A staggered space-time discontinuous Galerkin method for the three-dimensional incompressible Navier-Stokes equations on unstructured tetrahedral meshes MaurizioTavellia,1,MichaelDumbserb,2, aDepartmentofMathematics,UniversityofTrento, ViaSommarive14,I-38050Trento,Italy bLaboratoryofAppliedMathematics,DepartmentofCivil,EnvironmentalandMechanicalEngineering,UniversityofTrento,ViaMesiano77, I-38123Trento,Italy 6 1 Abstract 0 2 Inthispaperweproposeanovelarbitraryhighorderaccuratesemi-implicitspace-timediscontinuousGalerkinmethod n for the solution of the three-dimensional incompressible Navier-Stokes equations on staggered unstructured curved a tetrahedralmeshes. Astypicalforspace-timeDGschemes,thediscretesolutionisrepresentedintermsofspace-time J basisfunctions. Thisallowstoachieveveryhighorderofaccuracyalsointime, whichisnoteasytoobtainforthe 7 incompressibleNavier-Stokesequations. Similartostaggeredfinitedifferenceschemes,inourapproachthediscrete 2 pressure is defined on the primary tetrahedral grid, while the discrete velocity is defined on a face-based staggered ] dual grid. While staggered meshes are state of the art in classical finite difference schemes for the incompressible A Navier-Stokes equations, their use in high order DG schemes still quite rare. A very simple and efficient Picard N iterationisusedinordertoderiveaspace-timepressurecorrectionalgorithmthatachievesalsohighorderofaccuracy . intimeandthatavoidsthedirectsolutionofglobalnonlinearsystems. Formalsubstitutionofthediscretemomentum h equationonthedualgridintothediscretecontinuityequationontheprimarygridyieldsaverysparsefive-pointblock t a systemforthescalarpressure,whichisconvenientlysolvedwithamatrix-freeGMRESalgorithm. Fromnumerical m experimentswefindthatthelinearsystemseemstobereasonablywellconditioned,sinceallsimulationsshowninthis [ papercouldberunwithouttheuseofanypreconditioner,evenuptoveryhighpolynomialdegrees. Forapiecewise constantpolynomialapproximationintimeandifpressureboundaryconditionsarespecifiedatleastinonepoint,the 1 v resulting system is, in addition, symmetric and positive definite. This allows us to use even faster iterative solvers, 4 liketheconjugategradientmethod. 8 The flexibility and accuracy of high order space-time DG methods on curved unstructured meshes allows to 3 discretize even complex physical domains with very coarse grids in both, space and time. The proposed method is 7 verifiedforapproximationpolynomialsofdegreeuptofourinspaceandtimebysolvingaseriesoftypical3Dtest 0 . problems and by comparing the obtained numerical results with available exact analytical solutions, or with other 1 numericalorexperimentalreferencedata. 0 6 To the knowledge of the authors, this is the first time that a space-time discontinuous Galerkin finite element 1 method is presented for the three-dimensional incompressible Navier-Stokes equations on staggered unstructured : tetrahedralgrids. v i X Keywords: highorderschemes,space-timediscontinuousGalerkinfiniteelementschemes,staggeredunstructured meshes,space-timepressurecorrectionalgorithm,incompressibleNavier-Stokesequationsin3D r a 1. Introduction ThenumericalsolutionofthethreedimensionalincompressibleNavier-Stokesequationsrepresentsaveryimpor- tant and challenging research topic, both from a numerical and from an application point of view. In the literature, [email protected] (M. Tavelli) [email protected] (M. Dumbser) PreprintsubmittedtoJournalofComputationalPhysics January28,2016 there are many different approaches that have been proposed for the solution of the incompressible Navier-Stokes equations, for example using classical finite difference methods [1, 2, 3, 4] or continuous finite element schemes [5,6,7,8,9,10,11]. Veryrecently, alsodifferenthighorderdiscontinuousGalerkin(DG)methodshavebeenpre- sented for the solution of the incompressible and the compressible Navier-Stokes equations. The first DG schemes that were able to solve the Navier-Stokes equations were those of Bassi and Rebay [12] and Baumann and Oden [13, 14]. Many other methods have been presented in the meantime, see for example [12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27] for a non-exhaustive overview of the ongoing research in this very active field. In most DG schemes, the DG discretization is only used for space discretization, while the time discretization uses standardexplicitorimplicittimeintegratorsknownforordinarydifferentialequations,followingtheso-calledmethod oflinesapproach. ThemethodoflineshasalsobeenusedbyCockburnandShuintheirwell-knownseriesofpapers [28, 29, 30] on DG schemes for time-dependent nonlinear hyperbolic systems. In contrast to the method of lines approach,thefamilyofspace-timediscontinuousGalerkinfiniteelementschemes,whichwasintroducedforthefirst timebyvanderVegtetal. in[31,32,33],treatsspaceandtimeinaunifiedmanner. Thisisachievedbyusingtestand basisfunctionsthatdependonbothspaceandtime,see[34,35,36,37,38,39,40]foranoverviewofrecentresults. Foraveryearlyimplementationofcontinuousspace-timefiniteelementschemes,thereaderisalsoreferredto[41]. From an application point of view, it is very important to consider the fully three-dimensional Navier-Stokes equations,inordertocapturetherelevantflowfeaturesthatareobservedinlaboratoryexperiments,see[42,43,44, 45]. Thismeansthattheuseofatwo-dimensionalalgorithmisinmostcasesinappropriatetoreproducetheresults ofphysicalexperiments,evenforgeometriesthatcanbeconsideredessentiallytwo-dimensional. Theimportanceof fully three-dimensional computations has been shown, for example, in [45, 46, 47, 48, 49, 50, 51]. Unfortunately, themeshgenerationforcomplexandrealistic3Dgeometriesisstillnowadaysquitedifficult,andthecomputational cost of a fully three-dimensional simulation grows very quickly with increasing mesh resolution. In this context, it becomes crucial to use unstructured simplex meshes, since they help to simplify the process of mesh generation significantly compared to unstructured hexahedral meshes. Furthermore, it is at the same time also crucial to use very high order accurate methods in both space and time, since they allow to reduce the total number of elements significantly,comparedtolowordermethods,whilekeepingatthesametimeahighlevelofaccuracyofthenumerical solution. SincethesolutionoftheincompressibleNavier-Stokesequationsrequiresnecessarilythesolutionoflarge systems of algebraic equations, it is indeed very important to derive a scheme that uses a stencil that is as small as possible,inordertoimprovethesparsitypatternoftheresultingsystemmatrices.Itisalsodesirabletodesignmethods thatleadtoreasonablywellconditionedsystemsthatcanbesolvedwithiterativesolvers,liketheconjugategradient method[52]ortheGMRESalgorithm[53]. Forstructuredgrids,numericalschemescanbeusuallyderivedrathereasilyinmultiplespacedimensions,thanks totheparticularregularityofthemesh. Onthecontrary,thedevelopmentofnumericalschemesongeneralunstruc- tured meshes in three space dimensions is not as straightforward and requires some care in the derivation and the implementationofthemethod. ParticulardifficultiesoftheincompressibleNavier-Stokesequationsarisefromtheir nonlinearityandfromtheellipticnatureofthePoissonequationforthepressure,thatisalsoobtainedonthediscrete level when substituting the momentum equation into the discrete continuity equation. A unified analysis of several variantsoftheDGmethodappliedtoanellipticmodelproblemhasbeenprovidedbyArnoldetal. in[54]. While the use of staggered grids is a very common practice in the finite difference community, its use is not so widespreadinthecontextofhighorderDGschemes. ThefirststaggeredDGschemes,basedonavertex-baseddual grid,havebeenproposedin[55,56]. OtherrecenthighorderstaggeredDGschemesthatuseanedge-baseddualgrid havebeenforwardedin[57,58,59]. Theadvantageinusingedge-basedstaggeredgridsisthattheyallowtoimprove significantly the sparsity pattern of the final linear system that has to be solved for the pressure. Very recently, a newfamilyofstaggered semi-implicitDGschemesforthesolutionofthetwodimensionalshallowwaterequations waspresentedbyDumbser&Casulli[59]andTavelli&Dumbser[60]. Subsequently,thesesemi-implicitstaggered DG schemes have been successfully extended also to the two-dimensional incompressible Navier-Stokes equations by Tavelli & Dumbser in [61, 39]. One year later, a staggered DG formulation for the 2D incompressible Navier- Stokes equations has been reproposed independently also in [62]. Alternative semi-implicit discontinuous Galerkin schemesoncollocatedgridshavebeenpresented,forexample,in[63,64,65,66,67]. Thesesemi-implicitschemes trytocombinethesimplicityofexplicitmethodsfornonlinearPDEwiththestabilityandefficiencyofimplicittime discretizations. Inthispaperweproposeanew,arbitraryhighorderaccuratestaggered space-timediscontinuousGalerkinfinite 2 elementmethodforthesolutionofthethree-dimensionalincompressibleNavier-Stokesequationsoncurvedunstruc- turedtetrahedralmeshes,followingsomeoftheideasoutlinedin[39]forthetwo-dimensionalcase. Forthatpurpose we mimic the philosophy of staggered semi-implicit finite difference schemes, such as discussed and analyzed in [1,2,3,4,68,69,70,71,72,73,74,75,76,77],wherethediscretepressurefieldisdefinedontheprimarygrid,while thediscretevelocityfieldisdefinedonanedge-basedstaggereddualgrid. For the staggered space-time DG scheme proposed in this paper, we use a primal mesh composed of (curved) tetrahedral elements, and a face-based staggered dual mesh that consists of non-standard five-point hexahedral ele- ments that are obtained by connecting the three nodes of a face of the primal mesh with the barycenters of the two tetrahedra that share the common face. The face-based dual grid used here corresponds to the choice made also in [78,79,80,58].Thesespatialelementsarethenextendedtospace-timecontrolvolumesusingasimpletensorproduct inthetimedirection. Since all quantities are readily defined where they are needed, the staggered DG scheme does not require the useofRiemannsolvers(numericalfluxfunctions),apartfromthenonlinearconvectiveterms,whicharetreatedina standardway. FortheconvectivepartoftheincompressibleNavier-Stokesequations,weuseastandardDGscheme for hyperbolic PDE on the main grid, based on the local Lax-Friedrichs (Rusanov) flux [81]. For that purpose, the velocityfieldisfirstinterpolatedfromthedualgridtothemaingrid,assuggestedin[59]. Thisallowsustousethe same staggered space-time DG scheme again to discretize also the viscous terms, where now the velocity gradient thatisneededfortheevaluationoftheviscousfluxesiscomputedontheface-basedstaggereddualgrid. Inthisway, we can avoid again the use of numerical flux functions for the viscous fluxes, and furthermore, the structure of the resultinglinearsystemsfortheviscoustermsisverysimilartothepressuresystem. The discrete momentum equation is then inserted into the discrete continuity equation in order to obtain the discrete form of the pressure Poisson equation. Thanks to the use of a staggered grid, this leads to a very sparse five-point block system, with the scalar pressure as the only unknown quantity. Note that the same algorithm on a collocatedgrid wouldproducea17-pointstencil,sinceitwouldalsoinvolveneighborsofneighbors3. Ontheother hand,ifonedoesnotsubstitutethemomentumequationintothecontinuityequationonacollocatedgrid,onecould stillobtainafivepointstencil, butwiththepressureand thevelocityvectorasunknowns, hencethefinalsystemto solveisfourtimeslargerthanthecorrespondingsystemofourstaggeredDGscheme. Itisthereforeveryclearthat even in the DG context, the use of a staggered mesh is very beneficial, since it allows to produce a linear system with the smallest possible stencil and with the smallest number of unknowns, compared to similar approaches on a collocatedmesh. Oncethenewpressurefieldisknown,thevelocityvectorfieldcansubsequentlybeupdateddirectly.Averysimple Picarditerationthatembracestheentireschemeineachtimestepisusedinordertoachievearbitraryhighorderof accuracyintimealsoforthenonlinearconvectiveandviscousterms,withoutintroducinganonlinearityinthesystem forthepressure. The rest of the paper is organized as follows: in Section 2 we derive and present the new numerical method method. Section2.5containsthedetailsaboutthediscretizationofthenonlinearconvectivetermsonthemaingrid, while the velocity gradients for the viscous terms are discretized again on the face-based staggered dual mesh. In Section 2.7 we discuss the important special case of a high order DG discretization in space, while using only a piecewiseconstantpolynomialapproximationintime,leadingtosymmetricpositivedefinitesystemsforthepressure andtheviscousterms. Finally, inSection3thenewnumericalschemeproposedinthispaperisrunonasetof3D benchmarkproblems,comparingthenumericalresultseitherwithexistinganalyticalornumericalreferencesolutions, orwithavailableexperimentalresults. ThepapercloseswithsomeconcludingremarksprovidedinSection4. 3ThediscretecontinuityequationofaDGschemeonacollocatedgridinvolvesthevelocityintheelementitselfandinitsfourneighbor elements,duetothenumericalfluxontheelementboundaries.Furthermore,inthediscretemomentumequationthevelocityfieldineachtetrahedral elementdependsonthepressureinthecellitselfandinitsfourneighbors.Insertingnowthemomentumequationintothecontinuityequationon thediscretelevelinvolvesatotalof1+4+4·3=17elementsforthepressure! Onastaggeredmeshinstead,thediscretecontinuityequationinvolvesonlythevelocitiesofthefourdualelementsassociatedwiththefacesofthe primaryelement. Thediscretemomentumequationwrittenontheface-baseddualgridonlyinvolvesthepressureofthetwotetrahedrathatshare thecommonface.Hence,substitutingthemomentumequationintothecontinuityequationleadstoa1+4=5pointstencilforthepressure,which involvesonlytheelementanditsfourneighbors. 3 2. Staggeredspace-timeDGschemeforthe3DincompressibleNavier-Stokesequations 2.1. Governingequations Thethree-dimensionalincompressibleNavier-Stokesequationscanbewrittenas ∂v +∇·F +∇p=∇·(ν∇v)+S, (1) ∂t c ∇·v=0, (2) wherex = (x,y,z)isthevectorofspatialcoordinatesandtdenotesthetime; p = P/ρindicatesthenormalizedfluid pressure;Pisthephysicalpressureandρistheconstantfluiddensity;ν = µ/ρisthekinematicviscositycoefficient; v = (u,v,w) is the velocity vector; u, v and w are the velocity components in the x, y and z direction, respectively; S=S(x,t)isavectorofgivensourceterms;F =v⊗visthefluxtensorofthenonlinearconvectiveterms,namely: c Fc = uvuu uvvv uvww . wu wv ww Theviscositytermcanbegroupedwiththenonlinearconvectiveterm,i.e. themomentumEq. (1)thenreads ∂v +∇·F+∇p=S, (3) ∂t whereF=F(v,∇v)=F (v)−ν∇visthenonlinearfluxtensorthatdependsonthevelocityanditsgradient. c 2.2. Staggeredunstructuredmeshandassociatedspace-timebasisfunctions Throughoutthispaperweuseamaingridthatiscomposedof(eventuallycurved)tetrahedralsimplexelements, andastaggeredface-baseddualgrid,consistinginnon-standardfive-pointhexahedralelements. Thesespatialcontrol volumesarethenextendedtospace-timecontrolvolumesusingatensorproductintimedirection. Inthefollowing, thestaggeredmeshinspaceisdescribedindetailandissubsequentlyalsoextendedtothetimedirection. Themain notationistakenastheonepresentedforthetwodimensionalmethodproposedin[39]andissummarizedherefor thethreedimensionalcase. 2.2.1. Staggeredspace-timecontrolvolumes The spatial computational domain Ω is covered with a set of N non-overlapping tetrahedral elements T with e i i = 1...N . Bydenotingwith N thetotalnumberoffaces,the j−thfacewillbecalledΓ . B(Ω)denotesthesetof e d j indices jcorrespondingtoboundaryfaces. TheindicesofthefourfacesofeachtetrahedronT constitutethesetS i i definedbyS = {j ∈ [1,N ] | Γ isafaceofT}. Forevery j ∈ [1...N ]−B(Ω)thereexisttwotetrahedrathatshare i d j i d acommonfaceΓ . Weassignarbitrarilyaleftandarightelement,calledT andT ,respectively. Thestandard j (cid:96)(j) r(j) positivedirectionisassumedtobefromlefttoright. Let(cid:126)n denotetheunitnormalvectordefinedonthefacenumber j jandthatisorientedwithrespecttothepositivedirectionfromlefttoright. Foreverytetrahedralelementnumberi andfacenumber j∈S ,theindexoftheneighbortetrahedronthatsharesthecommonfaceΓ isdenotedby℘(i, j). i j For every j ∈ [1,N ]−B(Ω) the dual element (a non-standard5-point hexahedron) associated with Γ is called d j H anditisdefinedbythetwocentersofgravityofT andT andthethreeverticesofΓ ,seealso[78,80,60]. j (cid:96)(j) r(j) j WedenotebyT = H ∩T theintersectionelementforeveryiand j ∈ S . Figures1and2summarizethenotation i,j j i i usedonthemaintetrahedralmeshandontheassociateddualgrid. Weexdendourdefinitionsonthemaingridtothe dualone,namely: N isthetotalamountofsidesofH ;Γ indicatesthel-thside;∀j,thesetofsideslof jisindicated l j l withS ;∀l,(cid:96) (l)andr (l)aretheleftandtherighthexahedralelement,respectively;(cid:126)n isthestandardnormalvector j jl jl l definedonlandassumedpositivewithrespecttothestandardorientationonl(defined,asforthemaingrid,fromthe lefttotheright). Inthetimedirectionwecoverthetimeinterval[0,T]withasequenceoftimes0=t0 <t1 <t2...<tN <tN+1 =T. Wedenotethetimestepbetweentnandtn+1by∆tn+1 =tn+1−tnandtheassociatedtimeintervalbyTn+1 =[tn,tn+1],for n = 0...N. Inordertoeasethenotation,sometimeswewillsimplywrite∆t = ∆tn+1. Inthiswaythegenericspace- timeelementdefinedinthetimeinterval[tn,tn+1]isgivenbyTst =T ×Tn+1 forthemaingridand Hst = H ×Tn+1 i i j j forthedualgrid. 4 L Γ Γj3 j1 Hj i T i T i Γ j2 ℘(i, j) Γ Γ Γj3 j1 Hj i j4 r(j) T i (cid:126)n j T (cid:96)(j) i Γ j2 ℘(i, j) L Γ j Γ j4 r(j) Figure1:AtetrahedralelementoftheprimarymeshwithSi={j1,j2,j3,j4}(left)andthestandardorientationusedthroughoutthispaper(right). (cid:96)(j) n(cid:126)j Γ H j j Γ Γj3℘(i, j) j1 Hj i i T i Ti Ti Γ j2 ℘(i, j) Γ j4 r(j) ℘(i, j) Γ Γj3 j1 Hj i Figure2:Anexampleofadualelement(anon-standard5-pointhexahedron,highlightedinblue)associatedwiththefaceΓj. n(cid:126) j (cid:96)(j) T i T i 2.3. Space-timebasisfunctions Γ j2 ℘(i, j) Wefirstconstructthespatialbasisfunctionsandthenweextendthemtothetimedirectionusingasimpletensor Γj product. Fortetrahedralelements, thebasisfunctionsaregeneratedonastandardreferencetetrahedron, definedby Ti Γj4 Tref ={(ξ,η,ζ)∈R3 | 0≤ξ+η+ζ ≤1}. Wewritert(hje)basisfunctiononthereferenceelementas (cid:88)p (cid:88)p−r1p−(cid:88)r2−r1 Γj1 φk(ξ,η,(cid:96)ζ()j)=r1n=(cid:126)0j r2=0 r3=0 αkrξr1ηr2ζr3 :=αkrξr, (4) for some coefficients α and the multi-index r = (r ,r ,r ). We then define N = (p+1)(p+2)(p+3) nodal points ξ = kr 1 2 3 φ 6 j (ξ ,η ,ζ ) = (j /p, j /p, j /p), with the multi-index j = (j , j , j ) and 0 ≤ j + j + j ≤ p, as in standard cojn1forjm2 inj3gfinitee1leme2nts.W3ethenimposetheclassicalintΓerjpol1atio2nc3onditionforn1odal2finite3elementsφ (ξ )=δ , k j kj withtheusualKroneckersymbolδ . ThismeansthatwehavechosenanodalbasiswhichisdefinedbytheLagrange kj interpolationpolynomialsthatpassthroughthenodesgivenbythestandardnodesofconformingfiniteelements.This leadstothelinearsystemα ξr =δ forthecoefficientsα thatcanbesolvedanalyticallyforeverypolynomialdegree kr j kj kr 5 ponthereferencetetrahedron. InthiswayweobtainN basisfunctionsonT ,{φ } . Theconnectionbetween φ ref k k∈[1,Nφ] the reference coordinates ξ and the physical coordinates x is performed by the map T(·,T) = T : T −→ T for i i i ref everyi = 1...N anditsinverse,calledT−1(·,T) = T−1 : T ←− T . Themapsfromthephysicaltothereference e i i i ref coordinates can be constructed following a classical sub-parametric or a complete iso-parametric approach and in generalwewillwrite,foralli=1...N ,φ(i)(x,y,z)=φ (T(x,y,z)). e k k i Unfortunately,itisnotsoeasytoconstructasimilarnodalbasisonthedualmesh,duetotheuseofnon-standard 5-pointhexahedralelements. Asdiscussedin[82],thedefinitionofbasisfunctionsbasedonLagrangeinterpolation polynomialsonthiskindofelementisproblematic, sinceforspecialconfigurationsofthevertexcoordinatesofthe dual elements, the linear system to be solved for the classical interpolation condition of a nodal basis can become singular. ThisdoesnotallowtheconstructionofanodalpolynomialbasisforagenericelementH andthereforeone j hastopasstorationalfunctionsofpolynomialsinsteadofusingsimplepolynomialfunctionsinthatcase. Therefore, for the basis functions on the dual grid directly we choose a simple Taylor-type modal basis [83] directly in the physical space, hence the basis functions will consequently depend on the element j ∈ [1,N ]. The d basisfunctionsread (cid:16)x−x(j)(cid:17)k1(cid:16)y−y(j)(cid:17)k2(cid:16)z−z(j)(cid:17)k3 ψ(j)(x)= 0 0 0 , (5) k hk1+k2+k3 j wherexj =(x,y,z)(j)isthecenterofthedualelementandh isacharacteristiclengthofH usedforscalingthebasis. 0 0 j j Here, 0 ≤ k +k +k ≤ p, i.e. weusetheoptimalnumberofpolynomialsofdegree pinthreespacedimensions, 1 2 3 namelyN = N . Withthischoicewegetonlyamodalbasisforthedualhexaxedralelements,i.e. iftheconvective ψ φ termisdirectlycomputedonthedualmeshaccordingtothenaturalextensionofthemethodproposedin[39],thenit hastobecomputedaccordingtoamodalapproach,whichismoreexpensivethananodalone. Finally,thetimebasisfunctionsareconstructedonareferenceintervalI =[0,1]forpolynomialsofdegree p . In γ thiscasetheresulting N = p +1basisfunctions{γ } aredefinedastheLagrangeinterpolationpolynomials passing through the Gauγss-Leγgendre quadrature poinktsk∈f[o1,rNγt]he unit interval. For every time interval [tn,tn+1], the map between the reference interval and the physical one is simply given by t = tn +τ∆tn+1 for every τ ∈ [0,1]. Usingthetensorproductwecanfinallyconstructthebasisfunctionsonthespace-timeelementsTst and Hst suchas i j φ˜(ξ,η,ζ,τ) = φ(ξ,η,ζ)·γ(τ) and ψ˜(j)(x,y,z,t) = ψ(j)(x,y,z)·γ(τ(t)). The total number of basis functions becomes Nst = N ·N andNst = N ·N . φ φ γ ψ ψ γ 2.4. Staggeredsemi-implicitspace-timeDGscheme Thediscretepressure p isdefinedonthemaingrid,namely p (x,t)| = p(x,t)= p,whilethediscretevelocity h h Tst i i vectorfieldv isdefinedonthedualgrid,namelyv (x,t)| =v (x,t)=vi . h h Hst j j j Thenumericalsolutionof(2)-(3)isrepresentedineachspace-timeelementTst andHst betweentimestnandtn+1 i j bypiecewisepolynomialsas Nst (cid:88)φ p(x,t)= φ˜(i)(x,t)pˆn+1 =:φ˜(i)(x,t)pˆn+1, (6) i l l,i i l=1 Nst (cid:88)ψ v (x,t)= ψ˜(j)(x,t)vˆn+1 =:ψ˜(j)(x,t)vˆn+1, (7) j l l,j j l=1 wherethevectorofbasisfunctionsφ˜(x,t)isgeneratedfromφ˜(ξ,η,ζ,τ)onT ×[0,1]whileψ˜(j)(x,t)isdefinedfor std every j∈[1...N ]−B(Ω)directlyinthephysicalspace. d A weak formulation of the continuity equation (2) is obtained by multiplying it with a test function φ˜(i) and k integratingoverthespace-timecontrolvolumeTst,foreveryk=1...Nst. Theresultingweakformulationreads i φ (cid:90) φ˜(i)∇·vdxdt=0, (8) k Tst i 6 withdx = dxdydz. Similarly, multiplicationofthemomentumequation(3)bythetestfunctionψ˜(j) andintegrating k overacontrolvolumeHst yields j (cid:90) (cid:32) (cid:33) (cid:90) (cid:90) ∂v ψ˜(j) +∇·F dxdt+ ψ˜(j)∇p dxdt= ψ˜(j)Sdxdt, (9) k ∂t k k Hst Hst Hst j j j forevery j=1...N andk=1...Nst. UsingintegrationbypartsEq. (8)reads d ψ (cid:73) (cid:90) φ˜(i)v·(cid:126)n dSdt− ∇φ˜(i)·vdxdt=0, (10) k i k ∂Tst Tst i i where(cid:126)n indicatestheoutwardpointingunitnormalvector. Duetothediscontinuityof p andv ,equations(9)and i h h (10)havetobesplitasfollows: (cid:88)j∈SiΓ(cid:90)st φ˜(ki)vj·(cid:126)nijdSdt−T(cid:90)st ∇φ˜(ki)·vjdxdt=0, (11) j i,j and (cid:90) (cid:32) (cid:33) (cid:90) (cid:90) (cid:90) (cid:90) ∂v (cid:16) (cid:17) ψ˜(j) j +∇·F dxdt+ ψ˜(j)∇p dxdt+ ψ˜(j)∇p dxdt+ ψ˜(j) p −p (cid:126)n dSdt= ψ˜(j)Sdxdt, k ∂t k (cid:96)(j) k r(j) k r(j) (cid:96)(j) j k Hst Tst Tst Γst Hst j (cid:96)(j),j r(j),j j j (12) where(cid:126)nij =(cid:126)ni|Γsjt;Tis,tj =Ti,j×Tn+1;andΓsjt =Γj×Tn+1.NotethatthepressurehasadiscontinuityalongΓsjtinsidethe hexahedralelementHstandhencethepressuregradientin(9)needstobeinterpretedinthesenseofdistributions,asin j path-conservativefinitevolumeschemes[84,85].Thisleadstothejumptermspresentin(12),see[39].Alternatively, the same jump term can be produced also via forward and backward integration by parts, see e.g. the well-known workofBassiandRebay[12]. Usingdefinitions(6)and(7),werewritetheaboveequationsas (cid:88)j∈SiΓ(cid:90)st φ˜(ki)ψ˜(lj)(cid:126)nijdSdt ·vˆnl,+j1−T(cid:90)st ∇φ˜(ki)ψ˜(lj)dxdt ·vˆnl,+j1=0, (13) j i,j and (cid:90) (cid:90) (cid:90) (cid:90) ∂v ψ˜(j) j dxdt+ ψ˜(j)∇·Fdxdt+ ψ˜(j)∇φ˜((cid:96)(j))dxdt pˆn+1 + ψ˜(j)∇φ˜(r(j))dxdt pˆn+1 k ∂t k k l l,(cid:96)(j) k l l,r(j) Hst Hst Tst Tst j j (cid:96)(j),j r(j),j (cid:90) (cid:90) (cid:90) + ψ˜(j)φ˜(r(j))(cid:126)n dSdtpˆn+1 − ψ˜(j)φ˜((cid:96)(j))(cid:126)n dSdtpˆn+1 = ψ˜(j)Sdxdt, k l j l,r(j) k l j l,(cid:96)(j) k Γst Γst Hst j j j (14) wherewehaveusedthestandardsummationconventionfortherepeatedindexl. Integratingthefirstintegralin(14) bypartsintimeweobtain (cid:90) ∂v (cid:90) (cid:90) (cid:90) ∂ψ˜(j) ψ˜(j) j dxdt= ψ˜(j)(x,tn+1)v (x,tn+1)dx− ψ˜(j)(x,tn)v (x,tn)dx− k v (x,t)dxdt. (15) k ∂t k j k j ∂t j Hst H H Hst j j j j 7 In Eq. (15) we can recognize the fluxes between the current space-time element H ×Tn+1, the future and the past j space-timeelements,aswellasaninternalcontributionthatconnectsinanasymmetricwaythedegreesoffreedom inside the element Hst. Note that the asymmetry appears only in the volume contribution in (15). For the spatial j integral at time tn we will insert the boundary-extrapolated numerical solution from the previous time step, which correspondstoupwindingintimedirectionduetothecausalityprinciple. BysubstitutingEq. (15)into(14)andusing thecausalityprinciple,weobtainthefollowingweakformulationofthemomentumequation: (cid:90) ψ˜k(j)(x,tn+1)ψ˜(lj)(x,tn+1)dx− (cid:90) ∂ψ∂˜t(kj)ψ˜(lj)dxdtvˆnl,+j1−(cid:90) ψ˜(kj)(x,tn)ψ˜(lj)(x,tn)dxvˆnl,j+ (cid:90) ψ˜(kj)∇·Fdx H Hst H Hst j j j j (cid:90) (cid:90) (cid:90) (cid:90) + ψ˜(j)∇φ˜((cid:96)(j))dx pˆn+1 + ψ˜(j)∇φ˜(r(j))dx pˆn+1 + ψ˜(j)φ˜(r(j))(cid:126)n dS pˆn+1 − ψ˜(j)φ˜((cid:96)(j))(cid:126)n dS pˆn+1 k l l,(cid:96)(j) k l l,r(j) k l j l,r(j) k l j l,(cid:96)(j) Tst Tst Γst Γst (cid:96)(j),j r(j),j j j (cid:90) = ψ˜(j)Sdxdt, (16) k Hst j Foreveryiand j,Eqs. (13)and(16)canbewritteninacompactmatrixformas (cid:88) D vˆn+1 =0, (17) i,j j j∈Si and (cid:16) (cid:17) M+− M◦ vˆn+1− M−vˆn+Υ (v,∇v)+R pˆn+1−L pˆn+1 =S , (18) j j j j j j j r(j) j (cid:96)(j) j respectively,where: (cid:90) M+ = ψ˜(j)(x,t(1))ψ˜(j)(x,t(1))dx, (19) j k l H j (cid:90) M− = ψ˜(j)(x,t(0))ψ˜(j)(x,t(1))dx, (20) j k l H j (cid:90) ∂ψ˜(j) M◦ = k ψ˜(j)dxdt, (21) j ∂t l Hst j (cid:90) Υ = ψ˜(j)∇·Fdxdt (22) j k Hst j (cid:90) (cid:90) D = φ˜(i)ψ˜(j)(cid:126)n dSdt− ∇φ˜(i)ψ˜(j)dxdt, (23) i,j k l ij k l Γst Tst j i,j (cid:90) (cid:90) R = ψ˜(j)φ˜(r(j))(cid:126)n dSdt+ ψ˜(j)∇φ˜(r(j))dxdt, (24) j k l j k l Γst Tst j r(j),j (cid:90) (cid:90) L = ψ˜(j)φ˜((cid:96)(j))(cid:126)n dSdt− ψ˜(j)∇φ˜((cid:96)(j))dxdt, (25) j k l j k l Γst Tst j (cid:96)(j),j 8 (cid:90) S = ψ˜(j)Sdxdt. (26) j k Hst j Notehow M◦ introduces,for p > 0,anasymmetriccontributionthatwillleadtoanasymmetryofthemainsystem j γ for the discrete pressure. The action of matrices L and R can be generalized by introducing the new matrix Q , i,j definedas (cid:90) (cid:90) Q = ψ˜(j)∇φ˜(i)dxdt− ψ˜(j)φ˜(i)σ (cid:126)n dsdt, (27) i,j k l k l i,j j Tst Γst i,j j whereσ isasignfunctiondefinedby i,j r(j)−2i+(cid:96)(j) σ = . (28) i,j r(j)−(cid:96)(j) InthiswayQ =−L andQ =R ,andthenEq. (18)becomesintermsofQ (cid:96)(j),j j r(j),j j (cid:16) (cid:17) M+− M◦ vˆn+1− M−vˆn+Υ (v,∇v)+Q pˆn+1+Q pˆn+1 =S , (29) j j j j j j r(j),j r(j) (cid:96)(j),j (cid:96)(j) j or,equivalently, (cid:16) (cid:17) M+− M◦ vˆn+1− M−vˆn+Υ (v,∇v)+Q pˆn+1+Q pˆn+1 =S . (30) j j j j j j i,j i ℘(i,j),j ℘(i,j) j Inordertoeasethenotationwewilluse M = M+−M◦. Hence,thediscreteequations(17)-(18)readasfollows: j j j (cid:88) D vˆn+1 =0, (31) i,j j j∈Si M vˆn+1− M F(cid:99)v +Q pˆn+1+Q pˆn+1 =0, (32) j j j j r(j),j r(j) (cid:96)(j),j (cid:96)(j) where F(cid:99)v is an appropriate discretization of the nonlinear convective, viscous and source terms that will be pre- j sentedlater. Formalsubstitutionofthediscretevelocityfieldgivenbythemomentumequation(32)intothediscrete continuityequation(31),seealso[69,59],yields (cid:88) (cid:88) (cid:88) D M−1Q pˆn+1+ D M−1Q pˆn+1 = D F(cid:99)v . (33) i,j j i,j i i,j j ℘(i,j),j ℘(i,j) i,j j j∈Si j∈Si j∈Si Eq. (33)aboverepresentsablockfive-pointsystemforthepressure pˆn+1. i 2.5. Nonlinearconvectiveandviscousterms We now have to choose a proper discretization for the nonlinear convective and viscous terms. As discussed in [39]weintroduceasimplePicarditerationtoupdatetheinformationaboutthepressure,butwithoutintroducingany nonlinearityintothefinalsystemforthepressure. Hence,fork=1,N ,werewritesystem(33)as pic (cid:88)D M−1Q pˆn+1,k+1+(cid:88)D M−1Q pˆn+1,k+1 =(cid:88)D F(cid:99)vn+1,k+21. (34) i,j j i,j i i,j j ℘(i,j),j ℘(i,j) i,j j j∈Si j∈Si j∈Si TherightsideofEq. (34)canbecomputedbyusingthevelocityfieldatthePicarditerationkandincludingthe viscous effect implicitly, using a fractional step procedure detailed later. Once the new pressure field is known, the velocityvectorfieldatthenewPicarditerationvˆn+1,k+1 canbereadilyupdatedfromthediscretemomentumequation (32). Toclosetheproblemitremainstospecifyhowtoconstructthenonlinearconvective-diffusionoperatorF(cid:99)vn+1,k+21. j Atthispointonecantrytoextendtheprocedurealreadyusedin[39]to3D.However,inthiscasetherearesomeissues thathavetobetakenintoaccount. Inparticular,sinceweareusingamodalbasisonthestaggereddualnon-standard 5-pointhexahedralmesh,wecannotusethesimplenodalapproximationforthenonlinearconvectivetermFˆ =F (vˆ) c c thatconsistsinatrivialpoint-wiseevaluationofthenonlinearoperatorF . Inspiredbythegoodpropertiesobtained c 9 by the use of staggered grids, here we propose a new procedure for the computation of the nonlinear convective and viscous terms. For that purpose, the velocity field is first interpolated from the dual grid to the main grid. The nonlinearconvectivetermscanthenbeeasilydiscretizedwithastandard(space-time)DGschemeonthemaingrid. Then, the staggered mesh is used again in order to define the gradient of the velocity on the dual elements, which allowsustoproduceaverysimpleandsparsesystemforthediscretizationoftheviscousterms. Animplicitdiscretizationoftheviscoustermsonthedualgridleadstoalinearsystemforeachvelocitycomponent thatisaseven-pointnonsymmetricblocksystemwhichis, however, wellconditionedsinceitcanbewrittenasaν perturbationoftheidentitymatrix,seee.g. [39]. Here,wewilldevelopadiscretizationoftheviscoustermsthatleads onlytoafive-pointblocksystemand,moreimportantly,issymmetricandpositivedefiniteforν > 0and p = 0,but γ isstillbetterconditionedalsointhegeneralcase p >0. γ Givenadiscrete velocityfieldv onthedual gridinthetimeinterval[tn,tn+1], wecanproject thevelocityfield h fromthedualmeshtothemaingrid(denotedbyv¯)viastandardL projection, 2 (cid:88) v¯n+1 = M−1 M vˆn+1, ∀i∈[1,N ], (35) i i i,j j e j∈Si wherev¯n+1denotethedegreesoffreedomofthevelocityonthemaingridand i (cid:90) (cid:90) M = φ˜(i)φ˜(i)dxdt, M = φ˜(i)ψ˜(j)dxdt. (36) i k l i,j k l Tst Tst i i,j Theprojectionbackontothedualgridisgivenby vˆn+1 = M−1(cid:16)M(cid:62) vn+1+ M(cid:62) vn+1(cid:17). (37) j j (cid:96)(j),j (cid:96)(j) r(j),j r(j) with (cid:90) M = ψ˜(j)ψ˜(j)dxdt. (38) j k l Hst j Wecanrewritethenonlinearconvectiveandviscouspartofthemomentumequationbyintroducingtheviscous stresstensorσ=−ν∇vasauxiliaryvariable. Theconvectiveandviscoussubsystemofthemomentumequationthen reads ∂v +∇·F +∇·σ = 0, ∂t c σ = −ν∇v. (39) With the averaged velocity v¯n+1 = φ˜(i)v¯n+1 defined on the main grid and the viscous stress tensor σn+1 = ψ˜(j)σn+1 i l l,i j l l,j definedonthedualgrid,weobtainthefollowingweakformulationof(39): (cid:90) (cid:90) (cid:90) ∂φ˜(i) φ˜(i)(x,tn+1)v¯n+1dx− φ˜(i)(x,tn)v¯ndx− k v¯n+1dxdt+ k i k i ∂t i T T Tst i i i ∂T(cid:90)ist φ˜(ki)FRcS(cid:0)v¯−,v¯+(cid:1)·(cid:126)nidSdt−T(cid:90)ist ∇φ˜(ki)·Fc(v¯ni+1)dxdt+(cid:88)j∈SiΓ(cid:90)sjt φ˜(ki)σnj+1·(cid:126)nijdSdt−T(cid:90)is,tj ∇φ˜k(i)·σnj+1dxdt=0, (cid:90) ψ˜(kj)(x,tn+1)σnj+1dx=−ν (cid:90) ψ˜(kj)∇v¯n(cid:96)(+j1)dxdt+ (cid:90) ψ˜(kj)∇v¯nr(+j1) dxdt+(cid:90) ψ˜(kj)(cid:16)v¯nr(+j1) −v¯n(cid:96)(+j1)(cid:17)⊗(cid:126)njdSdt. (40) Hst Tst Tst Γst j (cid:96)(j),j r(j),j j 10