CHINA PARTICUOLOGY Vol. 3, Nos.

1-2, 69-77, 2005
COMPUTATIONAL FLUID DYNAMICS FOR DENSE GAS-SOLID
FLUIDIZED BEDS: A MULTI-SCALE MODELING STRATEGY
M. A. van der Hoef, M. van Sint Annaland and J. A. M. Kuipers*
Department of Science & Technology, University of Twente, 7500 AE Enschede, The Netherlands
*Author to whom correspondence should be addressed. E-mail: j.a.m.kuipers@utwente.nl
Abstract Dense gas-particle flows are encountered in a variety of industrially important processes for large scale
production of fuels, fertilizers and base chemicals. The scale-up of these processes is often problematic and is related to
the intrinsic complexities of these flows which are unfortunately not yet fully understood despite significant efforts made in
both academic and industrial research laboratories. In dense gas-particle flows both (effective) fluid-particle and (dissi-
pative) particle-particle interactions need to be accounted for because these phenomena to a large extent govern the
prevailing flow phenomena, i.e. the formation and evolution of heterogeneous structures. These structures have signifi-
cant impact on the quality of the gas-solid contact and as a direct consequence thereof strongly affect the performance of
the process. Due to the inherent complexity of dense gas-particles flows, we have adopted a multi-scale modeling ap-
proach in which both fluid-particle and particle-particle interactions can be properly accounted for. The idea is essentially
that fundamental models, taking into account the relevant details of fluid-particle (lattice Boltzmann model) and parti-
cle-particle (discrete particle model) interactions, are used to develop closure laws to feed continuum models which can
be used to compute the flow structures on a much larger (industrial) scale. Our multi-scale approach (see Fig. 1) involves
the lattice Boltzmann model, the discrete particle model, the continuum model based on the kinetic theory of granular flow,
and the discrete bubble model. In this paper we give an overview of the multi-scale modeling strategy, accompanied by
illustrative computational results for bubble formation. In addition, areas which need substantial further attention will be
highlighted.
Keywords dense gas-solid flow, gas-fluidized beds, multi-scale modelling
1. Introduction
Dense gas-particle flows are frequently encountered in a
variety of industrially important gas-solid contactors, of
which the gas-fluidized bed can be mentioned as a very
important example. Due to their favorable mass and heat
transfer characteristics, gas-fluidized beds are often ap-
plied in the chemical, petrochemical, metallurgical, envi-
ronmental and energy industries in large scale operations
involving i.e. coating, granulation, drying, and synthesis of
fuels and base chemicals (Kunii & Levenspiel, 1991). Lack
of understanding of the fundamentals of dense gas-particle
flows, and in particular of the effects of gas-particle drag
and particle-particle interactions (Kuipers et al., 1998;
Kuipers & van Swaaij, 1998), has led to severe difficulties
in the scale-up of these industrially important gas-solid
contactors (van Swaaij, 1990). To arrive at a better under-
standing of these complicated systems in which both
gas-particle and particle-particle interactions play a domi-
nant role, computer models have become an indispensa-
ble tool. However, the prime difficulty with modeling
gas-fluidized beds is the large separation of scales: the
largest flow structures can be of the order of meters; yet
these structures are found to be directly influenced by
details of the particle-particle collisions, which take place
on the scale of millimeters or less. Therefore, we have
adopted a multi-level modeling strategy (see Fig. 1), with
the prime goal to (i) obtain a fundamental insight in the
complex dynamic behavior of dense gas-particle fluidized
suspensions; that is, to gain an understanding based on
elementary physical principles such as drag, friction and
dissipation (ii) based on this insight, develop models with
predictive capabilities for dense gas-particle flows en-
countered in engineering scale equipment. To this end, we
consider gas-solid flows at four distinctive levels of mod-
eling.
At the most detailed level of description the gas flow field
is modeled at scales smaller than the size of the solid par-
ticles. The interaction of the gas phase with the solid phase
is incorporated by imposing “stick” boundary conditions at
the surface of the solid particles. This model thus allows us
to measure the effective momentum exchange between
the two phases, which can be used in the higher scale
models. In our model, the flow field between spherical
particles is solved by the lattice Boltzmann model (Succi,
2001; Ladd & Verberg, 2001) although in principle other
methods (such as standard computational fluid dynamics)
could be used as well. At the intermediate level of descrip-
tion the flow field is modeled at a scale larger than the size
of the particles, where a grid cell typically contains
O(10
2
)~O(10
3
) particles, which are assumed to be perfect
spheres (diameter d). This model consists of two parts: a
Lagrangian code for updating the positions and velocities
of the solid particles from Newton’s law, and a Eulerian
code for updating the local gas density and velocity from
the Navier-Stokes equation (Hoomans et al., 1996). The
advantage of this Discrete Particle Model (DPM) is that it
can account for particle-wall and particle-particle interac-
tions in a realistic manner, for system sizes of about O(10
6
)
particles, which is sufficiently large to allow for a direct
CHINA PARTICUOLOGY Vol. 3, Nos. 1-2, 2005

70

comparison with laboratory scale experiments. As a logical
consequence of this approach a closure law for the effec-
tive momentum exchange has to be specified, which can
be obtained from the aforementioned lattice Boltzmann
simulations. Note that in chemical engineering, to date
mainly empirical relations are used for the friction coeffi-
cient β (defined by (1) and (2)), such as the Ergun (1952)
correlation for porosities 0.8 ε < :
( ) ( )
2
2
1 1
150 1.75
d ε ε β
µ ε ε
− −
= + Re , (1)
and the Wen and Yu (1966) equation for porosities
0.8 ε > :
( )
2
2.65
d
3
1
4
d
C
β
ε ε
µ

= − Re ,
( )
0.687 3
d
3
24 1 0.15 / 10
0.44 10
C
¦
+ <
¦
=
´
> ¦
¹
Re Re Re
Re
,
(2)
where µ is the viscosity of the gas phase, Re is the particle
Reynolds number, and C
d
the drag coefficient, for which
the expression of Schiller and Nauman (1935) is used. At
an even larger scale a continuum description is employed
for the solid phase, i.e. the solid phase is not described by
individual particles, but by a local density and velocity field.
Hence, in this model both the gas-phase and the solid
phase are treated on an equal footing, and for both phases
an Eulerian code is used to describe the time evolution
(see Kuipers et al., 1992; Gidaspow, 1994, amongst oth-
ers). The information obtained in the two smaller-scale
models is then included in the continuum models via the
kinetic theory of granular flow. The advantage of this model
is that it can predict the flow behavior of gas-solid flows at
life-size scales, and these models are therefore widely
used in commercial fluid flow simulators of industrial scale
equipment. Finally, at the largest scale, the (larger) bub-
bles that are present in gas-solid fluidized beds are con-
sidered as discrete objects, similar to the solid particles in
the DPM model. This model is an adapted version of the
discrete bubble model for gas-liquid bubble columns. We
want to stress that this model, as outlined in section 5, has
been developed quite recently, and the results should be
considered as very preliminary. In this paper we will give an
overview of these four levels of modeling as they are em-
ployed in our research group. In the following sections we
will describe each of these models in more detail.
2. Lattice Boltzmann Model (LBM)
The lattice Boltzmann model (LBM) originates from the
lattice-gas cellular automata (LGCA) models (Frisch et al.,
Fig. 1 Multi-level modeling scheme for dense gas-fluidized beds.
Lattice Boltzmann
Model
Discrete Particle
Model
Continuum
Model
Discrete Bubble
Model
Larger geometry
Fluid-particle
interaction
Particle-particle
interaction
Particle-particle
interaction
Bubble behavior
Large scale motion
Industrial size
Larger scale phenomena
van der Hoef, van Sint Annaland & Kuipers: Computational Fluid Dynamics for Dense Gas-Solid Fluidized Beds

71

1986) for simple fluids. The LGCA model is basically a
discrete, simplified version of the molecular dynamics
model, which involves propagations and collisions of parti-
cles on a lattice. LGCA models have proved a simple and
efficient way to simulate a simple fluid at the microscopic
level, where it has been demonstrated both numerically
and theoretically that the resulting macroscopic flow fields
obey the Navier-Stokes equation. The lattice Boltzmann
model is the ensemble averaged version of the LGCA
model, so that it represents a propagation and collision of
the particle distributions instead of the actual particles as in
the LGCA models (McNamara & Zanetti, 1988). From a
macroscopic point of view, the LB model can be regarded
as a finite difference scheme that solves the Boltzmann
equation, the fundamental equation in the kinetic theory
which underlies the equations of hydrodynamics. In its
most simple form the finite difference scheme reads:
( )
eq
( , , ) ( , , ) ( , , ) ( , , )
i i i i
t
f t t t f t f t f t
τ
δ
+ δ + δ − = − r r r r v c c c c , (3)
where f is the single particle distribution function, which is
equivalent to the fluid density in the 6 dimensional veloc-
ity-coordinate space, and f
eq
represents the equilibrium
distribution. In Eq. (3), the position r and velocity c
i
are
discrete, i.e. the possible positions are restricted to the
sites of a lattice, and thus the possible velocities are the
vectors c
i
(i =1, b) connecting the b nearest neighbor sites
of this lattice. Note that Eq. (3) represents a propagation,
followed by a “collision" (relaxation to the equilibrium dis-
tribution). From the single particle distribution function, the
hydrodynamic variables of interest ÷ the local gas density
ρ and velocity c
i
÷ are obtained by summing up over all
possible velocities:
1
( , ) ( , , )
b
i
t f t ρ
=
=
∑ i
r r c ,
1
( , ) ( , ) ( , , )
b
i i
i
t t f t ρ
=
=

r u r c c r . (4)
It can be shown that the flow fields obtained from the LB
model are ÷ to order
2
t δ ÷ equivalent to those obtained
from the Navier-Stokes equation, where the viscosity is set
by the relaxation time τ. One of the advantages of the LB
model over other finite difference models for fluid flow, is
that boundary conditions can be modeled in a very simple
way. This makes the method particularly suit to simulate
large moving particles suspended in a fluid phase. An ob-
vious choice of the boundary condition is where the gas
next to the solid particle moves with the local velocity of the
surface of the solid particle, i.e. the so-called “stick”
boundary condition. For a spherical particle suspended in
an infinite three-dimensional system, moving with velocity
v, this condition will give rise to a frictional force on the
particle 3 d µ = π F v , at least in the limit of low particle
Reynolds numbers d ρ ε µ = Re v , where d is the hydrody-
namic diameter of the particle, and µ is the shear viscosity.
A particular efficient and simple way to enforce stick
boundary conditions for static particles in the LB model is
to let the distributions “bounce back” at the boundary
nodes (Ladd & Verberg, 2001); these nodes are defined as
the points halfway the two lattice sites which are closest to
the actual surface of the particle. The “bounce-back” rule
means that the distribution function moves back into the
direction that it comes from (see Fig. 2):
'
( , 1) ( , )
i i
f t f t + = r r , (5)

Fig. 2 Illustration of the bounce-back rule. The distribution at site r
that moves at time t into direction i, instead of arriving at the
(virtual) site s, is bounced at the boundary node, and thus ar-
rives back at site r at time t+1, but now headed in the opposite
direction.
where i and ' i are opposite links. This rule ensures that
the fluid velocity at the boundary node indeed vanishes:
the momentum at the boundary node at time t+1/2 is given
by:
' '
( , ) ( , 1)
b b i i i i
f t f t ρ = + + u r c r c , (6)
Inserting Eq. (5) and using
' i i
= − c c gives that 0
b b
ρ = u .
For non-static particles, the local fluid velocity must be set
equal to the local boundary velocity
b
v . This can be
achieved by a simple modification of the bounce-back rule:
'
( , 1) ( , )
i i i
f t f t + = + ⋅ r r a c , (7)
where a is chosen such that
b b
= u v . Note that only the
component of
b
v in the direction of the link can be set in
this way. For details we refer to Ladd and Verberg (2001).
The drag force F
d
can also be directly measured in the
simulation, from the change in gas momentum due to the
boundary rules. In this way, the average drag force
d
F
on a sphere in a static random array can be obtained,
where the gas flow is set at a constant velocity u
0
, accord-
ing to the desired Re number. The friction coefficient β
follows then from:
d
p 0
1
F
V u
ε
β

= ,
3
p
1
6
V d = π . (8)
By using this method, we found for low Reynolds num-
ber excellent agreement with data obtained by multipole
expansion methods (van der Hoef et al., 2005). By contrast,
it was found that the widely used empirical correlations (1)
and (2) significantly underestimate the drag force, at least
for low Reynolds numbers. Based on the Carman-Kozeny
approximation, we derived an expression for the correction
of the monodisperse drag force to account for bidispersity
which only depends on
i i
y d d = with d the average
CHINA PARTICUOLOGY Vol. 3, Nos. 1-2, 2005

72

diameter, for details see van der Hoef et al. (2005). In Fig. 3
we present some LBM results for a binary mixture at finite
Reynolds numbers. In this figure the individual drag force
F
i
divided by the drag force F(φ) of a monodisperse system
at the same solids volume fraction φ, is plotted as a func-
tion of the correction factor
2
(1 )
i i
y y φ φ − + that we derived,
where F(φ) is our best fit to LBM simulation data for mono-
disperse systems:
2
2
( ) 10 (1 ) [1 1.5 ]
(1 )
F
φ
φ φ φ
φ
= + − +

. (9)
As can be seen from Fig. 3, we find excellent agreement
between our data and theory. It should be noted here that
the assumption F
i
=F(φ), which is currently used in literature
can lead to differences with the LBM simulation data up to
a factor of 5. This finding indicates that one should be cau-
tious with relying on ad hoc modifications of drag laws for
monodisperse systems to extend their “validity” to
polydisperse systems. In addition, this result highlights the
usefulness of microscopic simulation methods, because
the experimental determination of the individual effective
drag force in a dense assembly would be extremely diffi-
cult.


Fig. 3 Example of particle configuration generated with a Monte Carlo
procedure for a binary system (upper) and dimensionless drag
force computed for small and large particles from LBM (lower).
3. Discrete Particle Model (DPM)
The discrete particle model is one level higher in the
multi-scale hierarchy. The most important difference with
the lattice Boltzmann model is that in this model the size of
the particles is smaller than the grid size that is used to
solve the equations of motion of the gas phase. This
means that for the interaction with the gas phase, the par-
ticles are simply point sources and sinks of momentum,
where the finite volume of the particles only comes in via
an average gas fraction in the drag force relations. A sec-
ond (technical) difference with the LB model is that the
evolution of the gas phase now follows from a finite differ-
ence scheme of the Navier-Stokes equation, rather than
the Boltzmann equation. A complete description of the
method can be found in Hoomans et al. (1996), however,
we will briefly discuss some of the basic elements here.
The discrete particle model consists of two parts: a La-
grangian part for updating the positions and velocities of
the solid particles, and an Eulerian part for updating the
local gas density and velocity. In the Lagrangian part, the
equation of motion of each particle i (velocity
i
v
G
, mass
i
m ,
volume
i
V ) is given by Newton’s law
( )
( )
pp pw
d
d 1
i i
i i i i i i
v V
m mg u v V p F F
t
β
ε
= + − − ∇ + +

G
G G
G G G
, (10)
where the RHS represents the total force acting on the
particle. This includes external forces (the gravitational
force
i
mg
G
), interaction forces with the gas phase (drag
force ( ) ~
i
u v β −
G G
and pressure force
i
V p ∇ ), and finally the
particle-particle forces
pp
i
F
G
and particle-wall forces
pw
i
F
G
,
which represents the momentum exchange during colli-
sions, and possible long-range attractions between the
particles, and particles and walls, respectively. There are,
in principle, two ways to calculate the trajectories of the
solid particles from Newton’s law. In a time-driven nu-
merical simulation, the new position ( d )
i
r t t +
G
and velocity
( d )
i
v t t +
G
are calculated from the values at time t, via a
standard integration scheme for ODE’s. Such type of
simulation is in principle suitable for any type of interaction
force between the particles. In an event-driven simulation,
the interactions between the particles are considered in-
stantaneous (“collisions”), and the systems evolves directly
(“free flight”) from nearest collision event to next-nearest
collision event, etc. This method is efficient for low-density
systems, however it is not suitable for dense-packed sys-
tems, or systems with long-range forces. In the Eulerian
part of the code, the evolution of the gas phase is deter-
mined by the volume-averaged Navier-Stokes equations:
( ) 0 u
t
ερ ερ

+ ∇⋅ =

G
, (11)
( ) u uu p S g
t
ερ ερ ε ετ ερ

+ ∇⋅ = − ∇ − ∇⋅ − +

G
G G G G
, (12)
where τ is the usual stress tensor, which includes the
van der Hoef, van Sint Annaland & Kuipers: Computational Fluid Dynamics for Dense Gas-Solid Fluidized Beds

73

coefficient of shear viscosity. Note that there is a full
two-way coupling with the Lagrangian part, i.e., the reac-
tion from drag and pressure forces on the solid particles is
included in the momentum equation for the gas phase via a
source term S
G
:
( ) ( )
1
d
1
i
i i
i
V
S u v r r V
V
β
δ
ε
= − −



G
G G G G
. (13)
Equations (11) and (12) are solved with a semi-implicit
method for pressure linked equations (SIMPLE-algorithm),
with a time step that is in general an order of magnitude
larger than the time step used to update the particle posi-
tions and velocities. The strength of the DP model is that it
allows to study the effect of the particle-particle interactions
on the fluidization behavior. In the most detailed model of
description, the interparticle contact forces includes normal
and tangential repulsive forces (modeled by linear springs),
and dissipative forces (modeled by “dash pots”), and tan-
gential friction forces (Walton, 1993). A DPM simulation
study by Hoomans et al. (1996) showed that the hetero-
geneous flow structures in dense gas-fluidized beds are
partly due to the collisional energy dissipation. More re-
cently, Li and Kuipers (2003) demonstrated that such flow
structures are also strongly influenced by the degree of
non-linearity of the particle drag with respect to the gas frac-
tion ε. Bokkers et al. (2004a; 2004b) studied the effect of the
closures for gas-particle drag on the bubble-induced mixing
in a pseudo 2D gas-fluidized bed and found that the best
agreement between theory and experiment was obtained in
case the LBM-generated drag closures reported by Hill et al.
(2001a; 2001b) were used in their DPM simulations.
One of the great advantages of discrete particle simula-
tions is that it allows to study properties of the system that
are very diificult to obtain via experimentation. A particu-
larly important example is the velocity distribution of the
particles, i.e. the probability of finding a particle with a ve-
locity component v
α
with α=x,y,z. It would be extremely
difficult to obtain reliable estimates for the velocity distribu-
tion from experiments; yet, this function is of great rele-
vance for the validity of the higher scale two-fluid model
(see next section) derived from the kinetic theory, where it
is assumed that the velocity distribution is both isotropic
and nearly Gaussian. The discrete particle simulations are
an ideal tool for testing this assumption, since it is relatively
straightforward to measure the velocity distribution as all
particle velocities are known at any moment in time.
We have studied the velocity distribution for two cases:
in Fig. 4 we show the result for a fluidized bed of ideal (i.e.
perfectly smooth and elastic) and non-ideal (i.e. rough and
inelastic) particles. The system contained 25000 particles
of 2.5 mm diameter, where the gas velocity is set to 1.5
times the minimum fluidization velocity. Details of the
sampling procedure for obtaining the velocity distributions
can be found in Goldschmidt et al. (2002). Figure 4 shows
that for both ideal and non-ideal particles, the velocity dis-
tributions do not deviate significantly from a Gaussian and
Maxwellian distribution. However, Fig. 4 reveals a clear
anisotropy of the distribution in case of non-ideal particles.
A possible explanation is the formation of dense particle
clusters in the case of inelastic collisions, which may dis-
turb the spatial homogeneity and thereby causing colli-
sional anisotropy. Analysis (Jenkins & Savage, 1983) of the
normal and tangential component of the impact velocity
indeed showed that, in dense gas-fluidized beds, not all
impact angles occur with the same frequency.


Fig. 4 DPM simulation data for the normalized particle velocity dis-
tribution f
x
(C
x
), f
y
(C
y
), f
z
(C
z
) and f(C), compared to a Gaus-
sian/Maxwellian distribution. Upper graph: ideal particles;
lower graph: non-ideal particles.
4. Two-Fluid Model (TFM)
The maximum number of particles that can be simulated
with the DP model, as described in the previous section, is
typically less than a million, whereas the number of parti-
cles that are present in an industrial size fluidized bed can
be two to three orders of magnitude higher. Since both the
CPU time and the required memory scales linear with the
number of particles, it is obvious that DPM simulations of
industrial size fluidized beds are beyond the capability of
commercially available computer facilities within the fore-
seeable future. Therefore, a different type of model is used
for simulations at larger scales, where the concept of a
solid phase consisting of individual, distinguishable parti-
cles is abandoned. This so-called two-fluid continuum
CHINA PARTICUOLOGY Vol. 3, Nos. 1-2, 2005

74

model (TFM) describes both the gas phase and the solids
phase as fully inter-penetrating continua, using a set of
generalized Navier-Stokes equations (Kuipers et al., 1992;
Gidaspow, 1994). That is, the time evolution of the gas
phase is still governed by (11) and (12); for the solid phase,
the discrete particle part (10) is now replaced by a set of
continuum equations of the same form as (11) and (12):
( )
s s s s
0 v
t
ε ρ ε ρ

+ ∇⋅ =

G
, (14)
( )
s s s s s s s s s s
v vv p p S g
t
ε ρ ε ρ ε ε τ ε ρ

+ ∇⋅ = − ∇ −∇ − ∇⋅ + +

G
G G GG
,
(15)
with
s
ρ , v
G
and
s
1 ε ε = − the local density, velocity, and
volume fraction of the solid phase, respectively. In this
description the source term S
G
is slightly different from
(13), namely
( ) S u v β = −
G
G G
. (16)
Obviously, the numerical scheme for updating the solid
phase is now completely analogous to (and synchronous
with) that of the gas phase. Since the concept of particles
has disappeared completely in such a modeling, the effect
of particle-particle interactions can only be included indi-
rectly, via an effective solids pressure and effective solids
viscosity. A description which allows for a slightly more
detailed description of particle-particle interactions follows
from the kinetic theory of granular flow (KTGF); such
theory expresses the diagonal and off-diagonal elements
of the solids stress tensor (i.e. the solids pressure and
solids shear rate) as a function of the granular temperature
for a monodisperse particle system, defined as:
p p
1
3
C C θ = ⋅
G G
, (17)
where
p
C
G
represents the particle fluctuation velocity and
the brackets indicate ensemble-averaging. The time evolu-
tion of the granular temperature itself is given by:
( ) ( )
( )
( )
s s s s
s s s s s
3
2
: 3 ,
v
t
p I v q
ε ρ θ ε ρ θ
ε τ ε βθ γ
∂ (
+ ∇⋅ =
(

¸ ¸
− + ∇ − ∇⋅ − −
G
G G
(18)
with
s
q
G
the kinetic energy flux, and γ the dissipation of
kinetic energy due to inelastic particle collisions. In equa-
tions (14)−(18) there are still a number of unknown quanti-
ties (pressure, stress tensor, energy flux), which must be
expressed in terms of the basic hydrodynamic variables
(density, velocity, granular temperature), in order to get a
closed set of equations. The derivation of such constitutive
equations follows from the KTGF, and can be found in the
books by Chapman and Cowling (1970) and Gidaspow
(1994) and the papers by Jenkins and Savage (1983) and
Ding and Gidaspow (1990). In this work, the constitutive
equations developed by Nieuwland et al. (1996) have been
used for the particle phase rheology.
In Fig. 5 we show the simulated bubble formation for a
pseudo two-dimensional (2D) bed (bed geometry:
0.57m×0.015m×1.0m (w × d × h)) operated with a central
jet (diameter 0.015m) at a velocity of 40 times the incipient
fluidization velocity. The bed contains ballotini with a parti-
cle diameter and density of 500µm and
s
ρ =2660kg⋅m
-3

respectively. Clearly, a very complex bubble pattern results
from the jet operation where the size and the shape of the
formed bubbles continuously change. It can also clearly be
seen that bubble coalescence occurs leading to a rapid
increase in the bubble size.

Fig. 5 Computed bubble formation in a pseudo 2D gas-fluidized bed with a central jet. Bed material: ballotini with d
p
=500 µm
and ρ
s
=2 660 kg⋅m
-3
. Jet velocity: 10.0 m⋅s
-1
(40u
mf
).
t =1.0000 s t =2.0000 s t =3.0000 s t =4.0000 s t =5.0000 s
t =6.0000 s t =7.0000 s t =8.0000 s t =9.0000 s
t =10.0000 s
van der Hoef, van Sint Annaland & Kuipers: Computational Fluid Dynamics for Dense Gas-Solid Fluidized Beds

75

5. Discrete Bubble Model (DBM)
Although the two fluid model can simulate fluidized beds
at life-size scales, the largest scale industrial fluidized bed
reactors (diameter 5 meters, height 16 meters) are still
beyond its capabilities. However, it is possible to introduce
yet another upscaling by considering the bubbles as dis-
crete entities, as observed in the DPM and TFM models of
gas-fluidized beds. This is the so-called discrete bubble
model, which has been successfully applied in the field of
gas-liquid bubble columns (Delnoij et al., 1997). The idea
to apply this model to describe the large scale solids cir-
culation that prevails in gas-solid reactors is new, however.
In this paper, we want to show some first results of the
discrete bubble model applied to gas-solid systems, which
involves some slight modifications of the equivalent model
for gas-liquid systems. To this end the emulsion phase is
modeled as a continuum, like the liquid in a gas-liquid
bubble column, and the larger bubbles are treated as dis-
crete bubbles. Note that granular systems have no surface
tension, so in that respect there is a pronounced difference
with the bubbles present in gas-liquid bubble columns. For
instance, the gas will be free to flow through a bubble in the
gas-solid systems, which is not the case for gas-liquid
systems. As far as the numerical part is concerned, the
DBM strongly resembles the discrete particle model as
outlined in section 3, since it is also of the Euler-Lagrange
type with the emulsion phase described by the vol-
ume-averaged Navier-Stokes equations:
( ) 0 u
t
ερ ερ

+ ∇⋅ =

G
, (19)
( ) u uu p S g
t
ερ ερ ε ετ ερ

+ ∇⋅ = − ∇ −∇⋅ − +

G
G G G G
, (20)
whereas the discrete bubbles are tracked individually ac-
cording to Newton’s second law of motion:
b
b tot
d
d
v
m F
t
=
G
G
, (21)
where
tot
F
G
is the sum of different forces acting on a single
bubble:
tot g d p L VM
F F F F F F = + + + +
G G G G G G
. (22)
As in the DPM model, the total force on the bubble has
contributions from gravity (
g
F
G
), pressure gradients (
p
F
G
)
and drag from the interaction with emulsion phase (
d
F
G
).
For the drag force on a single bubble (diameter d
b
), the
correlations for the drag force on a single sphere are used,
only with a modified drag coefficient C
d
, such that it yields
the Davies-Taylor relation
br b
0.711 v gd = for the rise ve-
locity of a single bubble. Note that in (21), there are two
forces present which are not found in the DPM, namely the
lift force
L
F
G
and the virtual mass force
VM
F
G
. The lift force
is neglected in this application, whereas the virtual mass
force coefficient is set to 0.5. An advantage of this ap-
proach to model large scale fluidized bed reactors is that
the behaviour of bubbles in fluidized beds can be readily
incorporated in the force balance of the bubbles. In this
respect, one can think of the rise velocity, and the tendency
of rising bubbles to be drawn towards the center of the bed,
from the mutual interaction of bubbles and from wall effects
(Kobayashi et al., 2000). Coalescence, which is an highly
prevalent phenomenon in fluidized beds, can also be easily
included in the DBM, since all the bubbles are tracked
individually.
With the DBM, two preliminary calculations have been
performed for industrial scale gas-phase polymerization
reactors, in which we want to demonstrate the effect of the
superficial gas velocities, set to 0.1 m⋅s
-1
and 0.3 m⋅s
-1
. The
geometry of the fluidized bed was 1.0 m×3.0 m×1.0 m
(w × h × d). The emulsion phase has a density of
400kg⋅m
-3
and the apparent viscosity was set to 1.0Pa⋅s.
The density of the bubble phase was 25 kg⋅m
-3
. The bub-
bles were injected via 49 nozzles positioned equally dis-
tributed in a square in the middle of the column.
In Fig. 6 snapshots are shown of the bubbles that rise in
the fluidized bed with a superficial velocity of 0.1m⋅s
-1
and
0.3m⋅s
-1
, respectively. It is clearly shown that the bubble
hold-up is much larger with a superficial velocity of
0.3m⋅s
-1
. However, the number of bubbles in this case
might be too large, since coalescence has not been taken
into account in these simulations. In Fig. 6 in addition time-
averaged plots are shown of the emulsion velocity after
Fig. 6 Snapshots of the bubble configurations (left) computed from the
DBM model without coalescence, and the time average vector
plots of the emulsion phase (right) after 100 s of simulation; top:
u
0
=0.1 m⋅s
-1
, d
b
=0.04 m; bottom: u
0
=0.3 m⋅s
-1
, d
b
=0.04 m.
CHINA PARTICUOLOGY Vol. 3, Nos. 1-2, 2005

76

100s of simulation. The large convection patterns, upfow in
the middle, and downflow along the wall, and the effect of
the superficial gas velocity, are clearly demonstrated. Fu-
ture work will be focused on implementation of closure
equations in the force balance, like empirical relations for
bubble rise velocities and the interaction between bubbles.
The model can be augmented with energy balances to
study temperature profiles in combination with the large
circulation patterns.
6. Summary and Outlook
In this paper we have presented an overview of the
multi-scale methods that we use to study gas-solid fluid-
ized beds. The key idea is that the methods at the smaller,
more detailed scale can provide qualitative and quantita-
tive information which can be used in the higher scale
models. A typical example of such qualitative information is
the insight (from the DPM simulations) that inelastic colli-
sions and nonlinear drag can lead to heterogeneous flow
structures. Even more important, however, is the quantita-
tive information that the smaller scale models can provide.
A typical example of this is the drag force relation obtained
from the LBM simulations, which finds its direct use in both
the DPM and TFM simulations. We should note here that
although the new drag force relations seem to give results
at the DPM/TFM level which compare better with the ex-
perimental findings, these relations are still far from optimal.
In particular, it should be borne in mind that these drag
force relations are derived for static, unbounded, homo-
geneous arrays of mono-disperse spheres. Yet, at the
DPM/TFM level these relations are applied to systems
which are, even locally, inhomogeneous and non-static;
furthermore, rather ad hoc modifications are used to allow
for polydispersity. In future work, we want to focus on de-
veloping drag force relations for systems which deviate
from the ideal conditions, where the parameters which
would quantify such deviation may be trivial to define
(polydispersity: width of the size distribution; moving parti-
cles: granular temperature) or not so trivial (inhomogenei-
ties). Our lattice Boltzmann results for the drag force in
binary systems (van der Hoef et al., 2005) revealed sig-
nificant deviations with the ad hoc modifications of the
monodisperse drag force relations, in which it is assumed
that the drag force scales linearly with the particle diameter.
At present, only qualitative information from the DPM
simulations is obtained, such as the aforementioned het-
erogeneous flow structures, which is caused by dissipative
forces.
Another example is the functional form of the velocity
distribution. It was found that that dissipative interaction
forces cause an anisotropy in the distribution, although the
functional form remains close to Gaussian for all three
directions (Goldschmidt et al., 2002). It would be interest-
ing to include the effect of anisotropy at the level of the
TFM, for instance along the lines of the kinetic theory de-
veloped by Jenkins and Richman (1988) for shearing
granular flows. Although the continuum models have been
studied extensively in the literature (e.g. Kuipers et al.,
1992; Gidaspow, 1994), these models still lack the capa-
bility of describing quantitatively particle mixing and seg-
regation rates in multi-disperse fluidized beds. An impor-
tant improvement in the modeling of life-size fluidized beds
could be made if direct quantitative information from the
discrete particle simulations could find its way in the con-
tinuum models. In particular, it would be of great interest to
find improved expressions for the solid pressure and the
solid viscosity, as they are used in the two fluid model,
however, it is a non-trivial task to extract direct data on the
solid viscosity and pressure in a DPM simulation. A very
simple, indirect method for obtaining the viscosity is to
monitor the decay of the velocity of a large spherical in-
truder in the fluidized bed. The viscosity of the bed follows
then directly from the Stokes-Einstein formula for the drag
force. Very preliminary results ÷ obtained from data of a
high velocity impact ÷ were in reasonable agreement with
the experimental values for the viscosity. More elaborate
simulations of these systems are currently underway.
Finally, the discrete bubble model applied to gas-solid
systems seems to be a promising new approach for de-
scribing the large scale motion in life-size chemical reac-
tors. Essential for this model to be successful is that reli-
able information with regard to rise velocities and mutual
interaction of the bubbles is incorporated, which can be
obtained from the lower scale simulations. In particular, the
TFM and DPM simulations will be used to guide the for-
mulation of additional rules to properly describe the coa-
lescence of bubbles, which is at present not incorporated
in the model. This will be the subject of future research.
Acknowledgment
The lattice Boltzmann simulations have been performed with the
SUSP3D code developed by Anthony Ladd. We would like to
thank him for making his code available, and Albert Bokkers for
performing the DPM, TFM and DBM simulations.
References
Bokkers, G. A., van Sint Annaland, M. & Kuipers, J. A. M. (2004a).
Mixing and segregation in a bi-dispersed gas-solid fluidized bed:
a numerical and experimental study. Powder Technol., 140,
176-186.
Bokkers, G. A., van Sint Annaland, M. & Kuipers, J. A. M. (2004b).
Comparison of continuum models using the kinetic theory of
granular flow with discrete particle models and experiments:
extent of particle mixing induced by bubbles. Proceedings of
Fluidization XI (pp.187-194), May 9-14, 2004, Naples, Italy.
Chapman, S. & Cowling, T. G. (1970). The Mathematical Theory
of Non-uniform Gases (Trial mode edition). Cambridge: Cam-
bridge University Press.
Delnoij, E., Kuipers, J. A. M. & van Swaaij, W. P. M. (1997). Com-
putational fluid dynamics applied to gas-liquid contactors. Chem.
Eng. Sci., 52, 3623. PhD thesis, University of Twente, Enschede,
The Netherlands.
Ding, J. & Gidaspow, D. (1990). A bubbling fluidization model
van der Hoef, van Sint Annaland & Kuipers: Computational Fluid Dynamics for Dense Gas-Solid Fluidized Beds

77

using kinetic theory of granular flow. AIChE J., 36, 523.
Ergun, S. (1952). Fluid flow through packed columns. Chem. Eng.
Process., 48, 89.
Frisch, U., Hasslacher, B. & Pomeau, Y. (1986). Lattice gas
automata for the Navier-Stokes equation. Phys. Rev. Lett., 56,
1505.
Gidaspow, D. (1994). Multiphase Flow and Fluidization: Contin-
uum and Kinetic Theory Descriptions. Boston: Academic Press.
Goldschmidt, M. J. V., Beetstra, R. & Kuipers, J. A. M. (2002).
Hydrodynamic modelling of dense gas-fluidised beds: com-
parison of the kinetic theory of granular flow with 3-D hard-
sphere discrete particle simulations. Chem. Eng. Sci., 57, 2059.
Hill, R. J., Koch, D. L. & Ladd, A. J. C. (2001a). The first effects of
fluid inertia on flow in ordered and random arrays of spheres. J.
Fluid Mech., 448, 213.
Hill, R. J., Koch, D. L. & Ladd, A. J. C. (2001b). Moderate-Rey-
nolds-number flows in ordered and random arrays of spheres. J.
Fluid Mech., 448, 243.
Hoomans, B. P. B., Kuipers, J. A. M., Briels, W. J. & van Swaaij, W.
P. M. (1996). Discrete particle simulation of bubble and slug
formation in a two-dimensional gas-fluidized bed: a hard sphere
approach. Chem. Eng. Sci., 51, 99.
Jenkins, J. T. & Savage, S. B. (1983). A theory for the rapid flow of
identical, smooth, nearly elastic particles. J. Fluid Mech., 130,
187.
Jenkins, J. T. & Richman, M. W. (1988). Plane simple shear flow
of smooth inelastic circular disks: the anisotropy of the second
moment in the dilute and dense limits. J. Fluid Mech., 192, 313.
Kobayashi, N., Yamazaki, R. & Mori, S. (2000). A study on the
behavior of bubbles and soldis in bubbling fluidized beds.
Powder Technol., 113, 327.
Kuipers, J. A. M., van Duin, K. J., van Beckum, F. P. H. & van
Swaaij, W. P. M. (1992). A numerical model of gas-fluidized
beds. Chem. Eng. Sci., 47, 1913.
Kuipers, J. A. M., Hoomans, B. P. B. & van Swaaij, W. P. M.
(1998). Hydrodynamic modeling of gas-fluidized beds and their
role for design and operation of fluidized bed chemical reactors.
Proceedings of the Fluidization IX conference (pp.15-30), Du-
rango, USA.
Kuipers, J. A. M. & van Swaaij, W. P. M. (1998). Computational
fluid dynamics applied to chemical reaction engineering. Adv.
Chem. Eng., 24, 227.
Kunii, D. & Levenspiel, O. (1991). Fluidization Engineering. But-
terworth Heinemann series in Chemical Engineering, London.


Li, J. & Kuipers, J. A. M. (2003). Gas-particle interactions in dense
gas-fluidized beds. Chem. Eng. Sci., 58, 711.
McNamara, G. R. & Zanetti, G. (1988). Use of the Boltzmann
equation to simulate lattice-gas automata. Phys. Rev. Lett., 61,
2332.
Nieuwland, J. J., van Sint Annaland, M., Kuipers, J. A. M. & van
Swaaij, W. P. M. (1996). Hydrodynamic modeling of gas/particle
flow in riser reactors. AIChE J., 42, 1569.
Schiller, L. & Nauman, A. (1935). A drag coefficient correlation.
V.D.I. Zeitung, 77, 318.
Succi, S. (2001). The Lattice Boltzmann Equation for Fluid dy-
namics and Beyond. Oxford: Oxford Science Publications.
van der Hoef, M. A., Beetstra, R. & Kuipers, J. A. M. (2005). Lat-
tice Boltzmann simulations of low Reynolds number flow past
mono- and bidisperse arrays of spheres: results for the perme-
ability and drag force. J. Fluid Mech., 528, 233-253.
van Swaaij, W. P. M. (1990). Chemical reactors. In Davidson, J. F.
& Clift, R. (Eds.), Fluidization. London: Academic Press.
Walton, O. R. (1993). Numerical simulation of inelastic, frictional
particle-particle interactions. In Roco, M. C. (Ed.), Particulate
Two-Phase Flow, Butterworth Heinemann series in Chemical
Engineering, London.
Wen, C. Y. & Yu, Y. H. (1966). Mechanics of fluidization. AIChE
Symp. Ser., 62, 100.
Manuscript received January 10, 2005 and accepted April 1, 2005.



Ladd, A. J. C. & Verberg, R. (2001). Lattice-Boltzmann simulations
of particle fluid suspensions. J. Stat. Phys., 104, 1191.

amongst others). Hence. 2. 3. At an even larger scale a continuum description is employed for the solid phase. In the following sections we will describe each of these models in more detail. but by a local density and velocity field. Re is the particle Reynolds number.e. Gidaspow. and the results should be considered as very preliminary.687 / Re  Cd =   0. This model is an adapted version of the discrete bubble model for gas-liquid bubble columns. (2) where µ is the viscosity of the gas phase. In this paper we will give an overview of these four levels of modeling as they are employed in our research group.15Re 0.65 . Note that in chemical engineering. such as the Ergun (1952) correlation for porosities ε < 0. . and these models are therefore widely used in commercial fluid flow simulators of industrial scale equipment. the (larger) bubbles that are present in gas-solid fluidized beds are considered as discrete objects. 2005 Larger geometry Lattice Boltzmann Model Discrete Particle Model Continuum Model Discrete Bubble Model Fluid-particle interaction Particle-particle interaction Particle-particle interaction Bubble behavior Large scale motion Industrial size Larger scale phenomena Fig.44  ( ) Re < 103 Re > 103 .8 : (1 − ε ) + 1. and for both phases an Eulerian code is used to describe the time evolution (see Kuipers et al. As a logical consequence of this approach a closure law for the effective momentum exchange has to be specified. the solid phase is not described by individual particles.8 : βd2 3 = CdRe (1 − ε ) ε −2.75 (1 − ε ) Re . Lattice Boltzmann Model (LBM) The lattice Boltzmann model (LBM) originates from the lattice-gas cellular automata (LGCA) models (Frisch et al. The advantage of this model is that it can predict the flow behavior of gas-solid flows at life-size scales.70 CHINA PARTICUOLOGY Vol. similar to the solid particles in the DPM model. i. at the largest scale.. for which the expression of Schiller and Nauman (1935) is used. has been developed quite recently. 1 Multi-level modeling scheme for dense gas-fluidized beds. 1-2. We want to stress that this model. µ 4  24 1 + 0. and Cd the drag coefficient. The information obtained in the two smaller-scale models is then included in the continuum models via the kinetic theory of granular flow. as outlined in section 5. to date mainly empirical relations are used for the friction coefficient β (defined by (1) and (2)). Nos. in this model both the gas-phase and the solid phase are treated on an equal footing. 1994. 1992. which can be obtained from the aforementioned lattice Boltzmann simulations.. comparison with laboratory scale experiments. Finally. βd2 = 150 µ ε ε 2 (1) and the Wen and Yu (1966) equation for porosities ε > 0.

where d is the hydrodynamic diameter of the particle.e. The drag force Fd can also be directly measured in the simulation. t )c i ' + fi (r . where it has been demonstrated both numerically and theoretically that the resulting macroscopic flow fields obey the Navier-Stokes equation. the possible positions are restricted to the sites of a lattice. (8) β = 6 Vp u0 By using this method. b) connecting the b nearest neighbor sites of this lattice. which involves propagations and collisions of particles on a lattice. this condition will give rise to a frictional force on the particle F = 3πµ dv . the average drag force Fd on a sphere in a static random array can be obtained. r . followed by a “collision" (relaxation to the equilibrium distribution). so that it represents a propagation and collision of the particle distributions instead of the actual particles as in the LGCA models (McNamara & Zanetti. A particular efficient and simple way to enforce stick boundary conditions for static particles in the LB model is to let the distributions “bounce back” at the boundary nodes (Ladd & Verberg. the fundamental equation in the kinetic theory which underlies the equations of hydrodynamics. t ) = ∑ f (c i . From the single particle distribution function. For details we refer to Ladd and Verberg (2001). In this way. is that boundary conditions can be modeled in a very simple way. t ) = f (c i . (5) τ ( ) where f is the single particle distribution function. simplified version of the molecular dynamics model. This makes the method particularly suit to simulate large moving particles suspended in a fluid phase. The distribution at site r that moves at time t into direction i. from the change in gas momentum due to the boundary rules. ρ ( r. and µ is the shear viscosity. By contrast. Note that Eq. (4) i =1 i =1 b b It can be shown that the flow fields obtained from the LB model are  to order δt 2  equivalent to those obtained from the Navier-Stokes equation. the so-called “stick” boundary condition. van Sint Annaland & Kuipers: Computational Fluid Dynamics for Dense Gas-Solid Fluidized Beds 1986) for simple fluids. t ) − f eq (c i . (3) represents a propagation. at least for low Reynolds numbers. 2005). Vp = πd 3 . From a macroscopic point of view. (7) where a is chosen such that u b = v b . the position r and velocity ci are discrete. i. where the gas flow is set at a constant velocity u0. For non-static particles. we found for low Reynolds number excellent agreement with data obtained by multipole expansion methods (van der Hoef et al. The “bounce-back” rule means that the distribution function moves back into the direction that it comes from (see Fig. LGCA models have proved a simple and efficient way to simulate a simple fluid at the microscopic level. it was found that the widely used empirical correlations (1) and (2) significantly underestimate the drag force.. and thus the possible velocities are the vectors ci (i =1. This can be achieved by a simple modification of the bounce-back rule: fi (r . r + c i δt . at least in the limit of low particle Reynolds numbers Re = ρ d εv µ . where i and i ' are opposite links. One of the advantages of the LB model over other finite difference models for fluid flow. (5) and using c i = −c i ' gives that ρbu b = 0 . moving with velocity v. according to the desired Re number. (3). t ) = ∑ c i f (c i .van der Hoef. which is equivalent to the fluid density in the 6 dimensional velocity-coordinate space. In its most simple form the finite difference scheme reads: δt f (v . (3) 71 the points halfway the two lattice sites which are closest to the actual surface of the particle. t ) . t + 1)c i . In Eq. Note that only the component of v b in the direction of the link can be set in this way. the LB model can be regarded as a finite difference scheme that solves the Boltzmann equation. we derived an expression for the correction of the monodisperse drag force to account for bidispersity which only depends on y i = di d with d the average ρ ( r. r. 1988). The lattice Boltzmann model is the ensemble averaged version of the LGCA model. where the viscosity is set by the relaxation time τ. i. For a spherical particle suspended in an infinite three-dimensional system. The LGCA model is basically a discrete. t ) . t ) . t )u( r. 2): fi (r . 2 Illustration of the bounce-back rule. r. t + δt ) − f (c i . and f eq represents the equilibrium distribution. (6) Inserting Eq. Based on the Carman-Kozeny approximation.e. t ) + a ⋅ c i . but now headed in the opposite direction. instead of arriving at the (virtual) site s. t + 1) = fi ' (r . the local fluid velocity must be set equal to the local boundary velocity v b . t + 1) = fi ' (r . r. and thus arrives back at site r at time t+1. t ) . This rule ensures that the fluid velocity at the boundary node indeed vanishes: the momentum at the boundary node at time t+1/2 is given by: ρbu b = fi ' (r . An obvious choice of the boundary condition is where the gas next to the solid particle moves with the local velocity of the surface of the solid particle. r . The friction coefficient β follows then from: 1 − ε Fd 1 . is bounced at the boundary node. the hydrodynamic variables of interest  the local gas density ρ and velocity ci  are obtained by summing up over all possible velocities: Fig. 2001). these nodes are defined as .

via a Fig. (9) As can be seen from Fig. and finally the particle-particle forces Fi pp and particle-wall forces Fi pw . In this figure the individual drag force Fi divided by the drag force F(φ) of a monodisperse system at the same solids volume fraction φ. respectively. and possible long-range attractions between the particles. the evolution of the gas phase is determined by the volume-averaged Navier-Stokes equations: ∂ (11) (ερ ) + ∇ ⋅ ερ u = 0 . standard integration scheme for ODE’s. two ways to calculate the trajectories of the solid particles from Newton’s law. This includes external forces (the gravitational force mi g ). the equation of motion of each particle i (velocity v i . (2005). the interactions between the particles are considered instantaneous (“collisions”). where the RHS represents the total force acting on the particle. This means that for the interaction with the gas phase. In the Lagrangian part. rather than the Boltzmann equation. which includes the . and the systems evolves directly (“free flight”) from nearest collision event to next-nearest collision event. In Fig. 3 Example of particle configuration generated with a Monte Carlo procedure for a binary system (upper) and dimensionless drag force computed for small and large particles from LBM (lower). ∂t ∂ (ερ u ) + ∇ ⋅ ερ uu = −ε∇p − ∇ ⋅ ετ − S + ερ g . 3 we present some LBM results for a binary mixture at finite Reynolds numbers. In a time-driven numerical simulation. This method is efficient for low-density systems. A complete description of the method can be found in Hoomans et al. interaction forces with the gas phase (drag force ~ β (u − v i ) and pressure force Vi ∇p ). volume Vi ) is given by Newton’s law mi dv i Vβ = mi g + i (u − v i ) − Vi ∇p + Fi pp + Fi pw . dt (1 − ε ) (10) φ (1 − φ )2 + (1 − φ )2 [1 + 1. (12) ∂t where τ is the usual stress tensor. 1-2. the particles are simply point sources and sinks of momentum. for details see van der Hoef et al. where the finite volume of the particles only comes in via an average gas fraction in the drag force relations. and particles and walls. This finding indicates that one should be cautious with relying on ad hoc modifications of drag laws for monodisperse systems to extend their “validity” to polydisperse systems. (1996). where F(φ) is our best fit to LBM simulation data for monodisperse systems: F (φ ) = 10 3. mass mi . and an Eulerian part for updating the local gas density and velocity. Nos. 2005 diameter.72 CHINA PARTICUOLOGY Vol. It should be noted here that the assumption Fi=F(φ). 3. There are. 3. In addition. however. this result highlights the usefulness of microscopic simulation methods. Such type of simulation is in principle suitable for any type of interaction force between the particles. we find excellent agreement between our data and theory. which is currently used in literature can lead to differences with the LBM simulation data up to a factor of 5. we will briefly discuss some of the basic elements here. The discrete particle model consists of two parts: a Lagrangian part for updating the positions and velocities of the solid particles. etc. Discrete Particle Model (DPM) The discrete particle model is one level higher in the multi-scale hierarchy. however it is not suitable for dense-packed systems.5 φ ] . which represents the momentum exchange during collisions. in principle. because the experimental determination of the individual effective drag force in a dense assembly would be extremely difficult. is plotted as a function of the correction factor (1 − φ )y i + φ y i2 that we derived. the new position ri (t + dt ) and velocity v i (t + dt ) are calculated from the values at time t. In an event-driven simulation. or systems with long-range forces. The most important difference with the lattice Boltzmann model is that in this model the size of the particles is smaller than the grid size that is used to solve the equations of motion of the gas phase. In the Eulerian part of the code. A second (technical) difference with the LB model is that the evolution of the gas phase now follows from a finite difference scheme of the Navier-Stokes equation.

this function is of great relevance for the validity of the higher scale two-fluid model (see next section) derived from the kinetic theory. (2002). Fig. However. 4 we show the result for a fluidized bed of ideal (i. In the most detailed model of description. Upper graph: ideal particles. not all impact angles occur with the same frequency. i. fy(Cy). compared to a Gaussian/Maxwellian distribution. it is obvious that DPM simulations of industrial size fluidized beds are beyond the capability of commercially available computer facilities within the foreseeable future. Since both the CPU time and the required memory scales linear with the number of particles. 4 reveals a clear 73 anisotropy of the distribution in case of non-ideal particles. as described in the previous section. Figure 4 shows that for both ideal and non-ideal particles. V i 1− ε Equations (11) and (12) are solved with a semi-implicit method for pressure linked equations (SIMPLE-algorithm). (1996) showed that the heterogeneous flow structures in dense gas-fluidized beds are partly due to the collisional energy dissipation.y. This so-called two-fluid continuum . The system contained 25 000 particles of 2. the reaction from drag and pressure forces on the solid particles is included in the momentum equation for the gas phase via a source term S : 1 Vβ (13) S = ∫ ∑ i ( u − v i ) δ ( r − ri ) dV . van Sint Annaland & Kuipers: Computational Fluid Dynamics for Dense Gas-Solid Fluidized Beds coefficient of shear viscosity. More recently.e. 2001b) were used in their DPM simulations. rough and inelastic) particles. A DPM simulation study by Hoomans et al. Two-Fluid Model (TFM) The maximum number of particles that can be simulated with the DP model. 1993).z. where the gas velocity is set to 1. a different type of model is used for simulations at larger scales. distinguishable particles is abandoned. 4. Fig. Note that there is a full two-way coupling with the Lagrangian part. whereas the number of particles that are present in an industrial size fluidized bed can be two to three orders of magnitude higher. where it is assumed that the velocity distribution is both isotropic and nearly Gaussian. Analysis (Jenkins & Savage. fz(Cz) and f(C). We have studied the velocity distribution for two cases: in Fig.5 times the minimum fluidization velocity. The strength of the DP model is that it allows to study the effect of the particle-particle interactions on the fluidization behavior. (2004a. 1983) of the normal and tangential component of the impact velocity indeed showed that. with a time step that is in general an order of magnitude larger than the time step used to update the particle positions and velocities.e. yet. the probability of finding a particle with a velocity component vα with α=x.e.5 mm diameter. 2004b) studied the effect of the closures for gas-particle drag on the bubble-induced mixing in a pseudo 2D gas-fluidized bed and found that the best agreement between theory and experiment was obtained in case the LBM-generated drag closures reported by Hill et al. the interparticle contact forces includes normal and tangential repulsive forces (modeled by linear springs). is typically less than a million. perfectly smooth and elastic) and non-ideal (i.van der Hoef. since it is relatively straightforward to measure the velocity distribution as all particle velocities are known at any moment in time. The discrete particle simulations are an ideal tool for testing this assumption. lower graph: non-ideal particles. where the concept of a solid phase consisting of individual. (2001a.. Therefore. i. Details of the sampling procedure for obtaining the velocity distributions can be found in Goldschmidt et al. 4 DPM simulation data for the normalized particle velocity distribution fx(Cx). One of the great advantages of discrete particle simulations is that it allows to study properties of the system that are very diificult to obtain via experimentation.e. and tangential friction forces (Walton. which may disturb the spatial homogeneity and thereby causing collisional anisotropy. A particularly important example is the velocity distribution of the particles. It would be extremely difficult to obtain reliable estimates for the velocity distribution from experiments. Bokkers et al. A possible explanation is the formation of dense particle clusters in the case of inelastic collisions. the velocity distributions do not deviate significantly from a Gaussian and Maxwellian distribution. Li and Kuipers (2003) demonstrated that such flow structures are also strongly influenced by the degree of non-linearity of the particle drag with respect to the gas fraction ε. and dissipative forces (modeled by “dash pots”). in dense gas-fluidized beds.

In this work.000 0 s (16) Obviously.015 m×1. Jet velocity: 10.000 0 s Fig. defined as: 1 (17) θ = Cp ⋅ Cp . ∂t (15) with ρs .000 0 s t =3. 3.0 m (w × d × h)) operated with a central jet (diameter 0. 5 we show the simulated bubble formation for a pseudo two-dimensional (2D) bed (bed geometry: 0. 2005 where Cp represents the particle fluctuation velocity and the brackets indicate ensemble-averaging. and volume fraction of the solid phase. the discrete particle part (10) is now replaced by a set of continuum equations of the same form as (11) and (12): ∂ (14) (ε s ρs ) + ∇ ⋅ ε s ρsv = 0 . velocity.000 0 s t =9. ∂t ∂ (ε s ρsv ) + ∇ ⋅ ε s ρsvv = −ε s∇p − ∇ps − ∇ ⋅ ε sτ s + S + ε s ρsg . the numerical scheme for updating the solid phase is now completely analogous to (and synchronous with) that of the gas phase. model (TFM) describes both the gas phase and the solids phase as fully inter-penetrating continua. Nos. The bed contains ballotini with a particle diameter and density of 500 µm and ρs =2 660 kg⋅m-3 respectively. the effect of particle-particle interactions can only be included indirectly. In Fig. (1996) have been used for the particle phase rheology. Since the concept of particles has disappeared completely in such a modeling. Clearly. for the solid phase. energy flux). Bed material: ballotini with dp=500 µm and ρs=2 660 kg⋅m-3. . and can be found in the books by Chapman and Cowling (1970) and Gidaspow (1994) and the papers by Jenkins and Savage (1983) and Ding and Gidaspow (1990). t =4. ( ) with qs the kinetic energy flux. granular temperature). The derivation of such constitutive equations follows from the KTGF. 1992. the constitutive equations developed by Nieuwland et al. namely S = β (u − v ) . the solids pressure and solids shear rate) as a function of the granular temperature for a monodisperse particle system. Gidaspow. respectively.000 0 s t =5. The time evolution of the granular temperature itself is given by: 3∂ (ε s ρsθ ) + ∇ ⋅ (ε s ρsθ v ) =  2  ∂t   (18) − ps I + ε sτ s : ∇v − ∇ ⋅ ( ε sqs ) − 3 βθ − γ . in order to get a closed set of equations.000 0 s t =6. A description which allows for a slightly more detailed description of particle-particle interactions follows from the kinetic theory of granular flow (KTGF). 1-2. and γ the dissipation of kinetic energy due to inelastic particle collisions.e.000 0 s t =2. which must be expressed in terms of the basic hydrodynamic variables (density. That is. It can also clearly be seen that bubble coalescence occurs leading to a rapid increase in the bubble size. 1994). a very complex bubble pattern results from the jet operation where the size and the shape of the formed bubbles continuously change.000 0 s t =7. 5 Computed bubble formation in a pseudo 2D gas-fluidized bed with a central jet.000 0 s t =8. velocity.0 m⋅s-1 (40umf).015 m) at a velocity of 40 times the incipient fluidization velocity.74 CHINA PARTICUOLOGY Vol. such theory expresses the diagonal and off-diagonal elements of the solids stress tensor (i. stress tensor. using a set of generalized Navier-Stokes equations (Kuipers et al.57 m×0. 3 t =1. In this description the source term S is slightly different from (13).. In equations (14)−(18) there are still a number of unknown quantities (pressure. v and ε s = 1 − ε the local density. via an effective solids pressure and effective solids viscosity.000 0 s t =10. the time evolution of the gas phase is still governed by (11) and (12).

only with a modified drag coefficient Cd. 6 snapshots are shown of the bubbles that rise in the fluidized bed with a superficial velocity of 0. the correlations for the drag force on a single sphere are used. db=0.0 m×3.1 m⋅s-1 and 0. and the time average vector plots of the emulsion phase (right) after 100 s of simulation. the DBM strongly resembles the discrete particle model as outlined in section 3. Note that granular systems have no surface tension. two preliminary calculations have been performed for industrial scale gas-phase polymerization reactors. In this paper. in which we want to demonstrate the effect of the superficial gas velocities. from the mutual interaction of bubbles and from wall effects (Kobayashi et al. (20) ∂t whereas the discrete bubbles are tracked individually according to Newton’s second law of motion: dv (21) mb b = Ftot .0 m (w × h × d). However. however. as observed in the DPM and TFM models of gas-fluidized beds. it is possible to introduce yet another upscaling by considering the bubbles as discrete entities. Coalescence. top: u0=0. For the drag force on a single bubble (diameter db). the gas will be free to flow through a bubble in the gas-solid systems. db=0. like the liquid in a gas-liquid bubble column. The idea to apply this model to describe the large scale solids circulation that prevails in gas-solid reactors is new. namely the lift force FL and the virtual mass force FVM .3 m⋅s-1. In Fig. 6 in addition timeaveraged plots are shown of the emulsion velocity after (22) As in the DPM model.. The bubbles were injected via 49 nozzles positioned equally distributed in a square in the middle of the column. An advantage of this approach to model large scale fluidized bed reactors is that Fig.04 m. the number of bubbles in this case might be too large. 2000).0 Pa⋅s.1 m⋅s-1 and 0. the total force on the bubble has contributions from gravity ( Fg ). there are two forces present which are not found in the DPM. In this respect. so in that respect there is a pronounced difference with the bubbles present in gas-liquid bubble columns. since all the bubbles are tracked individually. which is an highly prevalent phenomenon in fluidized beds. which has been successfully applied in the field of gas-liquid bubble columns (Delnoij et al. and the tendency of rising bubbles to be drawn towards the center of the bed. Note that in (21). 1997). the behaviour of bubbles in fluidized beds can be readily incorporated in the force balance of the bubbles. The geometry of the fluidized bed was 1. The density of the bubble phase was 25 kg⋅m-3. With the DBM.5. 6 Snapshots of the bubble configurations (left) computed from the DBM model without coalescence.04 m. we want to show some first results of the discrete bubble model applied to gas-solid systems. The emulsion phase has a density of 400 kg⋅m-3 and the apparent viscosity was set to 1. can also be easily included in the DBM. For instance. which involves some slight modifications of the equivalent model for gas-liquid systems. since coalescence has not been taken into account in these simulations.0 m×1.3 m⋅s-1.3 m⋅s-1. To this end the emulsion phase is modeled as a continuum. ∂t ∂ (ερ u ) + ∇ ⋅ ερ uu = −ε∇p − ∇ ⋅ ετ − S + ερ g . However. whereas the virtual mass force coefficient is set to 0. one can think of the rise velocity. since it is also of the Euler-Lagrange type with the emulsion phase described by the volume-averaged Navier-Stokes equations: ∂ (19) (ερ ) + ∇ ⋅ ερ u = 0 .1 m⋅s-1. such that it yields the Davies-Taylor relation v br = 0.711 gdb for the rise velocity of a single bubble. . pressure gradients ( Fp ) and drag from the interaction with emulsion phase ( Fd ).. set to 0. It is clearly shown that the bubble hold-up is much larger with a superficial velocity of 0.3 m⋅s-1. dt where Ftot is the sum of different forces acting on a single bubble: Ftot = Fg + Fd + Fp + FL + FVM . As far as the numerical part is concerned. The lift force is neglected in this application.van der Hoef. Discrete Bubble Model (DBM) Although the two fluid model can simulate fluidized beds at life-size scales. In Fig. which is not the case for gas-liquid systems. height 16 meters) are still beyond its capabilities. respectively. the largest scale industrial fluidized bed reactors (diameter 5 meters. This is the so-called discrete bubble model. van Sint Annaland & Kuipers: Computational Fluid Dynamics for Dense Gas-Solid Fluidized Beds 75 5. bottom: u0=0. and the larger bubbles are treated as discrete bubbles.

The model can be augmented with energy balances to study temperature profiles in combination with the large circulation patterns. 2002). (1970). 2005) revealed significant deviations with the ad hoc modifications of the monodisperse drag force relations. G. It was found that that dissipative interaction forces cause an anisotropy in the distribution. like empirical relations for bubble rise velocities and the interaction between bubbles. TFM and DBM simulations. D. J. M. J. A. 140. The key idea is that the methods at the smaller. G. rather ad hoc modifications are used to allow for polydispersity. In particular. upfow in the middle. more detailed scale can provide qualitative and quantitative information which can be used in the higher scale models. J. and downflow along the wall. Enschede. At present..187-194). References Bokkers. E. we want to focus on developing drag force relations for systems which deviate from the ideal conditions.. Chem. Yet. these relations are still far from optimal. the discrete bubble model applied to gas-solid systems seems to be a promising new approach for describing the large scale motion in life-size chemical reactors. Very preliminary results  obtained from data of a high velocity impact  were in reasonable agreement with the experimental values for the viscosity. Eng. M. & Kuipers. Ding. Nos. Proceedings of Fluidization XI (pp. M. 100 s of simulation. A. & Cowling. PhD thesis. which can be obtained from the lower scale simulations.g.. furthermore. In particular. Italy. Kuipers. & van Swaaij. Comparison of continuum models using the kinetic theory of granular flow with discrete particle models and experiments: extent of particle mixing induced by bubbles. as they are used in the two fluid model. and Albert Bokkers for performing the DPM. it would be of great interest to find improved expressions for the solid pressure and the solid viscosity. Gidaspow. M. T. van Sint Annaland. which is caused by dissipative forces. Finally. the TFM and DPM simulations will be used to guide the formulation of additional rules to properly describe the coalescence of bubbles. University of Twente. The viscosity of the bed follows then directly from the Stokes-Einstein formula for the drag force. Mixing and segregation in a bi-dispersed gas-solid fluidized bed: a numerical and experimental study. are clearly demonstrated. homogeneous arrays of mono-disperse spheres. and the effect of the superficial gas velocity. (1997). (1990). 3623. (2004a).. these models still lack the capability of describing quantitatively particle mixing and segregation rates in multi-disperse fluidized beds. at the DPM/TFM level these relations are applied to systems which are. It would be interesting to include the effect of anisotropy at the level of the TFM. 1994). Another example is the functional form of the velocity distribution. such as the aforementioned heterogeneous flow structures. unbounded. only qualitative information from the DPM simulations is obtained. in which it is assumed that the drag force scales linearly with the particle diameter. Our lattice Boltzmann results for the drag force in binary systems (van der Hoef et al. A bubbling fluidization model . 2004. We should note here that although the new drag force relations seem to give results at the DPM/TFM level which compare better with the experimental findings.. 2005 granular flows. In particular. 1-2. A. although the functional form remains close to Gaussian for all three directions (Goldschmidt et al. is the quantitative information that the smaller scale models can provide. Essential for this model to be successful is that reliable information with regard to rise velocities and mutual interaction of the bubbles is incorporated. 52. even locally. indirect method for obtaining the viscosity is to monitor the decay of the velocity of a large spherical intruder in the fluidized bed. Cambridge: Cambridge University Press. Kuipers et al. P. A typical example of such qualitative information is the insight (from the DPM simulations) that inelastic collisions and nonlinear drag can lead to heterogeneous flow structures. Even more important. Naples. Future work will be focused on implementation of closure equations in the force balance. however. which finds its direct use in both the DPM and TFM simulations. Bokkers. A very simple. J. Summary and Outlook In this paper we have presented an overview of the multi-scale methods that we use to study gas-solid fluidized beds. G. M. Delnoij. van Sint Annaland.. This will be the subject of future research. Powder Technol. A. it should be borne in mind that these drag force relations are derived for static. moving particles: granular temperature) or not so trivial (inhomogeneities). We would like to thank him for making his code available. Computational fluid dynamics applied to gas-liquid contactors. where the parameters which would quantify such deviation may be trivial to define (polydispersity: width of the size distribution. inhomogeneous and non-static. 6. The Netherlands. S. The Mathematical Theory of Non-uniform Gases (Trial mode edition). May 9-14. Sci. for instance along the lines of the kinetic theory developed by Jenkins and Richman (1988) for shearing Acknowledgment The lattice Boltzmann simulations have been performed with the SUSP3D code developed by Anthony Ladd. Although the continuum models have been studied extensively in the literature (e. A.. More elaborate simulations of these systems are currently underway. & Kuipers. The large convection patterns. 176-186.. W. An important improvement in the modeling of life-size fluidized beds could be made if direct quantitative information from the discrete particle simulations could find its way in the continuum models. it is a non-trivial task to extract direct data on the solid viscosity and pressure in a DPM simulation. (2004b).76 CHINA PARTICUOLOGY Vol. however. 3. In future work. M. A typical example of this is the drag force relation obtained from the LBM simulations. 1992. which is at present not incorporated in the model. Chapman. & Gidaspow.

(1988). van Beckum.. 99. D. M. Ser. Plane simple shear flow of smooth inelastic circular disks: the anisotropy of the second moment in the dilute and dense limits. B. Hoomans. Hill. Koch. London: Academic Press. A theory for the rapid flow of identical. & Nauman. Fluid Mech. M. 51. 61. B. Kuipers. Fluid Mech. 2005. Hydrodynamic modelling of dense gas-fluidised beds: comparison of the kinetic theory of granular flow with 3-D hardsphere discrete particle simulations. M. P. R. P. Rev.. McNamara. M.. Mechanics of fluidization. Ladd. AIChE J. J. Stat.). M. van Sint Annaland. Eng. Jenkins. 57... (1983).... (1993). W. V. & Yu. Chem. 243. & van Swaaij. M. Sci. & Verberg. Lattice Boltzmann simulations of low Reynolds number flow past mono. R. (1994). 448. A. 42. H. 528. (1935). van der Hoef. R. Kuipers. Y. J. USA. J. H.. T. J. 187. (Ed. Powder Technol. S. O. Moderate-Reynolds-number flows in ordered and random arrays of spheres. J. M. J. M. Chem. & Kuipers. 58. 192. & Clift. R. 2005 and accepted April 1. Multiphase Flow and Fluidization: Continuum and Kinetic Theory Descriptions. J. Butterworth Heinemann series in Chemical Engineering.. & Richman. Sci. & Mori. L.. V. (1998). nearly elastic particles. Durango. R. J.. Lett. W. Fluid Mech. Hasslacher. Goldschmidt. A. Phys. A drag coefficient correlation. J. M. AIChE J. J. (1996). (1986). J. Li. Particulate Two-Phase Flow. W. Kuipers.. Fluid Mech. Manuscript received January 10. 36. Lattice-Boltzmann simulations of particle fluid suspensions. Adv. 1505. (2003). Beetstra.and bidisperse arrays of spheres: results for the permeability and drag force.. A. L. Rev.. Gas-particle interactions in dense gas-fluidized beds.. Boston: Academic Press.. W. London. AIChE Symp. G. O. 77. London. J.van der Hoef. A. 24. 1913.. Chem. (1992). D. Kuipers... (2000). Computational fluid dynamics applied to chemical reaction engineering. 130. C. Ergun. J. Gidaspow. & Ladd. M. (1991). J. P. J. Hydrodynamic modeling of gas-fluidized beds and their 77 role for design and operation of fluidized bed chemical reactors. J. A. A. B. J. M. Kobayashi. Frisch. F. J. Hydrodynamic modeling of gas/particle flow in riser reactors. A. & Pomeau. van Swaaij. B. W. Schiller. P. 2332. (1990). Nieuwland. C. 213.. R. C.I. 227. M. Y. P. W. & Levenspiel. D. 62. & Zanetti. J. (1952). . & Savage. 313. A. A. R. J. S. & Ladd. Oxford: Oxford Science Publications. In Roco. 1191. T. 711. Koch. M. F. Proceedings of the Fluidization IX conference (pp. Chemical reactors. A. The Lattice Boltzmann Equation for Fluid dynamics and Beyond. Fluid flow through packed columns. 48. Chem. smooth. (2005). Jenkins. 523. S.).. Walton. B. Beetstra. 318. P. & van Swaaij. C. Discrete particle simulation of bubble and slug formation in a two-dimensional gas-fluidized bed: a hard sphere approach. B. 56. van Duin. A numerical model of gas-fluidized beds.. (2001b). The first effects of fluid inertia on flow in ordered and random arrays of spheres. Fluid Mech. J. (2001). 233-253. J. 113. frictional particle-particle interactions. (2001a). In Davidson.. S. Y. & Kuipers. J.. 2059. 448. R. G. (1988). M. & van Swaaij. K. M. & van Swaaij. (1966). Eng. J.. J. Numerical simulation of inelastic. A.D. van Sint Annaland & Kuipers: Computational Fluid Dynamics for Dense Gas-Solid Fluidized Beds using kinetic theory of granular flow. (2002). & Kuipers. M. A study on the behavior of bubbles and soldis in bubbling fluidized beds. M. Hill. Eng. (1996). L. P. Zeitung. 47. M. A.. 100. Chem. 89. J. Phys. Chem. W. W. Lattice gas automata for the Navier-Stokes equation. Eng. J. U. (Eds. Lett. Sci. C. Fluidization. 104.15-30). N. Butterworth Heinemann series in Chemical Engineering.. (1998). Eng. Fluidization Engineering. Wen. Hoomans. P. Succi. M. 1569. P. J.. A. Process. D. Kunii.. (2001). R. Eng. 327. Yamazaki. Phys. Use of the Boltzmann equation to simulate lattice-gas automata. Kuipers. Briels. Sci. & van Swaaij..

Sign up to vote on this title
UsefulNot useful