. the same parameter settings as used for the simulation in the previous paragraph.
2 s Figure 5. µ = 0.2 s. hardsphere. Snapshots of bubble formation simulations with ideal collision parameter settings (e = 1.0. β 0 = 0.1 s b) t = 0.0) compared with an experiment.Chapter 5 ______________________________________________________________________________________ a) t = 0. From left to right: experiment.1 s b) t = 0.7. 150 . softsphere a) t = 0.0.
7 the bubble seems to have disappeared which is no t in agreement with the snapshot at t = 0.fluidised bed have been developed. The results of the simulations did not change significantly when the time step for the integration of the contact forces was decreased. 151 .7 it can clearly be observed that the result presented earlier by Kuipers et al.6 as well as the experiment. SoftSphere ______________________________________________________________________________________ In Figure 5. Preliminary simulations with the softsphere code revealed that the time step used for the integration of the contact forces should be more than 3 times smaller than the contact time. This is found in both the hardsphere and softsphere granular dynamics simulations. The results of the simulations were also rather insensitive to the value of the normal spring stiffness. The value of the normal spring stiffness should be high enough compared to the net external force in order to prevent the particles from overlapping too much (more than roughly 10% of their diameter).2 s in Figure 5.6 and more importantly the experiment. But in the snapshot at t = 0. (1998) obtained with a hardsphere code is reproduced here with a softsphere code. These twodimensional models were compared with each other and validated experimentally using the formation of a single bubble at a central orifice as a test case. A simulation were the value of the tangential spring stiffness was set equal to the value of the normal spring stiffness ( t = k n ) showed hardly any difference with the k default simulations where k t = 2/7 k n . Conclusions A hardsphere and a softsphere granular dynamics model of a gas.HardSphere vs. 5.1 s is still quite similar to the snapshot of the simulation with the nonideal collision parameters presented in Figure 5. This stresses the importance of the collision parameters in granular dynamics simulations of gasfluidised beds. The snapshot at t = 0. Although the use of very high values for the spring stiffness does not change the results significantly it requires more CPU time since smaller time steps have to be used.2 s in Figure 5. When nonrealistic values for the collision parameters are specified nonrealistic behaviour is predicted.
The more collisions 152 . This could not be observed in the experiment nor in the simulation where a log.normal particle size distribution was taken into account. Also the gassolid boundary appeared to be sharper in the experiment than in the simulations. Applying periodic e boundary conditions in the zdirection would reduce the total number of particles required in the simulation but in that case the influence of the front and back walls would be eliminated as well. The shape of the bubble observed in the experiment was more round than the one observed in the simulations. Where the softsphere simulation progresses through the simulation at a rather constant speed the computational efficiency of the hardsphere simulation depends strongly on the dynamics of the system. Another improvement can be achieved by employing a higher order numerical scheme for the convective fluxes in the gas phase momentum equations that suffer less from numerical diffusion.000 particles which is extremely challenging even for state of the art computing r sources of today. However. The results of both the hardsphere and the softsphere simulation were nearly identical. A hardsphere and a softsphere simulation where a log. The main bubble size and the position of the bubble in the bed were found to be in good agreement with the experiment. there was a difference in computational efficiency. Therefore it can be concluded that the assumption in the hardsphere code that all collisions are binary is not limiting.normal particle size distribution was taken into account were compared with each other and with an experiment. However. Both simulations agreed rather well with the experiment especially during the initial stages.Chapter 5 ______________________________________________________________________________________ The incorporation of a particle size distribution turned out to improve the agreement between simulation and experiment. a full 3D system simulation of this system would require about 750. In a bubble formation simulation with particles of uniform size small ‘satellite’ bubbles appeared above and alongside the main bubble. This is probably due to the twodimensional nature of the simulations. Further improvement can be achieved by extending the model to three dimensions because although the bed used in the experiment was quasi twodimensional it was still 16 mm deep. Results of Eulerian simulations with such higher order schemes show sharper bubble boundaries and that could also be the case for Lagrangian simulations.
m vertical computational cell dimension. When the collisions were assumed to be fully elastic and perfectly smooth (e = 1.HardSphere vs. [] force. m/s2 moment of inertia. s horizontal computationa l cell dimension. This result was obtained with both the hardsphere and the softsphere model.0) no bubble could be observed after 0. [] normal unit vector. [] number of computational cells in xdirection. m time step.0. The overall CPU time requirements of the two codes were very similar for a simulation of 0.3 s. m coefficient of restitution.e. µ = 0. [] orifice width. kg number of particles. N/m particle mass. kgm2 spring stiffness. SoftSphere ______________________________________________________________________________________ occur the slower the progress. N/m3 153 .2 s. Notation Cd Do Dp DT DX DY e F g I k m Np NX NY n p Rp r Sp drag coefficient. coefficients of restitution and friction) turned out to have a key influence on the bubble formation process. N gravitational acceleration. m momentum source term. m particle diameter. Pa particle radius. m position vector. [] number of computational cells in ydirection. The collision parameters (i. [] pressure.
. m/s particle volume. [] coefficient of friction. and Jackson.. [] gas viscosity.B. 527. R. Ind. 6. Fundam. m Subscripts g mf p w gas phase minimum fluidisation particle wall References Anderson. kg/ms2 density. K time. kg/m3 displacement. (1967). Eng. A fluid mechanical description of fluidized beds (equations of motion). [] void fraction. kg/(m3 s) coefficient of tangential restitution. Ns/m gas phase stress tensor. T.Chapter 5 ______________________________________________________________________________________ T T t u v Vp torque.. m3 Greek symbols β β0 ε µ µg η τ ρ ξ volumetric interphase momentum transfer coefficient. kg/ms damping coefficient. Chem. s gas velocity vector. Nm temperature. m/s particle velocity vector. 154 .
and Horio. (1996). O. Multiphase Flow and Fluidization. 99. Engng Sci. van Duin K. 96. Numerical simulation of twodimensional fluidized beds using the discrete element method (comparison between the two. F.. Foerster S. W.M. 51. 15. (1998). D. (1979).Y. Hoomans. Chem.M. Kuipers. Discrete particle simulation of bubble and slug formation in a twodimensional gas.. W.J.M.. J. Kuipers.. Kawaguchi. J.. Gidaspow. 94.. The Influence of a Particle Size Distribution on the Granular Dynamics of Dense GasFluidized beds: a Computer Simulation Study”.. 47. W..B. (1992). 1927. B. Y.H. T. and van Swaaij.B.P. Briels. B. L. Chang. No.HardSphere vs.M.D. Ser.. (1998). W. Chem. and van Swaaij. Knowlton eds. W.B. Mikami. 1913 Kuipers. (1998). M. (1994). Fluids. Phys..and threedimensional models).M. Engng Sci.M. and Allia.. P. A numerical model of gasfluidized beds. Numerical simulation of cohesive powder behavior in a fluidized bed..M. T. 129. and Strack. J. M. AIChE Symp. 1108. Géotechnique..P.P. and van Swaaij.P. B. Tanaka. Powder Technol... Measurements of the collision properties of small spheres..P. van Beckum.. Hoomans. J. Academic Press. Fan and T..P. 29.P.L.P.. 318.. Kamiya. T. Hoomans.S. 47.M. 15.A. Kuipers. and Tsuji. Chem. 6.M.. USA.fluidised bed: a hardsphere approach. H.A. 155 . 53. (1994).A. Hydrodynamic Models of GasFluidized Beds and their Role for Design and Operation of Fluidized Bed Chemical Reactors”.. (1998).A. and van Swaaij.F.J. A discrete numerical model for granular assemblies. SoftSphere ______________________________________________________________________________________ Cundall. Engng Sci.H. in Fluidization IX. Boston.A. Louge. K..
(1995). Numerical simulation of inelastic. Richardson.. 62 (62). and Yu. University of Twente.. (1954). I France 6. Numerical simulation of the gassolid flow in a fluidized bed by combining discrete particle method with computational fluid dynamics.Y. Y. Dippel. (1992). 32. 156 . PhD Thesis. Force schemes in simulations of granular materials. J..L. in Particulate TwoPhase Flow. Powder Technol. J.H.. (1989). Boston.C.. Discrete particle simulation of two dimensional fluidized bed..T. and Zaki. C. O. Trans. Schwarzer. Phys. Roco.R. Inst. Appl. Prog. D. 1473.. edited by M. Walton. T. Chem. R. Hydrodynamic modelling of gassolid twophase flows. 6461. Part I. Symp.B. (1966). frictional particleparticle interactions. Chem. Sedimentation and flow through porous media: Simulating dynamically coupled discrete and continuum phase.E. (1996).J. Wen. Mech. Schäfer. Tsuji. 35. and Wolf. Y. (1994). J. 79.. 2785. Chem.. AIChE J. E.H.F. Eng. S. and Tanaka. and Mason. Gasparticle flow in a vertical pipe with particleparticle interactions. 52.Chapter 5 ______________________________________________________________________________________ Nieuwland. (1992). The Netherlands. Xu. 35. J. A. Twodimensional rigidbody collisions with friction. 52. Sinclair.. M. 884.N. S. Engng Sci. Wang. and Yu. Kawaguchi.. and Jackson.. Phys. W. 5.. 100. (1997).. T. 77. Rev. (1993). J. 635. Y. ButterworthHeinemann. B.. Sedimentation and fluidization: part I. Mechanics of fluidization. Ser. Eng. 59. J.
157 . GRANULAR DYNAMICS SIMULATION OF SEGREGATION PHENOMENA IN BUBBLING GASFLUIDISED BEDS Abstract: Granular dynamics simulations of gasfluidised beds were performed in order to simulate segregation phenomena in systems consisting of particles of different size and density. The (2D) model was applied to a system consisting of particles of three different densities but equal size and to systems consisting of particles of two different sizes but equal density. In the different density system segregation was observed over a time scale of several seconds. For each solid particle the Newtonian equations of motion are solved taking into account the particleparticle and particlewall interactions. The simulation predicts segregation at a lower gas velocity than used in the experiment.Segregation Phenomena ______________________________________________________________________________________ Chapter 6. A statistical analysis of the segregation in terms of mass fraction distributions is presented. When the particleparticle and particlewall interactions were assumed to be fully elastic and perfectly smooth. segregation occurred very fast and was almost complete due to the absence of bubbles. In the model the gasphase hydrodynamics is described by the spatially averaged NavierStokes equations for twophase flow. Preliminary experimental validation showed rather poor agreement between simulation and experiment. In the different size systems segregation was also observed but a clear steady state was not reached since the larger particles were continuously transported to the upper regions of the bed in bubble wakes.
(1999). Kuipers and W. Fan and T.S. B.M. van Swaaij. Discrete Particle Simulation of Segregation Phenomena in Dense GasFluidized Beds. (1998). in Fluidization IX.A. Hoomans.P.M. Kuipers and W.P.P. Granular Dynamics Simulation of Segregation Phenomena in Dense GasFluidised beds.M. 158 . Powder Technology (in press).M.M.P.B.A.Chapter 6 ______________________________________________________________________________________ Parts of this chapter are based on the following papers: B. J. Hoomans.B. Knowlton (eds). L. van Swaaij.485. J.
The smaller (lighter) particles show the tendency to float and reside at the bed surface. Segregation phenomena in gasfluidised beds have been the subject of several experimental studies reported in the literature. Introduction Segregation phenomena can play an important role i fluidised systems consisting of n particles of different size and/or density. In a system consisting of particles of equal density but different size the bigger (heavier) particles tend to reside at the bottom of the bed if the fluidisation velocity does not exceed the minimum fluidisation velocity (umf) of the big particles. Nienow and Naimer (1980) performed experiments on systems consisting of particles of equal size but different density. The big particles are in this case referred to as jetsam.Segregation Phenomena ______________________________________________________________________________________ 1. Rowe et al. At gas velocities much higher than the umf of the jetsam better mixing is normally achieved. In general the different density systems tend to segregate more easily than the polydisperse systems. Particleparticle and particlewall interactions are taken into account directly which is a clear advantage over two.fluid models that require closure relations for the solidsphase 159 . Due to increasing computer power discrete particle models have become a very useful and versatile research tool in order to study the hydrodynamics of gasfluidised beds. Nienow et al. In order to improve the performance of these processes detailed knowledge about the distribution of the different solids species throughout the bed in different operating conditions is required. In case of a system consisting of particles of equal size but different density the denser particles are referred to as jetsam whereas the lighter particles are referred to as flotsam. (1987). Typical examples of such processes are fluidised bed polymerisation and fluidised bed granulation among others. These particles are referred to as flotsam. (1993) and Wu and Baeyens (1998) studied systems consisting of particles of equal density but different size. In these models the Newtonian equations of motion for each individual particle are solved. (1972) presented experimental results for both types of systems. Hoffmann et al.
When simulating gas. Recently Kawaguchi et al. In addition. A discrete particle approach offers a more natural way to overcome these problems since each individual particle in the simulation is tracked. In this discrete particle model the particles interact through binary. (1993) to simulate liquid fluidised beds in which lubrication forces were also taken into account. inelastic collisions with friction. In this approach the particles are allowed to overlap slightly and this overlap is subsequently used to calculate the contact forces.fluidised beds with particles of different size and/or particles of different density multi. 1990) but several difficulties are encountered due to the fact that large sets of continuum equations have to be solved. (1998) presented results for a threedimensional version of this model. Kumaran and Koch (1993) and Manger (1996) applied this theory to binary mixtures and recently Mathiesen et al. Hoomans et al. (1998) extended the work of Manger (1996) in order to handle three solid phases.. Schwarzer (1995) used a model similar to that of Tsuji et al. Hence discrete particle models are very useful in order to study segregation phenomena in gas.fluidised bed. However the number of particles that can be taken into account in a simulation is limited (typically < 106 ) which implies that currently the method can only be applied to rather small systems consisting rather coarse particles. (1993) who developed a softsphere discrete particle model based on the work of Cundall and Strack (1979).fluidised beds consisting of particles of different size and/or density in detail. (1992) and Gidaspow (1994) among many others).Chapter 6 ______________________________________________________________________________________ stress tensor (Kuipers et al. and more fundamentally. instantaneous. Although the Kinetic Theory of Granular Flow (KTGF) offers a theoretical framework to overcome this problem this approach is not very flexible and the computationa l cost increases dramatically with each additional solid phase. The discrete particle approach for gasfluidised beds was pioneered by Tsuji et al. significant problems arise when closure laws for the mutual interaction of particles belonging to different classes have to be formulated. (1996) were the first to present a hardsphere granular dynamics model of a gas. Moreover they can be used to generate data which can subsequently be used to develop closure models for continuum models. Xu and Yu (1997) presented a hybrid technique where they used a contact force model in order to determine the interparticle forces and a collision detection 160 .fluid models can be used (Gidaspow et al.
These three collision parameters are all included in the model. Seibert and Burns (1998) were able to predict segregation phenomena in liquid fluidised beds using a Monte Carlo simulation technique. Mikami et al. A constant time step is used to take the external 161 . (1994) have shown that also the coefficient of tangential restitution (0 ≤ β0 ≤ 1) should be used in order to describe the collision dynamics more accurately. the Monte Carlo technique i only capable of predicting a certain s steady state and is not suitable to simulate the dynamics of segregation. (1998) recently extended the model of Tsuji et al. Foerster et al. (1993) in order to include cohesive forces between the particles. The majority of the simulations is concerned with systems consisting of particles of equal density but different size but also a system consisting of particles of uniform size but different density is included. However. inelastic collision with friction.Segregation Phenomena ______________________________________________________________________________________ algorithm in order to determine the precise instant at which the particles first come into contact. In the hardsphere approach a sequence of binary collisions is processed. the key features will be summarised only briefly here. 2. In this chapter twodimensional granular dynamics simulations will be used to study segregation phenomena in bubbling gasfluidised beds.1 Hardsphere granular dynamics The collision model as originally developed by Wang and Mason (1992) is used to describe a binary. Models Since a detailed description of the models is presented in Chapter 2 and Chapter 3. instantaneous. This implies that a collision list is compiled in which for each particle a collision partner and a corresponding collision time is stored. 2. The key parameters of the model are the coefficient of restitution (0 ≤ e ≤ 1) and the coefficient of friction (µ ≥ 0).
1. 2. In order to reduce the required CPU time neighbour lists are used. The neighbour list principle using two cut off distances.big Figure 6.small Dnblist . Dnblist.1. For each particle a list of neighbouring particles is stored and only for the particles in this list a check for possible collisions is performed. In softsphere models the following equations of motion are used: 162 .Chapter 6 ______________________________________________________________________________________ forces into account and within this time step the prevailing collisions are processed sequentially. The neighbour list consists of all the small particles which are found within the small square (coloured grey) and the big particles which are found within the big square (coloured grey). When simulating a binary system of particles of different size two cut off distances are used as schematically shown in Figure 6.2 Softsphere granular dynamics Although the majority of the simulations presented in this chapter were performed with the hardsphere code also a simulation with the softsphere code was performed and therefore that model is described briefly here as well. The shaded particles are stored in the neighbour list of the black one. By using this approach the number of particles in the neighbour list will never become too high which significantly reduces both CPU time and memory space requirements.
The contact forces are given by: Fn. (6.ab = − µ Fn. (6. dt 2 dω = T. the external forces will be discussed in the following paragraph.4) If however the following relation is satisfied: Ft. Fexternal is the net external force acting on the particle.1) (6.ab (6.5) then sliding occurs and the tangential force is given by: Ft. m is the mass of the particle. The resulting force and torque for particle a are now simply obtained by adding the pair wise contributions of all the particles that are in contact with a: 163 .ab t ab .3) Ft.ab (6. ω is the rotation velocity.ab = − k nξ n n ab − ηn v n . T is the torque acting on the particle and I is the moment of inertia of the particle.6) For contacts between particles and walls. ab .moving and of infinite mass just like in the hardsphere model. the walls were assumed to be non. dt (6.Segregation Phenomena ______________________________________________________________________________________ m I d 2r = Fcontact + Fexternal .ab > µ Fn. In this paragraph the focus will be on the contact forces between the particles.2) where r is the position vector of the centre of the particle. Fcontact is the contact force acting on the particle.ab = −k t ? t − η t v t.
b (6. In equation (6. The second term is due to the drag force where β represents an interphase momentum exchange coefficient as it usually appears in twofluid models.1DT the equations of motion were solved by taking only the contact forces into account.ab ) .fluid model presented by Kuipers et al. vp its velocity.Chapter 6 ______________________________________________________________________________________ Fcontact .7) Ta = ∑ (Ra n ab × Ft. (1996).9) the first term is due to gravity and the third term is the force due to the pressure gradient. b (6.ab + Ft. u the local gas velocity and Vp the volume of a particle. (1992) where. a = ∑ (Fn.9) where mp represents the mass of a particle. of course. For low void fractions (ε < 0. By default at each time step DT the external forces were taken into account while at 0. the forces now act on a single particle: m dv p p dt = mg+ p ( u − v ) − V ∇p (1 − ε ) p p p Vβ .ab ) .10) 164 . (6. (6.3 External forces The incorporation of external forces differs somewhat from the approach followed by Hoomans et al. 2.8) Since the contact forces are in general at least an order of magnitude larger than the external forces a separation of time scales was introduced.80) β is obtained from the wellknown Ergun equation: β = 150 (1 − ε ) 2 ε µg D 2 p + 1. In this work the external forces are used in accordance with those implemented in the two.75(1 − ε ) ρg Dp u − vp .
(6. 2.13) For the integration of equation (6.11) The drag coefficient Cd is a function of the particle Reynolds number and given by: 24 0. (1992).12) Re p ≥ 1000 where the particle Reynolds number (Re p ) in this case is defined as follows: ερg u − v p D p µg Re p = .Segregation Phenomena ______________________________________________________________________________________ where Dp represents the particle diameter. 165 . (6.4 Gas phase hydrodynamics The calculation of the gas phase hydrodynamics mainly follows the lines presented by Kuipers et al.65 .9) a simple explicit first order scheme was used to update the velocities and positions of the particles.15Re p ) Cd = p 0. µg the viscosity of the gas and ρg the density of the gas.44 Re p < 1000 . 4 Dp β= (6. It is based on the numerical solution of the following set of partial differential equations that can be seen as a generalised form of the NavierStokes equations for a gas interacting with a solid phase as originally derived by Anderson and Jackson (1967).80) the following expression for the interphase momentum transfer coefficient has been used which is basically the correlation presented by Wen and Yu (1966) who extended the work of Richardson and Zaki (1954): 3 ε (1 − ε ) Cd ρ g u − v p ε − 2.687 1 Re ( + 0. For high void fractions (ε ≥ 0.
98 m/s).14) Momentum equation gas phase: ∂ ερg u ∂t ( ) + (∇ ⋅ ερ uu ) = − ε∇p − S − (∇ ⋅ ε τ ) + ερ g . There is one important modification with respect to the model presented by Hoomans et al. Ternary density distribution A simulation was performed of a system that consisted of three types of particles of uniform size but of different density using the hardsphere code.1. 800 particles with a density of 1800 kg/m3 (u mf = 1.79 m/s). For the calculation of the void fraction (ε) the technique described in Chapter 3 was used. ( ) ∂t g (6.Chapter 6 ______________________________________________________________________________________ Continuity equation gas phase: ∂ ερg ( ) + ∇ ⋅ ερ u = 0 .15) In this work transient. (1996) and that deals with the way in which the twoway coup ling between the gas phase and the dispersed particles is established. The 166 . twodimensional. isothermal (T = 293 K) flow of air at atmospheric conditions is considered.42 m/s) and 800 particles with a density of 900 kg/m3 (umf = 0. No special modifications were made to take the effect of the particle size distribution into account. g p g g (6. The main parameter settings for this simulation can be found in table 6. In the present model the reaction force to the drag force exerted on a particle per unit of volume is fed back to the gas phase through the source term Sp which has the dimension of force per unit of volume N/m3 . The constitutive equations can be found in Chapter 3. The system was filled with 800 particles with a density of 2700 kg/m3 (umf = 1. 3.
2 snapshots of the particle configurations are presented.15 m. The bed was fluidised at the minimum fluidisation velocity of the densest particles (ug = 1. Table 6. In Figure 6. NY cell width. DT 0. Snapshots of the particle configurations in a system with a ternary density distribution.79 m/s).79 m/s) for 15 s starting from perfectly mixed initial conditions. DY time step.Segregation Phenomena ______________________________________________________________________________________ gas inflow velocity was set equal to the minimum fluidisation velocity of the densest (ρp = 2700 kg/m3 ) particles (umf = 1.3 2400 Bed: width height number xcells. NX number ycells.2.1. 167 .50 m 15 25 10 mm 10 mm 104 s The coefficients of tangential restitution were assumed to be equal to zero. DX cell height.total spherical 4. Parameter settings for the simulation with a different density mixture Particles: shape diameter ρp = 2700 kg/m3 ρp = 1800 kg/m3 ρp = 900 kg/m3 e = ew µ = µw Np.9 0. Figure 6.0 mm Np = 800 Np = 800 Np = 800 0. 0.
Figure 6.Chapter 6 ______________________________________________________________________________________ In Figure 6.3. The middle section of the bed clearly shows a peak in the mass f action of the r particles with ρp = 1800 kg/m3 . (1972) who reported that different 168 .2 the segregation effect can clearly be observed. In the upper regions of the bed the mass fraction jetsam is zero whereas the mass fraction flotsam approaches one. However. In Figure 6. Mass fractions as a function of the bed height at t = 15 s for the simulation with a ternary density distribution.3umf of the particles with ρp = 1800 kg/m3 . After only 15 seconds three distinct layers can be identified. It is remarkable that even though the two less dense types of particles are both fluidised well above their umf there is a clear segregation effect between these two layers as well.05 m height. The mass fractions are obtained by averaging over the full width of a bed section of 0.3 the mass fractions are plotted as a function of the bed height at t = 15 s. Even though the fluidisation velocity is 1. the segregation between the upper two l yers is rather clear. this is in agreement a with the experimental observations by Rowe et al. In this figure it can be observed that the mass fraction of the jetsam particles (ρp = 2700 kg/m3 ) is nearly one in the bottom regio n where the mass fraction of the flotsam particles (ρp = 900 kg/m3 ) is zero.
DY time step.big = 1. Snapshots of particle configurations are shown in Figure 6. DT 0.4.15 m. NY cell width.86 0.5 mm diameter (50/50 wt.15 5000 Bed: width height number xcells.0 mm e ew µ = µw Np. 169 .2. In this figure it can be observed that segregation does occur with the bigger particles accumulating at the bottom of the bed. The simulation was run for 50 seconds using the parameter settings presented in table 6. ρp Dp = 1. 0.1 Base case A simulation was performed of a system that consisted of two types of particles of different size but equal density using the hardsphere code. Since the particle size is rather large in this relatively small system wake effects due to bubbles are less pronounced. NX number ycells. DX cell height.2. 4.5 mm Dp = 4. Table 6.25 m 15 25 10 mm 10 mm 10–4 s The coefficients of tangential restitution were assumed to be equal to zero.total spherical 2480 kg/m3 Np = 4750 Np = 250 0. A homogeneous mixture of 250 particles of 4.0 mm diameter and 4750 particles of 1. A dispersive mechanism was therefore not dominantly present and this may have enhanced the segregation. The collision parameters for the glass particles were obtained from Hoomans et al. Parameter settings for the simulations with a binary size distribution Particles: shape density. Binary size distribution 4.) was fluidised at the minimum fluidisation velocity of the bigger particles (umf.Segregation Phenomena ______________________________________________________________________________________ density systems still show segregation above the umf of the jetsam particles. (1996).7 m/s).96 0.
Figure 6. this is a very dynamic situation as could be observed in animations where it became clear that the bigger particles are continuously transported to the upper regions of the bed in the wake of a bubble and descend again in the denser regions. This process continued throughout the simulation.4. Snapshots of the particle configurations for a binary system of particles of different size.Chapter 6 ______________________________________________________________________________________ However. 170 .
4 that there is no pure layer of jetsam particles being formed on the bottom of the bed. 171 . In experiments normally the gas supply is first shut off after which the bed is sectioned and each section is analysed separately.5 also results for higher gas velocities are included from which it can be observed that the solids mixing is better at increased gas velocities. This is not surprising since it could already be observed from the snapshots in Figure 6. Figure 6.Segregation Phenomena ______________________________________________________________________________________ A typical simulation took about 25 minutes of CPU time per simulated second on a Silicon Graphics Power Challenge with a 196 MHz R10000 processor. 1972) is confirmed here when Figure 6.3. Since all particle positions are known at each instant during a simulation such a plot can be calculated instantaneously.5 is included to guide the eye. In the figure an additional line at the average mass fraction of 0. The common knowledge that different density systems tend to segregate more easily than different size systems (Rowe et al. In Figure 6.5.. In Figure 6.5 the mass fraction jetsam at t = 20 s is plotted as a function of the height in the bed. Mass fraction jetsam as a function of the bed height at t = 20 s for the simulations with a binary size distribution at three different gas velocities (u_mf refers to the umf of the big particles) It can be observed that the mass fraction jetsam is indeed higher in the lower part of the bed although the fraction does not become one.5 is compared with Figure 6.
Figure 6. For the results presented in Figure 6. Hence a reliable average and a corresponding variance could be obtained from 5000 frames. Wu and Baeyens (1998)). Segregation profiles obtained from both single. This enables the determination of segregation profiles by averaging over a large number of particle configurations.2 Statistical analysis of segregation Although segregation could be observed in the base case simulation presented in the previous paragraph the system remained very dynamic. as was presented for the base case simulation. By repeating the experiments the reliability of the results can be improved which is however rather time consuming.6 an average segregation profile was obtained using the particle configurations at each 0. does not provide a complete understanding of the segregation phenomenon. A similar problem is encountered in segregation experiments where segregation profiles are determined after the gas supply is abruptly shut off and the bed is subsequently divided in sections that are separately analysed (Hoffmann et al.01 seconds for a total duration of 50 seconds. (1993). Therefore an analysis based on a single frame. also referred to as frames.6.Chapter 6 ______________________________________________________________________________________ 4.and multiple frame analysis 172 . A line indicating the average mass fraction jetsam is included to guide the eye. In granular dynamics simulations the positions of all the particles in the system are known at each instant.
It only gives an impression of an instantaneous situation and therefore does not provide a complete understanding of the segregated state of the system.7 the probability distribution of the mass fraction jetsam at a bed height of 0. Figure 6.0 s) is included as well. Since this profile differs significantly from the average profile it clearly shows that one has to be careful with the application of single frame analysis.7. Probability distribution of the jetsam mass fraction at 0. 173 . In Figure 6.075 m above the distributor plate obtained over 50 seconds The wide spread in this distribution again emphasises that a single frame analysis may lead to an incomplete understanding of the segregated state of the system.Segregation Phenomena ______________________________________________________________________________________ In this figure a segregation profile obtained from a single frame analysis (t = 20. It is important to note that the segregation profiles obtained from the simulation are based on the actual positions of the particles during fluidisation while in experiments the gas supply is always shut off before the collapsed bed is sectioned and analysed.075 m obtained over the full duration of the simulation is presented. The segregation profiles obtained in such experiments are therefore based on the particle positions in a collapsed bed rather than a bed in a fluidised state.
Snapshots of particle configurations for the simulation with ideal collision parameters (e = 1. Snapshots of this simulation are presented in Figure 6. µ = 0. which stresses the crucial role of bubbles for solids mixing in fluidised beds. All other parameters were set to the same values as presented in table 6. a segregated state is reached after a few seconds and this situation does not change significantly anymore.8.0) The result is striking. Figure 6. The simulation was 174 .0.3 Effects of collision parameters Since the collision parameters turned out to have a significant influence on the bed hydrodynamics in all the simulations presented in the previous chapters an additional simulation was performed assuming fully elastic (e = 1) and perfectly smooth (µ = 0) collisions.Chapter 6 ______________________________________________________________________________________ 4.2.8. Due to the absence of bubbles there is no mechanism present to transport the bigger particles to the upper regions of the bed.
15) and that the variance is a lot less as well.and multiple frame analysis for the simulation with ideal collision parameters (e = 1.ideal collision parameters (e = 0.0) In this figure it can be observed that the segregation is far more pronounced than in the base case simulation (Figure 6. This average segregation profile is presented in Figure 6.9.96. µ = 0.0. µ = 0.6) with non. Figure 6.10.Segregation Phenomena ______________________________________________________________________________________ continued for 20 seconds and an average segregation profile over 2000 frames was obtained.1 Experimental For the purpose of experimental validation of the segregation simulations a setup was designed and constructed that is schematically represented in Figure 6. 5. Segregation profiles obtained from both single.9. Experimental validation 5. 175 .
In preliminary measurements the minimum fluidisation velocities were determined separately for both types of particles.up the particles can be observed with a video camera during the fluidisation process since glass was used as the material for the construction of the bed.5 mm. the gas velocity in the jet was set equal to the uniform inflow velocity.Chapter 6 ______________________________________________________________________________________ width 150 mm PI PI PI PI depth 15 mm PI PI height 400 mm PI PI PI porous plate pore size ~10 micron 3mm glass beads steam PI height 100 mm steam gas gas (jet) Figure 6.5 mm diameter particles a minimum fluidisation velocity (umf) of 0.% mixture the 176 . For the 1.28 m/s was found. Schematic representation of the experimental setup In this quasi twodimensional set. Coloured glass particles obtained from Worf Glaskugeln GmbH in Mainz.5 mm.92 m/s was found whereas for the 2.5 mm diameter particles a umf of 1. particle mixture was used for the validation with an initial bed height of 0. The biggest particles were coloured red and had a diameter of 2. the smallest were coloured yellow and had a diameter of 1. The fluidising gas (air at ambient conditions) was humidified in order to minimise the influence of electrostatic forces. Since homogeneous gas inflow conditions were used during the segregation experiments the option to invoke a jet was not used. A 50/50 wt. Instead. For a 50/50 wt. Germany were used that can easily be distinguished by visual observation.15 m.10.
The values of the collision parameters that were obtained from the measurements at the Open University are presented in table 6. 5. Table 6. 177 . ρp Dp.total Spherical 2525 kg/m3 Np = 6481 Np = 1400 1000 N/m 7881 Bed: Width Height number xcells.40 m 15 40 10 mm 10 mm 10–4 s The collision parameters for the particles used in the experiment were measured at the Open University at Milton Keynes using the facility of the Impact Research group that is described in Chapter 2.small = 1. 1999).big = 2. NY cell width. NX number ycells.15 m. 0.2 Results The parameter settings used for the simulations that were to be compared with the experiment are presented in table 6.3.3 m/s. For the simulation presented here the code was extended to take the various collision parameters into account. In the simulations presented in the previous paragraphs the collision parameters for particleparticle collisions were identical: no distinction was made between collisions of particles of different types. Parameter settings for the simulation duplicating the experiment Particles: Shape density. DT 0.5 mm Dp.3.Segregation Phenomena ______________________________________________________________________________________ segregation turned out to be most pronounced when the gas inflow velocity was chosen equal to 1.4 (Gorham and Kharaz.5 mm kn Np. DY time step. DX cell height.
In the simulation with a gas inflow velocity of 1.97 0.3 m/s using the hardsphere code. In the video images of the experiment bubbling appears to be far more pronounced than in the simulation.97 0.10 0. For this simulation the softsphere code was extended to take the various collision parameters into account.33 bigwall 0.4.97 0.33 bigsmall 0. 178 . Video images of the experiment and snapshots of the simulation are presented together in Figure 6.33 At first a simulation was run with a homogeneous gas inflow velocity of 1.9 m/s. At lower gas velocities segregation was observed but for these simulations the hardsphere code could not be used due to stability problems. The normal spring stiffness was set equal to 1000 N/m for all sorts of contacts and a time step of 105 s was used for the integration of the contact forces.09 0. However. Collision parameters obtained by Gorham and Kharaz (1999) bigbig e µ β0 0. after 60 seconds still no segregation could be observed which is not in agreement with the results of the experiment. Therefore the softsphere code was used to perform a simulation with a homogeneous gas inflow velocity of 0.33 smallwall 0.33 smallsmall 0.11. The hardsphere approach is not able to handle static zones in the bed and at gas velocities below umf such static zones cannot be avoided.15 0.Chapter 6 ______________________________________________________________________________________ Table 6.10 0.11 0.3 m/s the bubbling was also much more pronounced but in that particular simulation segregation did not occur. It can clearly be observed that the segregation is more pronounced and faster in the experiment than in the simulation.97 0.97 0.
11.Segregation Phenomena ______________________________________________________________________________________ t = 0 s t = 20 st = 40 s Figure 6. Snapshots of the simulation compared with video images of the experiment 179 .
(continued) Snapshots of the simulation compared with video images of the experiment 180 .11.Chapter 6 ______________________________________________________________________________________ t = 60 st = 80 st = 100 s Figure 6.
12. Therefore the simulation was continued in order to check whether the segregation would become more complete.12 snapshots of the simulation at t = 150 s and t = 200 s are presented. 225 MHz. Since the simulation was very time consuming (280 minutes of CPU time per simulated second on a Silicon Graphics Origin200 server with a R10000 processor.Segregation Phenomena ______________________________________________________________________________________ At t = 100 s the segregation was almost complete in the experiment whereas this was by far not the case in the simulation. Figure 6. In Figure 6. Snapshots of the simulation at t = 150 s and t = 200 s Indeed the segregation is more pronounced than it was at t = 100 s but still not as complete as it was in the experiment. Also the bigger particles tend to accumulate in the centre of the bed whereas in the experiment there is a sharper horizontal division between the two particle layers. 181 . 2 Mb cache) the simulation was not repeated at different gas velocities.
On the other hand particles of 1.0 mm diameter could be used instead of 1. This is most likely due to the fact that the motion of the particles is restricted to two dimensions in the simulation. It is important to realise that in the present code the hydrodynamics of the gas phase is still resolved on a length scale that is larger than the particle size. assuming that the particle velocities are identical in both cases. Segregation is expected to be more pronounced in the case of a larger size difference as was used in the simulation in the previous paragraph (1. In Figure 6.0 mm diameter particles in a bed of 15 mm depth is questionable.5 mm).5 vs. Schematic representation of two computational cells with identical void fraction but different solids distribution In the present code these two configurations are treated in exactly the same manner as far as the twoway coupling is concerned.5 mm diameter but this would increase the number of particles in the simulation and hence the computational cost.Chapter 6 ______________________________________________________________________________________ Although the softsphere model is capable of predicting segregation for this system it is clear that the agreement with the experiment is rather poor. It is however to be expected that when the gas flow is solved on a length scale 182 . This affects the twoway coupling between the particle dynamics and the gas phase hydrodynamics. However the use of 4.0 mm). Also the particle sizes used in the experiment do not differ so much (1.5 vs.13. 4. 2. Figure 6.13 two computational cells are schematically represented with each a different particle configuration but identical void fraction.
In Chapter 3 some techniques are discussed that are capable of solving the gas phase hydrodynamics on this subparticle scale. The bigger (jetsam) particles were continuously transported to the upper regions of the bed by the bubbles and then moved down aga in in the denser regions. µ = 0.01 second for a total duration of 50 seconds). In the latter case the segregation effect was more pronounced which is in agreement with experimental findings reported in the literature. This stresses the important role of bubbles in the mixing of solids in fluidised beds. An average segregation profile was determined using 5000 frames (every 0. An analysis based on a single frame was found to lead to an incomplete understanding of the segregated state. This can eventually lead to improved agreement between simulation and experiment for the case presented in this section. When the collisions were assumed to be perfectly elastic and perfectly smooth (e = 1.96.0.fluidised beds consisting of particles of different size and density were simulated using a hardsphere discrete particle model. 183 . 6. Due to the absence of bubbles segregation became almost complete after only a few seconds and this situation did not change significantly anymore during the remainder of the simulation. Such microscopic simulations can subsequently be used to improve the twoway coupling in the EulerianLagrangian simulations by for instance an improved correlation for the drag force acting on a particle.Segregation Phenomena ______________________________________________________________________________________ smaller than the particle size these two configurations will show different results in terms of twoway coupling. The results showed a rather large variance.0) a completely different behaviour was observed.15) segregation was observed on a time scale of several seconds although the system never reached a clear steady state. Using for instance the Lattice Boltzmann technique the effects of the subgrid particle configuration on the flow behaviour can be studied on a length scale smaller than the particle size. Using realistic values for the collision parameters (e = 0. Conclusions Gas. µ = 0. Segregation was observed in the case of a binary system of particles of different size as well as in the case of a ternary system consisting of particles of different density.
3 m/s segregation was observed in the experiments over a time scale of less than a minute but in the simulations no segregation could be observed over the same time scale. [] number of computational cells in ydirection.9 m/s) segregation was observed in the simulations. [] 184 . For these simulations the softsphere code had to be used due to the presence of static zones. m/s2 moment of inertia. kg Number of particles. m vertical computational cell dimension.5 mm vs. [] number of computational cells in xdirection. m time step. N gravitational acceleration. At a lower gas inflow velocity (0. kgm2 spring stiffness. N/m particle mass. [] force.5 mm diameter) which renders a tough test for the model. [] width of neighbour list square.Chapter 6 ______________________________________________________________________________________ Finally an experimental setup was designed and constructed for the purpose of experimental validation. Furthermore the difference in the particle sizes used for the validation is rather small (1. However the time scale over which the segregation occurred and the final segregation pattern were in poor agreement with the experimental results. m coefficient of restitution. s horizontal computational cell dimension. 2. This is most likely due to the fact that the motion of the particles in the simulation is restricted to two dimensions in contrast with the experiment. m particle diameter. Notation Cd Dnblist Dp DT DX DY e F g I k m Np NX NY drag coefficient. At a gas inflow velocity of 1.
m Subscripts g mf n gas phase minimum fluidisation normal 185 . [] time. Pa particle radius. [] pressure. [] void fraction. K tangential unit vector. m/s particle velocity vector. s1 displacement. N/m3 torque. Ns/m gasphase stress tensor. [] coefficient of friction. kg/ms damping coefficient. m/s particle volume.Segregation Phenomena ______________________________________________________________________________________ n p Rp r Sp T T t t u v Vp normal unit vector. m momentum source term. kg/m3 angular velocity. s gas velocity vector. kg/m3 s coefficient of tangential restitution. m3 Greek symbols β β0 ε µ µg η τ ρ ω ξ defined in Eqs () and (). [] gas viscosity. m position vector. Nm temperature. kg/ms2 density.
Janssen.M. Ding. Boston. J..F.Y. (1999).J. Discrete particle simulation of bubble and slug formation in a twodimensional gas. Gorham. H.. Engng Sci. Ind. 47. B. A discrete numerical model for granular assemblies. Cundall. Géotechnique. O.H.A.B. Multiphase Flow and Fluidization. (1993)..Chapter 6 ______________________________________________________________________________________ nblist p t w neighbour list particle tangential wall References Anderson.P. and Jackson. Impact Research Group Report IRG 13.M.. D.A. Foerster S. 48. M. Particle segregation in fluidised binary mixtures. W. Engng Sci. (1994). Chem. UK. Chem. U.A. Gidaspow... (1967). Hoomans. A..K.B. 51..L.. Briels. 29. Gidaspow..D.B. (1996). Louge. D. Kuipers.C. 1108. W. and Jayaswal. Hoffmann. USA.P. 6. 47.. 1583.. K.. Milton Keynes. and Strack. P. Fluids. L. (1979). FED 91.. Phys. and Kharaz. Academic Press. J. A. Results of particle impact tests. (1990).M. Numerical Methods for Multiphase Flows. R. and Prins. 186 . Chang. and van Swaaij. The Open University. Measurements of the collision properties of small spheres.. Eng. T. and Allia. (1994). D.P. 527. Chem. 99. 6. A fluid mechanical description of fluidized beds (equations of motion).fluidised bed: a hardsphere approach. Multiphase NavierStokes equation solver. Fundam. J.
and van Swaaij. (1992). E. Chem. A...L. Engng Sci. Studies of segregation/mixing in fluidised beds of different size particles.N. and Naimer.D. Continuous mixing of two particulate species of different density in a gas fluidised bed. and Zaki. Trans IChemE. Norway. Tanaka. van Beckum F. and Koch.M. W. Proc. 62. Fluid Mech. T. (1954). T. 1913. A. Manger.J. Chem. and Tsuji. Nienow. Richardson.S. 1927. T. and Hjertager. V.. Telemark Institute of Technology.. Inst. Chem. 47.P. 187 . A computational study of multiphase flow behavior in a circulating fluidized bed with realistic particle size distribution. Y. Chem. 35. B. Trans. K. 247. J.. Naimer. 812 June 1998. (1998). Kamiya. H. Eng. N. Thesis. Eng.. Properties of a bidisperse particle..and threedimensional models). (1993). 58.P. Numerical simulation of twodimensional fluidized beds using the discrete element method (comparison between the two..W.H. J..S. Engng Sci.. France Mikami. (1996).. N.W. Mathiesen. Modelling and simulation of gas/solids flow in curvilinear coordinates. (1987). Porsgrunn. Collision time small compared with viscous relaxation time.. Comm. and Horio. (1998). Sedimentation and fluidization: part I. Ph. van Duin. T. Kumaran. 623. Lyon.. Kuipers.Segregation Phenomena ______________________________________________________________________________________ Kawaguchi. A numeric al model of gasfluidized beds.F. 53.A.M.. (1998). 129. On Multiphase Flow.gas suspension Part 1.. W. 32. 3rd Int. Powder Technol. 96. Nienow. and Chiba. D. M. V. 181.H. (1980).. J. Solberg T. Numerical simulation of cohesive powder behavior in a fluidized bed. Conf. 52.
. (1993). Instn Chem. Rev. M. and Mason. Prog. Ser. (1992). M..H. Wang. (1997). 79. Powder Technol. T.Chapter 6 ______________________________________________________________________________________ Rowe. 100. (1966). Twodimensional rigidbody collisions with friction. A.A. Appl. 98.. (1998). (1998). C. Seibert. 62 (62). A. Discrete particle simulation of two dimensional fluidized bed. S. K. Schwarzer.W. Engrs. Segregation by size difference in gas fluidized beds.. Symp. Wu. 310. Trans. and Tanaka.. Kawaguchi. Chem. Numerical simulation of the gassolid flow in a fluidized bed by combining discrete particle method with computational fluid dynamics. T.. and Agbim. S. Y. Powder Technol. 139. Mech. (1995).H. Nienow. 6461.Y. (1972).N. Phys. AIChE J. and Yu. Mechanics of fluidization..Y. 635. Y. 50. Y. E. 52. J. and Yu.B. A. J. The mechanism by which particles segregate in gas fluidised beds: binary systems of nearspherical particles.. 44. 52. Tsuji. Wen. B. and Burns.J. Eng. and Baeyens. 59. 528. Sedimentation and flow through porous media: Simulating dynamically coupled discrete and continuum pha se. Simulation of structural phenomena in mixedparticle fluidized beds.D.. Engng Sci. Chem. 77. 188 . 2785. P. Xu.T.
The gas phase hydrodynamics is described by the spatially averaged NavierStokes equations for twophase flow. A strong effect of the collision properties on the axial solids profile was found where a pronounced buildup of solids was observed in the simulation with realistic values for the collision parameters. This result is supported by the experimental findings of Chang and Louge (1992) where a smaller pressure drop over the riser was found when glass particles were used to which a coating was applied in order to decrease the coefficient of friction. Particlewall collisions turned out to have very little influence on the flow structure. coefficient of restitution (e) and coefficient of friction (µ)).0.94. µ = 0.28). µ = 0. THE INFLUENCE OF COLLISION PROPERTIES ON THE FLOW STRUCTURE IN A RISER Abstract: A twodimensional hardsphere discrete particle model was used in order to simulate the gassolid flow in the riser section of a circulating fluidised bed (CFB).e. Lift forces acting on the suspended particles turned out to have a slightly redispersive effect on the flow structure which made the radial segregation of the solids a little less pronounced.Riser Flow ______________________________________________________________________________________ Chapter 7. The results proved to be very sensitive with respect to the collision parameters (i. 189 . The Newtonian equations of motion are solved for each individual particle in the system while taking into account the particleparticle and particlewall collisions. In the case of fully elastic and perfectly smooth collisions (e = 1. In the simulation assuming fully elastic and perfectly smooth collisions no buildup of solids was observed.0) hardly any clustering of particles could be observed as opposed to the case where these collision parameters were assigned realistic values (e = 0.
Kuipers and W.). June 812. van Swaaij. Werther (Ed. Hoomans. Circulating Fluidized Bed Technology VI.P.B. J.A. van Swaaij. (1998). Kuipers and W.M. (1999). 190 .M.A.P.M. 266.P. B.M.P.Chapter 7 ______________________________________________________________________________________ Parts of this chapter are based on the papers: B. p. Discrete particle simulation of cluster formation in dense riser flow. J. Granular dynamics simulation of cluster formation in dense riser flow. Hoomans. Lyon. J. Third International Conference on Multiphase Flows 1998.B. France.
(1996)) did not focus particularly on cluster formation but more on the radial segregation of solids. (1998) demonstrated that the radial solids segregation in a riser (i.62 mm width and 76. Other TwoFluid approaches incorporating the kinetic theory of granular flow (Sinclair and Jackson (1989). Although a clear definition is lacking (Chen. 191 . After time averaging the typical coreannulus flow struc ture where the average solids concentration is considerably higher near the wall was obtained. Introduction Cluster formation is a characteristic feature of dense gas solid flow in the riser section of a Circulating Fluidized Bed (CFB).e. They reported the formation of cluster. They found that the clusters had a characteristic parabolic shape. Kuipers et al.like structures even though the spatial resolution in their simulation was rather low (computational cells of 7. Tsuo and Gidaspow (1990) used a TwoFluid approach with constant solids viscosity in order to simulate the riser section of a CFB. This was (among others) pointed out by Hrenya and Sinclair (1997) who also reported that when particle phase turbulence was included in the model the aforementioned dependency became far less pronounced. This type of model possesses a peculiar dependency on the magnitude of the coefficient of restitution where a small deviation from unity causes the flow structure to change completely while the agreement with experimental findings deteriorates. Since the existence of such clusters has a profound influence on the performance of a CFB unit as a chemical reactor they have been the subject of a number of studies.Riser Flow ______________________________________________________________________________________ 1. coreannulus flow) has a negative effect on the performance of the riser as a chemical reactor. Radial solids segregation in a riser is largely due to the existence of clusters that tend to accumulate near the wall. Nieuwland et al.2 mm height) compared to typical cluster sizes. Horio and Kuroki (1994) experimentally observed clusters in dilute riser flow using a laser sheet technique. 1995) clusters can be typified as regions of locally higher solid fraction.
2. that determine how much energy is dissipated in collisions. Hence this method does not account for actual particleparticle and particlewall interaction in a direct way. (1994) have shown that also the coefficient of tangential restitution (0 ≤ β0 ≤ 1) should be used in order to describe the collision dynamics more accurately. In a later paper Tsuji et al. (1996). turned out to be of key importance in their simulations. Model Since a detailed description of the model is presented in Chapter 2 and Chapter 3. Moreover the modified Nanbu method used in their work does not guarantee exact conservation of energy (Frezzotti. (1998) made a comparison between the work of Tsuo and Gidaspow (1990) and Tanaka et al. They employed the Direct Simulation Monte Carlo (DSMC) method to describe the particle dynamics. 1997). In this chapter a EulerianLagrangian simulation technique for dense gassolid twophase flow in a riser is presented which features direct incorporation of the particleparticle and particlewall interaction. (1996) performed simulations of gassolid flow in a vertical duct using the Lagrangian approach for the solid particles. instantaneous. Recently Helland et al. In the DSMC method the simulated particles actually represent a certain number of ‘real‘ particles. the key features will be summarised briefly here. Foerster et al. A random number generator is invoked to determine the collision partners and the geometry of the collision. These 192 .1 Granular dynamics The collision model as originally developed by Wang and Mason (1992) is used to describe a binary. This can be an important drawback especially since the collision parameters. (1999) and Ouyang and Li (1999) presented discrete particle approaches similar to the one used in this work but in their simulations collisions are detected at constant time intervals allowing the particles to slightly overlap. 2. The key parameters of the model are the coefficient of restitution (0 ≤ e ≤ 1) and the coefficient of friction (µ ≥ 0). inelastic collision with friction.Chapter 7 ______________________________________________________________________________________ Tanaka et al.
7 p 8 (7. For each particle a list of neighbouring particles is stored and only for the particles in this list a check for possible collisions is performed. In order to reduce the required CPU time neighbour lists are used.1) where mp represents the mass of a particle and vp its velocity. the coefficient of tangential restitution is assumed to be equal to zero. In equation 7. Since in riser flow the velocity differences between the particles is larger than in bubbling fluidised beds the neighbour list is chosen rather large (Dnblist = 8 Dp ) to ensure that all collisions are properly detected.Riser Flow ______________________________________________________________________________________ three collision parameters are all included in the model. This implies that a collision list is compiled in which for each particle a collision partner and a corresponding collision time is stored. In the hardsphere approach a sequence of binary collisions is processed.2) In equation 7. In all simulations presented in this chapter. A constant time step is used to take the external forces into account and within this time step the prevailing collisions are processed sequentially. It should be noted that the drag correction due to the presence of other particles has been accounted for via the wellknown Richardson and Zaki (1954) correction factor.2 External forces The equation of motion used for the simulations of riser flow is: mp dv p dt = m p g + Fdrag + F LG + F LM (7.2 ρg represents the density of the gas phase and u represents its velocity. 2.1 the first term on the right hand side is due to gravity and the second term is due to drag: 1 Fdrag = πD 2 C d ρ g u − v p (u − v p )ε − 2. The drag coefficient for an isolated particle depends on the particle Reynolds number as given by Rowe and Henwood (1961): 193 .
The second term on the right hand side of equation 7.687 Cd = Re p 0. 1 F LM = πD 2 C LM ρ g u − v p p 8 ωr (7.6) 194 . also known as the Magnus lift force: (u − v p ) × ω r . Following the lines of Tanaka et al. The lift forces are not used by default but since an additional simulation was performed in order to test their influence the lift forces will be discussed in more detail here.y)/∂x.3) where the particle Reynolds number is defined as: Re p = ερg u − v p D p µg .4) Gravity and drag force are implemented by default in all simulations shown in this work.1 is the lift force due to particle rotation. (7. x = πD 2 C LG ρ g u − v p sign ( χ ) p 8 (7. The third term on the right hand side in equation 7. The empirical relations used for the lift coefficient CLG are summarised in table 7.15(Re p )0 .1 is the lift force due to a velocity gradient. (1996) only the gradient in the xdirection is considered: 2 1 FLG .Chapter 7 ______________________________________________________________________________________ 24 1 + 0.1.44 ( ) Rep < 1000 Rep ≥ 1000 (7.5) where χ denotes the velocity gradient ∂(uy vp.
7) Table 7. 1994) 500 < Re p (Tsuji et al.2 α = 6. Sa Re p = 1 − 0. but also by the torque exerted on them by the fluid: 195 .2 α = 6. β = 32.40γ Cm = 1 −1 α Re ω 2 + β Re − 1 ω π 10 < Re p < 500 (Oesterlé.1.44. Empirical relations for the lift coefficients C LG C LG . Sa C LG C LG . 1980) Re ω = ρ g ω r D2 p 4 µg ( ) α = 5.1 The rotational velocity of particles is not only affected by collisions with the system walls or with other particles. ξ = Dp χ 2 u −vp γ = 2 and Re χ = ρ g χ D p 4µ g Rep < 10 with: D p ωr 2 u− vp C LM = 3. 1965). 0741 Re 1 / 2 χ with: C LG . 1985) with: 10< Reω ≤ 20 20< Reω ≤ 50 Re ω ≥ 50 (Dennis et al.3314 ξ 1 / 2 Re p ≤ 40 (Mei.. Sa = 8. (7. 1992) Re p > 40 = 0. β = 32.3314 ξ 1/ 2 exp − 10 ( ) + 0.45. β = 37.8γ Re p −1 / 2 C LM = 0.225 C LM = 2γ Re 1 / 2 χ Re (Saffman.32.Riser Flow ______________________________________________________________________________________ where ωr is defined as follo ws: 1 ωr =ω − ∇ ×u 2 The empirical relations used for the lift coefficient CLM can be found in table 7.1.
An initial axial velocity of 0.8 a simple explicit first order scheme was used to update the velocities and positions of the particles. The cross sectional area in this (twodimensional) case was assumed to be equal to the width of the system times the particle diameter.3 Inlet and outlet conditions In the riser section of a CFB there is a continuous transport of particles throughout the system that has to be taken into account in the simulation. It is based on the numerical solution of the following set of partial differential equations that can be seen as a generalised form of the NavierStokes equations for a gas interacting with a solid phase as originally derived by Anderson and Jackson (1967). (1992). p dt 64 (7. At each time step a specific number of particles was fed to the riser duct in accordance with a specified solids mass flux (Gs). including inlet and exit configurations. 196 .8) The empirical relations used for the rotational drag coefficient Cm are summarised in table 7. At the outlet particles which had crossed the upper boundary were simply removed from the system at each time step. only the middle section was simulated. Instead of simulating the entire riser section. For the integration of equation 7. 2.4 Gas phase hydrodynamics The calculation of the gas phase hydrodynamics mainly follows the lines presented by Kuipers et al.Chapter 7 ______________________________________________________________________________________ I dω 1 = − πD 5 C m ρ g ω r ω r .1 as well as equation 7. The particles were only allowed to leave the system at the upper boundary. At the inlet of the simulated system particles were placed at random in the bottom row of cells ensuring that there is no overlap with other particles or walls.4 m/s was assigned to each particle fed to the system. 2.1.
10) In this work transient. Firstly the main goal of this work is to study the influence of particle properties on the flow structure. (1999) did include an additional term in the momentum equation of the gas phase in order to account for sub. isothermal (T = 293 K) flow of air at slightly elevated pressure (p = 1.9) Momentum equation gas phase: ∂ (ερ g u ) ∂t + (∇ ⋅ ερ g uu ) = −ε∇ p − S p − (∇ ⋅ ετ g ) + ερ g g . However this term was taken directly from a singlephase flow model without any modification for the presence of the particles. There is one important modification with respect to the model of Hoomans et al. Note that no turbulence modelling is included in the present description. It cannot be denied that the assumption of laminar flow is questionable for the type of flow encountered in the simulations in this chapter but there is no doubt that the application of a singlephase flow turbulence model in this situation is questionable as well. (1996) and that deals with the way in which the 197 .grid scale effects of turbulent eddies.2 Bar) is considered. The void fraction (ε) is calculated from the particle positions in the bed. Secondly the particles used in the simulations in this chapter are rather coarse (500 µm) and are therefore not much influenced by turbulent eddies as for instance is the case for FCC particles (70 µm). ( ) ∂t g (7. And last but not least there is no generally accepted and experimentally validated turbulence model for the type of twophase flow that is the subject of study here. (7. Helland et al. twodimensional. The reasons for this can be summarised as follows. The constitutive equations and the boundary conditions used can be found in Chapter 3.Riser Flow ______________________________________________________________________________________ Continuity equation gas phase: ∂ ερg ( ) + ∇ ⋅ ερ u = 0 .
08 m 2.2 Bar 5. In the work of Hoomans et al. The general parameter settings are summarised in Table 7.00 m temperature. Therefore three different simulations were performed where the collision parameters were set according to the values presented in Table 7. General parameter settings particles system gas diameter. p velocity.2. Tanaka et al. 198 .3. Gs 500 µm 2620 kg/m3 25 kg/m2 s width height 0.0 m/s 3. In all simulations a time step of 104 s was used together with a computational grid of 20 (NX) by 100 (NY) cells.1 Effect of collision parameters At first the influence of the collision parameters was examined. 3. (1996) reported a strong dependency of the flow structure with respect to the coefficients of restitution (e) and friction (µ). Table 7.2. (1996) on bubbling and slugging fluidised beds these parameters also turned out to have a decisive influence on the flow behaviour. (1996) in their simulations. Dp density.Chapter 7 ______________________________________________________________________________________ twoway coupling between the gasphase and the particle motion is established. In the present model the reaction forces to the forces exerted on a particle by the gasphase per unit of volume is fed back through the source term Sp which has the dimension Nm3 . T pressure. Results As a first test case for the model the same system was chosen as was used by Tsuo and Gidaspow (1990) and Tanaka et al. ug 293 K 1. ρ solids mass flux.
ideal collision parameters it took more than 30 s before the outlet solids mass flux was equal to inlet solids mass flux.0 ew 1.28). µ = 0.1.3.0 0.28 0.0 0.ideal collision parameters.1 the outlet solids mass flux is presented as a function of time for the simulation with the non.94 1.94 0.28 e Ideal non.0 0.0 µw 0. Figure 7. 199 . Settings for the collision parameters µ 0. Once the steady state was reached the simulations were run for 10 s in order to obtain statistically reliable time averages. Outlet solids flux as a function of time for the simulation with nonideal collision parameters (e = 0.Riser Flow ______________________________________________________________________________________ Table 7.28 0.94.94 At first it was checked if each simulation had reached a steady state where the outlet solids mass flux was equal to the specified solids mass flux of 25 kg/m2 s. For the simulation with ideal collision parameters this steady state was reached within 10 s but for the simulation with non. In Figure 7.0 0.ideal Ideal particles/ nonideal walls 1.
1 it can be observed that the outlet solids mass flux has a very peaky nature but the time average value is equal to the specified mass flux (Gs = 25 kg/m2 s). In Figure 7.2.28) collision parameters are shown in Figure 7.2 m height).0. e = 1.0 to 1. µ = 0. µ = 0.Chapter 7 ______________________________________________________________________________________ Note that t = 0 s in this figure corresponds to the initial condition for the actual simulation of the system at steady state conditions. The solids flux during the time required to reach this steady state is not shown here.94 and µ = 0. The snapshot of the simulation with the ideal collision parameters shows a much more homogeneous distribution of the particles.2 a typical cluster can clearly be observed in the case of nonideal collision parameters.28.08 m).0) and nonideal (e = 0. µ = 0.0. In Figure 7. Figure 7. Left: ideal collision parameters.Two snapshots of simulations with ideal (e = 1.94. Right: nonideal collision parameters: e = 0.2. It should be noted that in the simulations presented by Tsuo and Gidaspow (1990) and Tsuji et al.0. (1998) the time averaged outlet solids mass flux was by far not equal to 25 kg/m2 s and therefore their results do not apply to fully developed riser flow. These are snapshots of the middle section of the riser (1. Snapshots of a small section of the simulated riser (width = 0. 200 .
causes the typical coreannulus structure since no cluster formation. 0. grow and break up. these can hardly be identified as clusters. and hence no coreannulus structure.04 0. Typical downward velocities for these clusters are in the order of magnitude of 1.0 m above the inlet of the simulated riser. This result is a strong indication that the presence of clusters.In Figure 7.06 0.0 m height is presented.ideal parameters for particlewall collisions.08 x.02 0. which is clearly observed in the non.00 ideal particles. For the case with non. Radial profile of the time averaged solids volume fraction at 1. m Figure 7.04 0.06 0. [] ideal particles and walls 0. In the case of ideal collision parameters the flow was much more homogeneous and although locally slightly denser regions can be distinguished. 201 .02 0.10 solids fraction.Riser Flow ______________________________________________________________________________________ From animations of the simulations it could be observed that in the simulation with the non. could be observed in the other two cases where no energy was dissipated in particleparticle collisions.3 the time averaged (10 s) radial solids profile at 1. In the other two cases the radial profile is flat.00 0.08 nonideal particles and walls 0. Similar results were obtained in a simulation with ideal collision parameters for particleparticle collisions but non.ideal collision parameters the clusters moved downward preferably along the walls and had a streaky. nonideal walls 0.0 m/s which is in agreement with the experimental findings of Tsuo and Gidaspow (1990). These clusters are rather dynamic in nature: they continuously form.ideal collision parameters the typical coreannulus flow structure can be observed where the solids fraction is relatively high near the wall and relatively low in the centre of the riser duct.3. semiparabolic shape.ideal case.
000.ideal collision parameters than in the simulations with ideal particleparticle collisions. The simulation with non. Figure 7.4. e =1. Right: nonideal collision parameters: e = 0.0. µ = 0. By the time the steady state was reached a total number of about 60.2 m. µ = 0.08 m) from 0. This had also a pronounced effect on the CPU time.28. This suggests that there is an effect of the collision parameters on the axial solids profile in the riser. Left: ideal collision parameters.0. Snapshots of the bottom section of the simulated riser (width = 0.ideal collision parameters required over 27 hours of CPU time per simulated second on a Silicon Graphics Origin200 server with a R10000 processor (180 MHz.0 to 0.Chapter 7 ______________________________________________________________________________________ It is also remarkable that the average solids hold.94. 1 Mb cache). 202 .up is higher in the simulation with the non.ideal collision parameters much longer to reach a steady state turned out to be the build up of a solids profile in the riser.2 Axial effects The reason why it took the simulation with the non. 3.000 particles was found inside the system whereas in the simulations with ideal particleparticle collisions this number was about 25.
10 solids fraction. Large strands of particles move down alongside the walls and then break up again after impinging on the bottom wall. In Figure 7. [] 0.08 0.00 0. Radial profile of the time averaged solids volume fraction at 0.2 but in the bottom section the difference is even more pronounced. 0. In this figure the same difference between ideal and nonideal collision parameters can be observed as in Figure 7. 203 .08 x.4 a snapshot of the bottom section of the riser is presented for the simulations with ideal and non.Riser Flow ______________________________________________________________________________________ The simulations with ideal particleparticle collisions required 2.2 m above the inlet of the simulated riser.Since clusters of particles move downward in the riser and the particles cannot leave the system at the bottom it can be understood that an axial solids profile will event ually be established.2 m above the bottom is presented for the three different cases. In Figure 7.ideal collision parameters.5 hrs of CPU time per simulated second on the same machine. 65 million collisions per second for the simulation with non. m Figure 7.02 0.ideal collision parameters.06 ideal particles and walls nonideal particles and walls ideal particles.5.04 0. nonideal walls 0. The reason for this huge difference is the total number of collisions that had to be processed in the simulations: 1 million collisions per second for the simulation with ideal collision parameters.00 0.04 0.5 the time averaged (10 s) radial solids profile at 0.02 0.06 0.
This suggests the existence of a nonuniform axial solids profile.ideal particlewall collisions indicating that particleparticle collisions dominate the flow behaviour. these results are very similar to the results of the simulation with ideal collision parameters.000).ideal collision parameters but not in the simulations with ideal particleparticle collisions.28. Again. ideal: e = 1. µ = 0.3 the coreannulus flow structure can be observed in the simulation with non.6 the time averaged (10 s) axial distribution of the solids volume fraction in the riser are presented for the two cases mentioned above. Axial profile of the timeaveraged solids volume fraction (nondeal: e = 0. For the non. It is also clear that the average solids hold.0 m (Figure 7.0) 204 . hardly any difference could be observed between the simulation with both ideal particleparticle and particlewall collisions and the simulation with ideal particleparticle but non.ideal particlewall collisions are not included for clarity reasons.up for the simulation with the non.ideal case clearly a build up of solids in the riser can be observed which is in agreement with the different total number of particles in the simulations reported earlier (app.94.0.3). However.ideal collisions at 0. 25. 60.2 m in the riser is higher than at 1. Figure 7.Chapter 7 ______________________________________________________________________________________ Similar to the radial solids profiles presented in Figure 7. In Figure 7. µ = 0.000 vs. In this figure the results of the simulation with ideal particleparticle but non.6.
ideal particle collisions whereas in the case of ideal collisions the axial solids distribution is nearly uniform.e. Chang and Louge (1992) reported results from experiments using a somewhat different system but operated at a rather similar mass loading and Froude number. They found a significant change in axial pressure profile in the riser when they applied a coating to glass particles in order to make them smoother. higher values for the coefficient of restitution) less clusters are being formed and hence there is less internal solids circulation within the riser.Riser Flow ______________________________________________________________________________________ In Figure 7. the more ideal system: lower values for the coefficient of friction. lower coefficient of friction) the pressure drop over the riser was much less than for the noncoated glass particles indicating a lower solids hold up in the system. Although the system appeared less homogeneous than at a solids flux of 25 kg/m2 s still the formation of clusters was far less pronounced than in the nonideal case with a solids flux of 25 kg/m2 s. In the case where less energy is dissipated in collisions (i. For the coated glass particles (i. A simulation was performed with ideal collision parameters at a solids flux of 50 kg/m2 s in order to investigate this. This results in a lower solids hold up in the riser and hence a lower pressure drop. It is therefore of great importance that the collision parameters (i.e. 205 .e. The results of the simulations presented above support the idea that the difference in the axial pressure profile found in these experiments is caused by the difference in particleparticle interaction in both cases. coefficient of restitution and friction) are taken into account in the scale up as well as the design of riser reactors. Obviously there is still a large difference between the simulated system with ideal collision parameters and the coated glass particles used in the experiments. Nevertheless the simulation results clearly show the important effect of the collision parameters on the flow structure in a riser and this result is supported by the experiments of Chang and Louge (1992).6 it can clearly be observed that a build up of solids exists in the case of the simulation with non. Chang and Louge (1992) pointed out that these effects become less pronounced at higher solids fluxes.
The case with the non.8) were not included in the simulations. an additional simulation was performed including all these effects in order to investigate their influence on the flow structure. From Figure 7. This can be understood since the lift forces tend to move the particles toward the centre of the riser and therefore have a redispersive effect. The radial segregation is slightly less pronounced when the lift forces are included in the equations of motion.2 m above the inlet of the simulated riser.1 and the effect of the gas flow on the rotation of the particles (equation 7. Influence of the lift forces on the timeaveraged radial distribution of the solids volume fraction at 0.7 it is apparent that the lift forces did not have a significant influence on the flow structure.Chapter 7 ______________________________________________________________________________________ 3. 0.3 Influence of lift forces By default the two lift forces in equation 7.ideal collision parameters and a superficial gas velocity of 5.06 0. The time averaged (10 s) radial distributions of the s olids volume fraction at 0. 206 .08 x. However.00 0.00 solids fraction.02 0.7 for the two simulations. [] without lift forces with lift forces 0. m Figure 7.04 0.7.04 0.2 m above the riser inlet are presented in Figure 7.02 0.10 0.08 0.0 m/s was chosen as a test case.06 0.
µ = 0) no cluster formation was observed.00 m height at a solids mass flux of 25 kg/m2 s. µ = 0. When the collision parameters were set to ideal values (e = 1.ideal case the solids holdup in the riser was much higher than in the ideal case as could be observed from axial solids volume fraction profiles.up) was found for the coated (smoother) particles which is in agreement with the trend observed in the simulations.ideal collision parame ters. These forces turned out to have a slightly redispersive effect on the flow structure which made the radial segregation of the 207 . This was also the case in a simulation where only particlewall collisions were considered nonideal whereas particleparticle collisions were still taken to be ideal. decrease the coefficient of friction). In those experiments a smaller pressure drop (i.e. The time averaged radial distribution of the solids volume fraction clearly revealed the typical coreannulus structure for the simulation with the non. smaller solids hold. These results are supported by the experimental findings of Chang and Louge (1992) where a large difference in axial pressure profile was found when a coating was applied to glass particles in order to change the surface roughness (i. In the non. For the simulations with ideal particleparticle collisions the time averaged radial distribution of the solids volume fraction was rather flat.e. By default the superficial gas velocity was set equal to 5. The simulations turned out to be rather sensitive to the collision parameters. i.28) cluster formation was observed.0 m/s.08 m width and 2. Particlewall collisions turned out to have very little influence on the flow structure. Finally the influence of the lift forces was investigated. When these parameters were assigned nonideal values (e = 0. Particles of 500 µm diameter and a density of 2620 kg/m3 were introduced in a riser of 0. Conclusions Granular Dynamics simulations of riser flow have been performed using a twodimensional hardsphere discrete particle model.94.Riser Flow ______________________________________________________________________________________ 4.e. the coefficients of restitution (e) and friction (µ).
m/s2 particle mass. kg/m2 s gravitational acceleration. [] number of computational cells in ydirection. m particle diameter. [] coefficient in expression Magnus lift force (table 7.Chapter 7 ______________________________________________________________________________________ solids a little less pronounced. [] pressure. [] rotational drag coefficient. [] . m y coordinate. m3 x coordinate. N solids mass flux. m moment of inertia. [] coefficient of restitution. m/s particle velocity vector. m/s particle volume. K time. s gas velocity vector. Pa momentum source term N/m3 temperature. kgm2 force. Since this was only a minor effect it could be regarded as a justification for the fact that the lift forces were neglected in the default simulations.1). kg number of computational cells in xdirection. [] width of neighbour list square. Notation Cd Cm e Dnblist Dp I F Gs g mp NX NY p Sp T t u vp Vp x y drag coefficient. m Greek symbols α β 208 coefficient in expression Magnus lift force (table 7.1).
A fluid mechanical description of fluidized beds (equations of motion). kg/ms2 density. and Louge.. 1/s Subscripts g LG LM p w gas phase lift force due to velocity gradient Magnus lift force particle wall References Anderson. kg/m3 dimensionless velocity gradient.Riser Flow ______________________________________________________________________________________ β0 χ ε µ µg τ ρ ξ ωp ωr coefficient of tangential restitution... 70. M. Chang. Clusters (The Thomas Baron Plenary Lecture 1995). [] particle rotation velocity. AIChE Symp. Chen.. Chem. Ser. Powder Technol. [] velocity gradient. H.C.. J. 1. 1/s relative rotation velocity gas/particle. 6. [] coefficient of friction. 259. 313). 1/s void fraction. kg/ms gas phase stress tensor. Fluid dynamic similarity of circulating fluidized beds. Eng. (1967).B. Ind. and Jackson. Fundam. R. 92 (No.. (1992). 209 .Y. 527. (1996). T. [] gas viscosity.
. K.). 49.. S. T. Singh. C. J.. and van Swaaij. H. Horio.J.B.H. Chem. Hrenya.B.A. Hjertager. (1997). (1999). Hoomans.H. Foerster S. p.. and Allia. in Fluidization IX. Briels. 1329.M. Louge. M. J. (1994).A. (1997). Fluids. E. 853. T. K.. 1108. 6. p. Threedimensional flow visualization of dilutely dispersed solids in bubbling and circulating fluidized beds. (1992).M. J. Frezzotti. B. and Mathiesen. 2413.B. 51.. Fan and T.P.. M. 99. J.. Effects of particle phase turbulence in gassolid flows... 43. Helland. Measurements of the collision properties of small spheres.M. and van Swaaij. Hydrodynamic models of gasfluidized beds and their role for design and operation of fluidized bed chemical reactors.J..C.. Fluids.M. W.. S. B. Chang.S. 257. and Sinclair. 9. 1913.A. L.. 261.. 15 210 . Kuipers.. Solberg. W..P. Phys.M.R. J. van den Moortel.P. Ingham. Engng Sci. Occelli..P. Kuipers. Hoomans.M. J. B.Chapter 7 ______________________________________________________________________________________ Dennis. van Beckum... W.fluidised bed: A hardsphere approach. Discrete particle simulation of bubble and slug formation in a twodimensional gas. Chem. (1994). The steady flow due to a rotating particle at low and moderate Reynolds numbers..Y. Kuipers.. Numerical/experimental study of gasparticle flow behaviour in a circulating fluidized bed.M. Engng Sci.. 47. A. Engng Sci. (1980).P.. F.F.. Werther (Ed. Fluid Mech. W. Tadrist. (1998).H. R. (1996). and van Swaaij. Chem. in Circulating Fluidized Bed Technology VI. V. Knowlton eds. 101 (2). D. Kuroki.. AIChE J. A numerical model of gasfluidized beds.P.M.... van Duin. Phys. A particle scheme for the numerical solution of the Enskog equation.N. L.
Engrs. W. 37. 39. M. Chem. Powder Technol. Experimental measurement of the Magnus lift force on a rotating sphere at low Reynolds numbers. Sinclair. Engng.. 54. (1992). Kuipers.. 107. 35. J. Ouyang. B. An approximate expression for the shear lift force on a spherical particle at finite Reynolds number. 18. T.P. Discrete simulations of heterogeneous structure and dynamic behavior in gassolid fluidization. G. P.Riser Flow ______________________________________________________________________________________ Nieuwland. Y. Trans.N.. 32. Tsuji. R.N.F. Trans. 22(2). Fluids Eng.. Richardson.. Kiribayashi. 79. J. (1999). (1985). Cluster formation and particle. (1954). R. Inst. Saffman. and Mizuno. 385.. (1996)... 39. Une étude de l’influence des forces transversales agissant sur les particules dans les écoulements gazsolide. van Sint Annaland. Instn Chem.. Morikawa. J. 145. 43. and Li. 35. Multiphase flow. Gasparticle flow in a vertical pipe with particleparticle interactions.. Tanaka.. Hydrodynamic modelling of gasparticle flows in riser reactors. JSME Int. Oesterlé.. 1473. 5427.. 239. and Henwood. Tsuji. Series B.induced instability in gassolid flows predicted by the DSMC method..A.M.G.M.. and Jackson. O. Y. AIChE J. Engng Sci.. and van Swaaij. Mei. P. 484. J... Sedimentation and Fluidisation: Part I. Drag forces in a hydraulic model of a fluidised bedpart I...A. Chem. (1996). J. K.. Rowe. (1994). 211 . W. AIChE J. J. Yonemura. S. (1961). Fluid Mech.J. J.. 1009. Y. (1989). J. The lift on a small sphere in a slow shear flow. (1965). and Zaki. J. J. 81. Int.
and Gidaspow... T. M. Tanaka. 95. Y. Appl. Y.. 59.fluid model). Powder Technol.P. Twodimensional rigidbody collisions with friction. D. Y.. Tsuo. Mech. 212 . (1992). and Mason. and Yonemura. J. Cluster patterns in circulating fluidized beds predicted by numerical simulation (discrete particle model versus two. (1990). Computation of flow patterns in circulating fluidized beds.. (1998). 885. 254.. 36.. 635. AIChE J. Wang. S.T.Chapter 7 ______________________________________________________________________________________ Tsuji.
5 mm were used as the bed material. 213 . The simulation data was time averaged over 45 s for each particle and subsequently ensemble averaged over all the particles in the simulation. independent measurements.000 particles was tracked during a period of 45 s. EXPERIMENTAL VALIDATION OF GRANULAR DYNAMICS SIMULATIONS OF GASFLUIDISED BEDS WITH HOMOGENOUS INFLOW CONDITIONS USING POSITRON EMISSION PARTICLE TRACKING Abstract: A hardsphere Granular Dynamics model of a twodimensional gasfluidised bed was experimentally validated using Positron Emission Particle Tracking (PEPT). The collision parameters required in the simulation were obtained from separate. occupancy plots and speed histograms. A quasi twodimensional bed of 0. The gas phase hydrodynamics is described by the spatially averaged NavierStokes equations for twophase flow. Glass particles (ρ p = 2435 kg/m3) with diameters ranging from 1. The comparison was made on the basis of averaged velocity maps. In the simulation the motion of 15.185 m width and 0.4 m height with homogenous inflow conditions at 1. When collisions were assumed to be fully elastic and perfectly smooth the agreement was much worse. The results showed good agreement between experiment and simulation when the measured values for the collision parameters were used.25 mm to 1.Experimental Validation using PEPT ___________________________________________________________________________________ Chapter 8. In the model the Newtonian equations of motion are solved for each solid particle while taking into account the particle particle and particlewall collisions. In the PEPT experiment the motion of a single tracer particle in the bed was tracked during the period of one hour.5 umf was chosen as a test case.
Chapter 8 ___________________________________________________________________________________ 214 .
It is not the objective to present a complete review of available experimental techniques in gasfluidised beds here. (1998) whereas Mikami et al. (1996).Experimental Validation using PEPT ___________________________________________________________________________________ 1. With increasing computer power Granular Dynamics simulations have become a very useful and versatile research tool to study the dynamics of dense gasparticle flows. The mutual interactions between particles and the interaction between particles and walls are taken into account directly. In Chapter 5 and Chapter 6 of this thesis experimental validation was performed by direct observation of transient phenomena that occurred in a transparent system by using a video camera. Xu and Yu (1997) developed a hybrid technique combining elements of both softsphere and hardsphere techniques. Introduction Experimental validation of flow models for dense gasfluidised beds. that form the basis for (amongst others) reactor models. In these models the Newtonian equations of motion are solved for each individual granular particle in the system. Hoomans et al. This was partly due to the fact that proper experimental techniques were not widely available. (1993) who used a soft sphere model to describe the interaction between the particles.fluidised beds previously reported in the literature experimental validation has received little attention. The trend in both experiments and models is that systems can be studied in more and more detail with continuously increasing accuracy. In the latter case it is required that the segregating species 215 . (1998) extended the (2D) model in order to include cohesive forces between the particles. In the studies on discrete particle simulation of gas. This is a very basic technique that nonetheless can provide useful information on for instance the formation of a single bubble at an orifice or the dynamics of segregation. A comprehensive review on measuring techniques in fluidised beds was presented by Werther (1999) whereas an extensive overview of existing techniques for the measurement of solids concentration and velocity was presented by Nieuwland et al. (1996) presented a hardsphere approach where collisions are assumed to be binary and instantaneous. is of cr ucial importance to arrive at models that can predict the performance of fluidised beds with confidence. A three dimensional version of this model was later presented by Kawaguchi et al. The discrete particle approach was pioneered by Tsuji et al.
Chapter 8 ___________________________________________________________________________________
can be visually identified. However, the direct observation technique is limited to systems that allow for direct observation (i.e. transparent quasi twodimensional setups) in contrast to for example Xray techniques (Yates et al., 1994). Furthermore the information that is obtained using direct observation covers macroscopic behaviour: Important quantities such as local solids volume fraction, local solids velocities etc. cannot be obtained using this technique.
Particle tracking techniques are noninvasive measurement techniques where the motion of a single tracer particle is tracked for a period of time. Such techniques are ideally suited to validate a discrete particle model since the motion of a single particle in the bed can be tracked and therefore it allows for relatively direct comparison between measured data and simulation results. Lin et al. (1985) presented a technique where a radioactive tracer particle was used that was made of scandium46 (Sc46). A total of 12 scintillation detectors were positioned around the fluid ised bed to enable determination of the location of the particle. More recently Larachi et al. (1996) applied the tracking technique using Sc46 tracers to gas liquidsolid flows and Mostoufi and Chaouki (1999) applied the technique to liquid fluidised beds.
In this work a cooperation between the University of Twente and the University of Birmingham was initiated in order to use the Positron Emission Particle Tracking (PEPT) technique developed at Birmingham to validate the granular dynamics model of a gas fluidised bed developed at Twente. PEPT has been developed at Birmingham since 1987 and was successfully applied to a large number of systems for solids processing including mixers and gas fluidised beds (Seville et al. 1995). A comprehensive introduction to PEPT can be found in Stein et al. (1997). Recently a PEPT system was also developed by Stellema et al. (1998). PEPT differs from other tracking techniques in the sense that it uses positron emitting radioisotopes which have the unique feature that their decay leads to simultaneous emission of a pair of backtoback γ rays. In a following section a more elaborate discussion of the PEPT technique will be presented.
216
Experimental Validation using PEPT ___________________________________________________________________________________
2. Model
Since a detailed description of the model is presented in Chapter 2 and Chapter 3, the key features will be summarised briefly here.
2.1
Granular dynamics
The collision model as originally developed by Wang and Mason (1992) is used to describe a binary, instantaneous, inelastic collision with friction. The key parameters of the model are the coefficient of restitution (0 ≤ e ≤ 1) and the coefficient of friction (µ ≥ 0). Foerster et al. (1994) have shown that also the coefficient of tangential restitution (0 ≤ β 0 ≤ 1) should be used in order to describe the collision dynamics more accurately. These three collision parameters are all included in the model.
In the hardsphere approach a sequence of binary collisions is processed. This implies that a collision list is compiled in which for each particle a collision partner and a corresponding collision time is stored. A constant time step is used to take the external forces into account and within this time step the prevailing collisions are processed sequentially. In order to reduce the required CPU time neighbour lists are used. For each particle a list of neighbouring particles is stored and only for the particles in this list a check for possible collisions is performed.
2.2
External forces
The incorporation of external forces differs somewhat from the approach followed by Hoomans et al. (1996). In this work the external forces are used in accordance with those implemented in the two fluid model presented by Kuipers et al. (1992) where, of course, the forces now act on a single particle:
m
p
dv Vβ = mg+ (u − v dt (1 − ε )
p p p
p
) − V ∇p
p
,
(8.1)
217
Chapter 8 ___________________________________________________________________________________
where m p represents the mass of a particle, v p its velocity, u the local gas velocity and Vp the volume of a particle. In equation (8.1) the first term is due to gravity and the third term is the force due to the pressure gradient. The second term is due to the drag force where β represents an interphase momentum exchange coefficient as it usually appears in twofluid models. For low void fractions (ε < 0.80) β is obtained from the wellknown Ergun equation:
β = 150
(1 − ε )2
ε
µg D
2 p
+ 1 .75(1 − ε )
ρg Dp
u − vp ,
(8.2)
where Dp represents the particle diameter, µg the viscosity of the gas and ρ g the density of the gas. For high void fractions (ε ≥ 0.80) the following expression for the interphase momentum transfer coefficient has been used which is basically the correlation presented by Wen and Yu (1966) who extended the work of Richardson and Zaki (1954): 3 ε (1 − ε ) Cd ρ g u − v p ε − 2 .65 . 4 Dp
β=
(8.3)
The drag coefficient Cd is a function of the particle Reynolds number and given by:
24 0.687 Re (1 + 0.15Re p ) Cd = p 0.44
Re p < 1000
,
(8.4)
Re p ≥ 1000
where the particle Reynolds number (Re p ) in this case is defined as follows:
ε ρg u − v p D p µg
Re p =
.
(8.5)
For the integration of equation (8.1) a simple explicit first order scheme was used to update the velocities and positions of the particles.
218
Experimental Validation using PEPT ___________________________________________________________________________________
2.3
Gas phase hydrodynamics
The calculation of the gas phase hydrodynamics mainly follows the lines presented by Kuipers et al. (1992). It is based on the numerical solution of the following set of partial differential equations that can be seen as a generalised form of the NavierStokes equations for a gas inte racting with a solid phase as originally derived by Anderson and Jackson (1967).
Continuity equation gas phase:
∂ ερ g
( ) + (∇ ⋅ ερ u) = 0 . ∂t
g
(8.6)
Momentum equation gas phase:
∂ ερg u ∂t
(
) + (∇ ⋅ ερ uu) = − ε∇p − S − ( ∇ ⋅ ετ ) + ερ g .
g p g g
(8.7)
In this work transient, twodimensional, isothermal (T = 293 K) flow of air at atmospheric conditions is considered. The constitutive equations can be found in Chapter 3. There is one important modification with respect to the model presented by Hoomans et al. (1996) and that deals with the way in which the twoway coupling between the gasphase and the dispersed particles is established. In the present model the reaction force to the drag force exerted on a particle per unit of volume is fed back to the gas phase through the source term S p which has the dimension of force per unit
3 of volume N/m .
3. Positron Emission Particle Tracking
Positron Emission Particle Tracking (PEPT) is a technique that allows noninvasive observation of the motion of a single radioactive tracer particle. The PEPT technique is schematically represented in Figure 8.1.
219
Chapter 8 ___________________________________________________________________________________
detector
detector
a) detection
b) reconstruction
c) particle location
Figure 8.1.
The principle of the PEPT technique
In the experiment reported here a glass particle taken from the sample of particles used in the fluidisation experiment was activated by direct irradiation in a cyclotron beam. 220
The positron then annihilates with an electron to produce a pair of back to back γ rays. 4. when the particle is moving the set of annihilation vectors should be large enough to locate the particle accurately but not so large that it has moved significantly during the time period over which the set was measured. the antiparticle of the electron. Comparison between PEPT data and simulation The PEPT experiment as performed in this work renders the trajectory of a single particle during one hour. This time scale cannot be reached by means of simulation on modern day computers. The decay of the 18 F isotope features the conversion of a proton to a neutron with the emission of a positron. Instead the motion of 15.000 simulated particles cover the whole of the fluidised system.3 by 0. In fact considerably more data is generated in the simulation since 15. The instantaneous particle velocity is inferred from the difference between successive locations. When the tracer particle does not move the more annihilation vectors are used the more accurate the particle position can be determined. the simulation data does not suffer from poorly sampled regions in contrast with the experimental data. By using a reconstruction algorithm the position of the particle can be obtained as the intersection point of successive annihilation vectors. And since the 15. The algorithm employs an iterative scheme to discard corrupt annihilation vectors that can be caused by γ ray scattering or random coincidences.Experimental Validation using PEPT ___________________________________________________________________________________ The glass particle was irradiated with the 3He beam from a cyclotron to produce the positron emitter 18 F from reactions involving the oxygen on the glass. The γ rays are detected by the positron cameras which consists of two positionsensitive γ ray detectors that have an active area of 0.000 times 45 seconds is a far larger number than 1 times 3600 seconds.6 m. The technique was not optimised for the use of a quasi twodimensional system and for the post processing the standard software was used. 221 . However.000 particles is tracked during a shorter period of time (45 s).
Chapter 8 ___________________________________________________________________________________
It is assumed that it is justified to compare the simulation results with the time averaged experimental data by first time averaging the simulation data for each of the 15,000 particles and subsequently taking an ensemble average over all the 15,000 particles. By doing this there is a risk that rare events, which occur over a time scale clo se to or greater than the duration of a simulation, are poorly sampled. On the contrary, the chance that such a rare event is experienced by the tracer particle in the experiment is rather low as well.
The actual comparison is made on the basis of occupancy plots, velocity maps and speed histograms which are standard outputs of the PEPT software (Tin et al., 1997) that was supplied by the University of Birmingham. To obtain an occupancy plot from the PEPT data the system is first divided into cells (10 mm width/height). In the occupancy plot the fraction of the total time that the tracer particle has spend in that particular cell is displayed using a colour code that is explained in the legend accompanying the plot. For the simulations a similar procedure was followed after all the individual particle trajectories were added together. In order to obtain the velocity maps the system was again divided into cells. The velocity vectors in the figure are averaged velocity for that particular cell. How the averages were obtained is explained above. Finally the PEPT data was compared with the results of the simulation using speed histograms. For the PEPT experiment the speed of the tracer particle was calculated on the basis of the 3 velocity components (x,y,z) at each instant. The histogram shows the speed distribution based on the data set covering the entire duration of the experiment of one hour. The speed histograms obtained from the simulation results are based on the two velocity components (x,y) taken into account in the simulations.
5. Results
As a test system for the experimental validation a gasfluidised bed (0.185 m width, 20 mm depth) with homogeneous inflow conditions (ug = 1.5 umf, umf = 0.9 m/s) was chosen. The parameter settings for the simulation are summarised in table 8.1. The bed was filled with the glass particles in such a way the static bed height was about 0.17 m.
222
Experimental Validation using PEPT ___________________________________________________________________________________
In the simulation a total number of 15,000 particles was used to match the bed height encountered in the experiment. The bed used in the experiment was assembled using glass as the construction material. It was fitted with a porous plate distributor with a sufficiently high pressure drop to ensure homogeneous inflow conditions. In preliminary experiments a malfunctioning distributor plate was used which caused a strongly nonsymmetrical solids flow pattern in the bed indicating a non uniform gas distribution. The actual PEPT experiment used for the experimental validation was run for one hour in order to collect sufficient data to calculate statistically reliable averages.
Table 8.1.
Parameter settings for the PEPT simulation
Particles: Shape density, ρp particle diameter, Dp e = ew µ = µw β0 = β 0,w Np spherical 2418 kg/m3 1.251.50 mm 0.97 0.10 0.33 15,000
Bed: Width Height Number xcells, NX Number ycells, NY cell width, DX cell height, DY time step, DT 0.185 m. 0.40 m 37 80 5 mm 5 mm 10 –4 s
The particleparticle collision parameters presented in table 8.1 were independently measured by Gorham and Kharaz (1999) using the facility at the Open University at Milton Keynes described in Chapter 2. The particlewall collision parameters were assumed to be equal to the particleparticle collision parameters. This is justified since earlier simulations (Chapter 4) showed that the influence of the particle wall collision parameters were negligible compared to the influence of the particleparticle collision parameters. A lognormal particle size distribution about an average diameter of 1.375 mm was taken into account in the simulation. Particle diameters lower than 1.25 mm and greater than 1.50 mm were rejected in order to mimic the effects of sieving and hence match the particle size distribution encountered in the experiment.
223
Chapter 8 ___________________________________________________________________________________
Figure 8.2.
Snapshots of particle configurations of the simulation.
An initial simulation was performed first to ensure that the actual simulation would not suffer from any startup effects. During this initial simulation the system was operated at a homogeneous gas inflow of 1.5 um f (um f = 0.9 m/s) with the parameter settings
224
Experimental Validation using PEPT ___________________________________________________________________________________
presented in table 8.1. After this initial simulation the actual simulation (duration 45 s) was started and the required data (positions and velocities) was stored at time intervals correspond ing to those used in the experiment. Snapshots of the simulation are presented in Figure 8.2. In this figure it can be observed that the bed is bubbling quite vigorously and this behaviour could also be visually observed during the PEPT experiment. In Figure 8.3 an example of the trajectory of one single (randomly chosen) particle in the simulation is presented. This trajectory covers the whole duration of the simulation of 45 s.
Figure 8.3.
Example of a trajectory of a single particle during the simulation.
Although the figure above gives some insight into the motion of a particle in a fluidised bed it does not provide a solid basis for a comparison. Therefore averaging techniques were applied as discussed in the previous section which yield data that permit a comparison between the PEPT data and the simulation results. In Figure 8.4 the velocity map obtained from the PEPT data is presented together with the velocity maps obtained from two simulations. In the centre the velocity map of the simulation using the measured values for the collision parameters (table 8.1) is presented and on the right the velocity map of a simulation assuming fully elastic, perfectly smooth collisions (e = 1, µ = 0, also referred to as ideal collisions) is presented. A reference vector indicating the magnitude of the velocity is included in all of the three velocity maps. In velocity map of the PEPT data a circulation pattern 225
Chapter 8 ___________________________________________________________________________________
can be observed where particles rise in the centre of the bed and descend near the walls. Two circulation cells can be distinguished which together form a rather symmetric picture indicating that the gas inflow was indeed homogeneous.
Figure 8.4.
Velocity map obtained from the PEPT data (left) compared with the velocity maps of the simulation using the measured collision parameters (centre) and the simulation assuming fully elastic, perfectly smooth collision (right)
From the simulation with the measured collision parameters a very similar velocity map was obtained. Although not perfectly symmetric two circulation cells can be distinguished which is in good agreement with the PEPT data. The velocity map of the simulation assuming fully elastic, perfectly smooth collisions shows a somewhat different behaviour. The velocity vectors are smaller indicating lower speeds and also two additional circulation cells on top of the two main cells can be observed. These additional circulation cells rotate in the opposite direction and were not present in the PEPT experiment.
In Figure 8.5 the occupancy plots obtained from the PEPT data and the two simulations are presented together. In this figure it can be observed that in the PEPT data the occupancy was higher near the walls which was also the case in the simulation with the measured collision parameters. This indicates that the particles spend relatively more time near the walls. In the simulation assuming fully elastic, perfectly
226
the simulation using the measured collision parameters (left) and the simulation assuming fully elastic. In other words the residence time of the particles is evenly distributed throughout the whole system. This is obviously due to the fact that in the experiment one single particle was tracked for one hour. Figure 8.5.Experimental Validation using PEPT ___________________________________________________________________________________ smooth collisions the occupancy is almost the same at each position in the system. This is due to the absence of bubbles in the latter case which causes a very homogeneous type of fluidisation. This does not guarantee that the whole system is covered and therefore it can be understood that 227 . perfectly smooth collisions (right) It should be noted that the occupancy plot obtained from the PEPT data is not as smooth as the ones obtained from the simulations. Occupancy plots obtained from the PEPT data (above).
Speed histogram obtained from the PEPT experiment (top) compared with the speed histograms obtained from the simulation using the measured collision parameters (left) and the simulation assuming perfectly smooth. Figure 8.6.Chapter 8 ___________________________________________________________________________________ some regions of the bed suffer from poor statistics. For the simulation results all the particles in the system were taken into account which renders a far smoother picture since the whole of the bed was automatically sampled. fully elastic collisions (right) 228 .
This once again emphasises the profound influence that these collision parameters have on the dynamics of a gas. In the latter simulation the average speed (about 0. Where the velocity map and the occupancy plot provide a basis for a more qualitative comparison between simulation and experiment. It is important to note that in both the experiment and the simulation particle speeds above 0. the agreement is encouraging. the speed histogram offers the possibility for a quantitative comparison. perfectly smooth collisions is much worse in this respect.4 m/s are observed.Experimental Validation using PEPT ___________________________________________________________________________________ In Figure 8.6 the speed histogram obtained from the PEPT experiment is presented together with the speed histograms obtained from the two simulations.25 m/s in the simulation and 0.4 m/s were observed. The reason for this is the absence of bubbles in the simulation with ideal collision parameters since the particles attain their maximum speeds when they are accelerated into the wake of a bubble.1 m/s) is actually closer to the average speed observed in the experiment but more importantly however. Although the simulation results show a somewhat higher average speed than the PEPT data (about 0. Although the speeds observed in the simulation are somewhat higher than the ones observed in the experiment for reasons discussed above. The agreement between the results of the PEPT experiment and the simulation assuming fully elastic. In Figure 8. 229 . The higher average speed is most likely due to the twodimensional nature of the simulation.15 m/s in the experiment) the shape of the distribution is rather similar. The absence of the front and back walls in the simulation implies that the particles in the simulation do experience a lower amount of wall friction that effectively results in a higher average speed.fluidised bed. Since these higher speeds are closely related to the bubbling behaviour in the bed it is important that good agreement between simulation and experiment is achieved on this matter.6 it can be observed that the results of the simulation using the measured collision parameters compare rather well with results of the PEPT experiment. the distribution of speeds in the histogram is far more narrow: no speeds above 0.
When collisions were assumed to be fully elastic and perfectly smooth the agreement was worse. The particle speeds observed in the simulation were somewhat higher than those observed in the experiment which is most likely due to the absence of the front and back wall in the simulation. From this first comparison between granular dynamics simulations and a PEPT experiment it can be concluded that PEPT is a powerful tool for the experimental validation of these simulations. The results showed good agreement between experiment and simulation when measured values for the collision parameters were used. The simulation data was time averaged over 45 s for each particle and subsequently ensemble averaged over all the particles in the simulation. In the PEPT experiment the motion of a single tracer particle in the bed was tracked during the period of one hour.4 m/s were observed whereas in the experiment as well as in the simulation with the measured collision parameters speeds up to 0. Glass particles (ρ p = 2435 kg/m3 ) with diameters ranging from 1. perfectly smooth collisions is not valid for fluidised bed simulations. In the simulation the motion of 15.4 m height with homogenous inflow conditions at 1. the PEPT software should be able reproduce these results. This demonstrates the profound influence of the collision parameters on the bed hydrodynamics since these higher speeds are closely related to the presence of bubbles in the bed.Chapter 8 ___________________________________________________________________________________ 6. A quasi twodimensional bed of 0.25 mm to 1.185 m width and 0. From a direct comparison with an experiment it was therefore shown that the assumption of fully elastic.8 m/s were found.000 particles was tracked during a period of 45 s. Since the simulation yields particle positions and corresponding velocities that are exact.5 u mf was chosen as a test case. No speeds higher than 0. 230 . Conclusions Granular dynamics simulations of gasfluidised beds with homogeneous inflow conditions were experimentally validated using the Positron Emission Particle Tracking facility at the University of Birmingham. In future work it should not only be attempted to perform the validation on a more detailed level but also the simulation data can be used to verify the PEPT software.5 mm were used as the bed material.
K time. [] number of computational cells in x direction. kg number of particles.Experimental Validation using PEPT ___________________________________________________________________________________ Notation Cd e Dp DT DX DY g mp Np NX NY p r Sp T t u vp Vp drag coefficient. m momentum source term N/m3 temperature. [] particle diameter. m gravitational acceleration. Pa position vector. [] number of computational cells in y direction. [] void fraction. [] coefficient of friction. kg/m3 Subscripts g gas phase 231 . kg/ms2 density. kg/ms gas phase stress tensor. m/s particle velocity vector. s gas velocity vector. [] pressure. m vertical computational cell dimension. s horizontal computational cell dimension. m/s 3 particle volume. [] gas viscosity. m Greek symbols β β0 ε µ µg τ ρ 3 volumetric interphase momentum transfer coefficient. m/s2 particle mass. [] coefficient of restitution. kg/(m s) coefficient of tangential restitution. m time step.
Flow structure of the solids in a 3D gasliquidsolid fluidized bed.and threedimensional models). (1985)..Chapter 8 ___________________________________________________________________________________ mf p w minimum fluidisation particle wall References Anderson. Louge. M. and Kharaz.P. Gorham. Chang.M. M. UK. R.P. 42 .P..T. 232 . and Tsuji. Chem. 6 . Kuipers.. Ind.. Discrete particle simulation of bubble and slug formation in a twodimensional gasfluidised bed: a hardsphere approach. van Duin K. D.. A numerical model of gasfluidized beds. van Beckum. and Guy. Chem. Hoomans. Milton Keynes. and Allia. Chaouki. 1913 Larachi. and Jackson. Lin.H.... A.. Engng Sci. Tanaka.A.. A novel radioactive particle tracking facility for measurement of solids motion in gas fluidized beds.P.. 129. J.J. 527.. 96. (1992). 31. Briels. Fluids. T. Measurements of the collision properties of small spheres. Engng Sci.M. and van Swaaij. Impact Research Group Report IRG 13. Foerster S.. J. Y. and van Swaaij.. Kawaguchi. Fundam.. (1996).. A fluid mechanical description of fluidized beds (equations of motion). 1108. B. J.. W. 6. (1996). Phys. B.H.H.A. Eng. T. (1994).. (1998).. Results of particle impact tests.. AIChE J. Chen. T. J.M.. W. F.Y.F. (1999).J. W. Chem.. 47. 99. 51 . Cassanello. Numerical simulation of twodimensional fluidized beds using the discrete element method (comparison between the two.. C. AIChE J.S. Kuipers. and Chao. M. K. 465. 2439. (1967). F.M.B.M. The Open University.B..A. Powder Technol..
. Powder Technol.R. J. Mudde.. Trans. (1995).. M cNeil. H. UK. 32 . Larachi. D. Tan.. Seville. R. in Noninvasive monitoring of multiphase flows. (1998). Martin.. 53 . and Horio. Vlek.. Numerical simulation of cohesive powder behavior in a fluidized bed. and van Swaaij. P. J.P. Kawaguchi.A. T. Parker.. Particle velocities in gasfluidised beds. Broadbent.P. (1998). J. Res.J. T. (1996).. 87..M.. R.. And Meth. Chaouki. Eng... Seville.. Stellema. In Phys. M.M. J. Discrete particle simulation of two dimensional fluidized bed. Parker. Fluidization VIII. T. Simons. Chem.S. 851.. Stein. Engng Sci. 1927. Preprints Tours. Positron emission particle tracking: particle velocities in gas fluidised beds. Measurements of solids concentration and axial solids velocity in gassolid twophase flows. (1997).. M. 35. 334. S.M.W. the Netherlands.P. C. mixers and other applications. and Zaki. and Tanaka. T. Development of an improved positron emission particle tracking system. J. (1997). Richardson. J. D. Tsuji. T.A. and Dududkovic. C. J. J.W.K. 233 . 54. F. Engineering Foundation. Engng Sci. (eds). (1993). Nieuwland. Instr. 77. D. (1954). 79.J. and Parker....J.J. PEPT data presentation software.. M. and Dee.. de Goeij. Kamiya.M. Meijer. W.K.J.. 127. Y.F.J. and Beynon. Amsterdam. Martin.N.. A. Prediction of effective drag coefficient in fluidized beds. Mostoufi.. W. Kuipers. Chem. Inst.. N.. (1999). M. P. 319.D. Elsevier Science. Sedimentation and fluidization: part I. and Chaouki. Chem.P. manual. J. and van den Bleek. Birmingham University. T. Powder Technol..F. C.J.. 404 .R. Nucl..Experimental Validation using PEPT ___________________________________________________________________________________ Mikami.
H. Experimental observations of voidage distribution around bubbles in a fluidized bed. (1999). (1992). Chem. Wen. Xu. Ser. Chem.G. Chem. (1994). 234 . Engng Sci. Yates. Y. D. 1885. 49 . 52.A. 2785..Chapter 8 ___________________________________________________________________________________ Wang. J. J. A. 100. and Mason. Mech. Y. Eng.H.T. 62 (62). (1997). Mechanics of fluidization... Werther. 635. B. Cheesman. 102 . Powder Technol. Appl.. and Yu. Y.. Numerical simulation of the gassolid flow in a fluidized bed by combining discrete particle method with computational fluid dynamics. Measurement techniques in fluidized beds.J. Engng Sci. Twodimensional rigidbody collisions with friction.Y.B. Prog. and Yu. Symp... J. (1966).. M. 15.. C. and Sergeev. 59.
B.M. Vol. Chem. Kuipers. (1998). J. van Swaaij. Briels and W.A. 3.B. B. Kuipers.H. 15. 4.M.P. Yu. Hoomans.P. Hoomans and W. Third International Conference on Multiphase Flows 1998. 318. in Fluidization IX. Brighton. van Swaaij. B. No. B.P.A..A. J. Discrete particle simulation of bubble and slug formation in twodimensional gasfluidised beds: a hardsphere approach.M. Engng Sci.M. 2. AIChE Symp.A.fluidized beds and their role for design and operation of fluidized bed chemical reactors.P.S.J.B. (1998). 6.P. van Swaaij. van Swaaij. B. Fan and T. France.M. W. J. (1998). The influence of particle properties on pressure signals in dense gasfluidised beds: a computer simulation study.A. 94. Knowlton (eds).M. World Congress on Particle Technology 3.M. J. L. 235 . Ser. Fan and T. (1996).P. van Swaaij.M. B. Hoomans. Kuipers.P. van Swaaij. 7. J. 5.B. Hoomans.J. Knowlton (eds). Kuipers and W.Publications ________________________________________________________________________ Publications 1.A. (1998). (1998). Lyon. Kuipers and W.B.M. June 812.M.P. Kuipers and W. Engng Sci. Xu and A. in Fluidization IX. 2645. Hoomans. Discrete particle simulation of segregation phenomena in dense gas.B. Hydrodynamic models of gas. 54. 15. The influence of a particle size distribution on the granular dynamics of dense gasfluidized beds: a computer simulation study.P.M.P. L. J.M. 99. van Swaaij.B. UK. 51.B. B..M.M.P.S.M. J. Comments on the paper ‘Numerical simulation of the gassolid flow in a fluidized bed by combining discrete particle method with computational fluid dynamics’ by B. Briels and W. Chem. July 79.M.B.P. Hoomans. Granular Dynamics Simulation of Cluster Formation in Dense Riser Flow.fluidized beds. Hoomans. Kuipers and W.A.P.P. 485. (1998). W.
).P. Kuipers and W. B. Kuipers and W. Werther (Ed. (2000). Granular dynamics simulation of segregation phenomena in dense gasfluidised beds. (1999).M. van Swaaij.P. B.A. Circulating Fluidized Bed Technology VI.A. Powder Technology (in press).B.M. Hoomans. J. Hoomans.P. J. J. .Publications ________________________________________________________________________ 8. 236 . van Swaaij. 9.P.M. Discrete particle simulation of cluster formation in dense riser flow.M.B.
afstudeerverslag Universiteit Twente. Peter Schinkelshoek (riserstroming). Patrick Huttenhuis en Olaf Veehof. Ook zonder Wim Briels was dit promotieproject niet van de grond gekomen. Wim van Swaaij volgde het onderzoek weliswaar van wat grotere afstand maar wist op beslissende momenten een zet in de goede richting te geven. Modellering van de hydrodynamica van meerfasenstromingen is een onderzoekslijn waarmee Hans Kuipers inmiddels wereldwijd een grote naam heeft opgebouwd. De experimenten beschreven in hoofdstuk 5 waren niet mogelijk geweest zonder de hulp van Jelle Nieuwland. Een vijftal afstudeerders heeft een zeer gewaardeerde bijdrage geleverd aan de diverse onderwerpen die in dit proefschrift aan bod komen. Mathijs Goldschmidt. 1996.Dankwoord (Ackowledgements) En zo komt er een einde aan een promotie met ballen* . Wim. Marco Stam (zachte ballen). Wim. Zijn vermogen om formules te doorgronden en die vervolgens te manipuleren is ook de kwaliteit van dit proefschrift ten goede gekomen. Hans. Reneke van SoestSeij (3D harde ballen) en Jan Gerard Schellekens (segregatie): allemaal ontzettend bedankt ! Veel steun heb ik gehad aan mijn collega’s Erik Delnoij. het idee voor het plaatje op de kaft van dit proefschrift is aan zijn brein ontsproten). ik zal altijd onthouden dat je me een sleepkabel aanreikte toen ik met pech langs de weg stond. 237 . ik denk met veel plezier terug aan de tijd die ik in jouw groep heb doorgebracht. Een promotie die zeker niet het werk is van slechts één persoon. De mensen die een bijdrage hebben geleverd aan de totstandkoming van dit proefschrift wil ik op deze plek dan ook graag bedanken. Ik wil Bart Lammers bedanken voor het ontwikkelingswerk aan de visualisatiesoftware dat hij heeft uitgevoerd tijdens * vrij naar Marco Stam. Ik ben je dankbaar voor het grote vertrouwen dat je in me had. Michiel Gunsing en Jie Li van de multiphase flow crew van Hans Kuipers waar we dankbaar van elkaars kennis en vaardigheden gebruik konden maken. Robert Heijnen (segregatie. het was me een w aar genoegen om deel uit te maken van jouw team. Allereerst gaat mijn dank uit naar mijn drie promotoren.
Met name de cursussen in het kader van de Powergroup en de directe interactie met experts op het gebied van supercomputing heb ik erg op prijs gesteld. Dennis Allen. and Aidan McCormack for their valuable contribution to this BirminghamTwente project. Peter Michielse en Bart Mellenberg: bedankt ! Secretariële ondersteuning werd altijd vriendelijk en correct verzord door (in chronologische volgorde): Greet van der VoortKamminga. Frank van de Scheur en AtzeJan van der Goot bedanken voor de mogelijkheid die ze Mathijs en mij geboden hebben om te participeren in een groots opgezet internationaal Unilever onderzoeksprogamma. De videoanimaties van mijn simulaties zijn inmiddels in alle werelddelen te zien geweest. Professor Jonathan Seville and Dr. David Parker of the University of Birmingham are gratefully acknowledged for their support of the PEPT validation project reported in Chapter 8. Ik wil Joop Olieman. Yvonne Bruggert. Deze metingen werden financiëel ondersteund door het Unilever Research Laboratorium Vlaardingen. Dit heb ik als zeer leerzaam ervaren. The collision parameters required for the validation simulations in Chapter 6 and Chapter 8 were experimentally obtained by Dr. Gery Stratingh. David Gorham and Ahmad Kharaz of the Open University at Milton Keynes. I wish Jie Li and Pranay Darda all the best with their continuation of the discrete particle work. Ronald van Pelt. Hierbij werd hij technisch ondersteund door Gerrit Schorfhaar. I want to thank Matthias Stein. Nicole Haitjema. PaulGuillaume Schmitt voerde in het kader van zijn stage een deel van de metingen met deze opstelling uit (merci beaucoup !) en dank zij Siebren Mellema konden de videobeelden digitaal verwerkt worden. Met veel plezier denk ik terug aan de contacten met Silicon Graphics (tegenwoordig SGI). Mathijs heeft een cruciale rol gespeeld bij de totstandkoming en het bedrijven van de opstelling die gebruikt is voor de experimenten beschreven in hoofdstuk 6.Dankwoord ________________________________________________________________________ zijn afstudeeropdracht bij Erik. Mohamad Amran Mohd Salleh. Ria Stegehuisde Vegte en 238 . Ruud van der Pas. Bartie Bruggink. Annemiek Vos.
Ties Bos. Reinier. het zaalvoetballen met Proceskunde en vooral ook het karten om de Briels trofee en De Dhont Cup met Computational Chemistry (nog altijd ongeslagen binnen CT !). een luisterend oor en een kritische noot op zijn tijd. kroegtochten en etentjes. Tinus. Peter en Ine. Niels Kruyt. Tinus heeft in de beginperiode een zeer belangrijke rol gespeeld door mij te attenderen op de simulatietechnieken uit de Moleculaire Dynamica. het was me een groot genoegen ! Tot slot wil ik mijn ouders. De basis voor dit proefschrift is in die periode gelegd. Roelof. Harald. Nico. Frank. Margie en José. Met groot plezier denk ik terug aan de diverse borrels. Allemaal ontzettend bedankt ! Bob. en mijn broers (paranimf) Tom & Rick bedanken die samen een solide basis vormen waar ik altijd op kon terugvallen. Verder is hij de afgelopen jaren altijd goed geweest voor veel gezelligheid. Ook buiten kantooruren om heb ik me altijd uitstekend vermaakt met mijn collega AIO’s in Enschede. Hein. Martin van der Hoef en Peter Roos hier nogmaals bedanken voor de bijdrage die ze geleverd hebben in de diverse afstudeercommissies binnen het promotieproject. Wouter.Dankwoord ________________________________________________________________________ Brigitte Sanderink. bedankt ! Een speciaal woord van dank is op zijn plaats voor mijn goede vriend en paranimf Martijn Oversteegen. Ik wil Ben Betlem. Rik Akse zorgde voor een keurige afhandeling van de financiële zaken. 239 . Johan. Wim.
Dankwoord ________________________________________________________________________ 240 .
In 1989 behaalde bij het VWO diploma aan het Thijcollege te Oldenzaal. In januari 1995 studeerde hij af bij de vakgroep Proceskunde op het onderwerp discrete deeltjessimulatie van gasgefluïdiseerde bedden. De resultaten van dit onderzoek vormen de basis voor dit proefschrift. In juni 1995 trad hij in dienst als promovendus (AIO) bij de vakgroepen Chemische Fysica (later Computational Chemistry) en Proceskunde waar hij onderzoek verrichte op het gebied van granulaire dynamica van gasvast tweefasenstromingen. In september 1989 werd begonnen met de studie Chemische Technologie aan de Universiteit Twente in Enschede. In oktober 1999 trad hij in dienst als CFD engineer bij DSM Research in Geleen. In het kader van deze opleiding liep hij van september 1993 tot maart 1994 stage bij het Joint Research Centre (JRC) in Ispra. 241 .Levensloop Bob Hoomans werd geboren op 2 augustus 1971 in Oldenzaal waar hij ook de lagere school bezocht. Italië.
242 .
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue listening from where you left off, or restart the preview.