Post on 17-Apr-2022
High-ordernumericalmethodsforLargeEddySimulation
INRIASeminarPalaiseau,France,November24,2016
J.Vanharen
JointworkwithG.Puigt,M.Montagnac,X.Vasseur,J.-F.BoussugeandP.Sagaut
High-ordernumericalmethodsforLargeEddySimulationINRIASeminarPalaiseau,France,November24,2016
J.VanharenJointworkwithG.Puigt,M.Montagnac,X.Vasseur,J.-F.BoussugeandP.Sagaut.
INTRODUCTION
2
High-ordernumericalmethodsforLargeEddySimulation
Introduction
3
Folie 8 > AIAA/CEAS-2008-AFN > Dobrzynski
Major Sources of Airframe Noise
◆ Trackvorticesoverlongdistance◆ Turbulence◆ Generationandpropagationofsoundwaves◆ Complexgeometries
Creation
Propagation
Interaction
DoorruptureseemstobecausedbytheNLGwake/MLGinterac:on.
High-ordernumericalmethodsforLargeEddySimulation
Industrialneeds
4
1)Gotowardscomplexgeometries, includingmoreandmoretechnologyeffect.
2)Captureunsteadyflowphysics.
3)Needforimprovementinhighperformancecomputingcapability.
High-ordernumericalmethodsforLargeEddySimulation
Standardmeshforindustrialproblems
5
◆ Structuredapproach(S):
▪ Standardtechniquetoalignmeshlinesandflowphysics,especiallyintheboundarylayertocaptureliftanddragcoefficients.
▪ Initialtechniqueintroducedinindustry.
▪ Directdataaddressing.▪ Efficientalgorithms(multigrid,linear
solvers…)usingadirectionalapproachfollowingmeshlines(i,j,k).
▪ Timeconsumingmeshgenerationprocess:
• Fromacoupleofhourstomanyweeks.• Dedicatedtoolsforspecific
configurations.
High-ordernumericalmethodsforLargeEddySimulation
Advancedmeshforindustrialproblems
6
◆ Chimera/Oversetgrid:▪ Assemblemeshesgeneratedseparately.▪ Quickermeshgeneration:independentmeshes.▪ AppliedtoexternalaerodynamicswithRANS/URANS.
▪ Dataexchangebetweengridsbyinterpolation:notconservative.▪ Inaccurateforinternalaerodynamics:massflowrateloss.
High-ordernumericalmethodsforLargeEddySimulation
Advancedmeshforindustrialproblems
7
◆ Nonconforminggridinterface(NGI)betweenstructuredblocks:▪ Meshesgeneratedseparatelyshareageometricinterface.▪ Geometricinterfacediscretizationsdiffer.▪ Conservativeapproachforplaneinterface.
▪ Usedformanyapplicationswithexternalandinternalflows.
High-ordernumericalmethodsforLargeEddySimulation
Advancedmeshforindustrialproblems
8
◆ Unstructuredapproach(U):▪ Quickgenerationofthemeshonregularcomputers,
evenforacomplexgeometry.▪ Less control on mesh quality and mesh can be inaccurate in some
specificpartsofthedomain.▪ IndirectdataaddressinginducesanincreaseofCPUcost.
◆ Hybridapproach(H):▪ Blend(S)and(U)blocksinasinglemesh.▪ Blend(S)and(U)algorithms
inasinglesolver.
▪ Takebenefitofanyapproachinits bestdomainofaccuracy.
▪ IntroduceNGIat(S)/(U)interface. Onera
High-ordernumericalmethodsforLargeEddySimulation
Overview
Introduction
I)Nonconforminggridinterfaceforunsteadyflows
II)LimitsinaFiniteVolumeFormalism
III)Alternative:theSpectralDifferenceMethod
Conclusion&perpectives
9
High-ordernumericalmethodsforLargeEddySimulationINRIASeminarPalaiseau,France,November24,2016
J.VanharenJointworkwithG.Puigt,M.Montagnac,X.Vasseur,J.-F.BoussugeandP.Sagaut.
I)NONCONFORMINGGRIDINTERFACEFORUNSTEADYFLOWS
10
High-ordernumericalmethodsforLargeEddySimulation
Spectralanalysis
11
◆ Mathematicalframework:cell-centeredfinitevolumeapproximation
◆ Modelproblem:linearadvectionequationonaregularstructuredgridwithfluxdensity
◆ In1D:
d(fV)dt
+ (cSf)i+1/2 � (cSf)i�1/2 = 0
d
dtfi + c ·
fi+1/2 � fi�1/2
�x= 0
High-ordernumericalmethodsforLargeEddySimulation 12
◆ In,needtodefineaninterfacevalue.Introduceacenteredsecond-orderapproximation:
◆ Andfinally:
◆ Thesameapproximationofthedivergencetermisconsideredinthefollowing.
d
dtfi + c ·
fi+1/2 � fi�1/2
�x= 0
fi+1/2 � fi�1/2
�x' fi+1 � fi�1
2�x
@f
@x
����i
' fi+1 � fi�1
2�x= 0
Spectralanalysis
High-ordernumericalmethodsforLargeEddySimulation 13
◆ Lastpoint:performaVonNeumann(orFourier)localanalysis.▪ Assumethatwithissolutionof
thelinearadvectionequation.▪ Injectthesolutionintheadvectionequationandconsider
auniformgrid:thestandardequivalentwavenumberisrecovered:
j2 = �1f = exp(jkx)
f 0i ' exp[jk(i+ 1)�x]� exp[jk(i� 1)�x]
2�x
' jk exp(jk · i�x) exp(jk�x)�exp(�jk�x)2jk�x
' jkfisin(k�x)
k�x
kmk
=sin(k�x)
k�x
Spectralanalysis
km =f 0ij fi
High-ordernumericalmethodsforLargeEddySimulation 14
Spectralanalysisforsecond-ordercenteredscheme.Thisschemeisnotdissipacve(centered)butdispersive.
Theoretical
SpectralanalysisR
e(k
mD
x )
Im(k
mD
x )
High-ordernumericalmethodsforLargeEddySimulation
Spectralanalysis
15
High-ordernumericalmethodsforLargeEddySimulation 16
◆ Ourgoal:▪ Performasimilaranalysisforan
unsteadyflowinthepresenceofaNGI.
▪ UnderstandtheunsteadybehaviouroftheNGI.
▪ Correctspuriouseffects.
▪ ThisanalysiswaspublishedintheJournalofComputationalPhysics[1].
Spectralanalysis
[1] J. Vanharen, G. Puigt,M.Montagnac. Theoreccal and numerical analysis of nonconforming grid interface forunsteadyflows.JournalofComputaconalPhysics,285:111–132,2015.
High-ordernumericalmethodsforLargeEddySimulation 17
Toyproblem
◆ In2D,linearadvectionequationisconsideredasin1D,withasecond-ordercentered(structured)approximation.
◆ Meshcomposedofrectangularelementsparametrisedby,andonsides.
◆ meansstandardjoininterface.
◆ Keypoint:computingtherightcontributionforthefluxincell.
�zh
h = 0, �zR = �zL
(m,n)
h = 0, �zR = �zL
High-ordernumericalmethodsforLargeEddySimulation
Toyproblem
18
◆ Toensureconservation,thefluxonisdecomposedintotwocontributions:onand.with
◆ Asaconsequence,theprincipleis:▪ Todefineintersectionfacets.▪ Tobuildtheinterfacefluxonfacets
usinganaverageofRandLcontributions.
[AB][AM ] [MB]
fAM =fm,n + fm0,n0+1
2, fMB =
fm,n + fm0,n0+2
2
fm+1/2,n =AM
AB· fAM +
MB
AB· fMB
High-ordernumericalmethodsforLargeEddySimulation
2Dspectralanalysis
Butthenonconforminggirdinterfacerequiresatwo-dimensionalstudy.Otherquanccesshouldbeploiedtoachievespectralanalysis.
characterisesdissipaRon(if<1)oramplificaRon(if>1).
characterisesdispersion(=0foranon-dispersivescheme).
tocomplytheNyquist-Shannonsamplingtheorem.
19
A (�x)
A(�x) = exp [=m [kxm + kym ]�x]
�(�x) =kx + ky �<e [kxm + kym ]
⇡�x
� (�x)
k�x 2 [0,⇡]
High-ordernumericalmethodsforLargeEddySimulation
2Dspectralanalysis
20
◆ OnaNGI,theapproachisvalidonLandRblocks.
◆ Breakinguptheanalytictime-integrationintotheleftandrightparts sequentially to show the combined effects of a wavetravellingtheNGI,theoveralleffectoftheNGIisintroduced:
◆ Amplitudeandphasedependonlocalmetrics.
f(t, x, z) = AL exp⇥j⇡�L
⇤| {z }
Left contribution
· AR exp⇥j⇡�R
⇤| {z }
Right contribution
· exp [j (kxx+ kzz � !t)
High-ordernumericalmethodsforLargeEddySimulation
Coarseningalongthez-axis
21
Left Block200 x 400
Right Block200 x 103
High-ordernumericalmethodsforLargeEddySimulation 22
0.0 0.5 1.0 1.5 2.0 2.5 3.0k · �x [�]
0.0
0.2
0.4
0.6
0.8
1.0
�L,R
( �x)
[�]
u = 1 Left Block
u = 1 Right Block
u = 2 Left Block
u = 2 Right Block
u = 4 Left Block
u = 4 Right Block
0.0 0.5 1.0 1.5 2.0 2.5 3.0k · �x [�]
0.6
0.8
1.0
1.2
1.4
1.6
1.8
2.0
AL,R
( �x)
[�]
Coarsening along the z-axis. Propagation non-orthogonal to the interface.
Coarseningalongthez-axis
High-ordernumericalmethodsforLargeEddySimulation
NGI-COVO
23
0.0 0.5 1.0 1.5 2.0x [m]
0.0
0.5
1.0
1.5
2.0
z[m
]
Initial V ortex Convected V ortex
Boundary conditions
Subsonic inlet
Subsonic outlet
Periodic
Join
-1.400
-1.145
-0.891
-0.636
-0.382
-0.127
0.127
0.382
0.636
0.891
1.145
1.400
u�
U0
[m/s
]
COnvection of a compressible and isentropic
VOrtex (CO-VO): exact solution of Euler equations.
An analytical solution is available.
Error analysis.
⇠ = kfanalytical � fcomputedk1
⌘ =kfanalytical � fcomputedk1
kfanalyticalk1
High-ordernumericalmethodsforLargeEddySimulation
Spuriousreflection
24
Aspectratiointhex-&z-direction:32
Thisspuriousreflectioncanbe correctedbyametric-dependentinterpolationforRiemannsolver
High-ordernumericalmethodsforLargeEddySimulation
Spuriousreflection
25
High-ordernumericalmethodsforLargeEddySimulationINRIASeminarPalaiseau,France,November24,2016
J.VanharenJointworkwithG.Puigt,M.Montagnac,X.Vasseur,J.-F.BoussugeandP.Sagaut.
II)LIMITINAFINITEVOLUMEFORMALISM
26
High-ordernumericalmethodsforLargeEddySimulation
elsAHworkflow
27
Hybridmesh
Pre-
elsAH
Post-processing
Hybridsplitter
1. Hybridmeshgeneration.▪ Notooltomanagehybridgrids.▪ Basealgebra:A+B=C,eitherbygeometric
reconstructionorbyfamilies.2. Hybridsplitterforparallelcomputation.
▪ Couplingofunstructuredandstructuredsplitter3. Pre-processing4. elsaH&co-processing.
▪ Co-processingismandatoryforLEScomputations.5. Post-processing.
▪ Antareslibrary.
High-ordernumericalmethodsforLargeEddySimulation
elsAHworkflow
28
Unstructured
Structured
Hybridnonconforminggridinterface
High-ordernumericalmethodsforLargeEddySimulation
Necessityofhigh-ordermethods
Convecconofanisentropicvortexinaboxwithperiodicboundarycondicons.Highorderversusloworderapproach,same#DOF.
Forthesameaccuracy,mul:plythe#DOFby64fora2ndorderscheme.
29
High-ordernumericalmethodsforLargeEddySimulation
Extensiontohigh-ordermethods
Forstructuredblocks,thehigh-orderschemeisasixth-ordercompactFiniteVolumescheme.
30
High-ordernumericalmethodsforLargeEddySimulation
Forunstructuredblocks,thesituationismuchmorecomplicated.
31
Extensiontohigh-ordermethods
ENO/WENO,k-exact&Least-Squares:definitionofalocalpolynomialrepresentationofquantitiestocomputethefluxesainterface.
DifficulttoimplementinelsAforgeneralunstructuredgrids.StenciltoolargeforHPC.
5thorder:104cells
6thorder:167cells
High-ordernumericalmethodsforLargeEddySimulation
Forunstructuredblocks,thesituationismuchmorecomplicated.
32
Extensiontohigh-ordermethods
INRIAapproachintroducedbyDervieuxetal.foracell-vertexcode.Schemesfromthesecond-ordertothesixth-order(forCartesiangrid).Reverttoasecond-orderontetrahedrabutwithlimiteddissipation&dispersion.
A PhD at Cerfacs investigated these schemes to adapt them in a cell-centered FiniteVolumeformalismbutwithoutsuccess.
Theaccuracyisdrivenbythegradientsonthecompactstencil.Asecond-ordergradientneedsathird-orderextrapolationofthesolution.Athird-orderofthesolutionneedsextendedstencilonunstructuredgrid.
AndthisleadstoHPC-inefficientschemes.[3]P.Cayot,G.Puigt,J.Vanharen,J.-F.Boussuge,P.Sagaut.Towardssecond-orderfinite-volumecell-centereddiffusionschemeonhybridunstructuredmeshes.Inpreparacon,AIAAJournal.
High-ordernumericalmethodsforLargeEddySimulationINRIASeminarPalaiseau,France,November24,2016
J.VanharenJointworkwithG.Puigt,M.Montagnac,X.Vasseur,J.-F.BoussugeandP.Sagaut.
III)ALTERNATIVE:THESPECTRALDIFFERENCEMETHOD
33
High-ordernumericalmethodsforLargeEddySimulation 34
◆ TheSpectralDifferenceMethodconsistsinprojectingequationsonthebasisofLagrange polynomials.
◆ Thesolutioniscontinuousinside acellbutcanbediscontinuous atcellinterfaces.
◆ Letusdefinep+1solutionpoints andp+2fluxpoints.
SpacetimediscretisationforSpectralDifferenceMethod
�x
High-ordernumericalmethodsforLargeEddySimulation 35
TheSpectralDifferenceMethoddiscretisestheadvectionequationwiththreesteps:
1)Extrapolationfromthesolutionpointstothefluxpoints(matrixE),
2)Fluxcomputationinfluxpoints(matrixF),
SpacetimediscretisationforSpectralDifferenceMethod
High-ordernumericalmethodsforLargeEddySimulation 36
3)Fluxpolynomialderivationtoobtainthefluxderivativeatsolutionpoints(matrixD),
Combiningallthesematrices,oneobtains:
whereisthevectorofsolutionspoints.
SpacetimediscretisationforSpectralDifferenceMethod
@Ui
@t= DFE · Ui
Ui
High-ordernumericalmethodsforLargeEddySimulation 37
Thegeneratingpatternforaone-dimensionalproblemisgivenbyonecell.Tointroduceanormalmode,onehastomultiplytheDFEmatrixbytheLmatrix:toobtain:
whereM=DFEL.Thissystemoflinearequationsissolvedwiththelow-storagesecond-ordersix-stageRunge-KuttaschemeofBogeyandBailly[5]tosolvetheadvectionequation.
SpacetimediscretisationforSpectralDifferenceMethod
[5]C.BogeyandC.Bailly.Afamilyoflowdispersiveandlowdissipacveexplicitschemesforflowandnoisecomputacons,J.Comput.Phys.194(1)(2004)194–214.
High-ordernumericalmethodsforLargeEddySimulation 38
◆ Thisequationcanbeseenasageometricprogression.
◆ Consideringtheeigenvalueproblem:
◆ Onefindstheeigenvalues:
◆ Onefindstheassociatedeigenvectorsspanning:
TheMatrixPowerMethod
Figure 6: Illustration of the spectral analysis. The mesh is composed of four cells. Each cell is delimited by the dotted lines.The black crosses + represent the solution point for each cell. The symbols • represent the initial solution at each solutionpoint and � the computed solution at each solution point.
4.5. Spectral analysis: the Matrix Power Method215
Using properties of geometric progessions and (34), one obtains:
Uni = Gn · U0
i . (38)
Carrying out an unsteady computation with the Spectral Di↵erence Method simply consists in computingthe matrix G to the nth power to obtain the solution at the nth iteration from the initial solution. Itis somewhat reminiscent of the Matrix Power Method. For estimating and computing eigenvalues, thisMethod generally gives the eigenvalue with the greatest absolute value of a given matrix. It consists incomputing the nth power of this matrix. The proof of this theorem can be found in [33]. The followingproof is considerably inspired from the previous reference but it will allow us to establish a general resultfor the spectral analysis of the Spectral Di↵erence Method. Note that this method could be applied apriori for other high-order methods which consider several degrees of freedom inside each cell. Considerthe eigenvalue problem G · Ui = �Ui, Ui 6= 0, G 2 Mp+1 (C), Ui 2 Cp+1, � 2 C, which we assume hasa complete normalized eigenvector space (v1,v2, ...,vp+1) spanning Cp+1, with corresponding eigenvalues
(�1,�2, ...,�p+1) where |�1| < |�2| < ... < |�p+1|. We assume that U0i is a given vector, for which
U0i =
p+1X
j=1
↵(0)j · vj , (39)
where ↵(0)n 6= 0. Let
8n 2 {1, 2, ...} , Uni = G · Un�1
i .
To analyze the convergence of the sequence Uni , one obtains
Uni =
p+1X
j=1
↵(0)j Gn · vj .
15
Un+1i = G · Un
i
Figure 6: Illustration of the spectral analysis. The mesh is composed of four cells. Each cell is delimited by the dotted lines.The black crosses + represent the solution point for each cell. The symbols • represent the initial solution at each solutionpoint and � the computed solution at each solution point.
4.5. Spectral analysis: the Matrix Power Method215
Using properties of geometric progessions and (34), one obtains:
Uni = Gn · U0
i . (38)
Carrying out an unsteady computation with the Spectral Di↵erence Method simply consists in computingthe matrix G to the nth power to obtain the solution at the nth iteration from the initial solution. Itis somewhat reminiscent of the Matrix Power Method. For estimating and computing eigenvalues, thisMethod generally gives the eigenvalue with the greatest absolute value of a given matrix. It consists incomputing the nth power of this matrix. The proof of this theorem can be found in [33]. The followingproof is considerably inspired from the previous reference but it will allow us to establish a general resultfor the spectral analysis of the Spectral Di↵erence Method. Note that this method could be applied apriori for other high-order methods which consider several degrees of freedom inside each cell. Considerthe eigenvalue problem G · Ui = �Ui, Ui 6= 0, G 2 Mp+1 (C), Ui 2 Cp+1, � 2 C, which we assume hasa complete normalized eigenvector space (v1,v2, ...,vp+1) spanning Cp+1, with corresponding eigenvalues
(�1,�2, ...,�p+1) where |�1| < |�2| < ... < |�p+1|. We assume that U0i is a given vector, for which
U0i =
p+1X
j=1
↵(0)j · vj , (39)
where ↵(0)n 6= 0. Let
8n 2 {1, 2, ...} , Uni = G · Un�1
i .
To analyze the convergence of the sequence Uni , one obtains
Uni =
p+1X
j=1
↵(0)j Gn · vj .
15
Figure 6: Illustration of the spectral analysis. The mesh is composed of four cells. Each cell is delimited by the dotted lines.The black crosses + represent the solution point for each cell. The symbols • represent the initial solution at each solutionpoint and � the computed solution at each solution point.
4.5. Spectral analysis: the Matrix Power Method215
Using properties of geometric progessions and (34), one obtains:
Uni = Gn · U0
i . (38)
Carrying out an unsteady computation with the Spectral Di↵erence Method simply consists in computingthe matrix G to the nth power to obtain the solution at the nth iteration from the initial solution. Itis somewhat reminiscent of the Matrix Power Method. For estimating and computing eigenvalues, thisMethod generally gives the eigenvalue with the greatest absolute value of a given matrix. It consists incomputing the nth power of this matrix. The proof of this theorem can be found in [33]. The followingproof is considerably inspired from the previous reference but it will allow us to establish a general resultfor the spectral analysis of the Spectral Di↵erence Method. Note that this method could be applied apriori for other high-order methods which consider several degrees of freedom inside each cell. Considerthe eigenvalue problem G · Ui = �Ui, Ui 6= 0, G 2 Mp+1 (C), Ui 2 Cp+1, � 2 C, which we assume hasa complete normalized eigenvector space (v1,v2, ...,vp+1) spanning Cp+1, with corresponding eigenvalues
(�1,�2, ...,�p+1) where |�1| < |�2| < ... < |�p+1|. We assume that U0i is a given vector, for which
U0i =
p+1X
j=1
↵(0)j · vj , (39)
where ↵(0)n 6= 0. Let
8n 2 {1, 2, ...} , Uni = G · Un�1
i .
To analyze the convergence of the sequence Uni , one obtains
Uni =
p+1X
j=1
↵(0)j Gn · vj .
15
Figure 6: Illustration of the spectral analysis. The mesh is composed of four cells. Each cell is delimited by the dotted lines.The black crosses + represent the solution point for each cell. The symbols • represent the initial solution at each solutionpoint and � the computed solution at each solution point.
4.5. Spectral analysis: the Matrix Power Method215
Using properties of geometric progessions and (34), one obtains:
Uni = Gn · U0
i . (38)
Carrying out an unsteady computation with the Spectral Di↵erence Method simply consists in computingthe matrix G to the nth power to obtain the solution at the nth iteration from the initial solution. Itis somewhat reminiscent of the Matrix Power Method. For estimating and computing eigenvalues, thisMethod generally gives the eigenvalue with the greatest absolute value of a given matrix. It consists incomputing the nth power of this matrix. The proof of this theorem can be found in [33]. The followingproof is considerably inspired from the previous reference but it will allow us to establish a general resultfor the spectral analysis of the Spectral Di↵erence Method. Note that this method could be applied apriori for other high-order methods which consider several degrees of freedom inside each cell. Considerthe eigenvalue problem G · Ui = �Ui, Ui 6= 0, G 2 Mp+1 (C), Ui 2 Cp+1, � 2 C, which we assume hasa complete normalized eigenvector space (v1,v2, ...,vp+1) spanning Cp+1, with corresponding eigenvalues
(�1,�2, ...,�p+1) where |�1| < |�2| < ... < |�p+1|. We assume that U0i is a given vector, for which
U0i =
p+1X
j=1
↵(0)j · vj , (39)
where ↵(0)n 6= 0. Let
8n 2 {1, 2, ...} , Uni = G · Un�1
i .
To analyze the convergence of the sequence Uni , one obtains
Uni =
p+1X
j=1
↵(0)j Gn · vj .
15
High-ordernumericalmethodsforLargeEddySimulation 39
◆ Projectingtheinitialsolutionontheeigenvectorsbasis:
◆ Usingthediagonalizationproperties:
◆ Factorising:
TheMatrixPowerMethod
n ! +10
Figure 6: Illustration of the spectral analysis. The mesh is composed of four cells. Each cell is delimited by the dotted lines.The black crosses + represent the solution point for each cell. The symbols • represent the initial solution at each solutionpoint and � the computed solution at each solution point.
4.5. Spectral analysis: the Matrix Power Method215
Using properties of geometric progessions and (34), one obtains:
Uni = Gn · U0
i . (38)
Carrying out an unsteady computation with the Spectral Di↵erence Method simply consists in computingthe matrix G to the nth power to obtain the solution at the nth iteration from the initial solution. Itis somewhat reminiscent of the Matrix Power Method. For estimating and computing eigenvalues, thisMethod generally gives the eigenvalue with the greatest absolute value of a given matrix. It consists incomputing the nth power of this matrix. The proof of this theorem can be found in [33]. The followingproof is considerably inspired from the previous reference but it will allow us to establish a general resultfor the spectral analysis of the Spectral Di↵erence Method. Note that this method could be applied apriori for other high-order methods which consider several degrees of freedom inside each cell. Considerthe eigenvalue problem G · Ui = �Ui, Ui 6= 0, G 2 Mp+1 (C), Ui 2 Cp+1, � 2 C, which we assume hasa complete normalized eigenvector space (v1,v2, ...,vp+1) spanning Cp+1, with corresponding eigenvalues
(�1,�2, ...,�p+1) where |�1| < |�2| < ... < |�p+1|. We assume that U0i is a given vector, for which
U0i =
p+1X
j=1
↵(0)j · vj , (39)
where ↵(0)n 6= 0. Let
8n 2 {1, 2, ...} , Uni = G · Un�1
i .
To analyze the convergence of the sequence Uni , one obtains
Uni =
p+1X
j=1
↵(0)j Gn · vj .
15
Figure 6: Illustration of the spectral analysis. The mesh is composed of four cells. Each cell is delimited by the dotted lines.The black crosses + represent the solution point for each cell. The symbols • represent the initial solution at each solutionpoint and � the computed solution at each solution point.
4.5. Spectral analysis: the Matrix Power Method215
Using properties of geometric progessions and (34), one obtains:
Uni = Gn · U0
i . (38)
Carrying out an unsteady computation with the Spectral Di↵erence Method simply consists in computingthe matrix G to the nth power to obtain the solution at the nth iteration from the initial solution. Itis somewhat reminiscent of the Matrix Power Method. For estimating and computing eigenvalues, thisMethod generally gives the eigenvalue with the greatest absolute value of a given matrix. It consists incomputing the nth power of this matrix. The proof of this theorem can be found in [33]. The followingproof is considerably inspired from the previous reference but it will allow us to establish a general resultfor the spectral analysis of the Spectral Di↵erence Method. Note that this method could be applied apriori for other high-order methods which consider several degrees of freedom inside each cell. Considerthe eigenvalue problem G · Ui = �Ui, Ui 6= 0, G 2 Mp+1 (C), Ui 2 Cp+1, � 2 C, which we assume hasa complete normalized eigenvector space (v1,v2, ...,vp+1) spanning Cp+1, with corresponding eigenvalues
(�1,�2, ...,�p+1) where |�1| < |�2| < ... < |�p+1|. We assume that U0i is a given vector, for which
U0i =
p+1X
j=1
↵(0)j · vj , (39)
where ↵(0)n 6= 0. Let
8n 2 {1, 2, ...} , Uni = G · Un�1
i .
To analyze the convergence of the sequence Uni , one obtains
Uni =
p+1X
j=1
↵(0)j Gn · vj .
15
Furthermore, 8j 2 {1, 2, ..., p+ 1}, Gn · vj = �nj vj yields
Uni =
p+1X
j=1
↵(0)j �n
j vj
or
Uni = ↵(0)
p+1�np+1
2
4vp+1 +pX
j=1
↵(0)j
↵(0)p+1
✓�j
�p+1
◆n
vj
3
5 .
Hence, using��� �j
�p+1
��� = 1� |�p+1|�|�j ||�p+1| , we find
�����
�����vp+1 �1
↵(0)p+1�
np+1
Uni
�����
�����1
6pX
j=1
�����↵(0)j
↵(0)p+1
����� (1� ap+1)k ,
where ap+1 = minj 6=p+1
��� |�p+1|�|�j |�p+1
���, that is, Uni is approximated by ↵(0)
p+1�np+1vp+1. The rate of convergence
depends on the gap between the absolute magnitude of eigenvalues. However, it is an exponential decreaseand during an unsteady computation, several thousands of iterations can be performed. More precisely, oneobtains:
Uni ⇠
n!+1↵(0)p+1�
np+1vp+1 (40)
This means that for n enough high, the solution at the iteration n� 1 is simply multiply by �p+1 to obtainthe solution at the iteration n. We will see that in practice that n has no need to be so high since the rate ofconvergence is exponentionally decreasing. This gives a way to perform the spectral analysis of the SpectralDi↵erence Method. The dispersion and the amplification are simply given by the spectral radius �p+1 of thematrix G. Between the iterations n and n+ 1, the dispersion is given by the argument ' = � arg (�p+1) /⌫220
and the amplification is given by the absolute value ⇢ = |�p+1|. The Fig. 7a represents the absolute value ofthe di↵erence between k�x and '. The more these quantity tends to zero, the less dispersive the numericalscheme is. This quantity is characterisitc of the dispersion error. The Fig. 7b represents the di↵erencebetween 1 and ⇢. The more these quantity tends to zero, the less there is dissipation. This quantity ischaracteristic of the dissipation error.225
(a) Dispersion error. (b) Dissipation error.
Figure 7: Spectral analysis for ⌫ = 0.1: e↵ect of the order of the solution reconstruction.
16
Furthermore, 8j 2 {1, 2, ..., p+ 1}, Gn · vj = �nj vj yields
Uni =
p+1X
j=1
↵(0)j �n
j vj
or
Uni = ↵(0)
p+1�np+1
2
4vp+1 +pX
j=1
↵(0)j
↵(0)p+1
✓�j
�p+1
◆n
vj
3
5 .
Hence, using��� �j
�p+1
��� = 1� |�p+1|�|�j ||�p+1| , we find
�����
�����vp+1 �1
↵(0)p+1�
np+1
Uni
�����
�����1
6pX
j=1
�����↵(0)j
↵(0)p+1
����� (1� ap+1)k ,
where ap+1 = minj 6=p+1
��� |�p+1|�|�j |�p+1
���, that is, Uni is approximated by ↵(0)
p+1�np+1vp+1. The rate of convergence
depends on the gap between the absolute magnitude of eigenvalues. However, it is an exponential decreaseand during an unsteady computation, several thousands of iterations can be performed. More precisely, oneobtains:
Uni ⇠
n!+1↵(0)p+1�
np+1vp+1 (40)
This means that for n enough high, the solution at the iteration n� 1 is simply multiply by �p+1 to obtainthe solution at the iteration n. We will see that in practice that n has no need to be so high since the rate ofconvergence is exponentionally decreasing. This gives a way to perform the spectral analysis of the SpectralDi↵erence Method. The dispersion and the amplification are simply given by the spectral radius �p+1 of thematrix G. Between the iterations n and n+ 1, the dispersion is given by the argument ' = � arg (�p+1) /⌫220
and the amplification is given by the absolute value ⇢ = |�p+1|. The Fig. 7a represents the absolute value ofthe di↵erence between k�x and '. The more these quantity tends to zero, the less dispersive the numericalscheme is. This quantity is characterisitc of the dispersion error. The Fig. 7b represents the di↵erencebetween 1 and ⇢. The more these quantity tends to zero, the less there is dissipation. This quantity ischaracteristic of the dissipation error.225
(a) Dispersion error. (b) Dissipation error.
Figure 7: Spectral analysis for ⌫ = 0.1: e↵ect of the order of the solution reconstruction.
16
Un+1i = G · Un
i
High-ordernumericalmethodsforLargeEddySimulation 40
◆ Onecouldfindtheasymptoticbehaviour:
◆ Thisasymptoticbehaviourisdrivenbythespectralradius.
◆ Wedefinecharacterisationsofdispersionanddissipation:
TheMatrixPowerMethod
Un+1i = G · Un
i
Furthermore, 8j 2 {1, 2, ..., p+ 1}, Gn · vj = �nj vj yields
Uni =
p+1X
j=1
↵(0)j �n
j vj
or
Uni = ↵(0)
p+1�np+1
2
4vp+1 +pX
j=1
↵(0)j
↵(0)p+1
✓�j
�p+1
◆n
vj
3
5 .
Hence, using��� �j
�p+1
��� = 1� |�p+1|�|�j ||�p+1| , we find
�����
�����vp+1 �1
↵(0)p+1�
np+1
Uni
�����
�����1
6pX
j=1
�����↵(0)j
↵(0)p+1
����� (1� ap+1)k ,
where ap+1 = minj 6=p+1
��� |�p+1|�|�j |�p+1
���, that is, Uni is approximated by ↵(0)
p+1�np+1vp+1. The rate of convergence
depends on the gap between the absolute magnitude of eigenvalues. However, it is an exponential decreaseand during an unsteady computation, several thousands of iterations can be performed. More precisely, oneobtains:
Uni ⇠
n!+1↵(0)p+1�
np+1vp+1 (40)
This means that for n enough high, the solution at the iteration n� 1 is simply multiply by �p+1 to obtainthe solution at the iteration n. We will see that in practice that n has no need to be so high since the rate ofconvergence is exponentionally decreasing. This gives a way to perform the spectral analysis of the SpectralDi↵erence Method. The dispersion and the amplification are simply given by the spectral radius �p+1 of thematrix G. Between the iterations n and n+ 1, the dispersion is given by the argument ' = � arg (�p+1) /⌫220
and the amplification is given by the absolute value ⇢ = |�p+1|. The Fig. 7a represents the absolute value ofthe di↵erence between k�x and '. The more these quantity tends to zero, the less dispersive the numericalscheme is. This quantity is characterisitc of the dispersion error. The Fig. 7b represents the di↵erencebetween 1 and ⇢. The more these quantity tends to zero, the less there is dissipation. This quantity ischaracteristic of the dissipation error.225
(a) Dispersion error. (b) Dissipation error.
Figure 7: Spectral analysis for ⌫ = 0.1: e↵ect of the order of the solution reconstruction.
16
' ⇠ arg (�p+1)
⇢ ⇠ |�p+1|
High-ordernumericalmethodsforLargeEddySimulation 41
TheMatrixPowerMethod
10−7
10−6
10−5
10−4
10−3
10−2
10−1
π/4 π/2 3π/4 π
|k·∆
x−ϕ|[−
]
k ·∆x [−]
SD2SD3SD4SD5
DispersionerrorCFL=0.1
High-ordernumericalmethodsforLargeEddySimulation 42
TheMatrixPowerMethod
10−12
10−10
10−8
10−6
10−4
10−2
π/4 π/2 3π/4 π
1−
ρ[−
]
k ·∆x [−]
SD2SD3SD4SD5
DissipationerrorCFL=0.1
High-ordernumericalmethodsforLargeEddySimulation 43
ExtensiontohighwavenumbersfortheSpectralDifferenceMethod
◆ Energylossestimationbetweentheiterationnandm(n>m):
◆ Phaseshiftestimationbetweentheiterationnandm(n>m):
High-ordernumericalmethodsforLargeEddySimulation 44
ComparisonwithstandardFiniteDifferencemethods
Dissipationfor100000iterationsatCFL=0.7.
RKo6s-SD5schemeisasaccurateasRKo6s-CF8-CS6scheme.
High-ordernumericalmethodsforLargeEddySimulation 45
ComparisonwithstandardFiniteDifferencemethods
Dispersionfor100000iterationsatCFL=0.7.
RKo6s-SD5schemeisasaccurateasRKo6s-CF8-CS6scheme.[2] J.Vanharen,G.Puigt,X.Vasseur, J.-F.Boussuge,P. Sagaut.Revisicng the spectral analysis forhigh-order spectraldisconcnuousmethods.Underreview,JournalofComputaconalPhysics,2016.
High-ordernumericalmethodsforLargeEddySimulation
HPCefficiency
46
High-ordernumericalmethodsforLargeEddySimulation 47
JAGUARCPUperformance(Euler)
GFLOPSobtainedarealmosttwiceasbigas standardCFDsolvers.
MeantimebyiterationandbyDoF:2.2µsin2DEuler(about8µsforNSin3D)
whichislowerthanstandardCFDsolvers.
High-ordernumericalmethodsforLargeEddySimulation
Partialconclusion
48
JAGUARisanalternativesolvertostandardFiniteVolumeones:
Unstructuredgridforcomplexgeometry.
Highperformancecapabilitywithverygoodscaling.
AsaccurateasstandardFiniteVolumeschemes.
Butonlyhexahedralmeshes…
High-ordernumericalmethodsforLargeEddySimulationINRIASeminarPalaiseau,France,November24,2016
J.VanharenJointworkwithG.Puigt,M.Montagnac,X.Vasseur,J.-F.BoussugeandP.Sagaut.
CONCLUSION&PERPECTIVES
49
High-ordernumericalmethodsforLargeEddySimulation
Conclusion&perpectives
Extensionofthenonconforminggridinterfaceforhybridmeshes.
Second-orderaccurateevenfornonconforminggridinterface.
Workflowcreationforhybridmeshes.
RelativeHPCefficiency.
Theextensiontohigh-orderschemesisimpossible forunstructuredmeshes.
50
High-ordernumericalmethodsforLargeEddySimulation
JAGUARisanalternativetostandardFiniteVolumesolvers.
Anyarbitraryorderofaccuracy.HPCefficient.
Butonlyforhexahedralmeshes…
Splittetrahedraintohexahedra…ExtensionoftheSpectralDifference
Methodontetrahedra.
51
Conclusion&perpectives
High-ordernumericalmethodsforLargeEddySimulation
LaBS:solvetheBoltzmannequation.Goodaccuracy.
(likeasixth-orderschemesfordissipation).HPCefficient.
Canhandleverycomplexgeometries withnonbody-fittedoctreemeshes.
Transitionresolution.Weaklycompressibleformulation.
52
Conclusion&perpectives
High-ordernumericalmethodsforLargeEddySimulation 53
Conclusion&perpectives
High-ordernumericalmethodsforLargeEddySimulation
◆ High-orderschemesbibliographyforLES.◆ elsAH,ANTARES&JAGUARdevelopmentsinPython,C/C++&Fortran90.
◆ Parallelprogramming:ArgonneTrainingProgramonExtreme-ScaleComputing(ATPESC2016).
◆ Airbusinternship:– fluide/structureinteraction,– gustresponse.
◆ AvailableinMay2017.
54
Conclusion&perpectives
High-ordernumericalmethodsforLargeEddySimulation 55
Thankyouforyourattention.JulienVanharen
PhDstudentAirbus/Cerfacsvanharen@cerfacs.fr
References:
[1] J. Vanharen, G. Puigt, M. Montagnac. Theoreccal and numerical analysis ofnonconforming grid interface for unsteady flows. Journal of Computaconal Physics,285:111–132,2015.[2] J.Vanharen,G.Puigt, X.Vasseur, J.-F.Boussuge,P. Sagaut.Revisicng the spectralanalysis for high-order spectral disconcnuous methods. Under review, Journal ofComputaconalPhysics,2016.[3] P. Cayot, G. Puigt, J. Vanharen, J.-F. Boussuge, P. Sagaut. Towards second-orderfinite-volume cell-centered diffusion scheme on hybrid unstructured meshes. Inpreparacon,AIAAJournal.[4] J. Vanharen, G. Puigt, J.-F. Boussuge, P. Sagaut. Opcmized Runge-Kuia cmeintegracon forSpectralDifferenceMethod. Inpreparacon, JournalofComputaconalPhysics.
[A] J. Vanharen, G. Puigt, M. Montagnac. Theoreccal and numerical analysis ofnonconforminggridinterfaceforunsteadyflows.AERO2015,Toulouse,France,March30th-31st,April1st2015.[B] J.Vanharen,G.Puigt,and J.-F.Boussuge.Revisicngthespectralanalysis forhigh-order spectral disconcnuous methods. In TILDA - Symposium, Toulouse, France,November21st-23rd2016.
High-ordernumericalmethodsforLargeEddySimulationINRIASeminarPalaiseau,France,November24,2016
J.VanharenJointworkwithG.Puigt,M.Montagnac,X.Vasseur,J.-F.BoussugeandP.Sagaut.
A)FLUXCONSERVATION
56
High-ordernumericalmethodsforLargeEddySimulation
A)Fluxconservation
57
Propertyoffluxconservationlostwhen:1.Noncoplanarfaces2. Curvedinterfaces
High-ordernumericalmethodsforLargeEddySimulationINRIASeminarPalaiseau,France,November24,2016
J.VanharenJointworkwithG.Puigt,M.Montagnac,X.Vasseur,J.-F.BoussugeandP.Sagaut.
B)SHOCK-CAPTURING
58
High-ordernumericalmethodsforLargeEddySimulation
B)Shock-capturing
59
High-ordernumericalmethodsforLargeEddySimulationINRIASeminarPalaiseau,France,November24,2016
J.VanharenJointworkwithG.Puigt,M.Montagnac,X.Vasseur,J.-F.BoussugeandP.Sagaut.
C)INTRODUCTION
60
High-ordernumericalmethodsforLargeEddySimulation 61
C)Industrialneeds
Complex geometries Accuracy
HPCPhysicalmodelling
High-ordernumericalmethodsforLargeEddySimulation
C)Introduction
Unsteady flow phenomena can occur on a large number of aircraqcomponents: -Landinggearsdoors, -Jet-flapinteraccons, -Airbrakesoutduringemergencydescent.
Theresulcngunsteadyaerodynamicforcesincreasethestructuralloadsthat may lead to vibracons and/or structural damage (facgue, limitloads).
The predicRon of these phenomena is tradiRonally based onexperiments,whichprovide resultsof appreciableaccuracy,but lateinthedesignprocess.
62
High-ordernumericalmethodsforLargeEddySimulation
C)Introduction
63
Trackingvorticeso
verlongdistancea
nd
propagatingacous
ticwavesoncomplex
geometriesarestillacha
llengeforCFDand
requireveryaccur
atenumericalmethods
intermsofdissipationand
dispersion.
High-ordernumericalmethodsforLargeEddySimulation 64
C)IndustrialCFDsimulationsatAirbus
Threeactiveresearchactivities:
Cabinnoise
Externalsourcenoise
Landinggear
High-ordernumericalmethodsforLargeEddySimulation
C)Summary
65
Statusandfutureneeds:◆ RANS/URANS, geometry not too complex, structured grid with mesh
movement(ALE).◆ Tendency to treat off-design configurations with complex geometry and
unsteadyflows(withURANSorLES).◆ Threekeypoints:complexgeometries,unsteadyflow&HPC.
Goalofthistalk:◆ Extend theNonconforming Grid Interface (NGI) inhybrid framework for
unsteadyflow.◆ FocusattentiononthelimitsofthisapproachinaFiniteVolumeformalism.◆ Exploreanewparadigms:theSpectralDifferenceMethodandtheLattice-
BoltzmannMethod.
High-ordernumericalmethodsforLargeEddySimulationINRIASeminarPalaiseau,France,November24,2016
J.VanharenJointworkwithG.Puigt,M.Montagnac,X.Vasseur,J.-F.BoussugeandP.Sagaut.
D)THESPECTRALDIFFERENCEMETHOD
66
High-ordernumericalmethodsforLargeEddySimulation 67
Thisexpressionwasobtainedbyaspace-timeFouriertransformbasedonthefollowingnormalmode:
Itis2periodic:
Inordertoavoidaliasing,shouldbelongto:thisistheNyquist-Shannontheorem.
p
kDx [0, p]
Un+1i = G · Un
i
D)SpacetimediscretisationforSpectralDifferenceMethod
High-ordernumericalmethodsforLargeEddySimulation 68
Finally,oneobtains:
D)SpacetimediscretisationforSpectralDifferenceMethod
Un+1i = G · Un
i
Firstremark:allthepropertiesofthespacetimediscretisationareinthismatrixG.
Secondremark:thisavectorexpressionsincethereareseveralsolutionpointspercell.