Recovering The Full Navier Stokes Equations With Lattice Boltzmann Schemes
Recovering The Full Navier Stokes Equations With Lattice Boltzmann Schemes
, , ,
François Dubois , Benjamin Graille , and Pierre Lallemand
Abstract. We consider multi relaxation times lattice Boltzmann scheme with two particle distributions for the thermal Navier
Stokes equations formulated with conservation of mass and momentum and dissipation of volumic entropy. Linear stability is
taken into consideration to determine a coupling between two coefficients of dissipation. We present interesting numerical results
for one-dimensional strong nonlinear acoustic waves with shocks.
Introduction
In this contribution, our program is to use the lattice Boltzmann schemes with multiresolution relaxation times [8] for
the approximation of the full compressible Navier Stokes equations. Interesting results have been obtained previously
in [11, 13] when using a single particle distribution. A popular idea proposed in [1, 10, 16, 17] is the use of several
particle distributions. One particle distribution is devoted to the conservation of mass and momentum and another one
to the conservation of energy. This framework has been used in the context of Bhatnagar Gross Krook [2] approxi-
mation [6, 9, 14, 15]. The adaptation to the multiresolution relaxation times approach is not straightforward [4] and
source terms must be added to a pure “collide-stream” algorithm in order to capture this nonlinear dissipation term.
In this contribution, we begin with elementary one-dimensional fluid flow with the D1Q3 multi relaxation times lattice
Boltzmann scheme. The Taylor expansion method allows us to determine the underlying partial differential equations.
We adapt this study to the case of only one advection-diffusion equation. The coupling of these two schemes for
the approximation of the Navier Stokes equations leads to difficulties. This defect has its origin in precise algebraic
properties of lattice Boltzmann schemes and associated moments. We then consider lattice Boltzmann schemes with
source terms. We focus on a formulation of the Navier Stokes equations that takes into consideration the production
of entropy. The coupling of two D1Q3 schemes is proposed and studied and first numerical results are presented.
Fluid flow with a multi relaxation times D1Q3 lattice Boltzmann scheme
We describe in this section a very simple but fundamental lattice Boltzmann scheme. We restrict ourselves to one space
dimension. The mesh is parameterized by a space step ∆x. The nodes are located at the vertices x = j ∆x, where j is
an integer. We suppose given three discrete velocities: v0 = 0, v+ = +1 and v− = −1. A first description of the degrees
of freedom is associated to the density f ≡ ( f0 , f+ , f− ) of particles. In the D1Q3 case, we have three kind of particles.
The motionless particles f0 (x, t), the particles f+ (x, t) going from x to x + ∆x during one time step ∆t and the particles
f− (x, t) going from x to x − ∆x during the time step. We adopt the so-called “acoustic scale” and the numerical velocity
λ ≡ ∆x∆t is supposed fixed for the
Pentire contribution. From this distribution of particle densities, we construct two first
moments: mass density ρ ≡ ρ0 j=0, +, − f j = ρ0 ( f+ + f− + f0 ) and momentum J ≡ ρ0 λ j=0, +, − v j f j = ρ0 λ ( f+ − f− ).
P
040003-1
The third moment is the so-called “energy” defined by e ≡ ρ0 λ2 ( f+ + f− − 2 f0 ). The second description of the degrees
of freedom is due to the moments m ≡ (ρ, J, e). An invertible matrix M is therefore defined between the particles and
the moments:
ρ
1 1 1
f0
m ≡ J = MD1Q3 f+ ≡ MD1Q3 f , MD1Q3 = ρ0 0 λ −λ .
(1)
−2 λ2 λ2 λ2
e f−
The evolution algorithm is composed by two steps: relaxation and advection. In the relaxation step, density and
momentum remain at equilibrium and do not change: ρ∗ = ρ, J ∗ = J. On the contrary, the equilibrium energy is a
given function of the two moments at equilibrium: eeq = eeq (ρ, J). The energy after relaxation is obtained by a
simple evolution: e∗ = e + se (eeq − e). The relaxation parameter se must be chosen satisfying 0 < se < 2 [12]. The
particle distribution f ∗ after relaxation is defined by f ∗ = MD1Q3
−1
(ρ, J, e∗ )t . The advection step is a free displacement
of the particles during one time step: f0 (x, t+∆t) = f0 (x, t), f+ (x, t+∆t) = f+∗ (x−∆x, t) and f− (x, t+∆t) = f−∗ (x+∆x, t).
∗
We conduct the analysis of the multi relaxation times with the Taylor expansion method [3]. This approach is a kind
of numerical Chapman-Enskog expansion with the space step ∆x as a small parameter. Equivalent partial differential
equations emerge from this analysis for the conserved variables. We have for the previous example up to order 2:
2 1 1
∂t ρ + ∂ x J = O(∆x2 ) , ∂t J + ∂ x λ2 ρ + eeq − σe ∆t ∂ x θe = O(∆x2 ) , (2)
3 3 3
with a coefficient σe = s1e − 12 introduced by Hénon [7] and θe ≡ ∂t eeq + λ2 ∂ x J ' 3 λ2 ρ ∂ x u for the previous fluid
D1Q3 scheme. We can compare these equations (2) with the unidimensional
“isentropic” Navier Stokes equations
∂t ρ + ∂ x J = 0 for mass conservation and ∂t J + ∂ x ρ u2 + p − ∂ x ρ ν ∂ x u = 0 for momentum conservation. Then the
energy at equilibrium follows the relation
eeq = 3 ρ u2 + p − 2 λ2 ρ . (3)
The relaxation parameter is defined in (8) and is related to kinematic viscosity: ν0 = λ ∆x σe .
In all generality, if the moments m are related to the particle distribution f via a relation of the type m = M f with a
invertible matrix M, we introduce [3] the momentum-velocity tensor Λ according to
X
Λk` ≡ λ Mk j v j (M −1 ) j` . (4)
j
For the previous D1Q3 model, with the ordering ρ, J, e of the moments m, we have
0 1 0
ΛD1Q3 = 23 λ2 0 13 .
(5)
0 λ2 0
The vector of conserved moments is denoted by W and the defect of conservation θ` is defined according to
X
eq eq
θ ` ≡ ∂ t m` + Λ`p ∂ x m p (6)
p
for the non-conserved moments m` . Then the kth conserved moment Wk satisfies asymptotically the following second
order partial differential equation
X X
eq
∂t Wk + Λk` ∂ x m` − ∆t σ` Λk` ∂ x θ` = O(∆x2 ) . (7)
` `
040003-2
Advection-diffusion with a multi relaxation times D1Q3 lattice Boltzmann scheme
We modify the notations of the previous D1Q3 “MRT” scheme, replacing the particle distribution f by the notation g:
g0 (x, t) is the motionless particles at x during the time step ∆t, g+ (x, t) the density of particles going from x to x + ∆x
during one time step and g− (x, t) the density of particles going from x to x − ∆x during ∆t. The numerical velocity λ
remains fixed. We introduce three moments: a conserved variable ζ, momentum ψ and associated energy ε in a way
analogous to (1): ζ ≡ ζ0 (g0 +g+ +g− ), ψ ≡ ζ0 λ (g+ −g− ) and ε ≡ ζ0 λ2 (g+ +g− −2 g0 ). An invertible matrix M e is defined
by the condition (ζ, ψ, ε) = M
t eD1Q3 (g0 , g+ , g− ) . It is analogous to the matrix introduced in (1) except that the scaling
t
factor ρ0 is replaced by ζ0 . For this model, only one moment is conserved and remains at equilibrium: ζ ∗ = ζ. The
momentum ψ and associated energy ε at equilibrium are given functions of the scalar ζ : ψeq = ψeq (ζ) and εeq = εeq (ζ).
The momentum and energy after relaxation ψ∗ and ε∗ are obtained by simple evolutions: ψ∗ = ψ + sψ (ψeq − ψ),
ε∗ = ε + sε (εeq − ε) with 0 < sψ < 2 and 0 < sε < 2. The particle distribution g∗ after relaxation is obtained from
the moments: (g∗0 , g∗+ , g∗− )t = ( M
eD1Q3 )−1 (ζ, ψ∗ , ε∗ )t . The advection step is identical to the corresponding one for the
fluid scheme: g0 (x, t + ∆t) = g0 (x, t), g+ (x, t + ∆t) = g∗+ (x − ∆x, t) and g− (x, t + ∆t) = g∗− (x + ∆x, t). We can analyze
∗
formally this scheme with the Taylor expansion method. With the matrix Λ introduced in (5), the equivalent partial
differential equation up to order 2 takes the form
∂t ζ + ∂ x ψeq − σψ ∆t ∂ x θψ = O(∆x2 ) (9)
with σψ = s1ψ − 12 and θψ ≡ ∂t ψeq + ∂ x ( 23 λ2 ρ + 13 εeq ). This “thermal” lattice Boltzmann model is well adapted
for the simulation of an advection-diffusion equation of the type ∂t ζ + ∂ x (u0 ζ) − ∂ x (κ ∂ x ζ) = 0. The second order
differential equation (9) simulated by the thermal lattice Boltzmann scheme is a good approximation to first order
of (9) if the momentum at equilibrium ψeq is given by ψeq = u0 ζ. The defect of equilibrium can be expanded:
θψ ≡ ( 32 λ2 − u20 ) ∂ x ζ + 31 ∂ x εeq + O(∆x) and the energy at equilibrium is proportional to the conserved variable:
εeq = α λ2 ζ. The identification of second order terms induces: ( 2+α3 λ − u0 ) ( sψ − 2 ) ∆t = κ with a typical constraint
2 2 1 1
for numerical stability: −2 < α < 1. We take simply sε = 1.5 for the third moment.
040003-3
eq eq
that m∗k = (1 − sk ) mk + sk mk and (iii) the “second non-conserved” moments such that m∗` = m` − s` (m` − m` ) +
eq
K`k (mk − mk ), with mk in the first family of nonconserved moments. Our results [4] are not entirely satisfying and we
have changed our point of view.
Lattice Boltzmann algorithm for the simulation of the thermal Navier Stokes equations
The important remark is that we can re-interpret the second equation of (11) as an usual advection of the particle distri-
bution with a source term. In the following, we approach with a lattice Boltzmann scheme the full Navier Stokesequa-
tions as a system of conservative partial differential equations with a source term: ∂t W + ∂ x F(W) − ∂ x Φ(W, ∇W) = S .
This choice is motivated by the fact that the equivalent equations of a lattice Boltzmann scheme are always conser-
vative. In this contribution, we have chosen as “conserved variables” the mass ρ, the momentum J ≡ ρ u and the
volumic entropy. The volumic entropy ζ is the product of the volumic mass ρ multiplied by the specific entropy s.
Then W = (ρ, J, ζ). The equation of state is the one of a polytropic perfect gas
ρ γ γ (s − s )
0
p = (γ − 1) ρ i = ρ r T = p0 exp . (12)
ρ0 cp
The sound velocity c satisfies c2 = γ ρp = γ (γ − 1) i as previously. With this (ρ, J, ζ) formulation, the Navier Stokes
equations express the conservation of mass and momentum, and the production of entropy:
κ ρ ν 2 κ 2
∂t ρ + ∂ x J = 0, ∂t J + ∂ x ρ u2 + p − ∂ x ρ ν ∂ x u = 0, ∂t ζ + ∂ x ζ u − ∂ x ∂x T = ∂ x u + 2 ∂ x T . (13)
T T T
We use two particle distributions f and g as previously. The first distribution f devoted to mass and momentum and
the second distribution g to volumic entropy. The conserved moments are defined from the double particle distribution
according to
ρ = ρ0 f0 + f+ + f− , J = ρ0 λ f+ − f− , ζ = ρ0 c p g0 + g+ + g− . (14)
The nonconserved moments are defined thanks to the previous considerations:
e = ρ0 λ2 f+ + f− − 2 f0 , ψ = ρ0 c p λ (g+ − g− ), ε = ρ0 c p λ2 g+ + g− − 2 g0 . (15)
We apply the general theory of multi relaxation times lattice Boltzmann schemes with the particle representation
fd ≡ ( f0 , f+ , f− , g0 , g+ , g− )t and the momentum representation m ≡ (ρ, J, ρ E, e, ψ, ε)t . We have m = MD1Q3Q3 fd .
The matrix MD1Q3Q3 between particles and moments is given by the relation
1 1 1 0 0 0
0 λ −λ 0 0 0
0 0 0 cp cp c p
MD1Q3Q3 = ρ0 . (16)
−2 λ λ2 2
λ2
0 0 0
0
0 0 0 c p λ −c p λ
0 0 0 −2 c p λ2 c p λ2 c p λ2
We must now specify the equilibrium functions for non-conserved moments:
eeq = eeq (ρ, J, ζ) , ψeq = ψeq (ρ, J, ζ) , εeq = εeq (ρ, J, ζ) . (17)
The coefficients se , sψ and sε for the relaxation of non-conserved moments, i.e.
e∗ = e + se (eeq − e) , ψ∗ = ψ + sψ (ψeq − ψ) , ε∗ = ε + sε (εeq − ε) . (18)
The time iteration of the D1Q3Q3 scheme can be simply written:
f0 (x, t + ∆t) = f0∗ (x, t), f+ (x, t + ∆t) = f+∗ (x − ∆x, t), f− (x, t + ∆t) = f−∗ (x + ∆x, t)
(
(19)
g0 (x, t + ∆t) = g∗0 (x, t), g+ (x, t + ∆t) = g∗+ (x − ∆x, t), g− (x, t + ∆t) = g∗− (x + ∆x, t) .
The adaptation to the presence of source terms is described in [5]. The gradients ∂ x u and ∂ x T are evaluated with
second order centered finite differences.
040003-4
We first decouple and simplify the system (20). We obtain after this operation a simple system:
ν0 2
∂t ρ + ∂ x J = 0, ∂t J + (c20 − u20 ) ∂ x ρ + 2 u0 ∂ x J − ν0 ∂2x ρ + ν0 u0 ∂2x J = 0, ∂t ζ + u0 ∂ x ζ − γ ∂ ζ = 0. (21)
Pr x
The system of equations (21) is nothing else than the juxtaposition of a linearized version of the system (2) of fluid
equations and an advection-diffusion equation of the type (9). Then it is easy to derive an equation for the equilibrium
moment: 32 λ2 ρ + 31 eeq = (c20 − u20 ) ρ + 2 u0 J. The defect of conservation θe can be evaluated in a pure algebraic way:
θe = 3 (λ2 − 3 u20 − c20 ) ∂ x J − 6 u0 (c20 − u20 ) ∂ x ρ ' 3 λ2 ∂ x J. The coefficient of relaxation se is determined through the
kinematic viscosity: ν0 = σe λ ∆x. From the scalar “thermal” equation, we deduce an expression for the equilibrium
moment ψ = u0 ζ. For the defect of conservation θψ , we have θψ = 3 λ − u0 ∂ x ζ + 13 ∂ x εeq . The compatibility at
eq 2 2 2
D1Q3Q3 lattice Boltzmann scheme for the volumic entropy Navier Stokes equations
The matrix MD1Q3Q3 between particles and moments is still given by the relation (16). The equilibrium of non-
conserved moments (17) is parameterized by nonlinear functions. We must also specify the coefficients se , sψ and
sε for the relaxation (18) of the non-conserved moments. The discrete time iteration of the D1Q3Q3 scheme follows
the relations (19). The Taylor expansion method at first order proposes partial differential equations satisfied by the
eq
conserved variables W = (ρ, J, ζ ≡ ρ s): ∂t Wk + Λk` ∂ x m` = O(∆t) with a momentum-velocity tensor Λ given in (4).
We have for this D1Q3Q3 lattice Boltzmann scheme
0 1 0 0 0 0
2 λ 2
0 0 1
0 0
3 3
0 0 0 0 1 0
Λ = . (25)
0 λ
2
0 0 0 0
0 2 3λ 0 0 13
0 2
0 0 0 0 λ2 0
The equivalent equations at first order take the form
2 λ2 1
∂t ρ + ∂ x J = O(∆x) , ∂t J + ∂ x ρ + eeq = O(∆x) , ∂t ζ + ∂ x ψeq = O(∆x) . (26)
3 3
They are compared to the Navier stokes equations of gas dynamics (13). By identification of first order terms, the
“energy” at equilibrium eeq is given by the relation (3) and
ψeq = ζ u . (27)
Thus two equilibria for nonequilibrium moments are fixed.
For the second order analysis, the Hénon’s coefficients [7] are defined according to
1 1 1 1 1 1
σe = − , σψ = − , σε = − . (28)
se 2 sψ 2 sε 2
040003-5
The defects of conservation (6) have to be estimated before the determination of the second order equivalent partial
differential equations (7). For the present D1Q3Q3 lattice Boltzmann scheme, we have
∆t
∂t ρ + ∂ x J = O(∆t) , ∂t J + ∂ x ρ u2 + p = σe ∂ x θe + O(∆x2 ) , ∂t ζ + ∂ x ζ u = ∆t σψ ∂ x θψ + O(∆x2 ) . (29)
3
The first defect of conservation θe ≡ ∂t eeq + λ2 ∂ x J can be approximated by 3 λ2 ρ ∂ x u. Then the relation
ν = σe λ ∆x (30)
determines the relaxation coefficient se thanks to (28). For the second defect of conservation θψ ≡ ∂t ψ + λ ∂ x ζ +
eq 2
3
2
h
3 ∂ x ε , we have, after integration by parts and using the first order equations (26), θψ = p ∂ x s + ∂ x 13 2 λ2 ζ + εeq −
1 eq
i h i
((ρ u2 + p) s) . This expression is approximated by ∂ x 31 2 λ2 ζ + εeq − (ρ u2 + p) s . The entropy equation (third
equation of (13)) is compared with the third equation of (29):
n h1 io
∂t ζ + ∂ x ζ u − ∂2x σψ ∆t 2 λ2 ζ + εeq − (ρ u2 + p) s = O(∆x2 ) . (31)
3
h i
Then σψ ∆t 13 2 λ2 ζ + εeq − (ρ u2 + p) s = κ log TT0 . With the relations (23) and (30), we obtain for the third
momentum for entropy
2 ρ c p λ2 T h i
εeq = log + 3 ρ u2 + p − 2 λ2 ρ s . (32)
γ T0
The relation (32) is clearly non-trivial. It is an interesting property of the present scheme to have determined this
algebraic relation in the full nonlinear
h framework.
For
the polytropic
ρ i h perfect
gas, the pressure
i is given by (12) and the
relation (32) becomes ε = 2 λ ρ (s − s0 ) + 1 − γ c p ρ log ρ0 + 3 ρ u + p − 2 λ ρ s.
eq 2 1 2 2
The present D1Q3Q3 algorithm can be implemented in the following way: (i) consider the particle distribution fd ≡
( f0 , f+ , f− , g0 , g+ , g− )t , (ii) compute the six moments m ≡ (ρ, J, ζ, e, ψ, ε)t , by m = MD1Q3Q3 fd with MD1Q3Q3
defined by the relation (16), (iii) fix the relaxation coefficients se , sψ , sε of the non-conserved moments e, ψ, ε through
the associated σ0 s thanks to (28) with σe determined by (30), σψ according to (23) and sε = 1.5, (iv) determine
the moments eeq , ψeq , εeq at equilibrium with the relations (3), (27) and (32) respectively, (v) compute the moments
e∗ , ψ∗ , ε∗ after relaxation with the relations (18), (vi) compute the particle distribution fd∗ after relaxation thanks
to fd∗ = (MD1Q3Q3 )−1 (ρ, J, ζ, e∗ , ψ∗ , ε∗ )t , (vii) evaluate the gradients ∂ x u and ∂ x T with second order centered finite
differences, add a source term to the momentum ζ as proposed by the third equation in (13) and the algorithm described
in [5], (viii) iterate the scheme in time according to (19).
FIGURE 1. Progressive linear wave, periodic boundary conditions, δρ = 0.001 ρ0 . Density and velocity (left). Entropy (right).
040003-6
FIGURE 2. Progressive nonlinear wave, periodic boundary conditions, δρ = 0.01 ρ0 . Density and velocity (left). Entropy (right).
FIGURE 3. Strong nonlinear wave, periodic boundary conditions, δρ = 0.1 ρ0 . Density and velocity (left). Entropy (right).
FIGURE 4. Strong nonlinear wave (δρ = 0.1 ρ0 ). Time evolution of entropy and total energy.
With the second experiment described in Figure 2, nonlinear effects are clearly visible for the density and velocity
results. In order to validate these effects, we have developed a finite difference software based on the formulation (13)
and using second order centered finite differences and explicit first order time integration. The variation of entropy is
ρ0
very small and is of good quality. With δρ = 100 , a shock wave is generated (see Figure 3). The production of entropy
is essentially localized in the region of high variation of the fields. On Figure 4, we present the transient evolution
of specific entropy and specific energy. Without the dissipation source term of (13), the total energy if not perfectly
conserved and the specific entropy is quasi constant. With this source term, the defect for total energy is very small
040003-7
and the production of entropy remains moderate in comparison to our finite difference simulator. The extension of
the previous work to two and three space dimensions is the next step of this work, typically with the coupled lattice
Boltzmann schemes D2Q9-D2Q5 and D3Q19-D3Q7.
Acknowledgements
This work is supported by the French “Climb” Oseo project. A part of this work has been realized during the stay
of two of us at the Beijing Computational Science Research Center. We thank the colleagues of CSRC for their
hospitality. Last but not least, the authors thank the referee who suggested several points in need of improvement.
REFERENCES
[1] F. Alexander, S. Chen and J. Sterling, “Lattice Boltzmann thermohydrodynamic”, Physical Review E, 47,
R2249-R2252 (1993).
[2] P. Bhatnagar, E. Gross and M. Krook, “A Model for Collision Processes in Gases. I. Small Amplitude Pro-
cesses in Charged and Neutral One-Component Systems”, Physical Review, 94, 511-525, 1954.
[3] F. Dubois, “Equivalent partial differential equations of a lattice Boltzmann scheme”, Computers and Mathe-
matics with Applications, 55, 1441-1449 (2008).
[4] F. Dubois, “Liens entre des schémas de Boltzmann sur réseau et les équations de Navier Stokes thermiques”,
Seminar, Institut Henri Poincaré, Paris, february 2016.
[5] F. Dubois, P. Lallemand, M.M. Tekitek, “Taylor expansion method for linear lattice Boltzmann schemes with
an external drift. Application to boundary conditions”, Springer Lecture Notes in Computational Science and
Engineering”, 99, R. Abgrall et al. Editors, 89-107 (2014).
[6] Z. Guo, C. Zheng and B. Shi, “Thermal lattice Boltzmann equation for low Mach number flows: decoupling
model”, Physical Review E, 75, 036704 (2007).
[7] M. Hénon, “Viscosity of a Lattice Gas”, Complex Systems, 1, 763-789 (1987).
[8] D. D’Humières, “Generalized Lattice-Boltzmann Equations”, in: AIAA Rarefied Gas Dynamics: Theory and
Applications, Progress in Astronautics and Aeronautics, 159, AIAA, Washington, D.C., 450-458 (1992).
[9] I.V. Karlin, D. Sichau and S.S. Chikatamarla, “Consistent two-population lattice Boltzmann model for thermal
flows”, Physical Review E, 88, 063310 (2013).
[10] B. Khobalatte and B. Perthame. “Maximum principle on the entropy and second-order kinetic schemes”,
Mathematics of Computation, 62, 119-131, 1994.
[11] P. Lallemand and F. Dubois, “Comparison of Simulations of Convective Flows”, Communications in Compu-
tational Physics, 17, 1169-1184 (2015).
[12] P. Lallemand and L.S. Luo, Theory of the lattice Boltzmann method: Dispersion, dissipation, isotropy,
Galilean invariance, and stability, Physical Review E, 61, 6546-6562, (2000).
[13] P. Lallemand and L.S. Luo, “Theory of the lattice Boltzmann method: Acoustic and thermal properties in two
and three dimensions”, Physical Review E, 68, 036706 (2003).
[14] Q. Li, Y.L. He, Y. Wang and W.Q. Tao, “Coupled double-distribution-function lattice Boltzmann method for
the compressible Navier-Stokes equations”, Physical Review E, 76, 056705 (2007).
[15] Q. Li, K.H. Luo, Y.L. He, Y.J. Gao and W.Q. Tao, “Coupling lattice Boltzmann model for simulation of
thermal flows on standard lattices”, Physical Review E, 85, 016710, (2012).
[16] G. McNamara and B. Alder, “Analysis of the lattice Boltzmann treatment of hydrodynamics”, Physica A, 194,
218-228 (1993).
[17] X. Shan and H. Chen, “Lattice Boltzmann model for simulating flows with multiple phases and components”,
Physical Review E, 47, 1815-1819 (1993).
040003-8