UNIVERSITY OF BOLOGNA
SCHOOLOF ENGINEERING ANDARCHITECTURE
Master of S ien e Degree inAerospa e Engineering
Testing of subgrid s ale (SGS)
models for large-eddy simulation
(LES) of turbulent hannel ow
Master thesis in Applied Aerodynami s
SUPERVISORS:
Prof. Alessandro Talamelli
Prof. Arne V. Johansson
ASSISTANT SUPERVISOR:
Dr. Geert Brethouwer
STUDENT:
Matteo Monte hia
A ademi Year 2013-2014
andthe onlywaytobetrulysatised
istodo what youbelieveisgreatwork.
And the onlywaytodo greatwork
istolovewhat youdo.
If youhaven'tfoundityet,keep looking.
Don'tsettle.
As withall mattersof the heart, you'llknow whenyound it.
SteveJobs
Vedi aroami o osatis rivo etidi o
e omesono ontento
di essere quiin questomomento,
vedi aroami o osa sideve inventare
per poter i ridere sopra,
Sub-grid s ale(SGS) modelsare requiredin order to model theinuen e of theunresolved small
s ales on the resolved s ales in large-eddy simulations (LES), the ow at the smallest s ales of
turbulen e.
In the following work two SGS models are presented and deeply analyzed in terms of a ura y
throughseveralLESswithdierentspatialresolutions,i.e. gridspa ings.
The rstpart ofthis thesisfo uses onthe basi theory ofturbulen e, the governingequations of
uiddynami sandtheiradaptationtoLES.Furthermore,twoimportantSGSmodelsarepresented:
oneistheDynami eddy-vis ositymodel(DEVM),developedby[Germanoetal.,1991℄,whilethe otheristheExpli itAlgebrai SGS model(EASSM),by[Marstorpetal.,2009℄.
In addition, somedetails about the implementation of the EASSM in a Pseudo-Spe tral
Navier-Stokes ode[Chevalieret al.,2007℄arepresented.
Theperforman eof thetwoaforementionedmodelswillbeinvestigatedin thefollowing hapters,
bymeansofLESofa hannel ow,withfri tionReynoldsnumbers
Re
τ
= 590
uptoRe
τ
= 5200
,with relatively oarse resolutions. Data from ea h simulation will be ompared to baseline DNS
data.
Resultshaveshownthat,in ontrasttotheDEVM, theEASSM haspromising potentialsforow
predi tionsathighfri tionReynoldsnumbers:thehigherthefri tionReynoldsnumberisthebetter
theEASSM willbehaveandtheworsetheperforman esoftheDEVMwillbe.
The betterperforman eof theEASSM is ontributedto theabilityto aptureowanisotropyat
thesmalls alesthrougha orre tformulationfortheSGSstresses.
Moreover,a onsiderableredu tionintherequired omputationalresour es anbea hievedusing
the EASSM ompared to DEVM. Therefore, the EASSM ombines a ura y and omputational
VersioneItaliana
Nelle Simulazioni agrandi vorti i (Large-eddy simulations, LES), dei modelli per la s ala di
sottogriglia(Sub-grid s ale, SGS) sono ne essariper riprodurrel'inuenza delle piúpi ole s ale
dellaturbolenza,quindinonrisolte,suquelle heinve evengonorisoltedirettamente.
Nel seguente lavorosono presentati due modelli SGS, la ui a uratezzaverrá poi analizzata
at-traversovarieLESadiverserisoluzionispaziali,equindidiversiintervallididierenziazione.
Laprima partedellatesisi on entrasullateoriadellaturbolenza,partendodalleequazioni
osti-tutivedellauidodinami a,noallaloroversioneperLES.
Due importanti modelli SGS sono stati presentati: il primo é il modello Dynami eddy-vis osity
model(DEVM),di[Germanoetal.,1991℄,ilse ondoéilmodelloSGSEspli itoAlgebri o,Expli it Algebrai SGS model (EASSM),di[Marstorpetal.,2009℄.
Sarannoinoltrefornitidettagliaggiuntivisull'implementazionedelmodelloEASSMsuun odi edi
Fluidodinami aComputazionale(ComputationalFluidDynami s,CFD),Pseudo-Spettrale
Navier-Stokessviluppato nel Linné Flow Centre del Dipartimento di Ingegneria Me ani a del KTH di
Sto olma,da[Chevalieret al.,2007℄.
I seguenti apitoli verterannosull'analisi dellastimadi unussoin un anale, hannel ow, fatta
daimodellides rittiinpre edenza,perunbassonumerodiReynoldsbasatosull'attrito,
Re
τ
= 590
,no a
Re
τ
= 5200
. I datiottenutida ias unasimulazioneverranno onfrontati ondatidiSimu-lazioni Numeri heDirette (Dire tNumeri al Simulations, DNS).
Dai risultati ottenuti si puó on ludere he, dierentemente dal modello DEVM, l'EASSM ha
promettenti potenzialitá nella stima del usso ad alti numeri di Reynolds
Re
τ
: piú alto é talenumero,piúilmodelloEASSMdarárisultatia urati,mentreleperforman esdelDEVM
peggior-eranno.
Lemiglioriperforman edelmodelloEspli itoAlgebri opossonosenz'altroessereattribuiteallasua
abilitádi al olareinmaniera orrettal'anisotropiaallepi oles aletramiteunaformulazione
or-rettadeglistressdisottogriglia,SGS.
In on lusione,datalaridottaquantitádirisorse omputazionaliri hiestapereettuaresimulazioni
rispetto alDEVM, tale modello ombina a uratezza ee ienza omputazionale,tanto he puo'
1 Turbulen e 5
1.1 Introdu tiontoTurbulen e . . . 5
1.2 Numeri alapproa hestoN-S solution . . . 12
1.3 Aparti ular ase: TurbulentChannelFlow . . . 17
2 Large Eddy Simulation 25 2.1 Thelteringoperation . . . 26
2.2 GoverningequationsofLESandthe losure problem . . . 27
2.3 An exampleofa losure: theEddyVis osityModel(EVM) . . . 27
2.4 Subgrid-s aledissipation . . . 28
3 Subgrid-s alestress modelsfor LES 31 3.1 TensorialpolynomialformulationoftheSGSstresstensor . . . 31
3.2 Thedynami pro edureforEVM(DEVM) . . . 33
3.3 Expli itAlgebrai SGS stressmodel . . . 34
4 Implementationin a CFD ode 39 4.1 Theneedfora ura y: spe tralmethods. . . 39
5.1 MeanValues . . . 49
5.2 Root-meansquaredValues,rms . . . 50
5.3 Vorti alstru tures . . . 51
6 Results 53
6.1 LESat
Re
τ
= 590
. . . 546.2 LESat
Re
τ
= 2000
. . . 646.3 LESat
Re
τ
= 5200
. . . 72Provaad immaginare unpenna hio di fumo fuorius ente dauna pipa di un uomopensante,
seduto su diuna sediaadondolo. An he in questopi oloaspettodellavita quotidiana la
turbu-lenza gio aunruolofondamentale.
Il noto si oRi hardFeynman denii laturbulenza ome il piúimportante problema dellasi a
lassi a hean oranonéstatorisolto.
Nel orsodegliannidiversis ienziatihannoprovatodi omprendereil vero omportamentoditale
fenomeno, la ui omplessitá gia e nelle equazioni he lodes rivono. Per unuido Newtoniano,
la turbolenzaé denita dalle equazioni di Navier-Stokes, unsistema di equazioni non-lineari alle
derivateparziali,la uisoluzioneanaliti aan oranonéstataottenuta.
Tuttavia, degli appro i alternativi per ottenere delle soluzioni sono stati sviluppati nora: un
metodo onsistenell'eettuaredegliesperimenti: tramitegalleriedelventosiamoingradodi
ripro-durreussiinsvariate ondizioni,dallesituazioni lassi hedistratolimite(motosudiunaparete)
eussoinun anale,noa asipiú ompli ati, omeilussoattornoadun orpotozzo( omeuna
automobile)eun orpoaerodinami o, omeunaeroplano.
Gliesperimentihanno omeneprin ipalel'analisidellequalitádelusso( omepressionee
velo -itá) attraversodelle sondeedelle te ni he di visualizzazione, omela Parti le ImageVelo imetry
(PIV).
[Talamellietal.,2009℄nell'Universitádi Bologna.
D'altro anto,unappro iodiversoéquellodisfruttarel'analisinumeri aepassarealla
implemen-tazionedelsistemaN-S su odi i diFluidodinami aComputazionale.
Tuttavia, la soluzione numeri a e ompleta fornita dalle Simulazioni Numeri he Dirette (DNS),
in grado di des riverela turbolenza aqualsiasi s ala, non ésemprepossibilea ausa dellegrandi
risorse omputazionaliri hieste,in parti olare adelevati numeridi Reynolds. Perquesto motivo,
l'obiettivoprin ipaledelprogettoCICLoPEéquellodi omprendereapienoilfenomenodella
tur-bolenzaqualoranonsiapossibileotteneredatitramiteDNS.
Comunque,late ni a DNSnon élasola in gradodi riprodurre numeri amente unpre iso usso.
Una se ondapossibilitási hiamala simulazione agrandi vorti i (Large-eddysimulation -LES);
questate ni aéingradodiraggiungere, onopportunimodelliperlepi oles aledellaturbolenza,
unasoddisfa entea uratezza omputazionale onunaragionevolequantitádirisorse
omputazion-ali.
A tale avviso, la tesi verterá sull'impiego e il test di due modelli innovativi per le pi ole s ale
dellaturbolenza;inparti olareverrámessaa onfrontol'a uratezzadi ias unmodello onl'altro,
rispettoan he i risultatiforniti da DNS. Lesimulazioni sonostate ompiuteatre diversi numeri
di Reynolds;in parti olarel'ultimo aso,dimaggioreimportanza, érelativoalmassimonumerodi
ReynoldsraggiuntonoradaDNS.
Laprimapartedellatesi omprenderálades rizionedellaturbolenza,siadaunpuntodivistasi o
hematemati o. Saranno onsiderateleequazionidiN-S perunuidoNewtoniano, onusso
in- omprimibileeturbolento, inun anale.
Nelse ondo apitololate ni aLESsaráespressaneidettagli,mentrenelterzoverrannopresentati
duemodellidis alasotto-griglia(sub-grids ale-SGS).Ilquarto apitolospiegherádeidettagli
te -ni isull'implementazionedellesimulazioni,deiduemodelli,edellaparallelizazioneMPIutilizzata.
Nel quinto apitolo enni di metodi statisti i per l'analisi della turbolenza saranno arontati. I
risultativerrannopresentatinelsesto apitolo,einnesarannofornitedelle on lusioninelsettimo
Imagineasmokingplume omingoutfrom apipeofathinking man,thisisatypi aleveryday
life s enariowhereturbulen eplaysamain role. Thephysi ist Ri hardFeynmandened itasthe
most important unsolvedproblemof lassi al physi s.
S ientistsalongtheyearshavetriedto understand therealbehaviourof thisphenomenon, whi h
gains its omplexitybe auseoftheequationsthat des ribesit. ForaNewtonian uid,turbulen e
istra edbyNavier-Stokesequations,asystemofnon-lineardierentialequationswhoseanalyti al
solutionhasnotbeenprovidedyet.
Anyway,alternativeapproa heshasbeendevelopeda rosstheyearsuptonow: themostintuitive
andoldwaytounderstanduidmotionistomakeexperiments: throughtheemploymentofwind
tunnelswe anreprodu etheowinseveral onditions,fromthebasi al hannelowandboundary
layerto themore omplexones,liketheowa rossablu body (likea ar)oranairplane. This
te hniqueinvolves theanalysis of the owproperties(likepressure and velo ity) throughprobes
andadvan ed visualizationte hniquessu hasParti leImageVelo imetry(PIV).
A re entexamplewhoseaimistodes ribehighReynoldsnumberturbulen ea rossapipeis
nev-erthelesstheCICLoPEexperiment,abigproje tdevelopedby
[Talamellietal.,2009℄inUniversityofBologna.
These ondapproa histhustousenumeri alanalysisinordertoimplement odesableto
mainCICLoPEproje t'saimistounderstandturbulen ephenomenafurthertheReynoldsnumber
limitdi tatedbyDire tNumeri alSimulations.
However, DNS is not the only te hnique able to reprodu e numeri allya given ow. A possible
se ond hoi eis alledLarge-eddysimulation;thiste hniqueisabletorea hasatisfying
omputa-tional a ura ywithareasonableamountof omputationalresour es.
InthisthesisNavier-StokesequationsforaNewtonianuidandanin ompressibleowina
turbu-lent hannelows enariowillbedes ribedintherst hapter,togetherwithaphysi aldes ription
ofturbulen e. Inthese ond haptertheLarge-eddysimulation(LES)te hniquewill beexplained
in detail,while in the third onetwoLES sub-grid s alemodelsare shown. In hapter foursome
te hni aldetails oftheimplementationof theLES,ofthemodels,with aparti ularfo usonMPI
parallelizationare pointedout. Sometheory aboutturbulen e statisti swill befa edin thefth
1
TURBULENCE
Everybodyisagenius. Butif youjudgeash byitsability to limb atree,
itwilllive itswholelifebelieving thatitisstupid. (A.Einstein)
1.1 Introdu tion to Turbulen e
Thephenomenonofturbulen eisfoundinseveralappli ations, forexample, ombustiontumbling
in Internal CombustionEngines(ICEs),thewakeofaFormula1 ar,thejetspreadbythenozzle
ofasupersoni air raftengine.
Inautomotiveengineering,forexample,thestudyofaerodynami sarounda arinvolvesthe
har-a terizationofaturbulentwall-boundedow, alledaBoundary Layer. Itwasintensivelystudied
byL.Prandtlin1904;hereturbulen eisthemain responsibleforfri tionandwakedrag.
In asolid ro ketmotor nozzle,there's ageneration ofaplume, whereturbulentmotionsofmany
s ales anbeobserved;fromeddiesandbulges omparableinsizetothewidthoftheplumetothe
and haoti nature,sin ethemotionofeveryeddy isunpredi table.
Figure1.1Large-eddysimulationofjetfromare tangularnozzle. There tangularnozzleis
showningraywithanisosurfa eoftemperature(gold) utalongthe enterplaneofthenozzle
showingtemperature ontours (red/yellow). Thea ousti eldisvisualizedby(blue/ yan)
ontoursofthepressureeldtakenalongthesameplane,fromP.Moin.
Whilelaminarowisa smoothand steadyowmotion,where any indu edperturbationsare
dampedoutduetotherelativelystrongvis ousfor es,inturbulentowsotherfor esmaybea ting
that ountera tthea tionofvis osity. Ifsu h for esarelargeenough,theequilibrium oftheow
is upset and theuid annot adapt suddenlyto vis osity. Thefor es that upset this equilibrium
anin lude buoyan y,inertia,oreven rotation. Ina hannel, vis ousandinertialfor esa tingon
theuidareproportionalto
F
v
∝ νL
(1.1)F
i
∝ V L
2
(1.2)where
ν
istheuidvis osity,andL
andV
arethe hara teristi velo ityand lengths ales. Ifthevis ous for esontheuidarelarge omparedwithothers, anydisturban esintrodu edin the
owwilltend to bedamped out. Ontheother hand,iftheinertialfor es be ome large,theuid
will tend to break upinto eddies. For greater inertialfor es, the eddies will break upinto even
smallereddies. This will ontinueuntilwerea hasmallenoughlengths ale(eddy size)onwhi h
The largest of these eddies will be onstrained by the physi al size onstraints on the ow (like
hanneldiameter);thesmallereddieswillbe onstrainedbythevis ousfor eswhi ha tstrongest
at the smallest length s ales. Therefore, one of the di ulties asso iated with the predi tion of
turbulentowis thattherangeoflengths ales anbeverylarge.
Thedes riptionofturbulen einvolvesdierent on eptsliketurbulen eenergyprodu tion,transfer
anddissipation. Ri hardson'sfamouspoemgivesagoodideaaboutturbulen e:
Big whorlshave littlewhorls
That feed ontheir velo ity
Andlittle whorls havelesserwhorls
Andso ontovis osity (in the mole ularsense)
Thisisthedes riptionoftheenergy as ade on ept. Itstatesthatturbulentows anbe
on-sideredasanagglomerateofeddies ofdierentsizes. Largeenergy ontainingeddiesareunstable
and break down andtransferenergy to smallereddies. The pro essgoesontill thesmallestone,
theKolmogorovs ale,where energyis dissipatedinto heatbyvis ousee ts.
This isthegreat on lusion thatKolmogorovmadein 1941,andhis rsthypothesis is ompletely
basedonthat: atsu ientlyhighReynoldsnumber,thestatisti softhesmalls alesareuniversal
andaredeterminedsolely byvis osity,
ν
,andtheenergydissipationrate,ε
.Usingdimensionalanalysis,itispossibleto derivetheKolmogorovlengths ale
η
,times alet
η
andvelo itys ale
v
η
:η =
ν
3
ε
!
1/4
,
t
η
=
ν
ε
!
1/2
,
v
η
= (νε)
1/4
(1.3)Then, in a ordan eto Ri hardson'spoem, Kolmogorovmade ase ond hypothesis, basedon
thefa t thatatsu ientlyhighReynoldsnumberthestatisti softhes aleswhi haresu iently
larger than
η
and mu h smallerthan the largestenergeti s ales are solely des ribed byε
. Thishypothesisreferstotheinertialrangeofs ales. Thekineti energyspe trumoftheses ales anbe
des ribedby
Figure1.2Turbulen eenergyvswavenumberspe trum,fromJ.M.M Donough.
where
k
isthewavenumberandC
k
≈ 1.5
[Sreenivasan,1995℄istheKolmogorov onstant. Fromthelastdiagram,it's learthatturbulen ehasdierentbehavioursa ordingtothewavenum-ber. Basingonthese on epts,we andistinguishfourdierentregions:
1. thelarges ale,determinedbytheproblem domaingeometry;
2. theintegrals ale
(Λ)
,whi h isanO(1)
fra tion(oftentakentobe∼ 0.2
)ofthelarges ale;3. the Taylor mi ros ale whi h is an intermediate s ale, found in the Kolmogorov's inertial
subrange
(η ≪
2π
κ
≪ Λ)
4. theKolmogorov(ordissipation)s ale
(η)
whi histhesmallestofturbulen es ales,theinners ale
1.1.1 3D Nature of Turbulen e
Turbulen e is rotational and a three-dimensional phenomenon. It is hara terized by large
u tuationsinvorti ity,whi hareresponsibleforvortexstret hingandlengths aleredu tion. These
hara teristi sareidenti ally zeroin twodimensionsand theseare thereasonswhyturbulen e is
hardtodes ribebothanalyti allyandnumeri ally.
Thesethree-dimensionaldynami alme hanismsarehighly omplexandnonlinear,howevertheow
anbeassumedas bi-dimensionalforlarges ale2Dstru tures. These stru turesplayadominant
inthesmallers ale,wheretheyarefundamentalformixing,mostofallatmole ulars ales(e.g. in
ombustionproblems).
1.1.2 Order & Randomness
Despiteturbulen e is haoti , it onsists of ompletely randommotionsthat anaggregatein
oherentstru tures. Typi alexamplesareturbulentboundarylayersand homogeneousturbulent
shearows,whi hexhibithorseshoe,orhairpinvorti es(seegure1.3)thatappeartobeinherent hara teristi s. Freeshearowslikethemixinglayerreveal oherentvortexstru turesvery learly,
againevenforveryhighturbulen eintensities.
The on epts of order and randomness have also led to some new analyti approa hes and new
interpretations in the study of turbulen e. The names of these dis iplines are known as Chaos,
Bifur ation Theory, and Dynami alSystems [M Donough,2007℄. These theorieshavebeenfa ed forthestudyofturbulen e,in parti ularintheareaofhydrodynami stabilityandtransitionfrom
laminarowtoturbulentone. Cometoattentionofmathemati ians,physi istsandengineers,these
phenomenaisofaremarkablenon-linearity,whi hmakesturbulen eunpredi tableand omplexto
des ribe.
Asanonlinearproblem,it anbeseenthatthesolutionstotheseproblemswiththesamenonlinear
equationswithonlyslightdieren esininitial onditions,willrapidlydiverge. Therefore,asuitable
denition ofturbulen e must ne essarilyinvolvea omplexdynami al systemwith many degrees
offreedom.
1.1.3 The Reynolds number
Turbulen e an be seen also a play between inertial for es and vis ous. Therefore the ratio
betweenthemis ru ialinorderto hara terizeaow. TheReynoldsnumberplaysthat role,and
in turbulentowsholds
Re
def=
U L
ν
≫ 1.
(1.5)Formany ows of pra ti alimportan e (e.g. a ow on airplanewings) the Reynoldsnumber
anbeontheorderof
Re ∼ 10
6
. Thismeansthat thevis ousfor es,thataremole ularfor es,a t
in smallers alesthanin thelargeones. However,in anyturbulentowthemole ularvis osityis
alwaysimportantatsomes ale. AstheowReynoldsnumberin reases,theregionwhere vis ous
ee tsareremarkable,de reasesinthi knessandthevelo ityoftheow hangesveryrapidlyfrom
zeroatthesurfa eto thefree-streamvelo ityat theouteredgesoftheboundarylayer. Again,we
see thetenden yofthenonlinearinertialtermstogeneratedis ontinuitiesat highwavenumbers.
On e more,theReynoldsnumber anbeinterpretedalso intermsoflengthandtimes aleratios.
Let's onsideradu tofwidth
L
,withaowvelo ityU
. Thetimeauidparti le,withtransversevelo ity
u
′
takesto rossthedu t is alledtheinertial time,
T
i
∼ L/u
′
. Atthesameway,vis ous
for eshaveatimes ale,
T
v
∼ L
2
/ν
.
Inaturbulentow,theinertialtime-s alewillbemi h lessthanthediusivetime-s ale,
T
v
T
in
=
u
′
L
ν
> 1.
(1.6) 1.1.4 Navier-Stokes equationsTurbulen e behaviouris ompletely des ribed by the governing equations of uid me hani s,
i.e. the ontinuity and Navier-Stokes(N-S) equations. In ase of in ompressible ows, they are
expressedin thefollowingwayonnon-dimensionalform:
∂u
i
∂x
i
= 0;
∂u
i
∂t
+
∂u
i
u
j
∂x
j
= −
∂p
∂x
i
+
1
Re
∂
2
u
i
∂x
j
∂x
j
(1.7)where
Re
isthe hara teristi Reynoldsnumberoftheow,u
i
, i = 1, 2, 3
arethevelo ityom-ponents,
p
isthepressure. NotethatEinstein'ssummation onventionisusedhere,wherei = 1, 2, 3
.Together withthemainoweld,sometimes weneedalsoto understandphenomenasthat
Bydenition,thewordpassiverefersto the onditionthattheresultingdensitydieren es areso
small that theee t from the s alaron theowis negligible. Soapassives alar anbe heat or
temperatureinaowora on entrationofasubstan e.
ThereforeisalsopossibletouseNavier-Stokesequations1.7to des ribethedevelopmentofa pas-sives alar,
θ
:∂θ
∂t
+
∂u
i
θ
∂x
j
=
1
ReP r
∂
2
θ
∂x
j
∂x
j
(1.8)where
P r
is the Prandtl number, dened as the ratio of momentum diusivity to thermaldiusivity:
P r
def=
µc
p
k
.
1.2 Numeri al approa hes to N-S solution
Thereareseveralnumeri alapproa hestosolvethesystem1.7. Themostintuitiveoneis alled Dire tNumeri al Simulation, DNS,wherebythe governingeuquationsare solvedwithoutmaking
anyassumption, resolvingallthes alesfrom thesmallesttothelargestone. Thereforeitprovides
alltheinformationofaturbulentow,withoutanyapproximations. Sin ethe omputational ost
ofDNSs aleswiththeReynoldsnumberis
∼ Re
37/14
[ChoiandMoin,2012℄,thisisnotaordable forpra ti alengineeringanalysesathigh Reynoldsnumber.
Soamorepra ti alapproa hhasbeendeveloped,basedontheReynolds'de omposition[Reynolds,1894℄:
u
i
= u
i
+ u
′
i
,
θ = θ + θ
′
(1.9)
wheretheoverlinerepresentstheensembleaveragedquantity,and
u
′
i
andθ
′
arethevelo ityand
s alar omponentsu tuations,respe tively.
UsingtheReynoldsde ompositionin equations1.7and1.8,andtakinganensembleaverageofall terms,theReynolds-averagedN-S (RANS)equationsarederived:
∂u
i
∂t
+
∂u
i
u
j
∂x
j
= −
∂p
∂x
i
+
1
Re
∂
2
u
i
∂x
j
∂x
j
−
∂u
′
i
u
′
j
∂x
j
,
∂u
i
∂x
i
= 0
(1.10)∂θ
∂t
+
∂u
i
θ
∂x
j
=
1
ReP r
∂
2
θ
∂x
j
∂x
j
−
∂u
′
j
θ
′
∂x
j
(1.11)Note that here turbulen e is solely des ribed by the Reynolds stress tensor
u
′
i
u
′
j
and s alar ux ve toru
′
j
θ
. These terms have to beproperlymodelled in order to lose the problem. Ea h model involvesapproximationswhi h limitsits a ura y. Therefore, this approa h providesonlyanapproximatesimulationofthemeanow.
The rst simple model was developed by Boussinesq in 1877. It is based on an eddy vis osity
formulation
u
′
i
u
′
j
−
2
3
Kδ
ij
= −2ν
T
S
ij
,
S
ij
=
1
2
∂u
i
∂x
j
+
∂u
j
∂x
i
!
(1.12) whereK = u
′
i
u
′
i
/2
isthemeanturbulentkineti energy,S
ij
isthemeanstrain-ratetensor,the symmetri partofthemeanvelo itygradienttensor,ν
T
istheeddyvis osityweusetomodelastheprodu tofa ertains alewith length
Λ
of theeddiesandvelo ityV
. Usingdimensional analysisν
T
∼ ΛV.
(1.13)Eddyvis osityisthenmodeleda ordinglyto some hara teristi softheow.
Forinstan e, algebrai models(orzero equationmodels) relatelength
Λ
andvelo ityV
to themeanvelo ityeldandtheowgeometry hara teristi slikevelo itygradient,distan etothewall,
thi knessoftheshearlayeret . Thesekindofmodelsworkquitewellforthespe i asethatthey
are designedfor,e.g. atta hedboundarylayersanddierenttypesof thin shearlayers. However,
theydon'tgivesatisfa toryresultsforgeneral ases.
Better results an be a hieved using one-equation models, they typi ally solve an additional
transport equation for the turbulent kineti energy,
K
, or the eddy vis osity,ν
T
. One-equationmodels give good results for atta hed boundary layers and other thin shear layer ows, but for
omplexows. A goodexampleistheSpalart-Allmaras[SpalartandAllmaras,1992℄model (SA), that solvesfor theeddy vis osity. This isverysuitablefor aeronauti alappli ationsand a tually
isthestandardmodelforexternalaerodynami sCFDanalysesatBoeing.
Two-equations models solve two transport equations for two quantities that an be used for
determiningthelengthandvelo itys aleneededto omputetheeddyvis osity. Themost ommon
are
K − ε
andK − ω
models,wheretransportequationsfortheturbulentkineti energy,K
andforthedissipationrate,
ε
orthe turbulen efrequen yω
aresolved. Nowadaystheimplementationofthese modelsin ommer ialCFD odes(e.g., ANSYSFluent)presentsadditional orre tionsthat
might be dependent on non-lo al quantities su h as thewall distan e. Oneimportant and most
re ent example is the Menter SST
K − ω
model [Menter,1994℄, whi h is suitable for separated ows;itis thestandardturbulen emodelusedatAirbus.Despiteeddyvis ositytwo-equationsmodelsarestilldominatinginindustrialCFD,there'sabig
de-mandformorea uratepredi tionof omplexowsituations,in ludingonsetofseparation,highly
urvedows,rapidly rotatingowset . Inthesesituations, eddyvis osityBoussinesq'hypothesis
(1.12)doesnotdes ribetherealphysi swell. Anee t ofde orrelation aused byrotationo urs at high rotation rates, and generally the alteration of produ tion to dissipation ratio is a dire t
onsequen eofthat.
This phenomena, for instan e, is animportant aspe twhi h eddy vis ositymodels doesnot take
into a ount,be ausethemodelisinsensitivetosystemrotation.
A better alternative to these models are the Reynolds Stress Models. They solve transport
equationsforea hReynoldsstress omponentsderivedfrom themodelledN-S equations,in order
ompu-tationally moreexpensiveand ompli ated than theothers. However,those dierentialReynolds
stressmodels anbesimpliedusingtheweakequilibriumassumptionbyRodi[Rodi,1992℄. Details aboutthiswillbeshownin 2. Thealgebrai relationisimpli itin Reynoldsstresses,butthereare some expli it solutions ([Pope,1975℄; [GatskiandSpeziale,1993℄; [WallinandJohansson,2000℄). These modelsare alledExpli itAlgebrai ReynoldsStress Models(EARSM).
Inparti ular,in theEA model[Wallin andJohansson,2000℄ theowanisotropy(
a
ij
)isdes ribed asanexpli itexpressionintermsofthe(normalized)meanstrainandrotationtensorswithaddi-tional s alarparameters. This leadsto a omparable omputational eorts, as ompared to eddy
vis ositytwo-equationmodels.
There'salsoaninterestinganalogyforthes alar
θ
modeling. Taylordevelopedananaloguewaytoformulate theeddydiusivitymodel(EDM)forthemeanturbulents alarux
u
′
i
θ
′
[Taylor,1915℄u
′
i
θ
′
= −D
T
∂θ
∂x
i
,
D
T
=
ν
T
P r
T
(1.14)where
D
T
istheeddy diusivity oe ientandP r
T
theturbulentPrandtlnumber.Inananalogouswayto
K − ε
model,Nagano&
Kim[Kim,1988℄developedtheK
θ
− ε
θ
model. Thetimes aleτ
θ
= K
θ
/ε
θ
,isusedto omputeD
T
∼ Kτ
θ
.Still, model 1.14 is not ompletely orre t. A ording to Bat helor, eddy diusivity assumes an alignmentbetweenthe
u
′
i
θ
′
ve torandthemeans alargradient,soithasto be onsidereditselfa
tensor. Thefollowingexpressionwillholdthen:
u
′
i
θ
′
= −D
ij
∂θ
∂x
j
(1.15)
Wheretheeddydiusivitytensor anberewrittenas[DalyandHarlow,1970℄
D
ij
= −C
θ
τ
θ
u
′
i
u
′
j
∂θ
∂x
j
(1.16)and
C
θ
isamodel oe ient.Inthesameway, s alar anbemodelled withanExpli it Algebrai model,in this ase alled the
Expli it Algebrai S alarFluxModels(EASFM).
Atrade-obetweena ura yand omputationaleortisLarge-eddysimulations(LES)ofturbulent
ows. InLES there'saseparationofs ales, inthesense that onlylarge-s aleeddiesareresolved,
while the remaining small s ales (whi h are alled sub-grid s ales, SGS) are modelled, on e the
resolveds ales havebeen omputed. Theseparationof s alesis generallydoneusing agrid. T
of the Navier-Stokesequations. This is thereforeaphysi ally- oherentapproa h sin eturbulen e
isunstationary.
Despitemore omputationallyexpensivethanRANS,LESgivesabetterdes riptionofturbulen e,
and unlikeDNS,is abletoprovideagoodresolution oftheowin ana eptableamountof
om-putational time. Re ently, that timehas been estimatedby Choi&Moin[ChoiandMoin,2012℄, to s aleas
∼ Re
26/14
.Afairandsimplieddistin tionbetweenDNSandLES anbenoti edhavingalooktothevorti al
stru tures, for bothof the ases,in gure 1.4. Ata rstglan e, we ansee that in LES vorti al stru tures are underestimatedand fewer, ompared to the DNS,whi h is able,instead, to givea
ompleteanddetailed des riptionofthem.
Therst LESmodelwasdevelopedbySmagorinsky[Smagorinsky,1963℄formeteorologi al appli- ationsusinganeddyvis osityassumptionintheSGSmodel. Thismodelhasbeenimprovedlater
on by Germano [Germanoetal.,1991℄ introdu ing the dynami pro edure, whi h givesa orre t asymptoti near-wallbehaviouroftheeddyvis osity,andimprovedtransitionalowspredi tions.
However, eddy-vis osity models are not anisotropi , that is, they are not able to apture ow
anisotropywell,afeatureoftheow,whi h isnotnegle tablenearthewalls.
Forthat reasonseveralnon-linear models (whi h are of ourse anisotropi ),have been developed
re ently,and someof themwill bedes ribed in detailin hapter 3. In parti ular,the aimof this thesis is to test the a ura y in ow predi tion of the Expli it Algebrai SGS model (EASSM),
a)
0
0.2
0.4
0.6
0.8
vel_u
-0.0112
0.864
b)0
0.2
0.4
0.6
0.8
vel_u
-0.0112
0.864
Figure 1.4 Vorti al stru tures inturbulent hannel ow at
Reτ
= 590
, visualized by isosurfa es ofλ2
, olored by the velo ity magnitude, froma) DNS by P.S hlatter b) LES simulationwiththeExpli itAlgebrai model.1.3 A parti ular ase: Turbulent Channel Flow
Mathemati almodelsforuiddynami shavebeenalreadydened. Inordertosolvethemodels,
amathemati alproblemneedsto bedened inapropertimeandspa edomains.
A ordingtothespa edomain,inadierentgeometry,physi alquantitiesandtheowwillbehave
inadierentway. Forthisreasonweneedtosetaparti ular ase,sothattheow anbeunivo ally
lassied.
Inthisthesisweare onsideringthe aseofChannelFlow,withthefollowingproperties:
turbulent,in thesense that theReynoldsnumberis su ientlyhigh su hthat the regime
anbeassumedas turbulent;
fully developed, sothat velo itystatisti s are onstant along
x
-axis. Inother words,theow is statisti ally stationary and statisti ally one-dimensional [Pope,2000℄, with velo ity statisti sonlyvariablealongthe
y
-axis. Inotherwords:hui = U = U(y),
hvi = V = 0,
u
′
v
′
= u
′
v
′
(y)
(1.17)
Channelowbelongsto the wall-boundedshear ows lass: owmotionis ontainedbetween
twosolidsurfa es. Therefore,no-slip onditions areimposedonthewalls,wheretheuidvelo ity
isassumedto bezero. Thefollowingpi tureshowsthequalitativebehaviourofthevelo ity.
Figure1.5ChannelFlowmeanvelo ityprole,with
u
′
u tuations ontour plotinthe
Furthermore,astatisti allysymmetri owgeometryw.r.t. themid-plane
y = δ
is onrmedbyexperiments;thereforethestatisti sof
(u, v, w)
aty
arethesameasthoseof(u, −v, w)
at2δ − y
.Reynolds number is always used to hara terize the ow, in this ase we will refer to two
parti ularReynoldsnumbers,
Re
τ
def=
u
τ
δ
ν
(1.18)Re
b
def=
uδ
ν
(1.19)where1.18is basedonthefri tionvelo ity
u
τ
,denedasfollows:u
τ
def=
p
τ
W
/ρ
(1.20)τ
W
isthemeanwallshearstressandδ
isthe hannelhalf-width. Thebulkvelo ityu
in 1.19,isdened asu =
1
2δ
2δ
Z
0
huidy
(1.21)Foraturbulent hannelow,thefollowingresultholds:
Re
τ
≈ 0.166Re
0.88
b
(1.22)Notethat thespe iedformula 1.22willbeused forthederivation of
Re
b
, whi h will be on-sideredasaninputquantityinthe omputations.As previouslystated, hannel ow is a wall-bounded shear ow. In a boundary layer ora
wall-bounded shear ow the hara teristi lengthfor streamwisedevelopmentis mu hlarger thanthe
ross-streamextentoftheregionwithsigni antvelo ityvariation.
So its behaviour an be observed studying a two-dimensional steady ows enario with the thin
shearlayerapproximation [JohanssonandWallin, 2012℄. Thisapproximationstatesthatthe har-a teristi streamwisedevelopmentlength,
L
ismu hlargerthanδ
,theshearlayerthi kness. Aswewill onrm later,in hannelowthere willbealayer,whose thi kness issmall omparedto the
hara teristi length,wherebyvis ousee ts fri tiondependson.
Thankstothisapproximation,theNavier-Stokesequations anbesimpliedtothethinshearlayer
U
∂U
∂x
+ V
∂U
∂y
= −
1
ρ
dP
0
dx
+
∂
∂y
ν
∂U
∂y
− u
′
v
′
!
|
{z
}
total shear stress
(1.23)
where
P
0
if thepressureat thewall. Notethat allthe velo ities in 1.23aremean values,and theonly omponentresponsibleforturbulen eisthelastoneontheright-hand-side,whi his alledtheReynoldsstress. Togetherwiththevis ousstress,theReynoldsstressgeneratesthetotalshear
stress.
In parti ular, usingthe hannel owassumptions given in 1.17, the thin-shearlayerequation be- omes
0 = −
1
ρ
dP
dx
0
+
d
dy
ν
dU
dy
− u
′
v
′
!
(1.24)Integratedinthewall-normaldire tionitreads
0 = −
1
ρ
dP
dx
0
y + ν
dU
dy
− ν
dU
dy
y=0
| {z }
u
2
τ
−u
′
v
′
+ 0
(1.25)Atthe enterline
(y = δ)
thetotalshearstressiszero,thereforewehavethefollowing ondition:1
ρ
dP
0
dx
= −
u
2
τ
h
(1.26)meaningthatthepressuregradientisrelatedtothefri tionvelo ityandthewidthofthe hannel.
Plugging this relation in 1.25, we ansee that thetotal shear stress developslinearly a rossthe hannel:
ν
dU
dy
− u
′
v
′
= u
2
τ
1 −
y
δ
!
(1.27)whi h inwallunitsbe omes
dU
+
dy
+
− u
′
v
′ +
=
1 −
y
+
δ
+
!
(1.28)wherethequantities
y
+
and
δ
+
ares aled bytheinner(vis ous)lengths ale
y
+
def=
y
l
∗
=
yu
τ
ν
Considerations. Dependingontheregion onsidered,equation1.28assumesdierentforms. Intheouterlayer,wherevis ous ee tsarenegligible,theleft-hand-sideof1.28be omes
− u
′
v
′ +
= 1 −
y
+
δ
+
(1.29)Ontheotherhand, loseto thewall,
y/δ << 1
,vis ous ee tswill benotnegligibleanymore,thereforethefollowingrelationholds
dU
+
dy
+
− u
′
v
′ +
≈ 1
(1.30)thatis ompatiblewiththelawofthewall,sin ethere'snotinuen e oftheReynoldsnumber.
It'sa onstantstress region:
U
+
def=
U
u
τ
= Φ
1
(y
+
)
(1.31)u
′
v
′ +
= Φ
2
(y
+
)
(1.32)Moreover,for largeReynoldsnumbers,we analsoassume that thereis an overlap region for
walldistan es
y
,ℓ
∗
<< y << δ
where
δ
isboundarylayerthi knessandℓ
∗
thevis ous lengths ale. This isaparti ularregionwhere outerandinnerlayerdes riptionshold simultaneously.
Derivatingthefollowing1.31w.r.t.
y
+
we'llhaveanexpressionwhi hisindependentoflengths ale.
Thereforeholds
y
+
dΦ
1
(y
+
)
dy
+
≡ const
(1.33)sothat,on eintegrated,itgivesalogarithmi law:
Φ(y
+
) =
1
κ
lny
+
+ B
(1.34)
where
κ = 0.38
istheKármán onstantandB=4.1,a ordingtoobservations. Inthefollowinggure6.22ameanvelo itydiagramisshown,togetherwiththelawof the wall andtheloglaw, at
PSfragrepla ements
y
+
hui
+
10
0−
2
10
−
1
10
0
10
1
10
2
10
3
10
4
5 10 15 20 25 30 Figure1.6:< u >
+
vsy
+
,at
Reτ
= 5200
,from[LeeandMoser,2014℄,.:lawofthe wall,:loglawAswe anseefromthepreviouspi ture,therstvelo ityregion alledvis oussublayer,follows
thelawofthewall,anditextendsoutto approximately
y
+
= 5
. Despite theabsolutemagnitude
of the turbulent u tuations are small in this region, the relative (wall-parallel) intensities are
large. Aswein rease
y
+
,wewillhaveabuer region, wherethemaximumturbulen eprodu tion
is at
y
+
= 12
and the maximum turbulen e intensity at
y
+
= 15
. Log-layer starts between
50 < y
+
< 200
, andit extendsto
y/δ ≈ 0.15
, whereδ
isthe hannel half-width. Beyondthe loglayer,there'snallytheouterregion.
The maximumturbulen e produ tion and Reynolds stress.
Theturbulen eprodu tioninthenear-wallregionofwall-bounded ows anbeformulatedas
P
+
= −
u
′
v
′
u
2
τ
dU
+
dy
+
≃
1 −
dU
+
dy
+
!
dU
+
dy
+
(1.35)Thelatterapproximationis validifvis ousee tsare small. Fromthis we anderivethatthe
maximumprodu tionisfoundwhere
dU
+
dy
+
=
1
2
(1.36) whi h leadstoP
max
+
=
1
4
(1.37)Thereforetheturbulentprodu tionismaximumwhereasbothvis ousandReynoldsstressesare
exa tlythesame,thatisinthenear-wallregion;itgenerallyhappenswhen
y
+
≈ 12
alsopossibletoestimatewherethemaximumReynoldsshearstresso urs. Derivingequation1.28
w.r.t.
y
+
, and denotingthe Reynoldsshear stress as
τ
, normalized by inner units, the followingrelationholds:
d
2
U
+
dy
+
2
+
dτ
+
dy
+
= −
1
δ
+
(1.38)ifwealsoassumethattheReynoldsstressismaximumin thelog-region,taking thederivative
1.34w.r.t
y
+
wehaved
2
U
+
dy
+
2
= −
1
κy
+
2
(1.39)meaningthat themaximumoftheReynoldsstressisat
y
max
+
=
r
h
+
κ
=
r
Re
τ
κ
(1.40)Then,intermsofouters alewe ansaythatthepositionwheretheReynoldsstressrea hesits
maximumisproportionalto
Re
−
1/2
τ
:y
max
h
= κ
−
1/2
Re
−
1/2
τ
∝ Re
−
1/2
τ
(1.41)Thisalso anbeprovedusingDNSresultsatdierentfri tionReynoldsnumbers. Anexample
ofReynoldsstressprolesis showningure1.7. PSfragrepla ements
y/δ
−
u
′
v
′
u
2
τ
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 Figure1.7−u
′
v
′
normalizedby
uτ
vsy/δ
,from[LeeandMoser,2014℄,· · ·
:Reτ
= 550
,Themaximum Reynoldsshear stress anbe derived aswell, using the relation1.38and 1.40, togetherwiththelog-law:
τ
max
+
= 1 −
y
+
max
h
+
−
1
κy
max
+
= 1 −
2
√
κh
+
= 1 −
2
√
κRe
τ
.
(1.42)Wealreadyhavetalkedaboutturbulen eanditsprodu tioninaqualitativesense,but,inorder
todes ribeitproperlyandtounderstandthephenomenonina ompleteanddetailedsense,afo us
on the three omponents of thevelo ity, theyare
u
′
,
v
′
and
w
′
, is ne essary. As we'll see in the
next hapters,the omputationsofthese omponentsplayanimportantrolein thevalidation ofa
LESmodel.
Inparti ular,it'spossibletoderivetheReynoldsstresstensormakingthesquareofthe
root-mean-squarevalueofthesethree omponents:
u
′
i
u
′
j
=
u
′
u
′
u
′
v
′
u
′
w
′
u
′
v
′
v
′
v
′
v
′
w
′
u
′
w
′
v
′
w
′
w
′
w
′
.
(1.43)Inthefollowingpi tureallthe omponentsoftheReynoldsstressareshown. Importantto
un-derlineistherelationbetweenthemaximumturbulen eprodu tion,andthepeaksintheReynolds
stressproles. PSfragrepla ements
y
+
0 5 10 15 20 25 30 35 40 45 50 0 1 2 3 4 5 6 7 8 9Figure1.8Varian esofthe velo ityu tuations, normalizedinwall units,fromDNS of
[LeeandMoser,2014℄at
Reτ
= 1000
,:u
′
u
′
+
,:v
′
v
′
+
,:w
′
w
′
+
'Anisotropy of the ow. A fundamental aspe t, whi h is stronglyrelevantfor the omplete
understandingoftheaforementionedwork,is the on eptofanisotropyin turbulen e.
Fromaphysi alpointof view,anisotropyis aproperty of theowwhi h is notalignedwith the
velo itydire tion andthevelo itygradient.
Forafully-developed hannelowthequantities
u
′
w
′
and
v
′
w
′
= 0
arezero. TheReynoldsstress
tensorbe omes:
τ
ij
= −ρ
τ
11
u
′
v
′
0
u
′
v
′
τ
22
0
0
0
τ
33
.
(1.44)in ase ofisotropi turbulen e, the diagonal termsare equal, i.e.
τ
11
= τ
22
= τ
33
andall thedeviatori termsarezero. Therefore,foraspe i LESmodel,theabilityto apturetheanisotropy
oftheow onsistsin reprodu ingdierentdiagonalterms,i.e.
u
′
u
′
6= v
′
v
′
6= w
′
w
′
.2
LARGE EDDY SIMULATION
Anxietyisthe handmaiden of reativity. (T.S.Eliot)
Reynolds-averaged Navier Stokes (RANS) equations-based simulations are able to solve only
themeanvelo ityeld oftheow. RANSsimulationsrelyheavilyonmodellingsin eallturbulent
motionsaremodelledandthereforetheyarenotalwaysa urate.
InDire tNumeri alSimulations(DNS), theunsteadyNavier-Stokesequationsaresolvedwithout
using models. Therefore, DNS is very a urate. However, the omputational powerdemand for
DNSistoolargeforindustrialappli ations. Therefore,anewmethodwhi h ombinesareasonable
owpredi tiona ura y withalimitedamount of omputational ost hasbeenstronglyrequired
in re entyears.
Large-eddysimulation(LES)representsanalternativethattsthoserequirements. Thisparti ular
methodemploysaseparationofs ales[Rasam,2014℄: alteringoperationde omposesthevelo ity eld (generallytogether withas alarone) intoaresolvedpart,representedbythe omputational
grid,andanunresolvedpart,whi hismodelledthroughphysi allyrealisti models. Inotherwords,
1. thesolutionofthelarges alesofturbulen eonarelatively oarsegrid;
2. themodelling ofthe smallerunresolveds ales, the so- alledsubgrid-s ales (SGS), basedon
theresolvedvelo ityeld.
2.1 The ltering operation
Step1. Let's onsidernowageneraltimeand spa e-dependentfun tion
φ(x, t)
. Thelteringoperation onsistsofa onvolutionofakernel
G
∆
onthatfun tion(Leonard,1975),overthedomainD
ofthegrid:e
φ(x, t) =
Z
D
φ(x, t)G
∆
(x − ξ)dξ
(2.1)There are several optionsfor the lter: the ommonly used ones are spe tral uto, box and
Gaussianlters. Inspe tral odes,spe tralltersarethemostsuitableonesforltering,sin ethey
a t onaspe tralspa e(where the odeworkson)and aremorepre ise. Figure below showsthe
dieren ebetweenthedierentkindsoflteringmethods.
Figure2.1.: Sharp-spe trallter,-: Gaussianlter, : Boxlter,
r = ξ
,from [Pope,2000℄eld
e
u(x, t)
is notadeterministi variable, implyingthateeu(x, t) 6= eu(x, t), eu
′
(x, t) 6= 0.
2.2 Governing equations of LES and the losure problem
Thelteringoperationleadsto thelteredNavier-Stokesequations:
∂
e
u
i
∂x
i
= 0;
∂
e
u
i
∂t
+
∂
e
u
i
u
e
j
∂x
j
= −
∂
p
e
∂x
i
+
1
Re
∂
2
e
u
i
∂x
j
∂x
j
−
∂τ
SGS
ij
∂x
j
(2.2)here
τ
ij
isthe SGS stress tensor. Leonard[Leonard,1974℄ proposed a possiblede omposition ofnon-linearterms(i.e. theSGSstresstensor),inthefollowingway:τ
SGS
ij
=
u
g
i
u
j
− e
u
i
u
e
j
= L
ij
+ C
ij
+ R
ij
(2.3) whereL
ij
= g
u
e
i
u
e
j
− e
u
i
e
u
j
,
(2.4)C
ij
= g
e
u
i
u
j
′
− g
u
e
j
u
′
i
(2.5)R
ij
= g
u
′
i
u
′
j
,
(2.6) andu
′
i
= u
i
− e
u
i
.In the Leonard de omposition
R
ij
is the Reynoldssubgrid tensor and represents the intera tionbetweensubgrid-s ales;
C
ij
is the ross-stresstensor anda ountsforlargevssmall s aleintera -tions;nally
L
ij
,theLeonardtensor,givestheintera tionsbetweenthelarges ales.τ
SGS
ij
is theunknown additivepartin theltered Navier-Stokesequation, and thereforeneeds to bemodelled.2.3 An exampleofa losure: theEddy Vis osity Model (EVM)
Step2. Thereareseveralwaystomodelthesubgrids alestresstensor. Aspreviouslywritten,
the rstandsimplemodel wasdevelopedby Smagorinsky[Smagorinsky,1963℄, formeteorologi al appli ations. ItoriginatesfromtheRANSmodel,takingintoa ountBoussinesq'shypothesis. The
EVM onsistsofalinearformulationofthedeviatori partof
τ
SGS
ij
,τ
ij
SGS
=
τ
kk
3
δ
ij
| {z }
isotropic
−2
ν
SGS
| {z }
EV M
contribution
e
S
ij
(2.7) whereν
SGS
= (C
s
∆)
e
2
| e
S|
(2.8)e
S
ij
and| e
S|
are theresolved strain-ratetensorand its magnitude, respe tively.∆
e
is thelter s ale,ν
SGS
istheSGSeddyvis osity,andC
s
isthemodel oe ient,theSmagorinsky oe ient.Notethat
ν
SGS
isa onstantratherthandependentofdire tion. A ordingtothere entpaperof Spalart [Spalart,2015℄, in a simple shear ow with two of the axes aligned with the velo ity dire tion and thegradientdire tion,su hthat thestraintensor
S
ij
havezeroandequaldiagonalterms, this model predi ts onstant and equal diagonal terms of the Reynolds stress tensor, i.e.
τ
11
= τ
22
= τ
33
.ThereforeEVM anbe onsideredanisotropi model,sin e
ν
SGS
doesn'ttakeintoa ountanyee t ofanisotropy.
As we'll dis uss further, this model an be improved using a dynami pro edure, where the
C
s
oe ientis omputedduringthesimulation.
2.4 Subgrid-s ale dissipation
Animportantaspe t ofLESistheimpa tofSGS ontheresolveds ales. Inotherwords,LES
predi tionsarestronglydependentonthe ontributionof
τ
SGS
ij
tensorontheresolvedkineti energyK =
u
e
i
u
e
j
/2
. Considerthekineti energyequation∂K
∂t
+
∂
∂x
j
(
e
u
j
K)
|
{z
}
advection
= − ν
∂
∂x
e
u
i
j
∂
u
e
i
∂x
j
|
{z
}
viscous
dissipation
−
∂x
∂
i
u
e
i
p + ν
e
∂K
∂x
i
− e
u
i
τ
ij
!
|
{z
}
dif f usion
+ τ
ij
SGS
S
e
ij
| {z }
−
SGS
dissipation
(2.9)Sin eLESbydenition,are arriedoutforlarges ales,andthegrids aleismu hlargerthan
Kolmogorovone,the vis ous dissipationterm is sosmall so that it is negligible ompared to the
others. Fromaphysi alpointofview,diusiontermtransfersenergyinspa e,butnotina
volume-averagedsense [Rasam,2014℄. Forthatreason,anadditionaltermisthereforerequiredinorderto reprodu ethe orre tenergy transferfrom thelargeto thesmallers ales. SGS dissipation overs
Π = −τ
SGS
ij
S
e
ij
(2.10)ThemeanSGS dissipationbehavesasasink term, whiletheinstantaneousonegivesnegative
(ba ks atter)andpositive(forwards atter) ontributionsinthetransfer.
The following pi ture showsthe SGS dissipation (red line), together with the dissipation of the
resolveds ales(blue line);theirtotal ontributionisshownbytheredline.
PSfragrepla ements
y
+
0 -100 -200 -300 -400 -500 -600 -0.25 -0.2 -0.15 -0.1 -0.05 0 0Figure2.2Dissipationoftheresolveds ales,
ε
,togetherwiththeSGSdissipation,Π
,and theirsumatin reasingresolutions,fromLESatReτ
= 590
shownin hapter6.Theprevious onsiderationleadstothefa tthataninvestigationonSGSdissipationbehaviour
isveryimportantinana ura yassessmentofLES.
Basedonthisprin iple,severalinvestigationshavebeen arriedoutbyChow&Moin[ChowandMoin,2003℄, Ghosal[Ghosal, 1996℄andKrav henko&Moin[Krav henkoetal.,1996℄.
Geurts &Fröhli h [Geurtsand Fröhli h, 2002℄introdu ed theSGS a tivityparameter, dened as follows
s =
<ε
SGS
>
<ε
SGS
> + <ε
µ
>
(2.11) where<ε
µ
> = 2µ e
S
ij
S
e
ij
isthevis ousdissipation.With in reasingresolution, SGS dissipation de reaseswhile vis ous onein reases, therefore
s
be- omes smaller. A remarkable aspe t isthat, the oarser is theresolution , the biggeris
s
. Thisan be proved having alook at 2.11: if we use a oarse grid, the resolution will be smallersu h that vis ous dissipationwill benegligible, and
s
will rea h itsmaximumunity value. This o urssmall.
Starting from the previous denition, Geurts & Fröhli h [GeurtsandFröhli h,2002℄ dened an errornormas
δ
E
=
E
LES
− e
E
DN S
e
E
DN S
(2.12)where
E
LES
isthemeanresolvedkineti energyintegratedovertheowdomain,whileE
e
DN S
is the samequantity, omputed from ltered DNS datainstead, with the samelter width asin
theLES.
The a ura y assessment onsists of omputing the relation between
δ
E
ands
: in the pi turebelow[Rasametal.,2011℄ it is shown that therelativeerror
δ
E
drops almost exponentiallywith de reasings
(i.e. high resolution); results referto a LES with the expli it algebrai SGS stressmodel (EASSM) [Marstorpetal.,2009℄ at
Re
τ
= 934
, for six dierent resolutions. This result showsthat with in reasing resolution theSGS ontribution be omessmaller and thea ura y ofthe LES higher. Inother words,the resolution of the LES mustbe su iently high to obtain an
a eptable solution.
Figure2.3Normalizederror
δE
oftheresolvedkineti energy,integratedoverthe hannel width[Rasametal.,2011℄w.r.t. the ltered DNSvalue vss
. Arrow points toin reasing resolution ases.3
SUBGRID-SCALE STRESS MODELS
FOR LES
Logi willgetyoufrom AtoB.
Imagination willtake youeverywhere. (A.Einstein)
Inthis haptertwodierentLESmodelsaregoingtobeanalyzed;therstoneisanimprovement
ofthe lassi Smagorinskymodel, theDynami EVM.
These ondoneisaparti ularnon-linearmodelusinganexpli italgebrai formulationfortheSGS
stresstensor.
Inthefollowingse tionade omposition oftheSGS stresstensorisgiven.
3.1 Tensorialpolynomial formulation of the SGS stresstensor
A usefulapproa h that leadsto several non-linear models is basedon ade omposition of the
SimilartoRANSapproa hforReynoldsstress losures,SGSstresstensor anbeexpressedinterms
ofstrainandrotationratetensors.
Lund &Novikov[LundandNovikov,1993℄ expressedthedeviatori partof theSGS stresstensor
τ
ij
SGS,d
asageneraltensorialfun tion of theltered strain-rateS
e
ij
androtation-rateΩ
e
ij
tensors, theKrone kerdeltaδ
ij
andtheltersize∆
e
asτ
ij
SGS,d
= τ
ij
SGS
−
1
3
τ
kk
δ
ij
= f ( e
S
ij
, e
Ω
ij
, δ
ij
, e
∆)
(3.1) wheree
S
ij
=
1
2
∂
e
u
i
∂x
j
+
∂
e
u
j
∂x
i
!
,
Ω
e
ij
=
1
2
∂
e
u
i
∂x
j
−
∂
e
u
j
∂x
i
!
.
(3.2)AsPope[Pope,1975℄showedtheReynoldsstressesinRANS,
τ
ij
anbeformulatedasatensor polynomialwithtenelementsofdierentspowersofS
e
ij
andΩ
e
ij
andtheir ombination.Coe ientsarefun tions of
S
e
ij
andΩ
e
ij
invariantsorboth,andtheyarederivedusingtheCayley-Hamiltontheorem:
τ
d
=
10
X
k=1
β
k
T
(k)
,
(3.3)sothatthetenpolynomialtensorsare:
T
(1)
= e
S
T
(2)
= e
S
2
−
1
3
II
S
I
T
(3)
= e
Ω
2
−
1
3
II
Ω
I
,
T
(4)
= e
S
Ω
e
− e
Ω
S
e
T
(5)
= e
S
2
Ω
e
− e
Ω
S
e
2
T
(6)
= e
S
Ω
e
2
+ e
Ω
2
S
e
−
2
3
IV I
T
(7)
= e
S
2
Ω
e
2
+ e
Ω
2
S
e
2
−
2
3
V I
T
(8)
= e
S
Ω
e
S
e
2
− e
S
2
Ω
e
S
e
T
(9)
= e
Ω
e
S
Ω
e
2
− e
Ω
2
e
S
Ω
e
T
(10)
= e
Ω
e
S
2
Ω
e
2
− e
Ω
2
S
e
2
Ω
e
and
β
k
ares alar oe ientsthat arefun tionsofvetensorialinvariantsII
S
= tr(e
S
2
),
II
Ω
= tr( e
Ω
2
),
V = tr(e
S
2
Ω
e
2
).
Thus, equation3.3isthemostgenerlaformulationfor
τ
d
in termsof
S
˜
ij
andΩ
˜
ij
.3.2 The dynami pro edure for EVM (DEVM)
Regardingto themodel oe ient
C
s
forthe eddy-vis osity model, it hasbeendemonstratedthat adynami omputation of
C
s
(that is, during the simulation), an signi antly improvetheowpredi tiona ura y.
The omputation of
C
s
is done by taking into a ount the resolved s ales, a ordingto a s aleinvarian eassumption.
The ru ialpointofthisdynami pro edure istheso- alledGermanoidentity. Let'snowdenotea
testlter,
∆ = 2 e
b
∆
,∆ =
e
3
√
Ω
,whereΩ
isthevolumeofa omputational ell. TheGermanoidentity isdened asfollows:L
ij
= T
ij
− b
τ
ij
,
(3.4)where
T
ij
isthe SGSstress tensorlteredat thetest lterlevel,andL
ij
= d
e
u
i
e
u
j
− be
u
i
beu
j
ThenL
ij
isappliedin thisway:L
ij
−
1
3
L
kk
δ
ij
= −2C
s
M
ij
(3.5)where
M
ij
= b
∆
2
|be
S|be
S
ij
− e
∆
2
| e
\
S| e
S
ij
.
(3.6)Thesystemofequations3.5isover-determined. Inordertohaveauniquevalueof
C
s
Germano ontra teditin:C
s
= −
1
2
hL
ij
S
e
ij
i
h e
S
ij
M
ij
i
(3.7)Moreover,tomake
C
s
variationssmoother,aspatial averagingh.i
hasbeenapplied.Togetherwithabetterperforman eof theEVM,theGermanoidentitygivesa orre tasymptoti
near-wallbehaviourfor
ν
SGS
.Ithasalsobeenshownthatispossibleto applyGermanoidentityforthedynami omputationof
3.3 Expli it Algebrai SGS stress model
Anisotropi ee ts of turbulen e are important in several onditions: examplesare near-wall
owbehaviourandboundarylayesseparation aseswith urvature,swirl,rotation.
Asitwasdis ussedbefore,theDynami Smagorinskymodelisanisotropi model,inthesensethat
theSGSvis osity
ν
SGS
isdire tion-independent.Therefore,in orderto aptureowanisotropy,DEVM isnotsuitable.
InthesamespiritasReynoldsstress-basedmodelsarene essaryforRANS,non-linearSGSmodels
areneededforLES.
A re ent example of those is the nonlinear dynami SGS stress model by Wang & Bergstrom
[WangandBergstrom,2005℄, whi h onsists of three base tensorsand three dynami oe ients.
OneofthetermsinthemodelissimilartotheDEVMmodel. Wang&Bergstrom[Wangand Bergstrom,2005℄ showedthatthedynami non-linearmodelpredi tsamorerealisti tensorialalignmentoftheSGS
stressthaneddy-vis ositymodelsand anprovideforba ks atterwithout lipping oraveragingof
thedynami modelparameters.
The model here dis ussed is alled Expli it Algebrai SGS stress model (EASSM), was
devel-oped by Marstorp [Marstorpetal.,2009℄ and is similar to the EARSM by Wallin & Johansson [WallinandJohansson,2000℄, whi h is based on a modelled transport equation of the Reynolds stressesandontheassumption thattheadve tionanddiusionof theReynoldsstressanisotropy
arenegligible.
AnalogoustotheReynoldsstressanisotropytensorwedenetheSGSstressanisotropytensoras
a
ij
=
τ
ij
K
SGS
−
2
3
δ
ij
(3.8)Forsimpli ity,we'll onsider
τ
SGS
ij
= τ
ij
. Moreover,K
SGS
= (
u
g
i
u
i
− e
u
i
e
u
i
)/2
istheSGSkineti energy.Inanon-rotatingframethetransportequationfor
a
ij
readsK
SGS
Da
ij
Dt
|
{z
}
advection
−
∂D
τ
ij
ijk
∂x
k
−
τ
ij
K
SGS
∂D
K
SGS
k
∂x
k
!
|
{z
}
dif f usion
= −
K
τ
ij
SGS
(P − ε) + P
ij
− ε
ij
+ Π
ij
,
(3.9) where−D
K
SGS
k
= −D
τ
ij
ijk
/2
is thesumoftheturbulent andmole ularuxes oftheSGS stress andSGS kineti energy,respe tively.Although the produ tion of the SGS stress