ContentslistsavailableatScienceDirect
Computers and Fluids
journalhomepage:www.elsevier.com/locate/compfluid
Benchmark solutions
Unexpected convergence of lattice Boltzmann schemes R
Bruce M. Boghosian
a, Francois Dubois
b,c,∗, Benjamin Graille
c, Pierre Lallemand
d, Mohamed Mahdi Tekitek
eaDepartment of Mathematics, Tufts University, Bromfield-Pearson Hall, Medford, MA 02155, USA
bConservatoire National des Arts et Métiers, Laboratoire de Mécanique des Structures et des Systèmes Couplés, Paris F-75003, France
cDepartment of Mathematics, University Paris-Sud, Orsay Cedex F-91405, France
dBeijing Computational Science Research Center, Zhanggguancun Software Park II, Haidian District, Beijing, 10 0 094, China
eDepartment of Mathematics, Faculty of Sciences of Tunis, University of Tunis El Manar, Tunis 2092, Tunisia
a rt i c l e i n f o
Article history:
Received 6 November 2017 Revised 3 April 2018 Accepted 27 April 2018 Available online 28 April 2018 MSC:
76M28 Keywords:
Heat equation Damped acoustic Dispersion equation Taylor expansion method
a b s t r a c t
Inthiswork,westudynumericallytheconvergenceofthescalarD2Q9latticeBoltzmannschemewith multiplerelaxationtimeswhenthetimestepisproportionaltothe spacestepand tendstozero.We do thisbyacombinationoftheoryand numerical experiment.The classicalformalanalysis whenall therelaxationparametersarefixedand thetimesteptendstozeroshowsthatthenumericalsolution convergestosolutionsoftheheatequation,withaconstraintconnectingthediffusivity,thespacestep andthe coefficientofrelaxationofthemomentum.Ifthediffusivityisfixedandthe spacesteptends tozero,therelaxationparameterforthemomentumisverysmall,causing adiscrepancybetweenthe previousanalysisandthenumericalresults.Weproposeanewanalysisofthemethodforthisspecific situationofevanescent relaxation,based onthedispersionequationofthelattice Boltzmannscheme.
Anewasymptoticpartialdifferentialequation,thedampedacoustic system,isemergentas aresultof thisformalanalysis.ComplementarynumericalexperimentsestablishtheconvergenceofthescalarD2Q9 latticeBoltzmann schemewith multiple relaxationtimes and acoustic scaling inthis specificcase of evanescentrelaxationtowardsthenumericalsolutionofthedampedacousticsystem.
© 2018ElsevierLtd.Allrightsreserved.
1. Introduction
LatticeBoltzmannmodelsaresimplificationsofthecontinuum Boltzmannequationobtainedbydiscretizinginbothphysicalspace andvelocityspace.Thediscretevelocitiesviretainedtypicallycor- respond to lattice vectors of the discrete spatial lattice. That is, each lattice vertexx is linked to a finite number ofneighboring verticesbylatticevectorsvit.Aparticledistributionfistherefore parameterizedbyitscomponentsineachofthediscretevelocities, the vertexxof thespatiallattice,andthediscrete timet.Atime stepofaclassicallatticeBoltzmannscheme[15]thencontainstwo steps:
(i)Arelaxationstepwherethedistributionfateachvertexxis locallymodifiedintoanewdistributionf∗,and
R Contribution to be published in Computers and Fluids , edition 31 March 2018.
∗ Corresponding author at: Department of Mathematics, University Paris-Sud, Or- say Cedex F-91405, France.
E-mail addresses: [email protected] (B.M. Boghosian), [email protected] (F. Dubois), [email protected] (B. Graille), [email protected] (P. Lallemand), [email protected] (M.M.
Tekitek).
(ii) an advection step based on the method of characteristics as an exact time-integration operator. We employ the multiple- relaxation-timeapproach introducedby d’Humières [10],wherein thelocalmapping f−→ f∗ isdescribedby adiagonaloperatorin aspaceofmoments.
In[6],wehavestudiedtheasymptoticexpansionofvariouslat- ticeBoltzmannschemeswithmultiple-relaxation timesfordiffer- entapplications.Weusedthe so-calledacoustic scaling,inwhich theratio
λ
≡x/tiskeptfixed.Wesupposedalsothattherelax- ationoperatorremainsfixed.Inthismanner,wedemonstratedthe possibility ofapproximating diffusion processes describedby the heatequation.The importance of using small values of relaxation parame- ters was recognized for linear viscoelastic fluids by Lallemand et al. [14]. Independently, unexpected results in simulations for advection-diffusionprocesses havebeen describedby Dellacherie in [4]. We have studied experimentally in [3] the curious con- vergence ofthe D1Q3 multiple-relaxation time lattice Boltzmann schemewithoneconservedvariablewhenusingtheacousticscal- inginonespatialdimension.Theasymptoticequationofthelattice Boltzmannschemeisnolongeranadvection-diffusionmodelbuta dampedacousticmodel.Inthiscontribution,weshowandanalyze https://doi.org/10.1016/j.compfluid.2018.04.029
0045-7930/© 2018 Elsevier Ltd. All rights reserved.
Fig. 1. Particle distribution f jfor 0 ≤j ≤8 of the D2Q9 lattice Boltzmann scheme.
an analogous phenomenon for two spatial dimensions with the scalarD2Q9latticeBoltzmannscheme.Thedifficultyconcernsthe highlightingoftheconvergencewiththenumericalexperiments.
In Section 2, we recall some fundamentals relative to the D2Q9 lattice Boltzmann scheme for scalar conservation laws. In Section3, westudyconvergence ofthisschemefordiffusive and acousticscaling. Aformal analysisis proposed in Section 4,with thedispersion equationmethod,initially proposedin[16].Wees- tablish that with acoustic scaling, the convergence of the scalar D2Q9schemeisnot theheat equationbutanunexpectedmodel!
Finally,westudytheexperimentalconvergenceofthescalarD2Q9 schemeinseveralsituationsinSection5.
2. ScalarD2Q9latticeBoltzmannschemeforthermalproblems
The D2Q9lattice Boltzmannscheme usesa setofdiscrete ve- locitiesdescribedinFig.1.Adensitydistributionfjisassociatedto eachvelocityvj≡
λ
ej,whereλ
≡ xt is the fixednumerical lattice velocity.The firstthree momentsforthedensityandmomentum aredefinedaccordingto⎧ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎨
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎩ ρ
=8 j=0
fj=m0,
Jx≡
ρ
ux=8j=0
λ
e1jfj=m1,Jy≡
ρ
uy=8j=0
λ
e2jfj=m2,(1)
wheretheeαj are the
α
thcartesiancomponents ofthevectorsejintroducedpreviously.Wecompletethissetofmomentsandcon- structavectormofmomentsmkaccordingto
m=M f, (2)
withaninvertiblefixedmatrixMusually[15]givenby
M=
⎛
⎜ ⎜
⎜ ⎜
⎜ ⎜
⎜ ⎜
⎜ ⎝
1 1 1 1 1 1 1 1 1
0
λ
0 −λ
0λ
−λ
−λ λ
0 0
λ
0 −λ λ λ
−λ
−λ
−4
λ
2 −λ
2 −λ
2 −λ
2 −λ
2 2λ
2 2λ
2 2λ
2 2λ
20
λ
2 −λ
2λ
2 −λ
2 0 0 0 00 0 0 0 0
λ
2 −λ
2λ
2 −λ
20 −2
λ
3 0 2λ
3 0λ
3 −λ
3 −λ
3λ
30 0 −2
λ
3 0 2λ
3λ
3λ
3 −λ
3 −λ
34
λ
4 −2λ
4 −2λ
4 −2λ
4 −2λ
4λ
4λ
4λ
4λ
4⎞
⎟ ⎟
⎟ ⎟
⎟ ⎟
⎟ ⎟
⎟ ⎠
.
ForscalarlatticeBoltzmannapplications,thedensity
ρ
isthe“con-servedvariable”.
• The particle distribution atequilibrium feq is a function only ofthisconservedvariable.ForthisthermalD2Q9latticeBoltz- mannscheme,thevectorofequilibriummomentsmeq isgiven by
meq=
ρ
,0,0,αλ
2ρ
,0,0,0,0,λ
4βρ
t. (3)
Inmostapplications,thecoefficients
α
andβ
areusually takentobe
α
=−2,β
=1. (4)The lattice Boltzmann scheme is comprised of two fundamental steps : Relaxation andadvection. During the relaxationstep, the conservedvariable
ρ
isnotmodified,andthenon-conservedmo- mentsm1tom8 relaxtowardsan equilibriumvalue:meqk =ψ
k(ρ
) fork≥1, wheretheψ
k arethe linearfunctionsofthe conserved momentgivenby (3).Thespecificationofthisstepalsoneedsre- laxationratessk:Fork≥1suchthatm∗k=mk+sk
meqk −mk
,
where thesuperscript ∗ denotes themoment mk after the relax- ationstep.Thetableofrelaxationparametersskchoseninoursim- ulationsisasfollows
[s]=
sJ,sJ,se,sx,sx,sq,sq,sε
. (5)
Weintroduce alsothe8×8diagonalmatrixSwhosediagonalel- ementsarethecomponentsofthevector[s].Inourcomputations, wetakethefollowingnumericalvalues
se=1.7,sx=1.1,sq=1.1,sε=1.7. (6) Onlytherelaxationcoefficient sJforthefirst ordermomentumis allowedtovaryinournumericalexperiments.
Then using the matrix M−1 the relaxation step becomes in f space:
fi∗
(
x,t)
=M−i1m∗. (7)Duringtheadvectionstepfi(xj)istransportedfromthenodexjby the discretevelocity vito the node xj+
v
it.Thus the evolution ofpopulationsfifor0≤i≤8atinternalnodexisdescribedby:fi
(
x,t+t
)
=fi∗(
x−v
it,t
)
,0≤i≤8. (8)•In[6],we haveanalyzed severallatticeBoltzmannmodelswith theTaylor-expansionmethod,includingthepresentonedefinedby Eqs.(2),(3),(5) and(8).Thehypothesis used wasthat therefer- encevelocity
λ
andtherelaxation coefficientssJ,se,sx,sq andsε remainconstantasthespatialstepxtendstozero.Thenthecon- servedvariableρ
satisfies(atleastformally!)theheatequation:∂ρ ∂
t −κ ρ
=O(
x2)
, (9)wherethethermaldiffusivity
κ
isgivenbytherelationκ
≡4+α
6
σλ
x,σ
≡1sJ−1 2
. (10)Thecoefficient
σ
isknownasthe“Hénonparameter” inreference tothepioneeringworkofHénon[9].Observethatwhentherelax- ationcoefficientsJ andthemeshvelocityλ
arefixed,thethermaldiffusivitytends tozero asthespacestep xtends tozero.This latticeBoltzmannschemeisstableinthefluidcase(see [15]) un- derthecondition:
−4<
α
<2.Forthescalarcase,thecondition
α
+4>0iscleartoassumethat the thermal diffusivityκ
is positive (see (10)) and theconditionα
<2 corresponds to our experimental know how. Observe thatwiththesechoices,thevalue oftherelaxationparametersJhasto befitwiththephysicaldiffusivity
κ
andthemeshsizexthrough the relation(10)ifthespacestep andtime step arevarying pro- portionately.Inparticular,wehavetheexpansionsJ=4+
α
6
κ λ
x+O(
x2)
(11)asxtendstozero.
• Diffusive scaling can alsobe used andwe refer, e.g., to the workofJunketal.[11].Inthiscase,theratio
(
x)
2t =
λ
xremains fixed. This diffusive scaling is intensively used with the explicit finite difference methodfor solving the heat equation. It iswellknown[17]thatthetimestepmustbeproportionaltothe square of the spacial step in order for the method to be stable.
An asymptotic analysiscan be done forthis simplelattice Boltz- mannthermic model,as, e.g.,inourcontribution[7],andwe ob- tain againtheheatEq.(9)asthescalinglimit ofthemodel.With thisdiffusive scaling, theparameters
σ
andsJ remain constant if the thermal diffusivity is given and the mesh size x tends to zero. Remark also that the convergence of the lattice Boltzmann schemewasrigorouslyprovedforthediffusivescalingforNavier–Stokesflowsinperiodicandboundeddomainsin[12]andforone dimensionalconvection-diffusion-reactionequationsin[13]. 3. Firstnumericalexperiments
We studythe diffusion of a Gaussian profile in a square do- main.Inordertocontrol thecomputercost duringthenumerical experimentandtobecertainthatthenumericalexperimentisnot polluted by the boundary scheme, we impose periodic boundary conditions.We usetwo variants ofthe scalarD2Q9lattice Boltz- mannscheme:Diffusiveandacousticscaling.
• ScalarD2Q9numericalexperimentswithdiffusivescaling
Wesolvenumericallytheheatequation
∂ρ ∂
t −κ ρ
=0, (12)inthesquare=[−1,1]2,withperiodicboundaryconditions.The initialconditionisaGaussian:
ρ
0(
x,y)
=exp−x2+y2 0.09
,−1≤x,y≤1. (13)Thecoefficients
α
andβ
oftheequilibriumare fixedaccordingto (4)andwekeepfixedtherelaxationcoefficientformomentum: sJ=32. (14)
Weusetheparticulardiffusivetimestept=x2.Then
σ
≡ s1J −1
2= 16 andthediffusivityfollowstherelation
κ
=σ3J andκ
= 181. (15)We have chosen an odd number ofmesh cells inthese numeri- cal experiments.With theconstraintt=x2, itis notpossible toobtain exactlythesameexactfinal time.We haveadaptedthe number of time steps in order to havevery close values forthe finaltimewiththedifferentmeshes.
• Comparisonwithfinite-differenceapproximation
Remarkthatthesolutionoftheheatequationonasquarewith an initial Gaussian and periodic boundary conditions has to our knowledgenoanalyticalsolution.Inconsequence,wecomparethe
Fig. 2. Two-dimensional heat Eq. (12) , κ= 181. D2Q9 scheme with diffusive scaling (left) vs. explicit finite differences (right) ; mesh 111 ×111, time = 0.19479.
Fig. 3. Two-dimensional heat Eq. (12) , κ= 181. D2Q9 scheme with diffusive scaling (left) vs. explicit finite differences (right) ; mesh 223 × 223, time = 0.16473.
Table 1
D2Q9 numerical experiments with diffusive scaling, s Jgiven by (14) and diffusivity κby (15) .
Number of cells 13 ×13 27 ×27 55 ×55 111 ×111 223 ×223 Nb. of time steps D2Q9 8 36 128 600 2048 Final time 0.18935 0.19753 0.17741 0.19479 0.16473
solution obtained by the lattice Boltzmann scheme withthe re- sultcomputedwithtwo-dimensionalfinitedifferences,centeredin spaceandexplicitintime.The degreesof freedomare locatedat half-integerpositions,exactly asdone withthelattice Boltzmann scheme:
ρ
in+12,j+12≈
ρ
i+1 2
x,
j+1 2
x,n
t
.
WefinitedifferencetheheatEq.(12)inthefollowingway:
⎧ ⎪
⎪ ⎪
⎨
⎪ ⎪
⎪ ⎩
1
t
ρ
in++112,j+12−
ρ
in+12,j+12
−
κ
1x2
ρ
in+32,j+12
−2
ρ
in+12,j+12
+
ρ
ni−12,j+12
+ 1
y2
ρ
in+12,j+32−2
ρ
in+12,j+12+
ρ
ni+12,j−12
=0.Weuseexactly thesamegrid inspaceforboth schemesandex- actly the sametime step(and in consequencethe same number oftimesteps).Theparameters forbothschemesare comparedin Table1.
Theresultsfollowwhatisexpected.Theapproximatesolutions ofbothschemesare verysimilar asobserved inFigs.2and3for 111×111 and223×223 meshes. The difference betweenthetwo schemesatthefinaltime ispresentedinFig.4.Theorderofcon- vergenceofthisresidualisapproximatelyoforder4.Sincethefi- nitedifferencemethodisofsecond orderaccuracy[17],thisindi- catesthatthelatticeBoltzmannmethodapproachestheheatequa- tionwithsecond-orderaccuracy.
Table 2
D2Q9 numerical experiments with acoustic scaling. The diffusivity κ= 181 is imposed in all the sim- ulations.
Number of cells 13 ×13 27 ×27 55 ×55 111 ×111 223 ×223
D2Q9 s Jparameter 1.5 1.182 0.830 0.52 0.298
Nb. of time steps D2Q9 8 16 32 64 128
Nb. of time steps, finite differences 8 32 128 512 2048
Final time 0.18935 0.18234 0.17902 0.17741 0.17661
Fig. 4. Two-dimensional heat Eq. (12) , κ= 181. Difference of the numerical results computed with the D2Q9 scheme with diffusive scaling and explicit finite differ- ences at the final times presented in Table 1 . The order of convergence for this residual in the L ∞norm is equal to 3.41 and in the L 2norm it is 3.96.
Fig. 5. Two-dimensional heat Eq. (12) , κ= 181. D2Q9 lattice Boltzmann scheme with acoustic scaling (left) vs. explicit finite differences (right) ; results at time = 0.17741 for a 111 × 111 mesh.
• ScalarD2Q9numericalexperimentswithacousticscaling
We still wish to solve the heat Eq. (12) in the square = [−1,1]2 withperiodic boundary conditions. The initial condition isagain givenby a Gaussian profile (13).The given diffusivity is imposed by the value (15). We adopt an acoustic scaling with t=xfortheD2Q9latticeBoltzmannsimulations.Wecompare the results with explicit finite differences; in this case, we take tࣃx2 and the time step is chosen in order to obtain exactly thesamefinaltimethanwiththelatticeBoltzmannmethod.
• ThenumericalresultspresentedinFigs.5and6forthetwo meshesof111×111and223×223seemcorrect.Butaquantitative examinationoftheresults(Fig.7) showsthat afteraconvergence similartotheoneobtainedfordiffusivescaling(seeFig.4),aper- sistentdifferenceappears.Thisqualitativebehaviorisverysimilar towhathasbeenobservedin[3]inonespatialdimension.
Fig. 6. Two-dimensional heat Eq. (12) , κ= 181. D2Q9 lattice Boltzmann scheme with acoustic scaling (left) vs. explicit finite differences (right) ; results at time = 0.1766 for a 223 ×223 mesh.
Fig. 7. Two-dimensional heat Eq. (12) , κ= 181, D2q9 scheme with acoustic scaling vs. explicit finite differences at the final times presented in Table 2 . There is no nu- merical evidence of coherence between the two methods when the mesh is refined.
Anewanalysisoftheschemeisnecessary toexplainthislack ofconvergencetowardstheexpecteddiffusivemodel.
4. Dispersionequationforanevanescentrelaxation
Inthissection, weproposeafirst-orderanalysiswhenthelast relaxationcoefficientsin (6)remainfixed orwhen therelaxation coefficientsJ forthemomentum Jfollowsthechoicepresentedin Eq.(11),idest
sJ= 4+
α
6
λ
2κ
t +O(
t2)
. (16)• Fixedrelaxations
Wewrite the relation(8)interms ofthemoments mdefined in(2):
mk
(
x,t+t
)
=j
Mk jM−j1m∗
(
x−v
jt,t
)
. (17)Before doinga Taylorexpansionatorder 1,weintroduce thefol- lowing“momentumvelocity” operatormatrixdefinedaccording to
k ≡ −
j α
Mk j
v
αjM−j1∂
∂
xα. (18)For the D2Q9 scheme, this matrix can be explicitly calculated [5]andwehave
(19)
Wesplitthemomentvectorintotwoblocks:
m=
W Y(20)
withW=
ρ
inourscalarexampleandYa columnvector with8components.Wedecompose alsotheoperatormatrixintofour blocksthatrespectthedecomposition(20):
≡ A B
C D
. (21)
In our case, A is a scalar 1×1 matrix, B has one line and 8 columns,Ciscomposedby8linesand1columnandDisa8×8 squarematrixasshownintheright-handsideofrelation(19).We can alsointroduce a constantmatrix E with8linesandone col- umnsuchthattherelation(3)canbewrittenintheform
Yeq≡Em. (22)
Therelation(17)isexpandedatfirstorder:
m+
t
∂
tm+O(
t2)
=m∗+t
m∗+O
(
t2)
(23)anddueto(22),wehave m∗=
I 0 SE I−Sm. (24)
Therelation(23)canbewrittenintheform Lm≡m∗−m+
t
−
∂
tm+m∗
=O
(
t2)
, (25) withL≡
0 0 SE −S+
t −
∂
t 00 −
∂
t+
A BC D
I 0 SE I−S. (26)
The dispersion relation associated with the relation (25) can be writteninasimpleway:
detL=0. (27)
We expand this determinant in order to eliminate the non- conserved moments Y. Moreover, due to the right-hand side of Eq.(25),wecanneglectallthetermsofsecondorthirdorderrel- ative tot.We write the expression(26) ofthe matrix Linthe form
L=
t
(
−∂
t+A+BSE)
tB(
I−S)
SE+
t
(
C+DSE)
−S+t
(
−∂
t+D(
I−S))
.
WeapplyGaussian eliminationinordertomakeexplicitthecon- dition(27).WemultiplythismatrixatleftbytheregularmatrixK definedby
K=
ItB
(
I−S)
S−10 I
. (28)
Thenwehave,aftersomelinesofalgebra, KL=
ItB
(
I−S)
S−1 0 I×
t
(
−∂
t+A+BSE)
tB(
I−S)
SE+
t
(
C+DSE)
−S+t
(
−∂
t+D(
I−S))
=
t
(
−∂
t+A+BSE)
+tB
(
I−S)
S−1SE O(
t2)
SE+O
(
t)
−S+O(
t)
andwehavethefollowingtriangularformfortheproductKL: KL=
t
(
−∂
t+A+BE)
O(
t2)
SE+O(
t)
−S+O(
t)
.
Thentherelation(27) isequivalentatfirstorderto thefollowing setoffirstorderpartialdifferentialequations:
(
−∂
t+A+BE)
W=O(
t)
, (29) recoveringthefirststepoftheBerlinalgorithmpresentedinAugier etal.[2].Forthescalardiffusionproblem,thisequation expresses simplythat∂
tρ
=O(
t)
.Thisresultisconsistentwiththesecond-order analysispresented attherelationin(9).
Whenweusediffusivescaling,thisdispersionequationcanbe adaptedinordertorecovertheheatequation atzeroorderofac- curacy.It isthenequivalenttotheTaylorexpansionmethodwith thediffusivescaling,asusedin[7].
• Evanescentrelaxations
Whentandxtendtozerowiththeacousticscaling,these twoinfinitesimalsareofthesameorder.Theexpansion(16)ofthe relaxationcoefficient sJ implies that theprevious asymptoticcal- culus has to be mademore precise. The coefficient sJ is now at first order proportional to the time step t. We decompose the non-conservedmomentsYintotwofamilies:The quasi-conserved momentsUidestthetwocomponentsofthemomentumJinthe scalarcase– andtheothertrulynon-conservedmomentsZ: Y=
U Z. (30)
The 8-componentvector Yis split into a first vectorU∈R2 and asecond oneZ with6components.Inother words,thefamilyof momentsissplitintothreecomponents:
m=
WU Z
.
Then the8×8relaxation matrixS can be decomposedintotwo blocks:
S=
tS+O
(
t2)
00 SZ
. (31)
Thetopleftblockintherighthandsideof(31)tendstozeroasthe meshisrefined.TheequilibriumvectorEisnaturallysplitintothe quasi-conservedcomponentEU andthe trulyrelaxingcomponent EZ:
E=
EUEZ
. (32)
We have: SE=
tS 00 SZ
EU EZ=
tSEUSZEZ
and the relation (24)takestheform
m∗=
I 0 0tSEU I−
tS 0 SZEZ 0 I−SZ
m. (33)
Then the momentum velocity operator matrix is split into 9 blocks:
= A A
2 B1
A3 A4 B2 C1 C2 D4
. (34)
Thisblockstructure (34)isexplicitlygivenforourthermal D2Q9 intheform
Then L=
0 0 0tSEU −
tS 0 SZEZ 0 −SZ
−
t
∂
t 0 00
∂
t 0 0 0∂
t+
t A A
2 B1
A3 A4 B2
C1 C2 D4
I 0 0
tSEU I−
tS 0 SZEZ 0 I−SZ
.
This expression can be expanded tofirst order int without anychangeintheresultoftheGaussianelimination.Thenwecan neglectthe terms of order one int inthe last product oftwo matrices.Weobtain
A A2 B1 A3 A4 B2
C1 C2 D4
I 0 0
0 I 0
SZEZ 0 I−SZ
=
A+B1SZEZ A2 B1
(
I−SZ)
A3+B2SZEZ A4 B2(
I−SZ)
C1+D4SZEZ C2 D4(
I−SZ)
, and,uptoorderO(t),wehave
L=
t
(
−∂
t+A+B1SZEZ)
tA2tB1
(
I−SZ)
t(
SEU+A3+B2SZEZ)
t(
−S−∂
t+A4)
tB2(
I−SZ)
SZEZ+
t
(
C1+D4SZEZ)
tC2 −SZ+t
(
−∂
t+D4(
I−SZ)
.
(35) WiththemethodofGaussianeliminationusedpreviously,wemul- tiplythematrixLobtainedin(35)ontheleftbythefollowingma- trix
K=
I 0tB
1
(
I−SZ)
SZ−1 0 ItB2
(
I−SZ)
SZ−10 0 I
whosedeterminantis equal to1. After some elementaryalgebra, weobtain
KL=
t
(
−∂
t+A+B1EZ)
tA2 O(
t2)
t(
SEU+A3+B2EZ)
t(
−S−∂
t+A4)
O(
t2)
SZEZ+O(
t)
tC2 −SZ+O(
t)
.
Ononehand,detK=1andontheotherhand,thelastcolumnof thematrixKLiscomposedofnegligibletermsexceptforthelast
one. Thenwe havethe condition(27) ifandonlyifthe determi- nant ofthe 2×2 upperblock matrix isnull. In other terms,this matrixhasanontrivialkernelatorderonerelativeto tandwe have
−
∂
t+A+B1EZ A2SEU+A3+B2EZ −
∂
t−S+A4 W U=O
(
t)
. (36)Then the equivalent partial differential equations are written as a system involving theconserved variable W andthe quasi con- servedmomentsU:
∂
tW =(
A+B1EZ)
W+A2U+O(
t)
∂
tU+SU=(
A3+B2EZ+SEU)
W+A4U+O(
t)
. (37) Thisresultgeneralizesthe firstanalysisdone in[3]fortheD1Q3 scheme. When we replace the block matrices introduced in the relations (31), (32) and (34) by their D2Q9 values, we estab- lishthat withthe acousticscaling, the scalarD2Q9 lattice Boltz- mannscheme withacoustic scaling admitsthe following asymp- toticdampedacousticmodel⎧ ⎪
⎨
⎪ ⎩
∂ρ ∂
t +divJ=O(
x)
∂
Jα∂
t +c20∂ρ
∂
xα +gJα=O(
x)
,1≤α
≤2, (38)withasoundvelocityc0 andadampingcoefficientggivenbythe relations
c20=
λ
26
(
4+α )
, g= c20κ
. (39)The above is a very interesting analysis, and clearly the correct two-dimensional analog ofthe earlier resultfor D1Q3. We point outthatitisequivalenttoadampedwaveequation.
5. ScalarD2Q9schemeconvergingtowardsdampedacoustic
Wehave nowtwo partialdifferentialequations withwhichto comparethenumericalsolutionobtainedwiththescalarD2Q9lat- ticeBoltzmannscheme: Theinitialheat Eq.(12)andthedamped acoustic system (38). We first consider numerical experiments done in Section 3 and compare our previous results with this new model. We also study in detail the eigenmodes of the sys- tem (38)and propose a simple numerical experimentwith a si- nusoidal analytic solution.The evolution of an initial Gaussian is againperformed,withtwodiffusioncoefficientsvaryingbyoneor- derofmagnitude.
• Dampedacousticsasalimitingmodelforthepreviousnumeri- calexperiments?
We wish to approximate the system of damped acoustic Eqs.(38).Thesoundvelocityisgivenby(39).Withthechoice(4), we obtain the classical value c0= √λ3. The imposed diffusivity
κ
andtherelation(39)fixthevalueg=6forthezero-orderdamp- inginthemomentumequationof(38).Thegeometryisthesquare =[−1,1]2withperiodicboundaryconditions.Theinitialdensity isstill givenbya Gaussianprofile(13). Becausethemomentum J atequilibriumisidenticallynull,wehavetakenthisspecificvalue asinitialconditionofourlattice Boltzmannsimulations.Wesup- poseinconsequencethat theinitialconditionforthemomentum issimplyJ(x,t=0)=0.
We adoptacoustic scaling with t=xfor the D2Q9 lattice Boltzmann simulations. Forthe acoustic system(38), we useex- plicit finite differences withstaggered grids, hereafter named as
“HaWAY” method and described with some details in the Ap- pendix.Inthiscase,theacoustictimesteptaforfinite-difference
Table 3
Numerical experiments with the D2Q9 lattice Boltzmann test case studied in Section 3 com- pared with the damped acoustic model (38) simulated with the HaWAY method.
Number of cells 13 ×13 27 ×27 55 × 55 111 ×111 223 ×223 D2Q9 s Jparameter 1.5 1.182 0.830 0.52 0.298
Nb. of time steps d2q9 8 16 32 64 128
Idem, finite differences 32 64 128 256 512
Final time 0.18935 0.18234 0.17902 0.17741 0.17661
Fig. 8. Numerical results for the damped acoustic model with the experimental plan proposed in Table 3 .
Fig. 9. Two-dimensional wave with wave vector k = 2 π(1 , 1). Initial condition.
simulations is proportional to the spatial step x, with a sta- bility constraint. The corresponding experimentsare describedin Table3.
The resultsobtained withthisnewexperimentare very simi- lar tothe one obtainedin Section 3.In particular, the numerical resultscomputedwiththedampedacoustic modelareveryclose totheonespresentedinFigs.5and6.Whenwelooktothecon- vergence withquitefinegrids (Fig.8),thesignalisbetterthan in Fig.7butthisexperimentisstillnotentirelyconvincing.
• Wavesforthedampedacousticmodel Wesearchmodesofthetype
ρ
=ρ
0exp(
−γ
t+ik·x)
J=J0exp
(
−γ
t+ik·x)
(40)Fig. 10. Two-dimensional wave with wave vector k = 2 π(1 , 1), g ࣃ 6, 2| k | c 0ࣃ 5.924.
Autocorrelation of density.
Fig. 11. Two-dimensional wave with wave vector k = 2 π(1 , 1), g ࣃ 6, 2| k | c 0ࣃ 5.924.
Convergence towards the damped acoustic model (38) .
forthedampedacousticmodel38–(39).Thenwehavetosolvethe followingill-posedlinearsystem:
−
γ ρ
0+ik·J0=0, ik·ρ
0+(
g−γ )
J0=0. (41) Afirstsolutionisatransversestationarywave withγ
=g,ρ
0=0 andk·J0=0.We donot considerthismode inthiscontribution.Thentheothermodessatisfythefollowingdispersionrelation
γ
2−gγ
+|
k|
2c20=0. (42) Thisequationhascomplexpropagativerootswheng<2
|
k|
c0, (43)Fig. 12. Two-dimensional wave with wave vector k = 2 π(1 , 1), g ࣃ 5.6470. Autocor- relation of density with 2| k | c 0ࣃ 5.924 for various meshes.
Fig. 13. Two-dimensional wave with vector k = 2 π(1 , 1), g ࣃ 5.6470. Convergence towards the damped acoustic model (38) with 2| k | c 0ࣃ 5.924. The order of conver- gence is 1.27 for the L 2norm and 1.30 in norm L ∞.
i.e.,whenthediffusivity
κ
issufficientlylargemeasuredinascalesystembasedonthesoundvelocityandwavenumber:
κ
> c02
|
k|
.Inthatcase,theeigenvalue
γ
takestheformγ
=2g∓iω
,ω
=|
k|
2c20−g24. (44)
Theeigenvectorsarefinallygivenaccordingto
⎧ ⎨
⎩
ρ
=ρ
0exp−g 2t
exp
i
(
k·x±ω
t)
J=i k
|
k|
2ρ
0 −g 2∓iω
exp
−g 2texp
i
(
k·x±ω
t)
. (45)
Weconsiderapureanalyticaltestcaseasthenextexperiment.
• Atwo-dimensionalsinusoidalwave
We keep the value
κ
= 181 0.05555 of the diffusivity intro- ducedin(15). We usethe traditionalvalue c0= √13 andthe dis- sipationcoefficient g (see (39)) is still equalto g=6.We change thedomain and consider[0, 2π
]2 with the initial conditionρ
=cos
(2
π
(x+y)andJ=0.Then k=2
π
(1,1) andthe right-hand side of (43) is 2|k|c0ࣃ5.924. In this case, the damped acoustic model(38)exhibitsanon-propagativemode.Theinitial conditionispresentedinFig.9.Theautocorrelation ofdensity
(
t)
≡ρ (
x,t) ρ (
x,0)
dx| ρ (
x,0) |
2dxis typical ofa diffusion process asshownin Fig.10. The conver- genceforsimpledyadicmeshesispresentedinFig.11.
Asecond numericalexperimenthasbeenconducted. We keep the same domain [0, 2
π
]2 with the same initial conditionρ
= cos(2
π
(x+y). Then k=2
π
(1,1) and the right-hand side of (43)is equal to2|k|c0ࣃ5.924. We change thevalue of thediffu- sivityκ
introducedin(15)toκ
=28817 0.05903.Wekeepthetra- ditionalvaluec0= √13.Thenthedissipationcoefficientg(see(39)) isnowgࣃ5.6470.Thenthedampedacousticmodel(38)exhibitsa propagativemodeinthiscase.Theautocorrelationfunctionispre- sentedinFig.12.TheconvergencecurveisdepictedinFig.13.We observe that this convergence is not regular. An extra-fine mesh withdimensions1024×1024hasbeennecessaryinordertocon- firmtheorderofaccuracy.• ComplementaryexperimentsforaninitialGaussian
We havecomparedthe scalarD2Q9lattice Boltzmannscheme with acoustic scaling with numerical solutions of the heat Eq. (12) as presented in Section 3 and with HaWAY simulations of the damped acoustic system (38) in Section 4. We consider again the first geometry studied in this contribution, id est the square =[−1,1]2 withperiodic boundaryconditions.An initial Gaussianprofile(13)isgivenatt=0.Twonumericalexperiments havebeenconsidered:Aquiteviscousonewithimposeddiffusivity
κ
=0.15andanotheronewithκ
= 0.015.Thenumericalparame- tersaredisplayedinTable4.The resultsfor the firsttest case with
κ
=0.15 are presented in Figs. 14–17. In Fig.14, a qualitative view of the numericalre- sultonagivenmeshshowsthatthescalarD2Q9schemeandthe HaWAYschemefordampedacousticareclosertoeachotherthan theyaretothesolutionoftheheatequation.Thethreeprofilesof densityareshowninFig. 15andacomparisonofautocorrelation functionsinFig.16.Evenonarelativelycoarsemesh,theconclu- sion is the same andour new asymptoticanalysis of the acous- tic system(38)is consistent withthe numericalresults. Lastbut notleast,boththeerrorbetweenD2Q9andthermicsononehand, andthatbetweenD2Q9anddampedacoustics ontheotherhand are displayedinFig.17.The errorbetweenthe latticeBoltzmann schemeandthedampedacousticresultstendstozerowhereasthe errorbetweenD2Q9andthethermicmodelremainsstationary.Thesecond numericalexperiment with
κ
=0.015ispresented inFigs.18–20.Attime =2onarelatively coarsemesh,thethree numericalsolutionscan notbe distinguishedasshowninFig.18. Itisalsothecasefortheautocorrelationfunctionaspresentedin Fig. 19. The numerical convergence is delicate for this test case.Duringonedecadeofmeshrefinement,thethreemethodspresent very close results as shownin Fig. 20. Two additional computa- tionson895×895and1791×1791refinedmesheshavebeennec- essarytodemonstratetheconvergenceofthescalarD2Q9scheme towards the damped acoustic system. Observe that the most re- finedmeshcontainsmorethan3millionscells!
6. Conclusion
We have first considered the scalar D2Q9 lattice Boltzmann scheme with diffusive scaling. Our experiments confirm numeri-