• Aucun résultat trouvé

An efficient high-order compact scheme for the unsteady compressible Euler and Navier–Stokes equations

N/A
N/A
Protected

Academic year: 2021

Partager "An efficient high-order compact scheme for the unsteady compressible Euler and Navier–Stokes equations"

Copied!
23
0
0

Texte intégral

(1)

HAL Id: hal-02874872

https://hal.archives-ouvertes.fr/hal-02874872

Submitted on 19 Jun 2020

HAL is a multi-disciplinary open access

archive for the deposit and dissemination of

sci-entific research documents, whether they are

pub-lished or not. The documents may come from

teaching and research institutions in France or

abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est

destinée au dépôt et à la diffusion de documents

scientifiques de niveau recherche, publiés ou non,

émanant des établissements d’enseignement et de

recherche français ou étrangers, des laboratoires

publics ou privés.

An efficient high-order compact scheme for the unsteady

compressible Euler and Navier–Stokes equations

Alain Lerat

To cite this version:

Alain Lerat.

An efficient high-order compact scheme for the unsteady compressible Euler and

Navier–Stokes equations.

Journal of Computational Physics, Elsevier, 2016, 322, pp.365-386.

(2)

An

efficient

high-order

compact

scheme

for

the

unsteady

compressible

Euler

and

Navier–Stokes

equations

A. Lerat

DynFluidLab.,ArtsetMetiersParisTech,151Boulevarddel’Hopital,75013Paris,France

a

b

s

t

r

a

c

t

Keywords:

Compactschemes Highorder

Unsteadycompressibleflows Taylor–GreenVortex

Residual-BasedCompact(RBC)schemesapproximatethe3-DcompressibleEulerequations witha5th- or7th-orderaccuracyona5×5×5-pointstencilandcaptureshockspretty well without correction. For unsteady flows however, they require a costly algebra to extract thetime-derivative occurring at severalplacesinthe scheme.A newhigh-order time formulationhasbeen recently proposed[13] for simplifyingthe RBC schemes and increasingtheirtemporalaccuracy.Thepresentpapergoesmuchfurtherinthisdirection anddeeplyreconsidersthemethod.Anavatarofthe

RBC schemes

ispresentedthatgreatly reducesthecomputingtimeandthememoryrequirementswhilekeepingthesametypeof successfulnumericaldissipation.Twoand three-dimensionallinearstabilityareanalyzed and themethodis extendedtothe 3-Dcompressible Navier–Stokes equations.Thenew compactschemeisvalidatedforseveralunsteadyproblemsintwoandthreedimension.In particular,anaccurateDNSatmoderatecostispresentedfortheevolutionoftheTaylor– GreenVortexatReynolds1600andPrandtl0.71.Theeffectsofthemeshsizeandofthe accuracyorderintheapproximationofEulerandviscoustermsarediscussed.

1. Introduction

Among high-order methods forcomputing compressible flows on structured meshes, compact schemes are attractive becauseoftheirnarrowgrid-stencilthatsignificantlyreducesthetruncationerrorforagivenaccuracy-orderandmakesthe treatmentofdeformingmeshesandboundaryconditionseasier.Compactschemesforcompressibleflowshavebeenmainly developedascenteredapproximationsinspace,notablyintheworksbyLele[1],CockburnandShu[2],Yee[3]andVisbal andGaitonde[4],withnumerical dissipationbased onartificialviscosities, numericalfilters orlimiters. Upwindcompact schemeshavealsobeenproposedbyTolstykh[5]andFuandMa[6].AnotherinterestingoptionistheuseofResidual-Based Compact(RBC)schemes.Inasuchascheme,theconsistentpartandthenumericaldissipationareexpressedonlyinterms ofcompactapproximationsofthecompleteresidual,i.e. thesumofthetermsinthegoverningequationsincludingthetime derivative.ThesecompactapproximationsarededucedfromPadéformulasinwhichthe inverseoperators areeliminated. OnaCartesianmesh,aRBC schemecanapproximateahyperbolicsystemofconservationlawsind-dimensionwitha 5th-or7th-orderaccuracyona 5d-point stencilandcaptureshocksprettywell withoutcorrection.Descriptionandanalysisof

theseschemescan befound in[7–13]. A relatedapproachdevelopedon unstructuredmeshesistheresidual-distribution method of Abgrall, Deconinckand Ricchiuto [14–18] in which the residuals are distributed to the nodes of triangles or tetrahedrons.

(3)

A peculiarity ofthe RBC schemesis themultiple occurrenceof thetime derivative in thescheme (

w

/∂

t occursd

+

1 times in d-dimension). Besides, due to compactness, discrete spatial-operators are applied to each time-derivative. In the first applications ofthe RBC schemesto unsteady problems (see [19,10,11,20]), the time formulation was based on the Gearthree-levelmethodwhichis A-stable. Thisapproachisefficientforcomputingcompressibleflowsinsteadyand slow unsteadyregimes.However, itrequires an iterativemethodtoadvance thesolution(dual-time approachorNewton sub-iterations) and its time accuracy is limited to order 2, which is not sufficient for some unsteady applications. This is the reason whyan explicit time-formulation of highaccuracy has beenproposed in[13].Since a directextraction of thetime-derivativerequiresthesolutionofalargelinearalgebraic-system,an approximatespace-factorizationofthetime operator has been done witha correction to preserve rigorously the high order in space. This leads to the solution of simplelinearsystemsthatareblocktridiagonalforRBC schemesspatially-accurateatorder3or5andblockpentadiagonal for the RBC schemeof order7.Then, an ABM time-integration of orderq (explicitcombination ofAdams–Bashforthand Adams–Moulton methods)is usedto advance the solution.The ABM method offers the advantage ofonly requiringtwo computationspertime-stepofthefluxbalanceandofthetime-derivativeextraction,whatevertheorderq.

Inthe presentpaperwe gomuchfurtherinthequestofsimplifyingthealgorithmandreducing theCPU-timeforthe same accuracylevel.We constructan avatarofthe RBC schemesby dissociating thecomputation ofthe numerical dissi-pation fromthat oftheconsistentpart.Doingso,theresidual-basedfeatureispartiallylost,butthenumericaldissipation of thenew schemediscretizes thesame partial derivative operatorasthe RBC scheme andonlyrequires thesolution of ordinary tridiagonal(uptoorder 5)orpentadiagonallinearsystemswithouttheneedforaspacefactorization.The time-integrationusedforthissimpleschemeistheoptimizedlow-storageRungeKuttamethodR K o6 ofBogeyandBailly[21].

The paperisorganized asfollows.Section2presentsthevariousPadécompactapproximationsusedinthispaperand recalls the construction ofthe RBC schemeand thespecific features ofits numerical dissipationfor solvinga hyperbolic systemofconservationlawsintwo-dimension.Italsorecallsthedirectextractionofthetimederivativewiththe factoriza-tion techniqueandtheRBC

ABM schemes.Section 3describesthenewcompactschemefortheEulerequationsandits combinationwith R K o6 in whichthenumericaldissipationis takeninto accountatthelast stageonly. Thenewformof numerical dissipationiscompared tothat ofnon-compactupwindschemesofhighorder. Amultidimensionalanalysisof linearstabilityoftheoverallalgorithmispresentedforthevariousspatialordersandprovesstabilityforCFL numbersequal toorslightlygreaterthanone.Section4presentstheextensionofthenewcompactschemeoforder5tothecompressible Navier–Stokesequations.Todiscretizethesecondderivativeswithvariablecoefficientsona5-pointstencil,wedonotapply twicethePadéapproximationofafirstderivative,butusespecificPadéformulas.Theschemeheredescribedisspatially ac-curateatorder5fortheEulerfluxesandatorder4forgeneralviscousfluxes(E5V 4 scheme)andonlyrequiresthesolution oftridiagonallinearsystemswithconstantcoefficients.InSection5,thenewschemeisappliedtoseveraltest-problems.We firstconsidertheadvectionofavortexduringalongtimeinthehorizontalordiagonaldirection.These2-DEulerproblems allow avalidationoftheaccuracyorderof E5

R K o6 forseveralCFL numbers andameasureoftheimportantreduction in computingtime over the RBC5

ABM4 scheme. Then the3-D scalar problemofthe diagonaladvection ofa spherical

Gaussian-shape iscomputedbythenewschemeinordertoconfirmtheaccuracyorderin3Daswell asthestability do-main. Successfulcomputationsare done up to theCFL theoretical-limitof 2

/

3. Thenext problemis a directnumerical simulation of the Taylor–GreenVortex atRe

=

1600 and Pr

=

0

.

71.Several accuracy orders are considered forthe Euler andviscoustermsonaseriesofmeshes.EnstrophyevolutionsarecomparedtoareferencesolutionandtheCPU-costsare presented.The E5V 4

R K o6 schemeproducesaveryaccuratesolutionona2563 meshforamoderateCPU-time.Finally, an Eulercomputation ofshock–vortexinteraction allowsa firstassessmentoftheshock-capturingcapabilitiesofthenew compactscheme,stillusedwithoutanykindofcorrection.ConclusionsaredrawnandfutureworkisplannedinSection6. 2. RBCschemes

2.1. Reminderoncompactapproximations

Aderivative

f

/∂

x canbeapproximatedatanyorderonaregular meshxj

=

j

δ

x usingtwo simplediscrete operators, i.e. adifferenceandanaverageoveronemeshinterval:

f

)

j+1

2

=

fj+1

fj

(

μ

f

)

j+12

=

1

2

(

fj+1

+

fj

)

where j isanintegerorhalfinteger.Forinstance:

μ

f

)

j

= (

μ

f

)

j+1 2

− (

μ

f

)

j−12

=

1 2

(

fj+1

fj−1

)

2f

)

j

= (δδ

f

)

j

=

fj+1

2 fj

+

fj−1

4f

)

j

=

fj+2

4 fj+1

+

6 fj

4 fj−1

+

fj−2

(4)

Table 1

Error-coefficientK2poftheapproximationofaderivativeatorder2p=4, 6, 8.

Error coefficient K4 K6 K8 Non-compact −301 1 140 − 1 630 Compact − 1 180 1 2100 − 1 44100



f

x



j

=



I

1 6

δ

2

+

1 30

δ

4

1 140

δ

6



δ

μ

f

δ

x



j

+

O

x8

)

(1)

where I is the identity operator. This writingalso contains the approximationsof lower orders. Indeed,by successively removing theterms

1401

δ

6, 301

δ

4 and

16

δ

2,werespectivelygetthe approximationsoforder6,4and2on a7-,5- and 3-pointstencil.

Toobtain thesameaccuracy orders ona shorterstencil,we canuse aPadé compactapproximation ofthe derivative, brieflycalledaPadéderivative,thatis:



f

x



j

=



f

δ

x



j

+

O

x2p

)

(2) with

∇ =

D−1

(

I

+

a

δ

2

) δ

μ

,

D

=

I

+

b

δ

2

+

c

δ

4

wherea,b andc arescalarparameters.Theaboveformulainvolves5pointsatmost.However,toextractthePadéderivative

f

x,wehavetosolvethelinearalgebraicsystem:



D

f

δ

x



j

=



(

I

+

a

δ

2

)

δ

μ

f

δ

x



j (3)

ThematrixassociatedtotheD-operatorispentadiagonalingeneral.

ThePadéderivativeissecond-orderaccurateatleast.Itbecomesoffourthorderatleastifa

=

b

16.Bychoosing

a

=

0

,

b

=

1

6

,

c

=

0 (4)

bothsidesof(3)havea3-pointstencil.

ThePadéderivativeisaccurateatorder6atleastiftheparameterssatisfy

a

=

1

30

+

6c

,

b

=

1

5

+

6c (5)

Bychoosing c

=

0,theD-operatorhasa3-pointstencil,whichleadstoatridiagonallinearsystem. ThePadéderivativeisaccurateatorder8for

a

=

5 42

,

b

=

2 7

,

c

=

1 70 (6)

The shortgrid-stencilofcompactformulasmakes thetreatment ofdeforming meshesandboundary conditionseasier.In addition,itsignificantlyreducestheerror.Indeed,thetruncationerrorofthePadéderivativeoforder2p

=

4

,

6

,

8 is:



f

δ

x

f

x



j

=

K2p

δ

x2p



2p+1f

x2p+1



j

+

O

x2p+2

)

The truncation error of the non-compact formula (1)has the same form, butthe error-coefficients K2p ofthe compact

formula are much smaller than those ofthe non-compact one, especially as the order increases(see Table 1). Thisis a strongargumentinfavorofcompactness.Notethatasimilarresultholdsforsecondderivatives(see[9]).

Todescribe thenumerical dissipation of theRBC schemes andto approximatethe viscous fluxes, we alsoneed Padé approximationsatmidpointsforaderivativeandanaveraging.Theyaredefinedas:

(5)

where

=

I

+

δ

2

+

δ

4

,

=

I

+

δ

2

+

δ

4

ThemidpointPadéapproximations δx and

μ

aresecond-orderaccurateatleast.Theybecomeoffourthorderatleastif

=

1

24

,

b

μ

aμ

=

1

8 (8)

Theyareaccurateatorder6atleastifinaddition

1 24b δ

cδ

=

3 640

,

1 8b μ

cμ

=

3 128 (9)

ToconstructthenumericaldissipationoftheRBC schemes,midpointapproximationsoforders2,4and6willbesufficient. Besideswewillchoose

=

,i.e. bδ

=

bμ andcδ

=

cμ,toobtainagreatsimplificationinthedissipationterm(vanishing of Dμ andDδ).Thesimplestrelevantsetsofparametersarethenthefollowingones.For2nd-order:

=

=

=

=

=

=

0 (10)

For4th-order,wesatisfy(8)with

=

andforsimplicity

=

=

0,whichgives

=

1

12

,

a

μ

=

0

,

bδ

=

bμ

=

1

8

,

c

δ

=

cμ

=

0 (11)

For6th-order,wesatisfy(8)and(9)with

=

,whichgives

=

11 60

,

a μ

=

1 10

,

b δ

=

bμ

=

9 40

,

c δ

=

cμ

=

3 640 (12)

Now concerningtheapproximationofviscous terms,thechoicecriterion isnolonger

=

butcδ

=

=

0 toproduce tridiagonallinearsystemsonly.For4th-order,wesatisfy(8)with

=

=

0 andforsimplicity

=

=

0,whichgives:

=

=

0

,

=

1

24

,

b

μ

=

1

8

,

c

δ

=

cμ

=

0 (13)

For6th-order,wesatisfy(8)and(9)with

=

=

0,whichgives

=

17 240

,

a μ

=

1 16

,

b δ

=

9 80

,

b μ

=

3 16

,

c δ

=

cμ

=

0 (14)

2.2. RBC spaceapproximationona5

×

5-pointstencil

Considerthehyperbolicsystemofconservationlaws:

w

t

+

f1

x1

+

f2

x2

=

0 (15)

where t is the time, x1 and x2 are the space coordinates, w is the state vector and f1

=

f1

(

w

)

, f2

=

f2

(

w

)

are flux

componentsdependingsmoothlyonw.TheJacobianmatricesofthefluxaredenoted A1

=

d f1

/

dw and A2

=

d f2

/

dw.

System(15)isapproximatedinspaceon auniformCartesianmesh

(

j1

δ

x1

,

j2

δ

x2

)

usingaresidual-based-compact(RBC)

scheme.Suchaschemeisacompactdiscreteformof

r

=

δ

x1 2

∂(

1r

)

x1

+

δ

x2 2

∂(

2r

)

x2 (16)

wherer istheexactresidual:

r

=

w

t

+

f1

x1

+

f2

x2

andtheright-handsideisaresidual-basednumericaldissipationinwhich



1 and



2arematrices,eachdependingonthe

eigensystems oftheJacobian matrices A1 and A2 andon theratio ofthespace steps.Thesematrices aredefinedin the

sequel.Theycontainnotuningparametersorlimiters.

Todescribethespaceapproximationof(16),weintroducethebasicdiscreteoperatorsineachspacedirection:

1v

)

j1+1 2,j2

=

vj1+1,j2

vj1,j2

2v

)

j1,j2+12

=

vj1,j2+1

vj1,j2

(

μ1

v

)

j 1+12,j2

=

1 2

(

vj1+1,j2

+

vj1,j2

)

(

μ2

v

)

j1,j2+12

=

1 2

(

vj1,j2+1

+

vj1,j2

)

(6)

ReplacingtheexactresidualinEq.(16)bydifferentcompactcenteredapproximationsr

˜

j1,j2,

(

r

˜

1

)

j1+12,j2 and

(

r

˜

2

)

j1,j2+12,

weobtaintheresidual-basedcompactscheme:

˜

rj1,j2

=

1

2

1

(

1

˜

r1

)

+ δ

2

(

2r

˜

2

)

]

j1,j2 (17)

Inspiteofappearances,thenumericaldissipationin(17)isnotsimplyoforderonebecause

(

r

˜

1

)

and

(

r

˜

2

)

approximatethe

exactresidualr

=

0.Clearly,ifr

˜

j1,j2 approximatesr atorder2p and

(

˜

r1

)

j1+12,j2,

(

r

˜

2

)

j1,j2+12 approximater atorder2p

2,

thenthescheme(17)isaccurateatorder2p

1 inspaceanddenotedasRBC2p−1.

Wenowrestricttheschemestencilto5

×

5 points.Forconstructingthemainresidualr

˜

j1,j2,wefirstreplacethespace

derivativesintheexactresidualr byPadéapproximations ∇1

δx1 and

∇2

δx2 oforder2p,wheresimilarlyasin2.1:

l

= (

Dl

)

−1

(

I

+

a

δ

l2

) δ

l

μ

l

,

Dl

=

I

+

b

δ

2l

+

c

δ

4

l

,

l

=

1

,

2 (18)

Thenweapplytheoperator D1D2 toallthetermsintheresidualandobtain

˜

rj1,j2

=



D1D2

w

t

+

D2

(

I

+

a

δ

2 1

)

δ

1μ1f1

δ

x1

+

D1

(

I

+

a

δ

22

)

δ

2μ2f2

δ

x2



j1,j2 (19)

For constructing the two residuals in numerical dissipation

(

r

˜

1

)

j1+1

2,j2 and

(

r

˜

2

)

j1,j2+12, we use Padé approximations at

midpoints. Moreprecisely, for

(

r

˜

1

)

j1+12,j2 we first replace

∂f1

∂x1 by thePadé derivative atmidpoints

∇1

δx1 andweapply the

Padéaverageatmidpoints

μ

1 to ∂∂wt and

∂f2 ∂x2,whichgives



μ

1

w

t

+

1f1

δ

x1

+

μ

1

f2

x2

j1+12,j2 with

l

= (

Dlδ

)

−1

(

I

+

δ

l2

) δ

l

,

μ

l

= (

D μ l

)

− 1

(

I

+

aμ

δ

2 l

)

μ

l Dlδ

=

Dlμ

=

I

+

δ

l2

+

δ

l4

,

l

=

1

,

2

where,aμ,bδ andcδ aretheparameters(10)(12)givingtheorder2p

2. Thenwereplace ∂f2

∂x2 byastandardPadéapproximation

(

∇ 2

δx2

)

j1,j2 with

l

= (

Dl

)

−1

(

I

+

a

δ

l2

) δ

l

μ

l

,

Dl

=

I

+

b

δ

l2

+

c

δ

l4

,

l

=

1

,

2 (20)

wherea,b andcaretheparametersgivingtheorder2p

2. Finally,weapplytheoperator

1D2 toallthetermsandobtain

(

r

˜

1

)

j1+1 2,j2

=

D2



(

I

+

δ

21

)

μ1

w

t

+ (

I

+

a δ

δ

2 1

)

δ

1f1

δ

x1



+ (

I

+

δ

12

)(

I

+

a

δ

22

)

δ

2μ2μ1f2

δ

x2

j1+12,j2

Notethatthe operator

l hascompletely disappearedandthus theparameters and play norole.Similarlywe

con-structthesecondresidualindissipation:

(

r

˜

2

)

j 1,j2+12

=

D1



(

I

+

δ

22

)

μ2

w

t

+ (

I

+

a δ

δ

2 2

)

δ

2f2

δ

x2



+ (

I

+

δ

22

)(

I

+

a

δ

21

)

δ

1μ1μ2f1

δ

x1

j1,j2+12

Thus,theresidualsindissipationdependonthefiveparameters,aμ,a,bandc.TheparametersoftheRBC3-dissipation areallzero,thoseoftheRBC5-dissipationsatisfy(11)and(4)andthoseoftheRBC7-dissipationsatisfy(12)and(5).

2.3. RBC dissipation

WenowdiscussthedissipationtermoftheRBC scheme:

˜

dj1,j2

=

1

2

1

(

1

˜

r1

)

+ δ

2

(

2r

˜

2

)

]

j1,j2 (21)

Thematricesinvolvedaredefinedas:

(7)

where,owingtohyperbolicityofSystem(15),TAl isaninvertiblematrixhavingtherighteigenvectorsoftheflux Jacobian-matrix Al (l

=

1

,

2)ascolumnvectorsandDiag

l(i)

)

denotesadiagonalmatrixwithdiagonalentries:

φ

1(i)

=

sgn

(

a(1i)

)

min



1

,

δ

x2

|

a (i) 1

|

δ

x1m

(

A2

)



,

φ

(2i)

=

sgn

(

a(2i)

)

min



1

,

δ

x1

|

a (i) 2

|

δ

x2m

(

A1

)



whereal(i)isan eigenvalueof Al andm

(

Al

)

=

min

i

|

a

(i)

l

|

,l

=

1

,

2.Eigenvaluesandeigenvectorsof A1 at

(

j1

+

1

2

,

j2

)

andof

A2at

(

j1

,

j2

+

12

)

,i.e. oncellfaces,arecomputedusingtheRoeaverage[22].

The matrices



1 and



2 havebeen designedin [23,7] to introduce some kind ofupwinding. Fora one-dimensional

problem,



1 reducestothesignmatrix:

(

1

)

j 1+12,j2

=

[sgn

(

A1

)

]j1+12,j2

=

TA1Diag

(

sgn

(

a (i) 1

)

T− 1 A1

j1+12,j2 (22)

Foratwo-dimensionalscalarequation:

w

t

+

A1

w

x1

+

A2

w

x2

=

0 (23)

with Al

=

Al

(

w

)

=

0,l

=

1

,

2,



1 and



2 reducesto

(

1

)

j1+12,j2

=



sgn

(

A1

)

min

(

1

,

1

θ

)



j1+12,j2

,

(

2

)

j1,j2+12

=

[ sgn

(

A2

)

min

(

1

, θ )

]j1,j2+12 (24)

where

θ

characterizestheadvectiondirectionwithrespecttothemesh:

θ

=

δ

x1

|

A2

|

δ

x2

|

A1

|

For

θ

=

1, the advection is along a mesh diagonal. For

θ <

1, the advection is closer to the x1-direction than to the

x2-direction.For

θ >

1,itistheopposite.

Thedissipation term(21)hasbeenanalyzedin[11] forunsteadyproblems.Forsteps

δ

x1 and

δ

x2 ofthesameorderof

magnitude,say

O(

h

)

,aTaylorexpansionof(21)foraRBC schemeoforder2p

1 gives:

˜

dj1,j2

=(−

1

)

p−1

κ

x1





1

x2p1 −1

2p−1f1

x2p1 −1

+

χ

δ

x1

δ

x 2p−2 2

2p−1f2

x2p2 −1

)



+

x2





2

(

χ

δ

x2p1 −2

δ

x2

2p−1f1

x2p1 −1

+ δ

x 2p−1 2

2p−1f2

x2p2 −1

)



j1,j2

+

O

(

h2p+1

)

(25)

where

κ

>

0 and

χ

aretwoconstantcoefficientsdependingonlyontheparametersindissipation.Notethat(25)doesnot containtimederivativessincetheyhavebeenreplacedbyspacederivativesusingtheexactsystem(15).

FortheRBC3-dissipation:

κ

=

1 24

,

χ

=

12



b

1 6



FortheRBC5-dissipation:

κ

=

1 240

,

χ

=

20



b

1 5



FortheRBC7-dissipation:

κ

=

1 5600

,

χ

=

280 3



c

1 70



Consideringthescalarequation(23),ithasbeenproved[11]thatanecessaryandsufficientconditionforthepartial differ-entialoperatorin(25)bedissipative,foralladvectiondirections

(

A1

,

A2

)

andfor(



1

,



2)oftheform(24),is

χ

=

0,i.e. no

crossedderivativesin(25).Asimilarresulthasalsobeenprovedin3-D[11]. FortheRBC7-dissipation,

χ

=

0 meansc

=

701.Togetherwiththecondition

a

=

1

30

+

6c



,

b

=

1

5

+

6c

(8)

ensuringthe6th-orderaccuracyof ∇2 δx2 in

˜

r1and ∇ 1 δx1 inr

˜

2,weobtain: a

=

5 42

,

b 

=

2 7

,

c 

=

1 70

that is a 8th-order accuracy on these

-Padé derivatives as forthe

-Padé derivatives in the main residual. A similar situationoccurswithRBC5- andRBC3-dissipation.Generallyspeaking,acorrectdissipationrequires



= ∇

,thatis

a

=

a

,

b

=

b

,

c

=

c

Fortunately,thisaccuracyenhancementofsometermsinthedissipationoperatordoesnotextendtheschemestencil. Finally,thedissipativeRBC schemes ona 5

×

5 stencilaredefinedbythemainresidual(19)andtheresidualsin dissi-pation:

(

r

˜

1

)

j1+1 2,j2

=

D2



(

I

+

δ

21

)

μ1

w

t

+ (

I

+

a δ

δ

2 1

)

δ

1f1

δ

x1



+ (

I

+

δ

12

)(

I

+

a

δ

22

)

δ

2μ2μ1f2

δ

x2

j1+12,j2 (26)

(

r

˜

2

)

j1,j2+1 2

=

D1



(

I

+

δ

22

)

μ2

w

t

+ (

I

+

a δ

δ

2 2

)

δ

2f2

δ

x2



+ (

I

+

δ

22

)(

I

+

a

δ

12

)

δ

1μ1μ2f1

δ

x1

j1,j2+12 (27)

withthefollowingparameters:

RBC3

:

a

=

0

,

b

=

1 6

,

c

=

a μ

=

aδ

=

0 (28) RBC5

:

a

=

1 30

,

b

=

1 5

,

c

=

a μ

=

0

,

aδ

=

1 12 (29) RBC7

:

a

=

5 42

,

b

=

2 7

,

c

=

1 70

,

a μ

=

1 10

,

a δ

=

11 60 (30)

In theseschemes, the dissipation error (of order 2p

1) dominates the dispersiveerror (of order 2p), whichis a good featureforrobustness.Spectralpropertiesoftheseschemesaredescribedin[12].

SincetheRBC5 schemeismostly usedinournumericalapplications,wewritedownitsspecificresiduals(omittingthe

subscripts):

˜

r

= (

I

+

1 5

δ

2 1

)(

I

+

1 5

δ

2 2

)

w

t

+ (

I

+

1 5

δ

2 2

)(

I

+

1 30

δ

2 1

)

δ

1

μ

1f1

δ

x1

+ (

I

+

1 5

δ

2 1

)(

I

+

1 30

δ

2 2

)

δ

2

μ

2f2

δ

x2

˜

r1

= (

I

+

1 5

δ

2 2

)



μ1

w

t

+ (

I

+

1 12

δ

2 1

)

δ

1f1

δ

x1



+ (

I

+

1 30

δ

2 2

)

δ

2μ2μ1f2

δ

x2

˜

r2

= (

I

+

1 5

δ

2 1

)



μ2

w

t

+ (

I

+

1 12

δ

2 2

)

δ

2f2

δ

x2



+ (

I

+

1 30

δ

2 1

)

δ

1μ1μ2f1

δ

x1

TheRBC5-dissipationapproximates:

˜

d

=

1 240



δ

x15

x1

(

1

5f1

x15

)

+ δ

x25

x2

(

2

5f2

x25

)



+

O

x17

, δ

x27

)

2.4. Directextractionofthetime-derivative

A difficulty inherent to the RBC schemesis that the time derivative occurs, through linear discrete operators dueto compactness,not onlyinthemainresidualbutalsointheother tworesiduals (orthreeresiduals in3-D)involvedinthe numericaldissipation.Toovercomethisdifficulty,thefollowingapproachhasbeenproposedin[13].First,themainresidual andthedissipationaresplitintoapartcontainingthetimederivativeandapurely-spatialpart(stillomittingthesubscripts

(9)

˜

r0

=

D2

(

I

+

a

δ

12

)

δ

1μ1f1

δ

x1

+

D1

(

I

+

a

δ

22

)

δ

2μ2f2

δ

x2

˜

r01

=

D2

(

I

+

δ

21

)

δ

1f1

δ

x1

+ (

I

+

δ

12

)(

I

+

a

δ

22

)

δ

2μ2μ1f2

δ

x2

˜

r02

=

D1

(

I

+

δ

22

)

δ

2f2

δ

x2

+ (

I

+

δ

22

)(

I

+

a

δ

12

)

δ

1μ1μ2f1

δ

x1

Sincethediscreteoperators

δ

1,

δ

2,

μ

1 and

μ

2commutetwobytwo,wecanalsowrited as

˜

˜

d

= (

M1D2

+

M2D1

)

w

t

+

1 2

1

(

1

˜

r 0 1

)

+ δ

2

(

2r

˜

02

)

]

with M1

=

1 2

δ

1



1

(

I

+

δ

12

)

μ1

.

,

M1

=

1 2

δ

2



2

(

I

+

δ

22

)

μ2

.

Owing to the matrix-coefficients



1 and



2, thematricesassociated to thediscrete operators M1 and M2 have ablock

structure. Theyareblock-tridiagonalif

=

0,i.e. forRBC3 andRBC5,orblock-pentadiagonalforRBC7.Theirdimensionis

thenumberofmeshpointsin x1- orx2-directionwithablock-sizeequaltothenumberofequationsintheexactsystem

(15).Forconsistency,thesameblock-structureisgiventothematricesassociatedto D1 and D2 withdiagonalblocks.The

matricesassociatedtoD1andD2 areblock-tridiagonalifc

=

0,i.e. forRBC3 andRBC5,orblock-pentadiagonalforRBC7.

TheRBC schemer

˜

= ˜

d canthusberewrittenas

w

t

=

H0 (31)

where

denotesthediscreteoperator:

=

D1D2

M1D2

M2D1

andH0 isthespace-fluxcontribution:

H0

= −˜

r0

+

1 2

1

(

1r

˜

0

1

)

+ δ

2

(

2

˜

r02

)

]

(32)

The time derivative can be obtained by solving the linear system (31). However, the matrix associated to

is rather complicated.Tosimplifyit,adimensionalfactorizationof

isattemptedas:

=

1

2

c

where

1

=

D1

M1

,

2

=

D2

M2

,

c

=

M1M2

As D1 and D2,theone-dimensionaloperators

1 and

2 are locally

O(

1

)

,butsince M1

=

O(δ

x1

)

andM2

=

O(δ

x2

)

,the

correctiveoperator

c islocally

O(δ

x1

δ

x2

)

andcannot beneglected.Thereforethelinearsystem(31)isconsideredinthe

form:

1

2

w

t

=

H0

+

c

w

t (33)

andsolvedapproximatelyinafewiterationsas:

1

2

w

t (m+1)

=

H0

+

c

w

t (m)

,

m

=

0

,

1

, ... ,

mf (34) startingfrom

w

t (0)

=

0

.

(35)

(10)

we haveonlyto solvelinearsystemsthatare block-tridiagonalforRBC3 andRBC5 orblock-pentadiagonalforRBC7.It has

beenprovedin[13]thatoneiteration(mf

=

1)forRBC3,twoiterationsforRBC5andthreeiterationsforRBC7 aresufficient

topreservethespaceaccuracyorder.

Thenumericalsolutionisadvancedbysolvingtheordinarydifferentialequation(ODE):

w

t

=

F

(

w

)

(37)

where F

=

−1H

0.Clearly, the maincost per time-step comes fromthecalculation of F asthesolution ofthe problem

(36).SoweareinterestedinODEmethodsrequiringfewevaluationsofF pertime-step.Severalexplicitmethodsoforder4 werecomparedin[13],namelytheAdams–Bashforthmethod A B4,theextrapolatedBackwardDifferentialFormulae B D F4,

theclassicalRunge–Kuttamethod R K4 andan explicitcombinationoftheAdams–BashforthandAdams–Moultonmethods

calledABM4.ThelatterwasfoundtobetheleastconsuminginCPU-time.ThisABM4 methodwrites:

˜

wn+1

=

wn

+

t 12

(

23Fn

16Fn−1

+

5Fn−2

)

wn+1

=

wn

+

24t

(

9F

˜

n+1

+

19Fn

5Fn−1

+

Fn−2

)

(38) wherewn

= (

wn

j1,j2

)

denotesthenumericalsolutionattimelevelt

n

=

n

t,Fn

=

F

(

wn

)

andF

˜

n+1

=

F

(

w

˜

n+1

)

.

Themaximaltime-stepforstabilityofABM4islowerthanthatofR K4 butmuchgreaterthanthoseofe B D F4 andA B4.

Ontheotherhand,ABM4 requirestwoevaluationsof F pertimestep(Fn andF

˜

n+1)while R K 4 requiresfour.Finally,the

shortestcomputing-timefortheRBC schemeswasobtainedwithABM4.

3. Anewcompactformulation

In order to simplify the RBC scheme (17), the idea is to compute the time derivative in dissipation from the non-dissipativescheme

˜

r

=

0

Fromthedefinition(19)ofthemainresidual

˜

r,weobtain

w

ˇ

t

= −

1f1

δ

x1

2f2

δ

x2 (39)

usingthePadécompactderivativesdefinedby(18).Theresidualsindissipation

˜

r1 andr

˜

2become

ˇ

r1

=(

D2

)

−1r

˜

1

= (

I

+

δ

21

)

μ1

(

w

ˇ

t

+

2f2

δ

x2

)

+ (

I

+

δ

21

)

δ

1f1

δ

x1

ˇ

r2

=(

D1

)

−1r

˜

2

= (

I

+

δ

22

)

μ2

(

w

ˇ

t

+

1f1

δ

x1

)

+ (

I

+

δ

22

)

δ

2f2

δ

x2

Since

˜

r1,

˜

r2 approximate 0 at a high order and

(

D2

)

−1

=

I

+

O(δ

x22

)

,

(

D1

)

−1

=

I

+

O(δ

x12

)

, the new residuals r

ˇ

1, r

ˇ

2

approximate0atthesameorderas

˜

r1,r

˜

2.

Using(39)toeliminate

w

ˇ

/∂

t,thenewresidualsindissipationreducetothesimpleexpressions:

ˇ

r1

=(

I

+

δ

12

)

δ

1f1

δ

x1

− (

I

+

δ

12

)

μ1

1f1

δ

x1

ˇ

r2

=(

I

+

δ

22

)

δ

2f2

δ

x2

− (

I

+

δ

22

)

μ2

2f2

δ

x2 (40)

andthenewcompactschemetakestheform:

w

t

+

1f1

δ

x1

+

2f2

δ

x2

=

1 2

1

(

1r

ˇ

1

)

+ δ

2

(

2

ˇ

r2

)

]

(41)

Forthe parametersets (28),(29) and(30),thisscheme remains respectivelyaccurate atorder3,5 and 7,butit ismuch simplertoimplement:

– thescheme (41)only requires thesolution ofalgebraic linearsystems withconstant coefficientsto compute the Padé derivatives

1f1

x1and

2f2

x2andthesesystems aretridiagonalforscheme-order3and5orpentadiagonalfororder 7,

withoutanyblock.

(11)

– Furthermore,thenewschemeisstablewithsimpler



1and



2matricesthanthoserecalledinSection2.3fortheoriginal

RBC scheme.Allthepresentcomputationswiththenewschemehavebeenrunusing:



l

=

sgn

(

Al

)

=

TAlDiag

[

sgn

(

a (i)

l

)

] (

TAl

)

1

,

l

=

1

,

2 (42)

The linearstability analysispresented below forthe new scheme with(42) justifies thissimplification for 2-D and3-D scalarproblems.

Althoughthisisnotrecommendedforanefficientprogramming,wenowdevelopthedissipationresiduals(40)inorder topreciselyidentifythenewformofdissipation.Fromthedefinition(18)of

l:

Dl

μ

l

l fl

δ

xl

=

μ

lDl

l fl

δ

xl

=

μ

2 l

(

I

+

a

δ

l2

)

δ

lfl

δ

xl

,

l

=

1

,

2

Usingtherelation

μ

2

l

=

I

+

1 4

δ

2 l,weobtain: Dl

μ

l

l fl

δ

xl

= (

I

+

1 4

δ

2 l

)(

I

+

a

δ

l2

)

δ

lfl

δ

xl

,

l

=

1

,

2

Theresiduals(40)canthusbewrittenas:

ˇ

rl

= (

Dl

)

−1Nl

δ

lfl

δ

xl

,

l

=

1

,

2 where Nl

=

Dl

(

I

+

δ

l2

)

− (

I

+

δ

2l

)(

I

+

1 4

δ

2 l

)(

I

+

a

δ

l2

),

Dl

=

I

+

b

δ

l2

+

c

δ

l4

AsimplecalculationreducesNl to

Nl

=

n2

δ

l2

+

n4

δ

4l

+

n6

δ

l6 with n2

=

b

a

+

1 4

,

n4

=

b a δ

a aμ

+

c

a 4

4

,

n6

=

c a δ

a aμ 4

Fortheparameters(28)ofRBC3: n2

= −

121

,

n4

=

0

,

n6

=

0

Fortheparameters(29)ofRBC5: n2

=

0

,

n4

=

1201

,

n6

=

0

Fortheparameters(30)ofRBC7: n2

=

0

,

n4

=

0

,

n6

= −

28001

Inserting

ˇ

rl

= (

I

+

b

δ

l2

+

c

δ

4 l

)

− 1

(

n 2

δ

l2

+

n4

δ

l4

+

n6

δ

6l

)

δ

lfl

δ

xl

,

l

=

1

,

2

intheright-handsideof(41)andreplacing

l by(18)intheleft-handside,thenewcompactschemetakestheform:

at order 3:

w

t

+

2



l=1

(

I

+

1 6

δ

2 l

)

−1

δ

l

μ

lfl

δ

xl

= −

1 24 2



l=1

δ

l

δ

xl



sgn

(

Al

)(

I

+

1 6

δ

2 l

)

−1

δ

l3fl



(43) at order 5:

w

t

+

2



l=1

(

I

+

1 5

δ

2 l

)

−1

(

I

+

1 30

δ

2 l

)

δ

l

μ

lfl

δ

xl

=

1 240 2



l=1

δ

l

δ

xl



sgn

(

Al

)(

I

+

1 5

δ

2 l

)

−1

δ

l5fl



(44) at order 7:

w

t

+

2



l=1

(

I

+

2 7

δ

2 l

+

1 70

δ

4 l

)

−1

(

I

+

5 42

δ

2 l

)

δ

l

μ

lfl

δ

xl

= −

1 5600 2



l=1

δ

l

δ

xl



sgn

(

Al

)(

I

+

2 7

δ

2 l

+

1 70

δ

4 l

)

−1

δ

l7fl



(45)

Clearly, thedissipationtermofeachabove schemeexpandsasin(25)with

χ

=

0 andthesamecoefficient

κ

.Ittherefore behaves infirstapproximationasthedissipationofthebasicRBC scheme.Notethatthenewschemeremainscompactin itsmainpart(left-handside),butnotinitsdissipativepart.Inpractice,thiscompacitylossisnottooserious,becauseitis easiertolocallymodifythedissipationthantheconsistentpart(forinstanceneara boundary)andalso,aswe seebelow, thisdissipationisonlyusedoncepertimestep,atthelaststageofaRunge–Kuttamethod.

(12)

at order 3:

w

t

+

2



l=1

(

I

1 6

δ

2 l

)

δ

l

μ

lfl

δ

xl

= −

1 12 2



l=1

δ

l

δ

xl



|

Al

l3w



(46) at order 5:

w

t

+

2



l=1

(

I

1 6

δ

2 l

+

1 30

δ

4 l

)

δ

l

μ

lfl

δ

xl

=

1 60 2



l=1

δ

l

δ

xl



|

Al

l5w



(47) at order 7:

w

t

+

2



l=1

(

I

1 6

δ

2 l

+

1 30

δ

4 l

1 140

δ

6 l

)

δ

l

μ

lfl

δ

xl

= −

1 280 2



l=1

δ

l

δ

xl



|

Al

l7w



(48) where

|

Al

| =

TAlDiag

[|(

a (i) l

|] (

TAl

)

− 1

,

l

=

1

,

2

Notethatthedissipationcoefficientoftheupwindschemeisalwayshalfofthelastcoefficientintheleft-handside(see[9]

foraproof).

Comparingcentered approximationsofafirst derivative inSection 2.1,we havementionedthat a compactformulais moreaccuratethananon-compactone ofthesameaccuracyorder.Thisyieldsthattheconsistentpart(left-hand-side)of thenewschemes(43)–(45)produceslessdispersive-errorthantheconsistentpartoftheupwindschemes(46)–(48).Now concerning the dissipation, there is some analogy betweenthe right-hand side of the two types ofhigh-order schemes. For a linear problem(constant matrices A1 and A2), these two scheme-types have the same form of first

differential-approximation,buttheschemes(43)–(45)arelessdissipativebecausetheirdissipationcoefficientsaresmaller,namely2,4 and20timessmalleratorder3,5and7,respectively.Sowe canconcludethatboth dispersiveanddissipativeerrorsare smallerfortheschemes(43)–(45).

Wereturntotheusual form(40)–(41)ofthenewschemesandpresenttheir timeintegration.Duetothesimplicityof theseschemes,themostefficienttime-integrationisnolongerABM4 asinSection 2.4forthebasicRBC-schemes,buthas

beenfound intheclassofRunge–Kuttamethods.Specifically, weusethe low-storageexplicitRunge–Kuttamethod R K o6

proposed by Bogey and Bailly [21]. This six-stage method was constructed by optimizing its dispersion and dissipation properties.Ithasbeenshown[21]to bemorestableandmore accuratethanthestandard low-storagemethodwithfour stages(R K s4).

Foradifferentialequationoftheform(37),the R K o6 methodreads:

w(0)

=

wn

w(k)

=

wn

+

α

k

t F

(

w(k−1)

),

k

=

1

,

2

, ... ,

6 (49)

wn+1

=

w(6)

withthefollowingcoefficients:

α1

=

0

.

117979901657

,

α2

=

0

.

184646966491

,

α3

=

0

.

246623604310

,

α4

=

0

.

331839542736

,

α5

=

1

2

,

α6

=

1

.

Moreprecisely, thetime integrationofthe newschemewillbe doneby takingintoaccount thenumericaldissipation at thelaststageonly.Thisreducesthecomputingtimebutnottheaccuracyorderbecausethenumericaldissipationisofthe orderoftheschemespatial-accuracy(3,5or7).Sowereplace(49)by

w(k)

=

wn

+

α

k

t Fk

(

w(k−1)

),

k

=

1

,

2

, ... ,

6 (50)

(13)

Fk

= −

1 f1

δ

x1

2f2

δ

x2

+

χ

k 2

1

(

1

ˇ

r1

)

+ δ

2

(

2r

ˇ

2

)

]

(51)

where

χ

k

=

0 fork

=

1

,

2

,

...,

5.Thecoefficient

χ

6isnormallyequalto1,butitmaybereducedinsmoothflows.

WenowpresentaL2-stabilityanalysisofthemethod(50)–(51)fora2-Dadvectionproblem( f1

=

A1w and f2

=

A2w

where A1 and A2 are scalar constants). The Fouriersymbol (in space) of the Padé derivative

t Al

l

xl is

ı ˙

AlPl, with

˙

Al

=

t Al

xl and Pl

=

[

1

2a

(

1

cos

ξ

l

)

]

sin

ξ

l 1

2b

(

1

cos

ξ

l

)

+

4c

(

1

cos

ξ

l

)

2

,

l

=

1

,

2

where

ξ

l isthereducedwave-numberinthexl-direction.

Concerningthedissipation,theFouriersymbolof 12

t

δ

l

(

lr

ˇ

l

)

is

−| ˙

Al

|

Ql,where

Ql

= [

1

2aδ

(

1

cos

ξ

l

)

](

1

cos

ξ

l

)

1 2

[

1

2a

μ

(

1

cos

ξ

l

)

]

sin

ξ

lPl

,

l

=

1

,

2

TheFouriersymbolof

t Fk istherefore

t



Fk

= −(

Q(k)

+ ı

P

)

where P

= ˙

A1P1

+ ˙

A2P2 andQ(k)

=

χ

k

(

| ˙

A1

|

Q1

+ | ˙

A2

|

Q2

)

.

Theamplificationfactor g

=

g

1

,

ξ

2

)

governingtheFourier-transformevolutionofwn (



wn+1

=

g



wn)satisfies

g(0)

=

1

g(k)

=

1

α

k

(

Q(k)

+ ı

P

)

g(k−1)

,

k

=

1

,

2

, ...,

6

g

=

g(6)

Denoting g(rk)

= (

g(k)

)

andg(ık)

= (

g(k)

)

,weobtain:

gr(0)

=

1

,

g(ı0)

=

0

gr(k)

=

1

α

k

(

Q(k)g(rk−1)

P gı(k−1)

),

gı(k)

= −

α

k

(

P g(rk−1)

+

Q(k)gı(k−1)

),

k

=

1

,

2

, ...,

6

|

g

|

2

= (

g(r6)

)

2

+ (

g(ı6)

)

2

Thestability domainofthescheme(50)–(51),thatisthepartofthe

( ˙

A1

,

A

˙

2

)

-planeinwhich

|

g

1

,

ξ

2

)

|

1 forall

1

,

ξ

2

)

,

is plotted on Fig. 1 forvarious space-orders corresponding to scheme coefficients(28)–(30)and forvarious dissipation-coefficients

χ

6 inthelast R K o6-stage.Thestabilityconditionisoftheform:

t

(

|

A1

|

δ

x1

+

|

A2

|

δ

x2

)

η

(52)

The stability domain increases as the space accuracy increases or asthe dissipation coefficient decreases. With

χ

6

=

1,

the

η

-boundis equalto 1

,

1

.

30 and 1

.

80 for thespace-orders3,5and7,respectively. Fora5th-order accuracyinspace,

η

=

1

.

30

,

1

.

82 and 1

.

98 for

χ

6

=

1

,

0

.

5 and 0

.

2, respectively. Note that in a square mesh (

δ

x1

= δ

x2

= δ

x), a sufficient

conditionforstabilityis

t

δ

x



A2 1

+

A22

η

2

Considernowthethree-dimensionalhyperbolicsystem

w

t

+

f1

x1

+

f2

x2

+

f3

x3

=

0 (53) with f3

=

f3

(

w

)

and A3

=

d f3

/

dw.

The extensionofthenewcompactschemeto(53)isreallystraightforward.Thetimeintegration(50)hassimplytobe appliedto Fk

= −

1 f1

δ

x1

2f2

δ

x2

3f3

δ

x3

+

χ

k 2

1

(

1

ˇ

r1

)

+ δ

2

(

2r

ˇ

2

)

+ δ

3

(

3r

ˇ

3

)

]

(54)

where

χ

k

=

0 fork

=

1

,

2

,

...

,

5.ThethirdPadéderivative

3f3

x3 isdefinedsimilarlyasin(18)andthethirdresidualin

dissipationr

ˇ

3 asin(40)withamatrix



3 givenby(42)forl

=

3.

The3-Dlinearstabilityanalysisisquitesimilartothe2-Doneandproducesastabilitydomainwithinaregular octahe-dron inthe

( ˙

A1

,

A

˙

2

,

A

˙

3

)

-space.The boundarypointsofthisoctahedrononthethreeaxisare xl

= ±

η

,l

=

1

,

2

,

3.The3-D

Références

Documents relatifs

Adapting at the discrete level the arguments of the recent theory of compressible Navier-Stokes equations [17, 6, 20], we prove the existence of a solution to the discrete

Abstract We propose a sixth-order staggered finite volume scheme based on polynomial reconstructions to achieve high accurate numerical solutions for the incompressible Navier-

is a discrete counterpart of a diffusion term for the density, provides an additional (mesh dependent) estimate on the discrete gradient of the density. As we will explain in details

The convection term in the momentum balance equation is discretized by a finite volume technique, in such a way that a solution obeys a discrete kinetic energy balance, and the

Des bactéries isolées à partir des nodules de la légumineuse du genre Phaseolus avec deux variétés ; P.vulgariset P.coccineus sont caractérisés par une étude

The main idea is to find a correct definition of the Stokes operator in a suitable Hilbert space of divergence-free vectors and apply the Fujita-Kato method, a fixed point procedure,

An essential feature of the studied scheme is that the (discrete) kinetic energy remains controlled. We show the compactness of approximate sequences of solutions thanks to a

Theorem 4.1 (Unconditional stability of the scheme) Let us suppose that the discrete heat diffusion operator satisfies the monotonicity property (3.20), that the equation of