Contents lists available atScienceDirect
Journal
of
Computational
Physics
www.elsevier.com/locate/jcp
A
simple
diffuse
interface
approach
on
adaptive
Cartesian
grids
for
the
linear
elastic
wave
equations
with
complex
topography
Maurizio Tavelli
a,
Michael Dumbser
a,
∗
,
Dominic
Etienne Charrier
b,
Leonhard Rannabauer
c,
Tobias Weinzierl
b,
Michael Bader
caDepartmentofCivil,EnvironmentalandMechanicalEngineering,UniversityofTrento,ViaMesiano77,I-38123Trento,Italy bDepartmentofComputerScience,UniversityofDurham,LowerMountjoy,SouthRoad,DurhamDH13LE,UnitedKingdom cDepartmentofInformatics,TechnicalUniversityMunich(TUM),Boltzmannstr.3,D-85748Garching,Germany
a
r
t
i
c
l
e
i
n
f
o
a
b
s
t
r
a
c
t
Articlehistory: Received25April2018
Receivedinrevisedform6February2019 Accepted7February2019
Availableonline18February2019 Keywords:
Diffuseinterfacemethod(DIM) Complexgeometries Highorderschemes
DiscontinuousGalerkinschemes Adaptivemeshrefinement(AMR) Linearelasticityequationsforseismicwave propagation
Inmost classicalapproaches of computational geophysicsfor seismic wavepropagation problems,complexsurfacetopographyiseitheraccountedforbyboundary-fitted unstruc-turedmeshes,or, where possible,by mappingthecomplex computational domainfrom physicalspacetoatopologicallysimple domaininareference coordinatesystem. How-ever, all these conventional approaches face problems if the geometry of the problem becomessufficientlycomplex.Theyeitherneedameshgeneratortocreateunstructured boundary-fittedgrids, whichcan become quite difficultand may require alot of man-ual user interactionsin orderto obtain ahigh quality mesh, orthey need the explicit computationofanappropriatemapping functionfromphysical toreference coordinates. ForsufficientlycomplexgeometriessuchmappingsmayeithernotexistortheirJacobian couldbecomeclosetosingular.Furthermore,inbothconventionalapproacheslowquality gridswillalwaysleadtoverysmalltimestepsduetothe Courant-Friedrichs-Lewy(CFL) conditionforexplicitschemes.Inthispaper, weproposeacompletelydifferentstrategy thatfollowstheideasofthesuccessfulfamilyofhighresolutionshock-capturingschemes, wherediscontinuities canactually beresolvedanywhereonthe grid, withouthaving to fitthemexactly.Weaddresstheproblemofgeometricallycomplexfreesurfaceboundary conditionsforseismicwavepropagationproblemswithanoveldiffuseinterfacemethod (DIM)on adaptive Cartesianmeshes (AMR)that consists inthe introduction ofa char-acteristicfunction0≤
α
≤1 which identifiesthe locationofthe solidmediumand the surroundingair(or vacuum) andthus implicitlydefinesthe locationofthe freesurface boundary.Physically,α
representsthevolumefraction ofthesolidmediumpresent ina controlvolume.Ournewapproachcompletelyavoids theproblemofmeshgeneration,since allthatisneededforthedefinition ofthecomplexsurfacetopographyistoset ascalar colorfunctiontounityinsidethe regionscoveredbythe solidandto zerooutside. The governingequations are derived from ideastypically used inthemathematical descrip-tionofcompressiblemultiphaseflows.AnanalysisoftheeigenvaluesofthePDEsystem showsthatthecomplexityofthegeometryhasnoinfluenceontheadmissibletimestep sizeduetotheCFLcondition.Themodelreducestotheclassicallinearelasticityequations insidethesolidmediumwherethegradientsofα
arezero,whileinthediffuseinterface zoneatthefreesurfaceboundarythegoverningPDEsystembecomesnonlinear.We can provethatthesolutionoftheRiemannproblemwitharbitrarydataandajumpinα
from*
Correspondingauthor.E-mailaddresses:m.tavelli@unitn.it(M. Tavelli),michael.dumbser@unitn.it(M. Dumbser),dominic.e.charrier@durham.ac.uk(D.E. Charrier), leo.rannabauer@tum.de(L. Rannabauer),tobias.weinzierl@durham.ac.uk(T. Weinzierl),bader@in.tum.de(M. Bader).
https://doi.org/10.1016/j.jcp.2019.02.004
0021-9991/©2019TheAuthor(s).PublishedbyElsevierInc.ThisisanopenaccessarticleundertheCCBY-NC-NDlicense (http://creativecommons.org/licenses/by-nc-nd/4.0/).
unitytozeroyieldsaGodunov-stateattheinterfacethatsatisfiesthefree-surface bound-aryconditionexactly,i.e. thenormalstresscomponentsvanish.Inthegeneralcaseofan interfacethatis notalignedwiththe gridandwhichisnot infinitelythin,asystematic studyonthedistributionofthevolumefractionfunctioninsidetheinterfaceandthe sen-sitivitywith respecttothethicknessofthe diffuseinterfacelayer hasbeencarried out. Inordertoreducenumericaldissipation,weusehighorderdiscontinuousGalerkin(DG) finiteelementschemesonadaptiveAMRgridstogetherwithasecondorderaccuratehigh resolutionshockcapturingsubcellfinitevolume(FV)limiterinthediffuseinterfaceregion. WefurthermoreemployalittledissipativeHLLEMRiemannsolver,whichisabletoresolve thesteadycontactdiscontinuityassociatedwiththevolumefractionfunctionandthe spa-tiallyvariablematerialparametersexactly.Whilethemethodislocallyhighorderaccurate intheregionswithoutlimiter,theglobalorderofaccuracyoftheschemeisatmosttwoif thelimiterisactivated.Itislocallyoforderoneinsidethediffuseinterfaceregion,which istypicalforshock-capturingschemes atshocks andcontact discontinuities.We showa large setofcomputationalresultsintwo andthreespacedimensionsinvolvingcomplex geometrieswherethe physicalinterfaceisnot alignedwiththegridorwhereitiseven moving.Foralltestcasesweprovideaquantitativecomparisonwithclassicalapproaches basedonboundary-fittedunstructuredmeshes.
©2019TheAuthor(s).PublishedbyElsevierInc.Thisisanopenaccessarticleunderthe CCBY-NC-NDlicense(http://creativecommons.org/licenses/by-nc-nd/4.0/).
1. Introduction
The numerical solution of linear elastic wave propagation is still a challenging task, especially when complex three-dimensionalgeometriesareinvolved.Inthepast,alargenumberofnumericalschemeshasbeenproposedforthesimulation ofseismicwavepropagation.Madariaga[1] andVirieux[2,3] introducedfinitedifferenceschemesforthesimulationof pres-sure(P)andshear(SVandSH)wavepropagation.Theseschemeswerethenextendedtohigherorder,see[4],threespace dimensions[5,6] andtoanisotropicmaterial[7,8].Forfinitedifference-likemethodsonunstructuredmesheswerefertothe workofMagnieretal. [9] andKäser&Igel[10,11].Therearealsoseveralapplicationsinthecontextoffinitevolume(FV) schemes[12–17],which,however,werealllimitedtosecond orderofaccuracyinspaceandtime. Thefirstarbitraryhigh orderADERfinite volumescheme forseismicwave propagationwas introduced in[18]. Forrealapplicationsit iscrucial thatanumericalschemeisabletoproperlycapturecomplexsignalsoverlongdistancesandtimes.Incontrasttoclassical loworderschemes,highordermethodsinspaceandtime areabletobetterreproducethetimeevolutionofthesolution. Aquantitativeaccuracyanalysisofhighordernumericalschemesforlinearelasticity,basedonthemisfitcriteriadeveloped in[19,20], canbe found in[21,22]. Spectral finiteelement methods [23] were successfullyappliedto linearelastic wave propagationina well-knownseriesofpaperofKomatitschandcollaborators[24–28].ForChebyshevspectralmethodsfor wavepropagationwerefertotheworkofTessmeretal. [29,8] andIgel[30].Foralternativedevelopmentsintheframework ofstabilizedcontinuous finiteelements appliedto elasticandacousticwave propagationwerefer to[31–33].Apartfrom wave propagationinthemedium, alsotheproperrepresentationofcomplexsurfacetopographyisachallenging task.For thispurpose,severalhighordernumericalschemesonunstructuredmesheswereintroducedinthepast.Aseriesofexplicit high order discontinuous Galerkin (DG) schemes for elastic wave propagation on unstructured mesheswas proposed in [34–39],whiletheconceptofspace-timediscontinuousGalerkinschemes,originallyintroducedandanalyzedin[40–46] for computationalfluiddynamics(CFD),waslateralsoextendedtolinearelasticityin[47–49].Thespace-timeDGmethodused in[49] isbasedon the novelconcept of
staggered discontinuous
Galerkin finiteelement schemes,which was introduced forCFD problemsin [50–56]. Inany case, all previous methods requirea boundary-fitted mesh that properly represents thegeometryofthephysicalproblemtobesolved.Thegenerationofthismeshisingeneralahighly non-trivialtaskand usuallyrequirestheuseofexternalmeshgenerationtools.Moreover,themeshgenerationprocessinhighlycomplex geom-etrycanleadtoverysmallelements withbadaspectratio,so-calledsliver elements [
57–59].Thiswellknownproblemcan oftenbeavoided,butnotalways,seee.g.[60,61].Forexplicittimediscretization,sliverelementscanonlybetreatedatthe aidoflocal time stepping (LTS),
see,forexample,[36,62–64],butcurrentlyonlyveryfewschemesusedinproductioncodes employedincomputational seismologysupporttime-accuratelocaltimestepping. Alternatively,implicitschemes like[49] requiretheintroductionofaproperpreconditionerinordertolimitthenumberofiterationsneededtosolvetheassociated linearalgebraicsystem.The key idea of this paper is therefore to completely avoid the mesh generation problem associated with classical
approachesusedincomputational seismology.Thisisachievedby extendingthelinearelasticwave equationsviaa char-acteristic (color)function
α
,whichis nothingelse than thevolume fractionofthe solid medium,andwhich determines ifapoint x islocatedinside thesolid material(α
(
x)
=
1)oroutside(α
(
x)
=
0). Inthiswaythe scalarparameterα
sim-plydeterminesthephysicalboundarythroughadiffuseinterfacezone,insteadofrequiringaboundary-fittedstructuredor unstructuredmesh.Withthisnewapproach,evenverycomplexgeometriescanbe easilyrepresentedonregular adaptive Cartesianmeshes,i.e.viatheuseofadaptivemeshrefinement(AMR).Furthermore,theintroductionofthenewparametertheCFLcondition.Tobemoreprecise:theadmissibletime stepsizeofthenewapproachpresentedinthispaperdepends ofcourseon thechosen
mesh spacing of
theregular AMRgridandonthesignal speeds in
thePDEsystem,butitdoesnot
explicitlydependonthemesh quality and
thegeometric complexity of
thecomputationaldomain,asitisthecaseformany otherapproachesincomputationalseismology.Inthecontextoffinitedifferenceschemesforseismicwavepropagation,immersedboundarymethodsforthetreatment ofcomplexfree surface topologiescanbefound, forexample,in[65–68] andreferencestherein.However, theunderlying mathematical models usedthereare
different from
the one proposed inthispaper,which isbasedon adiffuseinterface approachthatisusedforthedescriptionofcompressiblemulti-phaseflowsincomputationalfluiddynamics.Our newmethod isinspired by the work concerning the modeling andsimulation ofcompressible multiphaseflows, see [69] and[70–73]. Itcan alsobe interpretedasa specialcaseofthemoregeneralsymmetrichyperbolic and thermo-dynamically compatiblemodel of nonlinear hyperelasticity of Godunov & Romenski andcollaborators [74–80]. A diffuse interface approach,similar totheoneused inthispaper,hasalreadybeensuccessfullyappliedto nonlinearcompressible fluid-structureinteractionproblemsinaseriesofpapers[81–83],buttheemployednumericalmethodswereonlyloworder accurateinspaceandtimeandthereforenotsuitableforseismicwavepropagationproblems.Otherapplicationsofdiffuse interface methods forcompressiblemulti-phaseflows canbe foundin[84–86], but,to thebestofour knowledge,thisis thefirsttimethatadiffuseinterfaceapproachisderivedandvalidatedforlinearseismicwavepropagationincomplex ge-ometries.Withinthepresentpaper,weusehighorderaccurateADER-DGschemesonCartesianmesheswithadaptivemesh refinement (AMR). Thenumericalmethod hasalreadysuccessfullybeenapplied toother hyperbolic PDEsystems [87,78]. The use ofadaptivemesh refinementallows to increase theresolution locallywhereneeded, especially closeto thefree surfaceoratinternalmaterialboundaries.Toavoidspuriousoscillationsandtoenforcenonlinearstability,weuseasimple but very robust
a posteriori subcell
finite volume(FV) limiter [88]. Here,asecond order total
variation diminishing(TVD) finite volumeschemeisadoptedinthe limitedDGcells.Inorderto maintain accuracy,thesubgridofthelimiter isbya factor of2N+
1 times finercompared tothegrid ofan unlimitedDG schemewithpolynomial approximationdegree N. Theideaofusingana posteriori approach
tolimithighorderschemeswasfirstproposedbyClain,DiotandLoubèrewithin the so-calledMulti-dimensionalOptimal OrderDetection (MOOD)paradigm in thecontext offinitevolume schemes, see [89,90] formoredetails.Finally,inournumericalschemewemakeuseoftheHLLEMRiemannsolverintroducedin[91,92], whichisabletoresolvethesteadycontactdiscontinuities associatedwiththespatiallyvariablematerialparametersλ
andμ
(theLaméconstants),themassdensityρ
andthevolumefractionα
.Thenumericalresultspresentedlaterinthispaper show that theproposed methodologyseems tobe avalidalternativeto existingapproachesincomputational seismology that arebasedonboundary-fittedstructuredorunstructuredmeshes.Atthispointwewouldliketostressthattheuseof a secondordershockcapturingTVDfinitevolumeschemeinsidethediffuseinterfaceregionatthefreesurface boundary limits theglobal order
of accuracy ofthe schemeto atmosttwo. Numerical experimentsfurther show that the method is locallyfirstorder accurateinside thediffuseinterface region,whichis well-knownfromshockcapturingfinite volume schemes inCFD,whichalsoreducetofirstorderofaccuracyatshocksandcontactdiscontinuitiesthatare notexactly re-solvedonthegrid,see[93].Nevertheless,theunlimited ADER-DG
schemethatisusedinsidethesolidmediumandfarfrom thefreesurfaceboundaryislocally high
orderaccurateandthusbeneficialconcerningphaseandamplitudeerrorsforwave propagationoverlongdistancesandtimesinside the
solidmedium.Notethatthemanifolddescribingthefreesurfaceisof onedimensionlessthanthecomputationaldomain,hencemostcellscanactuallyusethehighorderaccurateunlimited DG
schemeandonlyveryfewcellsrequiretheuseofthesecondorderaccuratesubcellfinitevolumelimiter.Inordertoreduce thenumericalerrorsinthediffuseinterfaceregionasfaraspossible,weproposetouseadaptivemeshrefinement(AMR) withtime-accuratelocaltimestepping(LTS)combinedwithasubcellFVlimiter,wherethesubgridisbyafactorof2N+
1 timesfinerthanthegridoftheunlimitedADER-DGschemewithpolynomialapproximationdegreeN.
However,wewould liketoemphasizethatthemathematicalmodelproposedinthispaperisnot strictly
linked tothenumericalschemesthat areusedinthispaperforitssolution(ADER-DGwithAMRcoupledwithsubcellfinitevolumelimiter).Anystandardfinite difference schemeinsidethe solidtogether withanonlinearENO/WENOscheme atthe freesurface where∇
α
=
0 could havebeenappliedequallywell.Therestofthepaperisorganized asfollows:inSection2weintroducethegoverningPDEofthenewdiffuseinterface approachforlinearelasticity.Wealsoshowthecompatibilityofourmodelwiththefreesurfaceboundaryconditioninthe casewhere
α
jumpsfrom1 to0.InSection 3we brieflysummarizethehighorderADER-DGschemesusedinthispaper. InSection4weshownumericalresultsforalargesetoftestproblemsintwoandthreespacedimensions,alsoincludinga realistic3DscenariowithcomplexgeometrygivenbyrealDTMdata.Finally,inSection5wegivesomeconcludingremarks andan outlookonfuturework, whichwillconcernnonlinear large-strainelasto-plasticityanddynamic ruptureprocesses inmovingmediabasedonthetheoryofnonlinearhyperelasticityofGodunovandRomenski[74,94,77].2. Mathematicalmodel
Theequationsoflinearelasticity[95] canbewrittenas
∂
∂
tσ
xx− (λ +
2μ
)
∂
∂
xu− λ
∂
∂
yv− λ
∂
∂
zw=
Sxx,
∂
∂
tσ
y y− λ
∂
∂
xu− (λ +
2μ
)
∂
∂
yv− λ
∂
∂
zw=
Sy y,
∂
∂
tσ
zz− λ
∂
∂
xu− λ
∂
∂
yv− (λ +
2μ
)
∂
∂
zw=
Szz,
∂
∂
tσ
xy−
μ
∂
∂
xv+
∂
∂
yu=
Sxy,
∂
∂
tσ
yz−
μ
∂
∂
zv+
∂
∂
yw=
Syz,
∂
∂
tσ
xz−
μ
∂
∂
zu+
∂
∂
xw=
Sxz,
∂
∂
t(
ρ
u)
−
∂
∂
xσ
xx−
∂
∂
yσ
xy−
∂
∂
zσ
xz=
ρ
Su,
∂
∂
t(
ρ
v)
−
∂
∂
xσ
xy−
∂
∂
yσ
y y−
∂
∂
zσ
yz=
ρ
Sv,
∂
∂
t(
ρ
w)
−
∂
∂
xσ
xz−
∂
∂
yσ
yz−
∂
∂
zσ
zz=
ρ
Sw,
(1)where
λ
andμ
arethesocalledLaméconstantsandρ
isthemassdensity.Inmorecompactformtheabovesystemreads∂
σ
∂
t−
E(λ,
μ
)
· ∇
v=
Sσ,
(2)∂
ρ
v∂
t− ∇ ·
σ
=
ρ
Sv,
(3)where v
= (
u,
v
,
w
)
isthevelocityfield,ρ
isthematerialdensity, S ρ and S σ arevolumeorpointsources,σ
isthe sym-metricstresstensor,andE(λ,
μ
)
isthestiffnesstensorthatconnectsthestraintensorkl tothestresstensor
σ
accordingtotheHookelaw
σ
=
E.Thestresstensor
σ
isgivenbyσ
=
⎛
⎝
σ
σ
xxyxσ
σ
xyy yσ
σ
xzyzσ
zxσ
zyσ
zz⎞
⎠
(4)with the symmetry
σ
i j=
σ
ji. The normal stress components areσ
xx,
σ
y y andσ
zz, while the shear stress isrepre-sented by
σ
xy,
σ
yz andσ
xz. The stress tensorσ
can thus be written in terms of its six independent components(
σ
xx,
σ
y y,
σ
zz,
σ
xy,
σ
yz,
σ
xz)
. In the following we propose a new model that follows the ideas used in the simulation ofcompressiblemultiphaseflows[69–71,73].InordertoderivethemodelwestartfromaBaer-Nunziato-typesystemforthe descriptionofcompressiblemulti-phaseflows,whereforthesolidphase(index
s)
thepressuretermhasbeenappropriately replacedbythestresstensorσ
s,andwheretheusualpressureandvelocityrelaxationsourcetermshavebeendropped:∂
∂
t(
α
sρ
s)
+ ∇ · (
α
sρ
svs)
=
0,
∂
∂
t(
α
sρ
svs)
+ ∇ · (
α
sρ
svs⊗
vs+
α
sσ
s)
−
σ
I∇
α
s=
α
sρ
sSv,s,
∂
∂
t(
α
sρ
sEs)
+ ∇ · (
α
sρ
sEsvs+
α
sσ
svs)
−
σ
I∇
α
s·
vI=
α
sρ
sSv,s·
vs,
∂
∂
tα
gρ
g+ ∇ ·
α
gρ
gvg=
0,
∂
∂
tα
gρ
gvg+ ∇ ·
α
gρ
gvg⊗
vg+
α
gσ
g−
σ
g∇
α
g=
α
gρ
gSv,g,
∂
∂
tα
gρ
gEg+ ∇ ·
α
gρ
gEgvg+
α
gσ
gvg−
σ
I∇
α
g·
vI=
α
gρ
gSv,g·
vg,
∂
∂
tα
s+
vI∇
α
s=
0.
(5)Hereindex
s refers
tothesolidphaseandindexg refers
tothegasphasesurroundingthesolid;ρ
kisthemassdensityandEkisthespecifictotalenergyofphase
k,
vk isthephasevelocity,vI istheso-calledinterfacevelocityandσ
I isthestresstensorattheinterface,whichisageneralizationoftheinterfacepressureusedinstandardBNmodels.Wenowmakethe followingsimplifyingassumptions:
(i) Theinterfacebetweenthesolidandthegasismovingonlyatanegligiblespeed,hencewecanassumevI
=
0.(ii) Compared to the original Baer-Nunziato model [69,96,86], all pressure and velocity relaxation source terms are ne-glected.
(iii) Themassdensityofthegasphaseismuchsmallerthantheoneofthesolidphase(
ρ
gρ
s),hencethetimeevolutionofthegas phaseis notrelevantforour purposes.Therefore,all evolutionequationsrelatedto thegasphase canbe neglectedinthefollowing,similartotheapproachusedin[97–99] inthecontextofnon-hydrostaticfreesurfaceflow simulationsbasedonadiffuseinterface approach.Toeasenotation,theremaining index
s for
thesolidphasecan be dropped.(iv) Weassumethedensity
ρ
softhesolidphasetobeconstantintime,sotherelatedmassconservationequationcanbeneglected.
(v) Thestress tensorofthesolid canbedirectlycalculatedviaHooke’s law(2),soitisnotnecessarytoevolve thetotal energyconservationlawforthesolid.
(vi) Thenonlinear convectiveterm
α
sρ
svs⊗
vs,whichisquadraticinthesolid velocity, canbeneglected, sincethesolidvelocityisassumedtobesmallinthelinearelasticitylimit.
(vii) Last but not least, the free surface boundarycondition atthe interface betweensolid andsurrounding gasleads to
σ
I· ∇
α
s=
0.As aresultofthesesimplifyingassumptions,the
reduced governing
PDEsystemofthenewdiffuseinterface approachfor linearelasticityincomplexgeometryreads:∂
σ
∂
t−
E(λ,
μ
)
· ∇
v=
Sσ,
(6)∂
αρ
v∂
t− ∇ · (
ασ
)
=
αρ
Sv,
(7)∂
α
∂
t=
0.
(8)Since
∂
tρ
=
0,thepreviousequationsarethenrewrittenas∂
σ
∂
t−
E(λ,
μ
)
·
1α
∇(
α
v)
+
1α
E(λ,
μ
)
·
v⊗ ∇
α
=
Sσ,
(9)∂
α
v∂
t−
α
ρ
∇ ·
σ
−
1ρ
σ
· ∇
α
=
Sv,
(10)∂
α
∂
t=
0.
(11)Furthermorethefollowingequationsforthematerialparametersareaddedtothesystem:
∂λ
∂
t=
0,
∂
μ
∂
t=
0,
∂
ρ
∂
t=
0.
(12)The material parameters
λ,
μ
andρ
are assumed to be constant in time butnot in space, i.e.λ
= λ(
x),
μ
=
μ
(
x)
andρ
=
ρ
(
x)
,forwhichwewilluseahighorderpolynomial representationasfortheother variablesofthePDEsystem.The samediffuseinterfacemodel(6)-(8) canalsobeobtainedbycombiningthenonlinearhyperelasticityequationsofGodunov andRomenski[74,94,100] withthecompressiblemulti-phasemodelofRomenskietal.[75,101],assuming linearmaterial behaviorandneglectingnonlinearconvectiveterms.System (9)-(12) isthenrewritteninthefollowingform:∂
Q∂
t+
B1(
Q)
∂
Q∂
x+
B2(
Q)
∂
Q∂
y+
B3(
Q)
∂
Q∂
z=
S(
x,
t),
(13)wherethethreematricesB1,B2andB3 arespecifiedinEqs. (15)-(17).ThevectorQ isgivenby
Q
=
σ
xx,
σ
y y,
σ
zz,
σ
xy,
σ
yz,
σ
xz,
α
u,
α
v,
α
w, λ,
μ
,
ρ
,
α
,
(14)B1
=
⎛
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎝
0 0 0 0 0 0−
α1(λ
+
2μ
)
0 0 0 0 0 α1(λ
+
2μ
)
u 0 0 0 0 0 0−
α1λ
0 0 0 0 0 α1λ
u 0 0 0 0 0 0−
α1λ
0 0 0 0 0 α1λ
u 0 0 0 0 0 0 0−
α1μ
0 0 0 0 α1μ
v 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0−
α1μ
0 0 0 α1μ
w−
α ρ 0 0 0 0 0 0 0 0 0 0 0−
ρ1σ
xx 0 0 0−
αρ 0 0 0 0 0 0 0 0−
ρ1σ
xy 0 0 0 0 0−
αρ 0 0 0 0 0 0−
ρ1σ
xz 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0⎞
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎠
,
(15) B2=
⎛
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎝
0 0 0 0 0 0 0−
α1λ
0 0 0 0 α1λ
v 0 0 0 0 0 0 0−
α1(λ
+
2μ
)
0 0 0 0 α1(λ
+
2μ
)
v 0 0 0 0 0 0 0−
α1λ
0 0 0 0 α1λ
v 0 0 0 0 0 0−
α1μ
0 0 0 0 0 α1μ
u 0 0 0 0 0 0 0 0−
α1μ
0 0 0 α1μ
w 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0−
αρ 0 0 0 0 0 0 0 0−
ρ1σ
xy 0−
αρ 0 0 0 0 0 0 0 0 0 0−
1ρσ
y y 0 0 0 0−
αρ 0 0 0 0 0 0 0−
ρ1σ
yz 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0⎞
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎠
,
(16) B3=
⎛
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎝
0 0 0 0 0 0 0 0−
α1λ
0 0 0 α1λ
w 0 0 0 0 0 0 0 0−
α1λ
0 0 0 α1λ
w 0 0 0 0 0 0 0 0−
α1(λ
+
2μ
)
0 0 0 α1(λ
+
2μ
)
w 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0−
α1μ
0 0 0 0 α1μ
v 0 0 0 0 0 0−
α1μ
0 0 0 0 0 α1μ
u 0 0 0 0 0−
αρ 0 0 0 0 0 0−
ρ1σ
xz 0 0 0 0−
αρ 0 0 0 0 0 0 0−
ρ1σ
yz 0 0−
αρ 0 0 0 0 0 0 0 0 0−
ρ1σ
zz 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0⎞
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎠
.
(17)TheeigenvaluesassociatedwiththematrixB1 are
λ
1= −
cp,
λ
2,3= −
cs,
λ
4,5,6,7,8,9,10=
0,
λ
11,12= +
cs,
λ
13= +
cp,
(18) where cp=
λ
+
2μ
ρ
and cs=
μ
ρ
(19)are the p
−
ands−
wave velocities,respectively. The matrixof righteigenvectors ofthe matrix B1 asdefined in (15) isR
=
⎛
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎝
ρ
c2p 0 0 0 0 0 0 0 0−
σ
xx 0 0ρ
c2pρ
(
c2p−
2c2s)
0 0 1 0 0 0 0 0 0 0 0ρ
(
c2p−
2c2s)
ρ
(
c2p−
2c2s)
0 0 0 1 0 0 0 0 0 0 0ρ
(
c2p−
2c2s)
0ρ
c2s 0 0 0 0 0 0 0−
σ
xy 0ρ
c2s 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0ρ
c2s 0 0 0 0 0 0−
σ
xzρ
c2s 0 0 cp 0 0 0 0 0 0 0 0α
u 0 0−
cp 0 cs 0 0 0 0 0 0 0α
v 0−
cs 0 0 0 cs 0 0 0 0 0 0α
w−
cs 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0α
0 0 0⎞
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎠
.
(20)The expressions for the eigenvalues andeigenvectors of B2 and B3 are very similar andcan be obtainedfrom those of B1,sincethePDE systemisrotationallyinvariant.Forthisreason,we donot givetheir explicitexpressionshere. Wenow
wanttoshow thattheproposed modelsatisfiesthefree surfaceboundarycondition
σ
·
n=
0 exactlywhenconsidering a Riemannproblemthatincludesajumpofα
fromα
L=
1 toα
R=
0.Forthis,considertheleftandrightstateofaRiemannprobleminthe
x-direction
givenbyQL
= (
σ
xxL,
σ
y yL,
σ
zzL,
σ
xyL,
σ
yzL,
σ
xzL,
uL,
vL,
wL, λ,
μ
,
ρ
,
1),
(21)QR
= (
σ
xxR,
σ
y yR,
σ
zzR,
σ
xyR,
σ
yzR,
σ
xzR,
0,
0,
0, λ,
μ
,
ρ
,
0).
(22)Byusingasimplestraightlinesegmentpath
ψ (
s)
=
QL+
s(
QR−
QL) ,
(23)wecandefineageneralizedRoe-averagedmatrixB
˜
1inx direction
accordingto[102–104] asfollows:˜
B1=
1 0 B1(ψ (
s))
ds.
(24)TheexactsolutionofthelinearizedRiemannproblembasedontheRoe-averagedmatrixB
˜
1= ˜
R˜˜
R−1 aboveandthesimi-laritycoordinate
ξ
=
x/
t readsQRP
(ξ )
=
1 2R˜
I−
sign(ξ
I− ˜)
R˜
−1·
QL+
1 2R˜
I+
sign(ξ
I− ˜)
R˜
−1·
QR,
(25)with I beingthe identity matrix.From QRP
(ξ )
we can obtainthe following Godunov state QGod=
QRP(
0)
at theinterface(
ξ
=
0) QGod=
0,
σ
L xxc2p+
2σ
xxLc2s+
σ
y yc2p c2p,
σ
L xxc2p+
2σ
xxLc2s+
σ
zzc2p c2p,
0,
σ
yzL,
0,
cpρ
uL−
σ
xxL cpρ
,
csρ
v L−
σ
L xy csρ
,
csρ
w L−
σ
L xz csρ
, λ,
μ
,
ρ
,
1,
fromwhichitisclearthatallthecomponentsofthenormalstressin
x-direction
(σ
xx,
σ
xy andσ
xz)arezero,whichmeansthat thefreesurface boundarycondition
σ
·
n=
0 isindeedrespected bymerelyimposing ajumpinthevolumefraction functionfromα
=
1 toα
=
0.Asonecannote,themodel(9)-(11) involvesdivisionsby
α
thatcanbeasourceofinstabilitiesattheinterface,sincethe colorfunctionα
isideallysettozero,oratleastclosetozero,outsidethesolidmedium.Inordertoaddressthisproblem, we introduceasimpletransformationthat avoidsthedivisionsbyzeros.Inparticular,wesubstituteall multiplicationsbyα
−1=
1/
α
,withα
−1∼
=
α
α
2+
(
α
)
,
(26)where
=
(
α
)
hastosatisfy(
1)
=
0 and(
0)
=
0
>
0 inordertobe consistentwiththelinearelasticityequations.Inourcasewe takeasimplelinearfunction
=
0
(
1−
α
)
with0
=
10−3.Theintroductionofthisnewparameterwiththismethodismandatorytoobtainastablesolution.Theneweigenvaluesare
˜λ =
fλ
,where f=
αα2+0(1−α) thatfor
α
∈ [
0,
1]
satisfies f∈ [
0,
1]
and f=
1 forα
=
1.Assoonasweuseanon-trivialgeometrywechooseadiffuseinterfaceof
finite width I
D forthetransitionbetweenthesolid medium
α
=
1 and the surroundinggas / vacuum(α
=
0). For arelatively large width ID ofthe diffuseinterface,therearesomequestionsthatarisenaturallyconcerningthedistributionofthecharacteristicfunction
α
insidethediffuse interface andtheresultingeffectivepositionofthefree surfaceboundary. Ingeneral,it isimportanttosetupthe diffuse interfaceshapesuchthat∇
α
isorientedasthenormalvectortothephysicalsurface,i.e.∇
α
≈
n.Asimplewaytodothis istorepresentthetransitionregionbyapiecewisepolynomial.Letr
=
r(
x)
bethesigneddistancebetweentherealphysical interfacelocationandagenericpointx under consideration.Wethendefine theshape ofthediffuseinterface asfunction ofafiniteinterfacethicknessI
D≥
0,ashiftingparameterη
andtheauxiliaryfunction:ξ(
r)
=
⎧
⎨
⎩
1 if r>
(
1+
η
)
ID,
0 if r<
−(
1−
η
)
ID,
r+(1−η)ID 2ID if r∈ [−(
1−
η
)
ID, (
1+
η
)
ID].
(27) Wefinallydefinethesolidvolumefractionasα
(
r)
= (
1− ξ(
r))
p,
(28)where p
>
0 isan exponentthat determines theshape ofthe diffuseinterface. The widthofthe interface ID should berelatedtothelocalsize
h of
thecomputationalmesh,i.e.onewouldtypicallychooseI
D∼
h. InordertoreduceI
D asmuchaspossible,wewillmakeuseofadaptivemeshrefinement (AMR)incombinationwitha subcellfinitevolumelimiter, as discussedinthenextsection.
3. Numericalscheme
ThenumericalmethodthatweuseinordertosolvethePDEsystemintroduced inthepreviousSection2isan explicit ADER-DGschemeofarbitraryhighorderofaccuracyinspaceandtimeonadaptiveCartesiangrids (AMR).Thenumerical methodwas presentedfordifferentPDE systemsin[88,87,78],henceinthe followingweonlygive abriefsummary.The PDEsystem(9)-(11) canbewrittenincompactmatrix-vectornotationas
∂
Q∂
t+
B(
Q)
· ∇
Q=
S(
x,
t),
(29)where Q is the state vector, B
(
Q)
· ∇
Q is a non conservative product (see [105,102,103])and S(
x,
t)
is a known source term.Inregions whereα
=
1 andthus∇
α
=
0,thePDEsystem(29) reducestotheclassicallinearelasticwaveequations (1), while for∇
α
=
0 the system becomes locally nonlinear and therefore requires a very robust numerical scheme as well ashighresolution tobeproperly solved.Withinthispaperweusethesimple andvery robustsubcellfinite-volume limiterapproachincombinationwithadaptivemeshrefinement(AMR).Adetaileddescriptionofthelimitercan befound in [88,87]. As suggested in [88], we employ Ns= (
2N+
1)
d subgrid cells for the finite volume limiter, where d is thenumberofspacedimensionsoftheproblemand
N is
thepolynomialapproximationdegreeusedinthehighorderADER-DG scheme.Notethattheuseofsuchafinesubgridwithinthesubcellfinitevolumelimiterdoesnot reduce
thetimestepof the overall scheme, since finitevolume schemes are stableup to CFL=
1, while DGschemes require CFL<
1/(
2N+
1)
. Thed-dimensional
computationaldomainisdiscretized withanadaptive Cartesiangrid composedofCartesian control volumes
T
iinspaceas=
Ne i=1 Ti,
(30)where Ne isthe total number ofelements. Since we are interested ina high orderscheme, we first define a piecewise
polynomialnodalbasis
{φ
k}
k=1...(N+1)d asthesetofLagrange polynomialspassing throughtheGauss-Legendrequadraturepointsona referenceunitelement
T
re f foragivenpolynomial degree N≥
0 anddimensiond.
AweakformulationofthePDEsystem isobtainedaftermultiplying Eq.(29) by atest function
φ
k fork
=
1. . . (
N+
1)
d andthen integratingoveraspace-timecontrolvolume Ti
× [
tn,
tn+1]
: tn+1 tn Tiφ
k∂
Q∂
t+
B(
Q)
· ∇
Q dx dt=
tn+1 tn Tiφ
kS(
x,
t)
dx dt.
(31)Werestrictthediscretesolutiontothespaceofpiecewisepolynomialsofdegree
N,
i.e. thenumericalsolutionuhiswritteninsideeachelement
T
i intermsofthepolynomialbasisasuh
(
x,
tn)
Ti=
(N+1)dk=1
forx
∈
Tiandi
=
1. . .
Ne.Thevectorofdegreesoffreedomofuh(
x,
tn)
isdenotedbyuˆ
ni.ThroughoutthepaperweusetheEinstein summationconventionoverrepeatedindices.Usingthedefinition(32) in theweakformulationgivenby Eq.(31) weobtain(seealso[106,107])
⎛
⎜
⎝
Tiφ
kφ
ldx⎞
⎟
⎠
uˆ
nl,+i1− ˆ
unl,i+
tn+1 tn ∂Tiφ
kD
−(
qh−,
q+h)
·
n dS dt+
tn+1 tn Ti◦φ
kB(
qh)
· ∇
qhdx dt=
tn+1 tn Tiφ
kS(
x,
t)
dx dt,
(33) wherewehaveintroducedthejumpcontributionD
−(
q−h,
qh+)
·
n ontheelementboundariesandthespace-timepredictor solution qh(
x,
t)
.Moredetails concerningthecomputation ofqh(
x,
t)
willbe reportedlater.Fortheapproximation ofthejumpterm
D
− weuseapath-conservativeschemeasintroducedbyParésin[103] andCastroetal. in[102].Weintroduce aLipschitzcontinuouspathfunctionψ(
q−h,
q+h,
s)
definedfors
∈ [
0,
1]
suchthatψ(
qh−,
q+h,
0)
=
q−h andψ(
q−h,
q+h,
1)
=
q+h, where qh− denotes the boundary-extrapolated state fromwithin the element Ti and q+h theboundary-extrapolated statefrom the neighbor element. The simplestpossible choice for
ψ
, whichwe usein thispaper, is thelinear segment path betweenthetwostatesq−h andq+h:ψ (
q−h,
q+h,
s)
=
qh−+
sqh+−
q−h.
(34) Following[105,103,102] wenowdefine thejumpcontributionD
−(
q−h,
q+h)
·
n sothatitsatisfiesthegeneralized Rankine-HugoniotconditionsD
−(
q−h,
q+h)
·
n+
D
+(
q−h,
q+h)
·
n=
1 0 B(ψ (
q−h,
q+h,
s))
·
n∂ψ
∂
sds.
(35)Thepreviouspathintegralcansimplybeevaluated
numerically using
asufficientnumberofGaussianquadraturepoints.As RiemannsolverweusethenewHLLEM-typeRiemannsolverfornon-conservativesystemsrecentlydescribedin[92],since wewanttoexactlypreservethecontactdiscontinuitiesofthematerialparametersandofthevolumefractionfunctionthat appearinthePDEsystem.Atthispointwewouldliketoemphasizethatinthispaperwedeliberatelyusefinitevolumeand discontinuous Galerkinfiniteelement schemesthat employapiecewise polynomial approximationspace, whichexplicitly permits jumps in the discrete solution atelement interfaces.In thiscontext the path-conservative schemes can properly deal withjumps inα
if they areexactly resolved
atan element interface, thusnaturally allowing todiscretize problems alsowith ID=
0,seealso[102,103] foradiscussion inthecontextofshallowwaterequationswithdiscontinuous bottomtopography.Inthecaseof
I
D=
0,ourmethodbecomesagainasharpinterfacemethodifthejumpsareexactlyresolvedonthegrid,butthisisactually
not the
mainobjectiveofthepresentpaper.Regarding the space-time predictor, we need to introduce a new polynomial basis of degree N in space and time
{θ
k}
k=1...(N+1)d+1 wherenowθ
k(
x,
t)
contains alsothe time. We represent qh(
x,
t)
in termsof thisnew space-timebasisas qh
(
x,
t)
=
(N+1)d+1 k=1θ
k(
x,
t)
qˆ
nk.
(36)Let
T
◦i=
Ti− ∂
Ti denotetheinteriorofT
i andTist=
Ti◦× [
tn,
tn+1]
denotethenewspace-timecontrolvolume.Thespace-timepredictoristhencomputedasanelement-localsolutionofthefollowingweakformulationofthePDEsystem(29):
Tst iθ
k∂
qh∂
t dx dt+
Tst iθ
kB(
qh)
· ∇
qhdx dt=
Tst iθ
kS(
x,
t)
dx dt,
(37)for
k
=
1. . . (
N+
1)
d+1.UsingintegrationbypartsinthefirsttermofEq.(37) weobtaintwospatialcontributionsonT
iattn+1 and
t
n andan internalone sinceθ
k
= θ
k(
x,
t)
contains explicitlythetime.Forthespatial contributionattimet
n weuse thenumericalsolution fromtheprevious time step.Notice that thiscorresponds toupwindingin thetime direction duetothecausalityprinciple.OnethusobtainsthefollowingweakformulationofthePDE
in the small [
108]: Tiθ
k(
x,
tn+1)
qh(
x,
tn+1)
dx−
Tiθ
k(
x,
tn)
uh(
x,
tn)
dx−
Tist∂θ
k∂
t qh(
x,
t)
dx dt+
Tistθ
kB(
qh)
· ∇
qhdx dt=
Tistθ
kS(
x,
t)
dx dt.
(38)SinceEq.(38) iselement-localitcanbesolvedusingasimpleandefficientPicardmethodwithoutanycommunicationwith theneighborelements,seee.g.Dumbseretal. [109].
Thenumericalschemeisconstrainedbya
local CFL-type
stabilitycondition,see[109,110,87],thatisgivenbyt
<
CFL d h 2N+
1 1|λ
max|
,
(39)where
h is
thelocal meshsize,λ
max isthe maximumeigenvalueofthe PDEsystem, andCFL<
1 is theCourantnumber,which should be chosen according to [109] in order to have linear stability. Concerning the adaptive mesh refinement (AMR) werely onthe ExaHyPEenginehttp://exahype.eu,which isbuiltupon thespace-tree implementationPeano [111,
112] realizing cell-by-cellrefinement [113]. Forfurther details aboutAMR incombination withhighorder finitevolume andDGschemeswithtime-accuratelocaltimestepping(LTS)usedinthiswork,see[110,114–116].Throughoutthispaper, weuseadaptivemesheswitharefinementfactorof
r
=
3 betweentwoadjacentlevelsofrefinement.Inordertoillustrate theusefulcombinationofAMR withthe subcellfinitevolumelimiter, weprovidethe followingexample:withtwo levels ofmeshrefinementandfora DGschemewithpolynomial approximationdegree N=
3,thefinitevolume subgridonthe finestAMR levelwillbeby afactorofr
2· (
2N+
1)
=
32· (
2·
3+
1)
=
63 finerthanthegrid usedfortheDGschemeon the coarsestgrid level.Thiscorresponds to amesh refinementof almosttwo orders ofmagnitude. However, thishasno negative impactatall onthetime stepsizeusedonthecoarsestgridlevel,thankstotheuseoftime-accuratelocaltime stepping(LTS),whichisstraightforwardforADERdiscontinuousGalerkinandADERfinitevolumeschemes,see[36,110,87]. ThecombinationofAMRwiththefinesubgridusedforthefinitevolumelimiterisnecessarytoalleviatethelossofformal orderofaccuracyinsidethelimitedcellsandinordertoallowasmallinterfacewidthID∼
h.Inordertodecidewheretorefine,weintroduceasimplerefinementindicatorfunctionnamed
ϕ
=
ϕ
(
x,
t)
thatdefines theobservedvariablefortherefinement/recoarsening processanda socalledreal-valuedestimator function
χ
=
χ
[
ϕ
]
,see again[110] formoredetails.Afterdefiningtheindicatorfunction,wedefinethecell-averagesofϕ
asˆ
ϕ
i=
1|
Ti|
Tiϕ
(
x,
t)
dx∀
i=
1. . .
Ne,
(40)andthenwecomputetheestimatorfunctionas
χ
i[
ϕ
] =
max c∈Viˆ
ϕ
c− ˆ
ϕ
i/
xc−
xi,
(41)where
V
i containsalltheVoronoineighborelementsofT
i.Ourestimatorfunctionχ
issimplybasedonanapproximationofthe gradientofthe solutionin severalspatial directions[110]. Withtheseingredients athand,we introducea simple rulefortherefinement/recoarseningprocessbasedontwothresholds
χ
+andχ
− asfollows:(i) if
χ
i[
ϕ
]
>
χ
+ thenT
i islabeledformeshrefinement;(ii) if
χ
[ϕ
]
<
χ
−thenT
iislabeledformeshrecoarsening.Withinthispaper,wealwaysuse
ϕ
(
x,
t)
=
ϕ
(
Q)
=
α
,χ
+=
0.
01 andχ
−=
0.
001.Wewillalsousethevolumefractionα
to specifythezoneswheretoactivatethesubcellfinitevolumelimiter[88].Inparticular,weactivatetheFVlimiterwheneverα
∈ [
/
,
1−
]
,with=
10−3.Aslongasthetopologyofthegeometrydescribedbyα
issupposedtobestationaryintime, wecanconsidertherefinementandthelimitedzonesalsoassteadyandthereforetheyneedtobeidentifiedonlyoncein themeshinitializationstep.We stress againthat theuseof atime-accurate localtime stepping strategy (LTS)is mandatory inorderto avoidthe reductionofthetimestepsize
t in regionsthatarefarawayfromthediffuseinterface.
4. Numericalresults
4.1. Reflected plane wave
The purposeofthisfirsttest problemisto systematically studytheinfluenceof thewidth ID of thediffuseinterface
layerontothenumericalresults.Wealsoshowthatthemodelindeedconvergestothecorrectsolutioninthelimit ID
→
0.We take a simple plane wave impulse in a domain
= [−
1,
1]
× [−
0.
1,
0.
1]
initially placed at x0= −
0.
25 andhittinga free surface boundary placed in xD
=
0. The Laméconstants are chosen asλ
=
2,μ
=
1 andρ
=
1. We define Q0=
(
0,
0,
0,
0,
0,
0,
0,
0,
0,
λ,
μ
,
ρ
,
α
(
x))
andδ
= (
0.
4,
0.
2,
0.
2,
0,
0,
0,
−
0.
2,
0,
0,
0,
0,
0,
0)
andsetQ
(
x,
y,
t=
0)
=
Q0+ δ ·
e− (x−x0)22
,
withthehalfwidth
=
0.
05.Thevolumefractionfunctionα
(
x)
isprescribedaccordingto(28) and(27).WeuseanADER-DG P4 schemeanda uniform Cartesian gridwith 100×
2 elements. The mesh resolutionis chosen fine enoughso that theFig. 1. Numericalresultsobtainedwiththenewdiffuseinterfaceapproachforaplanewavereflectionproblemonafreesurfacelocatedinx=0 usinga variableinterfacethicknessofID=0,ID=0.001,ID=0.01 andID=0.03.Inallfourcaseswereportthevelocitycomponentu comparedwiththeexact solutionoftheproblem(bottom)togetherwiththespatialdistributionof
α
(top).numerical results are grid-independent and only depend on the choice of the interface thickness ID. Since for this test
cp
=
2,theexactsolutionattimet
=
tend=
0.
25 isthereflected p-wavewhichislocatedagainintheinitialposition.Weconsiderfourcaseswithdifferentchoicesoftheinterfacewidth ID,rangingfrom
I
D=
0.
03 tothelimit ID=
0,wheretheinterfaceisexactlylocatedonacellboundary.FromtheresultsdepictedinFig.1wecanconcludethatthediffuseinterface methodisabletoreproducetheexactsolutionoftheproblemforsufficientlysmallvaluesoftheinterfacethickness
I
D.Wealsostressthattheuseofapath-conservativemethodallowsustoreduce theinterfacethicknessexactlyto
I
D=
0,whichleads toajumpin
α
atan elementinterface,butwhichisstillproperlyaccountedforthankstothejumptermsD
− used inthenumericalscheme.Forratherlargevaluesofthefiniteinterfacethickness
I
D,wheretheactualshapeofthespatialdistributionofα
startstoplayarole,wehavefoundempiricallythatagoodchoicefortheparameters
η
andp in (27) isη
= −
0.
6 and p=
0.
5.This choice allows toobtainstill acorrectreflectionofap-waveeven forvery thick interfaces.However, forsufficientlysmall valuesofI
D,thechoiceofη
and p has onlyverylittleinfluence.Weconsidernowasimilarsetupofap-wavetravelinginaheterogeneousmaterialwithperiodicboundarieseverywhere.TheLaméconstantsarespecifiedas
(λ,
μ
,
ρ
)
=
(
4,
0,
1)
0≤
x≤
0.
2(
2,
1,
1)
otherwise (42)Note that inthearea 0
≤
x≤
0.
2 thePoissonratioisν
=
1/
2.The abilityofADER-DG schemesto dealwithfluids(μ
=
0) properlyhas alreadybeendiscussed in [34,117]. We furthermoreplace a reflectivefree surface atxf s
=
0.
75 only byusing a change inthe parameter
α
sothat the exact solution attend=
1.
0 isagain thereflected p-wave, located in theinitial position.We usean ADER-DGP3 schemeona sequenceofuniformCartesian grids inorderto checkthe accuracy
of the scheme.In Table 1 we report the L2-error norm inthe case whereno free surface boundary appears (i.e.
α
=
1everywhere). In thiscasethewave passesthroughthe heterogeneousmaterial andreturnstothe initialstate dueto the periodicboundary.WeobservehighorderofconvergenceifweuseapureDGschemewithoutlimiter(leftcolumn),while we observetheexpecteddecayoftheorderifweartificiallyactivatethelimitercloseto x
=
xf s.Takingα
=
0 forx
≥
xf s,we observeafirstorder ofconvergence,since thejumpin
α
alsorepresentsajumpinthe discretesolution,seeTable2. ThisisobservedalsoforpositivevaluesofI
D anddifferentvaluesof0,asreportedintherightcolumnofTable2.
InprincipleonecankeeptheLaméparametersinsidethediffuseinterfaceandoutsidethesolidastheonesinsidethe solid.HoweverwehaveempiricallyfoundoutthatrescalingtheLaméconstantsaccordingtothelocalvalueof
α
improves thenumericalconvergenceandallowstoreachsecondorderofaccuracy,seeTable3.Table 1
ComputedL2-errornormforthecase
α
=1 withoutthelimiter(left)andlimitingtheelementsthatareclosetoxf s(right). Elements L2(u) 20×2 4.8988e-3 − 40×2 7.2321e-4 2.8 80×2 4.1602e-5 4.1 160×2 2.4629e-6 4.1 Elements L2(u) 20×2 6.6547e-3 − 40×2 1.3748e-3 2.3 80×2 2.1391e-4 2.7 160×2 3.7588e-5 2.5 Table 2
ComputedL2-errornormforthecaseof
α
alignedwiththemesh(left)andforpositivediffuseinterfacesizeID(right). Elements L2(u) 320×2 2.5851e-4 − 640×2 1.2104e-4 1.1 1280×2 4.6230e-5 1.4 2560×2 2.0770e-5 1.2 Elements ID L2(u),0=10−3 L2(u),0=10−4 40×2 0.02 3.4813e-3 − 3.2060e-3 − 80×2 0.01 2.0510e-3 0.8 2.0144e-3 0.7 160×2 0.005 1.1723e-3 0.8 1.1374e-3 0.8 320×2 0.0025 5.9087e-4 1.0 5.9268e-4 0.9 Table 3Computed L2-error norm for α
aligned with the mesh and rescaledLaméconstantsλandμ.
Elements L2(u) 40×2 1.4512e-2 − 80×2 2.0014e-4 2.9 160×2 3.2374e-5 2.6 320×2 8.9161e-6 1.9
Fig. 2. Setupofthescatteringtestproblem.AMRgridanddistributionofthecharacteristicfunction
α
(left).Detailofthefreesurfacelocation∂C shown viaadashedlineandα
colorcontours(center).Limitedcellshighlightedinredandunlimitedcellsshowninblue(right).(Forinterpretationofthecolors inthefigure(s),thereaderisreferredtothewebversionofthisarticle.)4.2. Scattering of a plane wave on a circular cavity
In thistest casewe consideran initially planar p-wave travelingin x-direction inside a solid medium andhitting an emptycircularcavity.Theinitialstateisgivenby
Q
(
x,
0)
= (
0,
0,
0,
0,
0,
0,
0,
0,
0, λ,
μ
,
ρ
,
α
)
+
0.
1· (
4,
2,
2,
0,
0,
0,
−
2,
0,
0,
0,
0,
0,
0)
sin(
2π
x),
(43) withλ
=
2,μ
=
1 andρ
=
1.The value ofα
is parameterized throughthecircularsurface C= {(
x,
y
)
|
x
2+
y2≤
0.
252}
sothat
α
(
x)
=
0 if x∈
C and α=
1 if x∈
/
C . Thewidthparameterofthe diffuseinterface issetto ID=
0.
01 on∂
C . Thecomputational domainis
= [−
3,
3]
2 andtheinitial Cartesian gridatlevel=
0 consistsof80×
80 cells.We then use onefurtherrefinementlevelmax
=
1 basedonthegradientofα
inordertorefinethemeshclosetothediffuseinterface.Furthermore,we usea fifth orderADER-DG methodbased onpiecewise polynomials ofdegree N