Lattice Gas Cellular Automata and Lattice Boltzmann Models Chapter6

5.
Lattice Boltzmann Models
5.1 From lattice-gas cellular automata to lattice

Boltzmann models
Lattice-gas cellular automata for Navier-Stokes equations are plagued by sev-

eral diseases. Only for some of them therapies and cures could be found
(compare Table 5.1.1).
Table 5.1.1. Diseases of lattice-gas cellular automata: its causes and therapies and
cures.
disease cause therapy/cure remarks

non-isotropic lattice higher symmetry HPP → FHP
advection tensor of 4th of lattice
term rank is add inner degree HPP → PI
non-isotropic of freedom
multi-speed models
violation of Fermi-Dirac rescaling FHP, FCHC, PI
the Galilei distributions (symptomatic
invariance treatment)
spurious regular as much collisions Zanetti
invariants lattices as possible invariants
noise Boolean averaging enormous memory
variables (coarse graining) demand
pressure depends multi-speed Chen et al. 1989
explicitly on models
velocity
Historically the following stages in the development of lattice Boltzmann

models1 (LBM) can be distinguished:
1
A lattice Boltzmann model encompasses a lattice, an equilibrium distribution
and a kinetic equation which is called lattice Boltzmann equation (LBE). In
D.A. Wolf-Gladrow: LNM 1725, pp. 159–246, 2000.

c Springer-Verlag Berlin Heidelberg 2000
160 5. Lattice Boltzmann Models
1. Lattice Boltzmann equations have been used already at the cradle of

lattice-gas cellular automata by Frisch et al. (1987) to calculate the vis-
cosity of LGCA.
2. Lattice Boltzmann models as an independent numerical method for hy-
drodynamic simulations were introduced by McNamara and Zanetti in
1988. The main motivation for the transition from LGCA to LBM was
the desire to get rid of the noise. The Boolean fields were replaced by
continuous distributions over the FHP and FCHC lattices. Fermi-Dirac
distributions were used as equilibrium functions.
3. Linearized collision operator (Higuera and Jiménez, 1989).
4. Boltzmann instead of Fermi-Dirac distributions.
5. The collision operator, which is based on the collisions of a certain LGCA,
has been replaced by the BGK (also called single time relaxation) approx-
imation by Koelman (1991), Qian et al. (1992) and others. These lattice
BGK models (LBGK) mark a new level of abstraction: collisions are not
anymore defined explicitly.
Multi-speed LBGK models are most popular today. However, more complex
collision operators are still in use in models of multi-phase flows (Rothman
and Zaleski, 1994).
5.1.1 Lattice Boltzmann equation and Boltzmann equation
The microdynamics of LGCA are described by kinetic equations of the type

ni (x + ci ∆t, t + ∆t) = ni (x, t) + ∆i (5.1.1)
(compare, for example, eq. 3.2.15 where ∆t was set to 1). ∆i is the colli-
sion function of the respective model. Discrete equations of the form (5.1.1)
are referred to as lattice Boltzmann equations. The correspondence to the
Boltzmann equation
∂f
+ v∇f = Q (5.1.2)
∂t
(Q is the collision integral; external forces have been neglected) can be shown
by expansion of the left hand side of eq. (5.1.1):
∂ni
2
ni (x + ci ∆t, t + ∆t) = ni (x, t) + ∆t + ci ∆t∇ni + O (∆t) .
∂t
Neglecting higher order terms, one obtains
∂ni ∆i
+ ci ∇ni =
∂t ∆t
the literature you will find other names like ‘lattice Boltzmann equation’ or
‘Boltzmann cellular automata’ instead of lattice Boltzmann model.
5.1 From lattice-gas cellular automata to lattice Boltzmann models 161
which by substituting ni → f , ci → v, ∆i /∆t → Q gives the Boltzmann

equation (5.1.2).
Sterling and Chen (1996) show that this is more than pure formal correspon-
dence. They derive the lattice Boltzmann equation as a special discretization
of the Boltzmann equation. The Boltzmann equation with BGK approxima-
tion reads
∂f 1
+ v∇f = − f − f (eq) .
∂t τ
The distribution function f depends on space, velocity and time: f (x, v, t).
The v-space is discretized by introducing a finite set of velocities, v i , and
associated distribution functions, fi (x, t), which are governed by the discrete
Boltzmann equation:
∂fi 1 (eq)

+ v i ∇fi = − fi − fi
∂t τ
(please note that this equation is different from the discretized Boltzmann
equation, see below). The discrete Boltzmann equation will be nondimen-
sionalized by the characteristic length scale, L, the reference speed, U , the
reference density, nr , and the time between particle collisions, tc ,
∂Fi
ˆ i = − 1 Fi − F (eq)
+ ci ∇F (5.1.3)
i
∂ t̂ τ̂
where ci = v i /U , ∇
ˆ = L∇, t̂ = t · U/L, τ̂ = τ /tc , Fi = fi /nr . The parameter
U
= tc
L
may be interpreted as either the ratio of collision time to flow time or as the
ratio of mean free path to the characteristic length (i.e., Knudsen number).
A discretization of eq. (5.1.3) is given by:

Fi x̂, t̂ + ∆t̂ − Fi x̂, t̂ Fi x̂ + ∆x̂, t̂ + ∆t̂ − Fi x̂, t̂ + ∆t̂
+ cix
∆x̂
∆t̂
Fi x̂ + ∆y, t̂ + ∆t̂ − Fi x̂, t̂ + ∆t̂
+ciy
∆y

Fi x̂ + ∆z, t̂ + ∆t̂ − Fi x̂, t̂ + ∆t̂
+ciz
∆z
1 (eq)

= − F i − Fi
τ̂
where ∆t̂ = ∆t · U/L. Lagrangian behavior is then obtained by the selection

of the lattice spacing divided by the time step to equal the lattice velocity
(∆x̂/∆t̂ = ci ):

Fi x̂, t̂ + ∆t̂ − Fi x̂, t̂ Fi x̂ + ci ∆t̂, t̂ + ∆t̂ − Fi x̂, t̂ + ∆t̂
+
∆t̂ ∆t̂
Fi x̂ + ci ∆t̂, t̂ + ∆t̂ − Fi x̂, t̂ 1
(eq)
= =− Fi − F i . (5.1.4)
∆t̂ τ̂
Thus two terms on the left hand side cancel each other and thereby the
method becomes explicit. Choosing ∆t = tc , multiplying eq. (5.1.4) by ∆t̂
and dropping all carets leads to the (BGK) lattice Boltzmann equation
1 (eq)

Fi (x + ci ∆t, t + ∆t) − Fi (x, t) = − Fi − Fi (5.1.5)
τ
Sterling and Chen (1996, p.200) give the following interpretation of Eq.
(5.1.5)2 :
“This equation has a particular simple physical interpretation in

which the collision term is evaluated locally and there is only one
streaming step or ‘shift’ operation per lattice velocity. This stream-
and-collide particle interpretation is a result of the fully Lagrangian
character of the equation for which the lattice spacing is the distance
travelled by the particle during a time step. Higher order discretiza-
tions of the discrete Boltzmann equation typically require several
‘shift’ operations for the evaluation of each derivative and a particle
interpretation is less obvious. ... It did not originally occur to the
authors” [i.e. McNamara and Zanetti in 1988] “that the LB method
could be considered a particular discretization for the discrete Boltz-
mann equation (G. McNamara, private communication).”
2
Readers not familiar with lattice-gas cellular automata should skip this
quotation.
5.1 From lattice-gas cellular automata to lattice Boltzmann models 163
Table 5.1.2. From the Boltzmann equation to the lattice Boltzmann equation.
Boltzmann equation:
∂f
+ v∇f = Q
∂t
Boltzmann equation (BGK approximation):
∂f 1
+ v∇f = − f − f (eq)
∂t τ
discrete Boltzmann equation:

∂fi 1 (eq)
+ v i ∇fi = − fi − fi
∂t τ
non-dimensional discrete Boltzmann equation:

∂Fi ˆ i = − 1 Fi − F (eq)
+ ci ∇F i
∂ t̂ τ̂
discretized Boltzmann equation:

Fi x̂, t̂ + ∆t̂ − Fi x̂, t̂ Fi x̂ + ∆x̂, t̂ + ∆t̂ − Fi x̂, t̂ + ∆t̂
+ cix ...
∆t̂ ∆x̂

1 (eq)
=− Fi − Fi
τ̂
lattice Boltzmann equation:

1 (eq)
Fi (x + ci ∆t, t + ∆t) − Fi (x, t) = − Fi − Fi
τ
5.1.2 Lattice Boltzmann models of the first generation
The following remarks are mainly of historical interest and should be skipped
by readers not familiar with LGCA.
LGCA are plagued by noise which can be suppressed by coarse graining
over large domains and/or time intervals. Thus low noise levels are costly in
terms of memory and computer time (see Dahlburg et al., 1987, for further
discussion). In order to get rid of this noise McNamara and Zanetti (1988)
proposed to use directly the mean occupation numbers instead of the Boolean
fields. The lattice Boltzmann equation (LBE) had been applied already before
by Wolfram (1986) and Frisch et al. (1987) as an analytical tool to calculate
the viscosity coefficients of LGCA. McNamara and Zanetti for the first time
used the LBE as a numerical scheme.
The mean occupation numbers Fi (0 ≤ Fi ≤ 1) develop in time according to

the following kinetic equation
Fi (r + ci , t + 1) = Fi (r, t) + Ωi ({Fj (r, t)}) , (5.1.6)
where the form of the collision operator Ωi is identical to the arithmetic form
of the microscopic collision operator of the corresponding LGCA. For FHP,
for example, the Boolean (discrete) variables ni in (3.2.17) are replaced by
mean (continuous) occupation numbers Fi :
Ωi (N ) = Fi+1 Fi+3 Fi+5 (1 − Fi )(1 − Fi+2 )(1 − Fi+4 )

−Fi Fi+2 Fi+4 (1 − Fi+1 )(1 − Fi+3 )(1 − Fi+5 )
+ξFi+1 Fi+4 (1 − Fi )(1 − Fi+2 )(1 − Fi+3 )(1 − Fi+5 ) (5.1.7)
+(1 − ξ)Fi+2 Fi+5 (1 − Fi )(1 − Fi+1 )(1 − Fi+3 )(1 − Fi+4 )
−Fi Fi+3 (1 − Fi+1 )(1 − Fi+2 )(1 − Fi+4 )(1 − Fi+5 ).
This type of LBM has been improved by Higuera and Jiménez (1989). They
could show that the nonlinear collision operator which evaluation is time
consuming can be approximated by a linear operator.
Further reading: Benzi et al. (1992) give an extensive review on LBMs
based on the FHP and FCHC models.
5.2 BGK lattice Boltzmann model in 2D 165
5.2 BGK lattice Boltzmann model in 2D
Lattice Boltzmann models (LBMs) of the first generation are plagued by the
same problems as the corresponding lattice-gas cellular automata except for
the noise. Modern LBMs with Boltzmann distribution functions, several lat-
tice speeds and BGK approximation of the collision operator are free of all
problems mentioned in Table 5.1.1. In this chapter a detailed discussion of
such a model in 2D will be given. The derivations of the equilibrium distri-
butions and the Navier-Stokes equation will be presented in full length. We
will refer to this model as the D2Q9-LBM.
A LBM has three main ingredients:
1. the lattice: D2Q9 (multi-speed model),

2. the equilibrium distributions: Maxwell (see below),
3. the kinetic equation: BGK approximation.
On the contrary, a LGCA is basically defined by the lattice and the collision
rules.
Koelman (1991) defines his LBM for the Navier-Stokes equation over the
lattice with the following lattice velocities ci
cK
0 = (0, 0)
cK
1,2,3,4 = (±a, ±b)
cK5,6 = (±2a, 0)
cK7,8 = (0, ±2b)
where the lattice constants a and b are restricted by a2 /3 ≤ b2 ≤ 3a2 . The

1
special choice a = b = √ c and rotation of the lattice velocities by 45◦ leads
2
to the D2Q9 lattice (compare Section 3.3 and Fig. 5.2.1) with
c0 = (0, 0)
c1,3 , c2,4 = (±c, 0), (0, ±c) (5.2.1)
c5,6,7,8 = (±c, ±c).
Fig. 5.2.1. The D2Q9 lattice.
'$ '$ '$
6 2 5
&% &% &%
I
@ 6
@
@
@
@
@
@
@
@
@
'$ @'$ '$
-
3 0 1
&% &% &%
@
@
@
@
@
@
@
@
@
@
'$
?
'$ R'$
@
7 4 8
&% &% &%
In what follows the D2Q9 lattice will be used exclusively.

The mass density, ρ, and the momentum density, j, are defined by sums over
the distribution functions Fi (x, t)

ρ(x, t) = Fi (x, t) (5.2.2)
i

j(x, t) = ρ(x, t)v(x, t) = ci Fi (x, t) (5.2.3)
i
For vanishing velocities a global equilibrium distribution Wi (“fluid at rest”)

is defined. In the vicinity (small Mach numbers) of this resting equilibrium,
distribution functions can be written as sums of the Wi and small perturba-
tions fi (x, t)
Fi (x, t) = Wi + fi (x, t) (5.2.4)
with |fi (x, t)| << Wi .
The Wi should be positive to assure positive mass density. They are chosen
of Maxwell type in the following sense. The lattice velocity moments up to
fourth order over the Wi shall be identical to the respective velocity moments
over the Maxwell distribution
!D/2
m ' (
wB (v) = ρ0 exp −mv 2 /2kB T (5.2.5)
2πkB T
(D dimension, ρ0 mass density, m particle mass, v particle speed, kB Boltz-
mann constant, T temperature). Thus the odd moments vanish

Wi ciα = 0
i

Wi ciα ciβ ciγ = 0
i
and the even moments read

Wi = dv wB (v) = ρ0 (5.2.6)
i

kB T
Wi ciα ciβ = dv wB (v)vα vβ = ρ0 δαβ (5.2.7)
i
m

Wi ciα ciβ ciγ ciδ = dv wB (v)vα vβ vγ vδ
i
!2
kB T
= ρ0 (δαβ δγδ + δαγ δβδ + δαδ δβγ ).(5.2.8)
m
Note that the constraint (5.2.8) is more rigorous than pure isotropy (compare
Section 3.3).
Exercise 5.2.1. (**)

Calculate the velocity moments up to fourth order over the Maxwell distri-
bution in D ≥ 2 dimensions.
Nonnegative solutions of Eqs. (5.2.6 - 5.2.8) for the Wi can be found whenever
the number of lattice velocities ci is large enough. For the D2Q9 lattice (5.2.1)
one obtains (see Section 5.2.1 for derivation):
4
W0 /ρ0 =
9
1
W1 /ρ0 =
9
1
W2 /ρ0 =
36
kB T c2
= .
m 3
The evolution of the LBM consists of the recurring alternation between tran-
sition to the local equilibrium and propagation of the distributions to neigh-
boring sites according to the lattice velocities. The BGK (compare Section
4.1.3) kinetic equation reads
∆t
(0)

Fi (x + ci ∆t, t + ∆t) − Fi (x, t) = − Fi (x, t) − Fi (x, t)
τ
∆tciα
+ [Kα (x, t) + Kα (x + ci ∆t, t + ∆t)]
12c2
(5.2.9)
or
(0)
Fi (x + ci ∆t, t + ∆t) = (1 − ω) Fi (x, t) + ωFi (x, t)
(5.2.10)
∆tciα
+ [Kα (x, t) + Kα (x + ci ∆t, t + ∆t)]
12c2
where τ is the collision time, ω = ∆t/τ is the collision frequency, and K is

(0)
an applied body force. The local equilibrium distributions Fi depend only
on the local values of mass and momentum density
(0) (0)
Fi (x, t) = Fi (ρ (x, t) , j (x, t)) .
They can be derived by applying the maximum entropy principle under the
constraints of mass and momentum conservation (see Section 5.2.2). Up to
second order in j one obtains
% &
(0) Wi m m m 2
Fi (ρ, j) = ρ+ ci · j + (ci · j) − j 2 (5.2.11)
ρ0 kB T 2ρkB T kB T
or more explicitly
% &
4 3 u2
Fi = ρ 1− i=0
9 2 c2
% &
1 ci · u 9 (ci · u)2 3 u2
Fi = ρ 1+3 2 + − i = 1, 2, 3, 4
9 c 2 c4 2 c2
% &
1 ci · u 9 (ci · u)2 3 u2
Fi = ρ 1+3 2 + − i = 5, 6, 7, 8.
36 c 2 c4 2 c2
Application of the multi-scale technique (Chapman-Enskog expansion) yields

the Navier-Stokes equation with pressure p = ρkB T /m, kinematic shear vis-
cosity
! ! !
kB T 1 1 c2 1 1 2−ω 2 c2 ∆t
ν= − ∆t = − ∆t = c ∆t = τ−
m ω 2 3 ω 2 6ω 3 2
(5.2.12)
and an advection term with Galilean invariance3 . The Galilean invariance
breaking g-factor (compare Section 3.2) never arises (see Section 5.2.3).
The above given presentation of the LBM contains all informations necessary
to set up the computer code. The algorithm proceeds as follows:
1. For given initial values of mass ρ(x, t) and momentum density j(x, t)
(0)
calculate the equilibrium distributions Fi (ρ(x, t), j(x, t)) according to
(0)
Eq. (5.2.11) and set Fi = Fi . Remark: Global equilibrium distributions
(5.2.11) but with different values of ρ and j are used for local initializa-
tion, i.e. in the beginning there is a patchwork of local equilibria which
is far from a global equilibrium.
2. Apply the kinetic equation (5.2.10), i.e. add the (non-equilibrium) dis-
tribution function Fi (x, t) and the equilibrium distribution function
(0)
Fi (x, t) with the appropriate weights (1−ω) and ω, and then propagate
it to the next neighbor (except for the distribution of ‘rest particles’ with
c0 = 0).
3. Calculate from the propagated distributions new values of ρ(x, t) and
j(x, t) according to the definitions (5.2.2) und (5.2.3).
4. The next time step starts with the calculation of new equilibrium distri-
butions. Thereafter proceed with the second step of the algorithm.
3
But: There is always a disturbing fly or two in the ointment. Qian and Orszag
(1993) and Qian and Zhou (1998) have shown that a higher order term (slightly
compressible regime) leads to frame-velocity-dependent viscosity.
In BGK-LBMs collisions are not explicitly defined. They are kind of fictive
and make themselves felt only by the transition to local equilibrium (the
(0)
term ω Fi (x, t) in the kinetic equation). On the other hand for LGCA the
collisions are explicitly defined and the distribution functions are theoretical
constructs which can not be observed on the lattice because of their contin-
uous nature.
5.2.1 Derivation of the Wi
The D2Q9 model includes three different speeds (compare Table 5.2.1). The
Wi for directions with identical speeds are equal for reason of symmetry.
Table 5.2.1. Lattice speeds, cells and Wi of the D2Q9 lattice.
c2i cells number of cells Wi

0 0 1 W0
1 1, 2, 3, 4 4 W1
2 5, 6, 7, 8 4 W2
The non-vanishing elements of the even moments up to fourth order read:
– 0. moment:
Wi = W0 + 4W1 + 4W2 = ρ0 (5.2.13)
i
Remark: The only equation which includes W0 . It will be used to calculate

W0 .
– 2. moment:
kB T
c2i1 Wi = 2c2 W1 + 4c2 W2 = ρ0 (5.2.14)
i
m
Remark: (kB T )/m will be calculated from Eqs. (5.2.14) and (5.2.15).
kB T
c2i2 Wi = 2c2 W1 + 4c2 W2 = ρ0
i
m
Remark: This constraint is identical with (5.2.14).

– 4. Moment:
!2
kB T
c4i1 Wi = 2c4 W1 + 4c4 W2 = 3ρ0 (5.2.15)
i
m
Remark: W1 will be calculated from Eq. (5.2.15).

!2
kB T
c4i2 Wi = 2c4 W1 + 4c4 W2 = 3ρ0
i
m
Remark: This constraint is identical with (5.2.15).

!2
kB T
c2i1 c2i2 Wi 4
= 4c W2 = ρ0 (5.2.16)
i
m
Remark: W2 will be calculated from Eq. (5.2.16).
The solution reads:

4
W0 /ρ0 =
9
1
W1 /ρ0 =
9
1
W2 /ρ0 =
36
kB T c2
= .
m 3
One readily confirms that the first and third moments of the lattice velocities
over Wi vanish.
5.2.2 Entropy and equilibrium distributions
Koelman defines the relative entropy4 density by
k (0)
(0)
F (ρ, j)
S(ρ, j) := − F (ρ, j) ln i . (5.2.17)
m i i Wi
The weighting by 1/Wi in the logarithmic factor will lead to equilibrium

distributions of the form Fi = Wi eh(ρ,j ,ci ) and implies S = 0 for Fi =
(0) (0)
(0)
Wi . The equilibrium distributions Fi will be calculated by maximizing the
entropy for given constraints which for the case under consideration are the
mass and momentum density
4
Kullback (1959) and Cover and Thomas (1991) are standard references on rela-
tive entropy.
(0)
ρ(ρ, j) = Fi (ρ, j)
i
(0)
j(ρ, j) = ci Fi (ρ, j).
i
The functional
Ŝ := S + Ãρ + B̃ · j
k (0) (0)
(0)
F (ρ, j) (0)
= − Fi (ρ, j) ln i + Ã Fi (ρ, j) + B̃ ci Fi (ρ, j)
m i Wi i i
encompasses the constraints coupled by Lagrange multipliers Ã und B̃. The

necessary conditions for an extremum of Ŝ read
+ ,
(0)
∂ Ŝ k Fi
∀i : (0)
=− ln + 1 + Ã + B̃ · ci = 0. (5.2.18)
∂F m Wi
i
The solutions of (5.2.18) are of the form
= Wi eA(ρ, j) + B(ρ, j) · ci
(0)
Fi
with m m
Ã − 1, and B = B̃.
A=
k k
(0)
A and B can be determined by Taylor series expansions of Fi around j = 0
(same procedure as for the FHP lattice-gas cellular automata). Because of
the symmetry of the D2Q9 lattice the ansatz
A(ρ, j) = A0 (ρ) + A2 (ρ)j 2 + O(j 4 )
B(ρ, j) = B1 (ρ)j + O(j 3 ).
(0)
is sufficient. The expansion of Fi around j = 0 reads
(0)
∂Fi (0)
= (2A2 jα + B1 ciα )Fi
∂jα
→ B1 ciα Wi eA0 at j = 0
2
∂ Fi
(0)
' ( (0)
= (2A2 jα + B1 ciα )2 + 2A2 Fi
∂jα2
→ (B12 c2iα + 2A2 )Wi eA0 at j = 0
2 (0)
∂ Fi (0)
= (2A2 jα + B1 ciα )(2A2 jβ + B1 ciβ )Fi
∂jα ∂jβ
→ B12 ciα ciβ Wi eA0 at j = 0
and finally up to terms of second order in j

B2
Fi = Wi eA0 1 + B1 ci · j + 1 (ci · j)2 + A2 j 2 .
(0)
(5.2.19)
2
Now the definitions of ρ and j will be exploited to calculate the unknown
coefficients A0 (ρ), A2 (ρ) and B1 (ρ):
(0)
B 2 kB T 2
Fi = ρ = eA0 ρ0 + 1 ρ0 j + ρ 0 A2 j 2 (5.2.20)
i
2 m
kB T
= j = eA0 B1 ρ0
(0)
c i Fi j. (5.2.21)
i
m
The vector equation (5.2.21) reduces to a scalar constraint

1 m
B1 = ,
eA0 ρ0 kB T
i.e. an auxiliary condition is required to solve for all three unknowns (A0 , A2 ,
B1 ). Solving Eq. (5.2.20) for A2 yields
!
B 2 kB T 1 ρ
A2 = − 1 + 2 −1 .
2 m j ρ eA0
0
(∗)
To obtain A2 independent of j (as implied by the ansatz) the expression (∗)

must vanish. This is the third constraint looked for. It immediately follows
that
ρ m 1 m
eA0 = , B1 = , and A2 = − 2 .
ρ0 ρkB T 2ρ kB T
Insertion into (5.2.19) finally yields the equilibrium distributions
% &
(0) Wi m m m
Fi (ρ, j) = ρ+ ci · j + (ci · j)2 − j 2 . (5.2.22)
ρ0 kB T 2ρkB T kB T
Koelman (1991) used 1/ρ0 instead of 1/ρ in the coefficient of the third term
which is a good approximation for small Mach numbers.
Further reading: Karlin et al. (1998) construct
.local equilibrium functions,
(0) (0) (0)
Fi , that maximize the entropy SK = i Fi Fi .
5.2.3 Derivation of the Navier-Stokes equations by multi-scale

analysis
The derivation of the macroscopic equations (Navier-Stokes) proceeds in close

analogy to the multi-scale analysis discussed in Section 4.2 and for the FHP
model in Section 3.2. For the BGK-LBM the calculation of terms of second
order in the expansion parameter is much simpler than for LGCA. Here the
complete derivation of the Navier-Stokes equation will be given.
The distributions Fi (x, t) are expanded around the equilibrium distributions
(0)
Fi (x, t)
(0) (1) (2)
Fi (x, t) = Fi (x, t) + Fi (x, t) + 2 Fi + O(3 ) (5.2.23)
with
(1)
(1)
Fi (x, t) = 0, ci Fi (x, t) = 0,
i i
(5.2.24)
(2)
(2)
Fi (x, t) = 0, ci Fi (x, t) = 0,
i i
(1) (2)
i.e. the perturbations Fi (x, t) and Fi (x, t) do not contribute to the mass
and momentum density. The small expansion parameter can be viewed as
the Knudsen number Kn which is the ratio between the mean free path and
the characteristic length scale of the flow.
The left hand side of the kinetic equation (5.2.10) and the forcing term are
expanded into a Taylor series up to terms of second order (∆xi = ∆t · ci ):
Fi (x + ci ∆t, t + ∆t) = Fi (x, t) + ∆t∂t Fi + ∆tciα ∂xα Fi

(5.2.25)
(∆t) ' (
2
+ ∂t ∂t Fi + 2ciα ∂t ∂xα Fi + ciα ciβ ∂xα ∂xβ Fi + O ∂ 3 Fi
2
and
∆t ciα
[Kα (x, t) + Kα (x + ci ∆t, t + ∆t)]
c2 12
(5.2.26)
∆t ciγ
2
(∆t) ciγ ' (

3
= Kγ (x, t) + ∂t Kγ + ciβ ∂xβ Kγ + O (∆t) .
c2 6 c2 12
As for the FHP model two time scales and one spatial scale with the following
scaling will be introduced
(1) (2)
∂t → ∂t + 2 ∂t
(5.2.27)
∂xα → ∂x(1)
α
.
Conservation of mass and momentum. The expansions given above are

substituted into the kinetic equation (Eq. 5.2.10)

(0)
0 = Fi (x + ci ∆t, t + ∆t) − Fi (x, t) + ω Fi (x, t) − Fi (x, t)
∆t ciα
− [Kα (x, t) + Kα (x + ci ∆t, t + ∆t)]
c2 12
which leads to
0 =
Fi (x, t) + ∆t∂t Fi + ∆tciγ ∂xγ Fi
(5.2.25)
(∆t) ' (
2
+ ∂t ∂t Fi + 2ciγ ∂t ∂xγ Fi + ciβ ciγ ∂xβ ∂xγ Fi
2

' ( (0)
+O ∂ 3 Fi − Fi (x, t) + ω Fi (x, t) − Fi (x, t)
(∆t) ciγ ' (

2
∆t ciγ
− 2
K γ (x, t) − 2
∂t Kγ + ciβ ∂xβ Kγ
c
6 c 12
3
+O (∆t)

(1) (0) (0)
=
∆t ∂t Fi + ciγ ∂x(1)
γ
F i
(5.2.23),(5.2.27)

(1) (1) (2) (0) (1)
+2 ∆t ∂t Fi + ∂t Fi + ciγ ∂x(1)
γ
F i
(∆t)
(1) (1) (0)
2
(1) (0)
+2 ∂t ∂t Fi + 2ciγ ∂t ∂x(1)
γ
Fi
2
(0) (1)
+ciβ ciγ ∂x(1)
β
∂x(1)
γ
Fi + ωFi
(2) ∆t ciγ
+2 ωFi − Kγ
c2 6
2
(∆t) ciγ
' 3(
(1)
−2 ∂t Kγ + ciβ ∂x(1)
β
K γ +O
c2 12
Please note that the leading order of the body forcing was set proportional
to . Sorting according to orders in yields
(0) (1) ' (
0 = Ei + 2 Ei + O 3 (5.2.28)
with
(0) (1) (0) ω (1) ciγ (0)
Ei = ∂t Fi F − 2 Kγ
+ ciγ ∂x(1) Fi + (5.2.29)
∆t iγ
6c
(1) (1) (1) (2) (0) (1) ∆t (1) (1) (0)
Ei = ∂t Fi + ∂t Fi + ciγ ∂x(1) Fi + ∂ ∂ F
γ
2 t t i
(1) (0) ∆t (0)
+∆tciγ ∂t ∂x(1)γ
Fi + ciβ ciγ ∂x(1)
β
∂x(1)
γ
Fi (5.2.30)
2
ω (2) ciγ (1) ciγ
+ Fi − ∆t ∂ Kγ − ∆t ciβ ∂x(1) Kγ .
∆t 12c2 t 12c2 β
(0)
We will now calculate the zeroth and first lattice velocity moments of Ei
(1)
and Ei (conservation of mass and momentum density).
Terms of first order in . The zeroth and first lattice velocity moments
(0)
of Ei can be readily calculated:
(0) / (1) (0) (0) ω (1) ciγ 0
Ei = ∂t Fi + ciγ ∂x(1) F i + F − K γ
i i
γ
∆t i 6c2
(1)
= ∂t ρ + ∂x(1)
γ γ
j
/ ω (1) ciβ 0
(0) (1) (0) (0)
ciα Ei = ciα ∂t Fi + ciβ ∂x(1)
β
F i + F i − K β
i i
∆t 6c2
(5.2.31)
(1) (0)
= ∂t jα + ∂x(1)
β
Pαβ − Kα
where we have used

ciα ciβ = 6c2 δαβ . (5.2.32)
i
The momentum flux tensor in first order of reads

# $
(0)
(0) 1 j12 j1 j2
Pαβ := ciα ciβ Fi = + pδαβ (5.2.33)
i
ρ j1 j2 j22
kB T (0)
where p = ρ is the pressure (a detailed calculation of Pαβ will be given
m
below).
Thus to first order in we obtain the continuity equation
∂t ρ + ∇ · j = 0 (5.2.34)
and (in the incompressible limit: ρ = constant) the Euler equation (Navier-
Stokes equation without friction):
∂u 1
+ u∇u = − ∇p + K. (5.2.35)
∂t ρ
(0)
Calculation of the momentum flux tensor. In the calculation of Pαβ the
detailed form of the equilibrium distribution has to be taken into account for
(0.1) (0.4)
the first time. Four different summands Pαβ to Pαβ have to be evaluated.
The first term reads:
(0.1) ρ kB T
Pαβ := ciα ciβ Wi = ρ δαβ = pδαβ .
ρ0 i m
=p
kB T c2
For the pressure p = ρ = ρ the speed of sound cs is given by
m 3
1
dp c
cs = =√ .
dρ 3
The second summand

(0.2) m
Pαβ := ciα ciβ (ci · j)Wi = 0
ρ0 kB T i
vanishes because it is an odd moment in ci over Wi . The third summand

reads: !2
(0.3) 1 m 2
Pαβ := ciα ciβ (ci · j) Wi .
2ρ0 ρ kB T i
Prefactor: !2 !2
1 m kB T 1
· ρ0 =
2ρ0 ρ kB T m 2ρ
Tensor of fourth rank:
Tαβγδ = δαβ δγδ + δαγ δβδ + δαδ δβγ (5.2.36)

– α = β = 1:
1 2
* γ = δ = 1 → T1111 = 3 → 3j
2ρ 1
1 2
* γ = δ = 2 → T1122 = 1 → j
2ρ 2
* γ = δ → T1112 = T1121 = 0
1 2
– α = β = 2: → j1 + 3j22
2ρ
– α = 1 β = 2:
– γ = δ → T1211 = T1222 = 0
1
– γ = δ → T1212 = T1221 = 1 → · 2j1 j2
2ρ
# $
(0.3) 1 2 1 j12 j1 j2
Pαβ = j δαβ +
2ρ ρ j1 j2 j22
The fourth summand reads:

(0.4) 1 m 2
Pαβ := − j ciα ciβ Wi
2ρ0 ρ kB T i
1 2
= − j δαβ
2ρ
Summing up all terms yields

# $
(0) 1 j12 j1 j2
Pαβ = + pδαβ (5.2.37)
ρ j1 j2 j22
The pressure of the D2Q9-LBM does not depend explicitly on the flow speed
(0.4)
(see to the contrary the FHP model) because the summand Pαβ is exactly
(0.3)
compensated by a part of Pαβ .
Terms of second order in : mass. As for lattice-gas cellular automata

one has to take into account terms of second order in to derive the dissipative
terms of the macroscopic equations. Mass conservation leads to
(1)
0 = Ei
i
(1) (1) (2) (0) (1) ∆t (1) (1) (0)
= ∂t Fi + ∂t Fi + ciγ ∂x(1) Fi + ∂ ∂ F
i
γ
2 t t i
∆t
(1) (0) (0)
+∆tciγ ∂t ∂x(1)
γ
Fi ciβ ciγ ∂x(1)
β
∂x(1)
γ
+Fi (5.2.38)
2
ω (2) ciγ (1) ciγ
+ Fi − ∆t ∂ Kγ − ∆t ciβ ∂x(1) Kγ .
∆t 12c2 t 12c2 β
The following summands vanish:

(1)
(1)
∂t Fi = 0,

i (5.2.24)
(1)
∂x(1)
γ
ciγ Fi = 0,

i (5.2.24)
ω (2)
F = 0,

∆t i i
(5.2.24)
and
∆t (1)
2
∂t Kγ ciγ = 0.
12c i
The second term of Eq. (5.2.38) is the time derivative of the density:
(2)
(0) (2)
∂t Fi = ∂t ρ. (5.2.39)
i
The spatial gradient of the body force yields

∆t (1) ∆t
− ∂xα Kβ ciα ciβ = − ∂x(1) Kα . (5.2.40)
12c 2
i
2 α
In the transformation of the following summands the time derivatives are

substituted by spatial derivative using (5.2.34) and (5.2.31):
∆t (1) (1) (0) ∆t (1) (1) ∆t (1)
∂t ∂t Fi = ∂t ∂t ρ = − ∂t ∂x(1) β
ρuβ
2 i
2 2
∆t (1) (1) (0) ∆t (1)
= ∂ ∂ P − ∂ Kα
2 xα xβ αβ 2 xα
(1)
(0) (1)
∆t∂t ∂x(1)
α
ciα Fi = ∆t∂x(1)
α
∂t (ρuα )
i
(0)
= −∆t∂x(1)
α
∂x(1)
β
Pαβ + ∆t∂x(1)
α
Kα
∆t (1) (1) (0) ∆t (1) (1) (0)
∂ ∂ ciα ciβ Fi = ∂ ∂ P
2 xα xβ i 2 xα xβ αβ
(1)
The sum of these three terms yields ∆t∂xα Kα /2 which cancels with the term
derived from the body force gradient (Eq. 5.2.40), i.e. inclusion of the spatial
gradient of the body force is essential to ensure that there is no mass diffusion:
(2)
∂t ρ = 0. (5.2.41)
Terms of second order in : momentum. Conservation of momentum

leads to
(1)
0 = ciα Ei
i
(1) (1) (2) (0) (1) ∆t (1) (1) (0)
= ciα ∂t Fi + ciα ∂t Fi + ciα ciβ ∂x(1)
β
Fi + ciα ∂t ∂t Fi
i
2
(1)∆t (0) (0)
+∆tciα ciγ ∂t ∂x(1)
γ
Fi ciα ciβ ciγ ∂x(1)
β
+∂x(1)
γ
Fi
2
ω (2) ciγ (1) ∆t (1)
+ ciα Fi − ∆tciα ∂t Kγ − ciα ciβ ciγ ∂ Kγ .
∆t 12c 2 12c2 xβ
(1) (0)
An approximation of Fi can be derived from the first order in : Ei =0
⇒
(1) ∆t (1) (0) ∆t (0) ∆t ciγ
Fi (x, t) = − ∂t Fi − ciγ ∂x(1)
γ
Fi + Kγ , (5.2.42)
ω ω 6c2 ω
i.e. deviations from local equilibrium are driven by gradients in time and
space (compare Eq. 4.2.25) and by applied body forces.
In the transformation of the following summands terms of the order O(j 2 ) will
be neglected (indicated by ≈ instead of =). Thus the Navier-Stokes equation
will be recovered in the limit of low Mach numbers only.
(1)
(1)
∂t ciα Fi = 0

i (5.2.24)
(2)
(0) (2)
∂t ciα Fi = ∂t jα (5.2.43)
i
(1) ∆t (1) (1) (0)

∂x(1) ciα ciβ Fi =
− ∂ ∂ ciα ciβ Fi
β
i
ω t xβ i
(5.2.42)
(∗)
∆t (0)
− ∂x(1)
β
∂x(1)
γ
ciα ciβ ciγ Fi
ω i

(∗∗)
∆t (1)
+ 2
∂xβ Kγ ciα ciβ ciγ
6c ω i

=0
∆t (1) (1) (0) ∆t (1) (1)

∂ ∂ ciα Fi = ∂ ∂ jα
2 t t i
2 t t
∆t (1) (1) (0) ∆t (1)
= − ∂ ∂ P + ∂ Kα
2 t xβ αβ 2 t
∆t kB T (1) (1) ∆t (1)
≈ − ∂t ∂xβ ρδαβ + ∂ Kα
2 m 2 t
∆t kB T ∆t
⇒ ∇∇ · j + ∂t K (5.2.44)
2 m 2
∆t (1) (1) (0)

∂ ∂ ciα ciβ ciγ Fi
2 xβ xγ i

(∗∗)
(1)
(0)
∆t∂x(1)
β
∂t ciα ciβ Fi
i

(∗)
ω (2)
ciα Fi = 0

∆t i
(5.2.24)
The term derived from the time derivative of the body force
∆t (1) ∆t (1) ∆t
− ∂t K β ciα ciβ = − ∂t Kα ⇒ − ∂t K (5.2.45)
12c2 i
2 2
cancels with the term in Eq. (5.2.44), i.e. inclusion of the time derivative
of the body force is essential to obtain the exact form of the Navier-Stokes
equation.
∆t (1)
− 2
∂xβ Kγ ciα ciβ ciγ = 0
12c i
Summation of the (∗)-terms results in

! !
1 (1) (1) (0) 1 kB T
∆t 1 − ∂xβ ∂t ciα ciβ Fi ⇒ −∆t 1 − ∇∇ · j. (5.2.46)
ω i
ω m

kB T
≈ ρδαβ
m
The (∗∗)-terms lead to
!
1 1 (0)
∆t − ∂x(1)
β
∂x(1)
γ
ciα ciβ ciγ Fi
2 ω i
!
1 1 m 1 (1) (1)
= ∆t − ∂ ∂ Wi ciα ciβ ciγ (ci · j)
2 ω kB T ρ0 xβ xγ i
!
1 1 kB T 2
⇒ ∆t − ∇ j + 2∇∇ · j . (5.2.47)
2 ω m
By adding up (5.2.43 – 5.2.47) one finally obtains

!
(2) 1 1 kB T 2
∂t j = ∆t − ∇ j + ∇∇ · j , (5.2.48)
ω 2 m
i.e. dynamic shear and compression viscosity are equal:

!
1 1 kB T
µS = µK = ∆t − . (5.2.49)
ω 2 m
The viscosities can be tuned by an appropriate choice of the collision param-
eter ω. The equation for iterative solution of linear systems by the simultane-
ous over-relaxation (SOR) is of the same form as the BGK kinetic equation.
The SOR method is convergent for 0 < ω < 2 (Kahan, 1958). This range of
ω corresponds to positive viscosities in the BGK-LBM.
The sum of the first and second order terms yields (in the incompressible
limit) the Navier-Stokes equation:
∇·u=0 (5.2.50)
and
1
∂t u + (u∇) u = − ∇p + ν∇2 u + K (5.2.51)
ρ
with the kinematic shear viscosity

! !
1 1 c2 c2 ∆t
ν = ∆t − = τ− . (5.2.52)
ω 2 3 3 2
5.2.4 Storage demand
For the special case ω = 1 the kinetic equation reduces to

(0)
Fi (x + ci ∆t, t + ∆t) = Fi (x, t)
(eq)
and the storage demand for ρ, j, 9 distributions Fi and their propagated
values Fi , i.e. 21 2D arrays, seems to be quite large. The required memory
can, however, be drastically reduced by the following two procedures:
(eq)
1. One after the next Fi can be evaluated, propagated and will contribute
(independent of the other Fj ) to the values of ρ and j at the next time
level. Only 7 arrays, namely ρ and j at two time levels and one array Fi ,
are required, i.e. the reduction factor is 21/7 = 3.
2. Because the collisions are strictely local and the propagation is almost
local (to next neighbor nodes) the domain can be partitioned into subdo-
mains which can be updated one after the other using smaller auxiliary
arrays.
Procedure 1 leads to some increase in the length of the code but seems to be
worth the effort especially in 3D where the reduction factor is even larger.
The second procedure requires some additional coding to treat the bound-
aries of the subdomains but offers a further reduction in memory by a factor
of almost 2.
The storage demand for BGK-LBM with ω = 1 is substantially larger be-
cause one has to keep all (non-equilibrium) distributions Fi . The adventage
is the tunable viscosity which allows simulation of flows with high Reynolds
numbers.
5.2.5 Simulation of two-dimensional decaying turbulence
In 1994 Martinez et al. published a milestone paper. They simulated two-

dimensional decaying turbulence over a square with periodic boundary con-
ditions by the BGK lattice Boltzmann model over the D2Q9 lattice and
compared this method to a spectral model5 which is very efficient over this
simple domain. The initial velocity distribution consists of two shear layers
(realized by a truncated spectral representation of δ functions at y = π/2
and y = 3π/2) plus some noise over the whole domain. The characteristic
velocity U is defined as the root mean square value
"
U = u2 = 0.04 (5.2.53)
und the characteristic length L is given by
512
L= (5.2.54)
2π
where 512 is the number of grid points in each dimension. The viscosity
parameter ω has been chosen such that the Reynolds number
UL
Re = (5.2.55)
ν(ω)
5
The spectral model actually does not solve the Navier-Stokes equation but the
∂
vorticity equation +v·∇ = ν∇2 where = (∇ × v)z is the z-component
∂t
of the relative vorticity.
is 10000.
The following Figures (5.2.2 - 5.2.5) show isocontours of the vorticity a =
(∇ × u)z at four dimensionless times. Dashed lines indicate negative values.
The upper plots show the results of the spectral model and the lower plots
those of the lattice Boltzmann model. The features compare very well which
gives confidence in both methods.
What is most surprising: the LBM is as fast as the spectral model! And
the LBM keeps its efficiency in more complex geometries (porous media, for
example) whereas spectral models may not be applicable anymore. These
simulations clearly demonstrated the great potential of BGK-LBMs.
Fig. 5.2.2. Isovorticity contour plots for time 1. Dashed lines correspond to neg-
ative values of vorticity. The values for the contours are the same for all cases
(Figures 5.2.2 – 5.2.5). Strikingly similar features can be found for the lattice Boltz-
mann simulation, as compared with the spectral simulation. (SP refers to the spectral
sumulation and LBE to the lattice Boltzmann simulation; Martinez et al., 1994)
Fig. 5.2.3. Same as in Fig. 5.2.2 for time 5 (Martinez et al., 1994).
5.2.6 Boundary conditions for LBM
“To a certain degree, achieving self-consistent boundary conditions

with a given accuracy is as important as developing numerical
schemes themselves.”
Chen et al. (1996)
The implementation of boundary conditions for LGCA has been discussed

in Section 3.2.7. The same princles as outlined there can be applied also to
LBM. In the following a self-contained discussion of no-slip and slip boundary
conditions shall be given even risking the repetition of parts of Section 3.2.7.
LGCA and LBM seem to be very attractive because the apparent ease to
implement boundary conditions even in complicated geometry like porous
media. You often can find quotes like ‘difficult geometrical boundary con-
ditions are readily handled’ in the introduction of an article only to find
out that the authors performed some simulation over a square with periodic
boundary conditions. Much more realistic is the remark of He et al. (1997)
that “the real hydrodynamic boundary conditions have not been fully un-
derstood”. Accordingly various implementations of boundary conditions will
be discussed here. Some of them have to be ruled out because they are only
of first order of accuracy. You have to find out which of the remaining ones
gives satisfying results for the actual flow problem considered.
In general, there are two ways to define a boundary: the boundary curve may
include grid nodes (‘node boundary’; the nodes on the boundary are called
‘boundary nodes’; Skordos, 1993; Inamuro et al., 1995; Noble et al., 1995a,b,
1996) or passes through the midpoints of links between nodes (‘link bound-
ary’; Cornubert et al., 1991; Ginzbourg and Adler, 1994; Ladd, 1994a,b).
Node boundaries are appropriate for periodic and inflow boundary condi-
tions.
No-slip boundary condition. Inamuro et al. (1995) He et al. (1997)
1. The complete bounceback scheme: Instead of collision assign each Fi the

value of the Fi of its opposite direction, i.e. for a point on the lower
boundary:
in-state: F0 , F1 , F2 , F3 , F4 , F5 , F6 , F7 , F8
out-state: F0 , F1 , F4 , F3 , F2 , F7 , F8 , F5 , F6 .
2. The half-way wall bounceback: Consider a channel with periodic bound-

ary conditions in x-directions and walls at the lower and upper boundary.
The first and last nodes in y-direction, i.e. with indices j = 1 and j = N ,
are ‘dry’ (dry or wall nodes), i.e. outside the ‘wet’ domain which encom-
passes nodes with j = 2 to j = N − 1 (wet or interior or fluid nodes).
The lower (upper) boundary is located half-way between the first and
second (last but one and last) nodes in y-direction. Particle distributions
are propagated between wet and dry nodes and vice versa. On the dry
nodes no collision or forcing is performed and the distributions are all
bounced back. Note that the width of the channel is one unit smaller
than that with the complete bounceback scheme.
Plane Poiseuille flow: analytic solution. Only few analytical solutions

of the Navier-Stokes equation are known. One is the plane Poiseuille flow in
a channel of width 2L where the flow is steady (∂/∂t = 0), in x-direction u =
(u, v) = (u(y), 0), with constant pressure (p and ρ = constant) and without
variations in x-direction (∂/∂x = 0). The flow is driven by a constant force
K = Kex . Accordingly the Navier-Stokes equation reduces to an ordinary
differential equation for u(y)
d2 u
ν +K =0
dy 2
du(y)
and the continuity equation = 0 is satisfied. At the channel walls no-
dx
slip boundary conditions apply, i.e. u(y) = 0 at y = −L and y = L. The
analytical solution is a parabola
K 2
u(y) = L − y2 . (5.2.56)
2ν
Plane Poiseuille flow: numerical simulation. In the numerical simula-

tion the velocity is initialized to zero and the mass density to a constant value
of ρ = 1. In order to start the fluid flow a constant force has to be applied at
every time step. Here the microscopic forcing method is applied.
First the simplest no-slip scheme, namely bounceback will be used. The result
of an integration over a domain with 20 time 20 nodes is shown in Fig. 5.2.6
together with the analytical solution. The fluid is resting in the beginning and
then is slowly accelarated. After t = 1200 time steps the mean x-momentum
becomes steady (plot at lower right). The final velocity profile u(y) (+ in plot
at upper left) is compared with the analytical solution (Eq. 5.2.56) assuming
a channel width 2L = Y M AX − 1 = 19 where Y M AX = 20 is the number of
nodes in y-direction. The numerical solution is zero on the boundary points
(nodes j = 1 and j = Y M AX or y = ±9.5) and the profile looks similar to
a parabola. The most striking difference compared to the analytical solution
is the lower value of the maximum velocity. One can fit a parabola to the
numerical data with equal values of the maximum velocity (plot at upper
right). Deviations between this fit and the numerical values or most obvious
near the boundaries. The plot at the lower left shows the numerical solu-
tion together with the analytical solution but the latter one now based on a
smaller channel width 2L = Y M AX − 2 = 18. This solution compares very
well with the analytical solution except for the two boundary nodes. This fig-
ure suggests the following interpretation. The boundary is located half-way
between the first and second node respectively half-way between the second
to last and the last node. Therefore the ’wet’ channel has only a width of
2L = Y M AX − 2 = 18. The nodes at j = 1 and j = Y M AX are auxiliary
nodes which take care of the correct boundary conditions at j = 3/2 and
j = Y M AX − 1/2. The velocity values on these auxiliary nodes should not
be interpreted as flow velocities (these nodes are located ‘on land’ and could
by called ‘dry nodes’).
Fig. 5.2.6. No-slip boundary conditions and Poiseuille flow (see text for discus-
sion).
0.06 0.06
0.04 0.04
u(y)
u(y)
0.02 0.02
0 0
−10 0 10 −10 0 10
y y
0.06 0.03
0.04 0.02
j (t)
u(y)
0.02
x
0.01
0
0
−10 0 10 0 1000 2000
y Time t
Numerical experiments with different channel widths 2L = 16, 32, 64, 128 (dif-
ferent spatial resolutions of the parabola) have been performed while keeping
constant the viscosity ν and the product of channel width and maximum
speed (by varying the forcing). Thus the Reynolds number is kept constant.
The Mach number varies but for L ≥ 16 is small compared to 1. The error
has been calculated by

2
3N 2
13 (n)
Error := 4
(a)
ûi − ûi
N i=1
(n) (a)
where the summation is over N = 14 inner points and ûi and ûi are the
normalized numerical and analytical solutions with maximum speed equal to
one. Fig. (5.2.7) shows the error as a function of L. The slop of the curve is
close to −2 which indicates that this scheme is of second order in the spatial
discretization.
Fig. 5.2.7. Error of the numerical compared to the analytical solution of the
Poiseuille flow as a function of spatial resolution. The bounceback scheme is ap-
plied and the resulting data are interpretated such that the boundaries are located
between the first and second respectively between the second to last and the last node
(compare also Fig. 5.2.6C). The slop of the curve is close to −2 which indicates
that this scheme is of second order in the spatial discretization.
−3
10
slope = −2.1279
−4
10
Error
−5
10
−6
10 0 1 2
10 10 10
L
∂ut
Slip boundary condition. To test slip boundary condition ( = 0 at
∂xn
|y| → ∞, i.e. the normal derivative of the tangential velocity component ut
vanishes on the boundary) a shear layer with

−U for y < 0
u = , v=0
+U for y > 0
is initialized at t = 0. The analytical solution of the Navier-Stokes equation

in −∞ < y < ∞ for this initial condition is known:
!
y
u(y, t) = U erf √ , v=0
4νt
where erf (arg) is the error function.
The slip conditions have been implemented by reflection of the distribution
Fi on boundary nodes. The result of a numerical simulation (U = 0.1) in
a channel of finite width (2L = 18) are shown together with the analytical
solution in Fig. 5.2.8. The implementation of the simple slip scheme works
very well. Small deviations between numerical and analytical solution are to
be expected due to the finite width of the channel in the simulation.
Fig. 5.2.8. Shear layer flow in a channel: test of slip boundary conditions (see
text).
0.1
0.08
0.06
0.04
0.02
u(y)
0
−0.02
−0.04
−0.06
−0.08
−0.1
−10 −5 0 5 10
y
Further reading: boundary conditions for LBM

Skordos (1993), Ziegler (1993), Ginzbourg and Adler (1994), He and Zou
(1995), Noble et al. (1995a,b), Inamuro et al. (1995), Noble et al. (1996),
Chen et al. (1996), Ginzbourg and d’Humiéres (1996), Maier et al. (1996),
Filippova and Hänel (1997), Gallivan et al. (1997), Stockman et al. (1997),
Kandhai et al. (1999).
5.3 Hydrodynamic lattice Boltzmann models in 3D 195
5.3 Hydrodynamic lattice Boltzmann models in 3D
BGK-LBMs have been discussed so far only in two dimensions. The extension
to three dimensions is straightforward. One just has to choose a 3D lattice
and calculate appropriate equilibrium distributions. In this section we will
calculate equilibrium distributions by the method of Koelman (1991) for 3D
lattices with 19 and 15 velocities.
(eq)
The derivation of the equilibrium distribution functions Fi (5.2.11) by
minimizing the functional (5.2.18) is independent of the lattice velocities ci
and the dimension D. Only the constraints (5.2.6 to 5.2.8) for the veloc-
ity moments were used in Section (5.2). Thus one can use the form of the
(eq)
equilibrium distribution functions Fi (5.2.11) with the appropriate lattice
velocities ci and weights Wi .
5.3.1 3D-LBM with 19 velocities
For hydrodynamic simulation d’Humières et al. (1986) proposed a multi-

speed lattice-gas cellular automata over a cubic lattice with 19 velocities
(compare Section 3.3). This cubic lattice D3Q19 is defined by the following
velocities
c0 = (0, 0)
c1,2 , c3,4 , c5,6 = (±1, 0, 0), (0, ±1, 0) (0, 0, ±1)
c7,...,10 , c11,...,14 , c15,...,18 = (±1, ±1, 0), (±1, 0, ±1), (0, ±1, ±1).
The calculation of the Wi for the 3D model proceeds in close analogy with
the 2D case. There are three different speeds: 1 √
time√speed 0 (rest particle),
6 times speed 1 (1-particles) and 12 times speed 2 ( 2-particles). The even
moments yield four independent equations for the calculation of W0 , W1 , W2
and kb T /m:
– 0. moment:
Wi = W0 + 6W1 + 12W2 = ρ0
i
– 2. moment: kT
c2αx Wi = 2W1 + 8W2 = ρ0
i
m
– 4. moment:
!2
kT
c4αx Wi = 2W1 + 8W2 = 3ρ0
i
m
!2
kT
c2αx c2αy Wi = 4W2 = ρ0
i
m
The solution reads

ρ0
W0 =
3
!2
ρ0 kT ρ0
W1 = =
2 m 18
!2
ρ0 kT ρ0
W2 = = (5.3.1)
4 m 36
kT 1
= .
m 3
The odd velocity moments over Wi vanish.
5.3.2 3D-LBM with 15 velocities and Koelman distribution
The lattice velocities of the D3Q15 lattice read
c0 = (0, 0, 0) rest particle

c1,2 , c3,4 , c5,6 = (±2, 0, 0), (0, ±2, 0) (0, 0, ±2) 2-particles
√
c7,...,14 = (±1, ±1, ±1) 3-particles.
There are three

√ different speeds: 1 time speed 0, 6 times speed 2 and 8
times speed 3. The even moments yield four independent equations for the
calculation of W0 , W2 , W3 and kb T /m:
– 0. moment:
Wi = W0 + 6W2 + 8W3 = ρ0
i
– 2. moment:
kT
c2αx Wi = 8W2 + 8W3 = ρ0
i
m
– 4. moment: !2
kT
c4αx Wi = 32W2 + 8W3 = 3ρ0
i
m
!2
kT
c2αx c2αy Wi = 8W3 = ρ0
i
m
5.3 Hydrodynamic lattice Boltzmann models in 3D 197
The solution reads

7
W0 = ρ0
18
!2
ρ0 kT ρ0
W2 = =
16 m 36
!2
ρ0 kT ρ0
W3 = =
8 m 18
kT 2
= .
m 3
The odd velocity moments over Wi vanish.
5.3.3 3D-LBM with 15 velocities proposed by Chen et al. (D3Q15)
The equilibrium distributions are not unique and other choices are possible.
As an example we give the equilibrium distributions proposed by Chen et al.
(1992):
F0eq = d(0) + δ (0) v 2

Fieq = d(1) + β (1) ci · v + γ (1) (ci · v)2 + δ (1) v 2 α = 1, ..., 6
Fieq = d (2)
+β (2)
ci · v + γ (2)
(ci · v) + δ
2
v
(2) 2
α = 7, ..., 14
where
ρ ρ
d(0) = d(1) = , d(2) =
11 22
ρ ρ
α(1) = , α(2) =
24 12
ρ ρ
γ (1) = , γ (2) =
32 16
7 ρ ρ
δ (0) = − ρ, δ (1) =− , δ (2) = − .
24 48 24
These distributions will not be considered in the following.

5.4 Equilibrium distributions: the ansatz method
“The question that we are most often asked about cellular automata
is the following.
‘I’ve been shown cellular automata that make surprisingly good mod-
els of, say, hydrodynamics, heat conduction, wave scattering, flow
through porous media, nucleation, dendritic growth, phase separa-
tion, etc. But I’m left with the impression that these are all ad hoc
models, arrived at by some sort of magic.’
‘I’m a scientist, not a magician. Are there well-established correspon-
dence rules that I can use to translate features of the system I want to
model into specifications for an adequate cellular-automaton model
of it?’
Physical modeling with cellular automata is a young discipline. Sim-
ilar questions were certainly asked when differential equations were
new - and, despite three centuries of accumulated experience, mod-
eling with differential equations still requires a bit of magic.”
Toffoli und Margolus (1990)
The problem addressed for CA by Toffoli and Margolus exists also for lattice
Boltzmann models. However, in what follows it will be shown that a LBM
for the Navier-Stokes equation can be developed by an almost straightfor-
ward method. In chapter 5.2 global equilibrium functions Wi for vanishing
velocity (u = 0) and constant density ρ = ρ0 were determined in close anal-
ogy to the Maxwell distribution (more precise: the velocity moments of Wi
up to fourth order are equal to the corresponding velocity moments over
the Maxwell distribution; compare Eqs. 5.2.6 - 5.2.8). Subsequently the local
equilibrium distributions Fi have been derived using the maximum entropy
principle. These ‘Koelman-distributions’ lead to the Navier-Stokes equation
with isotropic advection, Galilean invariance and pressure which does not
explicitly depend on flow speed.
To be sure, these are not the only distributions over that lattice which in the
macroscopic limit yield the Navier-Stokes equation. This is not a problem
because the equilibrium distributions of our artificial microworld are not of
interest by themselves.
In the current section I will discuss an alternative approach to suitable equi-
librium distributions Fi whereby an ansatz for Fi will be used. After the
multiscale analysis the free parameters in the ansatz will be chosen such that
isotropy etc. are assured. This alternative approach will be of central impor-
tance for the development of lattice Boltzmann models for given differential
equations.
5.4 Equilibrium distributions: the ansatz method 199
To keep things simple the presentation will be restricted to a twodimensional

model. The extension to 3D is straightforward. An LBM is defined by three
ingredients:
1. A kinetic equation: here the meanwhile well-established BGK equation

is chosen.
2. A lattice with sufficient symmetry: in order to compare results with those
of the previous chapter the D2Q9 lattice is chosen; for a model based on
D2Q7 see exercise (5.4.3).
3. Equilibrium distributions: see below.
Instead of using the maximum entropy principle to derive equilibrium dis-

tributions one may propose an ansatz with free parameters. How should the
ansatz look like?
1. From lattice-gas cellular automata (FHP, FCHC, PI) and the Koelman
model we know that terms quadratic in u and ci ·u in the equilibrium dis-
tributions yield the nonlinear advection term of the Navier-Stokes equa-
tion; no higher moments are required.
2. The free parameters of the ansatz should depend only on the mass density
and the speeds |ci | but not on the directions of the ci .
This suggests the following ansatz with ten free parameters:
Fi = A0 + D0 u2 i=0
Fi = A1 + B1 ci · u + C1 (ci · u)2 + D1 u2 i = 1, 2, 3, 4 (5.4.1)
Fi = A2 + B2 ci · u + C2 (ci · u)2 + D2 u2 i = 5, 6, 7, 8.
The definitions of mass and momentum densities give three scalar constraints

ρ := Fi = A0 + 4(A1 + A2 ) +u2 (D0 + 4D1 + 4D2 + 2C1 + 4C2 ) (5.4.2)

i
=ρ =0

j := ci Fi = (2B1 + 4B2 ) u. (5.4.3)

i
=ρ
5.4.1 Multi-scale analysis
The multi-scale analysis proceeds as in Section 5.2.3 up to Eq. (5.2.31) inclu-

(0)
sively. Not until the calculation of the momentum flux tensor Pαβ the specific
form of the equilibrium distribution has to be taken into account:
(0)
(0)
Pαβ := ciα ciβ Fi
i

4
' (
= ciα ciβ A1 + B1 ci · u + C1 (ci · u)2 + D1 u2
i=1

8
' (
+ ciα ciβ A2 + B2 ci · u + C2 (ci · u)2 + D2 u2
i=5

= (2A1 + 4A2 ) δαβ + 2C1 u2α + 4C2 u2 δαβ
+8C2 uα uβ (1 − δαβ ) + (2D1 + 4D2 ) u2 δαβ
with u = (u, v) = (u1 , u2 ). The goal is to transform the tensor

 
2A1 + 4A2 + 2C1 u2 8C2 uv
 
 +(4C2 + 2D1 + 4D2 )u2 
 
Pαβ =  
(0)
 
 
 8C2 uv 2A1 + 4A2 + 2C1 v 2 
+(4C2 + 2D1 + 4D2 )u2
into
# $
u2 uv
ρ + p δαβ .
uv v2
The portion of the tensor which is independent of the flow velocity u yields
the pressure p:
p = 2A1 + 4A2 .
All other terms lead to the following constraints for the free parameters
(0)
4C2 + 2D1 + 4D2 = 0 (u2 term in P11 must vanish)
(0)
2C1 = ρ (u2 term in P11 must yield ρu2 )
(0)
8C2 = ρ (uv term in P12 must yield ρuv)
and therefore
ρ ρ
C1 =and C2 = .
2 8
The remaining eight unknowns (A0 , A1 , A2 , B1 , B2 , D0 , D1 , D2 ) are con-
strained by only four linear equations
A0 + 4(A1 + A2 ) = ρ
2B1 + 4B2 = ρ
3
D0 + 4D1 + 4D2 = −2C1 − 4C2 = − ρ
2
ρ
2D1 + 4D2 = − .
2
Therefore in addition some arbitrary restrictions can be imposed. Here is just

one possibility:
A0 A1 B1 D0
= = = =: r (5.4.4)
A1 A2 B2 D1
where r is a free parameter. Thus the coefficients are functions of r:
r2 r 1
A0 = ρ, A1 = ρ, A2 = ρ
(r + 2)2 (r + 2)2 (r + 2)2
r 1
B1 = ρ, B2 = ρ
4 + 2r 4 + 2r
r 1 r−2
D0 = − ρ, D1 = − ρ, D2 = − ρ
2+r 2+r 16 + 8r
2
p= ρ = c2s ρ (5.4.5)
r+2
The speed of sound 1 5
dp 2
cs = =
dρ r+2
is tunable by the parameter r (compare Fig. 5.4.1).
The additional restriction
D1
=r
D2
leads to a quadratic equation for r with solutions r1 = 4 and r2 = −2.
The solution r2 yields negative equilibrium distribution even at small Mach
numbers and an infinite pressure. The solution r = 4 (by the way: C1 /C2
also equals 4) leads to
4 1 1 1 1
A0 = ρ, A1 = ρ, A2 = ρ, B1 = ρ, B2 = ρ,
9 9 36 3 12
2 1 1 1
D0 = − ρ, D1 = − ρ, D2 = − ρ, p= ρ = c2s ρ
3 6 24 3
with the speed of sound
1
cs = √ .
3
The equilibrium distribution read
Fig. 5.4.1. The speed of sound cs as a function of the parameter r.

2
1.8
1.6
s
Speed of sound c
1.4
1.2
0.8
0.6
0.4
−2 0 2 4 6 8 10
r
% &
4 3
Fi = ρ 1 − u2 i=0
9 2
% &
1 9 3 2
Fi = ρ 1 + 3ci · u + (ci · u)2 − u i = 1, 2, 3, 4
9 2 2
% &
1 9 3 2
Fi = ρ 1 + 3ci · u + (ci · u)2 − u i = 5, 6, 7, 8.
36 2 2
The distribution functions are identical to those derived by the maximum

entropy principle (compare Eq. 5.2.12 with c = 1)! These equilibrium distri-
butions were also used by Martinez et al. (1994).
In summary, the equilibrium distribution have been derived in the current
section by adjusting the free parameters of a plausible ansatz after the multi-
scale analysis such that the desired macroscopic equations will result. For the
model discussed above, there are not enough constraints to derive a unique
solution for all free parameters. This freedom can be used, for example, to
tune the speed of sound. The equilibrium distribution derived by the maxi-
mum entropy principle is a special case of the above calculated solutions.
The alternative approach to derive equilibrium distributions opens up the
prospect to develop lattice Boltzmann models for other differential equations.
From an ansatz for the equilibrium distributions and the conservation laws
the desired properties of the macroscopic equations can be implemented into
the microscopic model by appropriate choice of the free parameter after the
multi-scale analysis. The ansatz method will be applied in the development
of thermal LBM (Section 5.5) and LBM for the diffusion equation (Section
5.8).
5.4.2 Negative distribution functions at high speed of sound
Eq. (5.4.5) suggests that the speed of sound cs might be tuned to arbitrary
values by variing the parameter r. For small or even negative r, however, the
distribution functions Fsi become negative which leads to numerical instabil-
ity very quickly. Let rc be the critical r where one of the Fsi first vanishes. It
can be shown (see Exercise 5.4.2) that rc depends on the velocity u as follows
2u2
rc = (5.4.6)
1 − u2
and therefore "
cs < cs (rc ) = 1 − u2 . (5.4.7)
The Navier-Stokes equation has been derived in the small Mach number limit.
Thus
|u| |u|
Ma := =√ (5.4.8)
cs 1 − u2
should remain small compared to 1. Consequently u2 must remain small

compared to 1 (singularity of M at u2 = 1) and therefore cs < 1 (compare
Fig. 5.4.2).
Fig. 5.4.2. The Mach number Ma and the speed of sound cs as a function of the
speed |u|.
1.4
1.2
Ma (solid line) and cs (dashed line)
0.8
0.6
0.4
0.2
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Speed | u |
Exercise 5.4.1. (***)

Show that even when skipping the auxiliary constraints (5.4.4) while keeping
Fsi > 0 the speed of sound cannot exceed 1.

Derive eq. (5.4.6).
Exercise 5.4.3. (***)

Calculate equilibrium distributions for an LBM for the Navier-Stokes equa-
tion in 2D over the D2Q7 lattice (FHP with rest particles). This model is
called pressure-corrected LBM (PCLBM) because the pressure does not de-
pend explicitly on the flow speed.
5.5 Hydrodynamic LBM with energy equation 205
5.5 Hydrodynamic LBM with energy equation
Using the ansatz method described in Section 5.4 it is possible to develop hy-
drodynamic LBMs including an energy equation (so-called thermal models).
Alexander, Chen and Sterling (1993) proposed a thermal LBM over the
D2Q13-FHP lattice (multi-speed FHP) with the following lattice velocities
ci (compare Fig. 3.3.8):
ci = (0, 0) ! i=0
2πk 2πk
ci = cos , sin i = 1, 2, ..., 6; k=i
6 6 !
2πk 2πk
ci = 2 cos , sin i = 7, 8, ..., 12; k =i−6
6 6
For the equilibrium distributions they made an ansatz including terms up to

third order in the flow velocity u
Fieq = Aσ + Bσ ci · u + Cσ (ci · u)2 + Dσ u2 + Eσ (ci · u)3 + Gσ (ci · u)u2 , (5.5.1)
whereby the 14 free parameters Aσ to Gσ depend only on mass and internal
energy density. The index σ refers to the square of the speed which is equal
to zero, one, or two.
The BGK kinetic equation
Fi (x + ci , t + 1) = (1 − ω)Fi + ωFieq (5.5.2)
is applied.
Mass density (ρ), velocity (u) and internal energy (εI ) are defined as follows

ρ = Fi (5.5.3)
i

ρu = c i Fi (5.5.4)
i
(ci − u)2
ρεI = Fi . (5.5.5)
i
2
The desired form of the macroscopic equations

∂ρ ∂
+ (ρuα ) = 0 (5.5.6)
∂t ∂xα
! % !&
∂uα ∂uα ∂p ∂ ∂uγ ∂ ∂uβ ∂uα
ρ +ρuβ =− + λ + µ + (5.5.7)
∂t ∂xβ ∂xα ∂xα ∂xγ ∂xβ ∂xα ∂xβ
! ! !2
∂εI ∂εI ∂uγ ∂ ∂T ∂uβ ∂uα ∂uβ ∂uγ
ρ +ρuβ = −p + κ +µ + +λ
∂t ∂xβ ∂xγ ∂xβ ∂xβ ∂xα ∂xβ ∂xα ∂xγ
(5.5.8)
is obtained by the following choice for the coefficients (see Appendix 6.4)
!
5 4 1
A0 = ρ 1 − εI + 2ε2I , A1 = ρ εI − ε2I , A2 = ρ −εI + 4ε2I
2 9 36
(5.5.9)
4 1
B1 = ρ (1 − εI ) , B2 = ρ (−1 + 4εI ) (5.5.10)
9 36
4 1
C1 = ρ (2 − 3εI ) , C2 = ρ (−1 + 6εI ) (5.5.11)
9 72
1 2 1
D0 = ρ (−5 + 8εI ) , D1 = ρ (−1 + εI ) , D2 = ρ (1 − 4εI ) (5.5.12)
4 9 72
4 1
E1 = −ρ , E2 = ρ (5.5.13)
27 108
G1 = G2 = 0. (5.5.14)
Alexander et al. (1993) give no statement concerning the uniqueness of the
solution. The shear viscosity µ and the thermal conductivity κ are given by
!
1 1
µ = ρεI − (5.5.15)
ω 2
!
1 1
κ = 2ρεI − (5.5.16)
ω 2
(the compressional viscosity λ vanishes). Thus the Prandtl number is
µ 1
P r := = . (5.5.17)
κ 2
The pressure p reads p = ρεI . The temperature T is proportional to the
internal energy εI .

Is it possible to define a thermal LBM over the D2Q9 lattice?
Exercise 5.5.2. (***)

Is it possible to choose the coefficients such that the Prandtl number is 1 or
even becomes a tunable parameter?
5.5 Hydrodynamic LBM with energy equation 207
Further reading:
Chen et al. (1995a,b) introduced an additional parameter in order to tune the
Prandtl number. McNamara et al. (1995) discussed numerical instabilities of
the models proposed by Alexander et al. (1993) and Chen et al. (1995a,b).
Vahala et al. (1996) applied the model of Alexander et al. (1993) to free-
decaying turbulence in 2D. McNamara et al. (1997) proposed a 3D thermal
LBM with 27 velocities. They apply a Lax-Wendroff scheme in order to im-
prove numerical stability. They performed simulations of Rayleigh-Bénard
convection in 2D and compared the results with those of an explicit finite-
difference (FD) solver. The computer times of LBM and FD are comparable
whereas LBM requires significantly more memory and stability is significantly
poorer. Thus they conclude that currently there is no potential advantage in
using a thermal LBM over a conventional FD solver. Hu et al. (1997) con-
sider the energy levels, εσ , of the three different speeds as free parameters and
thereby are able to tune the ratio of specific heats, γ. Shan (1997) proposed a
two-component LBGK in which temperature is advected as an passive scalar
and simulated Rayleigh-Bénard convection. Pavlo et al. (1998a,b) investigate
linear stability of thermal LBMs. The model of Alexander et al. (1993) has
been revisited by Boghosian and Coveney (1998). They showed “that it is
possible to achieve variable (albeit density-dependent) Prandtl number even
within a single-relaxation-time lattice-BGK model”. He et al. (1998), Vahala
et al. (1998b).
5.6 Stability of lattice Boltzmann models
Lattice Boltzmann schemes do not have an H-theorem and therefore are

subject to numerical instabilities (Sterling and Chen, 1996). Linear stability
analysis has been performed for various LB models (D2Q7, D2Q9, D3Q15)
and different background flows (homogeneous or shear flow) by Sterling and
Chen (1996) and Worthing et al. (1997). The stability does not only depend
on the background flow but also on the mass fraction parameters (α, β) of the
equilibrium distributions and the grid size (stability decreases with increasing
grid size).
5.6.1 Nonlinear stability analysis of uniform flows
The time evolution of BGK lattice Boltzmann models is described by the

kinetic equation

Fm (x + cm , t + 1) = Fm (x, t) − ω Fm (x, t) − Fm
(0)
(x, t) (5.6.1)
Inspection of Eq. (5.6.1) shows that an initially uniform flow, in the sense
Fm (x, t0 ) = Fm (t0 ),
will remain uniform at all later times and thus

Fm (t + 1) = Fm (t) − ω Fm (t) − Fm
(0)
(t) .
Furthermore, mass and momentum density are conserved and retain their
initial value
ρ(x, t) = ρ(t0 ) = ρ0 , j(x, t) = j(t0 ) = j 0
and therefore

Fm (t + 1) = Fm (t) − ω Fm (t) − Fm
(0)
(ρ0 , j 0 ) .
(0)
Subtracting Fm (ρ0 , j 0 ) on both sides leads to
F̃m (t + 1) = (1 − ω)F̃m (t)

(0)
with F̃m (t) = Fm (t) − Fm (ρ0 , j 0 ). The evolution is stable in the sense that
the magnitude of F̃m does not increase with time (|F̃m (t + 1)| ≤ |F̃m (t)|) if
1 1
|1 − ω| < 1 → 0<ω<2 or =τ >
ω 2
(compare Fig. 5.6.1).
5.6 Stability of lattice Boltzmann models 209
Fig. 5.6.1. Stability range of BGK models.
∞ τ 0.5
2
1.8
1.6
1.4
|1−ω| = |1−1/τ|
1.2
0.8
0.6
0.4
0.2
0
−1 −0.5 0 0.5 1 1.5 2 2.5 3
ω

Consider the kinetic equation

(0)
Fm (x + cm , t + 1) = Fm (x, t) + Ωik Fk (x, t) − Fk (x, t)
and show that for uniform flows the following equation holds
F̃ (t + 1) = [I + Ω]F̃ (t)
(0)
where the vector F̃ (t) has components F̃m (t) = Fm (t) − Fm (ρ0 , j 0 ).
In reality, a flow is never uniform in the strict sense because some noise
is always around. Therefore we have to investigate whether small spatially
varying perturbations grow or are damped with time. No general methods are
known that allow stability analysis of arbitrary nonlinear systems. In each
case you have to find special tricks (like, for instance, the construction of a
Liapunov function). To get an idea of the stability properties of the nonlinear
system one usually expands the system up to linear terms and investigates the
stability of the resulting linear system. Before attacking the kinetic equation
(5.6.1) we will discuss the von Neumann stability analysis for two simpler
equations.
5.6.2 The method of linear stability analysis (von Neumann)
The Fourier method of John von Neumann is the standard method for sta-
bility analysis because it is applicable to arbitrary linear systems. As a first
example we will consider the linear advection equation
∂u ∂u
+c = 0, c>0
∂t ∂x
which shall be solved numerically by the following finite difference scheme
uj,n+1 = uj,n − µ (uj,n − uj−1,n ) (5.6.2)
where uj,n = u(xj , tn ), xj = j∆x, tn = n∆t, and µ = c · ∆t/∆x. For the

difference equation (5.6.2) we make the following ansatz (‘poor man’s Fourier
transform’)
uj,n = Un eikx = Un eikj∆x (5.6.3)
(the actual solution is the real part) where k is the wave number of the spatial
perturbation. Inserting of (5.6.3) into (5.6.2) leads to
Un+1 = (1 − µ)Un + µUn e−ik∆x = λUn
with
λ := 1 − µ + µe−ik∆x .
It follows that
|Un+1 | = |λ| · |Un | = |λ|n · |U1 |.
Stability in the sense that |Un+1 | is bounded (i.e. limn→∞ |Un+1 | < B < ∞)
yields the constraint
|λ| ≤ 1.
Here we have
|λ|2 = 1 − 2µ(1 − µ)(1 − cos k∆x)
which is ≤ 1 when
∆t
0≤µ=c ≤1
∆x
or
∆x
≥c (5.6.4)
∆t
or
∆x
0 ≤ ∆t ≤ . (5.6.5)
c
This is the stability condition we were looking for. According to Eq. (5.6.4)
the ‘grid speed’ cgrid := ∆x/∆t has to be as fast or faster than the advection
speed c (this holds in very similar form also for several spatial dimensions,
for other explicit numerical schemes and other hyperbolic equations whereby
the advection speed is eventually replaced by the speed of the fastes wave of
the system).
As a second example we consider the diffusion equation
∂T ∂2T
=κ 2.
∂t ∂x
A finite difference approximation reads
∆t
Tj,n+1 = Tj,n + κ 2 (Tj+1,n − 2Tj,n + Tj−1,n ) .
(∆x)
Inserting the ansatz
Tj,n = An eikj∆x
leads to
An+1 = An + µ An eik∆x − 2 + e−ik∆x = λ An
with
∆t
µ = κ 2
(∆x)
and
λ = 1 − 2µ (1 − cos k∆x)

The condition |λ | ≤ 1 yields the stability condition
∆t 1
0 ≤ µ = κ 2 ≤
(∆x) 2
or
2
(∆x)
0 ≤ ∆t ≤ . (5.6.6)
2κ
5.6.3 Linear stability analysis of BGK lattice Boltzmann models
The following discussion is based on Worthing et al. (1997). The kinetic

equation including external body forces Km (x, t) reads

Fm (x + cm , t + 1) = Fm (x, t) − ω Fm (x, t) − Fm
(0)
(x, t) + Km (x, t) (5.6.7)
(0)
where Fm are the distribution functions and Fm are the equilibrium distri-
bution functions. Expansion about time-independent but otherwise arbitrary
(bf )
distribution functions Fm (‘background flow’) leads to
(bf )
Fm (x, t) = Fm (x) + fm (x, t)
Fm (x + cm , t + 1) = Fm
(bf )
(x + cm ) + fm (x + cm , t + 1)
(0) (0)
Fm (x, t) = Fm ({Fs (x, t)})
# $
∂F (0)
m
≈ Fm {Fs (x)} +
(0) (bf )
fn (x, t)
∂Fn
(x)}
(bf )
n {Fs
Km (x, t) = Km ({Fs (x, t)})
!
∂Km
≈ Km {Fs(bf ) (x)} + fn (x, t)
n
∂Fn (bf )
{Fs (x)}
Inserting the expansions into the kinetic equation (5.6.7) results in

fm (x + cm , t + 1) = Gm (x) + (1 − ω)fm (x, t) + Jmn (x)fn (x, t)
n
where

Gm (x) (bf )
= Fm (x) − Fm (bf )
(x + cm ) − ω Fm
(bf )
(x) − Fm
(0)
{Fs(bf ) (x)}

(bf )
+Km Fm (x)
and + ,
(0)
∂Fm ∂Km
Jmn = ω + . (5.6.8)
∂Fn ∂Fn (bf )
{Fs ( x)}
Here we are interested only in instabilities with exponential growth. Therefore
we will neglect the time-independent term Gm (x) which can lead to linear
growth at most. The rectangular domain of length L and width W comprises
a lattice with nodes at (xr = 0, 1, 2, ..., L; ys = 0, 1, 2, ..., W ). We will now
Fourier transform the linear equation

fm (x + cm , t + 1) = (1 − ω)fm (x, t) + Jmn (x)fn (x, t). (5.6.9)
n
(k,l)
The functions fm (t) and fm (x, y, t) are connected by the Fourier transform

(k,l)
fm (x, y, t) = fm (t)eikx2π/L eily2π/W . (5.6.10)
k,l
Consequently one obtains
fm (x + cx,m , y + cy,m , t + 1)

(k,l)
= fm (t + 1)eik(x+cx,m )2π/L eil(y+cy,m )2π/W
k,l

(k,l)
= fm (t + 1)eikx2π/L eily2π/W eikcx,m 2π/L eilcy,m 2π/W
k,l
The analysis is further simplified by the assumption that the background

flow (a shear layer or jet stream, for example) as well as the external forces
depend only on y:
(q) iqy2π/W
Jmn (y) = Jmn e .
q
Inserting of the Fourier transforms into Eq. (5.6.9) one obtains

eikcx,m 2π/L eilcy,m 2π/W fm
(k,l)
(t + 1) = (1 − ω)fm
(k,l)
(t) + (q) (k,l−q)
Jmn fn (t)
n,q
Since the x modes remain uncoupled, they are considered independently via

eikcx,m 2π/L eilcy,m 2π/W fm
(l)
(t + 1) = (1 − ω)fm
(l)
(t) + (q) (l−q)
Jmn fn (t) (5.6.11)
n,q
It can be shown (for some fine points see Worthing et al., 1997) that Eq.
(5.6.11) can be written as a matrix iteration
f t+1 = Af t (5.6.12)
(l)
where the vector f has components fm . If the spectral radius (= magnitude
of the largest eigenvalue) of A, ρ(A), is larger than unity, then the system
is said to be linear unstable (this corresponds to the case |λ| > 1 in the two
examples with one component each discussed in Subsection 5.6.2). In case of
uniform background flow and no external forcing considered by Sterling and
Chen (1996) the matrix A reads
A = D [(1 − ω) I + J ] (5.6.13)
where D is a diagonal matrix with components

Dmn = e−ikcx,m 2π/L e−ilcy,m 2π/W δmn , (5.6.14)
I is the identity matrix (diagonal matrix with value 1 of all diagonal compo-
nents) and the components of J are given by
(0)
∂Fm
Jmn = ω . (5.6.15)
∂Fn
The matrix A (and therefore also its spectral radius ρ(A)) depends on the
relative wave numbers (θx = 2πk/L, θy = 2πl/W ), the uniform inital velocity
(U ), the collision parameter (ω) and the rest mass parameters (α, β; see
below). Instability occurs if the maximal spectral radius (for given values of
U , ω, α and β) becomes larger than 1 for any wave number:
max ρ (A [θx , θy , U, ω, α, β]) > 1. (5.6.16)

θx ,θy
The components of J may be calculated as follows. The equilibrium distri-

butions for the models D2Q7, D2Q9 and D3Q15 are all of the form
' (
Fm(0)
= ρ Am + Bm cm · u + Cm (cm · u)2 + Dm u2
with constants Am , ..., Dm . The derivatives of mass and momentum density

with respect to Fn are easy to calculate:
∂ρ
ρ= Fn , → = 1,
n
∂Fm
∂j
j = ρu = Fn c n , → = cn
n
∂Fn
and consequently
∂j ∂ρ 2
2j ρ− j
j2 ∂(ρu2 ) ∂Fn ∂Fn
ρu = ,2
→ = = 2cn · u − u2 .
ρ ∂Fn ρ2
Finally we obtain
(0)
∂Fm ' (
= Am + Bm cm · cn + Cm 2(cm · cn )(cm · u) − (cm · u)2
∂Fn
' (
+Dm 2cn · u − u2 .
The coefficients for the various models are listed below. The coefficients Am
include free parameters α and β (denoted as rest mass parameters) which
can be tuned in order to improve stability.
FHP (D2Q7).
Am = α, Bm = 0, Cm = 0, Dm = −1; m=0
1−α
Am = , Bm = 13 , Cm = 23 , Dm = − 16 ; m = 1, ..., 6.
6
D2Q9.
Am = α, Bm = 0, Cm = 0, Dm = − 23 ; m=0
Am = β, Bm = 13 , Cm = 12 , Dm = − 16 ; m = 1, ..., 4
1 − α − 4β
Am = , Bm = 1
12 , Cm = 18 , Dm = − 24
1
; m = 5, ..., 8
4
D3Q15.
m=0:
Am = α, Bm = 0, Cm = 0, Dm = − 12 .
m = 1, ..., 6 :
Am = β, Bm = 13 , Cm = 12 , Dm = − 16 .
m = 7, ..., 14 :
1 − 6β − α
Am = , Bm = 1
24 , Cm = 1
16 , Dm = − 48
1
.
8
The eigenvalues of A have been calculated with a standard routine form

MATLAB. The maximum eigenvalues for the D2Q7 (FHP) model with 1.)
(u, v) = (0.2, 0), α = 0.2, θy = 0 (solid line) and 2.) (u, v) = (0.23, 0), α = 0.3,
θy = 0 (broken line) are shown in Fig. 5.6.2 as a function of the relative wave
number θx (the figure is identical to Figure 1 in Sterling and Chen (1996); this
problem served as a test of the MATLAB script). For certain wave numbers
the maximum eigenvalues become larger than unity and therefore the model
is linear unstable for this choice of parameters. Detailed results on stability
boundaries depending on the model parameters can be found in Sterling and
Chen (1996) and Worthing et al. (1997).
5.6.4 Summary
The main results of Sterling and Chen (1996) and Worthing et al. (1997) can
be summarized as follows:
– For the D2Q7 (FHP) model and homogeneous flow the wavenumber vector
k of the most unstable mode is parallel to the velocity u.
Fig. 5.6.2. The maximum eigenvalues for the D2Q7 (FHP) model with 1.) (u, v) =
(0.2, 0), α = 0.2, θy = 0 (solid line) and 2.) (u, v) = (0.23, 0), α = 0.3, θy = 0
(broken line) are shown as a function of the relative wave number θx (compare Fig.
1 in Sterling and Chen, 1996).
1.2
1.18
1.16
Maximum eigenvalue
1.14
1.12
1.1
1.08
1.06
1.04
1.02
1
1.5 2 2.5 3 3.5
Relative wave number θx
– For the D2Q9 model and homogeneous flow the wavenumber |k| of the
most unstable mode is equal to about 2.3 (k is not necessary parallel to
u). An explanation for the ‘magic’ value 2.3 is not known.
– The stability domain as a function of the rest mass parameters α and β of
the D2Q9 model shrinks with increasing lattice size. The ‘canonical’ values
α = 4/9, β = 1/9, which have been derived from maximum entropy prin-
ciple (compare Section 5.2), lie inside one of the stability islands (compare
Fig. 5.6.3).
– The D2Q7 model is less stable than the D2Q9 model in the sense that insta-
bility occurs already at smaller flow velocities (compare Fig. 2 in Sterling
and Chen, 1996).
– The D2Q9 lattice is a projection of the D3Q15 lattice. Therefore it may
be come at no big surprise that their linear stability properties for homo-
geneous background flows show some similarities. The wavenumber |k| of
the most unstable mode is again equal to about 2.3 (Sterling and Chen,
1996).
– The stability domain shrinks further when the background flow includes
shear. Please note that the usual linear stability analysis assumes a time-
independent background flow whereas free shear layers decay by momen-
tum diffusion. Therefore predictions from linear stability analysis are less
reliable at high viscosities where the shear flow may decay before an in-
stability has enough time to develop. Thus the unstable region is smaller
than predicted (see, for example, Fig. 9 in Worthing et al., 1997). How-
ever, the linear stability analysis works fine in the low viscosity region (high
Reynolds numbers) which is of most interest.
To the best of our knowledge linear stability analysis for D3Q19 and for
thermal LB models is not available yet.
Exercise 5.6.2. (***)

Propose equilibrium distributions with rest mass parameters α and β for the
D3Q19 model, investigate linear stability and compare the results with the
linear stability properties of D3Q15.
Fig. 5.6.3. Contour plot of the spectral radius as a function of the rest mass pa-
rameters α and β for the D2Q9 model. On a lattice with 64 times 64 grid points
the stability region shrinks to small islands (domains inside the contour lines). The
‘canonical’ values α = 4/9, β = 1/9, which have been derived from maximum en-
tropy principle, lie inside one of the stability islands.
0.2
0.18
0.16
0.14
0.12
0.1
β
0.08
0.06
0.04
0.02
0
0.4 0.45 0.5 0.55 0.6
α
5.7 Simulating ocean circulation with LBM 219
5.7 Simulating ocean circulation with LBM
Additional forces - wind stress, Coriolis force, and frictional forces - were
implemented into a LB model with 9 lattice velocities in two dimensions
(D2Q9). With this extended model it is possible to simulate the wind-driven
circulation of a barotropic ocean. Model results are compared with analytical
solutions of the linearized problem by Munk and with a finite difference
model of the full nonlinear problem. The implementation of various boundary
conditions and of body forces is discussed in some detail.
5.7.1 Introduction
Despite the rapid development of computer technology during the last

decades, simulation of global ocean circulation is still limited by computer
resources because of several small scale processes which are relevant for large
scale features. Increasing computer power will not solve these problems dur-
ing the next few years. More promissing is the development of new methods
like ‘active nesting’ (Spall and Holland, 1991, Fox and Maskell, 1995 and
1996) to deal with small scale dynamics in selected areas or the application
of other numerical methods like finite elements or LB models.
Because of the strict locality of the ‘collisions’ in LB models they are es-
pecially well suited for massive parallel computers. The code is extremely
simple compared to typical ocean circulation models like MOM (Pacanowski
et al., 1991). No elliptic equation has to be solved. Here the simplest case will
be considered, namely the wind-driven circulation of a barotropic ocean in
a rectangular domain. The main goal of this work is to test whether the LB
model yields results which are in quantitative agreement with the analytical
solution of the linearized problem and with the numerical (finite differences,
finite volumes) solution of the nonlinear problem.
5.7.2 The model of Munk (1950)
One of the grand challenges of physical oceanography in the first half of the
20th century was the explanation of the western boundary currents like the
Gulf Stream, the Agulhas or the Kuroshio. Ekman (1905 and 1923), Sver-
drup (1947) and others had made important contributions to the theory of
wind-driven circulation but could not explain the intensification of the flow
near the western boundaries of oceanic basins. In the late 40ties in the time
span of only two years different approaches for the basin-wide circulation
were proposed by Stommel (1948) and Munk (1950) .
With the North Atlantic in mind Munk defined the following problem. Con-
sider a rectangular (L × H) flat-bottom barotropic (vertically integrated)
ocean which is driven by wind stress of the form (locally cartesian coordi-
nates x, y where x is eastward and y northward)
π
Tx = −T0 cos y and Ty = 0, (5.7.1)
H
corresponding to westerly wind in mid latitudes and easterly wind in low
latitudes. The Navier-Stokes equation contains only molecular diffusion as
dissipative process. For large scale oceanic circulation molecular diffusion
does not play a role. To get rid of the vorticity imparted by the wind Munk
replaced the molecular viscosity coefficient, ν, by the so-called eddy viscos-
ity coefficient, A, which is several orders of magnitude larger than ν. The
Laplacian friction, A∇2 u, can be interpreted as a simple parameterization of
subscale processes. The equation of motion thus reads
∂u 1
+ (u∇) u + f û + ∇p − A∇2 u − T = 0 (5.7.2)
∂t ρ
with û = (−v, u). The Coriolis parameter f is approximated by f f0 + βy
(β-plane), where
!
∂f 2Ω cos ϕ0
f0 = 2Ω sin ϕ0 and β = = . (5.7.3)
∂y ϕ0 R
R = 6371 km is the mean radius of the Earth, Ω = 7.29·10−5 s−1 the angular
velocity of the Earth and ϕ0 the reference latitude. A ≈ 104 m2 s−1 is the
horizontal eddy viscosity coefficient. This value corresponds to a typical value
of 200 − 250 km for the widths of the western boundary currents (see Munk,
1950). The velocity u of an incompressible (∇ · u = 0) two-dimensional flow
can be calculated from a stream function ψ(x, y)
∂ψ ∂ψ
u=− and v= . (5.7.4)
∂y ∂x
Taking the curl of Eq. (5.7.2) to eliminate the pressure gradient one obtains
the vorticity equation
!
∂ 2 ∂ψ ∂Ty ∂Tx
∇ ψ + J(ψ, ∇2 ψ) + β − A∇4 ψ + − =0 (5.7.5)
∂t ∂x ∂x ∂y
where
∂a ∂b ∂a ∂b
J(a, b) = −
∂x ∂y ∂y ∂x
is the Jacobi operator.
The analytical solution of the linear Munk problem. For the station-
ary and linear case the vorticity equation (5.7.5) simplifies:
β ∂ψ T0 π π
∇4 ψ − =− sin y (5.7.6)
A ∂x AH H
The characteristic length (Munk scale)
!1/3
A
WM = (5.7.7)
β
gives an estimate of the width of the western boundary current (see below).
In order to resolve this current in the numerical simulation the grid spacing,
ci ∆t, has to be smaller than WM (= 80 km for β = 2 · 10−11 m−1 s−1 and
A = 104 m2 s−1 ).
Ideally one would like to derive an analytical solution with noslip conditions
(u = 0) on all boundaries (an ocean basin bounded by continents on all
sides). In terms of the streamfunction ψ this would mean that ψ = 0 on all
boundaries, ∂ψ/∂x = 0 on the west and east boundary, and ∂ψ/∂y = 0 on
the south and north boundary. These conditions cannot easily be fulfilled ex-
actly (consider the Fourier expansion of the meriodinal variation of the zonal
wind stress: for each term Yn of the series a solution Xn of the differential
equation has to be constructed and the sum of all these terms has to fulfill
the boundary conditions). Thus we will somewhat relax the constrains and
∂2ψ
allow slip conditions (ψ = 0, ) at the southern and northern boundaries
∂y 2
(these boundary conditions are appropriate for an ocean gyre bounded on
the west and east by continents, and on the south and north by other gyres
circulating in the opposite direction).
The exact solution of Eq. (5.7.6) reads
T0 H 3 π
ψM,e = − sin y
A π3 H
(5.7.8)

· 1+e αr kx
[p1 cos (αi kx) − p2 sin (αi kx)] + p3 eα3 kx
+ p4 eα4 kx
where α1,2 = αr ± iαi , α3 , and α4 are roots of the characteristic equation

2 2
α − γ2 = α (5.7.9)
and the coefficients p1 to p4 are derived from the boundary conditions. The
αn and pn have to be calculated numerically (Exercise 5.7.3).

Derive the exact solution (Eq. 5.7.8) and the characteristic Eq. (5.7.9) of the
linear Munk model (hint: use a separation ansatz).
Exercise 5.7.2. (*)

For small γ find approximate solutions of the characterstic Eq. (5.7.9).

Calculate the αm and pm (m = 1, ..., 4) for A = 103 m2 s−1 , mean latitude
φ0 = 30◦ , L = H = 2 · 106 m.
5.7.3 The lattice Boltzmann model
We apply the lattice Boltzmann model introduced in Section 5.2, i.e. the
BGK kinetic equation (Eq. 5.2.9) over the D2Q9 lattice with the equilibrium
distributions given in Eq. (5.2.12). In the simulations of the linear Munk
problem the nonlinear terms of the distribution functions have been dropped:
4
Fi = ρ i=0
9
1
ci · u
Fi = ρ 1+3 2 i = 1, 2, 3, 4
9 c
1
ci · u
Fi = ρ 1+3 2 i = 5, 6, 7, 8.
36 c
Here we will discuss some details of the coding of propagation, boundary

conditions and forcing.
The grid consists of XMAX × YMAX sites where (XMAX-2) times (YMAX-
2) points are ‘wet’ and the other sites are ‘dry’, i.e. they are located outside
the domain. The boundary between land and ocean lies halfway between the
dry and the neighboring wet site (Fig. 5.7.1).
Fig. 5.7.1. Ocean model: lattice, wet (inside the dashed box) and dry (outside the
dashed box) sites, slip or no-slip boundary conditions.
North (slip)
b b b b b b b b b y = Y M AX
b b b b b b b b b
b b b b b b b b b
b b b b b b b b b
6 I
@ 6
@
R
@ @
b
I
@
@b - b b @b - b b b b
West - East
@ @
@ @
(no-slip) b b? @
Rb b b? @
Rb b b b (no-slip)
b b b b b b b b b
I
@ 6
@
b b b b @b - b b b b
I
@ @
@ ?6
R
@
b b b b b b b b b y=1
x=1 South (slip) x = XM AX
Propagation and boundary conditions. In the propagation step all dis-

tributions Fi (x, t) (except the ’rest’ distribution F0 (x, t)) of wet points pro-
ceed along the corresponding link (direction ci ) to the neighbor site and are
stored in a second array, say Fi :
Fi (x, t) ⇒ Fi (x + ∆tci , t). (5.7.10)
Near the boundary on wet sites not all Fi are occupied because no distribu-
tions are propagated from dry sites. On the other hand some distributions
from sites near the boundary make it to dry sites. The latter distributions
propagate back to wet sites according to the local boundary conditions, i.e.
in case of no-slip boundary conditions the Fi bounce back:
Fi (x, t) ⇒ Fi (x + ∆tci , t) ⇒

FIi (x, t) (5.7.11)

where FIi is the distribution in the direction of −ci ; in case of slip boundary
conditions the distributions are reflected at the boundary:
Fi (x, t) ⇒ Fi (x + ∆tci , t) ⇒

FRi (x + ∆tcRi , t) (5.7.12)
where cRi is the lattice velocity that results from reflection of the lattice

velocity ci at the boundary and FRi is the distribution in the direction of cRi .
After the propagation all dry sites are empty and all wet sites are occupied.
This procedure obviously conserves total mass.
Forcing. As shown in Section 5.2 the inclusion of body forces requires
changes of the distributions whereby spatial and temporal variations of the
forcing have to be taken into account. Two difficulties immediately arise.
When the time dependence of the forcing is not explicitely given in ad-
vance, the scheme becomes implicit, i.e. unknown values at the future time
step are requested in the forcing term. This is indeed the case for the
Coriolis force where the velocity at the new time step is required to cal-
culate K(x + ∆tci , t + ∆t) = f (−u, v)x+∆tci ,t+∆t . This problem can be
solved by applying a predictor-corrector method. In the predictor, we use
[K(x, t) + K(x + ∆tci , t)] /2. In the corrector we evaluate the forcing term
by using the previous iterate. One corrector seemed to be sufficient; without
the corrector a numerical instability occurs after an integration time of sev-
eral weeks.
The forcing should respect conservation of total mass. Mass is conserved at
each lattice site in the case of the ’local forcing’ because

ci K(x, t) = K(x, t) ci = 0. (5.7.13)
i i
On the other hand mass is in general not conserved locally under ’non-local
forcing’
ci K(x + ∆tci , t) = 0. (5.7.14)
i
Nevertheless, total mass should (and can) be conserved. Problems arise near
the boundaries where some forcing terms are not compensated (in terms of
mass) by contributions from other sites. This can best be illustrated by an
onedimensional example (Fig. 5.7.2; all forcings at t + ∆t). K(x1 ) does not
exists because x1 is a dry site; it is set to zero. K(x2 ) is not compensated. In
order to compensate this term and to add some forcing at x2 , K(x2 ) is added
at x2 in such a way that after propagation the forcing at x2 is comparable
to that at neighboring sites (the details of the algorithm depend on the type
of boundary conditions).
Fig. 5.7.2. External forcing: a onedimensional example. The arrows indicate com-
pensation in terms of mass.
dry wet
b b b b b b
Forcing I: +K(x3 ) +K(x4 ) +K(x )

5 6 +K(x )7 +K(x )
X
y
XXX X
yXXX X yXXX
(formal) XX X X X X
−K(x1 ) z−K(xX
−K(x2 )X
X 3)
z−K(xX
X 4)
z−K(x5 )
X
does not not
exist compensated
Forcing II: +K(x3 ) +K(x )

4 +K(x )
5 6 +K(x ) 7 +K(x )
X
yXXX X yXX y
XXX
(conserving XXX XXXX XXXX
total mass) - −K(x2 ) X
−K(x2 ) z−K(x3 ) X
z−K(x4 ) X
z−K(x5 )
LB simulation of the linear Munk problem. The linear Munk problem

was integrated in a 2000 km × 2000 km domain with central latitude at
30◦ N (f0 = 7.29 · 10−5 s−1 , β = 1.98 · 10−11 s−1 m−1 ). The eddy diffusivity
coefficient A = 104 m2 s−1 results in a width of the Munk layer WM = 80 km
which is almost twice the grid spacing (∆x = 50 km). The characteristic speed
U = 5 · 10−4 m s−1 (linear regime!) is consistent with a wind stress coefficient
U · ∆x
T0 = 7.9 · 10−10 m s−2 . The grid Reynolds number (Re,g = = 2.5 ·
A
−3
10 ) is small compared to one.
In explicit numerical schemes the time step is limited by the fastest waves in
the system (Courant, Friedrichs und Lewy, 1928). Rossby (planetary) waves
are excited by the sudden onset of wind forcing. The free undamped waves
are governed by
∂ 2 ∂ψ
∇ ψ+β = 0. (5.7.15)
∂t ∂x
Inserting the ansatz ψ = ψ0 ei(kx+ly−σt) (wavenumber k = (k, l), k, l ≥ 0)
into Eq. (5.7.15) one can readily derive the dispersion relation
βk
σ(k, l) = − (5.7.16)
k2+ l2
which shows that Rossby waves live on the ‘β effect’, i.e. the fact that the
rotation rate (as measured by the Coriolis parameter f ) varies with latitude.
The phase speed
σ βk "
vph = = − 3 with κ = k 2 + l2 (5.7.17)
κ κ
is always negativ (westward propagating phase). Its magnitude becomes max-
imal for k = 2π/L and l = 0 and reads
βL2
|vph |max = = 2 m s−1 . (5.7.18)
4π 2
Thus the time step has to below ∆x/ |vph |max ≈ 2500 s. Integration with
∆t = 200 s was numerically stable over the whole simulation time (15 weeks).
The global kinetic energy (Fig. 5.7.3) increases over two weeks until it reaches
a quasi steady state with oscillations due to Rossby waves. The mean (over
several weeks in quasi steady state) velocity shows excellent agreement with
the analytical solution (Fig. 5.7.4).
LB simulation in the nonlinear regime. In order to test the LBM under

nonlinear conditions, simulations were performed in a closed (no-slip bound-
ary conditions everywhere) basin of size 4000 km × 4000 km with central
latitude at 30◦ N. The flow can be characterized by two dimensionless num-
bers: the Rossby number
U
Ro = (5.7.19)
βL2
Fig. 5.7.3. Time series of the global kinetic energy (LBM simulation of the linear
Munk problem). The oscillations in the quasi steady state are due to Rossby waves.
−5
x 10
3
2.5
2
Kinetic energy
1.5
0.5
0
0 5 10 15
Time (weeks)
Fig. 5.7.4. Isocontours of the velocity component u (times 104 ) show excellent
agreement between the analytical solution (left) and the LBM result (right).
1500 1500 0.5

y (km)
1000 1000
−0.5 −0.5
500 500
500 1000 1500 500 1000 1500

x (km) x (km)
measures the ratio between the advection (u∇u) and the β term (βyû), and
the Reynolds number
UL
Re = (5.7.20)
A
is the ratio between the advection and the friction term (A∇2 u). The ratio
of the Rossby and Reynolds number gives the Ekman number
A
EA = (5.7.21)
βL3
which measures the ratio between the friction and the β term. The Reynold
U ·L U
number, Re = , was set to 80 and the Rossby number, Ro = ,
A β · L2
to 1.28 · 10−3 . U = 0.4 m s−1 and A = 20300 m2 s−1 where calculated from
Re and Ro. The grid spacing, ∆x = 40 km, was chosen small than the Munk
width, WM = 100 km. The applied wind stress
πy
Tx = −T0 sin2 , Ty = 0 (5.7.22)
L
with T0 = 8 · 10−7 m s−2 leads to the formation of a double gyre (Fig. 5.7.5).
Fig. 5.7.5. LB simulation of a double gyre: isocontours of the stream function

ψ(x, y) (times 10−4 ).
1 1
5
3500
1
5
1
10
3000
2500 5
5
y (km)
1
1 1
2000
−1 −1
−1
−5
1500 −5
1000
−1
−5
500
−1
−5
−1 −1
500 1000 1500 2000 2500 3000 3500

x (km)

A flow without any external body force can be characterized by a single di-
mensionless number (Reynolds number). A flow with one type of external
body force requires two dimensionless numbers (Reynolds and Froude num-
ber; compare discussion of the similarity law in Section 1.3). Here we consider
a flow with two types of body forces (Coriolis and wind stress). Do we need
three (independent) dimensionless numbers to characterize the flow?
Exercise 5.7.5. (***)

The Navier-Stokes equation contains only molecular diffusion as dissipative
process. For large scale oceanic circulation molecular diffusion does not play
a role. To get rid of the vorticity imparted by the wind Stommel (1948) intro-
duced bottom friction that is linear in the velocity u = (u, v). The appropriate
equation of motion reads
∂u 1
+ (u∇) u + f û + ∇p + ks u − T = 0 (5.7.23)
∂t ρ
with the (bottom) friction coefficient ks . Taking the curl of Eq. (5.7.23) one
obtains the vorticity equation
!
∂ 2 ∂ψ ∂Ty ∂Tx
∇ ψ + J(ψ, ∇2 ψ) + β + ks ∇2 ψ + − = 0. (5.7.24)
∂t ∂x ∂x ∂y
For the stationary and linear case the vorticity equation (5.7.24) simplifies:
!
∂ψ ∂Ty ∂Tx
ks ∇2 ψ + β =− − . (5.7.25)
∂x ∂x ∂y
The analytical solution of Eq. (5.7.25) in a rectangular ocean basin of length
L and width H with 0 ≤ x ≤ L and 0 ≤ y ≤ H and the boundary conditions
ψ(0, y) = ψ(L, y) = ψ(x, 0) = ψ(x, H) = 0. (5.7.26)
reads
ψ(x, y)
(5.7.27)
 
β(L − x) βx
−
π  e 2ks sinh(αx) + e 2ks sinh(α(L − x)) 
T0 H  
= sin y 1 − 
ks π H  sinh(αL) 
where 1
β2 π2
α= + . (5.7.28)
4ks2 H2
This solution describes an asymmetrical gyre with a narrow intense western

boundary current and a wide slow southward drift in the eastern part of the
basin.
1. Write a LBM code for the Stommel problem. 2. Compare results of the
LBM code with the analytical solution of the linear Stommel model. 3. Dis-
cuss the pattern of the gyre as a function of Rossby number.
Further reading: Salmon (1999).

5.8 A lattice Boltzmann equation for diffusion
The formulation of lattice-gas cellular automata (LGCA) for given partial

differential equations is not straightforward and still requires ‘some sort of
magic’. Lattice Boltzmann models are much more flexible than LGCA be-
cause of the freedom in choosing equilibrium distributions with free param-
eters which can be set after a multi-scale expansion according to certain
requirements. Here a LBM is presented for diffusion in an arbitrary num-
ber of dimensions (Wolf-Gladrow, 1995). The model is probably the simplest
LBM which can be formulated. It is shown that the resulting algorithm with
relaxation parameter ω = 1 is identical to an explicit finite differences (EFD)
formulation at its stability limit. Underrelaxation (0 < ω < 1) allows stable
integration beyond the stability limit of EFD. The time step of the explicit
LBM integration is limited by accuracy and not by stability requirements.
The creation of LGCA for certain partial differential equations still seems
to require ‘some sort of magic’ (compare quotation of Toffoli and Margulos,
1990, in Section 5.4). Here a simple LBM for diffusion is presented and it
is shown how straightforward it is to derive such a model. In addition, the
resulting algorithms are compared with an explicit finite difference (EFD)
scheme.
5.8.1 Finite differences approximation
An explicit finite difference scheme for the diffusion equation
∂T
= κ∇2 T (5.8.1)
∂t
(T is the concentration of a tracer, κ is the diffusion coefficient and ∇2 is

the Laplace operator in D dimensions in Cartesian coordinates) results from
forward approximation in time and central differences in space
(n+1) κ∆t (n) (n)
Tk1 ,k2 ,...,kD = Tk1 +1,k2 ,...,kD + Tk1 −1,k2 ,...,kD +
(∆x)2

(n) (n)
... + Tk1 ,k2 ,...,kD +1 + Tk1 ,k2 ,...,kD −1
!
κ∆t (n)
+ 1 − 2D Tk1 ,k2 ,...,kD
(∆x)2
5.8 A lattice Boltzmann equation for diffusion 233
where equidistant and equal spacing in all dimensions have been assumed.
The scheme is stable for
1 (∆x)2
0 < ∆t ≤
2D κ
(see, for example, Ames, 1977; compare Subsection 5.6.2 for the case D = 1).
At the upper stability limit the scheme becomes especially simple

(n+1) (n) (n)
Tk1 ,k2 ,...,kD = Tk1 +1,k2 ,...,kD + Tk1 −1,k2 ,...,kD
6
(n) (n)
... + Tk1 ,k2 ,...,kD +1 + Tk1 ,k2 ,...,kD −1 (2D)
that is T at the new time level is given by the mean over all neighbor values
at the previous time level.
5.8.2 The lattice Boltzmann model for diffusion
“... it is well known that 90◦ rotational invariance is sufficient to yield

full isotropy for diffusive phenomena.”
Toffoli and Margulos (1990)
According to Toffoli and Margulos (1990) it is sufficient to use a square or a

cubic lattice in two or three dimensions, respectively. The following model is
applicable in an arbitrary number of dimensions. The grid velocities (vectors
connecting neighboring grid points) are defined by
c2n−1 = (0, 0, ..., 0, 1, 0, ..., 0), c2n = (0, 0, ..., 0, −1, 0, ..., 0) n = 1, 2, ..., D
where D is the dimension.

(0)
In general, the equilibrium distributions6 Tm depend on the conserved quan-
tities (here only T ), a number of parameters γk (k = 0, 1, ..., N ) and on the
direction (index m). Here, grids with only one speed are considered and the
(0)
equilibrium distribution functions Tm do not independent on m. T is given
as the sum over the distribution functions Tm

(0)
T (x, t) = Tm (x, t) = Tm (x, t) (5.8.2)
m m
6
We will use the notation Tm insteed of Fm for the distribution functions. The
(0)
equilibrium distribution functions for pure diffusive problems, Tm , (diffusion of
(0)
temperature or tracers) are simpler than those for flow problems, Fm .
where the summation runs over all directions (m = 1, 2, ..., M = 2D).

The diffusion equation is a linear differential equation. Hence it is reasonable
(0)
to use a linear ansatz for Tm
(0)
Tm = γ0 + γ1 T . (5.8.3)
Inserting (5.8.3) into (5.8.2) yields
(0) T
Tm = (5.8.4)
2D
that is, all free parameters are already fixed by the definition of the tracer
concentration. The diffusion coefficient κ will result from the multi-scale ex-
pansion as described below.
5.8.3 Multi-scale expansion
(0)
The LBM is defined by the grid, the equilibrium distribution Tm and the
kinetic equation
Tm (x + cm , t + 1) = (1 − ω)Tm (x, t) + ωTm

(0)
(x, t) (5.8.5)
which states that the distribution at the new time level (t + 1) at the neigh-
boring site (x + cm ) is a weighted sum of the distribution Tm (x, t) and the
(0)
equilibrium distribution Tm (x, t). Models with parameter ω go under various
names: ‘enhanced collision’ (Higuera et al., 1989), BGK (named after Bhatna-
gar, Gross and Krook, 1954; compare, for example, Qian et al., 1992), STRA
(‘single time relaxation approximation’, Chen et al., 1991) or SOR (‘succes-
sive over-relaxation’, Qian et al., 1992). The LBM is stable for 0 < ω < 2. Now
the macroscopic equations will be derived by a multi-scale analysis (compare
Frisch et al., 1987, for an analogous procedure for LGCA). The distribution
functions are expanded up to linear terms in the small expansion parameter

Tm = Tm (0) (1)
+ Tm + O(2 ).
(1)
From the kinetic equation (5.8.5) one can calculate an approximation of Tm
Tm (x + cm , t + 1) = Tm (x, t) + ∂xα cmα Tm + ∂t Tm + O(2 )

= (1 − ω) Tm (x, t) (0)
+ωTm (x, t)

(0)
= Tm (1)
+ Tm + O(2 )
→
1 1
(1)
Tm = − ∂xα cmα Tm − ∂t Tm + O(2 ).
ω ω
Diffusion is a slow process on large spatial scales which suggests the following
scaling (same as for the derivation of the Navier-Stokes equations in Frisch
et al., 1987)
(2)
∂t → 2 ∂t
∂xα → ∂x(1)
α
.
The components of the grid velocities obey the following equations

cm = 0
m

cmα cmβ = 2δαβ
m
and therefore T
cm Tm
(0)
= cm = 0.
m
2D m
Inserting the expansion and the scalings into the conservation relation for
tracer concentration, one obtains up to second order in

0 = [Tm (x + cm , t + 1) − Tm (x, t)]
m
(2) 1
= [Tm (x, t) + 2 ∂t Tm +∂x(1) c T + 2 ∂x(1)
α mα m
∂x(1) c c T (0)
β mα mβ m
m
2 α
→ ∂t T
−Tm (x, t) + O(3 )]
and

∂x(1) c T
α mα m
= ∂x(1)
α
(0)
cmα Tm + 2 ∂x(1) c T (1) + O(3 )
α mα m
m m m

=0
1 2 (1) (1)
= − (0)
∂xα ∂xβ cmα cmβ Tm + O(3 ) (5.8.6)
ω m
1 1 2
= − δαβ ∂x(1)
α
∂x(1)
β
T +O(3 )
ωD
→∇ T 2
1 1 2
2 ∂x(1)
α
∂x(1) c c T (0) =
β mα mβ m
δαβ ∂x(1)
α
∂x(1)
β
T
m
2 2D
and finally
∂T
= κ∇2 T
∂t
with !
1 1 1
κ= − . (5.8.7)
ω 2 D
5.8.4 The special case ω = 1
For ω = 1 the kinetic equation (5.8.5) reduces to

Tm (x + cm , t + 1) = Tm
(0)
(x, t)
1
and the diffusion coefficient is κ = . This LBM is identical to the finite
2D
difference scheme at the stability limit. The right hand side of the kinetic
equation is just the mean value of the nearest neighboring sites and the
diffusion coefficient is the maximal value allowed by the stability condition.
For the LBM the diffusion coefficient is expressed in the units ∆t = ∆x = 1;
the diffusion coefficient at the stability limit of the EFD reads
1 (∆x)2 1
κ= = .
2D ∆t 2D
This scheme requires only two arrays in memory: the tracer concentrations
at two time levels.
5.8.5 The general case
In the general case one has to store M = 2D distributions in addition to the

tracer concentrations at two time levels. What do we gain from this extra
cost? In the range 0 < ω < 1 (underrelaxation) the diffusion coefficient κ is
larger than the value at the stability limit of the EFD scheme while we still
keep ∆t = ∆x = 1. In contrast to EFD, the LBM is stable in this parameter
range.
5.8.6 Numerical experiments
To test the predictions of the LBM outlined above the one-dimensional dif-
fusion equation was integrated. As initial conditions, values of an analytical
solution were used, namely
% &
1 x2
T (x, ti ) = √ exp − .
2 πκti 4κti
The integration starts at ti = 15/κ(ω) and ends at tf = 75/κ(ω), thus the

time interval depends on κ(ω), but in each case the integration starts with the
same numerical values and ends after the maximum decreases from ≈ 0.073
to ≈ 0.033. Fig. 5.8.1 shows the results of such an integration for ω = 0.3
together with the analytical solution and the initial values. By appropriate
choice of ω, one can keep ∆t = 1 for ‘arbitrarily’ large diffusion coefficients:
the scheme is stable but the numerical error increases with increasing diffusion
coefficient (compare Fig. 5.8.2). Thus, we have an explicit scheme (BGK-
LBM) where the length of the time step is no longer limited by stability
requirements. The large error at small values of ω stems from the fact that
explicit approximations of parabolic equations act like a hyperbolic system
with two real finite difference characteristics instead of only a single real
characteristic of the continuous system (Ames, 1977).
5.8.7 Summary and conclusion
A very simple LBM for diffusion in an arbitrary number of dimensions is

proposed. For ω = 1 the resulting algorithm is identical to an explicit finite
difference scheme at its stability limit. Thus the LBM scheme is not only
stable, but automatically picks the maximal allowed diffusion coefficient κ to
ensure stability of the EFD scheme.
For LGCA the transport coefficients depend on the collision rules which are
never optimal in the sense that they yield only a certain approximation of the
(continuous) local equilibrium functions (compare the various FHP models
with and without rest particles in Frisch et al., 1987, or the various collision
rules proposed for FCHC by Hénon, 1987, Rem and Somers, 1989, and van
Coevorden et al., 1994) whereas for LBM, the collisions (which do not show
up explicitly) can create local equilibrium at each time step. By reducing
the number of collision in LGCA one obtains models with higher diffusion
coefficients while stability is assured. This can be regarded as a kind of un-
derrelaxation.
In the BGK-LBM the diffusion coefficient κ is an adjustable parameter. Of
special interest is the parameter range 0 < ω < 1. The use of information
contained in the nonequilibrium distribution functions allows explicit stable
integration beyond the stability limit of the EFD scheme. Thus, the time step
is limited by accuracy and not by stability requirements.
Fig. 5.8.1. Integration of the diffusion equation in one dimension by the BGK-
LBM with ω = 0.3. The % integration
& starts at time ti = 15/κ(ω) with initial values
1 x2
T (x, ti ) = √ exp − (dotted line) and ends at tf = 75/κ(ω). The figure
2 πκti 4κti
shows the numerical results (broken line) together with the analytical solution (solid
line).
0.08
0.07
0.06
num
and T
0.05
0.04
ana
,T
0.03
init
T
0.02
0.01
0
−50 0 50
x
Fig. 5.8.2. Integrations of the diffusion equation in one dimension by the BGK-
LBM. The integration
% & starts at time ti = 15/κ(ω) with initial values T (x, ti ) =
1 x2
√ exp − and ends at tf = 75/κ(ω). The plot shows the logarithm of
2 πκti 4κti
the maximum error (max{|Tnumerical solution − Tanalytical solution |}) at the end
of the integrations as a function of ω. The error increases at small values of ω (large
values of the diffusion coefficients).
−1
10
−2
10
error
−3
10
−4
10
−5
10
0 0.5 1 1.5
ω
5.8.8 Diffusion equation with a diffusion coefficient depending on

concentration
By a small modification the lattice Boltzmann model proposed above can be

generalized to a model for a diffusion equation with a diffusion coefficient
depending on concentration: The constant ω will be replaced by a function
ω(T ). The multi-scale analysis proceeds as before except for Eq. (5.8.6) where
1/ω and the spatial derivative must not be exchanged:

∂x(1) c T
α mα m
= ∂x(1)α
cmα Tm(0)
+ 2 ∂x(1) c T (1) + O(3 )
α mα m
m m m

=0
1
= − 2 ∂x(1) ∂ (1) cmα cmβ Tm
(0)
+ O(3 )
m
α
ω(T ) xβ
1 1
= − 2 δαβ ∂x(1) ∂ (1) T +O(3 ).
D α
ω(T ) xβ

% &
1
→ ∇ ∇T
ω(T )
Thus the macroscopic equation reads
∂T
= ∇ [κ(T )∇T ] (5.8.8)
∂t
with % &
1 1 1
κ(T ) = − . (5.8.9)
ω(T ) 2 D
Comparison with an analytical solution. A few analytical solutions are

known for the nonlinear diffusion equation
% & !2
∂T ∂ ∂T ∂κ ∂T ∂2T
= κ(T ) = + κ(T ) 2 . (5.8.10)
∂t ∂x ∂x ∂T ∂x ∂x
For κ(T ) = T (Logan, 1994, p. 143; a misprint has been corrected) a solution
reads
1 2 2/3
T (x, t) = A t − x2 for |x| < At1/3 (5.8.11)
6t
and T (x, t) = 0 otherwise. Fig. 5.8.3 shows the comparison between the
numerical solution by the lattice Boltzmann model and the analytical solution
at t = 210 for initial conditions according to eq. (5.8.11) with A = 5, t = 10
inside the interval |x| < At1/3 and T = 0 otherwise. The agreement is very
good.
Fig. 5.8.3. Integration of the nonlinear diffusion equation (5.8.10) by the lattice
Boltzmann model. The initial distribution is marked by circles. The numerical solu-
tion at t = 210 (solid line) is indistinguishable from the analytical solution (dashed-
dotted line; not visible). The broken line shows the difference between numerical and
analytical solution multiplied by 100.
2
1.5
1
T(x)
0.5
−0.5
−50 0 50
x
5.8.9 Further reading
Elton et al. (1990) proposed a LBM for a nonlinear diffusion equation in 2D.
The collision operator is based on variants of the HPP model (extended by
one- and two-particle collisions which conserve mass but not momentum).
The diffusion coefficients are functions of the tracer density.
The LBM for reaction-diffusion systems by Dawson et al. (1993) lives on
the FHP lattice (D2Q7) and applies the BGK approximation of the collision
operator.
Qian and Orszag (1995) developed a LBM for reaction-diffusion equations.
Their diffusion model is identical to that proposed by Wolf-Gladrow (1995).
Further reading: Sun (1998), van der Sman and Ernst (1999).
Finite differences: Teixeira (1999).
5.9 Lattice Boltzmann model: What else? 243
5.9 Lattice Boltzmann model: What else?
– Books: Rothman and Zaleski (1997).

– Review articles: Benzi, Succi and Vergassola (1992), Rothman and Zaleski
(1994), Qian, Succi and Orszag (1995), Biggs and Humby (1998), Chen
and Doolen (1998).
– Curvilinear coordinates, finite volumes, irregular grids, mesh refinement:
Filippova and Hänel (1998a,b), He and Doolen (1997a,b), He (1997), Karlin
and Succi (1998), Mei and Shyy (1998), Peng et al. (1998, 1999), Renwei
and Wei (1998), Tölke et al. (1998, 1999), Karlin et al. (1999), Xi et al.
(1999a,b).
– Flow past obstacles: Bernsdorf et al. (1998).
– Flow through porous media: Cancelliere et al. (1990), Aharonov and Roth-
man (1993), Gunstensen and Rothman (1993), Sahimi (1993), Buckles et al.
(1994), Heijs and Lowe (1995), Martys and Chen (1996), Shan and Doolen
(1996), Grubert (1997), Spaid and Phelan (1997; 1998), Angelopoulos et
al. (1998), Bosl et al. (1998), Coveney et al. (1998), Dardis and McCloskey
(1998a,b), Freed (1998), Kim et al. (1998), Koch et al. (1998), Koponen et
al. (1998a,b), Maier et al. (1998a,b), Noble and Torczynski (1998), Waite
et al. (1998; 1999), Berest et al. (1999), Bernsdorf et al. (1999), Inamuro
et al. (1999), Verberg and Ladd (1999).
– Granular flow: Herrmann (1995), Tan et al. (1995), Herrmann et al. (1996).
– Multiphase flows: Gunstensen et al. (1991), Holme and Rothman (1992),

Flekkøy (1993), Shan and Chen (1993, 1994), d’Ortona et al. (1995),
Flekkøy et al. (1995), Orlandini et al. (1995), Flekkøy et al. (1996b), Hal-
liday (1996), Halliday et al. (1996), Sofonea (1996), Swift et al. (1996),
Gonnella et al. (1997; 1998; 1999), Hou et al. (1997), Kato et al. (1997),
Wagner and Yeomans (1997; 1998; 1999), Angelopoulos et al. (1998), Chen
et al. (1998a), Halliday et al. (1998), Holdych et al. (1998), Lamura et al.
(1998; 1999), Masselot and Chopard (1998a), Theissen et al. (1998), Wag-
ner (1998b), Yu and Zhao (1999).
– Magnetohydrodynamics (MHD): Chen et al. (1991), Succi et al. (1991),
Martinez et al. (1994), Sofonea (1994).
– Compressible flows, Burgers equation: Alexander et al. (1992).
– Flow past fractal obstacles: Adrover and Giona (1997).
– Turbulence and large eddy simulation (LES): Frisch (1991), Benzi et al.
(1996), Amati et al. (1996), Fogaccia et al. (1996), Hayot and Wagner
(1996), Amati et al. (1997a,b), Succi (1998), Teixeira (1998).
– Flow in dynamical geometry (blood flow): Fang et al. (1998), Krafczyk et
al. (1998a).
– Glacier flow: Bahr and Rundle (1995).

– Rayleigh-Bénard convection: Bartoloni et al. (1993), Massaioli et al. (1993),
Benzi et al. (1994), Pavlo et al. (1998a) Vahala et al. (1998a).
– Rayleigh-Taylor instability: Nie et al. (1998), He et al. (1999a,b).
– Korteweg-de Vries equation: Yan (1999).
– Droplets: Schelkle et al. (1999), Xi and Duncan (1999).
– Crystal growth: Miller and Böttcher (1998).
– Wave propagation: Chopard and Luthi (1999).
– Maxwell’s equations: Simons et al. (1999).
– Decoupling of spatial grid from the velocity lattice: Cao et al. (1997), Pavlo
et al. (1998a).
– H-theorem for LBM: Karlin and Succi (1998), Karlin et al. (1998), Wagner
(1998a), Karlin et al. (1999).
– Quantum mechanics: Succi and Benzi (1993), Succi (1996), Meyer (1997a,b;
1998).
– Related methods: Junk (1999).
– Further reading: Lavallée et al. (1989), Kingdon et al. (1992), McNamara
and Alder (1992, 1993), Nannelli and Succi (1992), Qian (1993), Qian and
Orszag (1993), Ladd (1993, 1994a,b), Chen et al. (1994), Punzo et al.
(1994), Succi and Nannelli (1994), Elton et al. (1995), Hou et al. (1995),
McNamara et al. (1995), Miller (1995), Ohashi et al. (1995), Reider and
Sterling (1995), Succi et al. (1995), Wagner and Hayot (1995), Zou et al.
(1995a,b), Flekkøy et al. (1996a), Elton (1996), (Burgers eq.) He et al.
(1996), Kaandorp et al. (1996), Lin et al. (1996), Orszag et al. (1996),
Rakotomalala et al. (1996), Sterling and Chen (1996), Chen and Ohashi
(1997), Filippova and Hänel (1997, 1998c), Giraud and d’Humières (1997),
Luo (1997a,b), Maier and Bernard (1997), Qi (1997; 1999), Qian (1997),
Qian and Chen (1997), Stockman et al. (1997; 1998), Succi et al. (1997), van
der Sman (1997), Warren (1997), Ahlrichs and Dunweg (1998), Aidun et al.
(1998), Buick and Greated (1998), Buick et al. (1998), Chen (1998), Chen
and Ohashi (1998), Chen et al. (1998b), Chenghai (1998), De Fabritiis
et al. (1998), Giraud et al. (1998), He and Zhao (1998), Kandhai et al.
(1998a,b; 1999), Krasheninnikov and Catto (1998), Luo (1998), Succi and
Vergari (1998), Takada and Tsutahara (1998), Ujita et al. (1998), Xu and
Luo (1998), Yan (1998), Yan et al. (1998; 1999), Kendon et al. (1999), Klar
(1999).
5.10 Summary and outlook 245
5.10 Summary and outlook
Because LGCA and LBM are still in rapid development it is not possible
to give an actual and complete picture of the whole field. Instead I have
tried to introduce the basic models (HPP, FHP, FCHC, D2Q9) plus some
personal favorites (like PI) together with methods from statistical mechanics
(Chapman-Enskog expansion, BGK approximation, maximum entropy prin-
ciple) which are necessary for the theory of LGCA or LBM but which are
usually not part of the curriculum for students of physics or mathematics.
Knowledge of these special methods is usually taken for granted in articles
and even in reviews.
I am still fascinated by these new methods, and I guess, I am not alone. Al-
though the emergence of new numerical methods is often driven by practical
requirements by engineers and natural scientists, (applied) mathematicians
usually take the lead in the development of the schemes. However, many re-
searchers working on LGCA and LBM would call themselves most probably
physicists. Why are these kind of schemes so attractive to them? At least
three reason come to my mind. Firstly, both approaches are based on conser-
vation laws. Physicists feel at home with conservation laws. The continuity
equation and the Navier-Stokes equation express the conservation of mass
and momentum. Certain conserved quantities are related (Noether theorem)
to symmetry groups which allow us to derive equations of the fundamen-
tal field theories by the gauge principle. Secondly, LGCA and LBM require
more (physical) theory than many other numerical methods. Whereas you
can teach to a beginner how to create and apply the simplest schemes of
finite differences or spectral methods to partial differential in a few hours,
this seems not possible for LGCA, LBM, and finite elements. And the the-
ory required stems from statistical mechanics (Chapman-Enskog expansion,
maximum entropy principle) whereas for finite elements, for example, weak
convergence in Sobolev spaces is more the backyard of mathematicians. Last
but not least, it is fascinating to see the role of symmetry. The symmetry of
the underlying lattice is still important on the macroscopic level. It took more
than 10 years to discover that it is sufficient to replace a lattice with four-
fold (HPP) by one with hexagonal (FHP) symmetry in order to obtain the
correct form of the nonlinear advection term of the Navier-Stokes equation.
The same symmetry requirement even led us into four dimensions (FCHC).
Chiral symmetry has to be established by random processes (FHP). Spuri-
ous invariants have to be detected (still not solved in general) and then to be
destroyed (for example: three-particle collisions in FHP) or not to be initial-
ized or generated (Zanetti invariants). And finally consequences of missing
symmetry has to be scaled away (symptomatic treatment of the violation of
Galilean invariance).
As a final remark let me ask the question: What will be the future of LGCA
and LBM? Will these schemes have outcompeted the traditional methods (fi-
nite differences, finite volumes, finite elements, spectral, semi-Lagrange, ...)

in say five years from now? Or will we throw away these new schemes because
they cannot compete with the fastest of the other schemes? The application
of a wide variety of methods at the same time already indicates that currently
(forever?) there does not exist a best scheme for all problems in the numeri-
cal solution of partial differential equations. Although finite elements are well
established since decades (the first edition of the monograph by Zienkiewicz
appeared already in 1967) the Modular Ocean Model (the work horse of phys-
ical oceanography) still applies finite volumes (Bryan, 1969; Semtner, 1997).
And the traditional methods are still under development (the same is true for
languages such as FORTRAN). Rood (1987) lists about 100 different numer-
ical advection schemes and several new have been proposed since. LGCA and
LB models have disadvantages (how to construct models for given differential
equations, noise of LGCA, stability of LBM, large memory requirement of
LBM) which prohibit the general replacement of well-established traditional
methods.
On the other hand, LGCA and LB models have several advantages like flexi-
bility with respect to domain geometry, locality of the collisions (well adapted
to massively parallel computer) and low complexity of the code. Because LB
models are based on conservation principles they are especially suited for
long-time integrations (for instance: climate models). It has been shown al-
ready that LGCA and LB models can compete with traditional schemes in
certain classes of problems. Therefore, it is almost certain that both methods
will find their niches.

Lattice Gas Cellular Automata and Lattice Boltzmann Models Chapter6

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

Lattice Gas Cellular Automata and Lattice Boltzmann Models Chapter6

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lattice Gas Cellular Automata and Lattice Boltzmann Models Chapter6

Uploaded by

Copyright:

Available Formats

5.

Lattice Boltzmann Models

5.1 From lattice-gas cellular automata to lattice

Lattice-gas cellular automata for Navier-Stokes equations are plagued by sev-

disease cause therapy/cure remarks

Historically the following stages in the development of lattice Boltzmann

D.A. Wolf-Gladrow: LNM 1725, pp. 159–246, 2000.

1. Lattice Boltzmann equations have been used already at the cradle of

5.1.1 Lattice Boltzmann equation and Boltzmann equation

The microdynamics of LGCA are described by kinetic equations of the type

which by substituting ni → f , ci → v, ∆i /∆t → Q gives the Boltzmann

where ∆t̂ = ∆t · U/L. Lagrangian behavior is then obtained by the selection

“This equation has a particular simple physical interpretation in

5.1.2 Lattice Boltzmann models of the ﬁrst generation

The mean occupation numbers Fi (0 ≤ Fi ≤ 1) develop in time according to

Fi (r + ci , t + 1) = Fi (r, t) + Ωi ({Fj (r, t)}) , (5.1.6)

Ωi (N ) = Fi+1 Fi+3 Fi+5 (1 − Fi )(1 − Fi+2 )(1 − Fi+4 )

5.2 BGK lattice Boltzmann model in 2D

A LBM has three main ingredients:

1. the lattice: D2Q9 (multi-speed model),

where the lattice constants a and b are restricted by a2 /3 ≤ b2 ≤ 3a2 . The

Fig. 5.2.1. The D2Q9 lattice.

'$ '$ '$

In what follows the D2Q9 lattice will be used exclusively.

For vanishing velocities a global equilibrium distribution Wi (“ﬂuid at rest”)

Exercise 5.2.1. (**)

where τ is the collision time, ω = ∆t/τ is the collision frequency, and K is

Application of the multi-scale technique (Chapman-Enskog expansion) yields

5.2.1 Derivation of the Wi

Table 5.2.1. Lattice speeds, cells and Wi of the D2Q9 lattice.

c2i cells number of cells Wi

The non-vanishing elements of the even moments up to fourth order read:

Remark: The only equation which includes W0 . It will be used to calculate

Remark: This constraint is identical with (5.2.14).

Remark: W1 will be calculated from Eq. (5.2.15).

Remark: This constraint is identical with (5.2.15).

Remark: W2 will be calculated from Eq. (5.2.16).

The solution reads:

5.2.2 Entropy and equilibrium distributions

Koelman deﬁnes the relative entropy4 density by

The weighting by 1/Wi in the logarithmic factor will lead to equilibrium

encompasses the constraints coupled by Lagrange multipliers Ã und B̃. The

The solutions of (5.2.18) are of the form

and ﬁnally up to terms of second order in j

The vector equation (5.2.21) reduces to a scalar constraint

To obtain A2 independent of j (as implied by the ansatz) the expression (∗)

5.2.3 Derivation of the Navier-Stokes equations by multi-scale

The derivation of the macroscopic equations (Navier-Stokes) proceeds in close

Fi (x + ci ∆t, t + ∆t) = Fi (x, t) + ∆t∂t Fi + ∆tciα ∂xα Fi

Conservation of mass and momentum. The expansions given above are

(∆t) ciγ ' (

where we have used 

The momentum ﬂux tensor in ﬁrst order of  reads

The second summand

vanishes because it is an odd moment in ci over Wi . The third summand

Tαβγδ = δαβ δγδ + δαγ δβδ + δαδ δβγ (5.2.36)

The fourth summand reads:

Summing up all terms yields

Terms of second order in : mass. As for lattice-gas cellular automata

where we have used

The momentum ﬂux tensor in ﬁrst order of reads

(1) ∆t (1) (1) (0)

∆t (1) (1) (0) ∆t (1) (1)

∆t (1) (1) (0)