MATERIALS

SCIENCE

APPLYING THE LATTICE BOLTZMANN

EQUATION TO MULTISCALE FLUID
PROBLEMS
The authors discuss the theory and application of the Lattice Boltzmann equation to
multiscale physics in fluids. They present two examples relevant to real-life applications:
airflow around an airfoil at high Reynolds numbers and reactive flow in micropores.

I
n the quest to model complex physical to a single-scale regime. This approach’s central
phenomena, overstating the need for simu- issue is dealing with the “hand-shaking” regions:
lational tools that can handle multiple space At these regions, the different schemes that han-
and time scales is hard. The capability of dle individual scales need to exchange informa-
addressing problems across several length and tion in a way that is physically meaningful, math-
time scales is a hallmark of modern computa- ematically consistent, and computationally
tional science, the goal of which is to tackle is- efficient. Two-level schemes combing atomistic
sues that straddle various traditional disciplines and continuum methods for crack propagation
of science and engineering. Some interesting ex- in solids, or strong shock fronts in rarefied gases,
amples of such phenomena that have received made their appearance in the early 1990s.1–3
considerable attention recently are drug design, More recently, researchers have applied three-
the brittle or ductile failure of structural mate- level schemes to study crack dynamics, combin-
rials, heterogeneous catalysis, and turbulent ing a finite-element (FE) treatment of contin-
combustion. We have come to know computa- uum mechanics in regions far from the crack, a
tional schemes aimed at such complex applica- molecular dynamics (MD) treatment of atomic
tions that involve multiple levels of physical and motion near the crack, and a quantum mechan-
mathematical descriptions as multiscale methods. ical (QM) description of bonding in the crack
One approach to multiscale modeling meth- tip’s immediate neighborhood.4 This FE-MD-
ods is to integrate schemes, which usually apply QM implementation represents a concrete ex-
ample of composite algorithms—that is, meth-
ods that involve seamless interfaces between the
1521-9615/01/$10.00 © 2001 IEEE different mathematical models associated with
different physical levels of description.
SAURO SUCCI An alternative approach is to explore methods
National Research Council, Italy that can host more than one physical level of de-
OLGA FILIPPOVA scription—for example, atomistic, kinetic, and
Duisburg University, Germany
fluid—within the same mathematical frame-
work. A potential candidate is the Lattice Boltz-
GREG SMITH AND EFTHIMIOS KAXIRAS
mann equation method. The LBE is a minimal
Harvard University form of the Boltzmann kinetic equation in

26 COMPUTING IN SCIENCE & ENGINEERING

which all details of molecular motion are re-
moved except those that are strictly needed to
recover hydrodynamic behavior at the macro-
scopic scale—mass, momentum, and energy
conservation.5–11 The result is an elegant and
simple equation for a set of discrete distributions
fi(x, t) describing the probability of finding a par-
ticle at lattice site x at time t with speed vi = ci.
These discrete distributions relate to the stan-
dard one-body Boltzmann distribution function
of continuum kinetic theory through the fol-
lowing expression:

f ( x, v, t ) = ∑i f i ( x, t )δ ( v − c i ). (1)

The LBE combines the power of continuum Figure 1. The set of discrete speeds in a 3D lattice.
methods with the geometrical flexibility of atom- The set consists of one particle with zero speed
istic methods: It propagates smooth information, represented by the cube center, six particles with
the discrete populations f(x, ci, t), along simple speed unity represented by the arrows pointing
rectilinear trajectories prescribed by the lattice to the centers of the cube faces, and 12 particles
directions. The LBE could potentially play a with speed √2 represented by the arrows pointing
twofold function—as a telescope for the atom- to the cube-edge centers.
istic scale and a microscope for the macroscopic
scale. These aspects are useful and continue to
expand vigorously, even 10 years after the basic conditions consists of one particle with speed
theory’s inception, in both academic12 and com- zero, six particles with speed unity (connecting a
mercial environments.13 However, as we tackle cube’s center to its nearest-neighbor face cen-
multidisciplinary problems of growing com- ters), and 12 particles with speed √2 (connect-
plexity, it’s increasingly evident that we need to ing the cube’s center to its edge centers, as in
upgrade the method to extend the range of scales Figure 1).
the LBE can access and to couple it downward In Equation 2, the left-hand side represents a
or upward with micro- and macroscopic meth- forward discretization of the Boltzmann free-
ods. This article offers a cursory view of efforts streaming operator
in both directions, but let’s first review the LBE
∂
theory’s salient features. Dt ≡ + v ⋅∇ (3)
∂t
along the discrete particle trajectories dxi = vidt.
Basic LBE theory The right-hand side represents intermolecular
This is the simplest and by now most popular collisions in terms of an effective relaxation to a
form of the LBE (often referred to as the lattice local equilibrium distribution fieq in a time in-
BGK14,15 for Bhatnagar, Gross, and Krook16): terval of the order ω–1. This local equilibrium is
usually taken in the form of a quadratic expan-
fi ( x + δt c i , t + δt ) − fi ( x, t )
sion of a Maxwellian
[
= −ωδ t f i ( x , t ) − f ie ( x , t ) ] (2)
eq  u⋅c uu ⋅ ( c i c i − c 2s I ) 
fi = ρw i 1 + 2 i + , (4)
where fi(x, t) ≡ f(x, v = ci, t), i = 1,n is the proba-  cs 2c 4s 
bility of finding a particle at lattice site x at time
t, moving along the lattice direction defined by where ρ is the fluid density, u the flow speed, and
the discrete speed ci, and δt is the time unit. To I the identity tensor; cs is the lattice sound speed,
recover faithful fluid dynamics, we must choose defined by the condition
the set of discrete speeds so that it guarantees
mass, momentum, and energy conservation as c 2s I = ∑i wi c i c i (5)
well as rotational invariance. Only a limited sub-
class of lattices satisfies these conditions. For ex- and wi a set of directional weights normalized to
ample, in 3D space, the lattice that satisfies these unity.

NOVEMBER/DECEMBER 2001 27
 Local equilibria must obey the following con- skog multiscale expansion. Fortunately, we can
servation relations (for simplicity, we take the liken such a purely numerical effect to a nega-
mass to be unity): tive viscosity. In particular, by choosing ω near
and smaller than ω = 2/δt, we can achieve small
∑ f ieq = ρ (6) viscosities, a basic prerequisite to simulating
i highly turbulent flows.
Historically, the LBE was derived as the one-
∑i f ieq c i = ρu (7) body kinetic equation resulting from many-body
lattice gas automata.17 Here, we wish to moti-
∑i f ieq c i c i = ρ[uu + c 2s I ] (8)
vate the LBE’s physical meaning without delv-
ing into the details of its derivation. In doing so,
we must invoke the values of certain dimension-
In the limit of long wavelengths, where a par- less numbers that are typically employed in fluid
ticle’s mean-free-path sets the scale, the fluid dynamics. Table 1 defines all the dimensionless
density and speed obey the Navier-Stokes equa- numbers mentioned in this article in terms of
tions for a quasi-incompressible fluid: physical quantities relevant to fluid flow: lm is the
particle mean-free path, lc is the flow’s charac-
∂ρ teristic length, δx is the grid spacing (shortest
+ ∇ ⋅ ρu = 0 (9)
∂t fluid-dynamic length scale), u is the fluid veloc-
∂ρu ity, cs is the speed of sound in the fluid, ν is the
+ ∇ ⋅ ρuu kinematic viscosity, D is the diffusivity of the
∂t
= −∇P + ∇ ⋅ [ µ ( ∇u ) S + λ ( ∇ ⋅ u ) I ], (10) molecular species of interest, and τc is a charac-
teristic time scale for chemical reactions. L and
where P = ρcs2 is the fluid pressure, (∇u)S is the H are a tube’s length and height and both repre-
symmetrized strain tensor, µ = ρν is the dynamic sent macroscopic quantities.
viscosity, and λ is the bulk viscosity. For all prac- The dimensionless numbers relevant to deriv-
tical purposes, we can neglect the last term in- ing LBE are the Knudsen number Kn and the
volving the bulk viscosity when dealing with Mach number Ma. The major conceptual steps
quasi-incompressible fluids. According to the in first deriving the LBE were low-Kn expansion
definition of the pressure P, the LBE fluid obeys around local equilibria and low-Ma expansion of
an ideal equation of state, which is reasonable the local equilibria around global equilibria. The
because molecules, as represented by the LBE, low-Kn approximation corresponds to the stan-
have no potential energy. Using standard linear dard picture of fluid dynamics as the mean-field
transport theory, with careful handling of arti- approximation of kinetic theory in the limit of
facts the lattice introduces, the LBE fluid’s kine- weak departures from local thermodynamic
matic viscosity becomes equilibrium. The low-Ma expansion is specific
to the lattice framework. We can readily confirm
1 δ 
ν = c 2s  − t  . (11) that due to the finite number of discrete speeds,
ω 2 there is no way to fulfill the conservation con-
The appearance of the factor 1/ω in this expres- straints in Equation 4 with a Galilean-invariant
sion is not surprising, because it corresponds to local equilibrium of the form fieq = feq(vi – u) for
the relaxation time around local equilibria. In a generic value of the fluid speed u. However,
other words, cs /ω is the lattice fluid’s mean-free expanding around global equilibria (u = 0) and
path. The factor δt /2 is a genuine lattice effect, guaranteeing Galilean invariance term-by-term
which results from an underlying Chapman-En- in the expansion is possible. Because the Navier-

Table 1. Definitions of dimensionless numbers mentioned in this article.

Name Symbol Definition Description

Knudsen Kn lm/δx Ratio of particle mean-free path to shortest fluid dynamic scale
Mach Ma u/cs Normalization of fluid velocity u by speed of sound cs in fluid
Reynolds Re ulc/ν Ratio of advective to dissipative effects
Peclet Pe uH/D Ratio of advective to diffusive time scales
Damkohler Da H2/Dτc Ratio of diffusive to chemical reaction time scales

28 COMPUTING IN SCIENCE & ENGINEERING

 Figure 2. (a) A locally embedded
grid in which the local
connectivity (in this example,
four), is unchanged, but the
lattice spacing is refined in the
central region—the refinement
shown is by a factor of four. (b)
An unstructured grid in which
the local connectivity as well as
the spacing are varied—in this
example, the connectivity
ranges from five in the outer
region to eight in the inner.

(a) (b)

Stokes equations are quadratic in the fluid speed, computational grid at the expense of significantly
a second-order expansion sufficiently fulfills this more complex data structures. Another popular
requirement but restricts the LBE approach’s va- option comes from locally embedded grids—reg-
lidity to low-Ma, quasi-incompressible flows. ular grids in which the local connectivity is un-
Researchers have developed modern variants changed but the lattice spacing is refined or
dealing with high-Ma flows in recent years, but coarsened locally, typically in steps of two for
we aren’t concerned with that in this article, nor practical purposes. Figure 2 shows examples of
will we discuss flows with thermal effects. unstructured and locally embedded grids.
A few years after its introduction, researchers Local embedding is a specific instance of a
realized that they could obtain the LBE mathe- more general framework known as multiscale (or
matically by standard projection of the contin- multigrid) algorithms. Other work has proposed
uum BGK equation on Hermite polynomials multigrid LBE schemes,19 and researchers have
and subsequent numerical evaluation of the ki- subsequently tested and validated for moderate
netic moments by Gaussian quadrature.18 This and high Reynolds number flows around cylin-
alternative route highlights the role of the dis- ders and blades.20,21 Keep in mind that people
crete speeds ci as Gaussian knots and shows that used LBE schemes that had finer resolution in
Gaussian integration works as an effective im- regions of larger gradients in commercial soft-
portance-sampling device of velocity space. This ware for years,13 and we have examples of sev-
property is indeed the key for capturing the eral LBE schemes working on single-grid
complexities of hydrodynamic flows by means of nonuniform meshes.22–25 In what follows, how-
only a handful of discrete speeds. ever, we concentrate on LBE schemes with lo-
cal embedding that use a specific algorithm, pre-
sented elsewhere.19–21
Multigrid Lattice Boltzmann High-Re flows present a major challenge to
simulations computational physics mainly because of fluid
Much of the LBE method’s simplicity stems turbulence. The range of active scales in a tur-
from the fact that within its formulation, discrete bulent flow goes from the macroscopic length L
physical particles move and interact on a regular down to the so-called Kolmogorov dissipative
uniform lattice. Many phenomena of fundamen- length ld, at which point dissipation takes over
tal and practical interest—such as shocks, bound- and turbulence dies out. We can estimate the
ary layers, and thin interfaces—involve wide Kolmogorov length as ld ~ L [Re]–3/4, and because
ranges of time and length scales. An accurate de- values of Re around 106 are commonplace in
scription of these phenomena on a mesh de- practical applications (for instance, a car moving
mands highly clustered grids that are locally at 100 Km/h), it is immediately evident that a
adapted to the problem’s physics. Unstructured fully resolved numerical simulation of flow past
grids—that is, grids in which the lattice coordi- a car involves four to five decades in scale,
nation number might change from place to namely 1012 – 1015 degrees of freedom. This is
place—provide a powerful response to this need. beyond the capabilities of any foreseeable elec-
They allow much stronger distortions of the tronic computer probably for decades to come

NOVEMBER/DECEMBER 2001 29
 1. Move and collide F on coarse grid
2. For k = 0, ..., n – 1: δ xc  1 δt 
ν=  − . (12)
2.1. Interpolate F (Equation 18) on interface coarse-to-fine grid 3δ t  ω 2
2.2. Scale F to f (Equation 17) on interface coarse-to-fine grid
2.3. Using scaled interpolated values move and collide f To achieve the same viscosity on both coarse
3. Scale back f to F on interface fine-to-coarse grid. and fine grids, we must rescale the relaxation
parameter in the LBGK scheme as
Figure 3. The algorithm for updating the populations F and f on the −1
coarse and fine grids through Equations 17 and 18. δ  1 δt  
ω f =  t + n −  , (13)
 2 ωc 2  

and explains why, for most practical purposes, where ωf and ωc are relaxation parameters on the
we must resort to turbulence models based on fine and coarse grids, respectively, and n is the
closure approximations for the effects of unre- refinement parameter. We can split the discrete
solved scales on the resolved ones. distribution function into equilibrium and non-
Regardless of whether we use turbulence equilibrium components:
models or fully resolved simulations, the need
fi = fieq + fineq. (14)
for selective clustering of computational degrees
of freedom in the computational domain’s hot Using the LBGK equation, we can obtain the
spots remains, hence the motivation for multi- nonequilibrium component:
grid simulations. The starting point for the
1∂  eq
multigrid LBE theory is the LBGK formulation fi
neq
=− 
ω  ∂t 
( )
+ c i ⋅∇  f i + O [ Kn ]2 , (15)
of fluid dynamics.14 Dividing the lattice spacing
by a refinement factor n refines the grid. The which is accurate to second-order [Kn] = δx/lc.
kinematic viscosity, defined in the LBGK It is important to realize that the scale lc can
model’s frame, depends on the lattice spacing δx: change significantly from place to place in the
flow. The task of the multigrid method is pre-
cisely to adapt the grid resolution to this change
1.2e–02 3.7e–02 6.1e–02 8.6e–02 1.1e–01 of scales. In the low-frequency limit δt/τ ~ [Kn]2,
where τ is a typical hydrodynamic time scale,19
1.0e–05 2.4e–02 4.9e–02 7.3e–02 9.8e–02 the distribution function’s nonequilibrium com-
160
ponent simplifies to

140
fi
neq 1
=−
ω i
eq
(
c ⋅∇f i + O [ Kn ]2 . ) (16)

Combining this relation with the continuity of

120 the hydrodynamic variables—and of their deriv-
atives at the interface between the two grids—
100 leads to this relation between the coarse and fine
grid populations:

( )
80
f i′ = F˜i + F˜i′− F˜i Ω −1
eq eq
(17)

+ ( f ′− f )Ω,
60 eq eq
Fi′ = f i i i (18)
40 where capital symbols stand for coarse grid,
primed symbols stand for post-collision, and
20 symbols with a tilde stand for coarse-to-fine grid
interpolation. The variable Ω is defined as
50 100 150
(1 − δ tω c )ω f
Ω=n . (19)
r (1 − δ tω f )ω c
Figure 4. Modulus of the mean velocity u of the flow around the
NACA 4412 airfoil at Re = 1.5 × 106 and angle of attack α = 13.87°. The relations of Equations 17 and 18 define a
This calculation involves grid refinement between coarse and fine mapping between coarse- and fine-grid distrib-
grids (the region for the latter indicated by the black rectangle) of utions that clearly reduces to the identity in the
n = 5. Qref denotes the point of reference for the pressure and velocity. limit n → 1. We can cast these formal expres-

30 COMPUTING IN SCIENCE & ENGINEERING

 r
sions into an algorithm (see Figure 3): the algo- Figure 4 shows the velocity magnitude u and
rithm’s flow is straightforward, with the last step streamlines in the whole computational domain
implementing the correction of the coarse-grid for the case n = 5. Figure 5 shows the surface
population with dynamic effects derived from scaled-pressure distribution, defined as
the fine-scale grid. For steady-state flows, we can
accelerate this algorithm by using only one sub- P − Pref
CP = 2 , (20)
cycle (n = 1) in step 2.26 r 2
ρ u ref
We recently coupled the multigrid LBE
method with a finite-difference treatment of r
model equations for turbulence in relatively com- where Pref, uref are values at a chosen reference
plex geometries such as high-Re flows past air- point, as obtained numerically with the LBE
foils and compressor blades. Our goal was to de- code (for n = 5 and n = 7), as well as from a dif-
velop a robust treatment of the boundary ferent simulation28 and from experiment.29 Be-
conditions between the turbulent fluid and the sides a general agreement in the far-field region,
wall.27 We used this coupled approach to study we see a certain discrepancy between the nu-
the flow around the NACA 4412 airfoil profile merical and experimental results near the pres-
(a standard benchmark profile out of a suite pro- sure peak at the airfoil’s leading edge—the de-
vided by the National Advisory Committee for velopment of laminar and transitional local flow
Aeronautics) by using a boundary-fitting formu- regimes hampers a fully turbulent treatment of
lation for wall functions through which the as- the boundary. The correct treatment of such
ymptotic solution describing near-wall turbu- critical regions is a standing challenge for cur-
lence is matched with simulation data outside the rent computational aerodynamics.
boundary layer.21 To date, the LBE method has proven competi-
The flow geometry and parameters are defined tive with state-of-the-art unstructured adaptive
elsewhere.28,29 In our numerical simulations, we finite-volume formulations while retaining a sig-
took into account the confined wind tunnel’s nificantly simpler computational structure. Much
blockage effect.28 The numerical experiment is
as follows: The flow is injected at the inlet in x 2
direction with Ma = 0.07 and angle of attack α =
1
13.87° with respect to the inclined airfoil. We im-
pose slip conditions on the computational do- 0
main’s lower and upper boundaries. We extrap-
olate the pressure at these sections and at the –1
inlet along the normal to the sections from the
outer flow, while we assume that the pressure at –2
the outlet is constant and the velocity is extrapo-
CP

–3
lated along the normal from the outer flow. The
Reynolds number value Re = 1.5 × 106 is related –4
to the length of the chord lc, which, in the pre-
sent example, is given by lc = 50δxc, where δxc is –5
the coarse grid’s spacing. The trailing edge’s po-
–6
sition in respect to the lower-left corner of the
computational domain is 27.4391δxc, 73δxc. The –7
computational domain consists of 170 × 170
nodes. We apply grid refinement—defined by the –8
0.25 0.50 0.75 1
parameter n = δxc/δxf, where δxf is the fine-grid x/lc
spacing—to a box (66δxc, 25δxc) surrounding the
profile with the lower-left corner (22δxc,70δxc). Figure 5. Surface scaled pressure Cp distribution for NACA 4412
The values of the relaxation parameters on the airfoil at Re = 1.5 × 106 and angle of attack α = 13.870, as a function
coarse and fine grids are δtωc = 1.9, δtωc = 1.8, re- of scaled position x/lc, along the airfoil cord. The small magenta
sulting in the artificial numerical viscosity stabi- dots represent results from our runs with grid refinement between
lizing the solution of inviscid flow.18 The spac- coarse and fine grids n = 5, while the small blue dots correspond to
ing between the slip surface (where boundary grid refinement n = 7. The large white circles are numerical
conditions are applied), and the airfoil’s geomet- results,28 and the large blue squares are experimental data from
rical surface is uniform and equal to 0.0036 lc. other research.29

NOVEMBER/DECEMBER 2001 31
Table 2. Conversion efficiency η, defined in Equation 21, of the 2D reactive channel flow as a function of the Peclet (Pe)
and Damkohler (Da) numbers.

τc (chemistry) D (diffusion) Pe Da η
2 –1
10 (faster) 10 (faster) 50 250 0.47
102 (faster) 10–2 (slower) 500 2,500 0.18
103 (slower) 10–1 (faster) 50 25 0.09
103 (slower) 10–2 (slower) 500 250 0.07

work remains to be done to assess whether multi- centration of the chemical species obeys a con-
grid LBE simulations can take the lead in this im- vection-diffusion equation of the form
portant sector of computational aerodynamics.
∂C
+ u ⋅∇C = D∇ 2C + Rδ ( x − x w ), (22)
∂t
Coupling to macroscopic grids where C is the density of the reactive species
The LBE method can easily couple to finite- (pollutant) passively advected by the flow, D is
difference–volume–element methods for con- the pollutant’s molecular diffusivity, and R is a
tinuum partial differential equations partly be- catalytic reaction term that is only active at the
cause of its small time steps and geometrical solid wall boundary, labeled w.
flexibility. For example, take a millimetric flow In engineering practice, the fluid-wall chemi-
with, say, 100 grid points per side, yielding a cal interaction is expressed with a simple empir-
mesh spacing δx = 10 microns. Assuming a sound ical rate equation,
speed of the order of 300 m/s, we obtain a time
dC w C f − C w C w
step of order δt = 30 × 10–9 sec = 30 nsec. Such a = − , (23)
small time step permits the handling of relatively dt τw τc
fast reactions without having to resort to implicit where the subscripts w and f stand for the wall and
time stepping, thus avoiding the solution of large fluid cells that exchange mass, and τw and τc de-
systems of algebraic equations. The LBE’s flexi- note typical time scales for fluid-to-wall mass den-
bility is especially suited to heterogeneous catal- sity transfer and chemical reactions, respec-
ysis, because the simplicity of particle trajecto- tively—for simplicity, we take τw = τc. The relevant
ries permits the description of fairly irregular fluid scales are the advective transit time scale τA =
geometries and boundary conditions. L/u and the cross-channel diffusive time scale τD
Because of these two advantageous features, = H2/D, where H and L are the channel’s height
many researchers are using the LBE method to and length. These time scales define the major di-
simulate reactive flows over microscopically cor- mensionless parameters that control the conver-
rugated surfaces—an application of great interest sion efficiency—that is, the Pe and Da numbers,
for the design of chemical traps, catalytic con- which in this case take the form
verters, and related devices.30
uH L τD H 2 τD
The general problem of a fluid flow carrying Pe = = , Da = = . (24)
chemical species over a reactive surface is cen- D H τA Dτ c τc
tral to a variety of industrial and technological Because pollutant molecules must make it to
applications, such as the design of catalytic con- the wall to be chemically consumed, high chem-
verters for cars. An important quantity in these ical efficiencies of the device go with low-Pe and
systems is the chemical efficiency high-Da values. On the other hand, for fluids
moving at sizeable speeds—say, tens of meters
Φ out
η = 1− , (21) per second—molecular diffusivity is generally
Φ in too small to achieve sufficiently small Pe values.
where Φout and Φin are the chemical fluxes at the This motivates the use of corrugations to en-
inlet (where the chemical species is injected) and hance the fluid-to-wall mass transfer and in-
the outlet of the fluid domain, respectively. Here crease conversion efficiency. The corrugations
we consider 2D reactive flows to understand that can be manufactured into a channel are by
how the system’s efficiency is affected by the necessity of macroscopic dimensions. Their ef-
geometry, diffusivity, and surface reaction of the fect must be coupled to the microscopic scale at
chemical species at the catalytic wall. The con- which diffusion and chemical reactivity take

32 COMPUTING IN SCIENCE & ENGINEERING

place. We solve the fluid flow equations in the tion. For each set of parameter values, one simu-
Lattice Boltzmann form, whereas we handle the lation uses a smooth channel and the other a
reaction-diffusion equation for the chemical channel with a trapezoidal obstacle that is seven
species with a modified Lax-Wendroff finite-dif- grid units high: the longer (bottom) base of the
ference scheme on a simple nearest-neighbor trapezoid is between grid points 119 and 176, and
square lattice.31 the shorter (top) base is between grid points 125
The channel is 50 grid units high and 300 units and 170. In all flows, the Reynolds number is
long. With a lattice spacing of δx = 10 microns, this Re = uH/ν ≈ 500, the Peclet number takes the val-
gives a channel of dimensions 3 mm × 0.5 mm. We ues Pe = 50, 150, and 500, and the Damkohler
align the x-axis along the channel’s length and the number takes the values Da = 25, 250, and 2,500.
z-axis along its height. We choose initial condi- The simulations last 3 × 105 time steps, which
tions as follows: The flow is initialized at a uni- corresponds to about 10–2 seconds. This is longer
form density ρ = 1 and speed ux = u0, uy = 0, uz = 0, than the longest diffusive time scale, and we
with u0 a constant throughout the fluid domain. chose it to ensure that we achieve steady-state
The reactant concentration is initially zero every- conditions for both flow and species configura-
where. Boundary conditions for the flow are pe- tions. Table 2 summarizes these results.
riodic at the inlet/outlet and no-slip (zero fluid Not surprisingly, the simulations with the
speed) at the top and bottom rigid walls. The faster chemistry and diffusion show the highest
chemical reactant is continuously injected at the efficiency, η = 0.47. When we reduce D by an
inlet with a constant profile C(x = 0, y) =1 and flows order of magnitude, the efficiency decreases to
freely away from the flow domain at the outlet sec- 0.18. The difference between the faster and
tion. At the top and bottom walls the reactant con- slower diffusion simulations is not as pro-
centration obeys the dynamic boundary condition nounced in the case of slower chemistry (τc =
implied by the rate Equation 23. 103)—the efficiency is 0.09 and 0.07 in the two
To illustrate the LBE’s capabilities, let’s look at cases, respectively. In the slower chemistry sim-
four pairs of simulations, with combinations of ulations, diffusion is always fast enough to get
the values τc = 102 or τc = 103 and D = 10–2 or D = the chemical species to the wall, but the bottle-
10–1. We will refer to the simulations with the neck is waiting for the wall to consume the pol-
lower value of τc as the faster chemistry simula- lutants before advection sweeps them away.
tions and to those with the higher value of τc as For a given set of D and τc values, there is a
the slower chemistry simulations. Similarly, we small but significant difference between the effi-
will refer to the simulations with the higher value ciencies of the smooth channel and the channel
of D as the faster diffusion simulations and those with the obstacle. Figure 6 compares the con-
with the lower value of D as the slower diffusion centration of the pollutant along the channel
simulations. We use the terms faster and slower with and without the obstacle, and Figure 7
to reflect the comparative changes in the system shows a more detailed look at the differences in
because, in all cases, chemical interactions take concentration in the two geometries (for the
place on a rather fast pace compared to advec- faster chemistry case). Figure 6 shows the overall Figure 6.
Concentration
C
C of chemical
1.00 species along
40
the length of
30
the smooth (a)
z

20 0.65
and corrugated
10 (b) channels.
0.30
0 Near the top
25 75 125 175 225 275
(a) x and bottom
C
walls—
1.00 particularly
40
around the
30
obstacle—the
z

20 0.65
concentration
10
falls markedly
0 0.30
from its initial
25 75 125 175 225 275
(b) x value.

NOVEMBER/DECEMBER 2001 33
Figure 7. The D = 10–2; τc = 102)
difference in 50
∆C
concentration
40 0.20
∆C of the
chemical species
30
in the fluid,
z
0.00
between the 20
smooth and
corrugated 10 –0.20
channels, for the
faster chemistry 0
100 150 200
simulations (τc = (a) x
102), at slower D = 10 ; τc = 102)
–1

(a) and faster 50

(b) diffusion. ∆C
40 0.20
Only a portion
of the channel 30
along its length
z

0.00
is shown, close 20
to the
trapezoidal 10
–0.20
obstacle.
0
100 150 200
(b)
x

effect of the obstacle in the concentration, which a positive flux difference develops before the cor-
on this scale is indeed small. Figure 7 indicates rugation, an effect that is much more pro-
the spatial distribution of changes in the concen- nounced in the faster chemistry simulations.
tration ∆C due to the obstacle’s presence: a neg- This results from the perturbation of the flow
ative value of ∆C indicates lower concentration field far from the obstacle. Although the chan-
in the corrugated channel. As Figure 7 shows, the nel is four times longer than the obstacle, near
concentration is significantly reduced in the the inlet and the outlet the velocity field has a
neighborhood of the obstacle (the small increase vertical component with a magnitude approxi-
in concentration in the region far above the ob- mately 10–5 of the maximum velocity in the x di-
stacle is an effect for the conservation of the total rection. This small perturbation, coupled with
pollutant in the two simulations). Remarkably, the rapid chemical reactions at the wall, results
the changes in concentration are as pronounced in the flux difference’s initial rise.
in front of the obstacle as behind it. There is also Figure 8 indicates that the flux difference be-
a significant change in concentration at the top gins decreasing just before the front of the ob-
surface of the obstacle. The results for the slower stacle. This does not appear to be due to en-
chemistry simulations are qualitatively similar but hanced absorption in the wall in front of the
involve less pronounced changes between the obstacle. In fact, the values of the concentration
smooth and corrugated channels. at the wall indicate that the difference in con-
To gain a more quantitative picture of where centration between the two simulations de-
along the channel length the difference in effi- creases right before the beginning of the obsta-
ciency originates, Figure 8 shows the difference cle. Instead, this decrease appears to be the result
in flux between the smooth and corrugated of the chemical species diffusing within the bulk
channels as a function of x, the position along fluid toward the region above the corrugation
the channel. The flux is defined as the concen- where there is a relative deficit of chemical
tration integrated along the vertical z direction species concentration. Interestingly, the decrease
at a given value of x, normalized by the channel begins earlier and has larger magnitude in the
cross section. The flux difference is actually simulations corresponding to faster diffusion.
small in the two simulations—less than 1 per- Figure 8 also indicates that directly on top of the
cent at any point along the channel. In all cases, obstacle there is a steady decrease of the flux dif-

34 COMPUTING IN SCIENCE & ENGINEERING

 0.2 0.1 Figure 8. The percent difference
0.05 in chemical species flux between
0 smooth channel and corrugated
Flux % difference

Flux % difference
0 channel simulations as a function
–0.2
–0.05 of length x along the channel. A
–0.4 negative value indicates that the
–0.10
flux is smaller in the corrugated
D = 0.01 D = 0.01
–0.6 –0.15 channel simulation. (a) τc = 102
D = 0.1 D = 0.1
–0.20 for the faster chemistry
–0.8 simulation, and (b) τc = 103 for
–0.25
100 200 300 100 200 300 the slower chemistry simulation.
x x The arrows mark the beginning
(a) (b) and end of the obstacle.

ference, which results from the increase in con- include the values of Pe, Da, and Re under arbi-
centration along the upper surface of the obsta- trary geometrical conditions. To this purpose,
cle produced by the squeezing of hydrodynamic besides upgrading the computational techniques,
flow lines due to the obstacle’s presence. Finally, a synergistic interplay with analytical scaling the-
after the obstacle, the flux difference begins in- ories could also prove valuable.
creasing—most sharply in simulations with
slower diffusion. We can expect that well beyond

T
the extent of the obstacle the flux difference will he LBE method has already proven
tend to zero, because at that point the flow is its remarkable versatility in address-
similar in the smooth and corrugated channels. ing a wide range of problems that in-
However, the smooth channel has a higher con- volve complex fluid motion at dis-
centration of chemical species in that region due parate scales. Multiscale simulations based on
to the higher consumption of the pollutant the LBE approach, two examples of which we
around the obstacle, so the difference never discussed here, are just beginning to appear.
quite reaches a zero value in the simulations. There are good reasons to believe that further
These considerations give only a flavor of the upgrades of the LBE technique as well as better
type of questions that you can address by nu- communication with related fields of computa-
merically simulating microreactive flows. Ulti- tional physics will make these types of multiscale
mately, the goal would be to map out a sensible applications flourish in the years to come.
expression of the chemical efficiency as a func- Figure 9 gives a more concrete picture of our
tion of the relevant control parameters, which vision for future multiscale approaches. This

Fluid with reactant

Reactive substrate

LBE: ~1 mm KMC: ~1 µm QM: ~1 nm

Figure 9. Illustration of combined multiscale LBE/KMC/QM scheme. In the LBE regime, the system
length scale is ~ 1 mm, and the description is in terms of hydrodynamic fluid flows (gold lines). In the
KMC regime the scale of the cell is ~ 1µm, and the description is in terms of fluxes of reactants (blue
arrows) and reaction products (red arrows) determined by the diffusion, advection and chemical
reactivity rates. In the QM regime, the scale of the cell is ~1 nm, and the description is in terms of
individual molecules and atoms on the surface undergoing surface diffusion and chemical reactions.

NOVEMBER/DECEMBER 2001 35
 type of problem is the epitome of multiscale Acknowledgments
physical phenomena32 involving fluid motion This work was supported in part by the Materials
and reactant transport at macroscopic scales, Research Science and Engineering Center that the
fluid-wall interactions at intermediate scales, National Science Foundation funds at Harvard University.
and chemical reactions on the wall at micro- The NATO Collaborative Link Grant PST.CLG.976357
scopic scales. At the highest length scale, of or- greatly facilitated the interaction between the authors.
der 1 mm, LBE-based simulations can handle Sauro Succi thanks the Physics Department of Harvard
the flow of fluids carrying macroscopic reac- University for his Visiting Scholar appointment and for its
tants around the substrate’s geometric features. hospitality during his stay. We thank Achi Brandt for
These simulations will determine the concen- discussions on mathematical aspects of multiscale
tration fields of the reactants and reaction methods and Howard Stone for discussions on the
products. At the next lower scale, of order 1 physics of fluids.
µm, Kinetic Monte Carlo
(KMC) simulations33 will de-
The LBE has already termine the flux of reactants
toward the substrate and the
proven its remarkable flux of reaction products
away from it—given the References
versatility in macroscopically determined
1. W. Hoover, A.J. de Groot, and C. Hoover, ”Massively Parallel
concentration—as well as the Computer Simulation of the Plane-Strain Elastic-Plastic Flow via
addressing a wide reactivity and surface diffu- Non-Equilibrium Molecular Dynamics and Lagrangian Contin-
sion rates determined by mi- uum Mechanics,” Computers in Physics, vol. 6, no. 2, Mar./Apr.
1992, pp. 155–167.
range of problems. croscopic considerations.
2. E. Salomons and Mareschal, “Usefulness of the Burnett Descrip-
At present, it is possible to tion of Strong Shock Waves,” Physical Rev. Letters, vol. 69, no. 2,
determine a priori the rates July 1992, pp. 269–272.
for a range of relevant atom- 3. S.T. O’Connell and P. Thompson, “Molecular Dynamics-Con-
istic processes from ab initio quantum me- tinuum Hybrid Computations: A Tool for Studying Complex
Fluid Flows,” Physical Rev. E, vol. 52, no. 6, Dec. 1995, pp.
chanical (QM) calculations and feed them R5792–R5795.
into the KMC simulation.34 We can then cou- 4. F.F. Abraham et al., “Spanning the Length Scales in Dynamic
ple the KMC simulation to the LBE simula- Simulation,” Computers in Physics, vol. 12, no. 6, Nov./Dec.
tion in the manner discussed earlier. Ideally, 1998, pp. 538–546.

we would like to be able to calculate the rele- 5. G. McNamara and G. Zanetti, “Use of the Boltzmann Equation to
Simulate Lattice-Gas Automata,” Physical Rev. Letters, vol. 61, no.
vant atomistic rates on the fly, without having 20, Nov. 1988, pp. 2332–2335.
to make any assumptions about the likely 6. F. Higuera and J. Jimenez, “Boltzmann Approach to Lattice Gas
events at this scale. The large computational Simulations,” Europhysics Letters, vol. 9, no. 7, Aug. 1989, pp.
cost of QM calculations makes this unfeasible 663–668.

at present as well as for the foreseeable future 7. F.J. Higuera, S. Succi, and R. Benzi, “Lattice Gas-Dynamics with
Enhanced Collisions,” Europhysics Letters, vol. 9, no. 4, June 1989,
(barring some dramatic computational break- pp. 345–349.
throughs that could speed the QM calcula- 8. H.D. Chen, S.Y. Chen, and W.H. Matthaeus, “Recovery of the
tions by many orders of magnitude). Methods Navier-Stokes Equations Using a Lattice-Gas Boltzmann
for replacing the KMC-level simulations with Method,” Physical Rev. A, vol. 45, no. 8, Apr. 1992, pp.
R5339–R5342.
actual MD simulations are the subject of cur-
9. R. Benzi, S. Succi, and M. Vergassola, “The Lattice Boltzmann
rent research,35 but even those are not based Equation: Theory and Applications,” Physics Reports, vol. 222,
on QM calculations of the energy and forces. no. 3, Dec. 1992, pp. 145–197.
It is nevertheless intriguing to imagine that 10. S. Chen and G. Doolen, “Lattice Boltzmann Method for Fluid
multiscale simulations of the type outlined Flows,” Ann. Rev. Fluid Mechanics, vol. 30, 1998, pp. 329–364.

here will be carried out in the near future, of- 11. S. Succi, The Lattice Boltzmann Equation for Fluid Dynamics and
Beyond, Clarendon Press, Oxford, UK, 2001.
fering insight into complex physical
12. X. Shan and H. Chen, “Lattice Boltzmann Model for Simulating
processes. Allowing for the possibility that we Flows with Multiple Phases and Components,” Physics Rev. E, vol.
can realize this dream, the application to bio- 47, no. 3, Mar. 1993, pp. 1815–1819.
logically related problems—where both the 13. A. Anagnost et al., Digital Physics: Analysis of the Morel Body in
macroscopic geometric features of vescicles Ground Proximity, tech. report 970139, Int’l Congress and Expo-
sition, Detroit, 1997.
and the reaction of molecules at the micro-
14. S. Chen et al., “Lattice Boltzmann Model for Simulation of
scopic level are of crucial importance—seems Magnetohydrodynamics,” Physical Rev. Letters, vol. 67, no.
particularly promising. 27, Dec. 1991, pp. 3776–3779.

36 COMPUTING IN SCIENCE & ENGINEERING

15. Y.H. Qian, D. d’Humieres, and P. Lallemand, “Lattice BGK Mod- For more information on this or any other computing
els for Navier-Stokes Equation,” Europhysics Letters, vol. 17, no. 6,
Feb. 1992, pp. 479–484.
topic, please visit our Digital Library at http://computer.
org/publications/dlib.
16. P. Bhatnagar, E. Gross, and K. Krook, “A Model for Collisional
Processes in Gases I: Small Amplitude Processes in Charged and
Neutral One-Component System,” Physical Rev. B, vol. 94, no.
3, 1954, pp. 511–525.
17. U. Frisch, B. Hasslacher, and Y. Pomeau, “Lattice Gas Automata Sauro Succi is research director of the National Re-
for the Navier-Stokes Equations,” Physics Rev. Letters, vol. 56, no.
14, Apr. 1986, pp. 1505–1508. search Council’s Institute of Applied Computing in
18. X. He and L.-S. Luo, “Theory of the Lattice Bolztmann Equation:
Rome, Italy. His scientific interests cover a range of top-
From Boltzmann Equation to Lattice Boltzmann Equation,” ics associated with the mathematical and computa-
Physics Rev. E, vol. 56, no. 6, Dec. 1997, pp. 6811–6817. tional modeling of complex system dynamics, includ-
19. O. Filippova and D. Hänel, “Grid Refinement for Lattice-BGK ing fluid turbulence, microhydrodynamics, quantum
Models,” J. Computational Physics, vol. 147, no. 11, Nov. 1998,
fluids, and discrete kinetic systems. He received a PhD
pp. 219–228.
in plasma physics from the Swiss Polytechnic Institute
20. F. Mazzocco et al., “Multiscale Lattice Boltzmann Schemes: A
Preliminary Application to Axial Turbomachine Flow Simula- of Lausanne, Switzerland. Contact him at Istituto Ap-
tions,” Int’l J. Modern Physics C, vol. 11, no. 2, Mar. 2000, pp. plicazioni Calcolo, viale Policlinico 137, 00161, Roma,
233–245. Italy; succi@iac.rm.cnr.it; www.iac.rm.cnr.it.
21. O. Filippova et al., “Multiscale Lattice Boltzmann Schemes with
Turbulence Modeling,” J. Computational Physics, vol. 170, no. 2,
July 2001, pp. 812–829.
Olga Filippova is a researcher at the Institute of Com-
22. F. Nannelli and S. Succi, “The Lattice Boltzmann Equation in Ir-
regular Lattices,” J. Statistical Physics, vol. 68, nos. 3–4, Aug. bustion and Gas Dynamics at the University of Duis-
1992, pp. 401–407. burg affiliated with the Aerodynamical Institute RWTH
23. X. He, L. Luo, and M. Dembo, “Some Progress in Lattice Boltz- Aachen. Her scientific interests include the develop-
mann Method: Part I, Non-Uniform Grids,” J. Computational ment and application of the Lattice Boltzmann method
Physics, vol. 129, no. 2, Dec. 1996, pp. 357–363.
as a new efficient computational tool for studying low
24. H.D. Chen, “Volumetric Formulation of the Lattice Boltzmann
Mach number flows, in particular, turbulent and com-
Method for Fluid Dynamics: Basic Concept,” Physical Rev. E, vol.
58, no. 3, Sept. 1998, pp. 3955–3963. bustion flows in complex geometries. She received a
25. G.W. Peng et al., “Finite Volume Scheme for the Lattice Boltz- PhD from Moscow State University, Russia. Contact her
mann Method on Unstructured Meshes,” Physical Rev. E, vol. 59, at the Institute of Combustion and Gasdynamics, Univ.
no. 4, Apr. 1999, pp. 4675–4682. Duisburg, 47048 Duisburg, Germany; olga@vug.uni-
26. O. Filippova and D. Hänel, “Acceleration of Lattice-BGK Schemes duisburg.de; www.vug.uni-duisburg.de/STAFF/worker-
with Grid Refinement,” J. Computational Physics, vol. 165, no. 2,
Dec. 2000, pp. 407–427.
olga-main.html.
27. H. Chen, C. Teixeira, and K. Molvig, “Realization of Fluid Bound-
ary Conditions via Discrete Boltzmann Dynamics,” Int’l J. Mod-
ern Physics C, vol. 9, no. 8, Dec. 1998, pp. 1281–1292. Greg Smith did his graduate work in the Physics De-
28. C.M. Rhie and W.L. Chow, “Numerical Study of the Turbulent partment of Harvard University, where he earned a
Flow Past an Airfoil with Trailing Edge Separation,” AIAA J., vol. PhD. He also worked as a postdoctoral fellow on mul-
21, no. 11, 1983, pp. 1525–1532.
tiscale modeling of materials. He is currently studying
29. D. Coles and A.J. Wadcock, “Flying Hot-Wire Study of Flow Past
an NACA 4412 Airfoil at Maximum Lift,” AIAA J., vol. 17, no. 4,
art at Hunter College, New York City. Contact him at
1979, pp. 312–329. gregory_sam_smith@yahoo.com.
30. S. Succi et al., “Chemical Efficiency of Reactive Microflows with
Heterogeneous Catalysis: A Lattice Boltzmann Study,” to be pub-
lished in European J. Applied Physics.
Efthimios Kaxiras is a professor of physics and the
31. S. Succi et al., “An Integer Realization of a Lax Scheme for Trans- Gordon McKay Professor of Applied Physics in the
port Processes in Multiple Component Flows,” J. Computational
Physics, vol. 152, no. 2, July 1999, pp. 493–516.
Physics Department and the Division of Engineering
and Applied Sciences at Harvard University. His re-
32. E. Oran and J. Boris, Numerical Simulation of Reactive Flows, Else-
vier Science, New York, 1987. search interests include the physics of solids, surfaces,
33. M. Silverberg, A. Ben-Shaul, and F. Rebenstrost, “On the Effects clusters, and macromolecules, studied by first-princi-
of Adsorbate Aggregation on the Kinetics of Surface Reactions,” ples electronic structure methods. More recently, he
J. Chemical Physics, vol. 83, no. 12, Dec. 1985, pp. 6501–6513. has been involved in incorporating such methods in
34. M. Seel, “Modeling of Solid Rocket Fuel: From Quantum Chem- multiscale modeling of materials properties. He re-
istry to Fluid Dynamic Simulations,” Computers in Physics, vol. 5,
no. 5, Sept./Oct. 1991, p. 460.
ceived a PhD in physics from MIT. Contact him at Ly-
man Laboratory, Physics Dept., Harvard Univ., Cam-
35. A.F. Voter, “Hyperdynamics: Accelerated Molecular Dynamics of
Infrequent Events,” Physical Rev. Letters, vol. 78, no. 20, May bridge MA 02138; kaxiras@physics.harvard.edu;
1997, pp. 3908–3911. http://kriti.harvard.edu.

NOVEMBER/DECEMBER 2001 37