Tese Silo

KATHOLIEKE UNIVERSITEIT LEUVEN
FACULTEIT INGENIEURSWETENSCHAPPEN
DEPARTEMENT BURGERLKE BOUWKUNDE
Kasteelpark Arenberg 40, B-3001 Leuven
FLUID-STRUCTURE INTERACTION APPLIED TO
FLEXIBLE SILO CONSTRUCTIONS
Promotoren: Proefschrift voorgedragen tot
Prof. dr. ir. G. De Roeck het behalen van het doctoraat
Prof. dr. ir. G. Degrande in de ingenieurswetenschappen
door
David Dooms
Februari 2009
KATHOLIEKE UNIVERSITEIT LEUVEN
FACULTEIT INGENIEURSWETENSCHAPPEN
DEPARTEMENT BURGERLKE BOUWKUNDE
Kasteelpark Arenberg 40, B-3001 Leuven
FLUID-STRUCTURE INTERACTION APPLIED TO
FLEXIBLE SILO CONSTRUCTIONS
Jury: Proefschrift voorgedragen tot
Prof. dr. ir. D. Vandermeulen, voorzitter het behalen van het doctoraat
Prof. dr. ir. G. De Roeck, promotor in de ingenieurswetenschappen
Prof. dr. ir. G. Degrande, promotor
Prof. dr. ir. M. Baelmans door
Prof. dr. ir. J. Vierendeels
(Universiteit Gent) David Dooms
Prof. dr. ir. B. Blocken
(Technische Universiteit Eindhoven)
Prof. dr. rer. nat. M. Sch afer
(Technische Universitat Darmstadt)
U.D.C. 519.6:551.55:624.042.4:624.954
Februari 2009
Katholieke Universiteit Leuven Faculteit Ingenieurswetenschappen
Arenbergkasteel, B-3001 Leuven (Belgium)
Alle rechten voorbehouden. Niets uit deze uitgave mag worden vermenigvuldigd
en/of openbaar gemaakt worden door middel van druk, fotokopie, microlm,
elektronisch of op welke andere wze ook zonder voorafgaande schriftelke
toestemming van de uitgever.
All rights reserved. No part of the publication may be reproduced in any form by
print, photoprint, microlm or any other means without written permission from
the publisher.
D/2009/7515/19
ISBN 978-94-6018-036-1
Voorwoord
De afgelopen zeven jaren aan de afdeling bouwmechanica zn voor m bzonder
leerrk en verruimend geweest. Vandaag rolt het resultaat van al dit werk uit de
printer. Hoewel mn onderzoek vaak uit individueel werk bestond, droegen toch
vele mensen een steentje b. Ik wil hen dan ook graag bedanken voor hun bdrage.
Vooreerst gaat mn dank uit naar mn promotoren, Guido De Roeck en Geert
Degrande. Z boden m de mogelkheid om dit doctoraat aan te vatten en hebben
me gedurende de afgelopen jaren opgevolgd en ondersteund. Ze gaven me de
kansen en de vrheid om mn weg te zoeken binnen mn onderzoeksdomein en
stonden steeds met nuttige tips en raad klaar. Teksten voor publicaties werden
nauwgezet nagelezen. Met hun steun heb ik dit doctoraat tot een goed einde
gebracht. Ook op menselk vlak kon ik steeds b hen terecht.
Naast mn promotoren wil ik Tine Baelmans en Jan Vierendeels bedanken voor
hun nuttige inbreng als lid van mn begeleidingscommissie. Verder dank ik ook
Bert Blocken voor de interessante discussies en om deel uit te maken van mn
jury. I like to thank Michael Sch afer for giving me the opportunity to stay during
a short period at the TU Darmstadt and for his willingness to be a member of my
jury. Dirk Vandermeulen dank ik voor het vervullen van de voorzitterstaak.
Meer dan vf jaar heb ik de bureau gedeeld met Mattias Schevenels. De sfeer was er
steeds plezant en we hadden vele discussies over onderws, onderzoek, architectuur
of gewoonweg het leven zoals het is.
Ook de andere collegas van de afdeling bouwmechanica wil ik bedanken voor de
goede werksfeer: Danielle, Johan, Geert, Anne, Lincy, Serge, Ralf, Kathleen, Daan,
Stn, Edwin, Hamid, Shashank, Jaime,

Ozer, Ali, Kai, Eliz-Mari, Bram, Amin,
Saartje, Suzhen, Hans en de vele buitenlandse bezoekers. Ik beleefde vele leuke
momenten tdens het werk, op onze etstochtjes aan de zee, op de verschillende
bouwkunde-uitstappen en tdens de kerstfeestjes en etentjes met de afdeling. Met
plezier denk ik terug aan de meetcampagne met de hoogtewerker op de silo te
Antwerpen.
i
ii VOORWOORD
Ik heb gedurende de afgelopen jaren mogen deelnemen aan een aantal interessante
buitenlandse congressen en workshops, waarvan sptig genoeg slechts enkele in
gezelschap van collegas. De spannende momenten van het te korte congres op
Stantorini zal ik - en waarschnlk ook Geert en Stn - niet snel vergeten. Meer
ontspannen ging het eraan toe b onze verkenning van de steegjes van Venetie. Ook
ons bezoekje met de afdeling aan Geert in Pars behoort tot de beste herinneringen
die ik overhoud.
Ik wil ook graag enkele mensen bedanken die vooral onrechtstreeks bgedragen
hebben tot deze thesis. Mn ouders bedank ik voor de steun die ik altd heb
mogen ervaren tdens mn studies en gedurende mn doctoraat. Stan en Liese
verwelkomden m dagelks b mn thuiskomst van het werk enthousiast met een
portie glimlachjes: alle zorgen waren er op slag door vergeten! Kristien stond
steeds voor m klaar. Jouw bezorgdheid, steun en stimulans zn voor m goud
waard.
David Dooms,
Leuven, februari 2009
Fludum-structuur interactie
toegepast op exibele
siloconstructies
Inleiding
Het gedrag b windbelasting van bouwkundige constructies die vr exibel zn,
een complexe vorm hebben of in de nabheid van een ander gebouw liggen, is
moeilk te bepalen. Het gebrek aan kennis over wind-structuur interactie-eecten
heeft in het verleden tot een aantal rampen geleid: in 1940 stortte de Tacoma
Narrows brug (Washington, US) in ten gevolge van utter; in 1965 stortten drie
van een groep van acht koeltorens in te Ferrybridge (UK).
Tdens de storm op 27 oktober 2002 traden ovaliserende trillingen op b
verschillende lege silos van een groep van veertig silos in de haven van Antwerpen
(guur 1). De verplaatsingen waren naar schatting van de grootte-orde van
meerdere centimeters. Het ovaliseren is een aero-elastisch fenomeen waarb de
negatieve aerodynamische demping de structurele demping opheft. Gelkaardige
voorvallen in Duitsland en Schotland duiden erop dat de stormschade zich vooral
op de hoeken van de groepen voordoet.
Om deze fenomenen te bestuderen is vandaag de dag de gebruikelke aanpak een
reeks windtunneltesten uit te voeren. Het doel van deze thesis is om gekoppelde
numerieke simulaties te gebruiken om dit gedrag te voorspellen. De verschillende
numerieke methodes die toegepast worden in de verschillende velden (udum
en structuur) en in de koppelingsprocedure worden verkend. Aangezien deze
thesis binnen de onderzoeksgroep het eerste werk vormt dat gebruik maakt van
Computational Fluid Dynamics (CFD) en Fluid-Structure Interaction (FSI), is een
van de objectieven om in ieder hoofdstuk een uitgebreid en coherent overzicht van
de desbetreende theorie te geven.
iii
iv SAMENVATTING
Figuur 1: De silo groep in de haven van Antwerpen.
Dit werk focust op de simulatie van wind-structuur interactiefenomenen die
optreden b schaalconstructies. Het ovaliseren van de silos zal als voorbeeld
doorheen de tekst gebruikt worden. Voor dit probleem worden nauwkeurige
modellen voor het udum en de structuur opgesteld en ovaliserende trillingen
voor een silo gesimuleerd. In toekomstig werk kan de invloed van de smalle
tussenafstand tussen twee naburige silos op het optreden van ovaliserende
trillingen onderzocht worden en veranderingen aan het ontwerp voorgesteld worden
om het optreden van het fenomeen in de toekomst te vermden.
Methodologie
De methodes om wind-structuur interactiefenomenen die optreden b schaal-
constructies te simuleren worden gentroduceerd aan de hand van het voorbeeld
van door de wind genduceerde ovaliserende trillingen van silos. Dit
dynamisch udum-structuur interactieprobleem wordt voldoende representatief
beschouwd voor de beoogde problemen. Hoewel het een eenvoudige cilindrische
geometrie betreft, komen alle specieke aspecten van een wind-schaalstructuur
interactieprobleem aan bod: geometrisch niet-lineaire vervormingen van een
schaalstructuur, een turbulente windstroming met separatie en tenslotte het zelf-
geexciteerd fenomeen dat slechts numeriek bestudeerd kan worden met methodes
die de interactie tussen het udum en de structuur in rekening brengen. De
gebruikte methodologie is algemeen en kan eveneens toegepast worden op de
andere voorbeelden uit de inleiding.
Voor de numerieke simulatie van udum-structuur interactie wordt een
gepartitioneerde methode gebruikt: het udum en de structuur worden
SAMENVATTING v
beschouwd als gesoleerde velden die afzonderlk opgelost worden. De
interactie-eecten worden aan de individuele velden opgelegd als uitwendige
randvoorwaarden die uitgewisseld worden tussen de velden. Het voornaamste
voordeel van deze gepartitioneerde aanpak is dat gevestigde discretisatietechnieken
en oplossingsalgoritmes (en bgevolg bestaande software) ingezet kunnen worden
in de individuele velden. Deze technieken kunnen in hoge mate aangepast zn
aan het karakteristieke gedrag van dit veld. Aangezien deze thesis binnen de
onderzoeksgroep het eerste werk vormt dat gebruik maakt van CFD en FSI, is
er vanaf het begin bewust voor gekozen om gebruik te maken van bestaande
software: Ansys (Ansys, 2005b) voor de structuurberekening en Flotran (Ansys,
2005b) en CFX (Ansys, 2005a) voor de stromingsberekeningen. Bgevolg waren de
keuzes wat betreft algoritmes en modellen beperkt tot wat beschikbaar is in deze
programmas (appendix A). Aangezien de stromingsvergelkingen (bv. de Navier-
Stokes vergelkingen) niet-lineair zn en het niet-lineair gedrag van de structuur
in rekening gebracht moet worden, wordt directe tdsintegratie gebruikt in beide
velden.
Alvorens de studie van het udum-structuur interactiefenomeen aan te vatten, is
een grondige kennis van de individuele velden, nl. udum en structuur, nodig.
Nauwkeurige modellen zn vereist voor deze velden.
De structuur wordt gediscretiseerd aan de hand van de eindige-elementen methode
en gentegreerd in de td met het Newmark schema. De formulering brengt
geometrisch niet-lineair gedrag in rekening en laat grote rotaties en grote
verplaatsingen toe. Vermits de reele waarde van de structurele modale demping
van groot belang is voor het voorspellen van ovaliserende trillingen, is een in
situ experiment uitgevoerd om deze te bepalen. De gemeten eigenfrequenties en
modevormen worden gebruikt om een drie-dimensionaal eindige-elementenmodel
van de silo te valideren.
Voor het udum wordt een eindige-volumediscretisatie toegepast. De oplossing
wordt gentegreerd in de td met de drie-punts-achterwaartse dierentiemethode.
De twee-dimensionale stroming rond een starre cilinder wordt berekend. Het eect
van de groepsopstelling op de stroming wordt bestudeerd voor een groep starre
cilinders zoals in de haven van Antwerpen.
Tdens de FSI berekeningen wordt het rooster van het udum gealigneerd met de
vervormende rand van de structuur. De stromingsberekeningen worden uitgevoerd
op een vervormend rooster en de vergelkingen worden daarom geformuleerd in een
arbitraire Lagrangiaanse-Euleriaanse (ALE) beschrving. De vervorming van het
rooster wordt berekend aan de hand van een diusievergelking met een variabele
diusiviteit en moet er b voorkeur voor zorgen dat de kwaliteit en de verjningen
van het rooster behouden blven.
Tussen Ansys en Flotran, en Ansys en CFX is een koppelingsalgoritme commercieel
beschikbaar. Twee gepartitioneerde algoritmes worden gebruikt: een zwak
vi SAMENVATTING
gekoppeld serieel staggered algoritme dat elk veld een maal oplost per tdstap
en een sterk gekoppeld algoritme dat itereert tussen de twee velden binnen een
tdstap. De roosters van het udum en de structuur zn niet samenvallend aan de
interface. Een consistente interpolatie wordt gebruikt voor de overdracht van de
verplaatsingen, terwl een knoop-naar-element conservatieve methode toegepast
wordt b de overdracht van de belastingen. De interactie tussen de windstroming
en een silo wordt berekend.
Originele bdragen van de thesis
Een drie-dimensionaal eindige-elementenmodel van een cilindervormige silo is
gevalideerd aan de hand van in situ metingen op een silo. De radiale versnellingen
ten gevolge van de heersende windbelasting zn gemeten in 10 punten op de
silo. Aan de hand van de stochastische systeemidenticatietechniek zn de
modale parameters bepaald op basis van uitgangssignalen. De invloed van de
randvoorwaarden op de eigenfrequenties en modevormen is onderzocht.
Er wordt een overzicht gegeven van de huidige state-of-the-art methodes voor de
simulatie van turbulente stromingen in functie van hun performantie en de vereiste
rekenkracht. De invloed van turbulentiemodellen, wandmodellen en niet-stationair
gedrag is bestudeerd voor de twee-dimensionale stroming rond een cilinder in het
postkritische regime. De resultaten worden vergeleken met experimentele data en
numerieke resultaten uit de literatuur. De twee-dimensionale stroming rond een
groep van 2-b-2 en van 8-b-5 cilinders toont dat er een grote invloed is van de
groepsopstelling op de drukverdeling rond de cilinders.
Er wordt een coherent en uitgebreid overzicht gegeven van alle technieken vereist
voor de simulatie van dynamische udum-structuur interactieproblemen. De
vergelkingen voor de stroming in een Lagrangiaanse-Euleriaanse beschrving
zn afgeleid. De gevolgen van deze beschrving voor de nauwkeurigheid en de
stabiliteit van de tdsintegratie worden besproken. De verschillende methodes om
de snelheid van de roosterpunten te berekenen worden toegelicht. Het gekoppelde
probleem wordt opgelost via een gepartitioneerd algoritme. In het geval van
onsamendrukbare stromingen worden de nauwkeurigheid, stabiliteit en ecientie
van zwak en sterk gekoppelde algoritmes besproken. De verschillende methodes
voor de overdracht van krachten en verplaatsingen tussen niet-samenvallende
roosters worden geevalueerd aan de hand van hun nauwkeurigheid en van de mate
waarin de behoudswetten gerespecteerd worden.
Als praktische toepassing wordt de interactie tussen de windstroming en een
cilinder berekend. Het drie-dimensionale eindige-elementenmodel van de silo is
gekoppeld met de drie-dimensionale onsamendrukbare turbulente stroming om
ovaliserende trillingen te voorspellen.
SAMENVATTING vii
Structureel gedrag van de silo
Het structureel gedrag van de silo in de haven van Antwerpen wordt bestudeerd.
De silos zn cilindrische schaalconstructies met een diameter van 5.5 m en een
hoogte van 25 m. Een cilinder bestaat uit 10 aluminium platen met een hoogte
van 2.5 m en een dikte die afneemt met de hoogte van 10.5 mm aan de onderkant
naar 6 mm aan de bovenkant.
Vermits de reele waarde van de structurele modale demping van groot belang
is voor het voorspellen van ovaliserende trillingen, is in situ een experiment
uitgevoerd om deze te bepalen. Om de meetopstelling af te stemmen op de
te verwachten eigenmodes en eigenfrequenties werden deze vooraf berekend met
een harmonisch eindige-elementenmodel. De modevormen bestaan uit m halve
sinussen over de hoogte en n sinussen over de omtrek. Terwl b balken en platen
de complexiteit van de modevormen toeneemt b stgende eigenfrequenties, is dit
hier niet het geval. De laagste eigenfrequentie wordt namelk teruggevonden b
m = 1 en n = 4.
De radiale versnellingen ten gevolge van de heersende windbelasting werden
gemeten in 10 punten op de silo. Aan de hand van de stochastische systeem-
identicatietechniek zn de modale parameters (eigenfrequenties, modevormen
en modale dempingsfactoren) bepaald. Er werden 60 eigenmodes met een
eigenfrequentie lager dan 20 Hz gedenticeerd. De modale dempingsfactoren
varieren tussen 0.07 % en 1.32 %, wat realistische waarden zn voor een gelaste
aluminium structuur. De eigenmode met de laagste eigenfrequentie (3.94 Hz,
m = 1, n = 3 of 4) heeft de grootste bdrage in de gemeten respons. Figuur 2
toont een bovenaanzicht en een drie-dimensionaal zicht van drie gedenticeerde
eigenmodes.
De laagste eigenfrequenties van de silo worden sterk overschat door het harmonisch
eindige-elementenmodel. De eigenfrequenties en modevormen van een cilindrische
schaalstructuur zn heel gevoelig voor de randvoorwaarden opgelegd aan de axiale
verplaatsingen u
z
, terwl de invloed van de randvoorwaarden voor de rotaties
bna verwaarloosbaar is (Forsberg, 1964; Koga, 1988; Leissa, 1993). Het

harmonisch eindige-elementenmodel belemmert de axiale verplaatsingen onderaan
over de hele omtrek, terwl in realiteit de silo slechts in vier punten over de
omtrek vastgebout is aan een achthoekige randbalk. Een drie-dimensionaal
eindige-elementenmodel van de silo maakt het mogelk de verbindingen met de
draagstructuur correct in te rekenen. Figuren 3b tot 3f tonen een bovenaanzicht en
een drie-dimensionaal zicht van vf berekende modevormen. Het bovenaanzicht
toont dat eigenmodes (1,5) en in mindere mate ook (1,3) een combinatie zn
van ovaliseren en van globale buiging van de silo. De eigenmodes met drie en
vier sinussen over de omtrek en een halve sinus over de hoogte hebben beide een
eigenfrequentie van 3.93 Hz.
viii SAMENVATTING
8
15 7
9
16
6
10
11 12
13
6
15 7
16
8
9
10
11
12
13
7
6
16
8
15
9
11
10
12
13
a. Mode 3
b. Mode 4
c. Mode 53
f = 4.00 Hz f = 4.01 Hz f = 17.95 Hz
= 0.77 % = 0.81 % = 0.10 %
Figuur 2: Bovenaanzicht en drie-dimensionaal zicht van het reeel (volle ln) en
het imaginair deel (streepln) van de gedenticeerde eigenmodes 3, 4
en 53 van de silo. De stippelln duidt de onvervormde silo aan.
De invloed van geometrisch niet-lineair gedrag op vervormingen van grootte-orde
0.1 m (zoals gedurende de storm) is niet verwaarloosbaar: ten opzichte van een
lineaire berekening worden de verplaatsingen groter waar de kromming vermindert
en kleiner waar de kromming toeneemt. Het geometrisch niet-lineair gedrag wordt
b de verdere berekeningen in rekening gebracht.
Turbulente windstromingen rond cilinders
Windstromingen rond gebouwen worden door heel grote Reynoldsgetallen
gekenmerkt. Om de rekentd op een processor beperkt te houden, worden
in dit werk enkel niet-stationaire Reynolds gemiddelde Navier-Stokes (RANS)
berekeningen toegepast. Om de invloed van de turbulentiemodellen, de
wandmodellen en niet-stationair gedrag te verduidelken, wordt de twee-
dimensionale stroming rond een cilinder in het postkritische regime b een
SAMENVATTING ix
a.Model
b.(1, 3) c.(1, 4) d.(1, 5) e.(1, 6) f.(1, 2)
3.93 Hz 3.93 Hz 5.25 Hz 7.37 Hz 7.75 Hz
Figuur 3: Drie-dimensionaal eindige-elementenmodel van de silo en een aantal
geselecteerde eigenmodes.
Reynoldsgetal 12.4 10
6
berekend. De stroming rond een cilinder is een
ideale test omdat het verschillende stromingspatronen combineert zoals separatie,
recirculatie, stagnatie en kromming van de stroomlnen. Bovendien is het een
vereenvoudiging van de stroming rond een silo. De resultaten worden vergeleken
met experimentele data en numerieke resultaten uit de literatuur. In Flotran
benvloedt het al dan niet gebruiken van wandfuncties de resultaten in hoge
mate, terwl de resultaten bekomen met CFX hiervoor vr ongevoelig zn.
Resultaten berekend met verschillende turbulente-viscositeit-modellen vertonen
grote verschillen, maar de minimale drukcoecient is steeds onderschat en de
basisdrukcoecient steeds overschat. Voor de stroming rond een cilinder b deze
Reynoldsgetallen bekomt het Shih-Zhu-Lumley turbulentiemodel in Flotran de
beste overeenkomst met de experimentele resultaten. In CFX levert het shear
stress transport turbulentiemodel de beste resultaten op. Een niet-stationaire
berekening met het shear stress transport turbulentiemodel laat toe de regelmatige
wervelafscheiding in het postkritische regime te berekenen en wordt daarom
verkozen boven een stationaire.
x SAMENVATTING
Figuur 4 vergelkt de drukcoecient van een stationaire berekening met de
tdsgemiddelde drukcoecient C
p
van een niet-stationaire berekening. De
maximale en minimale drukcoecient gedurende de niet-stationaire berekening
zn eveneens aangeduid. Vergeleken met de stationaire resultaten, voorspelt
het tdsgemiddelde van de niet-stationaire resultaten een lagere minimum
drukcoecient C
min
p
en een lagere basisdrukcoecient C
b
p
. Vermits het
tdsgemiddelde van de niet-stationaire berekening duidelk verschilt van de
stationaire oplossing, is de niet-stationaire berekening fysisch gezien meer correct
en biedt als voordeel extra informatie over de tdsvariatie van de druk (Iaccarino
et al., 2003). Daarom worden enkel niet-stationaire berekeningen uitgevoerd voor
de stroming rond een groep cilinders.
0 45 90 135 180
2.5
2
1.5
1
0.5
0
0.5
1
Angle []
P
r
e
s
s
u
r
e

c
o
e
f
f
i
c
i
e
n
t

[
]
(a)
0 45 90 135 180
2.5
2
1.5
1
0.5
0
0.5
1
Angle []
P
r
e
s
s
u
r
e

c
o
e
f
f
i
c
i
e
n
t

[
]
(b)
Figuur 4: Drukcoecient als functie van de hoek bekomen met een stationaire
berekening (streepln) en het tdsgemiddelde (volle ln), het minimum
(streep-stippelln) en het maximum (stippelln) van de drukcoecient
bekomen met een niet-stationaire berekening voor (a) rooster A
2
en (b)
rooster B. De lichtgrze zone bevat alle beschikbare experimentele data
b Reynoldsgetallen van 0.73 10
7
tot 3.65 10
7
.
In de haven van Antwerpen zn de silos opgesteld in vf ren van acht silos
met tussenruimtes van 30 cm tussen twee naburige silos. De twee-dimensionale
stroming rond een groep van 2-b-2 en van 8-b-5 dicht b elkaar staande cilinders
toont dat er een grote invloed is van de groepsopstelling op de drukverdeling rond
de cilinders. De resultaten voor de groep van 8-b-5 worden hier kort besproken.
Figuur 5 toont het tdsverloop en de frequentie-inhoud van de druk op het
oppervlak van een cilinder ter plekke van een smalle tussenruimte. Wervels worden
afgescheiden van de groep als geheel b 0.165 Hz en van de individuele cilinders
b 2.85 Hz. De eerste frequentie komt overeen met een Strouhalgetal van 0.24, als
de geprojecteerde breedte 45.89 m van de groep gebruikt wordt als karakteristieke
lengte, terwl de tweede frequentie een Strouhalgetal geeft van 0.49 als de cilinder
SAMENVATTING xi
diameter gebruikt wordt als karakteristieke lengte. Dit strookt met de hogere
Strouhalgetallen die gemeten worden in compacte buizenbundels (Blevins, 1990).
35 40 45 50
850
800
750
700
650
600
Time [s]
P
r
e
s
s
u
r
e

[
P
a
]
(a)
0 1 2 3 4 5
0
5
10
15
20
25
30
35
40
Frequency [Hz]
P
r
e
s
s
u
r
e

[
P
a
/
H
z
]
(b)
Figuur 5: (a) Tdsverloop en (b) frequentie-inhoud van de druk op het oppervlak
van een cilinder in de groep van 8-b-5 cilinders.
Figuur 6 toont de tdsgemiddelde drukcoecienten voor de groep. De streepln
geeft ter vergelking de drukcoecient rond een alleenstaande cilinder weer. De
groepsopstelling verandert de drukverdeling rond de cilinders drastisch. De
stuwpunten worden weg van de groep verschoven b cilinders 2 tot 8. De
zone met positieve druk verkleint b deze cilinders. Voor cilinders 9, 17, 25
en 33 verplaatst het stuwpunt zich in de richting van de stroomopwaarts gelegen
cilinder. De zone met positieve druk is uitgebreid voor cilinders 1, 9 en 17. De
minimum drukcoecient die optreedt vlak voordat de grenslaag loslaat, neemt
van cilinder 1 naar cilinder 8 toe. Cilinders 8 en 33 ondervinden heel grote zuiging
vlak voordat de grenslaag loslaat. De hoge snelheden in smalle tussenruimtes
veroorzaken uitgesproken minima in de drukverdeling op de plaats waar de
tussenruimte het smalst is. Alle cilinders die zich in het zog van andere cilinders
bevinden, hebben twee maxima op de posities waar de stroming terug aanhecht
aan de wand. Hoe meer stroomafwaarts de cilinders zich bevinden, hoe lager de
basisdrukcoecient. Sommige stroomopwaarts gelegen cilinders ondervinden een
positieve basisdrukcoecient.
De twee cilinders op de hoeken opz van de groep ondervinden vr grote
hefcoecienten. Eigenmodes met vier sinussen over de omtrek worden zwaarder
belast in de groepsopstelling dan in het geval van een alleenstaande cilinder
omdat iedere cilinder omringd wordt door 4 cilinders. Vooral voor de cilinders
op de hoeken opz van de groep worden de eigenmodes met drie of vier sinussen
over de omtrek sterk geexciteerd. Deze eigenmodes hebben vaak de laagste
eigenfrequenties. Dit verklaart waarom stormschade hoofdzakelk optreedt b
hoeksilos.
xii SAMENVATTING
1 9 17 25 33
2 10 18 26 34
3 11 19 27 35
4 12 20 28 36
5 13 21 29 37
6 14 22 30 38
7 15 23 31 39
8 16 24 32 40
Figuur 6: Tdsgemiddelde drukcoecienten C
p
voor de stroming rond een groep
van 8-b-5 cilinders (volle ln) en voor een alleenstaande cilinder
(streepln). De pl duidt de invalshoek van de wind aan.
SAMENVATTING xiii
Gekoppelde berekening van de windstroming rond een
silo
Het eindige-elementenmodel van de structuur wordt gekoppeld met de drie-
dimensionale onsamendrukbare turbulente windstroming. Tussen de structuur en
een cilinder met als diameter twee maal de diameter van de silo wordt de stroming
berekend op een vervormend rooster op basis van de arbitraire Lagrangiaanse-
Euleriaanse beschrving (sectie 2.2.1). De verplaatsingen van de roosterpunten
worden berekend aan de hand van een diusievergelking met een variabele
diusiviteit. Om de kwaliteit en de verjningen van het rooster te behouden,
wordt de diusiviteit gelk gekozen aan het omgekeerde van het volume van de
eindige volumes.
De roosters van het udum en de structuur zn niet samenvallend aan de
interface. Een consistente interpolatie wordt gebruikt voor de overdracht van de
verplaatsingen, terwl een knoop-naar-element conservatieve methode toegepast
wordt b de overdracht van de belastingen.
Als beginvoorwaarden worden de onvervormde structuur en de niet-stationaire
stroming rond een starre cilinder gekozen. De structurele demping wordt bepaald
aan de hand van de gemeten modale dempingsfactoren en toegevoegd onder de
vorm van Rayleigh demping. De structuur en het udum worden sequentieel
gekoppeld met twee van de beschikbare algoritmes: een zwak gekoppeld staggered
algoritme (A) dat elk veld een maal oplost per tdstap en een sterk gekoppeld
algoritme (B) dat itereert tussen de twee velden binnen een tdstap. Deze
iteraties zorgen ervoor dat het evenwicht op de interface op ieder tdstip voldaan
is. De overgedragen verplaatsingen en krachten worden niet gerelaxeerd. Er zn
maximaal vier iteraties vereist om de relatieve verandering van de overgedragen
grootheden te beperken tot 0.001.
Figuur 7a vergelkt het tdsverloop van de radiale verplaatsingen berekend met de
algoritmes A en B in drie punten op halve hoogte. Binnen dit korte tdsinterval
zn beide algoritmes stabiel. De verschillen tussen de resultaten van algoritme
A en B worden duidelk groter na verloop van td. De nauwkeurigheid van het
zwak gekoppelde algoritme is lager dan die van het sterk gekoppelde algoritme.
Figuur 7b toont de frequentie-inhoud van de radiale verplaatsingen berekend met
algoritme B. De resolutie f = 0.4 in het frequentiedomein is vr laag omwille
van het korte tdsinterval. De respons van de silo wordt gedomineerd door de
eigenmodes (1,3) en (1,4) rond 4 Hz. De piek b 2 Hz wst op de eecten van
de wervelafscheiding op de structuur. De kleinere pieken boven 4 Hz horen b
eigenmodes met hogere eigenfrequenties.
De vervormingen van de structuur benvloeden de drukverdeling rond de structuur:
b de windstroming rond een starre silo waren er in het tdsverloop van de druk
xiv SAMENVATTING
5.5 6 6.5 7 7.5
0.08
0.06
0.04
0.02
0
0.02
0.04
0.06
0.08
Time [s]
D
i
s
p
l
a
c
e
m
e
n
t

[
m
]
(a)
0 2 4 6 8 10
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
Frequency [Hz]
D
i
s
p
l
a
c
e
m
e
n
t

[
m
/
H
z
]
(b)
Figuur 7: (a) Tdsverloop van de radiale verplaatsingen op halve hoogte voor
= 66
( ), = 120
( ) en = 180
( )berekend met algoritme A

(streepln) en algoritme B (volle ln) en (b) frequentie-inhoud van de
radiale verplaatsingen berekend met algoritme B. De hoek = 0
valt
samen met het stuwpunt.
voornamelk bdrages b 2.16 en 4.31 Hz, terwl er b de gekoppelde berekening
eveneens bdrages zn b hogere frequenties. De belangrke bdrage rond 4 Hz
komt overeen met de eigenfrequenties van de eigenmodes (1,3) en (1,4). Door de
interactie is de grootte van de drukvariaties b 2 Hz duidelk toegenomen, wat
duidt op de versterking van de wervelafscheiding.
Aangezien de structuur plots belast wordt en de structurele demping vr laag is,
moet de respons gedurende een langere periode berekend worden om te evalueren
of de ovaliserende trillingen geleidelk uitgedempt worden of integendeel versterkt
worden. De rekentd voor deze berekeningen op een computer met een processor
is vr lang.
Suggesties voor verder onderzoek
In deze tekst wordt de stroming op vervormende rekenroosters berekend met
het commercieel softwarepakket CFX. De nauwkeurigheid en de stabiliteit van
stromingsberekeningen op vervormende roosters zou geevalueerd kunnen worden.
Ten eerste kan gecontroleerd worden of de tweede-orde nauwkeurigheid in de td
van het drie-punts-achterwaartse dierentieschema behouden blft in de ALE
beschrving. Bvoorbeeld kan de twee-dimensionale stroming rond een cilinder
die vervormt b een bepaalde frequentie met een bepaalde aantal sinussen over de
omtrek gesimuleerd worden. Op basis van berekeningen met verschillende groottes
van tdstap kan de nauwkeurigheid in de td bepaald worden.
SAMENVATTING xv
Vervolgens kan de implementatie van de geometrische behoudswet gecontroleerd
worden door een uniforme stroming te berekenen terwl de interne knopen van
het rooster op een arbitraire manier bewegen. De juiste oplossing moet bekomen
worden onafhankelk van de vervormingen van het rooster en aan de massabalans
moet steeds voldaan zn.
In deze thesis wordt de twee-dimensionale turbulente luchtstroming bestudeerd
rond een groep van 8-b-5 cilinders met een tussenafstand van 30 cm voor een
invalshoek van 30
. Een pertinente vraag is hoe gevoelig de berekende resultaten

zn voor veranderingen van de tussenafstand en van de invalshoek. Om deze
eecten te begroten, moet een systematische studie van de stroming uitgevoerd
worden voor een aantal invalshoeken en tussenafstanden.
De stroming rond een eindige cilinder met verhouding h/D = 4.55 is sterk drie-
dimensionaal met hoejzervormige wervels en boogvormige wervels. Het is te
verwachten dat soortgelke drie-dimensionale wervelstructuren terug te vinden zn
rond de groep silos als geheel. De numerieke simulatie van de drie-dimensionale
stroming rond een groep van veertig dicht b elkaar staande silos is een uitdagend
probleem wat betreft rekencapaciteit en kan slechts opgelost worden door de
berekeningen te paralleliseren.
Als overgeschakeld wordt op drie-dimensionale berekeningen kunnen de RANS
turbulentiemodellen vervangen worden door een hybride RANS/LES aanpak. Deze
methode is beter geschikt om de grootschalige turbulente structuren in het zog van
bouwkundige constructies te berekenen.
In het laatste hoofdstuk is de gekoppelde simulatie van de drie-dimensionale
onsamendrukbare turbulente windstroming rond een cilindervormige silo berekend
om het optreden van ovaliserende trillingen te voorspellen. Aangezien de rekentd
heel lang is, zn meer eciente koppelingsalgoritmes vereist. Het gebruik van
de commercieel beschikbare koppeling tussen Ansys en CFX heeft de keuze wat
betreft koppelingsalgoritmes beperkt. Een eigen implementatie van de koppeling
tussen de twee commerciele programmas moet het gebruik van meer geavanceerde
koppelingsalgoritmes mogelk maken.
De nauwkeurigheid van het conventionele staggered algoritme kan verbeterd
worden door geextrapoleerde structurele verplaatsingen op te leggen aan het
udum en de overgedragen udumkrachten overeenkomstig te corrigeren. Als
dit verbeterde algoritme stabiel zou zn voor zwak gekoppelde problemen met
onsamendrukbare stromingen, kan het gebruik van iteraties binnen een tdstap
vermeden worden en de rekentd substantieel verminderd worden.
B sterk gekoppelde problemen kan het algoritme met iteraties binnen een
tdstap verbeterd worden door gebruik te maken van automatisch bepaalde
relaxatiefactoren gedurende de iteraties tussen het udum en de structuur. Deze
techniek verbetert de convergentie en reduceert het aantal benodigde iteraties.
xvi SAMENVATTING
Een interessante denkpiste is het aanmaken van gereduceerde orde modellen voor
het udum en de structuur tdens de iteraties tussen het udum en de structuur
(Vierendeels, 2006). In elke iteratie wordt in het udum op de udum-structuur
interface de drukverdeling berekend die hoort b de opgelegde verplaatsingen.
Deze sets van opgelegde verplaatsingen en bhorende drukverdelingen kunnen
gebruikt worden om een gereduceerde orde model van het udum op te stellen.
Voor de structuur worden op een analoge manier in iedere iteratie de verplaatsingen
op de interface berekend die horen b een opgelegde drukverdeling. Deze sets
kunnen gebruikt worden om een gereduceerde orde model van de structuur op te
stellen. De resultaten van een gekoppelde berekening met de beide gereduceerde
orde modellen levert een startwaarde op voor de volgende iteratie. B elke iteratie
worden de gereduceerde orde modellen verbeterd. Deze techniek laat toe om sterk
gekoppelde problemen op te lossen met een klein aantal iteraties. De techniek
kan eventueel verbeterd worden door de gereduceerde orde modellen opgesteld
in de vorige tdstap te hergebruiken. Dit kan heel interessant zn b aero-
elastische problemen zoals utter, galloping en het ovaliseren van silos, waarb
veel berekeningen gemaakt worden voor bna identieke geometrieen.
Een globale multigrid techniek, die de gepartitioneerde koppeling tussen het
udum en de structuur berekent op verschillende roosters, kan de stabiliteit, de
ecientie en de nauwkeurigheid verhogen (Sch afer et al., 2006; van Zulen et al.,
2007). De gekoppelde berekening wordt eerst uitgevoerd op een ruw rekenrooster
voor zowel het udum als de structuur. Deze oplossing kan gebruikt worden als
startwaarde voor een gekoppelde berekening op jnere roosters. De laagfrequente
fouten in de oplossing op deze jnere roosters kunnen gecorrigeerd worden door
een nieuwe berekening op het ruwe rooster uit te voeren.
De ecientie, nauwkeurigheid en stabiliteit van de verschillende koppelings-
algoritmes kan geverieerd worden aan de hand van een reeks benchmarks bv. een
elastische balk achter een vierkant (Wall and Ramm, 1998) of achter een cilinder
(T urek and Hron, 2006). Bathe and Ledezma (2007) geven een overzicht van een
reeks benchmarks om koppelingsalgoritmes en de overdracht van belastingen en
verplaatsingen te evalueren.
De reductie van de dimensies van het probleem kan de rekentd eveneens
verminderen. In het geval van de silo kan de eindige-strookmethode (appendix B)
toegepast worden om een representatief twee-dimensionaal model van de structuur
te maken. De structurele verplaatsingen worden verondersteld als een halve sinus
of een halve cosinus te varieren over de hoogte. De eindige-strookmethode is
gemplementeerd als de combinatie van een twee-dimensionaal balkelement en
een door de gebruiker toegevoegd element. Een nadeel is dat het door de
gebruiker toegevoegd element onafhankelk is van de vervormingen, waardoor
geometrisch niet-lineair gedrag niet ingerekend wordt voor deze elementen.
Dit eindige-strookmodel kan gekoppeld worden met een twee-dimensionale
stromingsberekening. B de overdracht van de winddrukken naar de structuur
SAMENVATTING xvii
moet een aanname gemaakt worden voor de variatie van de drukken over de
hoogte. Dit verloop is noch constant, noch sinusodaal. Als validatie kunnen de
resultaten van een gekoppelde berekening met het eindige-strookmodel en de twee-
dimensionale stroming vergeleken worden met de resultaten van de gekoppelde
berekening van de schaalconstructie en de drie-dimensionale stroming. De invloed
van de groepsopstelling op het optreden van ovaliserende trillingen kan vervolgens
bestudeerd worden door het eindige-strookmodel van de silos te koppelen met
twee-dimensionale stroming rond de groep. Het aantal silos in de groep kan
hierb best geleidelk verhoogd worden, te beginnen met twee dicht b elkaar
staande silos.
Hoewel in het voorliggende werk enkel het ovaliseren van silos bestudeerd
wordt, kan dezelfde methodologie direct toegepast worden b het berekenen
van de respons van zeilconstructies onder windbelasting, van utter van
bruggen, van galloping van kabels en van trillingen van kabels geexciteerd door
wervelafscheiding.
Organisatie van de tekst
Alvorens de studie van het wind-structuur interactieprobleem aan te vatten, is
een grondige kennis van het gedrag van de structuur en van de stroming vereist.
Voor beide deelproblemen zn nauwkeurige numerieke modellen nodig. Daarom is
veel energie genvesteerd in de oplossing van deze afzonderlke problemen. Deze
worden uiteindelk aangewend in hoofdstuk 5 om het ovaliseren te simuleren aan
de hand van een gekoppelde berekening van de windstroming rond de structuur.
Hoofdstuk 1 situeert het onderwerp en licht de objectieven en eigen bdragen toe.
Het ovaliseren van cylindrische schaalconstructies wordt ingeleid. Dit dynamisch
udum-structuur interactieprobleem wordt doorheen de hele thesis als voorbeeld
gebruikt.
Hoofdstuk 2 behandelt de arbitraire Lagrangiaanse-Euleriaanse beschrving, die
vereist is voor de berekening van stromingen in gebieden met veranderende
begrenzingen. De stabiliteit en de tdsnauwkeurigheid van deze berekeningen
wordt besproken en gerelateerd aan de geometrische behoudswet. De verschillende
methodes voor de berekening van de vervormingen van het rekenrooster worden
beschreven.
Hoofdstuk 3 beschrft de vergelkingen voor lineair en geometrisch niet-lineair
gedrag van schaalconstructies. In situ is op een silo een experiment uitgevoerd
om de eigenfrequenties, eigenmodes en modale dempingsfactoren te bepalen. Aan
de hand van deze experimentele modale parameters wordt een drie-dimensionaal
xviii SAMENVATTING
eindige-elementenmodel van de silo gevalideerd. Het belang van geometrisch niet-
lineair gedrag tdens het ovaliseren van de silos wordt bestudeerd.
Hoofdstuk 4 behandelt de berekening van turbulente windstromingen rond starre
bouwkundige constructies. Bestaande turbulentiemodellen worden besproken.
Om een geschikt turbulentiemodel te selecteren, worden de resultaten van een
stationaire en een niet-stationaire Reynolds gemiddelde Navier-Stokes berekening
van de stroming rond een cilinder vergeleken met experimentele data en met
beschikbare numerieke resultaten. De berekening van de stroming rond een
groep van 8-b-5 dicht b elkaar staande cilinders illustreert de invloed van de
groepsopstelling op de wervelafscheiding en de drukverdelingen rond de cilinders.
Hoofdstuk 5 beschrft gepartitioneerde algoritmes die het gekoppelde udum-
structuur interactie probleem oplossen door de verschillende velden sequentieel
of parallel te berekenen. De nauwkeurigheid, stabiliteit en ecientie van deze
algoritmes wordt behandeld. De overdracht van verplaatsingen en belastingen
tussen de niet-samenvallende rekenroosters van de structuur en het udum op
de interface wordt besproken. Het behoud van de totale belasting en van de
energie op de interface is van groot belang. Een gekoppelde berekening van de
drie-dimensionale windstroming rond een vervormende silo wordt uitgevoerd. Meer
eciente koppelingstechnieken moeten het mogelk maken de respons van de de
silo gedurende veel langere periodes te simuleren.
Hoofdstuk 6 vat de belangrkste conclusies van deze thesis samen en geeft
aanbevelingen voor verder onderzoek.
Summary
For modern slender structures, wind loading is often a critical load case. In the
case of quite exible structures, the large structural displacements inuence the
wind pressure distribution around the structure. This may give rise to structural
instabilities, such as utter of bridges, galloping of cables and ovalling of silos. This
thesis studies wind induced ovalling oscillations of a group of forty closely spaced
silos by means of coupled numerical simulations of the uid and the structure.
First, a three-dimensional nite element model of the silo structure is validated
by means of modal parameters derived from an in situ experiment. The
boundary conditions strongly inuence the eigenfrequencies and mode shapes. The
eigenmodes with the lowest eigenfrequencies (3.93 Hz) have half a wave along the
height and three or four waves around the circumference. All eigenmodes have a
low damping ratio around 1%.
Next, the turbulent wind ow around a group of 8 by 5 silos at Reynolds number
12.4 10
6
is studied using two-dimensional unsteady incompressible Reynolds
averaged Navier-Stokes simulations. In order to clarify the inuence of the
turbulence models, the near-wall modelling and the unsteadiness, results for the
ow around a single cylinder are rst compared with experimental data and
numerical results reported in the literature. In the group conguration especially
the cylinders on the side corners are heavily loaded.
Finally, the nite element computation of the structure is coupled with the three-
dimensional computation of the ow around a single silo in order to predict
ovalling oscillations. At the uid-structure interface boundary conditions are
exchanged between both computations. Within each time step the method iterates
between the structure and the uid until the coupling conditions are fullled. The
ow is computed on a deforming mesh, using the arbitrary Lagrangian Eulerian
formulation. The grid point displacements of the uid mesh are obtained by
diusing the structural displacements at the interface through the uid domain.
Geometrically non-linear structural behaviour is taken into account through an
updated Lagrangian description. The computed response of the silo is dominated
by the eigenmodes with three or four waves around the circumference.
xix
List of Symbols
The following list provides an overview of symbols used throughout the text. The
physical meaning of the symbols is explained in the text. Vectors, matrices and
tensors are denoted by bold characters.
The general symbols, conventions and abbreviations are collected in the rst
two sections. The remaining symbols are categorized in sections referring to the
chapters or to the sections where they are rst introduced.
General symbols and conventions
(x, y, z) Cartesian coordinates
(r, , z) cylindrical coordinates
n
vector at time level n
(i)
vector at iteration i
discretization of a vector
/ rst order partial derivative with respect to the variable
2
/
2
second order partial derivative with respect to the variable
rst order partial time derivative of the variable with

variable xed
D
Dt
rst order material time derivative of the variable
rst order material time derivative of the variable
second order material time derivative of the variable
del operator
2
Laplace operator
gradient of a vector eld
divergence of a vector eld
1
inverse of a matrix
T
transpose of a matrix
det determinant of a matrix
tr trace of a matrix
I identity matrix
xxi
xxii LIST OF SYMBOLS
ij
Kronecker Delta
t time
f frequency
circular frequency
t time step
Acronyms
ALE Arbitrary Lagrangian-Eulerian
APR Adverse Pressure Recovery
CFD Computational Fluid Dynamics
DGCL Discrete Geometrical Conservation Law
DNS Direct Numerical Simulation
FEM Finite Element Method
FFT Fast Fourier Transform
FSI Fluid-Structure Interaction
GCL Geometrical Conservation Law
GLS Galerkin/Least Squares
LBB Ladyzhenskaya-Babuska-Brezzi
LES Large Eddy Simulation
PFEM Particle Finite Element Method
PSPG Pressure-Stabilizing/Petrov-Galerkin
RANS Reynolds averaged Navier-Stokes
RMS Root Mean Square
RSM Reynolds Stress Model
SST Shear Stress Transport
SUPG Streamline-Upwind Petrov-Galerkin
SZL Shih-Zhu-Lumley
TL Total Lagrangian
UL Updated Lagrangian
URANS Unsteady Reynolds averaged Navier-Stokes
X-FEM Extended Finite Element Method
Flow on a domain with deforming boundaries
x spatial coordinates
X material coordinates
referential coordinates
x
spatial domain
X
material domain
referential domain
LIST OF SYMBOLS xxiii
x
variable in the spatial domain
X
variable in the material domain
variable in the referential domain

area vector
n unit normal vector
variable concerning the mapping from the material to the
spatial domain
variable concerning the mapping from the material to the
referential domain
variable concerning the mapping from the referential to the
spatial domain
u displacement
v particle velocity (in the spatial domain)
v particle velocity in the referential domain
v velocity of the nodes of the mesh
c convective velocity
u displacement of the nodes of the mesh
x position of the nodes of the mesh
mapping
F deformation gradient
J Jacobian
M mass
M momentum
density
b body forces
t traction vector
Cauchy stress tensor
strain rate tensor
vorticity tensor
p pressure
deviatoric stress tensor
dynamic viscosity
second viscosity constant
K bulk viscosity
kinematic viscosity
p
kinematic pressure
k diusivity
xxiv LIST OF SYMBOLS
Shell structures
u virtual displacement
virtual strain
D displacement gradient
C right Cauchy-Green deformation tensor
b left Cauchy-Green deformation tensor
E Green-Lagrange strain tensor
e Almansi strain tensor
R orthogonal rotation tensor
U right stretch tensor
V left stretch tensor
Cauchy stress tensor
Kirchho stress tensor
P
X
rst Piola-Kirchho stress tensor
S second Piola-Kirchho stress tensor
rst Newmark parameter
second Newmark parameter
h height
L length
R radius
t thickness
D diameter
P center-to-center distance
rst Lame constant
second Lame constant
E Youngs modulus
Poissons ratio
m axial half wave number
n circumferential wave number
Finite element method
h
trial solution for the variable
1 residual
w weighting functions for the momentum equations
q weighting function for the continuity equation
N
v
i
nite element shape functions for the velocity
N
p
i
nite element shape function for the pressure
M mass matrix
K stiness matrix
LIST OF SYMBOLS xxv
K
T
tangent stiness matrix
K
geometric stiness matrix

C damping matrix
G discrete gradient operator
f load vector
w
m
modied weighting functions for the momentum equations
numerical diusion
diusivity tensor
stabilization parameter
ST stabilization term
Finite volume method
V conservative variables
Q source terms
F
d
diusive (viscous) uxes
F
c
convective (inviscid) uxes
F
e
convective uxes corresponding to an Eulerian description
F
ale
additional convective uxes corresponding to an ALE
description
F
d
numerical diusive uxes
F
c
numerical convective uxes
F
e
numerical convective uxes corresponding to an Eulerian
description
F
ale
additional numerical convective uxes corresponding to an
ALE description
Computation of turbulent wind ows
mean part of the variable
uctuating part of the variable

k turbulent kinetic energy
turbulent frequency
turbulence dissipation rate
T turbulence production
t
turbulent viscosity
y
+
dimensionless wall distance
Re Reynolds number
St Strouhal number
Ma Mach number
C
d
drag coecient
xxvi LIST OF SYMBOLS
C
l
lift coecient
C
p
pressure coecient
C
min
p
minimum of the pressure coecient
C
b
p
base pressure coecient
circumferential angle
s
separation angle
incidence angle of the ow
Coupled simulation of wind loading on structures
f
variable in the uid partition
s
variable in the structural partition
variable in the interior of the domain
variable on the uid-structure interface

t
f
traction vector in the uid on the uid-structure interface
t
s
traction vector in the structure on the uid-structure
interface
f
stress eld in the structural partition
s
stress eld in the uid partition
u
discretized displacements of the uid mesh on the

uid-structure interface
u
discretized displacements in the structure on the

uid-structure interface
f
f
nodal forces in the uid on the uid-structure interface
f
s
nodal forces in the structure on the uid-structure interface
E energy
relaxation factor
F uid solver
S structural solver
S Schur complement
H motion transfer matrix
L load transfer matrix
Contents
Voorwoord i
Samenvatting iii
Summary xix
List of Symbols xxi
Contents xxvii
List of Figures xxxi
List of Tables xxxix
1 Introduction 1
1.1 Problem outline and motivation . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Wind loading on civil engineering structures . . . . . . . . . 1
1.1.2 A wind-structure interaction problem: ovalling of silos . . . 2
1.2 State-of-the-art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.1 Response of structures to wind loading . . . . . . . . . . . . 4
1.2.2 Ovalling oscillations . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Objectives and original contributions . . . . . . . . . . . . . . . . . 11
1.3.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . 11
xxvii
xxviii CONTENTS
1.3.2 Original contributions . . . . . . . . . . . . . . . . . . . . . 13
1.4 Organization of the text . . . . . . . . . . . . . . . . . . . . . . . . 14
2 Flow on a domain with deforming boundaries 17
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 Flow on a domain with deforming boundaries . . . . . . . . . . . . 19
2.2.1 The arbitrary Lagrangian-Eulerian description . . . . . . . 19
2.2.2 Kinematics . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.2.3 Constitutive equations . . . . . . . . . . . . . . . . . . . . . 33
2.2.4 Governing equations . . . . . . . . . . . . . . . . . . . . . . 34
2.3 Spatial discretization . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.3.1 Finite Element Method . . . . . . . . . . . . . . . . . . . . 36
2.3.2 Finite Volume Method . . . . . . . . . . . . . . . . . . . . . 46
2.4 Time integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.4.1 The Eulerian description . . . . . . . . . . . . . . . . . . . . 48
2.4.2 The ALE description . . . . . . . . . . . . . . . . . . . . . . 50
2.5 Mesh deformation . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3 Shell structures 65
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.2 Geometric non-linear behaviour . . . . . . . . . . . . . . . . . . . . 66
3.2.1 Virtual work . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.2.2 Kinematics . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.2.3 Conjugate stress and strain tensors . . . . . . . . . . . . . . 73
3.2.4 Incremental virtual work . . . . . . . . . . . . . . . . . . . . 76
3.2.5 Newton-Raphson procedure . . . . . . . . . . . . . . . . . . 80
3.2.6 Time integration . . . . . . . . . . . . . . . . . . . . . . . . 82
CONTENTS xxix
3.3 Structural behaviour of a silo . . . . . . . . . . . . . . . . . . . . . 84
3.3.1 Description of the silo structure . . . . . . . . . . . . . . . . 84
3.3.2 Harmonic nite element model . . . . . . . . . . . . . . . . 85
3.3.3 The experimental setup . . . . . . . . . . . . . . . . . . . . 88
3.3.4 Three-dimensional nite element model of the silo . . . . . 93
3.3.5 Non-linear behaviour of the silos . . . . . . . . . . . . . . . 99
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4 Computation of turbulent wind ows 105
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.2 Computational approaches to turbulent ows . . . . . . . . . . . . 106
4.2.1 Direct numerical simulation . . . . . . . . . . . . . . . . . . 106
4.2.2 Reynolds averaged Navier-Stokes simulation . . . . . . . . . 107
4.2.3 Large eddy simulation . . . . . . . . . . . . . . . . . . . . . 109
4.3 RANS turbulence models . . . . . . . . . . . . . . . . . . . . . . . 111
4.3.1 Linear eddy viscosity models . . . . . . . . . . . . . . . . . 111
4.3.2 Non-linear eddy viscosity models . . . . . . . . . . . . . . . 117
4.3.3 Reynolds stress models . . . . . . . . . . . . . . . . . . . . . 118
4.4 Turbulent air ow around a single cylinder . . . . . . . . . . . . . . 119
4.4.1 Steady computation: problem domain, near-wall
modelling and eddy viscosity turbulence models . . . . . . . 123
4.4.2 Unsteady computation . . . . . . . . . . . . . . . . . . . . . 126
4.4.3 Mesh renement . . . . . . . . . . . . . . . . . . . . . . . . 128
4.4.4 Comparison with numerical results reported in the literature 132
4.5 Turbulent air ow around a cylinder group . . . . . . . . . . . . . 134
4.5.1 Group of 2 by 2 cylinders . . . . . . . . . . . . . . . . . . . 138
4.5.2 Group of 8 by 5 cylinders . . . . . . . . . . . . . . . . . . . 145
4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
xxx CONTENTS
5 Coupled simulation of wind loading on structures 153
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
5.2 Coupling algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . 155
5.2.1 The monolithic and the partitioned approach . . . . . . . . 155
5.2.2 Loosely coupled algorithms . . . . . . . . . . . . . . . . . . 158
5.2.3 Strongly coupled algorithms . . . . . . . . . . . . . . . . . . 169
5.3 Load and motion transfer . . . . . . . . . . . . . . . . . . . . . . . 174
5.3.1 Motion transfer or surface tracking . . . . . . . . . . . . . . 176
5.3.2 Load transfer . . . . . . . . . . . . . . . . . . . . . . . . . . 179
5.3.3 Conservation of energy . . . . . . . . . . . . . . . . . . . . . 184
5.4 Application: ovalling of silos . . . . . . . . . . . . . . . . . . . . . . 185
5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
6 Conclusions and recommendations for further research 195
6.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
6.2 Recommendations for further research . . . . . . . . . . . . . . . . 198
Bibliography 203
Curriculum vitae 221
A Possibilities of commercial software packages Flotran and CFX 225
B Finite strip model of the silo 229
List of Figures
1 De silo groep in de haven van Antwerpen. . . . . . . . . . . . . . . iv
2 Bovenaanzicht en drie-dimensionaal zicht van het reeel (volle ln)
en het imaginair deel (streepln) van de gedenticeerde eigenmodes
3, 4 en 53 van de silo. De stippelln duidt de onvervormde silo aan. viii
3 Drie-dimensionaal eindige-elementenmodel van de silo en een aantal
geselecteerde eigenmodes. . . . . . . . . . . . . . . . . . . . . . . . ix
4 Drukcoecient als functie van de hoek bekomen met een
stationaire berekening (streepln) en het tdsgemiddelde (volle ln),
het minimum (streep-stippelln) en het maximum (stippelln) van
de drukcoecient bekomen met een niet-stationaire berekening voor
(a) rooster A
2
en (b) rooster B. De lichtgrze zone bevat alle
beschikbare experimentele data b Reynoldsgetallen van 0.73 10
7
tot 3.65 10
7
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x
5 (a) Tdsverloop en (b) frequentie-inhoud van de druk op het
oppervlak van een cilinder in de groep van 8-b-5 cilinders. . . . . xi
6 Tdsgemiddelde drukcoecienten C
p
voor de stroming rond een
groep van 8-b-5 cilinders (volle ln) en voor een alleenstaande
cilinder (streepln). De pl duidt de invalshoek van de wind aan. . xii
7 (a) Tdsverloop van de radiale verplaatsingen op halve hoogte voor
= 66
( ), = 120
( ) en = 180
( )berekend met algoritme

A (streepln) en algoritme B (volle ln) en (b) frequentie-inhoud
van de radiale verplaatsingen berekend met algoritme B. De hoek
= 0
valt samen met het stuwpunt. . . . . . . . . . . . . . . . . . xiv

1.1 (a) Flutter of the Tacoma Narrows suspension bridge and (b) its
nal collapse (Champneys, 2008). . . . . . . . . . . . . . . . . . . . 2
xxxi
xxxii LIST OF FIGURES
1.2 (a) Collapse of one of the cooling towers at Ferrybridge (UK) and
(b) the nal destruction of three cooling towers (Norfolk, 2008). . . 3
1.3 The silo group in the port of Antwerp. . . . . . . . . . . . . . . . . 3
1.4 Plan and lateral view of the silo group. . . . . . . . . . . . . . . . . 4
2.1 (a) Explicit and (b) implicit description of a deforming boundary
of the uid domain. . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2 Representations of (a) the Lagrangian, (b) the Eulerian and (c) the
arbitrary Lagrangian-Eulerian descriptions in the spatial domain
x
at two dierent times t: continuum (dark gray), material particles
(black dots) and mesh (black lines). Dashed lines show the mesh
motion, while solid lines show the particle motion. . . . . . . . . . 20
2.3 Domains, meshes and mappings for an ALE description: xed
referential domain
, changing material domain

X
and changing
spatial domain
x
at time t (solid lines) and t + t (dashed lines). 21
2.4 Analytical solution for the velocity (black line), exact nodal solution
(black dashed line) (2.128), Galerkin solution (dark grey line) (2.126)
and full upwind solution (light grey line) (2.130) of a steady 1D
scalar convection-diusion equation with a constant convection c =
1 and a constant source term b = 1 using an element size h = 0.1
for a Peclet number of (a) 0.25 and (b) 5. . . . . . . . . . . . . . . 40
3.1 Octagonal beam with four bolted connections at the bottom of the
cylinder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.2 Harmonic nite element model of the silo and selected eigenmodes. 86
3.3 Eigenfrequencies of the harmonic nite element model of the silo as
a function of m and n. . . . . . . . . . . . . . . . . . . . . . . . . . 87
3.4 Dimensionless total strain energy (solid line) and dimensionless
strain energy due to bending (dashed-dotted line) and stretching
(dashed line) for m = 2 () and m = 3 () for a freely supported
cylinder with a height h = 25 m and a constant thickness t = 7 mm. 88
3.5 Location of the accelerometers on the silo wall. . . . . . . . . . . . 89
3.6 Power spectral density of the radial acceleration at the point HL09. 90
3.7 Top and three-dimensional view of the real (solid line) and
imaginary (dashed line) part of the identied eigenmodes 3, 4 and
53 of the silo. The dotted line shows the undeformed shape. . . . . 91
LIST OF FIGURES xxxiii
3.8 The radial displacements of mode 4 in the points HL6 to HL13 in
the complex plane. The displacements are scaled to unity in the
point HL6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
3.9 Measured eigenfrequencies of the silo as a function of m and n. . . 92
3.10 Time history of the radial acceleration at the point VL15 (top), the
nine modal contributions with the highest RMS value and the error
(bottom). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
3.11 The approximated relative error on nine selected eigenfrequencies
predicted with the quarter three-dimensional nite element model
as a function of the mesh renement factor. . . . . . . . . . . . . . 95
3.12 Top view of the eigenmode (1, 6) of the silo at a height of 15 m
computed with (a,b) the quarter three-dimensional nite element
model with SS and AA boundary conditions and (c) the full three-
dimensional nite element model. . . . . . . . . . . . . . . . . . . . 96
3.13 Top and lateral view of the eigenmode (1, 5) of the silo computed
with (a) the quarter three-dimensional nite element model with
SA boundary conditions and (b) the full three-dimensional nite
element model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
3.14 Three-dimensional nite element model of the silo and selected
eigenmodes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
3.15 Eigenfrequencies of the three-dimensional nite element model of
the silo as a function of m and n. . . . . . . . . . . . . . . . . . . . 98
3.16 Deformed shape of a cantilever beam of length L under a tip moment
M = EI/L. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
3.17 Imposed pressure distribution around the circumference of the silo. 101
3.18 (a) Finite element mesh and boundary conditions and (b)
deformations magnied by factor of 3. . . . . . . . . . . . . . . . . 102
4.1 Pressure coecient C
p
for a single cylinder. . . . . . . . . . . . . . 120
4.2 Measured pressure coecients at Reynolds numbers from 0.7310
7
to 3.65 10
7
(Zdravkovich, 1997). . . . . . . . . . . . . . . . . . . 121
4.3 Problem domain and mesh A
2
for a single cylinder. . . . . . . . . . 124
xxxiv LIST OF FIGURES
4.4 Pressure coecient as a function of the angle obtained with
Flotran (solid lines) and CFX (dashed lines) turbulence models
using various eddy viscosity turbulence models ( k , SST,
SZL) for (a) mesh A
2
and (b) mesh B. The light grey zone contains
all available experimental data at Reynolds numbers from 0.7310
7
to 3.65 10
7
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
4.5 (a) Turbulent kinetic energy and (b) wall shear stress as a function
of the angle at the cylinders surface using dierent eddy viscosity
turbulence models ( k , SST, SZL) for mesh B: Flotran (solid
lines) and CFX (dashed lines). . . . . . . . . . . . . . . . . . . . . 125
4.6 (a) Time history and (b) frequency content of the pressure at the
cylinders surface at = 172
for a single cylinder using mesh B. . 126

4.7 (a) Time average and (b) standard deviation of the pressure p for a
single cylinder using mesh B. . . . . . . . . . . . . . . . . . . . . . 127
4.8 Pressure coecient as a function of the angle obtained with the
steady state computation (dashed line) and the time average (solid
line), minimum (dash-dotted line) and maximum (dotted line) of
the pressure coecient obtained with the transient computation
for (a) mesh A
2
and (b) mesh B. The light grey zone contains all
available experimental data at Reynolds numbers from 0.73 10
7
to 3.65 10
7
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.9 (a) Pressure coecient obtained on meshes A
1
(black), A
2
(blue),
A
3
(green) and A
4
(red) and extrapolated solutions for mesh
renement study 1 (cyan) and 2 (magenta) and (b) local apparent
order of accuracy for mesh renement study 1 (solid line) and 2
(dashed line). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
4.10 (a) Approximated relative errors (dash dotted lines), extrapolated
relative errors (dashed lines), GCI using the local (dotted lines)
and the average (solid lines) apparent order of accuracy for mesh
renement study 1 (thin lines) and 2 (thick lines) and (b) time
history of drag (thin) and lift (thick) coecients obtained on meshes
A
1
(solid line), A
2
(dash dotted line) , A
3
(dashed line) and A
4
(dotted line). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
4.11 Flow patterns around two closely spaced cylinders for (a) small, (b)
medium and (c) large incidence angles (Sumner et al., 2000). . . . 135
4.12 Detail of the mesh for the group of 2 by 2 cylinders. . . . . . . . . 138
LIST OF FIGURES xxxv
4.13 Dimensionless distance y
+
of the nodes next to the cylinder wall
as a function of the angle for cylinder 1 (solid line), 2 (dashed
line), 3 (dotted line) and 4 (dash-dotted line) of the group of 2 by
2 cylinders in (a) mesh A and (b) mesh B. . . . . . . . . . . . . . . 139
4.14 Time average (thick line), maximum and minimum (thin lines) of
the pressure coecient C
p
for mesh A (black) and mesh B (grey)
as a function of the angle for (a) cylinder 1, (b) cylinder 2, (c)
cylinder 3 and (d) cylinder 4 of the group of 2 by 2 cylinders. . . . 139
cylinders surface in the point B of the group of 2 by 2 cylinders. . 140
cylinders surface in the point E of the group of 2 by 2 cylinders. . 141
4.17 Streamlines at t = 14.075 s from a transient computation around a
group of 2 by 2 cylinders. . . . . . . . . . . . . . . . . . . . . . . . 141
4.18 Flow patterns around two closely spaced cylinders for (a) = 25
and P/D = 1.3 and (b) = 45
and P/D = 1.1 (Alam et al., 2005). 142

4.19 Time averaged pressure coecients C
p
for the ow around a group
of 2 by 2 cylinders (solid line) and for the ow around a single
cylinder (dashed line). The arrow indicates the incidence angle . . 142
4.20 (a) Time averaged drag coecient C
d
, (b) uctuating drag
coecient C
d
, (c) time averaged lift coecient C
l
and (d)
uctuating lift coecient C
l
for the ow around the group of 2
by 2 cylinders. The arrow indicates the incidence angle . . . . . . 143
4.21 Decomposition of the time averaged pressure coecient C
p
into a
series of cosine functions with circumferential wavenumber n for the
ow around the group of 2 by 2 cylinders (2 by 2 circles) and for
the ow around a single cylinder (single circle). The arrow indicates
the incidence angle . . . . . . . . . . . . . . . . . . . . . . . . . . 144
4.22 Detail of the mesh for the group of 8 by 5 cylinders. . . . . . . . . 145
4.23 Dimensionless distance y
+
of the nodes next to the cylinder wall
as a function of the angle for all cylinders of the group of 8 by 5
cylinders. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
cylinders surface in the point B of the group of 8 by 5 cylinders. . 146
cylinders surface in the point E of the group of 8 by 5 cylinders. . 146
xxxvi LIST OF FIGURES
4.26 Streamlines at t = 50.94 s from a transient computation around a
group of 8 by 5 cylinders. . . . . . . . . . . . . . . . . . . . . . . . 147
4.27 Time averaged pressure coecients C
p
for the ow around a group
of 8 by 5 cylinders (solid line) and for the ow around a single
cylinder (dashed line). The arrow indicates the incidence angle . . 148
4.28 (a) Time averaged drag coecient C
d
, (b) uctuating drag
coecient C
d
l
and (d)
uctuating lift coecient C
l
for the ow around the group of 8
by 5 cylinders. The arrow indicates the incidence angle . . . . . . 149
4.29 Decomposition of the time averaged pressure coecient C
p
into a
ow around the group of 8 by 5 cylinders (8 by 5 circles) and for
the ow around a single cylinder (single circle). The arrow indicates
the incidence angle . . . . . . . . . . . . . . . . . . . . . . . . . . 150
5.1 The serial staggered algorithm. . . . . . . . . . . . . . . . . . . . . 159
5.2 Non-collocated algorithm. . . . . . . . . . . . . . . . . . . . . . . . 163
5.3 The parallel algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . 165
5.4 The enhanced parallel algorithm. . . . . . . . . . . . . . . . . . . . 165
5.5 Staggered algorithm with subcycling of the uid. . . . . . . . . . . 166
5.6 The iteratively staggered algorithm. . . . . . . . . . . . . . . . . . 170
5.7 Meshes and displacement proles at the uid-structure interface
in the structure ( ) and the uid partition ( ) obtained with the
consistent interpolation method when uid mesh is (a) ner and (b)
coarser than the solid mesh. . . . . . . . . . . . . . . . . . . . . . . 177
5.8 Meshes and displacement proles at the uid-structure interface
in the structure ( ) and the uid partition ( ) obtained with the
element to element mapping when uid mesh is (a) ner and (b)
5.9 Meshes and nodal forces at the uid-structure interface in the
structure ( ) and the uid partition ( ) obtained with the
conservative method when uid mesh is (a) ner and (b) coarser
than the solid mesh. . . . . . . . . . . . . . . . . . . . . . . . . . . 182
LIST OF FIGURES xxxvii
5.10 Meshes and nodal forces at the uid-structure interface in the
structure ( ) and the uid partition ( ) obtained with the element
to element conservative method when uid mesh is (a) ner and (b)
cylinders surface at mid-height for = 112
( ), = 174
( ) and
= 180
( ). The angle = 0
coincides with the stagnation point. 187

5.12 Model for the coupled simulation of the three-dimensional wind ow
around a cylinder and the response of the silo structure. . . . . . . 188
5.13 (a) Time history of the radial displacements at mid-height for =
66
( ), = 120
( ) and = 180
( ) computed with algorithm

A (dashed lines) and algorithm B (solid lines) and (b) frequency
content of the radial displacements computed with algorithm B. The
angle = 0
coincides with the stagnation point. . . . . . . . . . . 189

5.14 Deformations (enlarged with a factor 5) of the structure between
11.25 m and 13.75 m high at (a) t = 5.905 s, (b) t = 6.805 s and (c)
t = 7.250 s. The wind ows from the left. . . . . . . . . . . . . . . 190
( ), = 174
( ) and
= 180
( ) using algorithm B. The angle = 0
coincides with
the stagnation point. . . . . . . . . . . . . . . . . . . . . . . . . . . 190
5.16 Pressure eld on the vertical plane through the cylinder axis parallel
with the inlet ow direction at t = 7.9 s. . . . . . . . . . . . . . . . 191
5.17 Pressure at the cylinders surface along the height for (a) = 0
,
(b) = 112
and (c) = 180
at t = 5.4 s ( ), t = 5.475 s ( ),
t = 5.925 s ( ), t = 6.35 s ( ) and t = 6.605 s ( ). . . . . . . . . . . 191
5.18 (a) Time history and (b) frequency content of the rst (solid line)
and second principal component (dashed line) of the displacements
at mid-height with circumferential wavenumber n = 2 ( ), n = 3
( ) and n = 4 ( ). . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
B.1 The relative error on the nine lowest eigenfrequencies predicted with
the nite strip model as a function of the number of elements in the
circumferential direction. . . . . . . . . . . . . . . . . . . . . . . . . 232
B.2 Eigenmodes of a silo with a height h = 25 m and a thickness t =
7 mm, computed with a nite strip model. . . . . . . . . . . . . . . 233
List of Tables
3.1 Thickness of the aluminium plates of the silo as a function of the
height. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.2 Comparison of the eigenfrequencies computed with the harmonic
nite element model, the quarter three-dimensional nite element
model and the full three-dimensional nite element model with the
experimental eigenfrequencies. . . . . . . . . . . . . . . . . . . . . . 99
3.3 Horizontal and vertical tip displacement, number of load increments
and equilibrium iterations as a function of the element type. . . . . 101
4.1 The time averaged drag coecient C
d
, the RMS of the drag
coecient C
RMS
d
, the RMS of the lift coecient C
RMS
l
, the
uctuation C
d
of the drag coecient, the uctuation C
l
of the lift
coecient and the separation angle
s
obtained on meshes A
1
, A
2
,
A
3
and A
4
and the dierent approximated relative errors. . . . . . 131
4.2 Comparison of the Reynolds number Re, the Strouhal number
St, the separation angle
s
, the base pressure coecient C
b
p
, the
minimum pressure coecient C
min
p
, the drag coecient C
d
and
the adverse pressure recovery APR with numerical results from the
literature. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
5.1 Density ratios
s
/
f
between some typical civil engineering materials
and air or water. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
5.2 Eigenfrequencies (in Hz) computed with the coarser and the
validated three-dimensional nite element model. . . . . . . . . . . 186
xxxix
xl LIST OF TABLES
A.1 Overview of the dierent possibilities for CFD computations within
Flotran 10.0 and CFX 10.0. . . . . . . . . . . . . . . . . . . . . . . 226
A.2 Overview of the dierent possibilities to couple Ansys 10.0 with
Flotran 10.0 and CFX 10.0. . . . . . . . . . . . . . . . . . . . . . . 227
B.1 Comparison of the eigenfrequencies computed with the three-
dimensional nite element model and the nite strip models for
varying n and m = 1. . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Chapter 1
Introduction
1.1 Problem outline and motivation
1.1.1 Wind loading on civil engineering structures
The behaviour of civil engineering structures under wind loads is an important
matter of concern. Melchers (1987) mentions as prime cause of the failure
of structures the inadequate estimation of loading conditions or the inaccurate
assessment of structural behaviour. It is a trend in modern architecture to
construct more slender and lightweight structures with often complex geometries.
The actual wind distribution around structures with a complex shape cannot
be derived from the literature or from the Eurocode (BIN, 1995). Moreover,
the loadings specied in the Eurocode are generally based on wind tunnel tests
performed on free-standing structures in open surroundings. Wind loads on
buildings in realistic environments may be considerably dierent due to the
presence of neighbouring structures.
For lightweight structures, the ratio between the wind load and the dead load
carried by the structure is substantially larger than for conventional structures.
Therefore, wind loading has become an important design load case. The increase
in span often results in more exible structures with lower eigenfrequencies that
are more susceptible to dynamic vibrations. The shape of shell structures is often
designed to carry the static dead loads by normal forces, while the additional wind
loading causes considerable bending moments in the structure.
For relatively rigid structures, the wind load can be determined for the undeformed
shape of the structure. This is no longer the case for exible structures, where
1
2 INTRODUCTION
the wind pressures cause large structural displacements. The resulting change of
shape has an inuence on the wind pressure distribution around the structure.
This may give rise to structural instabilities, such as utter of bridges, vortex
induced vibrations of tall buildings, galloping of cables and ovalling of silos (Simiu
and Scanlan, 1986).
The lack of knowledge about wind-structure interaction eects has led to serious
disasters in the past. The Tacoma Narrows bridge (Washington, US) collapsed
in 1940 due to utter (gures 1.1(a) and (b)). The H-shape of the section of
the bridge resulted in a poor aerodynamic behaviour and made the bridge very
sensitive to wind excitation.
(a) (b)
Figure 1.1: (a) Flutter of the Tacoma Narrows suspension bridge and (b) its nal
collapse (Champneys, 2008).
In 1965, three of a group of eight cooling towers at Ferrybridge (UK) collapsed,
while the others were severely cracked (gures 1.2(a) and (b)). The cooling towers
had been built closer together than usual and the design wind loads were based
on experiments for one isolated tower (Gunn and Malik, 1966). The towers that
collapsed were not situated on the row facing the wind but in the row sheltered
from the wind (the wind was blowing from the right in gure 1.2b).
These two examples illustrate that the behaviour of civil engineering structures
under wind loading is dicult to predict.
1.1.2 A wind-structure interaction problem: ovalling of silos
During a storm on October 27, 2002, wind induced ovalling oscillations were
observed on several empty silos of a group consisting of forty silos (gure 1.3),
located in the port of Antwerp. These oscillations were estimated to have large
amplitudes of several centimeters. During the storm the hourly average wind speed
in Deurne (about 7 kilometers to the east) ranged from 61 to 68 km/h with peak
wind speeds up to 113 km/h. The wind direction was west-southwest.
PROBLEM OUTLINE AND MOTIVATION 3
(a)
(b)
Figure 1.2: (a) Collapse of one of the cooling towers at Ferrybridge (UK) and (b)
the nal destruction of three cooling towers (Norfolk, 2008).
The silos are circular cylindrical shell structures with a diameter of 5.5 m and a
height of 25 m. One cylinder consists of 10 aluminium sheets with a height of 2.5 m
and a thickness that decreases with the height from 10.5 mm at the bottom to 6 mm
at the top. The height-to-radius ratio h/R = 9.1 and the radius-to-thickness ratio
ranges from R/t = 262 at the bottom to R/t = 458 at the top.
Figure 1.3: The silo group in the port of Antwerp.
4 INTRODUCTION
The forty silos of this group are placed in ve rows of eight silos with gaps of 30 cm
between two neighbouring silos (gure 1.4). The spacing ratio of the distance P
between the center of two cylinders to the cylinder diameter is P/D = 5.8/5.5 =
1.05. The bottom of the silos is located at 16.66 m above ground level.
N
D
P
L
W
h = 25 m
16.66 m
Figure 1.4: Plan and lateral view of the silo group.
In ovalling, the cross section of a circular cylindrical structure deforms as a shell
without bending deformation of the longitudinal axis of symmetry. The deformed
shape consists of a number of waves in the circumferential direction and a number
of waves in the axial direction. Similar cases in Germany and Scotland indicate
that storm damage is mainly located on silos at the corners of the group.
The ovalling oscillations of this group of silos in the port of Antwerp will be used
as an example throughout the thesis. The ovalling phenomenon is described in
this introductory section in order to clarify the objectives of the experiments and
simulations in the following chapters.
1.2 State-of-the-art
1.2.1 Response of structures to wind loading
In this section an overview of possible methods to determine the response of
structures to wind loading is given. First the experimental techniques for wind
tunnel testing are reviewed. Next, the numerical simulations are described.
Experimental techniques: wind tunnel tests
Rigid structures For relatively rigid structures, the behaviour under wind loading
is split in two separate steps: the determination of the pressure distribution around
the structure and the response of the structure to this loading.
STATE-OF-THE-ART 5
Usually, wind tunnel tests on rigid scale models are used to determine wind loads
on buildings with complex geometries. The surroundings of the structure can
easily be included in this model. It is often impossible to correctly scale both the
mean wind prole and the turbulent Reynolds number Re = Dv/, where D is
a characteristic length, v the uid velocity and the kinematic viscosity. If the
ow is dependent on the Reynolds number (e.g. for circular or rounded shapes),
corrections should be made in the conversions to full scale results. To improve the
local ow similarity, curved surfaces are usually roughened. The preparation of
these tests is time-consuming and expensive. It is possible to cover a wide range
of wind speeds, directions and turbulence intensities in a short time. Only global
pressure distributions are obtained. A change in the design requires a new model
and a complete re-run of the experiments.
The response of the structure is computed by applying the measured pressure
distributions to a numerical model of the structure. The numerical model yields
stresses and forces in all structural members. Changes in the structural design
that do not inuence the outer shape of the structure are easily evaluated.
Orlando (2001) used a linear nite element model to compute the quasi-
static response of isolated and grouped cooling towers under measured pressure
distributions. Portela and Godoy (2007) performed an eigenvalue analysis and a
geometric non-linear analysis on a nite element model of one of a group of steel
storage tanks to determine the buckling load. The applied pressure distributions
were obtained from a wind tunnel test on the group conguration. Lazzari et al.
(2003) used pressure coecients obtained by wind tunnel tests to dene pressure
time histories that correspond to articially generated space-varying velocity time
histories. A dynamic geometrically non-linear nite element analysis determines
the response of a tensegrity stadium roof under these pressure time histories.
Flexible line-like structures For slender, line-like structures with a relatively
rigid cross section which is uniform along the length, as long-span bridges, the
wind tunnel test can be performed on a model which represents a fraction of
the total length. In the wind tunnel, a rigid section model of the structure is
moved harmonically with a controlled amplitude and frequency in a uniform ow
in order to determine the resulting aerodynamic forces. From these measurements
aerodynamic or utter derivatives are obtained for a number of dimensionless
frequencies. The reduction of the model to a part of the structure enables the use
of larger scale models at higher Reynolds numbers.
These aerodynamic derivatives are used to calculate the response of the structure
by means of an analytical or a numerical model. The simplest structural model
consists of a mass and a mass moment of inertia supported by a vertical spring
and a rotational spring. This model has only two eigenmodes and the response
is computed in the frequency domain (Simiu and Scanlan, 1986). Current state-
6 INTRODUCTION
of-the-art techniques use time domain methods in order to account for the non-
linearity in the aerodynamic forces due to the change in angle of incidence (Chen
and Kareem, 2001). Simple two degree-of-freedom models of the structure are
replaced by nite element models which include all contributing eigenmodes and
are able to account for non-linear structural behaviour (Zahlten and Eusani, 2006).
Alternatively, if the rigid section model is mounted dynamically with a vertical and
a rotational spring, the utter wind speed can be determined from wind tunnel
tests.
Flexible 3D structures If the structural deformations are large enough to
substantially change the pressure distribution around the structure, wind tunnel
tests on fully aeroelastic models are a possible approach. In the aeroelastic model,
the mass, stiness and damping of the structure should be scaled correctly. This
is especially dicult if several eigenmodes contribute to the structural response.
For shell structures, similarity requirements have to be fullled for the Poissons
ratio as well. Wind tunnel test results often have to be corrected to represent
the full-scale behaviour. Vickery and Majowiecki (1992) studied the response of
a cable supported stadium roof on an aeroelastic model. Niemann and Kopper
(1998) measured strains in a cooling tower by means of an aeroelastic wind tunnel
test. The pressures and deformations can generally only be measured in a limited
number of points.
Numerical simulations
CFD simulation of wind ow around rigid structures The uid ow around
rigid structures can be simulated numerically with Computational Fluid Dynamics
(CFD). Wind ows around buildings are very high-Reynolds-number ows with
unsteady separation and vortex shedding. This requires an appropriate turbulence
model. Large eddy simulations (LES) are perfectly suited to capture these large-
scale unsteadiness, but the computational cost is very high. Reynolds averaged
Navier-Stokes (RANS) simulations are much cheaper and very often used to
compute these ows, but are, especially in the wake, less successful (Stathopoulos,
2002). For a relatively simple shape, the averaged physical quantities, such
as the drag and the lift coecients, can be computed with sucient accuracy.
The numerical prediction of peak pressures or spatial correlations of uctuating
pressures with LES is still challenging. The application of hybrid RANS/LES
methods is very promising as they are able to capture the large scale turbulent
structures in the wake of blu bodies at a more aordable computational cost
(Spalart, 2000).
STATE-OF-THE-ART 7
These numerical simulations have the advantage of yielding much more detailed
pressure distributions than in the case of wind tunnel testing. However, due to the
high computational cost, quite coarse meshes and less expensive but less suited
turbulence models are often used. Therefore, the results are not always reliable
and accurate.
Numerical simulation of uid-structure interaction If the interaction between
the uid and the structure is substantial, a CFD model of the ow can be coupled
to a numerical model of the structure. Due to the increase in computing power,
the numerical simulation of uid-structure interaction (FSI) problems has gained
a lot of interest in the past decade. The main areas of interest are aeronautical
engineering, biomechanics and civil engineering. The aerospace applications
(Dowell and Hall, 2001; Piperno and Farhat, 2001; Willcox and Peraire, 2002;
Geuzaine et al., 2003) are mainly related to utter of wings, while in biomechanics
(Vierendeels et al., 2000; Baaens, 2001; Gerbeau and Vidrascu, 2003) blood ows
through the heart or the arteries or air ows in the lungs are studied. Within
the context of civil engineering, challenging problems include the wind response
of tensile structures (Gl uck et al., 2001, 2003; Haug et al., 2003; H ubner et al.,
2004; Bletzinger et al., 2006), utter of bridges (H ubner et al., 2002; Fourestey
and Piperno, 2004), vortex induced vibrations of marine riser pipes (Willden and
Graham, 2004) or cables (Meynen et al., 2005) and ovalling of silos (Dooms et al.,
2007).
Numerical methods used for uid-structure interaction can be classied according
to the complexity of the models and techniques employed in the structure
and the uid (L ohner and Cebral, 1996). For the structure, the following
models might be used (in increasing order of complexity): a system of springs,
masses and dampers; linear nite element model; non-linear nite element model.
For the uid, typical options would be: potential ow; inviscid ow (Euler
equations); Reynolds averaged Navier-Stokes equations; large eddy simulations;
direct numerical simulations. Nowadays, the available computational resources
limit the computations at high Reynolds numbers to the coupled simulation of
Reynolds averaged Navier-Stokes equations with a non-linear nite element model.
These numerical simulations have the advantage of yielding pressure distributions
and structural displacements and stresses for all elements, which is much more
detailed than in the case of wind tunnel testing. The main disadvantage is
the computational cost as the calculations do not run in real-time: they take
much longer than the simulated time period. The determination of appropriate
boundary conditions (e.g. velocity proles, wall roughnesses) and turbulence
model parameters is often dicult. The current accuracy and reliability of uid-
structure interaction computations implies that they are not competing with wind
tunnel testing, but are rather highly complementary: the use of these calculations
in an early design phase can reduce the required number of wind tunnel tests.
8 INTRODUCTION
1.2.2 Ovalling oscillations
As the ovalling oscillations of the group of silos in the port of Antwerp will be used
as an example throughout the thesis, the state-of-the-art with respect to ovalling
oscillations of cylindrical shell structures is briey described in this section.
Ovalling oscillations were rst reported in the fties for steel chimney-stacks
(Dockstader et al., 1956). The ovalling frequency was observed to be close to twice
the estimated vortex shedding frequency. Analogously to swaying oscillations of
beam-like structures, ovalling oscillations were believed to be excited by periodic
vortex shedding.
Wind tunnel tests
A lot of experimental research on ovalling has been carried out on scale models
in wind tunnels. Johns and Sharma (1974) concluded from wind tunnel tests
that the ovalling frequency of the structure is equal to an integer multiple of the
vortex shedding frequency. The vortex shedding frequency f was not measured but
assumed to lie in the range corresponding to Strouhal numbers St = fD/v between
0.166 and 0.2, where D is the diameter and v the ow velocity. The ovalling
frequency of the structure is very close or identical to one of its eigenfrequencies.
Small discrepancies are probably due to the static deformation of the structure
under the wind pressures or the structural geometrically non-linear behaviour.
The deformed shape has a node or an antinode facing the ow direction. Ovalling
starts at a threshold (or onset) ow velocity and its amplitude increases up to
a peak ow velocity, which is substantially higher than the threshold velocity.
Sometimes more than one eigenmode of the structure is excited simultaneously
and the eigenmode that dominates the response changes with the ow velocity.
Padoussis and Helleur (1979) measured the ovalling frequency as well as the
vortex shedding frequency in the wake. At the onset the ovalling frequency of
the structure was equal to an integer multiple of the vortex shedding frequency.
However, with increasing wind speeds, the ovalling frequency remained the same,
while the vortex shedding frequency increased in accordance with a constant
Strouhal number. The integer relationship between both frequencies ceased to
hold, which meant that no lock-in occurred as in the case of swaying oscillations
of beam-like structures, where vortices continue being shed at the eigenfrequency
of the structure if the ow velocity is increased. In a second series of experiments,
the periodic vortex shedding was suppressed by mounting a splitter plate behind
the cylinder. In this case, ovalling still occurred but at a slightly higher onset
velocity. These two observations suggested that periodic vortex shedding is not
the cause of the ovalling. In a third series of experiments the ow pattern in the
wake was altered by gluing exible plastic tubes on the cylinder wall in its wake.
Ovalling oscillations were totally suppressed.
STATE-OF-THE-ART 9
New experiments by Padoussis et al. (1982b) showed that, at the onset of ovalling,
the ovalling frequency of the structure is not always equal to an integer multiple of
the vortex shedding frequency, although this often occurs. It was suggested that
ovalling is an aero-elastic utter in a single eigenmode, which is associated with
negative aerodynamic damping (Padoussis et al., 1983). It should be distinguished
from classical utter of wings and bridges where two eigenmodes of the structure
couple together during the oscillations. If the energy extracted from the ow
exceeds the energy dissipated in the structure, the response of the structure to an
initial disturbance diverges. Ovalling especially occurs on thin welded structures
made of high strength steel or aluminium which have a low internal damping.
The often encountered integer relationship between the ovalling frequency and
the vortex shedding frequency at the onset of ovalling might be explained as
follows: ovalling starts at a slightly lower velocity because of the deformations
of the structure due to the periodic vortex shedding.
As also silos experience ovalling oscillations, Katsura (1985) performed wind
tunnel tests in a uniform ow at dierent turbulence intensities on cylinders with
height-to-radius and radius-to-thickness ratios typical for silos. The eective total
damping of the structure was measured at dierent ow velocities and reduced
to almost zero at the onset velocity where the vibration amplitudes increase
suddenly. At the lowest turbulence intensity, a negative phase lag between shell
displacements and the wind pressures occurred above the onset velocity, which
corresponds to a negative aerodynamic damping. At the peak ow velocity, the
frequency component in the pressure eld close to the cylinder at the ovalling
frequency becomes dominant compared to the component at the vortex shedding
frequency. At the highest turbulence intensity, the phase lags were always positive
and vibration amplitudes increase gradually in proportion with the pressure, which
is similar to turbulent bueting. Uematsu and Uchiyama (1985) determined
from the cross-correlation of uctuating pressures in two points in the wake of
a rigid cylinder that the pressure uctuations in the wake are convected towards
the separation point with a velocity that depends on their frequency and the
free stream velocity. This correlation might trigger the ovalling phenomenon
for eigenmodes with higher circumferential mode numbers. The eigenmode with
lowest eigenfrequency is not always excited rst.
Experiments by Panesar and Johns (1985) did not show any lock-in between the
ovalling frequency and the vortex shedding frequency. The presence of a splitter
plate in the wake changed the ovalling onset velocity, but surprisingly did not
suppress the vortex shedding. Padoussis et al. (1991) showed that stiening the
upstream or the side part of the cylindrical shell changes the onset velocity, while
stiening the leeward side totally eliminates the ovalling, which proves that the
behaviour in the wake is determining. Laneville and Mazouzi (1995) obtained
dierent amplitudes for the ovalling vibrations at a certain ow velocity depending
on if the velocity is increasing or decreasing with time.
10 INTRODUCTION
Semi-analytical models
Simultaneously, semi-analytical models were developed to predict the onset
velocity of ovalling (Padoussis et al., 1982a, 1991; Mazouzi et al., 1991; Laneville
and Mazouzi, 1996). The linear behaviour of a cylindrical structure made
of an elastic, homogeneous and isotropic material is described by Donnells
or Fl ugges theory (Kraus, 1967). The pressures acting on the structure are
obtained from the superposition of a steady and an unsteady ow. The steady
part is represented as a potential ow which is tuned to experimental values.
The separation angle is not aected by the oscillations. The unsteady part is
calculated as a potential ow caused by small amplitude vibrations. The eects
of periodic vortex shedding and turbulence are neglected. If the coupling between
dierent eigenmodes during ovalling is neglected, the resulting eigenvalue problem
yields the ovalling frequencies and the amount of aerodynamic damping. The
ovalling frequencies are almost identical to the eigenfrequencies of the structure.
Eigenmodes with an even circumferential wavenumber have negative aerodynamic
damping if a node faces the free stream direction and positive damping if an
antinode faces the free stream direction. Eigenmodes with an odd circumferential
wavenumber have negative aerodynamic damping if an antinode faces the free
stream direction and positive damping if a node faces the free stream direction.
The aerodynamic damping increases with the free stream velocity. If the energy
gained from the ow equals the energy dissipated by the structure, ovalling sets
on. The accurate experimental determination of the structural modal damping
is therefore very important. Padoussis et al. (1991) included in the wake of
the cylinder experimentally determined changes in the base pressure due to the
shell deformations and the phase lag between the shell motion and the pressure
in order to improve the predicted onset ow velocities. Mazouzi et al. (1991)
used for the steady ow the potential ow model of Parkinson and Jandali (1970)
which requires experimental values of the separation angle and the base pressure
coecient. Through these values, the dependence of the pressure distribution
on the Reynolds number is introduced. Laneville and Mazouzi (1996) simplied
these semi-analytical models and obtained an expression for the onset velocity
as a function of the eigenfrequency and the Scruton number. As the simplied
aerodynamic damping is independent of the circumferential wavenumber and the
energy dissipated by structural damping during ovalling is proportional to the
product of the modal logarithmic decrement and the square of the eigenfrequency,
the eigenmode with the lowest value for this product is excited rst. These
analytical models generally require a substantial amount of experimental data like
the structural modal damping ratio, the separation angle and the base pressure
coecient in order to obtain a good correspondence between measured and
predicted onset velocities. They do not provide physical insight in the occurring
ow phenomena.
OBJECTIVES AND ORIGINAL CONTRIBUTIONS 11
Numerical simulations
W uchner et al. (2005) and Bletzinger et al. (2006) numerically simulated ovalling
oscillations by coupling the computation of the two-dimensional wind ow with
a nite element model of a ring, which corresponds to an innitely long circular
cylinder. The three-dimensional behaviour in the structure, the ow and the
interaction are neglected. The eigenfrequencies and the sequence of the eigenmodes
of an innitely long cylinder dier considerably of cylinders of nite length.
Any choice made for the boundary conditions of the ring structure changes the
eigenfrequencies of some eigenmodes. The dierences in dimensions, material
properties and wind speed render the validation with the experimental results
from Johns and Sharma (1974) impossible.
1.3 Objectives and original contributions
The behaviour under wind loading of civil engineering structures that are quite
exible, have a complex geometry or are located in the presence of neighbouring
structures is very dicult to determine. Nowadays, the common approach is
to perform a series of wind tunnel tests. The aim of the thesis is to study
coupled numerical simulations of the uid and the structure to determine this
behaviour. The dierent numerical methods applied in the individual elds (uid
and structure) and in the coupling procedure are explored. More specically, the
work focusses on the simulation of wind-structure interaction phenomena occurring
on shell structures. As the present thesis is the rst work dealing with CFD and
FSI in the research group, one of the objectives is to give in each chapter an
extensive and coherent overview of the related theory.
With regard to the example of wind induced ovalling oscillations of a silo, the
current work will focuss on the creation of accurate models for the uid and
structural eld and on the computation of the ovalling oscillations for a single
silo. Future work should investigate the inuence of the small distance between
two neighbouring silos on the occurrence of ovalling and propose design alterations
in order to avoid the occurrence of the oscillations.
1.3.1 Methodology
The methods to simulate wind-structure interaction phenomena occurring on shell
structures will be introduced by means of the example of wind induced ovalling
oscillations of a silo. This dynamic uid-structure interaction problem is estimated
representative for the problems under consideration. Although it has a simple
circular cylindrical geometry, it covers all aspects specic for a general wind-shell
12 INTRODUCTION
interaction problem: the geometrical non-linear deformations of a shell structure,
the turbulent wind ow with separation and nally the self-excited phenomenon,
that can only be assessed numerically by methods which take the interaction
between uid and structure into account. The adapted methodology is general
and can also be applied to the problems mentioned in the previous sections.
For the numerical simulation of uid-structure interaction a partitioned approach
is used: the uid and the structure are treated as isolated elds and solved
separately. The interaction eects are applied on the individual elds as external
boundary conditions which are exchanged from one eld to the other. The
main advantage of this partitioned approach is that well-established discretization
techniques and solution algorithms (and by consequence existing software) which
are tailored to the characteristic behaviour of the individual elds can be used
in each eld. As the present thesis is the rst work dealing with CFD and FSI
in the research group, it was a conscious choice at the beginning to make use
of existing software: Ansys (Ansys, 2005b) for the structural computations and
Flotran (Ansys, 2005b) and CFX (Ansys, 2005a) for the ow computations. By
consequence, the possible choices regarding models and algorithms were restricted
to those available in these packages (appendix A). As the uid ow equations
(e.g. the Navier-Stokes equations) are non-linear and the non-linear behaviour of
the structure should be taken into account, direct time integration is used in both
elds.
Prior to the study of uid-structure interaction phenomena, a thorough knowledge
of the structural and the uid eld is required, for which accurate numerical models
are needed.
The structure is discretized by means of the nite element method and integrated
in time using the Newmark time integration scheme. The formulation takes into
account geometrically non-linear behaviour and allows for large rotations and large
displacements. As for the prediction of ovalling oscillations the exact value of the
structural modal damping is very important, an in situ experiment is performed
in order to determine the modal damping ratios of the silos. The measured
eigenfrequencies and eigenmodes are used to validate a three-dimensional nite
element model of a silo.
For the uid a nite volume discretization is applied and the solution is integrated
in time using the three-point backward dierence method. A two-dimensional
computation of the ow around one rigid cylinder is made. The eect of the
close spacing of cylinders on the ow pattern is studied for a group of rigid
cylinders as in the port of Antwerp. For the FSI computations, the mesh in the
uid eld is chosen to be aligned with the deforming boundary of the structure.
The uid computations are performed on a deforming mesh and the governing
equations are formulated in an arbitrary Lagrangian-Eulerian description. The
mesh deformation is computed by means of a diusion equation with a variable
OBJECTIVES AND ORIGINAL CONTRIBUTIONS 13
diusivity and should preferentially preserve the quality and the renements of
the mesh.
Between Ansys and Flotran and between Ansys and CFX, a coupling algorithm
is commercially available. Two partitioned solution algorithms are employed: the
loosely coupled serial staggered algorithm which solves each eld one time in each
time step and a strongly coupled staggered algorithm which iterates between the
two elds in each time step. The meshes of the uid and the structure are non-
matching at the uid-structure interface. A consistent interpolation method is
used for the displacement transfer, while a node to element conservative method
is applied for the load transfer. The interaction between the wind ow and a single
silo is computed.
1.3.2 Original contributions
A three-dimensional nite element model of a circular cylindrical silo has been
validated by means of in situ experiments performed on a single silo. Radial
accelerations at 10 points along the silo are measured under ambient wind loading
and modal parameters are extracted from the output-only data using the stochastic
subspace identication technique. The inuence of the boundary conditions on the
eigenfrequencies and mode shapes is investigated.
The current state-of-the-art methods for the simulation of turbulent ows are
reviewed as a function of their performance and their computational requirements.
The inuence of turbulence models, near-wall mesh renement and unsteadiness
is studied for the two-dimensional ow around a single cylinder in the post-critical
regime. The results are compared with experimental data and numerical results
available in the literature. The two-dimensional ow around a group of 2 by 2
and of 8 by 5 cylinders is computed and shows a strong inuence of the group
conguration on the pressure distribution around the cylinders.
A coherent and comprehensive overview of the whole set of numerical techniques
required for the simulation of dynamic uid-structure interaction is given. For
the uid a boundary-tted mesh is used together with the arbitrary Lagrangian-
Eulerian description. The governing equations for the ow in an arbitrary
Lagrangian-Eulerian description are derived. The implication of this description
on the time-accuracy and the stability of the time integration is discussed. The
dierent options to compute the uid mesh velocity and deformation are reviewed.
The coupled problem is solved by a partitioned algorithm, where the uid and the
structure are separately integrated in time and the interaction eects are applied
as external boundary conditions. The accuracy, stability and eciency of loosely
and strongly coupled algorithms is discussed in the case of incompressible ows.
Dierent methods for load and motion transfer between non-matching meshes are
described as a function of their accuracy and conservation properties.
14 INTRODUCTION
As a practical application, the interaction between the wind ow and a single silo
is computed. The three-dimensional nite element model of the silo is coupled
with the three-dimensional incompressible turbulent wind ow as to predict
ovalling oscillations. The results computed with the loosely coupled conventional
serial staggered algorithm show dierences that increase in time with the results
computed with the strongly coupled algorithm using subiterations, which is more
accurate.
1.4 Organization of the text
Prior to the study of wind-structure interaction phenomena, a thorough
understanding of the structural and the uid behaviour is required, for which
accurate numerical models are needed. Much eort has therefore been devoted to
the solution of the separate problems. These models are used in chapter 5 for a
coupled numerical analysis of the wind ow around the structure that simulates
the ovalling phenomenon.
Chapter 1 situates the subject of the thesis and highlights the objectives and
original contributions. The ovalling of cylindrical shell structures is introduced.
This dynamic uid-structure interaction problem is considered as an example
throughout the thesis.
Chapter 2 describes the arbitrary Lagrangian-Eulerian description for ow
computations on domains with deforming boundaries. The concepts of
stabilization for nite element discretizations and upwinding for nite volume
discretizations, which are needed for ow computations, are presented. The
stability and the time accuracy of computations on domains with deforming
boundaries are discussed and related to the geometric conservation law. The
methods to compute the mesh deformation are reviewed.
Chapter 3 reviews the governing equations for linear and geometrically non-
linear behaviour of shell structures. In situ experiments performed on a single
silo in order to obtain the eigenfrequencies, eigenmodes and modal damping ratios
are described. A three-dimensional nite element model of a silo is validated by
means of these experimental modal parameters. The importance of geometrically
non-linear behaviour during ovalling of the silos is evaluated.
ORGANIZATION OF THE TEXT 15
Chapter 4 treats the computation of turbulent wind ows around rigid civil
engineering structures. Existing turbulence models are reviewed. The results
of steady and unsteady Reynolds averaged Navier-Stokes simulations of the ow
around a single silo are compared with experimental data and few available
numerical results in order to select a suitable turbulence model. The computation
of the ow around a closely spaced group of 8 by 5 cylinders illustrates the inuence
of the group conguration on the vortex shedding and the pressure distributions
around the cylinders.
Chapter 5 describes the partitioned algorithms to solve the coupled dynamic
uid-structure interaction problem by computing the dierent elds sequentially
or in parallel. The accuracy, stability and eciency of these algorithms is discussed.
The transfer of displacements and loads between non-matching discretizations of
the uid and the structure at the interface is treated. The conservation of the total
load and of the energy at the interface is of importance. A coupled simulation of
a three-dimensional wind ow around a single deforming silo is performed. More
ecient coupling procedures should enable simulations during a much longer time
interval.
Chapter 6 summarizes the conclusions of the thesis and presents recommenda-
tions for further research.
Chapter 2
Flow on a domain with
deforming boundaries
2.1 Introduction
As the aim is to perform a coupled numerical simulation of the uid and the
structure, the position of the structure determines at least partially the uid
domain boundaries. If the structure undergoes large displacements, it is necessary
to perform the computations of the uid ow on a domain with moving boundaries.
Many dierent approaches exist for the computations of ows on a domain with
deforming boundaries.
In a rst approach, the deforming boundary is described explicitly. The mesh of the
uid is aligned with the deforming boundary (gure 2.1a). As the boundary moves,
the mesh should follow this boundary. This can be achieved by deforming the mesh
without modifying its topology. If, however, the mesh becomes too distorted, the
domain has to be remeshed. In order to compute ow on a deforming mesh,
the governing equations should be written in an arbitrary Lagrangian-Eulerian
description (Hughes et al., 1981; Donea et al., 1982). If this approach is used to
calculate free surface ows, it is called an interface tracking technique.
In a second approach, the deforming boundary is described implicitly. The ow
computation is performed in an Eulerian description on a grid which is xed in
space and covers the union of the uid and the solid (or empty) domain. This
extended domain has usually a simpler geometrical shape and does not change with
time. The solid domain is allowed to move with respect to the xed grid (gure
2.1b), so the uid elements that contain the solid domain change continuously.
17
18 FLOW ON A DOMAIN WITH DEFORMING BOUNDARIES
The need for remeshing or mesh deformation is eliminated. As the meshes are
not boundary-tted, structured rened boundary layer meshes next to the solid
boundary are not possible. The governing equations of the uid domain are altered
in order to reect the presence of the solid domain. In the immersed boundary
method (Peskin, 1972; Peskin and Mcqueen, 1980) a body force is added to the uid
equations. This force is equal to a Dirac delta function which diers from zero at
the uid boundary. The ctitious domain method (Glowinski et al., 1994; Baaens,
2001; De Hart et al., 2003) imposes the boundary conditions for the velocities at
the uid-solid interface using Lagrange multipliers. In the extended nite element
method (X-FEM) (Gerstenberger and Wall, 2008), the approximation space is
enriched with additional functions that might be discontinuous as well. In this
way the discontinuities in the pressure and the uid eld at the location where
a thin structure is present, can be included. Applied to free surface ows, these
techniques are called interface capturing techniques.
(a) (b)
Figure 2.1: (a) Explicit and (b) implicit description of a deforming boundary of
the uid domain.
Combinations of the above approaches exist as well. For uid ows around moving
and rotating rigid bodies, the Chimera technique (Steger et al., 1983) is often used.
The uid domain consists of two meshes which partially overlap: a background
mesh which is xed in space and a mesh which moves together with the rigid
body. This approach can be extended to a technique (Wall et al., 2006) where a
background mesh is xed in space and the other mesh moves and deforms together
with the structure using an arbitrary Lagrangian-Eulerian description.
The particle nite element method (PFEM) (Idelsohn et al., 2004) uses the
positions of a number of particles to generate a new mesh at every time step.
The particles correspond to the uid nodes. If the distance between the particles
is too large, there is no element created between these particles. In this way
the domain may split in several parts or individual particles may leave the uid
FLOW ON A DOMAIN WITH DEFORMING BOUNDARIES 19
and possibly join later again. As the uid ow is computed using a Lagrangian
formulation, convective terms are not present in the governing equations. This
method is especially suited to solve free surface problems, breaking waves and
uid particle separation.
For the applications in mind, no uid particles are leaving the uid domain nor
does the structural domain split due to fractures or explosions. The structural
deformations are expected to be large, but no rotations of 360 degrees will
occur. Therefore, a boundary-tted mesh can be used together with the arbitrary
Lagrangian-Eulerian description without frequent remeshing. Its ability to have
structured boundary layer meshes near the structure is very important. First,
the governing equations for the uid ow in an arbitrary Lagrangian-Eulerian
description are derived. The spatial discretization by means of the nite element
and the nite volume method is discussed. Using an Eulerian description, the time
integration is straightforward. In the case of an arbitrary Lagrangian-Eulerian
description, specic choices have to be made for the time integration in order to
preserve the time accuracy and the stability. Finally, dierent options to compute
the uid mesh deformations are reviewed.
2.2 Flow on a domain with deforming boundaries
2.2.1 The arbitrary Lagrangian-Eulerian description
Two classical kinematic descriptions are extensively used in continuum mechanics:
the Lagrangian and the Eulerian description. The arbitrary Lagrangian-Eulerian
description (ALE) was developed to combine the advantages of both descriptions.
In a Lagrangian description (gure 2.2a) an observer is xed to a material particle
and follows its motion. This permits easy tracking of free surfaces and interfaces
between dierent materials where boundary conditions are conveniently described.
Physical quantities of a material particle are easily obtained. As the observer
follows the motion of material particles, the material time derivative reduces to
a simple time derivative and convective terms do not occur in the governing
equations. After discretization of these governing equations, the nodes of the
mesh function as observers. Each element of the mesh contains at each time
the same material particles, which enables to incorporate time dependent material
behaviour. Conservation of mass is automatically satised. However, in non-linear
computations large distortions of the continuum result in severe deterioration of
the quality of the mesh and loss in accuracy. This description is most commonly
used in structural analysis.
In an Eulerian description (gure 2.2b) the observer is xed in space. Physical
quantities are examined in a xed point in space as material particles pass. The
material time derivative introduces convective terms in the governing equations.
The material particles move with respect to the mesh and the continuum is allowed
to undergo large distortions without loss of accuracy. Tracking of free surfaces and
interfaces is more dicult. For uid analysis, an Eulerian description is often used,
although a Lagrangian description might be used for contained uids which only
undergo small motions.
The ALE description (gure 2.2c) combines the advantages of the two descriptions:
the observer moves arbitrarily through space. At an interface between dierent
materials or at a free surface, the observer may follow the material particles, while
at an inlet or outlet of the domain the observer may stay xed in space. Inside the
domain, the movement of the nodes (or the observers) can be chosen arbitrarily but
should preferentially preserve the regularity of the mesh. The dierent possibilities
to automatically specify suitable movement of the nodes will be treated in section
2.5. In gure 2.2c the nodes of the left curved boundary follow the material
particles, while the nodes at the right straight boundary are xed in space. The
nodes in between these boundaries neither follow the material particles, neither
are xed in space.
t
x
1
x
2
(a)
t
x
1
x
2
(b)
t
x
1
x
2
(c)
Figure 2.2: Representations of (a) the Lagrangian, (b) the Eulerian and (c) the
arbitrary Lagrangian-Eulerian descriptions in the spatial domain
x
at two dierent times t: continuum (dark gray), material particles
(black dots) and mesh (black lines). Dashed lines show the mesh
motion, while solid lines show the particle motion.
The ALE description was rst developed in the framework of the nite dierence
method by Hirt et al. (1974). Hughes et al. (1981) and Donea et al. (1982)
described the rst implementations for the nite element method. The ALE
description was primarily used to compute ows on domains with moving
boundaries like free surfaces or uid-structure interfaces. Its use has been extended
to non-linear structural mechanics to deal with crack propagation, impacts,
explosions, penetrations and forming processes (Liu et al., 1988), for which the
ALE description copes with the signicant distortions of the material.
Computations using the ALE description are performed in the referential domain
where the mesh is xed at all time steps (gure 2.3): a node has the same
referential coordinates at each time t. The mapping denes for a node with
referential coordinates what the position x in space is of the material particle
present at this node at time t:
:
[t
0
, [
x
: (, t) x = (, t) (2.1)
The spatial domain
x
(t) which corresponds at time t + t to the referential
domain
is indicated by a dashed line. The mapping can be interpreted

as the position of the nodes in the spatial domain. The time derivative of this
mapping with the referential coordinate held xed is denoted as v and describes
the velocity of the nodes of the mesh:
v =
x(, t)
t
(2.2)
The initial positions X where the material present in the mesh at time t was
located at time t = 0 determine the material domain
X
. The material domain
X
(t) at time t + t is indicated by a dashed line. The mapping determines
for a material particle with initial position X at time t = 0 what the referential
coordinates in the mesh are at time t:
:
X
[t
0
, [
: (X, t) = (X, t) (2.3)

This mapping can be interpreted as the position of the material particles in
the referential domain. The time derivative of this mapping with the material
coordinate held xed is denoted as v and describes the particle velocity in the
referential domain:
v =
(X, t)
t
X
(2.4)
X

Figure 2.3: Domains, meshes and mappings for an ALE description: xed
referential domain
, changing material domain

X
and changing
spatial domain
x
at time t (solid lines) and t + t (dashed lines).
The material particle motion , typically used in structural mechanics, relates
the initial position of a material particle with its current position at time t and is
obtained as a combination of the mappings and :
= :
X
[t
0
, [
x
: (X, t) x = (X, t) = ( (X, t), t) (2.5)
The displacements u of a particle are obtained by subtracting the initial position
X form the spatial position x:
u = x X (2.6)
The time derivative of this mapping with the material coordinate held xed is
denoted as v and describes the particle velocity in the spatial domain:
v =
x(X, t)
t
X
(2.7)
In gure 2.3 the nodes of the left curved boundary follow the material particles and
coincide in material and referential domain, while the nodes at the right straight
boundary are xed in space and coincide in spatial and referential domain.
Both Lagrangian and Eulerian descriptions can be obtained as special cases of
the ALE description: in a Lagrangian description the material and the referential
domain coincide and = I. The mesh moves together with the material particles.
In an Eulerian description the spatial and the referential domain coincide and
= I. The mesh is xed in space.
The description of a scalar physical quantity f in the referential domain
involves
a complementary description g in the spatial domain:
g(x, t) = f(
1
(x, t), t) = f(, t) (2.8)
g = f
1
(2.9)
and a complementary description h in the material domain:
h(X, t) = f( (X, t), t) = f(, t) (2.10)
h = f (2.11)
For simplicity, the complementary descriptions g and h will be denoted as f(x, t) =
f(X, t) = f(, t) from here on.
The material time derivative (also known as the substantive derivative, the
substantial derivative, the total time derivative, the convective derivative, the
advective derivative or the Lagrangian derivative) appears in the basic conservation
laws for mass, momentum and energy. An expression for the material time
derivative of a scalar quantity f is obtained in terms of the referential time
derivative and a convective term taking into account the relative motion between
the material particle and the mesh:
f =
Df
Dt
=
f(X, t)
t
X
=
f(, t)
t
+
(X, t)
t
f(, t) (2.12)
The rst factor of the last term is equal to the particle velocity v in the referential
domain (equation (2.4)):
Df
Dt
=
f(, t)
t
+ v
f(, t) (2.13)
The application of this equation to the spatial coordinate x yields an expression
for the particle velocity v:
v =
Dx
Dt
=
x(, t)
t
+ ( v
)x(, t) (2.14)
The rst term on the right hand side is equal to the grid velocity v (equation
(2.2)):
v = v + ( v
)x(, t) (2.15)
The second term on the right hand side of equation (2.15) is equal to the convective
velocity c:
c = v v = ( v
)x(, t) =
x(X, t)
t
x(, t)
t
(2.16)
The convective velocity c is the relative velocity between material particles and
the mesh in the spatial domain and diers from the particle velocity v in the
referential domain. The convective velocity c and the particle velocity v in the
referential domain are equal if
x(, t) = I which is the case when the mesh

only translates without any rotation or deformation.
As mentioned before, both Lagrangian and Eulerian descriptions can be obtained
as special cases of the ALE description. If the grid velocity v is equal to zero,
the grid is xed in space and an Eulerian description is recovered. The convective
velocity c and the particle velocity v in the referential domain equal the material
particle velocity v. If, on the other hand, the grid velocity v is equal to the
material particle velocity v, the grid moves together with the material particles
and a Lagrangian description is retrieved. The convective velocity c and the
particle velocity v in the referential domain are zero.
In equation (2.12) the gradient of the quantity f in the referential domain has
to be computed. As constitutive relations are naturally expressed in the spatial
domain, equation (2.12) is rewritten as a function of the convective velocity c using
equation (2.16) and the chain rule:
f =
Df
Dt
=
f(X, t)
t
X
=
f(, t)
t
+ v
x
f(x, t)
x(, t) (2.17)
=
f(, t)
t
+c
x
f(x, t) (2.18)
This fundamental ALE equation shows that the material time derivative of a
scalar quantity f is equal to its referential time derivative and a convective term
consisting of the convective velocity c and its spatial gradient. Application to the
dierent components of the velocity v yields an expression for the material time
derivative of the particle velocity:
v =
Dv
Dt
=
v(, t)
t
+ (c
x
)v(x, t) (2.19)
The reader familiar with the derivation of the continuity and momentum equations
in the Eulerian, Lagrangian and ALE formulation may proceed immediately with
section 2.2.2.
Deformation gradients
The relation between an innitesimal vector dX in the material domain and in the
spatial domain dx is given by the deformation gradient F:
dx =
X
xdX = FdX (2.20)
An analogous deformation gradient is dened between the referential and the
spatial domain:
dx =
xd =

Fd (2.21)
and between the material and the referential domain:
d =
X
dX =

FdX (2.22)
Volume and area change
The innitesimal volume in the material domain with edges parallel to the
Cartesian axes is given as:
d
X
= dX
1
(dX
2
dX
3
) = dX
1
E
1
(dX
2
E
2
dX
3
E
3
) (2.23)
= dX
1
dX
2
dX
3
(2.24)
where E
1
, E
2
and E
3
are orthogonal unit base vectors. The same innitesimal
volume in the spatial domain is expressed as:
d
x
= dx
1
(dx
2
dx
3
) (2.25)
Using equation (2.20) the innitesimal vectors dx
1
, dx
2
and dx
3
in the spatial
domain are obtained in terms of the innitesimal vectors dX
1
, dX
2
and dX
3
in
the material domain:
d
x
= FdX
1
(FdX
2
FdX
3
) (2.26)
= (FE
1
(FE
2
FE
3
))dX
1
dX
2
dX
3
(2.27)
= det Fd
X
= Jd
X
(2.28)
where the determinant of the deformation gradient F denes the Jacobian J, which
relates the volume d
X
in the material domain to the volume d
x
in the spatial
domain.
Analogously, the Jacobian

J relates the volume d
in the referential domain to

the volume d
x
in the spatial domain:
d
x
= det
Fd
=

Jd
(2.29)
and the Jacobian

J relates the volume in the material domain d
X
to the volume
d
in the referential domain:

d
= det
Fd
X
=

Jd
X
(2.30)
Substituting equation (2.30) in equation (2.29) gives:
d
x
=

Jd
=

J

Jd
X
(2.31)
Comparison with equation (2.28) yields the following relation:
J =

J

J (2.32)
The time rate of change of the Jacobians is given by:
DJ
Dt
=
J
t
X
= J
x
v (2.33)
J
t
=

J
x
v (2.34)
D
J
Dt
=

J
t
X
=

J
v (2.35)
Expressing the innitesimal volumes d
x
and d
X
as the dot product of a vector
dl and an area d, equation (2.28) for the volume change between the material
and the spatial domain yields:
dl
x
d
x
= Jdl
X
d
X
(2.36)
Recalling equation (2.20), the vector dl
x
is expressed in terms of the vector dl
X
:
(Fdl
X
) d
x
= Jdl
X
d
X
(2.37)
As this expression is valid for any vector dl
X
, a relation between the innitesimal
areas in the material and the spatial domain is obtained:
F
T
d
x
= Jd
X
(2.38)
or
d
x
= JF
T
d
X
(2.39)
Analogously, a relation between the innitesimal areas in the referential and the
spatial domain is obtained:
d
x
=

J
F
T
d
(2.40)
The relation between the innitesimal areas in the material and the referential
domain is:
d
=

J
F
T
d
X
(2.41)
The material time derivative of an extensive property
An extensive property G(t) is dened as the volume integral of the mass density
x
times an intensive property g over the spatial domain:
G(t) =
_
x
gd
x
(2.42)
The material time derivative of the extensive property G(t) is:
DG(t)
Dt
=
D
Dt
_
x
gd
x
(2.43)
Recalling equation (2.28) for the change in volume between the spatial and the
material domain, the material time derivative is rewritten:
DG(t)
Dt
=
D
Dt
_
x
gJd
X
=
_
X
_
D
x
g
Dt
J +
x
g
DJ
Dt
_
d
X
(2.44)
Using the fundamental ALE equation (2.19) for the material time derivative of the
vector
x
g and equation (2.33) for the rate of change of the Jacobian J gives:
DG(t)
Dt
=
_
X
__
x
g
t
+ (c
x
)(
x
g)
_
J +
x
gJ
x
v
_
d
X
(2.45)
or
DG(t)
Dt
=
_
X
_
x
g
t
+ (c
x
)(
x
g) +
x
g
x
v
_
Jd
X
(2.46)
Equation (2.28) enables the integral to be written again as an integral over the
spatial domain:
DG(t)
Dt
=
_
x
_
x
g
t
+ (c
x
)(
x
g) +
x
g
x
v
_
d
x
(2.47)
Adding and subtracting the term
x
g
x
v to the integrand gives:
DG(t)
Dt
=
_
x
_
x
g
t
+ (c
x
)(
x
g) +
x
g
x
v
x
g
x
v +
x
g
x
v
_
d
x
(2.48)
Using equation (2.16) the second, third and fourth term are combined and by
means of equation (2.34) the last term is transformed:
DG(t)
Dt
=
_
_
_
x
g
t
+
x
(
x
g c) +
x
g
1

J
t
_
_
d
x
(2.49)
Equation (2.29) enables the integral to be written as an integral over the referential
domain:
DG(t)
Dt
=
_
_
_
J

x
g
t
+

J
x
(
x
g c) +
x
g

J
t
_
_
d
(2.50)
The rst and third term of the integrand are combined:
DG(t)
Dt
=
_
_
_
J
x
g
t
+

J
x
(
x
g c)
_
_
d
(2.51)
The referential time derivative in the rst term can be brought outside the integral:
DG(t)
Dt
=

t
_
_
J
x
gd
+
_
J
x
(
x
g c)d
(2.52)
Equation (2.29) enables the integrals to be written again as an integral over the
spatial domain:
DG(t)
Dt
=

t
__
x
gd
x
_
+
_
x
(
x
g c)d
x
(2.53)
Conservation of mass
If in equation (2.42) the intensive property g is equal to 1, the extensive property
is the mass M(t) contained in the spatial domain
x
:
M(t) =
_
x
d
x
(2.54)
Conservation of mass states that the mass contained in a material volume is
conserved or that the material time derivative of the the mass M(t) is equal to
zero:
DM(t)
Dt
=
D
Dt
_
x
d
x
= 0 (2.55)
Using equation (2.47) with g equal to 1, the expression for conservation of mass
is:
_
x
_
x
t
+c
x
x
+
x
x
v
_
d
x
= 0 (2.56)
Since the volume integral is zero for any arbitrarily chosen volume, the integrand
must be pointwise equal to zero in the spatial domain, which yields the dierential
form of the law of conservation of mass, known as the continuity equation:
x
t
+c
x
x
+
x
x
v = 0 (2.57)
The time derivative is taken holding the referential coordinate xed, while all
other derivatives are calculated in the spatial domain.
Second, equation (2.51) with g equal to 1 can be substituted in (2.55):
_
_
_

J
x
t
+

J
x
(
x
c)
_
_
d
= 0 (2.58)
As the integrand must be pointwise equal to zero in the referential domain, the
continuity equation becomes:
J
x
t
+

J
x
(
x
c) = 0 (2.59)
Third, equation (2.53) with g equal to 1 can be substituted in (2.55):
t
__
x
d
x
_
+
_
x
(
x
c)d
x
= 0 (2.60)
The advantage of this equation is that the referential time derivative of a volume
integral is taken and that both volume integrals are calculated in the spatial
domain.
From the continuity equations in the ALE description the continuity equation in
Eulerian description can be derived. In the Eulerian description the spatial and
the referential domain coincide and the convective velocity c and the material
particle velocity v are equal. From equation (2.57) the dierential form of the
law of conservation of mass in an Eulerian description, as used in uid mechanics,
emerges:
x
t
x
+v
x
x
+
x
x
v = 0 (2.61)
As to formulate conservation of mass (2.55) in the material domain, the integrand
can be rewritten using equation (2.28):
x
d
x
=
x
Jd
X
=
X
d
X
(2.62)
where the density
X
is dened as
X
= J
x
. The expression for conservation of
mass in the material domain becomes:
DM(t)
Dt
=
D
Dt
_
X
d
X
= 0 (2.63)
As the material time derivative is taken holding the material coordinate X xed,
the dierential form of the law of conservation of mass in Lagrangian description
in the material domain, as used in structural mechanics, is easily obtained:
X
t
X
= 0 (2.64)
This shows that the density
X
is constant in time in the material domain d
X
.
Conservation of momentum
If in equation (2.42) the intensive property g is equal to v, the extensive property
is the momentum M(t) contained in the spatial domain
x
:
M(t) =
_
x
vd
x
(2.65)
Conservation of momentum states that the change in momentum of a material
volume M(t) with respect to time is equal to the sum of all external forces on this
volume:
DM(t)
Dt
=
D
Dt
_
x
vd
x
=
_
x
td
x
+
_
x
bd
x
(2.66)
where t is the traction that act on the boundary of the volume and
x
b is the body
force. The traction vector t on a surface with unit normal vector n
x
is obtained
as:
t = n
x
(2.67)
where the second order tensor is called the Cauchy stress tensor. Using this
equation, Gauss theorem enables to write the integral over the boundary in the
rst term on the right hand side of equation (2.66) as a volume integral:
D
Dt
_
x
vd
x
=
_
x
d
x
+
_
x
bd
x
(2.68)
Using equation (2.44) with g equal to v, the expression for conservation of
momentum is:
_
X
_
D
x
v
Dt
J +
x
v
DJ
Dt
_
d
X
=
_
x
d
x
+
_
x
bd
x
(2.69)
The left hand side is elaborated:
_
X
_
x
Dv
Dt
J +v
D
x
Dt
J +
x
v
DJ
Dt
_
d
X
(2.70)
=
_
X
_
x
Dv
Dt
J +v
_
D
x
Dt
J +
x
DJ
Dt
__
d
X
(2.71)
The term in square brackets is equal to zero as this follows from the conservation
of mass (equation (2.44) with g equal to 1). Using equation (2.19) for the material
time derivative of the vector function v in the ALE description and equation (2.28)
for the volume change, gives:
DM(t)
Dt
=
_
x
_
v
t
+ (c
x
)v
_
d
x
(2.72)
From the combination with equation (2.69), emerges the expression for
conservation of momentum:
_
x
_
v
t
+ (c
x
)v
_
d
x
=
_
x
d
x
+
_
x
bd
x
(2.73)
Since the volume integrals are equal for any arbitrarily chosen volume, the
integrand must be pointwise equal in the spatial domain, which yields the following
the momentum equation:
x
v
t
+
x
(c
x
)v =
x
+
x
b (2.74)
Second, equation (2.51) with g equal to v can be substituted in (2.68) and the
volume integrals are transformed to the referential domain (2.29):
_
_
_

J
x
v
t
+

J
x
(
x
v c)
_
_
d
=
_
J
x
d
+
_
J
x
bd
(2.75)
As the integrand must be pointwise equal to zero in the referential domain, the
momentum equation becomes:
J
x
v
t
+

J
x
(
x
v c) =

J
x
+

J
x
b (2.76)
Third, equation (2.53) with g equal to v can be substituted in (2.68):
t
__
x
vd
x
_
+
_
x
(
x
vc)d
x
=
_
x
d
x
+
_
x
bd
x
(2.77)
The advantage of this equation is that the referential time derivative of a volume
integral is taken and that both volume integrals are calculated in the spatial
domain.
From the momentum equations (2.74) in the ALE description the momentum
equations in Eulerian and Lagrangian description can be derived. In the Eulerian
description the spatial and the referential domain coincide and the convective
velocity c and the material particle velocity v are equal. From equation (2.74) the
dierential form of the law of conservation of momentum in Eulerian description
in the spatial domain, as used in uid mechanics, emerges:
x
v
t
x
+
x
(v
x
)v =
x
+
x
b (2.78)
In the Lagrangian description the material and the referential domain coincide
and the convective velocity c and the particle velocity v in the referential domain
are zero. From equation (2.74) the dierential form of the law of conservation of
momentum in Lagrangian description in the material domain, as used in structural
mechanics, is easily obtained:
x
v
t
X
=
x
+
x
b (2.79)
As to formulate conservation of momentum (2.66) in the material domain, the
boundary integral in the rst term on the right hand side should be transformed
to the material domain. The integrand of this term is transformed using equation
(2.67):
td
x
= n
x
d
x
= d
x
(2.80)
where d
x
is the area vector. Recalling equation (2.39) for the relation between
an area vector in the spatial and the material domain, yields:
td
x
= JF
T
d
X
= JF
T
n
X
d
X
(2.81)
where n
X
denotes the unit normal vector to the same area in the material domain.
In the material domain the traction t
X
= JF
T
n
X
is dened such that td
x
=
t
X
d
X
. The rst Piola-Kirchho stress tensor P
X
= JF
T
is dened analogous
to the Cauchy stress tensor (2.67) in the spatial domain:
t
X
= P
X
n
X
(2.82)
The rst Piola-Kirchho stress tensor P
X
in the material domain relates the area
vector d
X
in the material domain to the corresponding force vector td
x
in the
spatial domain.
The volume integrals on the left hand side and in the second term on the right
hand side of equations (2.66) can be rewritten using equation (2.62). In the rst
term on the right hand side equations (2.81) and (2.82) are substituted.
D
Dt
_
X
vd
X
=
_
X
P
X
d
X
+
_
X
bd
X
(2.83)
This is the expression for conservation of momentum in the material domain.
The material time derivative is taken holding the material coordinate X xed.
Taking into account the conservation of mass (2.64), the momentum equation in
Lagrangian description in the material domain, which is useful for structures with
non-linear behaviour, is obtained:
X
v
t
X
=
X
P
X
+
X
b (2.84)
Conservation of moment of momentum
The conservation of moment of momentum is not elaborated here but implies the
symmetry of the Cauchy stress tensor:
=
T
(2.85)
2.2.2 Kinematics
The velocity increment between two neighbouring particles can be expressed as:
dv =
x
vdx (2.86)
The velocity gradient is decomposed into its symmetric and skew-symmetric parts:
x
v =
1
2
(
x
v + (
x
v)
T
) +
1
2
(
x
v (
x
v)
T
) = + (2.87)
The symmetric tensor is called the strain rate tensor and describes the
deformations of the uid volume, while the skew-symmetric tensor is called
the vorticity tensor and describes rigid body rotations. The strain rate can be
split into its isotropic and deviatoric part:
x
v =
1
3
(tr)I +
_

1
3
(tr)I
_
+ (2.88)
2.2.3 Constitutive equations
The stress is split into its isotropic and its deviatoric part :
=
1
3
(tr)I + (2.89)
The isotropic part is the sum of the negative pressure p and the product of the
bulk viscosity K with the trace of the strain rate tensor:
1
3
(tr)I = pI + K(tr)I (2.90)
In a uid at rest, the shear stresses are zero. For a Newtonian uid, a linear
relation between the deviatoric stress and the deviatoric strain rate is assumed:
= T
_

1
3
(tr)I
_
(2.91)
If the uid is isotropic, the relation becomes:
= 2
_

1
3
(tr)I
_
(2.92)
where denotes the dynamic viscosity. Combining equations (2.89), (2.90) and
(2.92) yields the expression for the Cauchy stress tensor as a function of the strain
rate:
= pI + K(tr)I + 2
_

1
3
(tr)I
_
(2.93)
Reordering the terms, yields:
= pI + (tr)I + 2 (2.94)
where = K 2/3 is the second viscosity constant.
2.2.4 Governing equations
The continuity equation (2.57) and the momentum equation (2.74) in the spatial
domain using an ALE description, together with the constitutive equation (2.94)
give:
x
t
+c
x
x
+
x
x
v = 0 (2.95)
x
v
t
+
x
(c
x
)v +
x
p = 2
x
+
x
(tr) +
x
b (2.96)
For incompressible ows (Mach number Ma < 0.3),
x
may be assumed to be
constant. An equation for conservation of energy is not needed. The continuity
equation (2.95) reduces to:
x
v = 0 (2.97)
From the continuity equation follows that the trace of the strain rate tr is equal
to zero and the momentum equation (2.96) becomes:
x
v
t
+
x
(c
x
)v +
x
p = 2
x
+
x
b (2.98)
Dividing the momentum equations by
x
and dening the kinematic viscosity
= /
x
and the kinematic pressure p
= p/
x
, yields:
v
t
+ (c
x
)v +
x
p
= 2
x
+b (2.99)
Together with the boundary and initial conditions, equations (2.97) and (2.99)
form the Navier-Stokes equations for an incompressible ow on a deforming
domain. This is a system of four non-linear second order dierential equations
in four independent variables and is formulated here as a function of the primitive
variables v and p
. In the continuum equation (2.97), no time derivative occurs.

This equation formulates a kinematical constraint on the velocity eld. In the
momentum equation (2.99), the rst term on the left hand side is the time
derivative of the velocity. The second term on the left hand side is a non-linear
convective (advective) term. Neglecting this term yields the Stokes equations for
highly viscous ow. The third term is the pressure gradient. Since only the
gradient of the pressure appears in the equations, the pressure is determined only
up to an arbitrary constant, which should be xed by a boundary condition. The
rst term on the right hand side is a linear viscous (diusive) term. Dropping this
term, the Euler equations for inviscid ow are obtained. Recalling the continuity
equation and the denition of the strain rate, this term can be formulated
alternatively as
2
x
v.
Dirichlet (or essential, kinematic) boundary conditions impose the velocity v on a
part
D
of the boundary:
v = v
D
(2.100)
Neumann (or natural, mechanical) boundary conditions prescribe the traction t
on a part
N
of the boundary:
t = n
x
= pn
x
+ 2n
x
= t
N
(2.101)
After division by
x
, an alternative expression for the Neumann boundary
conditions emerges:
t
n
x
= p
n
x
+ 2n
x
= t
N
(2.102)
Since no time derivative of the pressure appears in the governing equations, initial
conditions are only needed for the velocity eld at t = t
0
:
v = v
0
(2.103)
and should fulll the continuity equation:
x
v
0
= 0 (2.104)
For the computation of the convective velocity c (equation (2.16)) the velocities v
of the nodes of the mesh are still needed. The dierent possibilities to obtain the
velocities of the nodes will be treated in section 2.5.
2.3 Spatial discretization
2.3.1 Finite Element Method
The standard nite element method (Hughes, 1987; Bathe, 1996; Zienkiewicz et al.,
2005b; Zienkiewicz and Taylor, 2005; Zienkiewicz et al., 2005a) was originally
developed in structural mechanics. The typical applications (e.g. linear elasticity,
heat conduction) were governed by diusion-type dierential equations. The
solutions of these equations correspond to the minimum of an energy norm (e.g.
total potential/thermal energy). Application of the standard nite element method
to convection-dominated problems in uid mechanics caused several diculties
and motivated the development of stabilization techniques. A detailed review of
these methods is given by Donea and Huerta (2003), Gresho and Sani (2000a,b),
Zienkiewicz et al. (2005a) and L ohner (2001). The CFD program Flotran (Ansys,
2005b) is based on the nite element method. For ease of notation the subscripts
x referring to the spatial domain will be omitted in the following sections. First,
the standard nite element method is applied to the incompressible Navier-Stokes
equations (2.97) and (2.99):
v = 0 (2.105)
v
t
+ (c )v
1
b = 0 (2.106)
This is the strong form of the system of equations. If trial solutions for the velocity
v
h
and for the pressure p
h
which satisfy the Dirichlet boundary conditions (2.100)
are introduced, the residuals are given by:
1
c
= v
h
(2.107)
R
m
=
v
h
t
+ (c
h
)v
h

h
b
h
(2.108)
A weighted residual formulation of the governing equations uses the weighting
functions q and w for the continuity and momentum equations:
_
q v
h
d +
_
w
v
h
t
d +
_
w (c
h
)v
h
d
w
_
1

h
_
d
_
w b
h
d = 0 (2.109)
SPATIAL DISCRETIZATION 37
The integral of the divergence of the stress tensor over the domain is integrated
by parts:
_
q v
h
d +
_
w
v
h
t
d +
_
w (c
h
)v
h
d
(w
h
)d +
_
h
: wd
_
w b
h
d = 0 (2.110)
Gauss theorem enables to write the volume integral in the fourth term on the left
hand side as a boundary integral:
_
q v
h
d +
_
w
v
h
t
d +
_
w (c
h
)v
h
d
w
h
nd +
_
h
: wd
_
w b
h
d = 0 (2.111)
Introducing the Neumann boundary conditions (2.101) and recalling that the
weighting functions w vanish on the part
D
of the boundary, the weak form
of the problem is obtained:
_
q v
h
d +
_
w
v
h
t
d +
_
w (c
h
)v
h
d
N
w
1
t
N
d
N
+
_
h
: wd
_
w b
h
d = 0 (2.112)
Recalling the constitutive equation (2.94) yields, after reorganizing of the terms:
_
w
v
h
t
d +
_
w (c
h
)v
h
d +
_
2
h
: wd
p
h
wd +
_
q v
h
d =
_
w b
h
d +
_
N
w t
N
d
N
(2.113)
In the weak form, the continuity requirements for the trial solutions and weighting
functions are modied. Whereas in the strong form the second order derivative
of the velocity trial solutions v
h
and the rst order derivative of the pressure
trial solutions p
h
appeared, now only the rst order derivative of the velocity
trial solutions and the pressure trial solution must be square integrable. Whereas
no requirement existed for the weighting functions w in the strong form, now
their rst order derivative must also be square integrable. The trial solutions are
discretized using nite element approximations. The solutions for the velocity
v
h
(x) are approximated as:
v
h
(x) =
n
v
j
N
v
j
(x)v
j
(2.114)
where N
v
j
is the shape function (basis function) associated with node j and v
j
are the nodal unknowns, except on the Dirichlet boundaries where v
j
is equal to
the imposed value. n
v
nodes are used for the approximation of the velocity vector.
The shape function associated with node i is a piecewise polynomial function which
only diers from 0 in all connecting elements. Its value is equal to 1 for this node
and equal to 0 for all other nodes:
N
v
i
(x
j
) =
ij
(2.115)
The trial functions for the pressure p
h
(x) are approximated as:
p
h
(x) =
n
p
j
N
p
j
(x)p
j
(2.116)
The number n
p
of nodes and the shape functions used for the approximation of
the pressure may dier from those used for the approximation of the velocity.
In the Galerkin approach, every shape function is used once as a weighting
function:
w(x) = N
v
i
(x) (2.117)
q(x) = N
p
i
(x) (2.118)
Substituting the discretizations of the velocity (2.114) and the pressure (2.116)
and the weighting functions (2.118) in the weak form (2.113) yields:
n
v
j
__
N
v
i
N
v
j
d
_
v
j
t
+
n
v
j
__
N
v
i
(c
h
)N
v
j
d
_
v
j
+
n
v
j
__
(N
v
j
+ (N
v
j
)
T
) : N
v
i
d
_
v
j
n
p
j
__
N
p
j
N
v
i
d
_
p
j
+
n
v
j
__
N
p
i
N
v
j
d
_
v
j
=
_
N
v
i
b
h
d +
_
N
N
v
i
t
N
d
N
(2.119)
This can be rewritten as a matrix system:
_
M 0
0 0
_ _
v
p
_
+
_
K(v) G
G
T
0
_ _
v
p
_
=
_
f
0
_
(2.120)
where
M
ij
=
_
N
v
i
N
v
j
d (2.121)
K
ij
=
_
N
v
i
(c
h
)N
v
j
d +
_
(N
v
j
+ (N
v
j
)
T
) : N
v
i
d (2.122)
G
ij
=
_
N
p
j
N
v
i
d (2.123)
f
i
=
_
N
v
i
b
h
d +
_
N
N
v
i
t
N
d
N
(2.124)
For convection-dominated problems, the stiness matrix K is non-symmetric.
The advection-diusion stability: Streamline-Upwind Petrov-Galerkin method
(SUPG)
If the Galerkin approach is applied, the solutions for the velocities are sometimes
corrupted by spurious node-to-node oscillations (also called wiggles) which pollute
the whole computational domain. In order to illustrate the deciencies of
the Galerkin approach for convection-dominated problems, equation (2.99) is
simplied to a steady 1D scalar convection-diusion equation with a constant
convection c and a constant source term b:
c
v
x

2
v
x
2
= b (2.125)
The combination of linear shape functions with the Galerkin approach yields on
a uniform mesh with element length h a discretized equation for node i that is
identical to that obtained with second order central dierences:
c
v
i+1
v
i1
2h

v
i+1
2v
i
+ v
i1
h
2
= b (2.126)
The ratio of the convective to the diusive transport is expressed by the mesh
Peclet number Pe that can be interpreted as a local Reynolds number:
Pe =
ch
2
(2.127)
When the Peclet number exceeds unity, the solution of equation (2.126) using
the Galerkin approach is corrupted by spurious node-to-node oscillations (also
called wiggles) which pollute the whole computational domain. Figure 2.4a shows
the solution of a steady 1D scalar convection-diusion equation with a constant
convection c = 1 and a constant source term b = 1 using an element size h = 0.1
for a Peclet number smaller than one and gure 2.4b for a Peclet number larger
than one. The oscillations can be avoided by rening the mesh to obtain Peclet
numbers lower than one. The discretized equation which produces the exact nodal
solution with linear shape functions on a uniform mesh is (gure 2.4):
c
v
i+1
v
i1
2h
( +
ch
2
)
v
i+1
2v
i
+ v
i1
h
2
= b (2.128)
where is dened as:
= coth(Pe)
1
Pe
(2.129)
0 0.2 0.4 0.6 0.8 1
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
Distance [m]
V
e
l
o
c
i
t
y

[
m
/
s
]
(a)
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
Distance [m]
V
e
l
o
c
i
t
y

[
m
/
s
]
(b)
Figure 2.4: Analytical solution for the velocity (black line), exact nodal solution
(black dashed line) (2.128), Galerkin solution (dark grey line) (2.126)
and full upwind solution (light grey line) (2.130) of a steady 1D scalar
convection-diusion equation with a constant convection c = 1 and a
constant source term b = 1 using an element size h = 0.1 for a Peclet
number of (a) 0.25 and (b) 5.
To remove these oscillations without any requirements on the mesh size, two
possibilities exist. First, comparison of equations (2.126) and (2.128) shows that
the discretization error of the Galerkin nite element and the central dierence
discretization methods is equal to a negative diusion term with numerical
diusion =
ch
2
. This numerical diusion can be added to the Galerkin
discretization of equation (2.126) to balance the negative diusion inherent to
the Galerkin approach. The resulting discretization is identical to equation
(2.128) where may now dier from equation (2.129) and controls the amount of
numerical diusion.
Second, equation (2.128) can be rewritten as:
1
2
c
v
i+1
v
i
h
+
1 +
2
c
v
i
v
i1
h

v
i+1
2v
i
+ v
i1
h
2
= 1 (2.130)
The convective ux is discretized by a weighted average of the uxes from the left
and the right side, while the diusive ux is still discretized with a second order
central dierence scheme. may dier from equation (2.129) and controls the
weight given to the upstream and downstream element. The elements upstream
of a node can be more heavily weighted than those downstream for the convective
term. For a positive convective velocity c, a value of equal to one corresponds
to a full upwind discretization (gure 2.4), which is always stable but often too
diusive (Donea and Huerta, 2003). For equal to zero the equation reduces to
the standard Galerkin discretization. Using values between one and zero is called
hybrid dierencing. The modied weighting functions w
m
dier from the Galerkin
weighting functions w which are equal to the shape functions. A possible choice
is:
w
m
= w +
h
2
w
x
(2.131)
In both cases, the absolute value of should be greater than a critical value in
order to avoid oscillations:
[[ 1
1
[Pe[
(2.132)
The extension of the convection-diusion equation (2.125) to two dimensions gives:
c v
2
v = b (2.133)
c
1
v
x
1
+ c
2
v
x
2
2
v
x
2
1
+

2
v
x
2
2
_
= b (2.134)
Hybrid dierencing can be applied separately to the terms c
1
v/x
1
and c
2
v/x
2
,
which corresponds to adding the following numerical dissipation:
11
2
v
x
2
1
+
22
2
v
x
2
2
(2.135)
This technique introduces numerical diusion perpendicular to the convection
direction (crosswind diusion) which becomes large if the convective velocity is
not aligned with the mesh (Donea and Huerta, 2003). In order to add numerical
diusion only in the convection direction and not in the transversal direction, a
diusivity tensor is constructed:
=

|c|
2
c c (2.136)
This approach does not introduce any crosswind diusion. The magnitude of the
convective velocity is used to dene the numerical viscosity:
=
|c|h
2
(2.137)
Peclet numbers are computed using the magnitude of the convective velocity:
Pe =
|c|h
2
(2.138)
Adding the numerical diusivity tensor corresponds to the use of a modied
weighting function w
m
for the convective term:
w
m
= w+
h
2|c|
c w = w+ c w (2.139)
where is the stabilization parameter which is also called the intrinsic time.
Several options exist for computing the element length h in equation (2.139). The
choice largely inuences the amount of numerical diusion introduced. In the nite
element code Flotran (Ansys, 2005b), the length of the projection of the element
on the direction of the convection velocity c is used. To reduce the computational
costs the optimal blending function (2.129) is replaced by its doubly asymptotic
approximation:
=
_
Pe/3 if 0 Pe 3
1 if Pe > 3
(2.140)
Other approximations exist, but the inuence of the blending function is much
smaller than the inuence of the choice for the computation of the element length.
The approach where the modied weighting function of equation (2.139) is applied
to the convective term only, is called streamline upwind (SU). The exact solution of
the dierential equation is no longer a solution of the weak form. Generally, if the
same weighting function, dierent from the shape functions, is applied to all terms
in the weak form of equation (2.113), a consistent formulation is obtained, which is
called the Petrov-Galerkin approach. If the weighting function of equation (2.139)
is used in all terms, the Streamline-Upwind Petrov-Galerkin (SUPG) method
(Brooks and Hughes, 1982) is obtained, which can easily be implemented by adding
a stabilization term, emanating from the second term on the right hand side of
equation (2.139), to the Galerkin weak form of equation (2.113):
ST =
n
e
i
_
supg
(c
h
w) R
m
(v
h
, p
h
))d
i
(2.141)
This term should not be integrated by parts because then for linear elements the
stabilization eects disappear. For linear elements the stabilization contribution
in the diusion term vanishes. As a result the spurious oscillations are localized
and do not pollute the whole domain.
The pressure stability: Pressure-Stabilizing/Petrov-Galerkin method (PSPG)
If the Galerkin approach is applied, the solution for the pressure sometimes exhibits
spurious pressure modes (e.g. checkerboard modes). This pressure instability can
be illustrated by means of the steady 2D incompressible Stokes equations for highly
viscous ow, which are obtained by neglecting the transient and the convective
term in the equations (2.97) and (2.99):
v = 0 (2.142)
p
2
v = 0 (2.143)
By taking the divergence of the momentum equation (2.143) and inserting the
continuity equation into the second term on the left hand side, a Laplace equation
for the pressure is obtained:
(p
2
v) =
2
p
2
( v) =
2
p = 0 (2.144)
If the velocity and the pressure are discretized in 2D using second order central
dierences, the discrete Laplace equation becomes:
p
i+2,j
+ p
i2,j
+ p
i,j+2
+ p
i,j2
4p
i,j
4h
2
= 0 (2.145)
This equation is independent of the pressures at points i + 1, i 1, j + 1 and
j 1. This is called the odd-even decoupling as the solution for odd points
is independent of the solution for even points. This decoupling occurs if equal
order interpolations are used for the velocity and the pressure and results in
spurious pressure oscillations. The Ladyzhenskaya-Babuska-Brezzi (LBB) or
inf-sup condition (Babuska, 1971) species a relation between the velocity and
pressure interpolation to guarantee the stability of a mixed method. Combinations
of trial solutions for velocity and pressure that satisfy the LBB condition are
for instance obtained using quadratic shape functions for the velocity and linear
shape functions for the pressure. Stabilization methods enable to circumvent the
LBB condition (Hughes et al., 1986) and to use combinations of trial solutions for
velocity and pressure that are not stable in the Galerkin formulation e.g. equal
order interpolations for velocity and pressure. A stabilization term consisting
of a stability parameter
pspg
and the Laplacian of the pressure is added to the
continuity equation (2.97):
v
pspg
2
p = 0 (2.146)
If this stabilization term is integrated by parts in the weighted residual formulation,
it becomes:
_
pspg
q pd (2.147)
This can be viewed as an additional term resulting from a modied weighting of
the pressure gradient term in the momentum equation with the following weighting
function:
w
m
= w+
pspg
q (2.148)
A consistent approach, which is called the Pressure-Stabilizing/Petrov-Galerkin
(PSPG) method (Hughes et al., 1986), is obtained if this weighting function
is applied to all terms of the momentum equation. This yields the following
stabilization term:
ST =
n
e
i
_
pspg
q R
m
(v
h
, p
h
)d
i
(2.149)
The stability parameter
pspg
is dened empirically as:
pspg
=
h
2
2
(2.150)
where = 1/3 appears to be optimal for linear elements.
SUPG/PSPG method
The combination of the SUPG (2.139) and the PSPG method (2.148) leads to the
SUPG/PSPG method which employs the following weighting function:
w
m
= w+
supg
c
h
w+
pspg
q (2.151)
This results in the corresponding stabilization term:
ST =
n
e
i
_
i
_
supg
c
h
w+
pspg
q
_
R
m
(v
h
, p
h
)d
i
(2.152)
Galerkin/Least Squares method
The Galerkin/Least Squares (GLS) method (Hughes et al., 1989) is a generalization
of the SUPG and PSPG methods. It is a linear combination of the Galerkin and
the least squares method. The least-squares method minimizes the integral of the
squared residuals over the domain with respect to the nodal velocities v
i
:
v
i
_
R
2
m
(v
h
, p
h
)d = 2
_
R
m
(v
h
, p
h
)
R
m
(v
h
, p
h
)
v
i
d (2.153)
= 2
_
R
m
(v
h
, p
h
) (c
h
w2
h
(w))d
(2.154)
The integral of the squared residuals over the domain is minimized as well with
respect to the nodal pressures p
i
:
p
i
_
R
2
m
(v
h
, p
h
)d = 2
_
R
m
(v
h
, p
h
)
R
m
(v
h
, p
h
)
p
i
d (2.155)
= 2
_
R
m
(v
h
, p
h
) (q)d (2.156)
After multiplication with a stability parameter
gls
, the sum of equations (2.154)
and (2.156) yields the following stabilization term:
ST =
n
e
i
_
i
_
gls
(c
h
w2
h
(w) +q)
_
R
m
(v
h
, p
h
)d
i
(2.157)
The modied weighting function w
m
for the Galerkin/Least Squares (GLS) method
is:
w
m
= w+
gls
(c
h
w2
h
(w) +q) (2.158)
GLS and SUPG are identical for P1 elements, but dier for higher order elements.
2.3.2 Finite Volume Method
The nite volume method (Hirsch, 1995a,b; Anderson, 1995; Versteeg and
Malalasekera, 1995; Ferziger and Peric, 2002) was originally developed in uid
mechanics. The CFD program CFX (Ansys, 2005a) is based on the nite volume
method. In the nite volume method the conservation of mass (2.60) and the
conservation of momentum (2.77) are expressed for a control volume
i
:
t
__
i
d
i
_
+
_
i
(c)d
i
= 0 (2.159)
t
__
i
vd
i
_
+
_
i
(v c)d
i
=
_
i
d
i
+
_
i
bd
i
(2.160)
By splitting the stress tensor in its hydrostatic and deviatoric part (2.89), the
terms in equation (2.160) are reordered:
t
__
i
vd
i
_
+
_
i
(v c + pI)d
i
=
_
i
d
i
+
_
i
bd
i
(2.161)
Using Gauss theorem, the integrals of the divergence over the control volume in
equations (2.159) and (2.161) can be rewritten as integrals on the boundary
i
of
this control volume:
t
__
i
d
i
_
+
_
i
(c) nd
i
= 0 (2.162)
t
__
i
vd
i
_
+
_
i
(v c + pI)nd
i
=
_
i
nd
i
+
_
i
bd
i
(2.163)
The vectors V and Q and the tensor F
d
respectively represent the conservative
variables, the source terms and the diusive (viscous) uxes:
V =
_

v
_
Q=
_
0
b
_
F
d
=
_
0
_
(2.164)
The tensor F
c
contains the convective (inviscid) uxes. Using the denition of the
convective velocity (2.16), the convective uxes are split in a part F
e
corresponding
to an Eulerian description and a contribution from the ALE description:
F
c
=
_
c
v c + pI
_
=
_
v
v v + pI
_
_
v
v v
_
(2.165)
= F
e
V v = F
e
F
ale
(2.166)
Using these denitions the equations are rewritten:
t
__
i
Vd
i
_
+
_
i
F
c
nd
i
=
_
i
F
d
nd
i
+
_
i
Qd
i
(2.167)
In order to discretize these equations the computational domain is meshed. In the
cell-centered approach the mesh cells are used as control volumes. The unknowns
are located at the centroid of the mesh. In the vertex-centered approach the
unknowns are located at the vertices of the mesh. The control volumes are dened
by a dual mesh which connects the centroid of the mesh cells (centroid dual) or the
centroid of the mesh cells and the centroid of the faces of the mesh cells (median
dual). This latter approach is used in the nite volume code CFX (Ansys, 2005a).
In the cell-vertex approach the unknowns are located at the vertices of the mesh
and the mesh cells are used as control volumes. In the sequel of this section it
is supposed that the vertex-centered median dual approach is used, however the
discretization principles are similar for all approaches. One of the main dierences
is the treatment of the boundary conditions.
As the control volume
i
is closed by N
i
faces, the boundary integrals are
transformed into a sum over these faces:
t
__
i
Vd
i
_
+
N
i
j
_
ij
F
c
nd
ij
=
N
i
j
_
ij
F
d
nd
ij
+
_
i
Qd
i
(2.168)
The volume integrals are replaced by the product of the volume and the average
values:
i
V
i
t
+
N
i
j
_
ij
F
c
nd
ij
=
N
i
j
_
ij
F
d
nd
ij
+
i
Q
i
(2.169)
V
i
contains the average values of the conservative variables over the control volume
and Q
i
the average values of the source terms.
The integrals of the uxes F
c
and F
d
over the faces are numerically approximated
in two levels. First, the integration is performed numerically using a quadrature
rule. The simplest approximation is obtained using the midpoint rule: the integral
is approximated as the product of the integrand at the midpoint of the face and the
face area. This approach is second order accurate. Higher-order approximations
need the value of the integrand at the corners and centers of edges. Second, in
order to compute the uxes at the midpoint (and eventually the corners and the
centers of the edges), the uxes at these locations are interpolated from the uxes
at the vertices (ux averaging) or are computed from the conservative variables
at these locations, which are interpolated from the conservative variables at the
vertices (variable averaging). In CFX, nite element shape functions are used
to interpolate the values of the conservative variables at these locations. As the
integral conservation applies to each control volume, the surface integrals should be
computed in a way that they cancel out for surfaces with opposite normals in order
to guaranty global conservation over the domain. The numerical approximation
of the uxes nally yields:
i
V
i
t
+F
c
i
(V, x, v) = F
d
i
(V, x) +
i
Q
i
(2.170)
where F
c
i
and F
d
i
are equal to the sum of all the numerical uxes of the dierent
faces of the control volume. Collecting all equations into a single system of
equations gives:
V
t
+F
c
(V, x, v) = F
d
(V, x) +Q (2.171)
where is a diagonal matrix consisting of the volumes and V and Q are the
vectors formed by the collection of respectively V
i
and Q
i
.
2.4 Time integration
The system of equations which has already been discretized in space is now
integrated in time. First, the dierent time integration methods for an Eulerian
description are reviewed. Next, the implication of the ALE description on the
accuracy and the stability of the time integration is discussed and related to the
geometric conservation law. The methods are described for the nite volume
method, but analogous methods exist for the nite element method.
2.4.1 The Eulerian description
After spatial discretization a system of coupled rst order ordinary dierential
equations (2.171) is obtained. In the Eulerian description the control volumes are
xed in space and their volume is constant. The system of equations becomes:
V
t
+F
c
(V) = F
d
(V) (2.172)
This is simplied to:
V
t
= F(V) (2.173)
Direct time integration methods are used to integrate this initial value problem. A
group of commonly used methods for integrating rst order dierential equations
TIME INTEGRATION 49
are the one-step methods. One-step methods only refer to the values at the
previous time t and the current time t + t to determine the solution at time
t + t. In the one-step methods a weighted average of the right hand side of
equation (2.173) at time t and t + t is used:
V
n+1
V
n
t
= F(V
n+1
) + (1 )F(V
n
) (2.174)
For = 0 the forward Euler method is obtained, which is an explicit method as
V
n+1
can be directly computed. All other values of yield implicit methods as a
system of equations for V
n+1
has to be solved.
The forward Euler method is conditionally stable. For convection dominated
problems the stability is related to the Courant number C =
vt
h
where h is
the element size. As an example, for a one-dimensional convection equation
the combination of the forward Euler method for the time integration with the
rst order upwind dierence scheme for the spatial discretization is stable if the
Courant number C is smaller than one. This is known as the Courant-Friedrichs-
Lewy (CFL) condition and limits the time step so that the distance travelled by
a material particle during a the time step is smaller than element size. Methods
with 1/2 are unconditionally stable or A-stable. For = 1, the backward
Euler method is obtained and for = 1/2, the Crank-Nicolson method, which is
also called the trapezoidal method.
The forward and backward Euler are rst order accurate in time while the Crank-
Nicolson method is the only method which is second order accurate. A drawback of
the Crank-Nicolson method is that it has no numerical dissipation and oscillatory
solutions might arise.
The one-step methods belong to a larger group of linear multistep (LMS)
methods:
k
i=0
a
i
V
n+1i
= t
k
i=0
b
i
F(V
n+1i
) (2.175)
These multistep methods refer to several previous time steps. A linear combination
of the values at the previous time steps is used. A-stable linear multistep methods
are at most second order accurate. Among these methods, the Crank-Nicolson
method has the smallest truncation error. If b
0
= 0, the method is explicit, while
otherwise it is implicit. For a
0
= 3/2, a
1
= 2, a
2
= 1/2 and b
0
= 1 a second order
accurate, implicit three-point backward dierence method (BDF2) is obtained:
3
2
V
n+1
2V
n
+
1
2
V
n1
= tF(V
n+1
) (2.176)
This method is not self-starting as it requires the solutions at two previous
time steps. In Flotran and CFX this implicit backward dierence method is
implemented.
2.4.2 The ALE description
In the ALE description the system of coupled rst order ordinary dierential
equations (2.171) obtained after spatial discretization has some important
dierences in comparison with equation (2.173) for the Eulerian description:
V
t
= F
c
(V, x, v) +F
d
(V, x) (2.177)
The diagonal matrix with the volumes of the control volumes changes with
time. The convective uxes F
c
are integrated over a changing boundary and
depend on the mesh velocity. The diusive uxes F
d
are as well integrated over a
changing boundary. The possibilities to construct a time integration method for
computations using the ALE description are studied is this section.
The geometrical conservation law
A rst approach to construct a time integration procedure in an ALE description
is to require that a uniform ow can be computed exactly, independently of the
mesh deformations (Lesoinne and Farhat, 1996).
The time integration between t
n
and t
n+1
of the nite volume equations (2.167)
for a control volume
i
yields:
__
i
Vd
i
_
n+1
__
i
Vd
i
_
n
+
_
t
n+1
t
n
_
i
F
ale
nd
i
dt
=
_
t
n+1
t
n
_
i
F
d
nd
i
dt +
_
t
n+1
t
n
_
i
Qd
i
dt
(2.178)
For a uniform ow the source termQis equal to zero and the conservative variables
V are constant with time and denoted by V
:
__
i
V
d
i
_
n+1
__
i
V
d
i
_
n
+
_
t
n+1
t
n
_
i
F
ale
nd
i
dt
=
_
t
n+1
t
n
_
i
F
d
nd
i
dt (2.179)
TIME INTEGRATION 51
The convective uxes F
c
are split in two parts as in equation (2.166):
V
(
n+1
i

n
i
) +
_
t
n+1
t
n
_
i
(F
e
V v) nd
i
dt =
_
t
n+1
t
n
_
i
F
d
nd
i
dt
(2.180)
The integrals on a closed boundary of the ux of a constant function, like F
e
and
F
d
, are equal to zero:
V
(
n+1
i

n
i
)
_
t
n+1
t
n
_
i
(V
v)nd
i
dt = 0 (2.181)
As V
is constant, the equation is rewritten as:

(
n+1
i

n
i
) =
_
t
n+1
t
n
_
i
v nd
i
dt (2.182)
This is the geometrical conservation law (GCL) or space conservation law
(Demirdzic and Peric, 1988) for nite volume methods which is universal for all
time integration schemes: the change in volume of a control volume during a time
step should be equal to the volume swept by its boundaries during the same time
step. As the change in volume is computed exactly, any time integration scheme
should also compute the right hand of equation (2.182) side exactly. The GCL only
includes geometric quantities as the node positions x and the mesh velocities v.
It only provides information for the time integration of the ALE convective uxes
as the viscous uxes F
d
automatically disappear from the equations. For a nite
element discretization, an analogous GCL can be derived (Lesoinne and Farhat,
1996; Formaggia and Nobile, 2004) if equation (2.76) is used as a starting point.
The time derivative in this equation is applied to the product of the Jacobian
determinant and the variables (density and velocity). If, however, equation (2.74),
which only includes the time derivative of the variables, is chosen as a starting point
for the nite element discretization, the geometric conservation law is satised
independent of the time integration scheme (Formaggia and Nobile, 2004; Forster
et al., 2006). Space-time discretizations always satisfy the GCL.
The implications of the GCL will now be illustrated for two specic implicit time
integration schemes: the backward Euler method and the three-point backward
dierence scheme. At rst reading, the reader may wish to skip these two
paragraphs and continue with the paragraph on the importance of the GCL on
page 57.
If the semi-discretized equation (2.170) for one control volume is integrated in time
between t
n
and t
n+1
, the following discrete equation is obtained:
(
i
V
i
)
n+1
(
i
V
i
)
n
+
_
t
n+1
t
n
F
c
i
(V, x, v)dt =
_
t
n+1
t
n
F
d
i
(V, x)dt (2.183)
The backward Euler method The backward Euler method (equation (2.174)
for = 1) approximates these integrals as the uxes F
i
corresponding to V
n+1
multiplied with t. The question arises on which mesh conguration x these
convective uxes and diusive uxes should be evaluated: at x
n+1
, at x
n
or in
between those two congurations. A similar question is which mesh velocities v
should be used and how they have to be computed from the node positions. If a
uniform ow is assumed, V
i
= V
i
and the previous equation becomes:
n+1
i
V
i

n
i
V
i
+
_
t
n+1
t
n
F
c
i
(V
, x, v)dt =
_
t
n+1
t
n
F
d
i
(V
, x)dt (2.184)
For a uniform ow V
i
the diusive ux F
d
i
(V
, x) is equal to zero. The

convective ux is split in a part F
e
i
corresponding to an Eulerian description and
a contribution F
ale
i
from the ALE description:
n+1
i
V
i

n
i
V
i
+
_
t
n+1
t
n
F
e
i
(V
, x)dt
_
t
n+1
t
n
F
ale
i
(V
, x, v)dt = 0 (2.185)
For a uniform ow V
i
the ux F
e
i
(V
, x) is equal to zero as well. From the

mathematical consistency of the numerical uxes follows that the integral of the
ux F
ale
i
(V
, x, v) is given by:
_
t
n+1
t
n
F
ale
i
(V
, x, v)dt = V
i
_
t
n+1
t
n
_
i
v nd
i
dt (2.186)
By isolating the conservative variables V
i
from the convective uxes F
ale
i
(V
, x, v)
the function G
i
( x, v) is dened which should satisfy:
V
i
_
t
n+1
t
n
G
i
( x, v)dt = V
i
_
t
n+1
t
n
_
i
v nd
i
dt (2.187)
Substituting equation (2.187) in equation (2.184) yields:
n+1
i

n
i
=
_
t
n+1
t
n
G
i
( x, v)dt =
_
t
n+1
t
n
_
i
v nd
i
dt (2.188)
Generally the mesh positions x are only known at discrete time steps. The mesh
position between t
n
and t
n+1
is parameterized:
x(t) =
n+1
(t) x
n+1
+
n
(t) x
n
(2.189)
where
n
(t) = 1
n+1
(t). The function
n+1
(t) satises
n+1
(t
n+1
) = 1 and
n+1
(t
n
) = 0. The mesh velocity v becomes:
v(t) =

n+1
(t) x
n+1
+

n
(t) x
n
(2.190)
TIME INTEGRATION 53
The sequel depends on the space dimensions of the problem. Using the above
parameterizations, the integrand of the integral on the right hand side of equation
(2.187) becomes a quadratic function of
n+1
for three-dimensional computations.
This integral can be integrated exactly if a Gaussian quadrature with two
integration points is used:
n+1
1
=
1
2
_
1 +
1
3
_

n
1
=
1
2
_
1
1
3
_
w
c
1
=
1
2
(2.191)
n+1
2
=
1
2
_
1
1
3
_

n
2
=
1
2
_
1 +
1
3
_
w
c
2
=
1
2
(2.192)
w
c
1
and w
c
2
are the weight of the respective integration points. The Gaussian
quadrature with two integration points corresponds to the assumption that the
time derivatives

n+1
and

n
at the integration points are given by:
n+1
1
=

n+1
2
=
1
t
n
1
=

n
2
=
1
t
(2.193)
Combining equations (2.189) and (2.192), the mesh position at the integrations
points is:
x
1
c
=
1
2
_
1 +
1
3
_
x
n+1
+
1
2
_
1
1
3
_
x
n
(2.194)
x
2
c
=
1
2
_
1
1
3
_
x
n+1
+
1
2
_
1 +
1
3
_
x
n
(2.195)
Combining equations (2.190) and (2.193), the mesh velocities at the integrations
points are computed as:
v
1
= v
2
=
x
n+1
x
n
t
(2.196)
Using the numerical integration equation (2.188) becomes:
n+1
i

n
i
= t
K
c
k=1
w
c
k
G
i
( x
k
c
, v
k
) (2.197)
This is the discrete geometrical conservation law (DGCL) for the backward Euler
time integration scheme which is fullled for K
c
= 2 and with x
k
c
and v
k
respectively given in equations (2.195) and (2.196). Therefore, the convective
uxes should be integrated as:
_
t
n+1
t
n
F
c
i
(V, x, v)dt = t
K
c
k=1
w
c
k
F
c
i
(V
n+1
, x
k
c
, v
k
) (2.198)
Demirdzic and Peric (1988) describe an alternative way to satisfy the DGCL for
the backward Euler scheme. Writing the right hand side of equation (2.188) as a
sum over all the faces of the control volume, gives:
n+1
i

n
i
=
N
i
j
_
t
n+1
t
n
_
ij
v nd
ij
dt (2.199)
The integral on the right hand side should be computed so that:
_
t
n+1
t
n
_
ij
v nd
ij
dt =
n+1
ij
(2.200)
n+1
ij
is the volume swept by the face
ij
of the control volume
i
between t
n
and t
n+1
. The sum of all these volumes is equal to the change in volume of the
control volume. If these volumes
n+1
ij
are computed exactly based on the face
positions at t
n
and t
n+1
, they can be used to integrate the ux F
ale
i
in a way that
satises the DGCL:
_
t
n+1
t
n
F
ale
i
(V
, x, v)dt =
N
i
j
_
t
n+1
t
n
_
ij
V( v n)d
ij
dt (2.201)
=
N
i
j
I
ij
(V
n+1
)
n+1
ij
(2.202)
where I
ij
(V
n+1
) interpolates the values of the conservative variables V
n+1
at the
cell centers to the midpoint of the face
ij
. In this approach, the mesh velocity
does not have to be calculated explicitly. It is not specied how to integrate the
part F
e
i
corresponding to an Eulerian description of the convective uxes.
The three-point backward dierence scheme Koobus and Farhat (1999) derived
the DGCL for the three-point backward dierence scheme. If the semi-discretized
equation (2.170) for one control volume is integrated in time with the three-point
backward dierence scheme (2.176), the following discrete equation is obtained:
3
2
(
i
V
i
)
n+1
2(
i
V
i
)
n
+
1
2
(
i
V
i
)
n1
+
_
t
n+1
t
n
F
c
i
(V, x, v)dt (2.203)
=
_
t
n+1
t
n
F
d
i
(V, x)dt (2.204)
TIME INTEGRATION 55
Assuming a uniform ow V
i
= V
i
an analogous reasoning as for the backward
Euler scheme leads to:
3
2
n+1
i
2
n
i
+
1
2
n1
i
=
_
t
n+1
t
n
G
i
(V, x, v)dt (2.205)
The left hand side is reordered as:
3
2
(
n+1
i

n
i
)
1
2
(
n
i

n1
i
) =
_
t
n+1
t
n
G
i
(V, x, v)dt (2.206)
The terms in brackets on the left hand side are elaborated using the GCL (2.182).
The mesh positions x are now parameterized between t
n
and t
n+1
as:
x(t) =
n+1
(t) x
n+1
+
n
(t) x
n
+
n1
(t) x
n1
(2.207)
where
n1
(t) = 1
n+1
(t)
n
(t). The function
n+1
(t) satises
n+1
(t
n+1
) = 1
and
n+1
(t
n
) = 0 and the function
n
(t) satises
n
(t
n+1
) = 0 and
n
(t
n
) = 1.
The mesh velocity v becomes:
v(t) =

n+1
(t) x
n+1
+

n
(t) x
n
+

n1
(t) x
n1
(2.208)
In order to satisfy equation (2.206) for three-dimensional computations, four
integration points are used to compute the integral on the right hand side. An
innite number of choices exist for the integration points and the weights. One
particular choice is to use a two-point Gaussian quadrature between t
n+1
and t
n
and another two-point Gaussian quadrature between t
n
and t
n1
:
n+1
1
=
1
2
_
1 +
1
3
_

n
1
=
1
2
_
1
1
3
_

n1
1
= 0 (2.209)
n+1
2
=
1
2
_
1
1
3
_

n
2
=
1
2
_
1 +
1
3
_

n1
2
= 0 (2.210)
n+1
3
= 0
n
3
=
1
2
_
1 +
1
3
_

n1
3
=
1
2
_
1
1
3
_
(2.211)
n+1
4
= 0
n
4
=
1
2
_
1
1
3
_

n1
4
=
1
2
_
1 +
1
3
_
(2.212)
The weights are the product of the weight of a two-point Gaussian quadrature with
the coecients of the terms in brackets in the left hand side of equation (2.206):
w
c
1
=
1
2
3
2
w
c
2
=
1
2
3
2
w
c
3
=
1
2
_
1
2
_
w
c
4
=
1
2
_
1
2
_
(2.213)
The mesh position x
k
c
at the kth-integration point is given by:
x
k
c
=
n+1
k
x
n+1
+
n
k
x
n
+
n1
k
x
n1
(2.214)
The two Gaussian quadratures with two integration points correspond to the
assumption that the mesh velocities at the integrations points are computed as:
v
1
= v
2
=
x
n+1
x
n
t
v
3
= v
4
=
x
n
x
n1
t
(2.215)
Using the numerical integration equation (2.205) becomes:
3
2
n+1
i
2
n
i
+
1
2
n1
i
= t
K
c
k=1
w
c
k
G
i
( x
k
c
, v
k
) (2.216)
This is the DGCL for the three-point backward dierence scheme which is fullled
for K
c
= 4 and with x
k
c
and v
k
respectively given in equations (2.214) and (2.215).
Therefore, the convective uxes should be integrated as:
_
t
n+1
t
n
F
c
i
(V, x, v)dt = t
K
c
k=1
w
c
k
F
c
i
(V
n+1
, x
k
c
, v
k
) (2.217)
The approach of Demirdzic and Peric (1988) can be applied as well to the three-
point backward dierence scheme:
3
2
n+1
i
2
n
i
+
1
2
n1
i
=
N
i
j
_
t
n+1
t
n
_
ij
v nd
ij
dt (2.218)
The left hand side is reordered as:
3
2
(
n+1
i

n
i
)
1
2
(
n
i

n1
i
) =
N
i
j
_
t
n+1
t
n
_
ij
v nd
ij
dt (2.219)
The integral on the right hand side should be computed so that:
_
t
n+1
t
n
_
ij
v nd
ij
dt =
3
2
n+1
ij

1
2
n
ij
(2.220)
If these volumes
n+1
ij
and
n
ij
are computed exactly based on the face positions
at t
n+1
, t
n
and t
n1
, they can be used to integrate the ux F
ale
i
in a way that it
satises the DGCL:
_
t
n+1
t
n
F
ale
i
(V
, x, v)dt =
N
i
j
I
ij
(V
n+1
)
_
3
2
n+1
ij

1
2
n
ij
_
(2.221)
This approach is implemented in CFX.
TIME INTEGRATION 57
The importance of the geometric conservation law The importance of fullling
the DGCL, which ensures that the particular case of a uniform ow is computed
exactly whereas other ow solutions are still approximated, has been thoroughly
studied during the last years. Guillard and Farhat (2000) proved for the backward
Euler and the three-point backward dierence scheme that the fulllment of the
DGCL is a sucient condition to be at least rst order time accurate in the
ALE description. In the case of a non-linear scalar hyperbolic conservation law,
Farhat et al. (2001) show that, for the one-step schemes (2.174), the fulllment
of the DGCL is a sucient and a necessary condition to preserve in the ALE
description the non-linear stability properties that these schemes exhibit in
an Eulerian description. If the DGCL is violated, spurious oscillations develop
during the computations, which may lead to instabilities in the ow solver or
wrong predictions as for instance underestimated utter speeds in aeroelastic
computations. The magnitude of these oscillations is substantially larger for the
three-point backward dierence scheme than for the backward Euler method which
means that it is more critical for time integration schemes which are second order
time accurate in an Eulerian description to satisfy the DGCL than for rst order
time-accurate schemes. The magnitude of these oscillations increases with the
computational time step. For suciently small time steps and smooth mesh
motions the non-linear stability properties of time integration schemes are kept
despite the fact that they violate the DGCL. However, the estimation of the
largest time step which produces reasonable results is cumbersome and problem
dependent. As both the three-point backward dierence scheme and the backward
Euler method are implicit schemes, a limitation on the magnitude of the time step
is not desirable and might drastically increase the computational cost. Therefore, it
is recommended to use time integration schemes that fulll the DGCL. However,
Bo and Gastaldi (2004) and Formaggia and Nobile (2004) proved for a linear
advection-diusion equation of a form similar to equation (2.76) discretized with
nite elements that, even if the DGCL is satised, the Crank-Nicolson and the
three-point backward dierence scheme are only conditionally stable in an ALE
description.
In equations (2.198) and (2.217), the numerical uxes F
c
i
have to be evaluated
K
c
times on dierent mesh congurations. This raises the computational cost,
although the cost for the evaluation of the uxes is only small with respect to
the total cost. If, however, the convective uxes F
c
i
are computed using Roes
numerical ux, the computational costs can be reduced. Roes numerical ux
consists of a part that varies linearly with the mesh conguration and a part that
varies non-linearly with it. As the non-linear part is equal to zero for a uniform
ow, the DGCL can be satised as well by computing the ux once on a unique
averaged mesh conguration (Koobus and Farhat, 1999):
_
t
n+1
t
n
F
c
i
(V, x, v)dt = tF
c
i
(V
n+1
,
K
c
k=1
w
c
k
x
k
c
,
K
c
k=1
w
c
k
v
k
) (2.222)
The results computed with this equation generally dier from the results computed
with equation (2.217) as the ux is a non-linear function of the mesh conguration.
Time integration accuracy
A second approach to extend an Eulerian time integration procedure to an ALE
description is to require that the order of time accuracy obtained in the Eulerian
description is preserved in the ALE description. For aeroelastic computations,
the time accuracy is very important as it inuences the energy exchange between
structure and uid. As the fulllment of the DGCL is a sucient condition to
be at least rst order time accurate in the ALE description, the conditions for
the the three-point backward dierence scheme to be second order time accurate
in the ALE description are derived (Geuzaine et al., 2003; Farhat and Geuzaine,
2004). The DGCL only species rules (2.217) for the time integration of the
convective uxes. For the time integration of the diusive uxes an analogue
numerical integration is assumed. The three-point backward dierence scheme
(2.204) becomes:
3
2
(
i
V
i
)
n+1
2(
i
V
i
)
n
+
1
2
(
i
V
i
)
n1
+ t
K
c
k=1
w
c
k
F
c
i
(V
n+1
, x
k
c
, v
k
) = t
K
d
k=1
w
d
k
F
d
i
(V
n+1
, x
k
d
) (2.223)
The integration points for the diusive uxes may dier from the integration points
for the convective uxes as reected in the possibly dierent mesh positions x
k
c
and x
k
d
. From the analysis of the truncation errors the conditions are derived for
the scheme to be second order time accurate. The conditions for the parameters
associated with the convective uxes are independent of the conditions for the
parameters associated with the diusive uxes.
For the integration of the convective uxes the scheme with four integration points
dened by equations (2.212)-(2.215) satises these conditions for second order time
accuracy as well as its DGCL. An alternative scheme which satises the conditions
for second order time accuracy uses one integration point with a weight equal to
one. As mesh position the position at t
n+1
is taken:
n+1
1
= 1
n
1
= 0
n1
1
= 0 (2.224)
The coecients for the computation of the mesh velocity correspond to the
coecients of the rst three terms in equation (2.223):
n+1
1
=
3
2
1
t
n
1
= 2
1
t
n1
1
=
1
2
1
t
(2.225)
TIME INTEGRATION 59
As this scheme is second order time accurate, but does not satisfy its DGCL, it
shows that the fulllment of the DGCL is not a necessary condition to preserve
the order of time accuracy.
For the integration of the diusive uxes one of the possibilities is to use one
integration point which uses the mesh position at t
n+1
and has a weight equal to
one.
By analogy with equation (2.222) a time integration scheme which computes the
uxes once on a unique averaged mesh conguration might be proposed:
3
2
(
i
V
i
)
n+1
2(
i
V
i
)
n
+
1
2
(
i
V
i
)
n1
+ tF
c
i
(V
n+1
,
K
c
k=1
w
c
k
x
k
c
,
K
c
k=1
w
c
k
v
k
) = tF
d
i
(V
n+1
,
K
d
k=1
w
d
k
x
k
d
) (2.226)
The conditions for second order time accuracy for this scheme are a little bit
dierent from those of the scheme of equation (2.223), but all schemes that satisfy
the conditions related to equation (2.223) satisfy these conditions as well.
If the uxes F
ale
i
are integrated as in equation (2.221) and the diusive uxes are
integrated using one mesh position at t
n+1
, the following time integration scheme
is obtained:
3
2
(
i
V
i
)
n+1
2(
i
V
i
)
n
+
1
2
(
i
V
i
)
n1
+ tF
e
i
(V
n+1
, x
k
)
N
i
j
I
ij
(V
n+1
)
_
3
2
n+1
ij

1
2
n
ij
_
= tF
d
i
(V
n+1
, x
n+1
) (2.227)
Only the mesh position x
k
which is used for the uxes F
e
i
has to be determined.
Geuzaine et al. (2003) proved that this results in a rst order time-accurate scheme
for k = n and in a second order time accurate scheme for k = n+1 which satises
as well its DGCL.
For the nite element discretizations that satisfy the GCL independent of the
time integration scheme, the order of time accuracy can be recovered if the mesh
velocities are computed corresponding to the time integration scheme (Forster
et al., 2006). For the three-point backward dierence scheme the mesh velocity is
given by:
3
2
x
n+1
2 x
n
+
1
2
x
n1
= v
n+1
(2.228)
2.5 Mesh deformation
In section 2.2.1 the governing equations (2.97)-(2.99) for the ow on a domain with
deforming boundaries are derived using the ALE description. In order to compute
the convective velocity c (2.16) the velocities v of the nodes of the mesh still have to
be determined. The movement of these nodes can be chosen arbitrarily but should
preferentially preserve the quality and the renements of the mesh. Computations
on meshes consisting of distorted elements are less accurate and may require more
iterations and smaller time steps. Especially for the thin elements in the boundary
layer, the thin layer of uid near the structure in which the velocity changes
from the velocity of the structure to the free stream value, the limitation of the
distortions is very important and challenging. Ideally, these elements should move
together with the neighbouring structure with the least amount of deformation.
If the quality of the mesh is still deteriorated too much, a new mesh should be
generated and the ow solution has to be projected from the old onto the new mesh.
As this remeshing has a high computational cost and the projection introduces
errors in the ow solution, the challenge is to develop automatic mesh deformation
techniques that minimize the frequency of remeshing.
Meshes that deform to follow the change of the computational domain are called
dynamic meshes. A common method to obtain a suitable motion for the grid points
is the spring analogy by Batina (1990). All the edges of the elements are replaced
by ctitious linear springs. In order to prevent node collisions the spring stiness
is proportional to the inverse of the edge length, which means that closely spaced
nodes exhibit stronger spring forces. This analogy works well if the uid mesh is
not very ne and the mesh motion is relatively small. However, since the rotation
of the edges does not induce any force in the linear springs, strongly distorted
elements are easily obtained in the case of signicant deformations. The distorted
elements have very large or very small angles between adjacent edges. Ultimately,
elements might be inverted, which is also called snap-through or cross-over (e.g.
for two-dimensional triangular elements a vertex of the triangle passes through the
opposite edge). For nite elements the sign of the Jacobian determinant should
be positive within the element domain.
Farhat et al. (1998a) added torsional springs between adjacent edges in the case of
two-dimensional triangular unstructured meshes. The stiness k of the torsional
springs depends on the angle between the adjacent edges as:
k =
1
1 + cos
1
1 cos
=
1
sin
2
(2.229)
Through the term sin the stiness is related to the areas of triangles so that the
area of a triangle cannot become zero or negative. The addition of the torsional
springs makes the method more robust. Degand and Farhat (2002) extended this
method to three-dimensional unstructured meshes consisting of tetrahedra. By
MESH DEFORMATION 61
controlling the volume of the tetrahedra, all collapse mechanisms are prevented.
The volume control is achieved indirectly by torsional springs which are added in
twelve triangles constructed within the tetrahedron.
Analogously, Blom (2000) divided for two-dimensional triangular unstructured
meshes the linear spring stiness of an edge by the angle facing this edge.
As the spring analogy is essentially an elliptic problem, Saint-Venants principle
which states that local perturbations of the solution have only a local impact, is
applicable. Therefore, Blom (2000) multiplied the spring stiness locally with a
factor of two in order to conserve the mesh quality in the boundary layers.
In the spring analogy a discrete pseudo-structural system is used to compute
the mesh deformations. Johnson and Tezduyar (1994) used a continuous pseudo-
structural system for the deformation of the mesh based on the equations of linear
elasticity (2.79):

x
2
u
t
2
X
+ ( + )( u) + ( u) = 0 (2.230)
The displacements of the structure on the uid-structure interface are applied
as boundary conditions. The inertia term is usually taken equal to zero and a
quasi-static equation is used. During the computation of the element stiness
matrices, the Jacobian determinant resulting from the parametric transformation
between the natural coordinates and the physical coordinates, is dropped. This
modication stiens the smaller elements as compared to the larger elements. The
smaller elements maintain their shape, while the larger elements, usually further
away from the structure, take a larger part of the deformations. In order to
further reduce the distortions of the thin elements in the boundary layer near the
structure, the stiness of these elements is increased by Stein et al. (2004). For
the computation of the element stiness matrices a modied equation is used:
K
ji
=
_
B
e
T
j
DB
e
i
J
e
_
J
0
J
e
_
dd (2.231)
where J
0
is a constant. The case = 1 corresponds to dropping the Jacobian
determinant. By changing = 2 for the elements close to the structure, these
elements are more stiened and behave like an extension of the structural elements.
Alternatively, Bar-Yoseph et al. (2001) made the Youngs modulus dependent on
element shape quality measure through a distortion measure.
A quite similar choice to compute the mesh deformation, is to solve the Laplace
equations for the mesh displacements u (Bathe et al., 1999; Robertson and Sherwin,
1999):
2
u = 0 (2.232)
In CFX, a diusion equation with a variable diusivity is applied to obtain the
mesh displacements u:
(k u) = 0 (2.233)
The diusivity can be specied by the user. In order to maintain the quality of
the smaller elements, the inverse of the volume of the nite volumes might be used
as diusivity. As to conserve the mesh quality in the boundary layers the inverse
of the wall distance is a valuable alternative.
Lohner and Yang (1996) applied a Laplacian smoother with a variable diusivity
to the mesh velocities v. The diusivity k depends on the distance to the nearest
deforming boundary. Close to this boundary the diusivity is large and leads to
almost constant mesh velocities, while far away from this boundary the diusivity
is equal to one which yields a uniform deformation of those elements.
Helenbrook (2003) proposed to use a fourth-order dierential equation for the the
mesh displacements u which allows for the specication of two boundary conditions
along each boundary: the mesh displacement u and the normal mesh spacing. The
biharmonic equation is selected as a straightforward generalization of the Laplace
equation.
For all these approaches based on the connectivity between the nodes, a system of
equations has to be solved for all the nodes. Therefore, it is advisable to keep the
ALE part of the uid domain as small as possible and use the Eulerian description
in the other parts of the domain. On the contrary in point-by-point schemes the
displacements at the boundaries are interpolated to all nodes within the mesh
without solving a system of equations. Interpolation techniques as the transnite
interpolation (TFI) are only applicable to structured or block-structured meshes
(Sch afer et al., 2006). Kjellgren and Hyvarinen (1998) interpolated the mesh
velocities v based on an analytical function of the distance to the closest point on
the moving boundary. In a layer around the moving boundary the mesh velocity
is equal to the velocity at the deforming boundary. This method can be applied
to unstructured meshes and complex geometries. de Boer et al. (2007) used radial
basis functions to interpolate the displacements for unstructured meshes. Only a
small system of equations which consists of the nodes on the boundaries has to be
solved.
CONCLUSION 63
2.6 Conclusion
In this chapter the techniques to perform uid ow computations on a domain with
moving boundaries are described. The uid mesh is aligned with the deforming
boundary and should follow its deformations. Therefore, the governing equations
for the uid ow are derived in an ALE description. The convective velocity in
these equations is the relative velocity between uid particles and the mesh. The
mesh velocity is arbitrary but should preferentially preserve the quality and the
renements of the mesh.
The spatial discretization by means of the nite element (e.g. Flotran) and the
nite volume method (e.g. CFX) is discussed. The concepts of stabilization
for nite element discretizations and upwind discretizations for the nite volume
method, which are needed for ow computations, are presented.
The time integration techniques for an Eulerian description are described. The
implication of the ALE on the time accuracy and the stability of the time
integration is discussed and related to the geometric conservation law. The implicit
three-point backward dierence method will be used.
Finally, the dierent options to compute the uid mesh velocity and deformation
are reviewed. In CFX the mesh deformation is computed by means of a diusion
equation with a variable diusivity.
Chapter 3
Shell structures
3.1 Introduction
Shells are widely used in civil engineering structures as silos, chimneys, water
towers, cooling towers and tensile structures.
The present chapter focusses on the computation of the linear and the geometrical
non-linear response of shell structures. First, the virtual work formulation which
leads to the nite element method is briey reviewed. The conjugate stresses and
strains used for the computation of geometrical non-linear structures are described.
The Newton-Raphson technique solves the non-linear system of equations. In order
to compute the transient geometrical non-linear response of shell structures direct
time integration is applied.
As an application, the cylindrical shell structure of the silos in the port of Antwerp
is studied. As for the prediction of ovalling oscillations the exact value of the
in order to determine the modal damping ratios, eigenfrequencies and eigenmodes
of the silos. These modal parameters allow for the validation of a three-dimensional
nite element model of the silo. It is investigated if geometrical non-linear
behaviour should be taken into account during the ovalling of the silos.
65
66 SHELL STRUCTURES
3.2 Geometric non-linear behaviour
In structural mechanics, three types of non-linearity exist: material, geometric and
contact non-linearities. In the case of material non-linearities the relation between
stresses and strains is non-linear: the material properties become dependent on the
stress state (e.g. yield and creep). For geometric non-linearities, the deformations
are so large that the change in geometry inuences the stiness. The equilibrium
should be expressed with respect to the deformed conguration. Pressure loads
may change direction during deformation, which are called follower forces. In
the case of contact non-linearities, the boundary conditions change during the
computation. The structural non-linear behaviour (Bonet and Wood, 1997) can
be computed using the nite element method (Criseld, 1991, 1997; Bathe, 1996).
The current section focusses on geometrical non-linear behaviour. A further
distinction exists between cases where only the rotations and displacements are
large and cases where the strains are large as well. The reader familiar with the
theoretical background to geometrical non-linear behaviour may skip this section.
3.2.1 Virtual work
In structural analysis a Lagrangian description is commonly used. As the material
particles are followed in time, the representation in the material domain does not
change with time. The material conguration is therefore often called the initial
conguration, while the representation in the spatial domain at time t is called
the current conguration.
In the nite element method, a weighted residual formulation of the governing
momentum equation (2.79) in the spatial domain is used. In the context of
structural mechanics, the weighting function might be interpreted as a virtual
displacement u and the weighted residual formulation expresses the virtual work
done by the virtual displacement:
W =
_
x
(
x
2
u
t
2
X
+
x
+
x
b) ud
x
= 0 (3.1)
Using integration by parts for the second term, equation (3.1) becomes:
_
2
u
t
2
X
ud
x
+
_
x
:
x
ud
x
=
_
x
(
T
u)d
x
+
_
x
b ud
x
(3.2)
GEOMETRIC NON-LINEAR BEHAVIOUR 67
Gauss theorem enables to write the volume integral in the rst term on the right
hand side as a boundary integral:
_
2
u
t
2
X
ud
x
+
_
x
:
x
ud
x
=
_
T
u n
x
d
x
+
_
x
b ud
x
(3.3)
Writing the rst term on the right hand as a function of the traction (2.67), the
virtual work becomes:
_
2
u
t
2
X
ud
x
+
_
x
:
x
ud
x
=
_
x
t ud
x
+
_
x
b ud
x
(3.4)
Using the symmetry of the Cauchy stress tensor (2.85), the integrand of the second
term on the left hand side is written as:
:
x
u =
1
2
( +
T
) :
x
u = :
1
2
(
x
u + (
x
u)
T
) = : (3.5)
where is called the virtual strain. Substituting equation (3.5) in (3.4), the
expression for virtual work in the current conguration is obtained:
_
2
u
t
2
X
ud
x
+
_
x
: d
x
=
_
x
u td
x
+
_
x
b ud
x
(3.6)
The second term on the left hand side is the internal virtual work, while the right
hand side describes the external virtual work.
When the displacements, rotations and strains are small, the dierence between
the initial and the current conguration can be neglected. All integrations are
performed over the original volume and surface of the body and all derivatives are
taken with respect to the initial conguration:
_
2
u
t
2
X
ud
X
+
_
X
: d
X
=
_
X
utd
X
+
_
X
bud
X
(3.7)
Hookes law assumes a linear relation between the stress tensor and the small
strain tensor:
= C : (3.8)
In the case of an isotropic material, this relation can be described by two
independent Lame constants and :
= (tr)I + 2 (3.9)
68 SHELL STRUCTURES
The Lame constants are easily derived from the Youngs modulus E and the
Poisson coecient . After discretization of the displacements using locally dened
shape functions, the nite element system of equations is obtained from equation
(3.7):
M u +Ku = f (3.10)
In the case of geometric non-linear behaviour, the stiness matrix and, possibly
the load vector, become dependent on the displacement vector:
M u +K(u)u = f(u) (3.11)
3.2.2 Kinematics
As the conguration of the structure changes continuously, appropriate strain
measures should be dened. The deformation gradient (2.20) relates the relative
position of two particles in the material domain to their relative position in the
spatial domain. Substituting equation (2.6) for the spatial vector x into equation
(2.20) for the deformation gradient F gives:
F =
X
x =
X
(X+u) = I +
X
u = I +D (3.12)
where D is the displacement gradient with respect to the material coordinates.
The inverse deformation gradient F
1
relates an innitesimal spatial vector dx to
the corresponding innitesimal material vector dX:
dX =
x
Xdx = F
1
dx (3.13)
Using equation (2.6), the inverse deformation gradient F
1
is expressed in terms
of the displacement gradient with respect to the spatial coordinates:
F
1
=
x
X =
x
(x u) = I
x
u (3.14)
Strain tensors
The scalar product of two vectors includes the lengths of both vectors and the
angle enclosed between the two vectors. Therefore, the change in the scalar
product of two vectors between the material and the spatial domain is a measure
of deformation. Using equation (2.20), the scalar product of two spatial vectors is
expressed in terms of the scalar product of the two material vectors:
dx
1
dx
2
= dX
1
F
T
FdX
2
= dX
1
CdX
2
(3.15)
where C = F
T
F denotes the right Cauchy-Green deformation tensor. The change
in scalar product of two vectors from the material to the spatial domain is expressed
with reference to the material domain:
dx
1
dx
2
dX
1
dX
2
= (dX
1
CdX
2
) dX
1
dX
2
(3.16)
= dX
1
(CI)dX
2
(3.17)
= 2dX
1
1
2
(CI)dX
2
= 2dX
1
EdX
2
(3.18)
where E is called the Green-Lagrange strain tensor. Recalling equation (3.12) the
Green-Lagrange strain tensor is obtained in terms of the displacements:
E =
1
2
(CI) =
1
2
(F
T
FI) =
1
2
((I +
X
u)
T
(I +
X
u) I) (3.19)
=
1
2
(
X
u + (
X
u)
T
+ (
X
u)
T
X
u) (3.20)
=
1
2
(D+D
T
+D
T
D) (3.21)
The Green-Lagrange strain tensor is symmetric and objective, which means that
it is equal to zero if a rigid body rotation is applied to the body. The rst and
second terms are linear and only a quadratic higher-order term appears.
Alternatively, the scalar product of two material vectors is expressed in terms of
the scalar product of the two spatial vectors:
dX
1
dX
2
= dx
1
F
T
F
1
dx
2
= dx
1
b
1
dx
2
(3.22)
where b = FF
T
denotes the left Cauchy-Green deformation tensor. The change in
scalar product of two vectors from the material to the spatial domain is expressed
with reference to the spatial domain:
dx
1
dx
2
dX
1
dX
2
= dx
1
dx
2
dx
1
b
1
dx
2
(3.23)
= dx
1
(I b
1
)dx
2
(3.24)
= 2dx
1
1
2
(I b
1
)dx
2
= 2dx
1
edx
2
(3.25)
where e is the Almansi strain tensor:
e =
1
2
(I b
1
) =
1
2
(I (F
1
)
T
F
1
) (3.26)
=
1
2
(I (I
x
u)
T
(I
x
u)) (3.27)
=
1
2
(
x
u + (
x
u)
T
(
x
u)
T
x
u) (3.28)
70 SHELL STRUCTURES
The Almansi strain tensor is symmetric as well.
If the quadratic term is dropped, the small strain tensor is recovered:
=
1
2
(
x
u + (
x
u)
T
) (3.29)
The change in squared length of a vector between the material and the spatial
domain is obtained from the change in the scalar product (3.18):
dx dx dX dX = 2dX
T
EdX (3.30)
dl
2
dL
2
= 2dX
T
EdX (3.31)
dl
2
dL
2
2dL
2
=
dX
dL
E
dX
dL
(3.32)
dl
2
dL
2
2dL
2
= n
X
En
X
(3.33)
Generally the strain is dened as:
1
a
dl
a
dL
a
dL
a
(3.34)
Commonly used strain measures correspond to dierent choices of the parameter
a:
a = 2
dl
2
dL
2
2dL
2
= n
X
En
X
Green-Lagrange strain
a = 1
dldL
dL
Biot (engineering) strain
a = 0 ln dl logarithmic strain
Polar decomposition
Applying the right polar decomposition, the deformation gradient F is decomposed
as the product of an orthogonal tensor R and a symmetric tensor U:
F = RU (3.35)
The orthogonal tensor R corresponds to a rigid body rotation, while the symmetric
tensor U corresponds to stretching of the material. The spatial vector dx is
obtained by rst applying the right stretch tensor U to the material vector dX
and then the rotation tensor R. Substituting the right polar decomposition in the
denition of the right Cauchy-Green deformation tensor C and recalling that R
is orthogonal and U is symmetric, gives:
C = F
T
F = U
T
R
T
RU = U
T
U = UU = U
2
(3.36)
In order to obtain the stretch tensor U, the spectral decomposition of the right
Cauchy-Green deformation tensor C is calculated:
C =
3
=1
(3.37)
The eigenvalues
2
are equal to the square of the eigenvalues
of U and the
eigenvectors N
of the right Cauchy-Green deformation tensor C and the stretch

tensor U are identical. The stretch tensor is computed as:
U =
3
=1
(3.38)
Using equation (3.35) the rotation tensor R is computed from U and F.
Alternatively, applying the left polar decomposition, the deformation gradient F is
decomposed as the product of a symmetric tensor V times the orthogonal rotation
tensor R:
F = VR (3.39)
The spatial vector dx is now obtained by rst applying the rotation tensor R to
the material vector dX and then the left stretch tensor V. The left stretch tensor
V is expressed in terms of the right stretch tensor U:
V = RUR
T
(3.40)
Substituting the left polar decomposition in the denition of the left Cauchy-Green
deformation tensor C and recalling that R is orthogonal and Vis symmetric, gives:
b = FF
T
= VRR
T
V
T
= VV
T
= VV = V
2
(3.41)
The spectral decomposition of the left stretch tensor V is similar to the spectral
decomposition of the right stretch tensor U in equation (3.38):
V =
3
=1
(3.42)
Combining equations (3.40) and (3.38), the spectral decomposition of the left
stretch tensor V reads as:
V = R(
3
=1
)R
T
=
3
=1
(RN
) (RN
) (3.43)
72 SHELL STRUCTURES
The comparison of both expressions for the spectral decomposition of V yields:
n
= RN
(3.44)
The deformation gradient dened in equation (2.20) transforms a vector dX
1
=
dL
1
N
1
aligned with N
1
in the material domain into the vector dx
1
in the spatial
domain. Using the right polar decomposition, the vector dx
1
is expressed as:
dx
1
= FdX
1
= (RU)(dL
1
N
1
) = dL
1
(RUN
1
) = dL
1
(R
1
N
1
)
= dL
1
1
(RN
1
) = dL
1
1
n
1
= dl
1
n
1
(3.45)
The eigenvalue
1
gives the ratio between the length of a material vector in the
direction N
1
and the length of the corresponding spatial vector, which is aligned
with n
1
:
1
=
dl
1
dL
1
(3.46)
Substituting equations (3.36) and (3.41) in respectively the denitions of the
Green-Lagrange strain tensor E (3.21) and the Almansi strain tensor e (3.28)
and inserting the spectral decompositions gives:
E =
1
2
(CI) =
1
2
(U
2
I) =
1
2
3
=1
(
2
1)N
e =
1
2
(I b
1
) =
1
2
(I V
2
) =
1
2
3
=1
(1
2
)n
(3.47)
The extension of the dierent strain measures dened in equation (3.34) yields the
following strain tensors with respect to the material or the spatial domain:
E
(a)
=
1
a
(U
a
I) =
1
a
3
=1
(
a
1)N
e
(a)
=
1
a
(I V
a
) =
1
a
3
=1
(1
a
)n
(3.48)
3.2.3 Conjugate stress and strain tensors
In equation (3.6) the virtual work is expressed in the current conguration using
integrals over the current volume
x
and the current area
x
. For the computation
of the virtual strain (3.5) derivatives with respect to the current conguration
are needed. As the current conguration is unknown, an expression for virtual
work in the initial conguration is derived.
Kirchho stress tensor
Using the relation d
x
= Jd
X
between the initial volume d
X
and the current
volume d
x
and the relation d
x
= JF
T
d
X
between the initial area d
X
and
the current area d
x
, an expression with integrals over the initial volume
X
and
the initial area
X
emerges from equation (3.6):
_
2
u
t
2
X
uJd
X
+
_
X
: Jd
X
=
_
X
tu
d
x
d
X
d
X
+
_
x
buJd
X
(3.49)
The Kirchho or nominal stress tensor is dened as:
= J (3.50)
The Kirchho stress tensor is work conjugate to the strain with respect to the
initial volume. The Kirchho stress vector t
X
is dened as:
t
X
= t
d
x
d
X
(3.51)
which is the traction vector per unit undeformed area. Both traction vectors t
X
and t have the same direction. Finally the virtual work of equation (3.49) is
rewritten inserting the above denitions and using equation (2.62):
_
2
u
t
2
X
ud
X
+
_
X
: d
X
=
_
X
t
X
ud
X
+
_
X
bud
X
(3.52)
First Piola-Kirchho stress tensor
For the computation of the virtual strain (3.5) still derivatives with respect
to the current conguration are needed. The relation between the gradient of a
virtual displacement with respect to the current coordinates and the gradient with
respect to the initial coordinates is:
x
u =
X
uF
1
(3.53)
74 SHELL STRUCTURES
Substituting equations (3.5) and (3.53) into the second term on the left hand side
of equation (3.49) gives:
_
2
u
t
2
X
uJd
X
+
_
X
:
X
uF
1
Jd
X
=
_
X
t u
d
x
d
X
d
X
+
_
x
b uJd
X
(3.54)
Using the denition of the double product for the the second term on the left hand
side yields:
_
2
u
t
2
X
uJd
X
+
_
X
JF
T
:
X
ud
X
=
_
X
t u
d
x
d
X
d
X
+
_
x
b uJd
X
(3.55)
The rst factor in the double product is the rst Piola-Kirchho stress tensor
P
X
= JF
T
, which gives the current force per unit undeformed area and is not
symmetric. Its direction corresponds to the direction of the force in the current
conguration. The second factor is the virtual change of the displacement gradient
D (3.12):
D =
X
u (3.56)
Combining equations (3.52), (3.55), (2.82) and (3.56), an expression for virtual
work in the initial conguration is obtained:
_
2
u
t
2
X
ud
X
+
_
X
P
X
: Dd
X
=
_
X
t
X
ud
X
+
_
X
b ud
X
(3.57)
This expression can also be obtained by using a weighted residual formulation of
the governing momentum equation (2.84) in the material domain with an arbitrary
virtual displacement u as weighting function.
Second Piola-Kirchho stress tensor
The Kirchho stress vector t
X
as dened in equation (2.82) still points in the same
direction as the Cauchy stress vector t. Using the deformation gradient F, it is
transformed to the initial conguration:
F
1
t
X
= F
1
P
X
n
X
(3.58)
The second Piola-Kirchho stress tensor is dened as:
S = F
1
P
X
= JF
1
F
T
= F
1
F
T
(3.59)
The denition shows that the second Piola-Kirchho stress tensor is symmetric.
The stress vector t
X
is similarly dened as:
t
X
= F
1
t
X
= Sn
X
(3.60)
Introducing P
X
= FS in the second term on the left hand side of equation (3.57)
gives:
_
2
u
t
2
X
ud
X
+
_
X
FS :
X
ud
X
=
_
X
t
X
ud
X
+
_
X
bud
X
(3.61)
Using the denition of the double product yields for the integrand of the second
term on the left hand side:
FS :
X
u = tr(S
T
F
T
X
u) (3.62)
Recalling that the second Piola-Kirchho stress tensor is symmetric, gives:
FS :
X
u = tr(
1
2
(S +S
T
)
T
F
T
X
u) (3.63)
= tr(S
T
1
2
(F
T
X
u + (
X
u)
T
F)) (3.64)
Substituting the expression for the deformation gradient F (3.12), it becomes:
FS :
X
u = tr(S
T
1
2
((I +
X
u)
T
X
u + (
X
u)
T
(I +
X
u)) (3.65)
= tr(S
T
1
2
(
X
u + (
X
u)
T
+ (
X
u)
T
X
u
+ (
X
u)
T
X
u)) (3.66)
where the second factor is the virtual Green-Lagrange strain:
E =
1
2
(
X
u + (
X
u)
T
)
. .
E
l
+
1
2
((
X
u)
T
X
u + (
X
u)
T
X
u)
. .
E
nl
(3.67)
76 SHELL STRUCTURES
which nally yields:
FS :
X
u = S : E (3.68)
Combining equations (3.61) and (3.68), an expression for virtual work in the initial
conguration is obtained:
_
2
u
t
2
X
ud
X
+
_
X
S : Ed
X
=
_
X
t
X
ud
X
+
_
X
b ud
X
(3.69)
3.2.4 Incremental virtual work
Suppose all the congurations up to time t
n
are known and the unknown
conguration at time t
n+1
has to be computed. For the reference conguration
with respect to which the virtual work (3.69) is expressed, any known conguration
can be used. In the total Lagrangian (TL) formulation all variables are referred
to the initial conguration at time 0 for all time steps, while in the updated
Lagrangian (UL) formulation the reference conguration is updated for every time
step and all variables are referred to the last calculated conguration at time t
n
.
Total Lagrangian formulation
In the total Lagrangian (TL) formulation, the displacements at time t
n+1
are
obtained by adding a displacement increment to the known displacements at time
t
n
:
u
n+1
= u
n
+ u (3.70)
The subscript indicates at which time the displacement occurs. The second Piola-
Kirchho stress at time t
n+1
is decomposed in the initial conguration as:
S
n+1
0
= S
n
0
+ S
0
(3.71)
The subscript indicates the conguration with respect to which the stress is dened.
The stress increment S
0
is given by:
S
0
= C
n+1
0
: E
0
(3.72)
where C
n+1
0
is the incremental stress-strain tensor at time t
n+1
with respect to
the initial conguration. The Green-Lagrange strain at time t
n+1
with respect to
the initial conguration is:
E
n+1
0
= E
n
0
+ E
0
(3.73)
The Green-Lagrange strain increment is expressed as a function of the displacement
increment:
E
0
=
1
2
(
X
u + (
X
u)
T
)
. .
E
l
0
(3.74)
+
1
2
((
X
u
n
)
T
X
u + (
X
u)
T
X
u
n
. .
E
nl
0
+(
X
u)
T
X
u
. .
neglected
) (3.75)
As the virtual displacements are not a function of time, from the virtual Green-
Lagrange strain (3.67) the virtual strain increment is derived:
(E
0
) =
1
2
((
X
u)
T
X
u + (
X
u)
T
X
u) (3.76)
Using equations (3.71- 3.76) the internal virtual work in equation (3.69) is rewritten
as a function of the displacement increment:
_
X
S
0
: E
0
d
X
+
_
X
S
0
: (E
0
)d
X
(3.77)
Substituting equation (3.72) for the constitutive behaviour into the previous
equation gives:
_
X
E
0
: C
n+1
0
: E
0
d
X
+
_
X
S
0
: (E
0
)d
X
(3.78)
Splitting the virtual Green Lagrange strain (3.67) and the strain increment (3.75)
in their linear and non-linear parts yields:
_
X
E
l
0
: C
n+1
0
: E
l
0
d
X
+
_
X
E
l
0
: C
n+1
0
: E
nl
0
d
X
+
_
X
E
nl
0
: C
n+1
0
: E
l
0
d
X
+
_
X
E
nl
0
: C
n+1
0
: E
nl
0
d
X
+
_
X
S
0
: (E
0
)d
X
(3.79)
After discretization the rst term becomes the linear stiness matrix K
0
. The
second, third and fourth term form the initial displacement (slope) stiness matrix
K
L
and the last term is the geometric stiness (initial stress, stress stiening)
matrix K
.
78 SHELL STRUCTURES
Updated Lagrangian formulation
In the updated Lagrangian (UL) formulation, the reference conguration is
updated to the last calculated conguration. The updated material coordinates
become:
X
n+1
= X
n
+u
n
(3.80)
With respect to the updated reference conguration, the displacements at time
t
n+1
are:
u
n+1
= u (3.81)
The second Piola-Kirchho stress tensor at time t
n+1
with respect to the updated
reference conguration is:
S
n+1
n
=
n
+ S
n
(3.82)
where
n
are the Cauchy stresses computed in the previous time step. The stress
increment S
n
is given by:
S
n
= C
n+1
n
: E
n
(3.83)
where C
n+1
n
is the incremental stress-strain tensor at time t
n+1
with respect to
the updated conguration. The Green-Lagrange strain at time t
n+1
with respect
to the updated conguration is:
E
n+1
n
= E
n
(3.84)
The Green-Lagrange strain increment is expressed as a function of the displacement
increment:
E
n
=
1
2
(
X
u + (
X
u)
T
)
. .
n
+
1
2
((
X
u)
T
X
u)
. .
neglected
(3.85)
The rst term is equal to the engineering strain increment computed at time t
n
.
The virtual Green-Lagrange strain reduces to:
E
n
=
1
2
(
X
u + (
X
u)
T
)
. .
n
+
1
2
((
X
u)
T
X
u)
. .
neglected
(3.86)
The rst term corresponds to the virtual engineering strain as computed at time
t
n
. The expression for the virtual strain increment is identical to equation (3.76).
Using equations (3.82-3.85) the internal virtual work in equation (3.69) is rewritten
as a function of the displacement increment:
_
X
S
n
: E
n
d
X
+
_
n
: (E
n
)d
X
(3.87)
Substituting equation (3.83) for the constitutive behaviour into the previous
equation gives:
_
X
E
n
: C
n+1
n
: E
n
d
X
+
_
n
: (E
n
)d
X
(3.88)
Neglecting the higher order terms in the virtual Green Lagrange strain (3.86) and
the strain increment (3.85) yields:
_
n
: C
n+1
n
:
n
d
X
+
_
n
: (E
n
)d
X
(3.89)
After discretization the rst term becomes the linear stiness matrix K
0
. The
second term is the geometric stiness matrix K
. In contrast to the total

Lagrangian formulation, there is no initial displacement stiness matrix.
Follower forces
In the external virtual work, body forces (e.g. gravity) and nodal forces acting
in the global coordinate system are not deformation dependent. On the contrary,
the traction applied on a surface changes if the area or the normal of this surface
changes during deformation. Loads which are dependent on the deformations are
called follower forces. In the case of a uniform normal pressure p applied to a
surface, the external virtual work (3.6) in the current conguration becomes:
_
x
(p)n
x
ud
x
(3.90)
If the surface is parameterized using a single isoparametric element, the normal
and the innitesimal area are expressed as a function of
x
and
x
:
_
(p)
_
x
n+1

x
n+1
_
ud
(3.91)
Rewriting this equation as a function of the displacement increment gives:
_
(p)
_
x
n

_
u
u
_
x
n

_
u
u
__
d
(3.92)
After discretization this becomes the follower load stiness matrix K
p
, which is
nonsymmetric.
80 SHELL STRUCTURES
3.2.5 Newton-Raphson procedure
In order to solve the non-linear equations of the previous sections, the load is
applied in a number of load increments and an iterative procedure is used to obtain
the solution for each load increment. At each load level equilibrium equations are
performed in order to obtain convergence:
1. Set i = 1
2. Predict the displacements u
(0)
. Usually the converged solution of the
previous load increment is used for the displacement u
(0)
and the restoring
force f
(0)
r
. If a dierent prediction is used, compute the values corresponding
to this prediction. Compute the out-of-balance load vector:
r
(0)
= f
(0)
e
f
(0)
r
(3.93)
3. Compute the tangent stiness matrix:
K
(i1)
T
=
f
(i1)
r
u
(i1)
(3.94)
4. Compute the displacement increment u
(i)
from:
K
(i1)
T
u
(i)
= r
(i1)
(3.95)
and a new total displacement from:
u
(i)
= u
(i1)
+ u
(i)
(3.96)
5. Compute the restoring force f
(i)
r
corresponding to the element stresses.
For material non-linear behaviour a strain path has to be chosen and the
constitutive equation has to be integrated along this path.
6. Compute the external applied force f
(i)
e
and the out-of-balance load vector
r
(i)
:
r
(i)
= f
(i)
e
f
(i)
r
(3.97)
Only in the case of follower forces (as pressures) the external applied force
might change each iteration.
7. If the maximum number of equilibrium iterations is reached: i = i
max
, stop.
8. Check the convergence for the displacements:
|u
(i)
|
L
2
<
u
|u
(i)
|
L
2
(3.98)
and for the out-of-balance load vector:
|r
(i)
|
L
2
<
r
|f
(i)
e
|
L
2
(3.99)
The L
2
-norm consists of the square root of the sum of squares (SRSS) of
the nodal values. Quantities with dierent units should not be summed
in the convergence criteria, which are therefore checked separately for
displacements and rotations and for forces and moments. If the convergence
criteria are fullled, stop. Otherwise set i = i + 1 and go to step 3.
A number of modications can be made to this scheme.
In the modied Newton-Raphson method, the tangent stiness matrix K
(0)
T
of the
rst iteration is used for all equilibrium iterations during a load increment. The
tangent stiness matrix only has to be factorized once per load increment. In
the initial-stiness Newton-Raphson method, the tangent stiness matrix K
(0)
T
is
evaluated only once in the rst iteration of the rst load increment and is used in
all iterations of all load increments.
Adaptive descent uses a weighted average of the secant and the tangent stiness
matrix to compute the displacement increment (3.95):
K
(i1)
= K
(i2)
T
+ (1 )K
(i1)
T
(3.100)
The descent parameter is adapted according to the convergence behaviour.
The line search technique alternatively computes a new total displacement from:
u
(i)
= u
(i1)
+ u
(i)
(3.101)
where the underrelaxation factor is determined by minimizing the total potential
energy of the structure.
The magnitude of a new load increment is predicted based on the number of
equilibrium iterations needed at the previous load increment, the maximum
allowable strain increment and the response frequency for dynamic computations.
If the convergence criteria have not been satised within the maximum number of
equilibrium iterations, the load increment is bisected and the iterative procedure
is repeated.
If the behaviour of a structure beyond a (local) maximum in the load-deection
curve has to be studied, methods which allow for an increase in displacement with
decrease in load are needed. A possibility is to apply displacement increments
82 SHELL STRUCTURES
instead of load increments. A more general approach is provided by the arc-length
method, which uses a variable load increment during the equilibrium iterations. A
constraint equation relates the load increment to the displacement increment:
(f
(i)
e
)
2
+
(
i
u
(i)
)
T
(
i
u
(i)
)
2
= (l)
2
(3.102)
where l is the arc-length.
3.2.6 Time integration
The direct time integration is performed using the Newmark method. The
equilibrium equations are formulated at time t
n+1
:
M u
n+1
+C u
n+1
+Ku
n+1
= f
n+1
(3.103)
The acceleration u
n+1
and the velocity u
n+1
at time t
n+1
are expressed as a
function of the displacement u
n+1
at time t
n+1
and known quantities at time t
n
(Bathe, 1996):
u
n+1
=

t
_
u
n+1
u
n
_
u
n
2
2
t u
n
(3.104)
u
n+1
=
1
t
2
_
u
n+1
u
n
_
1
t
u
n
1 2
2
u
n
(3.105)
Substituting expressions (3.104) and (3.105) in the equilibrium equations (3.103)
at time t
n+1
yields:
_
M
t
2
+
C
t
+K
_
u
n+1
= f
n+1
+M
_
u
n
t
2
+
u
n
t
+
1 2
2
u
n
_
+C
_

t
u
n
+

u
n
+
2
2
t u
n
_
(3.106)
This is the general equation for all Newmark time integration schemes, which are
implicit schemes. For =
1
4
and =
1
2
the constant-average-acceleration method
or trapezoidal rule is obtained which is unconditionally stable and second-order
time accurate. In order to reduce the period elongation and obtain suciently
accurate results, the time step should be a fraction (e.g. 1/20) of the eigenperiods
that contribute to the response.
An alternative time integration method is the midpoint rule. The equilibrium
equations are formulated at time t
n+
1
2
:
M u
n+1/2
+C u
n+1/2
+Ku
n+1/2
= f
n+1/2
(3.107)
The acceleration u
n+1/2
, velocity u
n+1/2
and displacement u
n+1/2
at the midpoint
are approximated as:
u
n+1/2
=
1
2
_
u
n+1
+ u
n
_
(3.108)
u
n+1/2
=
1
2
_
u
n+1
+ u
n
_
(3.109)
u
n+1/2
=
1
2
_
u
n+1
+u
n
_
(3.110)
The combination of these equations with the equations (3.104) and (3.105) yields
the midpoint time integration scheme. In the linear case this method is identical
to the trapezoidal rule.
84 SHELL STRUCTURES
3.3 Structural behaviour of a silo
In this section the structural behaviour of the silos in the port of Antwerp is
analysed. As for the prediction of ovalling oscillations the exact value of the
in order to determine the modal damping ratios, eigenfrequencies and eigenmodes
of the silos. The mode shapes and frequencies of a single silo are rst computed
with a harmonic nite element model, in order to support the design of the in situ
experiment. Radial accelerations are measured in situ under ambient wind loading
and modal parameters are extracted from the output-only data in order to validate
a nite element model of the silo (Dooms et al., 2006b). A three-dimensional nite
element model of the silo accounts for the inuence of the boundary conditions
on the eigenfrequencies and mode shapes. Finally, the importance of geometrical
non-linear behaviour is studied for deformations of order of magnitude of 0.1 m.
3.3.1 Description of the silo structure
The silos are circular cylindrical shell structures with a diameter D = 5.5 m and
a height h = 25 m (gure 1.3). One cylinder consists of 10 aluminium sheets with
a height of 2.5 m and a thickness that decreases with the height (table 3.1). The
height-to-radius ratio h/R = 9.1 and the radius-to-thickness ratio ranges from
R/t = 262 at the bottom to R/t = 458 at the top.
At the top and the bottom, a cone is welded to the cylinder at an angle of 15
and 60
with the horizontal plane, respectively. An octagonal beam supports the

bottom of the cylinder and is bolted to the silo at 4 points around the circumference
(gure 3.1). The silos are made of aluminium with a Youngs modulus E = 67600
10
6
N/m
2
, a Poissons ratio = 0.35 and a density = 2700 kg/m
3
.
Height Thickness
[m] [mm]
0.0 - 2.5 10.5
2.5 - 5.0 9.0
5.0 - 7.5 8.5
7.5 - 10.0 7.5
10.0 - 12.5 7.0
12.5 - 15.0 6.5
15.0 - 25.0 6.0
Table 3.1: Thickness of the aluminium plates of the silo as a function of the height.
STRUCTURAL BEHAVIOUR OF A SILO 85
Figure 3.1: Octagonal beam with four bolted connections at the bottom of the
cylinder.
3.3.2 Harmonic nite element model
In order to design a proper experimental setup, the eigenfrequencies and mode
shapes are computed with a harmonic nite element model of a single silo. In a
cylindrical frame of reference, the displacements of an axisymmetric structure are
decomposed as (Zienkiewicz and Taylor, 2005):
_
_
u
r
(r, , z, t)
u
(r, , z, t)
u
z
(r, , z, t)
_
_
=
n=0
_
_
cos(n) 0 0
0 sin(n) 0
0 0 cos(n)
_
_
_
_
u
s
rn
(r, z, t)
u
s
n
(r, z, t)
u
s
zn
(r, z, t)
_
_
(3.111)
+
n=0
_
_
sin(n) 0 0
0 cos(n) 0
0 0 sin(n)
_
_
_
_
u
a
rn
(r, z, t)
u
a
n
(r, z, t)
u
a
zn
(r, z, t)
_
_
(3.112)
where n indicates the circumferential wave number and the superscripts s and a
refer to the symmetric and antisymmetric terms with respect to . For each value
of n, the eigenfrequencies and mode shapes for the symmetric case are calculated
using the harmonic 2-node SHELL61 element of the nite element program Ansys
(Ansys, 2005b). Each mode shape is referred to by a couple (m, n), where m
denotes the half wave number in the axial direction (m/2 is the number of axial
waves) and n is the number of circumferential waves. The antisymmetric case
yields identical eigenfrequencies and similar mode shapes.
86 SHELL STRUCTURES
Figure 3.2a shows the nite element model and the applied boundary conditions.
On the axis of symmetry, the radial displacement u
r
and the rotation
are
constrained. The vertical displacements u
z
are restricted around the circumference
at the bottom of the cylinder.
a. Model
b. (1, 4) c. (1, 5) d. (1, 3) e. (1, 6) f. (2, 5)
4.65 Hz 5.52 Hz 5.99 Hz 7.38 Hz 8.78 Hz
Figure 3.2: Harmonic nite element model of the silo and selected eigenmodes.
A nite element length l
e
= 0.83 m is used, except near the connections between
the cylinder and the cones, where local bending waves occur. At the top of the
cylinder, the axial wavelength is estimated as (Billington, 1965):
z
=
2
Rt
4
_
3(1
2
)
(3.113)
resulting into
z
= 0.634 m for a cylinder with a radius R = 2.75 m and a shell
thickness t = 0.006 m. The number of elements per wavelength is equal to 12, so
that the element length is equal to l
e
= 0.053 m near the connections.
Figures 3.2b up to 3.2f show ve mode shapes referred to by the couple (m, n).
Mode (1,4) at 4.65 Hz has the lowest frequency. The degree of clamping of the
bottom of the cylinder increases with increasing circumferential wave number n.
Figure 3.3 summarizes the eigenfrequencies as a function of m and n. For modes
corresponding to m = 1, the eigenfrequency is minimal for n = 4. Whereas for
beam and plate structures, the complexity of the modes shapes generally increases
for increasing eigenfrequencies, this appears not to be the case here.
Figure 3.3: Eigenfrequencies of the harmonic nite element model of the silo as a
function of m and n.
The dimensionless strain energy S of a freely supported cylinder is dened as
(Arnold and Warburton, 1949):
S =
4S(1
2
)
Eh[u
r
[
2
(3.114)
Figure 3.4 shows the dimensionless strain energy for a cylinder with a height
h = 25 m and a constant shell thickness t = 7 mm as a function of n for m = 2
and m = 3. The total strain energy associated with a mode shape is equal to
the sum of the bending and the stretching energy. The bending energy is almost
independent on m and strongly increases with n. The stretching energy increases
with m and strongly decreases with n. Consequently, the total strain energy has
a minimum as a function of n, which shifts to higher values of n for increasing
m. The order of the eigenfrequencies of the silo (gure 3.3) is related to the total
88 SHELL STRUCTURES
strain energy associated with a mode shape. The dotted line in gure 3.3 shows the
eigenfrequencies for such a tube, representing the silo without the cones. As only
the bending energy is important for a tube, this curve is a good approximation
for the eigenfrequencies of the silo for high values of n, for which the stretching
energy is small. This approximation is best at low values of m since the stretching
strain energy increases with m.
Figure 3.4: Dimensionless total strain energy (solid line) and dimensionless strain
energy due to bending (dashed-dotted line) and stretching (dashed
line) for m = 2 () and m = 3 () for a freely supported cylinder
with a height h = 25 m and a constant thickness t = 7 mm.
3.3.3 The experimental setup
Measurements (Dooms et al., 2003) have been performed on the silo on the corner
of the group where the largest amplitudes were observed during the October 2002
storm. Two lling pipes are attached to this silo in discrete points along the height
(gure 1.3).
In the case of wind excitation, the eigenmodes below 15 Hz have the highest
contribution to the structural response. The results of the harmonic nite element
calculation (gure 3.3) indicate that these eigenmodes incorporate maximum seven
circumferential waves (n = 8) and a single axial wave (m = 3).
Ten accelerometers are glued to the silo wall to measure the radial acceleration.
Eight accelerometers are placed along a horizontal arc of 7/24 radians at a height
of 15 m (gure 3.5); the angle between two accelerometers is equal to /24 radians,
corresponding to 6 accelerometers per wavelength for a circumferential wave with
n = 8. Two additional accelerometers are placed on a vertical line at a height of
2 m and 10 m (gure 3.5) in order to identify the number m.
Figure 3.5: Location of the accelerometers on the silo wall.
The accelerations are sampled at f
max
= 100 Hz and 18432 time steps are recorded,
resulting in a measurement time of about three minutes.
90 SHELL STRUCTURES
Figure 3.6 presents the power spectral density of the radial acceleration measured
at the point HL09. The spectrum clearly shows peaks around 4 Hz and 5.3 Hz.
0 5 10 15 20
0
0.05
0.1
0.15
0.2
Frequency [Hz]
P
o
w
e
r

s
p
e
c
t
r
u
m

[
(
m
/
s
/
H
z
)
]
Figure 3.6: Power spectral density of the radial acceleration at the point HL09.
The stochastic subspace identication method (Van Overschee and De Moor, 1996;
Peeters and De Roeck, 1999) is used to estimate the modal parameters of a
structure based on output-only time data. The unknown excitation is assumed
to be a realization of a stochastic process (white noise). From the measured
accelerations, stochastic state space models are identied of dierent order N,
ranging from 2 to 350 in steps of 2. N/2 modal parameters (eigenfrequencies,
complex mode shape vectors and modal damping ratios) are extracted from a
model of order N. If similar modal parameters are obtained with increasing model
order, a physical eigenmode is identied. Sixty eigenmodes are identied below
20 Hz. The modal damping ratios vary between 0.07 % and 1.32 %, which are
low values as expected for a welded aluminium structure. Figure 3.7 shows a top
view and an isometric view of three identied eigenmodes. Whereas the mode
shapes 3 and 53 are almost entirely real and correspond to standing waves along
the circumferential direction, mode 4 has an imaginary part with a comparable
magnitude as the real part, resulting in a propagating wave. The latter is expected
only for axisymmetric structures, while the axisymmetry is slightly disturbed here
by the lling pipes along the height of the silo. A perfect axisymmetric structure
has pairs of eigenmodes corresponding to identical eigenfrequencies, reecting that
these mode shapes are indenite and propagate in the circumferential direction.
The radial modal displacement u
rmn
(r, , z) at the eigenfrequency
mn
is equal
to u
rmn
(r, z)[cos(n) + i sin(n)], which is represented by a circle with radius
u
rmn
(r, z) and center at the origin of the complex plane. Figure 3.8 depicts the
displacements of mode 4 in the points HL6 to HL13 (corresponding to dierent
values of ) in the complex plane, resulting in an almost circular shape.
8
15 7
9
16
6
10
11 12
13
6
15 7
16
8
9
10
11
12
13
7
6
16
8
15
9
11
10
12
13
a. Mode 3
b. Mode 4
c. Mode 53
f = 4.00 Hz f = 4.01 Hz f = 17.95 Hz
= 0.77 % = 0.81 % = 0.10 %
Figure 3.7: Top and three-dimensional view of the real (solid line) and imaginary
(dashed line) part of the identied eigenmodes 3, 4 and 53 of the silo.
The dotted line shows the undeformed shape.
1 0.5 0 0.5 1
1
0.5
0
0.5
1
Real part
I
m
a
g
i
n
a
r
y

p
a
r
t
Figure 3.8: The radial displacements of mode 4 in the points HL6 to HL13 in the
complex plane. The displacements are scaled to unity in the point
HL6.
92 SHELL STRUCTURES
Radial modal displacements u
rmn
(r, , z) at a height z are written as a linear
combination of the radial displacements corresponding to the symmetric and
antisymmetric eigenmodes at this frequency:
u
rmn
(r, , z) = u
s
rmn
(r, z) cos(n) + u
a
rmn
(r, z) sin(n) (3.115)
A least squares analysis is performed to determine the values of the parameters
and and the circumferential wave number n that give the best correspondence
with the measured mode shapes at a height z = 15 m. As no analytical expression
is available for the mode shape as a function of the height z, the number of half
wavelengths m is estimated from the experimental results. Figure 3.9 shows the
experimental eigenfrequencies of the silo as a function of m and n. Comparison
of these results with the predicted eigenfrequencies (gure 3.3) reveals that the
harmonic nite element model overestimates the lowest eigenfrequencies. This
dierence cannot be explained by a variation of the Youngs modulus within
reasonable bounds. The eigenfrequencies and mode shapes of a circular cylindrical
shell structure are very sensitive to the boundary conditions imposed on the axial
displacements u
z
, while the inuence of the boundary condition for the rotation
is almost negligible (Forsberg, 1964; Koga, 1988; Leissa, 1993). The harmonic nite
element model imposes zero axial displacements around the whole circumference
at the bottom of the silo. In reality, each silo is only bolted to the octagonal beam
in four points around the circumference.
Figure 3.9: Measured eigenfrequencies of the silo as a function of m and n.
The time history of the radial acceleration in each point is decomposed into 68
modal contributions by a stochastic state space model of order N = 136 (Peeters
and De Roeck, 1999). Most of these modes are physical, although spurious modes
may exist, which are identied by a high damping ratio. The modal contributions
are ordered according to decreasing Root Mean Square (RMS) values of their time
history. Figure 3.10 shows the time history of the radial acceleration at the point
VL15, the nine modal contributions with the highest RMS value and the error,
which is dened as the dierence between the measured response and the sum of
68 modal contributions. In order to determine the importance of an eigenmode
in the global response of the silo, the average of the RMS values of the modal
contributions in all channels is calculated. The dominant eigenfrequencies are
identied as 3.94 Hz, 4.00 Hz and 4.49 Hz. These eigenfrequencies may slightly
dier from the previously identied eigenfrequencies, as the latter are selected at
dierent orders N, while the former are all selected at order N = 136.
3.3.4 Three-dimensional nite element model of the silo
In order to reduce the dierence between the measured and computed
eigenfrequencies and mode shapes, a realistic modelling of the four bolted
connections is required, necessitating a three-dimensional nite element model
of the silo.
A rst model exploits the symmetric geometry to reduce the model to one fourth
of the silo, and will be referred to as the quarter three-dimensional nite element
model. The silo is modelled with the 8-node quadrilateral SHELL93 element
(Ansys, 2005b). Four symmetry planes are dened by the vertical axis and the
position of the bolted connections, where all degrees of freedom are constrained.
The nite element mesh is similar to the three-dimensional nite element mesh of
the complete silo that will be introduced later in gure 3.14a. For the eight central
aluminium sheets (with a height of 2.5 m each), 15 shell elements are used along
an arc in the circumferential direction with an angle /2 and 16 elements are used
along the vertical direction. In the zones near the lower and upper edges of the silo
and on both cones, smaller nite elements are used as to fulll the requirement
dened in equation (3.113). The total number of shell elements in the quarter
three-dimensional nite element model is equal to 4747.
Convergence is checked by computing the eigenfrequencies of alternative nite
element models where the mesh has been uniformly rened with a factor of
2, 2
and 2
2. The approximated relative error

j
on an eigenfrequency
j
is dened
as:
j
=
[
j
res
j
[
res
j
(3.116)
94 SHELL STRUCTURES
Figure 3.10: Time history of the radial acceleration at the point VL15 (top), the
nine modal contributions with the highest RMS value and the error
(bottom).
where the reference frequency
res
j
is computed with the nest nite element mesh
(mesh renement factor 2
2).
Figure 3.11 shows the approximated relative error on the predicted eigenfrequen-
cies as a function of the mesh renement factor. For the original mesh (mesh
renement factor 1), all relative errors are below 10
2
and most of the values are
even below 10
3
. This is suciently accurate for the subsequent comparison with
the experimental eigenfrequencies, as the latter are also subject to measurement
and identication errors.
1 1.4 2
10
5
10
4
10
3
10
2
Refinement factor []
R
e
l
a
t
i
v
e

e
r
r
o
r

[
]
(1,4)
(2,6)
(3,10)
(1,3)
(1,5)
(4,9)
(1,4)
(1,2)
(3,10)
Figure 3.11: The approximated relative error on nine selected eigenfrequencies
predicted with the quarter three-dimensional nite element model
as a function of the mesh renement factor.
Four dierent models with symmetric and antisymmetric boundary conditions
on the symmetry planes are analysed. The two models with symmetrical and
antisymmetrical boundary conditions (SA and AS) result in similar mode shapes
with identical eigenfrequencies, corresponding to odd values of the circumferential
wave number n. The eigenmodes of the models with symmetrical-symmetrical
(SS) and antisymmetrical-antisymmetrical (AA) boundary conditions correspond
to even values of n and have dierent mode shapes and eigenfrequencies due to the
dierent relative position of the bolted connections and the ovalling mode shapes
(gures 3.12a and 3.12b). Although in these models the four bolted connections at
the bottom of the silo violate axisymmetry, still two eigenmodes can be associated
with every couple (m, n), as for the harmonic nite element model.
The disadvantage of the quarter three-dimensional nite element model is that it
xes the relative position of the four bolted connections and the ovalling mode
shapes. The eigenmodes (1,6), computed with the quarter three-dimensional
nite element models with SS (gures 3.12a) and AA boundary conditions (gure
3.12b), have xed relative positions between the ovalling mode shapes and the
bolted connections (mode shapes are shown of the part of the cylindrical shell at
96 SHELL STRUCTURES
a height between 15 and 17.5 m above the bottom cone). A three-dimensional
nite element model of the complete silo with constrained degrees of freedom at
the four bolted connections enables to realize other relative positions between the
ovalling mode shapes and the bolted connections (gure 3.12c). As in the quarter
three-dimensional model, two eigenmodes are associated with every couple (m, n).
a. SS model b. AA model c. 3D model
Figure 3.12: Top view of the eigenmode (1, 6) of the silo at a height of 15 m
computed with (a,b) the quarter three-dimensional nite element
model with SS and AA boundary conditions and (c) the full three-
dimensional nite element model.
Figure 3.13a shows a top and a lateral view of the eigenmode (1,5) computed with
the quarter three-dimensional nite element model with AS boundary conditions,
xing the relative position of the global bending modes and the ovalling modes.
a. SA model b. 3D model
Figure 3.13: Top and lateral view of the eigenmode (1, 5) of the silo computed
with (a) the quarter three-dimensional nite element model with SA
boundary conditions and (b) the full three-dimensional nite element
model.
The three-dimensional nite element model of the complete silo allows for any
combination of bending and ovalling (gure 3.13b) and is therefore preferred; it
is obtained by mirroring the quarter model on two symmetry planes through the
bolted connections (gure 3.14a).
Figures 3.14b up to 3.14f show a top view and a three-dimensional view of ve
calculated mode shapes of the silo. The top view corresponds to the displacements
of the part of the cylindrical shell at a height between 15 and 17.5 m above the
bottom cone and reveals that mode shape (1,5) and, to a lesser extent, mode shape
(1,3) are a combination of global bending of the silo and ovalling.
a.Model
b.(1, 3) c.(1, 4) d.(1, 5) e.(1, 6) f.(1, 2)
3.93 Hz 3.93 Hz 5.25 Hz 7.37 Hz 7.75 Hz
Figure 3.14: Three-dimensional nite element model of the silo and selected
eigenmodes.
98 SHELL STRUCTURES
Figure 3.15 summarizes the eigenfrequencies calculated with the full three-
dimensional nite element model. Two mode shapes and eigenfrequencies are
associated with every couple (m, n). Table 3.2 compares the eigenfrequencies
computed with the harmonic nite element model, the quarter three-dimensional
nite element model and the full three-dimensional nite element model
with the experimental eigenfrequencies. The eigenfrequencies of the quarter
model are computed with the dierent sets of symmetric and antisymmetric
boundary conditions. For the full three-dimensional nite element model, the
eigenfrequencies of the two dierent eigenmodes that can be associated with every
couple (m, n) are given. The correct representation of the bolted connections
strongly inuences the eigenfrequencies of eigenmodes with a low circumferential
wave number n, more specically those eigenmodes where the contribution of
the stretching energy to the total strain energy is large. Dierences between the
quarter model and the full three-dimensional nite element model are apparent
for the eigenfrequencies of the modes (1,3), (1,5) and (2,5), whereas all other
eigenmodes have the same eigenfrequencies. The eigenfrequencies of the full
three-dimensional model agree better with the experimental eigenfrequencies. The
eigenfrequencies of modes (1,4) and (1,3) are almost identical. Discrepancies can
be explained by the presence of the lling pipes, the assumed boundary conditions,
and the uncertainty on the modal parameters derived from the dynamic system
identication procedure.
Figure 3.15: Eigenfrequencies of the three-dimensional nite element model of the
silo as a function of m and n.
(m, n) Harm. Quarter 3D Full 3D Experimental
AA SS SA AS
(1,2) 10.8 7.75 8.48 7.75 8.48 7.87
(1,3) 5.99 4.19 4.19 3.93 3.93 3.91 4.49
(1,4) 4.65 3.93 4.04 3.93 4.04 3.91 4.00 4.08 4.49
(1,5) 5.52 5.42 5.42 5.25 5.25 5.31 5.34
(1,6) 7.38 7.37 7.37 7.37 7.37 7.07 7.23 7.49
(1,7) 9.73 9.72 9.72 9.72 9.72 9.84 10.7
(1,8) 12.5 12.5 12.5 12.5 12.5 12.8
(1,9) 15.8 15.7 15.7 15.7 15.7
(1,10) 19.4 19.4 19.4 19.4 19.4 19.8
(2,3) 16.5 13.9 13.9 13.9 13.9
(2,4) 10.7 8.71 8.94 8.71 8.94
(2,5) 8.78 8.08 8.08 8.01 8.01
(2,6) 9.47 9.29 9.39 9.29 9.39 9.05 9.25
(2,7) 11.5 11.5 11.5 11.5 11.5 11.1 11.4
(2,8) 14.1 14.1 14.1 14.1 14.1 14.2
(2,9) 17.1 17.1 17.1 17.1 17.1
(2,10) 20.5 20.5 20.5 20.5 20.5
(3,4) 20.1 17.5 17.7 17.5 17.7
(3,5) 14.7 13.0 13.0 13.0 13.0
(3,6) 12.9 12.0 12.2 12.0 12.2 11.9 12.0
(3,7) 13.8 13.6 13.6 13.6 13.6 13.6
(3,8) 16.1 16.1 16.1 16.1 16.1 15.9
(3,9) 19.0 19.0 19.0 19.0 19.0 19.0 19.1
(4,5) 22.8 20.8 20.8 20.8 20.8
(4,6) 18.2 16.8 17.0 16.8 17.0
(4,7) 17.0 16.3 16.3 16.3 16.3 16.3 16.5
(4,8) 18.4 18.3 18.3 18.3 18.3
Table 3.2: Comparison of the eigenfrequencies computed with the harmonic nite
element model, the quarter three-dimensional nite element model and
the full three-dimensional nite element model with the experimental
eigenfrequencies.
3.3.5 Non-linear behaviour of the silos
In this section, it is investigated if geometrical non-linear behaviour should be
taken into account during the ovalling of the silos. Based on a movie of the ovalling
of the silos taken during the storm in October 2002, the amplitude of the ovalling
oscillations is estimated to be maximally of an order of 0.1 m. Therefore, the
results of a static linear computation under a pressure load which yields a maximal
displacement of 0.1 m, is compared to the results of a non-linear computation with
the same loading.
In order to check the parameters and element types employed during this
geometrical non-linear computation, rst the well-known deformation of a
cantilever beam of length L under a tip moment M (Cook et al., 2002) is computed.
100 SHELL STRUCTURES
The analytical solution for the vertical tip displacement u
y
is:
u
y
=
EI
M
_
1 cos
ML
EI
_
(3.117)
If the tip moment M =
EI
L
, the deformed shape of the beam is half a circle (gure
3.16) and the horizontal and vertical tip displacement are respectively:
u
x
= L u
y
=
2L
(3.118)
The cantilever beam is modelled using 1 element along the width and 15 shell
elements along the length. Two dierent shell elements (Ansys, 2005b) were used
to compute the deformations.
The 8-node quadrilateral SHELL93 and the 4-node quadrilateral SHELL181
element are based on the Mindlin-Reissner plate theory, which includes transverse
shear deformations. An updated Lagrangian formulation which uses logarithmic
strains and Cauchy true stresses, is adopted. The formulation is suited to account
for large strains. In the case of the SHELL181 element, the follower load stiness
matrix is included in the tangent stiness matrix, while it is not included for the
SHELL93 element.
The geometric stiness matrix K
was not included in the non-linear computations.

The tolerance for the convergence check is set to
r
= 0.5% for forces and moments
and to
u
= 5% for the displacements.
Figure 3.16 shows the deformed shape. Table 3.3 lists the horizontal and vertical
tip displacement, number of load increments and equilibrium iterations as a
function of the element type. The displacements obtained with SHELL181 are
closer to the analytical solution, but a few more equilibrium iterations were
executed. For the geometrical non-linear computation of the deformation of the
silo, the SHELL181 element will be used.
Figure 3.16: Deformed shape of a cantilever beam of length L under a tip moment
M = EI/L.
u
x
u
y
increments iterations
SHELL93 0.9968L 0.6412L 10 78
SHELL181 1.0L 0.6378L 10 85
Table 3.3: Horizontal and vertical tip displacement, number of load increments
and equilibrium iterations as a function of the element type.
The comparison between a linear and a non-linear computation focusses on ovalling
oscillations in eigenmode (1,4). Therefore, a pressure distribution is imposed which
varies as p sin(4) around the circumference (gure 3.17). Within each element only
the linear part of the variation is taken into account. The silos are simplied to a
cylindrical shell with a diameter of 5.5 m, a height of 25 m and a constant thickness
of 7 mm. The cylinder is made of aluminium. The magnitude p of the loading is
equal to 1004.8 N/m
2
which yields in the case of a static linear computation a
maximal radial displacement of 0.1 m.
Figure 3.17: Imposed pressure distribution around the circumference of the silo.
The nite element model exploits the symmetry of the geometry and the loading to
reduce the model to one eighth of the cylinder in the circumferential direction and
half a cylinder along the height. Using the antisymmetry of the loading as well,
the model could be reduced to one sixteenth of the cylinder in the circumferential
direction for the linear computation. However for the geometrical non-linear
computation, antisymmetric boundary conditions should not be applied as large
structural deformations out of the plane of antisymmetry might occur and disturb
the validity of the symmetry of the geometry. At the bottom of the cylinder a
support is supposed which constrains around the circumference the radial and
circumferential displacements and the rotations around the z-axis. Figure 3.18a
shows the nite element mesh and the boundary conditions of the model. 8 shell
elements are used along an arc with an angle /4 in the circumferential direction
and 10 elements are used along the vertical direction.
102 SHELL STRUCTURES
(a) (b)
Figure 3.18: (a) Finite element mesh and boundary conditions and (b)
deformations magnied by factor of 3.
The geometric stiness matrix K
was not included in the non-linear computations.

The loading was applied in 12 increments and totally 24 equilibrium iterations were
performed.
Figure 3.18b shows the deformations of the cylinder magnied by factor of 3. In
the left symmetry plane the structure displaces outward. The curvature increases
and the structure stiens. The maximal radial displacement decreases from 0.1 to
0.093 m by including the non-linear behaviour. In the right symmetry plane the
structure displaces inward. The curvature decreases and the structure softens. The
magnitude of the radial displacement increases from 0.1 to 0.125 m by including
the non-linear behaviour. The inuence of geometrical non-linear behaviour is not
very large, nor negligible. The geometrical non-linear behaviour will be included
in further computations.
CONCLUSION 103
3.4 Conclusion
This chapter treats the computation of the linear and the geometrical non-linear
response of shell structures. First, the virtual work formulation which leads to
the nite element method is briey reviewed. The conjugate stresses and strains
used for the computation of geometrical non-linear structures are described. The
Newton-Raphson technique solves the non-linear system of equations. In order to
compute the transient geometrical non-linear response of shell structures direct
time integration is applied.
As an example, the cylindrical shell structure of the silos in the port of Antwerp
has been studied. The eigenmodes and eigenfrequencies of a single silo have
been rst computed with a harmonic nite element model in order to enable the
design of a proper experimental set-up. Measurements of the radial acceleration
under ambient wind loading have been carried out at 10 points on the silo.
Modal parameters (eigenfrequencies, mode shapes and modal damping ratios) were
extracted from the output-only data using the stochastic subspace identication
technique. The eigenmode with the lowest eigenfrequency at 3.94 Hz (m = 1, n = 3
or 4) has the largest contribution to the measured response. All modes have a low
damping ratio around 1%. A three-dimensional nite element model of the silo
using shell elements enables to model correctly the connections between the silo
and the supporting structure. The eigenmodes with three and four waves along
the circumference (n = 4) and half a wave along the height (m = 1) are both
situated at 3.93 Hz. Finally, the inuence of geometrical non-linear behaviour
on deformations of the silos of order of magnitude of 0.1 m is not negligible:
the magnitude of the displacements increases where the curvature decreases and
decreases where the curvature increases. The geometrical non-linear behaviour
will be included in further computations.
Chapter 4
Computation of turbulent wind
ows
4.1 Introduction
In this chapter, some aspects of the computation of turbulent wind ows around
civil engineering structures which are relevant to the work in this thesis, will be
treated. First current state-of-the-art methods for the simulation of turbulent
ows are reviewed as a function of their performance and their computational
requirements.
In order to evaluate the feasibility of numerical ow simulations and to clarify the
inuence of turbulence models, wall treatment, mesh renement and unsteadiness,
rst an example for which the solution is known is computed. The ow around
a single cylinder is an ideal test case because it combines several types of
ow structures as separation, recirculation, stagnation and streamline curvature.
Moreover it is a simplication of the wind ow around a single silo. The results are
compared with the pressure coecients from Eurocode 1 (BIN, 1995), experimental
data (Zdravkovich, 1997) and the few numerical results reported in the literature.
In the port of Antwerp the group of 8 by 5 silos (gure 1.3) has very small gaps
of 30 cm between two neighbouring silos. As this conguration is expected to
have a large inuence on the wind pressure distribution around the silos and
secondarily on the occurrence of ovalling oscillations, a numerical study of the
wind ow should provide a more realistic estimation of the pressure coecients
and forces. The inuence of surrounding structures on the ow pattern around
a structure is called interference eect. The ow around the group of silos is
105
106 COMPUTATION OF TURBULENT WIND FLOWS
simplied to the two-dimensional ow around a group of 8 by 5 cylinders. Before
the extensive computation for the group of 8 by 5 cylinders, the less comprehensive
case of a group of 2 by 2 cylinders is modelled.
4.2 Computational approaches to turbulent ows
The simulation of turbulent ows (Versteeg and Malalasekera, 1995; Wilcox, 1998;
Pope, 2000) is still a challenging task. Turbulent ows are characterized by a
velocity eld that varies randomly both in space and time. They are inherently
three-dimensional and time-dependent. A large range of time and length scales are
present in a turbulent ow. The incompressible Navier-Stokes equations (2.97)-
(2.99) are applicable for isothermal turbulent ows of Newtonian uids and are
recalled here in an Eulerian description:
v = 0 (4.1)
v
t
+ (v v) +
1
p =
2
v +b (4.2)
The challenge is to accurately compute the relevant statistics of such ows.
Turbulent kinetic energy is produced at the largest scales of motion, which are as
large as a characteristic dimension of the ow. These largest turbulent motions
(eddies) are directly aected by the geometry and the boundary conditions and
therefore anisotropic. As direct eects of viscosity are very small for the large
eddies, they are unstable and break up transferring their energy to successively
smaller eddies. This energy cascade continues up to scales that are small enough
to dissipate the energy by viscous forces. These smallest eddies are statistically
isotropic. The rate at which energy is transferred from larger to smaller scales is
determined at the largest scales and nearly equal to the dissipation rate . The
smallest length, velocity and time scales, which are called the Kolmogorov scales,
are determined by the kinematic viscosity and the dissipation rate . The ratio
of the smallest to the largest length scales is equal to Re
3
4
and for the time scales
to Re
1
2
(Pope, 2000). The turbulence energy is predominantly contained in the
larger anisotropic eddies.
4.2.1 Direct numerical simulation
In order to numerically simulate turbulent ows, a number of approaches exists.
Conceptually, direct numerical simulation (DNS) is the simplest approach. The
Navier-Stokes equations (4.1)-(4.2) are solved, resolving all length and time scales.
For a three-dimensional time-dependent computation, the number of grid points
COMPUTATIONAL APPROACHES TO TURBULENT FLOWS 107
has generally to be proportional to Re
9
4
and the number of time steps proportional
to Re
3
4
. As the computational cost increases drastically with the Reynolds number,
DNS is limited to ows of low and moderate Reynolds numbers. Therefore,
DNS is mainly a research tool to perform numerical experiments that enables the
improvement of turbulence models and increases insight into turbulence physics
(Moin and Mahesh, 1998).
4.2.2 Reynolds averaged Navier-Stokes simulation
In Reynolds averaged Navier-Stokes (RANS) simulation the velocity v is
decomposed in the mean velocity v and the uctuating velocity v
:
v = v +v
(4.3)
Analogously, the pressure p is decomposed the mean pressure p and the uctuating
pressure p
:
p = p + p
(4.4)
For statistically stationary ows the mean can be thought of as the time
average over a long period, while for statistically non-stationary ows the mean
corresponds to an ensemble average: the unsteady Reynolds averaged Navier-
Stokes (URANS) simulation produces the average over a large number of identical
experiments. Taking the mean of the Navier-Stokes equations (4.1)-(4.2) and
substituting (4.3) in the convective term, the Reynolds equations for the mean
velocity are obtained:
v = 0 (4.5)
v
t
+ (v v) + v
+
1
p =
2
v +b (4.6)
Because of the non-linear convective term in the Navier-Stokes equations, the
divergence of the velocity covariances v
appears in the momentum equations.

The velocity covariances are called the Reynolds stresses
1
, although they stem
from the convective term. These four equations contain, in addition to the three
components of the mean velocity and the mean pressure, the six Reynolds stresses.
Without additional equations for these Reynolds stresses, this set of equations
cannot be solved. The transport equations for the Reynolds stresses can be derived,
but a lot of additional unknowns are present in those equations. This is the
closure problem: the number of equations never balances the number of unknowns.
Therefore, some of the unknowns have to be modelled. In turbulent viscosity
1
Although the apparent stress is equal to v
, v
are conventionally called the

Reynolds stresses.
models the Reynolds stresses in equation (4.6) are determined by a turbulence
model. In Reynolds stress models (RSM) the unknown terms in the transport
equations for the Reynolds stresses are modelled. Using RANS all the turbulent
uctuations are modelled and consequently all the turbulent kinetic energy as well.
A huge number of turbulence models exist. In these models many constants appear,
which are calibrated by means of experimental results for simple shear ows.
As all turbulent wind ows are bounded by walls, the grid requirements near walls
in RANS are generally very important. The properties of a straight fully developed
boundary layer ow, as obtained from experiments and DNS, are very illustrative.
The total shear stress is the sum of the Reynolds and the viscous stresses. At the
wall the Reynolds shear stress is zero and the wall shear stress
w
is entirely due to
the viscous contribution. In the region close to the wall the viscous contribution
dominates, while further away from the wall the Reynolds stress is dominant. In
the former region, a viscous velocity scale and length scale are dened from the
wall shear stress
w
and the viscosity : the friction velocity v
=
_
w
/ and the
viscous length scale
v
=
_
/
w
= /v
.
Dividing the distance from the wall y by the viscous length scale
v
yields a
dimensionless expression for the distance y
+
= y/
v
= v
y/. This dimensionless

distance y
+
can be considered as a local Reynolds number, which gives the relative
importance of viscous and turbulent processes. Analogously, a dimensionless
velocity v
+
= v/v
is obtained.
Close to the wall, an inner layer exists where the dimensionless mean velocity
prole v
+
only depends on the dimensionless wall distance y
+
. The viscous
contribution to the shear stress decreases from 100 % at the wall to 50% at y
+
12
and less than 10% at y
+
= 50 (Pope, 2000). In the viscous sublayer y
+
< 5, where
viscosity eects dominate, v
+
equals y
+
. In the outer layer y
+
> 50, the eects
of the viscosity on the mean ow are negligible. At high Reynolds numbers, a
layer exists where the mean velocity v
+
is not inuenced by the viscosity and only
depends on the dimensionless wall distance y
+
:
v
+
=
1
ln y
+
+ B =
1
ln
v
+ B (4.7)
This is the logarithmic law of the wall (Pope, 2000) for smooth surfaces. The von
Karman constant is equal to 0.41 and B is equal to 5.2. The region of validity of
the logarithmic law can be extended from y
+
> 50 to y
+
> 30. In the logarithmic
law region, production and dissipation of turbulence are almost in equilibrium.
COMPUTATIONAL APPROACHES TO TURBULENT FLOWS 109
For rough surfaces the logarithmic law can be modied in order to account for the
eects of sand-grain roughness k
s
:
v
+
=
1
ln
v
y
(1 + 0.3k
s
)
+ B (4.8)
This roughness modication is inconsistent with the requirements for an
atmospheric boundary layer ow and might compromise the accuracy of these
simulations (Blocken et al., 2007).
For the computations there are two possibilities. In a low Reynolds formulation,
the RANS equations can be solved up to the wall. Ideally, ve nodes are located in
the viscous sublayer (y
+
1 for the node next to the wall). The turbulence models
for the Reynolds stress have to be modied in order to be valid in the region where
viscosity has direct eects. In a high Reynolds formulation, the nodes next to the
wall are placed in the logarithmic law region. So-called wall functions apply the
logarithmic law as a boundary condition on the nodes next to the wall. The near-
wall mesh is much coarser in comparison with the low Reynolds formulation. These
wall functions are not suitable if the logarithmic layer does not exist, for instance
for separated or impinging ows, or in case of curved streamlines. In these cases,
there is no equilibrium between production and dissipation of turbulence and the
solution might be dependent on the location of the node next to the wall.
4.2.3 Large eddy simulation
In large eddy simulation (LES) only the larger scale turbulent motions, which
depend on the ow geometry and are anisotropic, are directly computed. The
inuence of the smaller scale motions, which are isotropic and universal, can be
taken into account by a simple turbulence model. Therefore, a velocity eld v
where smaller scale motions are ltered out, is dened:
v(x) =
_
G(r, x)v(x r)dr (4.9)

where G(r, x) is a lter with
_
G(r, x)dr = 1 and the integration is performed
over the entire uid domain. The velocity eld v is decomposed into the ltered
component v and a residual component v
. If the specied lter width is equal

to the size of the smallest energy containing eddies, about 80 % of the turbulent
kinetic energy is resolved. The number of grid points has to be just large enough
to compute accurately this ltered velocity eld. The computational cost is much
smaller than for a DNS. Filtering the Navier-Stokes equations (4.1)-(4.2) with a
spatially uniform lter G(r) yields:
v = 0 (4.10)
v
t
+
_
v v +
v v
v +
_
+
1
p =
2
v +b (4.11)
To close the set of equations, a residual stress tensor
R
is dened as the dierence
between the term in brackets, which is equal to
v v, and the product of the

ltered velocities v v. The momentum equation becomes:
v
t
+ ( v v) +
R
+
1
p =
2
v +b (4.12)
This residual stress tensor
R
has to be modelled taking into account the lter
width and lter type. The ltered Navier-Stokes equations (4.10)-(4.12) are solved
for the ltered velocity v. The lter characteristics only enter indirectly in these
equations through the model for the residual stress tensor.
LES is especially suited for ows with large scale unsteadiness, like ows over blu
bodies with unsteady separation and vortex shedding. Unfortunately, close to a
solid wall, all eddies, including the energy containing eddies, are very small and
their size is dependent on the Reynolds number. As these small energy containing
eddies should be resolved, very ne grids and small time steps are required near
the wall. To reduce the computational cost two options exist. As in the wall
functions approach for RANS, appropriate boundary conditions can be applied at
the rst grid point away from the wall. This option is not satisfactory if there is
no equilibrium between production and dissipation, as in the case of separating
ows. The second possibility is to solve RANS boundary layer equations on an
embedded grid, which produce boundary conditions for the LES computation.
Due to the poor performance of RANS for massively separated ows and the
high computational cost of wall-resolved LES, hybrid RANS/LES models are
being developed, which are a generalization of the second approach for LES near
walls. These models switch to RANS or LES behaviour depending on the grid
resolution. In the RANS regions the turbulence model controls the solution, while
in the LES region the larger eddies are resolved. In very large eddy simulations
(VLES) (Speziale, 1998), the lter width (or grid spacing) is compared with a
turbulent length scale. If the lter width is larger than the size of the smallest
energy containing eddies, the computation switches to RANS behaviour, while
LES behaviour is activated when the lter width is smaller. In detached eddy
simulations (DES) (Spalart et al., 1997; Travin et al., 2000), the ratio of the
wall distance and the grid spacing determines if the RANS or LES mode is used.
The model operates in RANS mode near the wall if the grid spacing normal or
parallel to the wall is large compared with the boundary-layer thickness. The
detached eddies further from the wall are simulated in LES mode. The design of
RANS TURBULENCE MODELS 111
the grid controls the behaviour of these hybrid models. The computational costs
are considerably lower than for a wall-resolved LES.
The application of hybrid RANS/LES methods is very promising as they are able to
capture the large scale turbulent structures in the wake of blu bodies at a more
aordable computational cost. However, for the use in coupled uid-structure
problems, this computational cost today is still very high and in this thesis only
RANS models will be used. The next sections treats those models in more detail.
4.3 RANS turbulence models
4.3.1 Linear eddy viscosity models
The Reynolds stress tensor v
can be split in its isotropic and deviatoric part:

v
=
1
3
v
I +
_
v
1
3
v
I
_
(4.13)
The turbulent kinetic energy k is dened as half the trace of the Reynolds stress
tensor v
:
k =
1
2
v
(4.14)
The isotropic part of the Reynolds stress tensor is equal to 2/3kI. The turbulent
viscosity hypothesis by Boussinesq (Pope, 2000) supposes that the deviatoric part
of the Reynolds stress tensor is proportional to the mean strain rate = 1/2(v+
(v)
T
):
v
=
2
3
kI 2
t
=
2
3
kI
t
(v + (v)
T
) (4.15)
Equation (4.15) is analogous to the expression for the stress in a Newtonian uid
(2.94). The main assumption is that the mean velocity gradients determine the
deviatoric part of the Reynolds stress tensor. As the molecular motion, responsible
for the viscous stresses, rapidly adjusts to the mean straining, equation (2.94)
is appropriate and the viscous stresses are mainly isotropic. Turbulent motion
often does not rapidly adapt to the mean straining. The history of straining
is important and the Reynolds stresses are largely anisotropic. The turbulent
viscosity hypothesis is not generally applicable. As for simple turbulent shear
ows the mean velocity gradients change slowly and are a good measure for the
straining history, the hypothesis is in these cases reasonable. The production and
dissipation of turbulence are approximately equal.
More specically, in equation (4.15) a linear relation is assumed between the
deviatoric part and the mean strain rate. The scalar
t
is the turbulent viscosity.
In reality the principal axes of the deviatoric part of the Reynolds stress tensor
and the mean strain rate tensor are not aligned. For simple shear ows, only
one component of the Reynolds stress tensor is of interest and the hypothesis is
rather a denition of the turbulent viscosity
t
. For swirling ows or ows with
substantial streamline curvature, the linear relation is not valid.
Substituting the turbulent viscosity hypothesis (4.15) in the momentum equation
(4.6) yields:
v
t
+ (v v) +
1
(p +
2
3
k) = ( +
t
)
2
v +b (4.16)
This equation is similar to the Navier-Stokes momentum equation (2.99), but the
velocity, viscosity and pressure are replaced by the mean velocity, the eective
viscosity and the modied mean pressure, respectively.
The turbulent viscosity
t
still has to be determined. At high Reynolds numbers
and far away from the wall, it scales with a length scale and a velocity scale of the
ow. For algebraic models as the mixing-length model, no additional transport
equations need to be solved, but they are incomplete as an appropriate mixing
length should be specied which depends on the ow geometry. One- and two-
equation models require the solution of respectively one or two additional transport
equations to determine the turbulent viscosity
t
. The Spalart-Allmaras model
(Spalart and Allmaras, 1994) is a complete one-equation model which performs
well for aerodynamic ows with separation.
The standard k- model
The standard k- model (Jones and Launder, 1972) is a two-equation model. The
turbulent viscosity
t
is expressed as a function of the turbulent kinetic energy k
and the turbulent energy dissipation rate :
t
= C
k
2
(4.17)
Two additional transport equations are solved for the turbulent kinetic energy k
and the turbulent energy dissipation rate :
k
t
+ (kv) =
__
+

t
k
_
k
_
+T (4.18)
t
+ (v) =
__
+

t
_
+ C
1
k
T C
2
2
k
(4.19)
where the turbulence production T is, using the turbulent viscosity hypothesis
(4.15), equal to:
T =
t
v : (v + (v)
T
) (4.20)
The time derivative, the convective term and the production term of the transport
equation (4.18) stem from the exact equation for the turbulent kinetic energy k,
while the rst term on the right hand side models the diusion terms with the
gradient-diusion hypothesis. In the transport equation (4.19) for the turbulent
energy dissipation rate only the left hand side corresponds to the exact equation.
The right hand side is entirely modelled and mainly follows from dimensional
analysis. The standard k- model includes ve constants C
,
k
,
, C
1
and C
2
which are calibrated from a number of turbulent ows e.g. decaying homogeneous
turbulence, a homogeneous shear ow and a turbulent boundary layer. For a
particular ow the accuracy of the model can be improved by tuning the constants.
The standard k- model yields acceptable results for simple turbulent shear
ows, but has considerable shortcomings for more complex ows with streamline
curvature and boundary layer separation under adverse pressure gradients.
The k- model
In many two-equation models, the turbulent kinetic energy k is taken as one of
the variables. In the k- model (Wilcox, 1988) the turbulent frequency is the
second variable. The turbulent viscosity
t
is expressed as:
t
=
k
(4.21)
Comparison of equations (4.17) and (4.21) yields the relation between the turbulent
frequency and the turbulent energy dissipation rate :
=
1
C
k
(4.22)
The transport equations for the turbulent kinetic energy k and the turbulent
frequency are:
k
t
+ (kv) =
__
+

t
k1
_
k
_
+T
k (4.23)
t
+ (v) =
__
+

t
1
_
_
+

1
t
T
1
2
(4.24)
The transport equation for the turbulent kinetic energy k is, except for the values of
the constants, identical to one of the standard k- model. The transport equation
(4.19) for the turbulent energy dissipation rate can be transformed in a transport
equation for the turbulent frequency :
t
+(v) =
__
+

t
2
_
_
+
t
T
2
2
+
2
k
_
+

t
2
_
k (4.25)
Except for values of the constants, the only dierence between equations (4.24) and
(4.25) is the appearance of the cross diusion term on the right hand side (Wilcox,
1998). This cross diusion term deteriorates the quality of the computation
of turbulent boundary layers under adverse pressure gradients. The k- model
therefore predicts the onset and the amount of separation under adverse pressure
gradients much more accurate than the standard k- model. In contrast to the
standard k- model, it does not need any damping function to be used within
the viscous sublayer. However, in free shear ows the cross diusion increases
the production of , which results in an increased dissipation of kinetic energy k.
The absence of this term in the k- model makes the model very sensitive to free
stream turbulence levels, while the standard k- model is independent from the
free stream levels.
The shear stress transport model
The shear stress transport (SST) model (Menter, 1994) combines the independence
of the free stream turbulence levels of the k- model with the robust and accurate
computation of boundary layers under adverse pressure gradients of the k- model.
Therefore, the standard k- model is transformed into a k- formulation. The
equation for the turbulent kinetic energy k is, except for the values of the constants,
identical to equation (4.23). The transformation of the transport equation for the
turbulent energy dissipation rate is given in equation (4.25).
A blending function F
1
is dened which is equal to one in the near-wall region
and zero far away from the wall. The transport equations of the k- model are
multiplied with F
1
while the transport equations of the standard k- model are
multiplied with 1 F
1
. The corresponding transport equations are added. The
resulting model uses the k- model in the boundary layer and the standard k-
model far away from the walls.
In turbulent boundary layers under adverse pressure gradients the production can
be much larger than the dissipation. In this case the turbulent viscosity
t
is
overestimated by equation (4.21). This denition is slightly modied in order to
account for the transport of turbulent shear stresses:
t
=
a
1
k
max(a
1
, F
2
)
(4.26)
where is the absolute value of the mean vorticity. F
2
is, similar to F
1
, a blending
function which is equal to one in the near-wall region and zero far away from
the wall. Far away from the walls, the original denition (4.21) of the turbulent
viscosity
t
is recovered. Close to the walls, the original denition is as well
recovered if a
1
F
2
. For a
1
< F
2
the deviatoric part of the Reynolds
stresses becomes:
v
2
3
kI =
a
1
k
F
2
(v + (v)
T
) (4.27)
In boundary layers, where mainly one velocity gradient is of importance, this
corresponds to the assumption of Bradshaw (Menter, 1994) that in a turbulent
boundary layer the Reynolds shear stress is proportional to the turbulent kinetic
energy. This modication makes the SST model very appropriate for the prediction
of the onset and the amount of separation.
Finally, in order to avoid excessive build-up of turbulence in stagnation regions, a
limiter is applied to the productions terms:
T = min(T, c
lim
) (4.28)
where the limiter c
lim
defaults to 10.
The realizable k- model
In the Reynolds stress tensor v
the normal Reynolds stresses are always

positive:
v
i
v
i
0 (4.29)
and the correlations between uctuating velocities fulll the Cauchy-Schwartz
inequality:
v
i
v
j
2
v
2
i
v
2
j
1 (4.30)
By the turbulent viscosity hypothesis (4.15) these properties of the Reynolds stress
tensor are not automatically guarantied and for instance violated in the case of
large mean strains . Equations (4.29) and (4.30) are called the realizability
conditions. In order to fulll these conditions, C
(equation (4.17)) cannot be

a constant and should be related to the mean strain rate (Shih et al., 1995a):
C
=
1
A
0
+ A
s
k
: + :
(4.31)
where is the mean rotation rate:
=
1
2
(v (v)
T
) C
r
I (4.32)
The vector denes the angular velocity of the frame of reference. The parameters
C
r
and A
0
are respectively equal to 3 and 4 and A
s
is given by:
A
s
=
6 cos , =
1
3
arccos(
6W), W =
( ) :
T
_
:
_
3
(4.33)
In the realizable k- model also a new transport equation for the turbulent energy
dissipation rate is obtained from the dynamic equation for the mean-square of
the vorticity uctuation:
t
+ (v) =
__
+

t
_
+ C
1
C
2
2
k +
(4.34)
where is:
=
2 : (4.35)
If in the last term on the right hand side of equation (4.34) k >
, as is the case
at high Reynolds numbers, the production term is the only dierence with the
equation (4.19) of the standard k- model. The values of the constants
and C
2
in equation (4.34) dier from those of the standard k- model. C
1
is determined
by:
C
1
= max
_
0.43,

+ 5
_
(4.36)
where is the ratio of a time scale of the turbulence and the mean strain:
=
k
(4.37)
4.3.2 Non-linear eddy viscosity models
The turbulent viscosity hypothesis (4.15) supposes that a linear relation exists
between the deviatoric part of the Reynolds stress tensor and the mean strain rate
tensor. The deviatoric part of the Reynolds stress tensor could more generally be
related to the mean velocity gradient, which consists of the mean strain rate tensor
and the mean rotation rate tensor (4.32). After multiplication with a turbulence
time scale k/ the normalized mean strain rate and the normalized mean rotation
rate are obtained:
=
k
(4.38)
=
k
(4.39)
The Reynolds stress tensor is now modelled as:
v
=
2
3
kI + 2kB( , ) (4.40)
where B is a general function of the normalized mean rate of strain and
normalized the mean rate of rotation . An integrity basis for all symmetric and
deviatoric tensors formed from the normalized mean strain rate and rotation
rate consists of ten combinations. Craft et al. (1996) introduced a model where
all combinations of the strain rate and the rotation rate tensor up to the third
order are included, which corresponds to six combinations:
v
=
2
3
kI + 2kc
1
+ 2kc
2
( ) + 2kc
3
(
2
1
3
tr(
2
)I) (4.41)
+ 2kc
4
(
2
1
3
tr(
2
)I) + 2kc
5
(
2

2
) (4.42)
+ 2kc
6
(
2
+
2
2
3
tr(
2
)I) (4.43)
Generally the coecients c
1
till c
6
might depend on the ve independent tensor
invariants composed of the normalized mean strain rate and rotation rate .
The cubic terms enable to capture the streamline curvature and swirl eects.
Shih et al. (1993, 1995b) only retained combinations up to the second order (c
5
=
c
6
= 0). The realizability conditions (4.29) and (4.30) were used to determine the
coecients, which nally only depend on two tensor invariants:
=
2 : (4.44)
=
2 : (4.45)
where was already dened in equation (4.37). In the earliest version (Shih et al.,
1993) the coecient c
1
for the linear term is given by:
c
1
=
A
s1
A
s2
+ + A
s3
(4.46)
where A
s1
= 2/3, A
s2
= 1.25 and A
s3
= 0.9.
The Shih-Zhu-Lumley model (SZL) in Flotran uses this coecient, but drops all
higher order terms.
4.3.3 Reynolds stress models
In the previous sections the Reynolds stresses v
are modelled as a function of

the mean velocity gradient. As to avoid this assumption, Reynolds stress models
(RSM) start from the exact transport equations for the Reynolds stresses v
i
v
j
:
v
i
v
j
t
+ v
k
v
i
v
j
x
k
= 1
ij
+

x
k
_
i
v
j
x
k
+T
kij
_
+T
ij

ij
(4.47)
The production tensor T
ij
is in closed form:
T
ij
= v
i
v
k
_
v
j
x
k
_
v
j
v
k
_
v
i
x
k
_
(4.48)
The pressure-rate-of-strain tensor 1
ij
is equal to:
1
ij
=
p
_
v
i
x
j
+
v
j
x
i
_
(4.49)
This expression introduces new unknowns and is the most important term to be
modelled. The part T
kij
of the turbulent transport is given by:
T
kij
= v
i
v
j
v
i
p
jk
j
p
ik
(4.50)
This is usually modelled using a gradient-diusion model. The dissipation tensor
ij
is equal to:
ij
= 2
v
i
x
k
v
j
x
k
(4.51)
Because dissipation occurs at the smallest scales, often isotropy is assumed for the
dissipation tensor:
ij
=
2
3
ij
(4.52)
TURBULENT AIR FLOW AROUND A SINGLE CYLINDER 119
where the turbulent energy dissipation rate = 1/2
ii
is the same as appears
in the standard k- models. An additional transport equation is solved for the
turbulent energy dissipation rate :
t
+ v
i
x
i
=

x
i
_
C
i
v
x
j
_
+ C
1
k
T C
2
2
k
(4.53)
Globally seven transport equations (for the six Reynolds stresses v
i
v
j
and the
turbulent energy dissipation rate ) have to be solved instead of two equations for
the two-equation eddy viscosity models. The hypothesis that the Reynolds stress
tensor is determined by the velocity gradients is removed which enables the models
to capture the streamline curvature and swirl eects. The transport equations for
the Reynolds stresses make it possible to take the eects of the ow history into
account.
4.4 Turbulent air ow around a single cylinder
Circular cylinders are widely used in civil engineering constructions as silos,
chimneys, water towers, power transmission lines, oshore structures and
suspension bridge cables.
The turbulent air ow around a single cylinder is computed in order to clarify
the inuence of turbulence models, near-wall mesh renement and unsteadiness
(Dooms et al., 2006a).
The cylinder represents a silo in the port of Antwerp with a diameter D = 5.5 m.
For the air, a density = 1.25 kg/m
3
and a dynamic viscosity = 1.7610
5
Pa s
are used. The bottom of the silos is located at 16.66 m above ground level. Due to
the proximity of the river Scheldt, the surroundings are quite at and vast, which is
reected in terrain category II (BIN, 1995) with an aerodynamic roughness length
of 0.05 m. The mean wind velocity at half the height of the silo or z = 30 m above
ground level is equal to v = 31.8 m/s. The Reynolds number Re = Dv/ is equal
to 12.4 10
6
.
The ow around a nite circular cylinder of height-to-diameter ratio 4.55 (as the
silos in the port of Antwerp) is strongly three-dimensional. The critical ratio below
which the free end eect extends to the base of the cylinder and Karman vortex
shedding is suppressed, varies between the dierent studies in the literature from 2
to 6 (Pattenden et al., 2005). The three-dimensional ow around one nite cylinder
of small aspect ratio embedded in an atmospheric boundary layer has been often
studied experimentally. However, the numerical simulation of this ow (using
LES) is still challenging from a computational point of view at moderate Reynolds
numbers (Re = 20000 200000) (Frohlich and Rodi, 2004; Afgan et al., 2007;
Pattenden et al., 2007). Computations of ows around more complex geometries
in atmospheric boundary layers at high Reynolds number have been performed,
but with quite coarse near-wall meshes (Nozu et al., 2008).
Using a single processor, the numerical simulation of the three-dimensional ow
around forty closely-spaced nite circular cylinders at Re = 12.4 10
6
is out of
reach. Therefore, all ow computations around cylinders will be limited to two
dimensions, corresponding to an innite cylinder.
For the two-dimensional ow around a cylinder, the uid particles reach the
cylinder rst at the stagnation point, where the velocity is equal to zero and
the streamline is perpendicular to the wall. The angle = 0
coincides with
the upstream stagnation point. The pressure coecient C
p
= 2(p p
f
)/(v
2
f
)
(gure 4.1) is a dimensionless expression for the pressure at the cylinders surface,
where p
f
and v
f
are the free stream pressure and velocity, respectively. It is
maximal and equal to 1 at the stagnation point and gradually decreases around
the circumference. The boundary layers stay attached in this favourable pressure
region. The pressure coecient attains a minimum C
min
p
and starts to increase.
The boundary layers now experience an adverse pressure gradient and separates
from the wall at
s
. A recirculation zone is formed behind the cylinder. The
pressure is approximately constant in the recirculation zone. The base pressure
coecient C
b
p
is the pressure coecient at = 180
. The drag coecient

C
d
=
_
2
0
C
p
cos d is a dimensionless expression for the force experienced by
the cylinder in the ow direction. The ow around a cylinder is a an ideal test
case for the turbulence models because it combines several types of ow structures
as separation, recirculation, stagnation and streamline curvature.
C
min
p
C
b
p
1
s
v
Figure 4.1: Pressure coecient C
p
for a single cylinder.
As the Reynolds number Re is larger than 3.5 6 10
6
, the regime of the ow
around the cylinder is post-critical (Zdravkovich, 1997): the wake and the shear
layers are fully turbulent and the boundary layers become fully turbulent prior
to separation. In the boundary layers, the transition from laminar to turbulent
ow takes place between the stagnation and the separation point. Regular vortex
shedding reappears, while it is absent at lower Reynolds numbers.
Eurocode 1 (BIN, 1995) describes the pressure coecient as a function of the
circumferential angle for a Reynolds number Re = 10
7
. The ow around a circular
cylinder has been widely studied, mainly focussing on the regular vortex shedding
in the subcritical regime and the drag crisis in the critical regime. Zdravkovich
(1997) gives an overview of experimental data. There is a lack of detailed
experimental data at post-critical Reynolds numbers, while available data show
considerable scatter, which may be explained by the high sensitivity of the ow
to perturbations due to surface roughness and free-stream turbulence. Numerical
results in the present section are always compared with 12 experimentally obtained
pressure coecients at Reynolds numbers from 0.73 10
7
to 3.65 10
7
(gure
4.2). Because of the amount of scatter and the number of experimental curves, it
is reasonable to consider numerical results as accurate if they fall within the zone
dened by these curves.
0 45 90 135 180
2.5
2
1.5
1
0.5
0
0.5
1
Angle []
P
r
e
s
s
u
r
e

c
o
e
f
f
i
c
i
e
n
t

[
]
Figure 4.2: Measured pressure coecients at Reynolds numbers from 0.73 10
7
to 3.65 10
7
(Zdravkovich, 1997).
Few computations of the ow in the post-critical regime have been performed.
A number of two-dimensional URANS simulations are reported in the literature.
Celik and Shaer (1995) used URANS with an empirically xed transition point
to compute the ow for Reynolds numbers up to 3.6 10
6
. The predictions are
strongly inuenced by grid renement and especially by the grid distribution in
the boundary layer. The best results are obtained with the rst grid point located
in the viscous sublayer. Holloway et al. (2004) applied an URANS technique
capable of resolving boundary layer transition for Reynolds numbers up to 10
7
. At
Reynolds number 10
7
, similar results are obtained for the fully turbulent model and
the transition model. The magnitude of the time averaged drag coecient does
not agree with experimental values due to a rather small overestimation in the
separation angle. Saghaan et al. (2003) compared the ow computed with a non-
linear eddy viscosity model and with the standard k model. The cubic terms in
the non-linear model account for eects due to streamline curvature. The pressure
coecient around the circumference at Reynolds number 8.410
6
has a signicant
decrease between 165
and 180
, which is explained by the assumption that the

ow is two-dimensional, causing vortices to roll up close to the cylinder. Younis
and Przulj (2006) modied the k model to account for the direct energy input
at the vortex shedding frequency from the mean ow into the random turbulence
motions. This eect is introduced by an additional source term in the dissipation
rate equation. The results are compared with the RNG k model. The pressure
coecient around the circumference at Reynolds number 3.5 10
6
signicantly
decreases as well between 150
and 180
.
Travin et al. (2000) applied three-dimensional DES for Reynolds numbers up to 3
10
6
. For the turbulent separation cases, the results obtained with two-dimensional
URANS computations are very close to those obtained with three-dimensional
DES. Adding a curvature correction term to the turbulence model improves the
estimation of the separation angle, the base pressure and the drag coecient.
Catalano et al. (2003) used three-dimensional LES with the dynamic Smagorinsky
model to compute the ow for Reynolds numbers up to 2 10
6
and compared
it with three-dimensional RANS and URANS results. The solutions show
relative insensitivity to the Reynolds number and inaccurate predictions at higher
Reynolds numbers (2 10
6
) which are probably due to poor grid resolution.
Tamura et al. (1990) computed the ow in two and three dimensions without any
turbulence model, which corresponds to a LES using the discretization scheme
as a lter with numerical dissipation. Using this technique, grid renement in
the circumferential direction strongly inuences the results obtained with a two-
dimensional computation. At Re = 10
6
, the time averaged drag coecients C
d
of
the two-dimensional and three-dimensional computations are the same. For super-
critical and post-critical ow, the eects of unsteadiness and three-dimensional
vortex structures in the ow are less important and mainly located in the wake.
Singh and Mittal (2005) performed a similar analysis without any turbulence model
in two dimensions for Reynolds numbers up to 10
7
. At Reynolds number 10
6
, the
pressure coecient around the circumference has a signicant decrease between
140
and 180
.
Some results of these numerical simulations reported in the literature are listed in
table 4.2.
A two-dimensional RANS and URANS simulation will be performed with the
Flotran nite element code (Ansys, 2005b) and the CFX nite volume code (Ansys,
2005a) using a number of eddy viscosity turbulence models on two dierent meshes.
In Flotran the second order accurate Collocated Galerkin (COLG) approach is
used for the spatial discretization. To handle the coupling between the pressure
and momentum equations, an enhanced segregated solution algorithm (SIMPLEN)
(Wang, 2001) is employed. In CFX the High Resolution spatial discretization is
used, which is an automatically determined blend of a rst and a second order
accurate scheme. The coupled algorithm solves the momentum and continuity
equation as a single system. These options are used for all computations in this
thesis.
4.4.1 Steady computation: problem domain, near-wall
modelling and eddy viscosity turbulence models
At the inlet a uniform velocity v = 31.8 m/s, a turbulent kinetic energy
k = 0.1521 m
2
/s
2
and a turbulent energy dissipation rate =
C
k
3/2
0.99
=
0.0053929 m
2
/s
3
are imposed. This corresponds to a turbulence intensity I =
_
2
3
k/v = 1.00%. In atmospheric boundary layers, the turbulence intensity is
higher, but this low turbulence intensity enables to compare with the results from
wind tunnel tests. At the lateral boundaries symmetry is imposed. The walls are
considered smooth and no-slip boundary conditions are applied. In Flotran zero
pressure is imposed at the outlet, while in CFX the average of the pressure over
the outlet should be zero. For the steady computations, symmetry permits the
ow to be computed on one side of the cylinder, with reective symmetry imposed
along the centerline.
The boundaries of the uid domain should be suciently far from the region close
to the cylinder where the accuracy of solution is important. Behr et al. (1991,
1995) suggest a distance of at least 8D for the inlet and the lateral boundaries
and a distance of 22.5D for the outlet, with D the diameter of the cylinder. For
the present computations, a distance of 9D is adopted for the inlet and the lateral
boundaries and 30D for the outlet (gure 4.3). Results on a larger problem domain,
where the inlet and the lateral boundaries are located at 12D and the outlet at
40D, are comparable.
All simulations are performed on two dierent meshes that dier only in the near-
wall region. In mesh A
2
, a high Reynolds formulation is used and the nodes next
to the cylinder wall are predominantly placed in the logarithmic law region. Close
to the cylinder wall, the mesh is structured and consists of 138 quadrilaterals
around the circumference and 70 elements in the radial direction. Far away from
the cylinder wall, an unstructured mesh consisting of triangles is used. The mesh
9D 9D 21D
9D
Figure 4.3: Problem domain and mesh A
2
for a single cylinder.
consists of 26765 elements and 22899 nodes. The dimensionless wall distance y
+
of the nodes next to the cylinder wall varies from 0 at the stagnation points to 130.
The choice of the wall functions (table A.1) has no important inuence on the
pressure coecient, both within Flotran (equilibrium, Van Driest and Spalding
wall functions) and within CFX (standard and scalable wall functions).
In mesh B, a low-Reynolds formulation is used and the nodes next to the cylinder
wall are located in the viscous sublayer at y
+
1. Mesh B consists of 54694
elements and 51009 nodes. The structured mesh close to the cylinder wall consists
of 409 quadrilaterals around the circumference and 80 elements in the radial
direction. The dimensionless wall distance y
+
of the nodes next to the cylinder
wall is smaller than 1.6.
0 45 90 135 180
2.5
2
1.5
1
0.5
0
0.5
1
Angle []
P
r
e
s
s
u
r
e

c
o
e
f
f
i
c
i
e
n
t

[
]
(a)
0 45 90 135 180
2.5
2
1.5
1
0.5
0
0.5
1
Angle []
P
r
e
s
s
u
r
e

c
o
e
f
f
i
c
i
e
n
t

[
]
(b)
Figure 4.4: Pressure coecient as a function of the angle obtained with Flotran
(solid lines) and CFX (dashed lines) turbulence models using various
eddy viscosity turbulence models ( k , SST, SZL) for (a)
mesh A
2
and (b) mesh B. The light grey zone contains all available
experimental data at Reynolds numbers from 0.7310
7
to 3.6510
7
.
Figure 4.4 shows the pressure coecient computed for various turbulence models
with a steady simulation. Near-wall modelling strongly inuences the pressure
coecient computed with Flotran, while the results obtained with CFX are quite
similar for a high- and a low-Reynolds formulation. A light grey zone containing
all available experimental data at Reynolds numbers from 0.73 10
7
to 3.65 10
7
is superimposed on this and the following gures. In Flotran, changing from a
high- to a low-Reynolds formulation the correspondence between measurements
and results improves for the standard k and the SZL model, but not for the
SST model.
Figure 4.4 as well shows that the inuence of the dierent eddy viscosity turbulence
models (table A.1) on the pressure coecient is large. The results obtained
with the SZL turbulence model of Flotran have the best correspondence with the
experimental results. Within CFX, the SST turbulence model produces reasonable
results.
The standard k- model overestimates the turbulence production in front of the
cylinder (gure 4.5a), does not account for streamline curvature and predicts the
onset of separation too late (gure 4.5b). Dierent results are obtained with
Flotran and CFX using the same turbulence models with the same parameters.
These are probably caused by dierences in the near-wall modelling employed in
Flotran and CFX.
0 45 90 135 180
0
5
10
15
20
Angle[]
T
u
r
b
u
l
e
n
t

k
i
n
e
t
i
c

e
n
e
r
g
y

[
m
2
/
s
2
]
(a)
0 45 90 135 180
0
2
4
6
8
10
12
14
Angle[]
W
a
l
l

s
h
e
a
r

s
t
r
e
s
s

[
P
a
]
(b)
Figure 4.5: (a) Turbulent kinetic energy and (b) wall shear stress as a function
of the angle at the cylinders surface using dierent eddy viscosity
turbulence models ( k , SST, SZL) for mesh B: Flotran (solid
lines) and CFX (dashed lines).
The SZL turbulence model of Flotran produces the best overall correspondence
with the experimental results. Since computations converge faster in CFX due to
the use of a multigrid solver, the SST turbulence model of CFX, whose results
correspond, of all the turbulence models in CFX, best with the experimental
results, is used for all subsequent calculations.
4.4.2 Unsteady computation
In the previous section, steady state solutions have been computed, although
regular vortex shedding is present in the post-critical regime, which demands an
unsteady RANS computation (Iaccarino et al., 2003). In this section, the results
from a transient computation of the ow are compared with the steady state
results.
The vortex shedding frequency f
vs
is described by the dimensionless Strouhal
number St = f
vs
D/v. For ows with Re 10
7
, experimental values for the
Strouhal number range from 0.27 to 0.32 (Zdravkovich, 1997). Eurocode 1 suggests
a constant value of 0.2, independent of the Reynolds number.
The transient solution is integrated by the three-point backward dierence
scheme with a dimensionless time step t v/D = 0.029, which corresponds to
approximately 100 time steps per vortex shedding period. Within every time step,
maximum 5 iterations are performed in order to reduce the RMS of the normalized
residuals to 10
6
. The computed time window corresponds to ten vortex shedding
periods. For mesh A
2
, 937 time steps are computed, which results in a time
window of 4.68 s. With a shedding frequency of 2.13 Hz, a Strouhal number of 0.37
is calculated. For mesh B, 954 time steps are computed, which results in a time
window of 4.765 s. With a shedding frequency of 2.10 Hz, a Strouhal number of
0.36 is calculated.
Figure 4.6 shows the time history and the frequency content of the pressure at
= 172
at the cylinders surface for mesh B. The vortex shedding frequency and
some higher harmonics are clearly visible in the frequency content.
0 1 2 3 4
350
300
250
200
Time [s]
P
r
e
s
s
u
r
e

[
P
a
]
(a)
0 5 10 15 20
0
5
10
15
20
25
30
35
40
Frequency [Hz]
P
r
e
s
s
u
r
e

[
P
a
/
H
z
]
(b)
Figure 4.6: (a) Time history and (b) frequency content of the pressure at the
cylinders surface at = 172
for a single cylinder using mesh B.

Figure 4.7 shows the time average and the standard deviation of the pressure p.
The stagnation pressure at the windward side and the suction in the wake are
clearly visible. The largest time variations of the pressure occur in the wake.
(a) (b)
Figure 4.7: (a) Time average and (b) standard deviation of the pressure p for a
single cylinder using mesh B.
Figure 4.8 compares the pressure coecient of the steady state computation with
the time averaged pressure coecient C
p
of the transient computation. The
maxima and minima of the pressure coecient during the transient computation
are depicted as well. Compared to the steady state results, the time average of the
transient computation predicts a lower minimum pressure coecient C
min
p
and
a lower base pressure coecient C
b
p
, which is a deterioration for the minimum
pressure coecient C
min
p
and an improvement for the base pressure coecient C
b
p
with regard to the experimental data.
0 45 90 135 180
2.5
2
1.5
1
0.5
0
0.5
1
Angle []
P
r
e
s
s
u
r
e

c
o
e
f
f
i
c
i
e
n
t

[
]
(a)
0 45 90 135 180
2.5
2
1.5
1
0.5
0
0.5
1
Angle []
P
r
e
s
s
u
r
e

c
o
e
f
f
i
c
i
e
n
t

[
]
(b)
Figure 4.8: Pressure coecient as a function of the angle obtained with the
steady state computation (dashed line) and the time average (solid
line), minimum (dash-dotted line) and maximum (dotted line) of the
pressure coecient obtained with the transient computation for (a)
mesh A
2
and (b) mesh B. The light grey zone contains all available
experimental data at Reynolds numbers from 0.7310
7
to 3.6510
7
.
As the time average of the transient solution diers clearly from the steady state
solution, a transient approach is more physically sound and oers additional
information about the time variation of the pressure (Iaccarino et al., 2003).
Therefore, only transient computations will be performed for the ow around a
cylinder group.
4.4.3 Mesh renement
In order to study the inuence of mesh renement, the unsteady ow is computed
on four dierent meshes, with a mesh renement factor r equal to
2. The nest
mesh A
1
is a factor
2 ner than mesh A

2
. The coarsest mesh A
4
is a factor 2
coarser than mesh A
2
. Mesh A
3
is a factor
2 coarser than mesh A

2
. For each
set of three meshes a mesh renement study is made: study 1 starting from the
solutions on meshes A
2
, A
3
and A
4
and study 2 with the meshes A
1
, A
2
and A
3
.
For all meshes the time averaged pressure coecient at the cylinder wall is
computed (gure 4.9a). At the stagnation point, the lowest pressure coecient
1.015 is obtained using the nest mesh, while the coarsest mesh gives a value 1.038.
The sequence of the solutions changes between 55
and 67
. Near the minimum,

the nest mesh yields the highest value -2.444 and the coarsest mesh the lowest
value -2.468. The sequence of the solutions changes again around the separation
point, between 115
and 121
. In the recirculation zone the largest dierences

occur and the nest mesh produces a value that is about 0.05 lower than the value
of the coarsest mesh. The pressure coecient is symmetrical, so the conclusions
between 181
and 360
are similar to those for 0
to 180
.
0 90 180 270 360
2.5
2
1.5
1
0.5
0
0.5
1
Angle []
P
r
e
s
s
u
r
e

c
o
e
f
f
i
c
i
e
n
t

[
]
(a)
0 90 180 270 360
0
1
2
3
4
5
Angle []
O
r
d
e
r

o
f

a
c
c
u
r
a
c
y

[
]
(b)
Figure 4.9: (a) Pressure coecient obtained on meshes A
1
(black), A
2
(blue), A
3
(green) and A
4
(red) and extrapolated solutions for mesh renement
study 1 (cyan) and 2 (magenta) and (b) local apparent order of
accuracy for mesh renement study 1 (solid line) and 2 (dashed line).
The solutions on the three meshes should be known at a common location.
Therefore, the solutions on the ne and the intermediate mesh are interpolated to
the locations of the nodes of the coarse grid using piecewise cubic splines. Based on
three solutions
k
computed on meshes A
k
, the convergence ratio s is calculated:
s =

2
2
(4.54)
If s is between 0 and 1, the convergence is monotonic. For s smaller than 0,
the convergence is oscillatory. If s is larger than 1, the solution diverges. In
the zones where the sequence of the solutions changes (between 55
and 67
and
between 115
and 121
), the convergence is oscillatory or divergent. Near the

stagnation point (between 357
and 360
) the solution diverges as well. For mesh

renement study 1 and 2 respectively 3.6 % and 6.7 % of the points exhibited
oscillatory convergence or divergence. At the points with monotonic convergence,
the apparent order of accuracy p is computed:
p =
ln
_
1
_
ln r
(4.55)
Figure 4.9b shows the apparent order of accuracy. For mesh renement study 1
the local order of accuracy p ranges from 0.04 to 8.52 with an average of 1.27, while
for study 2 the local order of accuracy p ranges from 0.05 to 13.00 with an average
of 1.70. The variability of the apparent order is large, which is explained by the
fact that it is poorly estimated except in the asymptotic range (Stern et al., 2001).
For the High Resolution discretization, in the asymptotic range theoretically an
order of accuracy between 1 and 2 is expected. Ferziger and Peric (2002) state that
for turbulent ows the denition of order is often dicult. In the neighbourhood
of the separation points, very large and very small apparent orders occur. Near
the stagnation points the apparent order becomes very small. In the recirculation
zone an almost constant, but very low value of 0.5 is obtained. Around 49
and
311
the apparent order becomes very large.

Starting from the solutions
1
and
2
and the apparent order p, an extrapolated
solution
21
e
is computed:
21
e
=
r
p
2
r 1
(4.56)
If the apparent order is quite unrealistic, this leads to unrealistic extrapolated
values (Celik and Karatekin, 1997). Therefore, points where the apparent order is
smaller than 0.5 or larger than 5 are disregarded here. For mesh renement study
1 and 2 respectively 12.3 % and 7.2 % of the points are ignored. The extrapolated
solutions for both mesh renement studies are shown as well in gure 4.9a. The
dierences between the solution on the nest mesh and the extrapolated solutions
is small except in the recirculation zone, where the dierence increases to 0.05 due
to the small apparent order. The time averaged drag coecient can be estimated
as 0.40 using the extrapolated pressure coecient of mesh renement study 2.
As to quantify the errors, the approximated relative error
21
a
is computed:
21
a
=
(4.57)
Based on the extrapolated solution
21
e
the extrapolated relative error
21
e
is
calculated:
21
e
=
21
e

1
21
e
(4.58)
The grid convergence index GCI
21
is dened as (Celik and Karatekin, 1997):
GCI
21
=
1.25
21
a
r
p
1
(4.59)
The grid convergence index (GCI) is computed twice: once using the local apparent
order of accuracy and once with the average apparent order of accuracy.
Figure 4.10a shows the dierent errors. The relative errors become very large
between 28
and 35
and between 326
and 332
, as the pressure coecient is

there around zero. The largest errors are present in the recirculation zone. There,
the GCI using the local apparent order of accuracy predicts errors of about 20 %
and 14 % for respectively mesh A
2
and A
1
. The extrapolated relative errors are
respectively around 14 % and 10 %. Using the average apparent order of accuracy,
the GCI yields 8 % and 4 %. The approximated relative errors are about 3.5 %
and 2.5 %.
Figure 4.10b shows the time history of the drag and the lift coecients as obtained
on the dierent meshes. The time averaged drag coecient C
d
and the uctuation
C
d
=
_
(C
d
C
d
)
2
/N of the drag coecient increase when the mesh is rened.
The uctuation C
l
=
_
(C
l
C
l
)
2
/N of the lift coecient increases as well when
the mesh is rened. Table 4.1 gives the values obtained on the dierent meshes
and the approximated relative errors for the time averaged drag coecient C
d
, the
RMS of the drag coecient C
RMS
d
=
_
C
2
d
/N, the RMS of the lift coecient
C
RMS
l
=
_
C
2
l
/N, the uctuation C
d
of the drag coecient, the uctuation C
l
of the lift coecient and the separation angle
s
. None of the integral variables
0 90 180 270 360
0
10
20
30
40
50
Angle []
E
r
r
o
r

[
%
]
(a)
0 0.2 0.4 0.6 0.8 1 1.2 1.4
0.2
0.1
0
0.1
0.2
0.3
0.4
0.5
Time [s]
D
r
a
g
/
l
i
f
t

c
o
e
f
f
i
c
i
e
n
t

[
]
(b)
Figure 4.10: (a) Approximated relative errors (dash dotted lines), extrapolated
relative errors (dashed lines), GCI using the local (dotted lines)
and the average (solid lines) apparent order of accuracy for mesh
renement study 1 (thin lines) and 2 (thick lines) and (b) time history
of drag (thin) and lift (thick) coecients obtained on meshes A
1
(solid line), A
2
(dash dotted line) , A
3
(dashed line) and A
4
(dotted
line).
(C
d
, C
RMS
d
, C
RMS
l
, C
d
and C
l
) is converging (s > 1). The order of convergence for
integral quantities is clearly not the same as for eld values (Ferziger and Peric,
2002). The errors for the uctuating parts are still quite large. For the three nest
meshes the separation angle
s
converges to a value of 116
4

3

2

1

43
a

32
a

21
a
C
d
0.341 0.342 0.349 0.358 0.5 % 2.1 % 2.5 %
C
RMS
d
0.341 0.342 0.349 0.358 0.5 % 2.1 % 2.5 %
C
RMS
l
0.0753 0.0834 0.0923 0.1023 9.7 % 9.6 % 9.8 %
C
d
0.00155 0.00203 0.00260 0.00334 23.7 % 22.0 % 22.2 %
C
l
0.0753 0.0834 0.0923 0.1023 9.7 % 9.6 % 9.8 %
s
119 118 118 117 0.4 % 0.4 % 0.3 %
Table 4.1: The time averaged drag coecient C
d
, the RMS of the drag coecient
C
RMS
d
, the RMS of the lift coecient C
RMS
l
, the uctuation C
d
of
the drag coecient, the uctuation C
l
of the lift coecient and the
separation angle
s
obtained on meshes A
1
, A
2
, A
3
and A
4
and the
dierent approximated relative errors.
4.4.4 Comparison with numerical results reported in the
literature
Table 4.2 compares the Reynolds number Re, the Strouhal number St, the
separation angle
s
b
p
, the minimum pressure
coecient C
min
p
d
and the adverse pressure recovery APR =
C
b
p
C
min
p
from the present computations with numerical results from the literature.
The Reynolds numbers of the numerical results published in the literature are
slightly lower than the Reynolds number of the present calculations.
In the dierent experiments (Zdravkovich, 1997), Strouhal numbers lie within the
range of 0.27 St 0.32, and the boundary layer separates between 100
and
110
. After separation, the base pressure coecient C

b
p
is constant and lies within
the range 0.5 to 0.8. The adverse pressure recovery varies from 1.0 to 1.5. Drag
coecients coecients vary from 0.40 to 0.80.
According to Eurocode 1 (BIN, 1995), the minimum value of the pressure coecient
C
min
p
equals 1.5 at an angle of 75
. The boundary layer separates at 105
. After
separation, the base pressure coecient C
b
p
is constant and equal to 0.8. This
corresponds to an extremely low adverse pressure recovery of 0.7. Eurocode 1
mentions the value 0.2 for the Strouhal number, which corresponds to laminar
vortex shedding. The drag coecient is equal to 0.72 for a smooth surface (k/b =
10
5
).
Simulations with the standard k- model predict separation much too late, yielding
a smaller and shorter wake with vortices shed at a higher frequency. The suction
in the wake is too low, which results in a lower drag coecient.
If the present computations are compared with the 12 experimentally obtained
pressure coecients (gure 4.2), they overestimate the base pressure and
consequently underestimate the drag coecient, but some of the more advanced
models in the literature do as well. The Shih-Zhu-Lumley model yields relatively
good results on mesh B, though the base pressure coecient is quite high. In the
unsteady computations with the SST model, the boundary layer separates slightly
too late, and therefore the drag coecient is underestimated and the Strouhal
number is overestimated. The time averaged base pressure coecient is lower
than in steady computations.
Re St
s
-C
b
p
-C
min
p
C
d
APR
10
6
[
]
2-D RANS k Flotran mesh A
2
12.4 158 -0.08 2.6 0.30 2.7
2-D RANS k Flotran mesh B 12.4 143 0.08 2.5 0.31 2.4
2-D RANS SST Flotran mesh A
2
12.4 113 0.37 2.1 0.36 1.7
2-D RANS SST Flotran mesh B 12.4 118 0.31 2.3 0.30 2.0
2-D RANS SZL Flotran mesh A
2
12.4 116 0.35 2.0 0.35 1.7
2-D RANS SZL Flotran mesh B 12.4 108 0.41 1.8 0.42 1.4
2-D RANS k CFX mesh A
2
12.4 147 0.10 2.4 0.33 2.3
2-D RANS k CFX mesh B 12.4 149 0.10 2.4 0.34 2.3
2-D RANS SST CFX mesh A
2
12.4 114 0.38 2.2 0.29 1.9
2-D RANS SST CFX mesh B 12.4 113 0.39 2.2 0.30 1.8
2-D URANS SST CFX mesh A
2
12.4 0.37 117 0.45 2.4 0.35 2.0
2-D URANS SST CFX mesh B 12.4 0.36 116 0.46 2.4 0.36 2.0
Eurocode (BIN, 1995) 10.0 0.2 105 0.80 1.5 0.72 0.7
2-D URANS k transition 3.6 118 0.35 2.3 2.0
(Celik and Shaer, 1995)
2-D URANS realizable k 10.0 120 0.26
(Holloway et al., 2004)
2-D URANS transition 10.0 119 0.25
(Holloway et al., 2004)
2-D URANS RNG k 3.5 0.28 122 0.8 2.5 0.56 1.7
(Younis and Przulj, 2006)
2-D URANS modied k 3.5 0.28 120 1.25 2.5 0.72 1.3
(Younis and Przulj, 2006)
2-D URANS k 8.4 0.25 104 0.72 1.8 0.66 1.1
(Saghaan et al., 2003)
2-D URANS non-linear 8.4 0.33 125 1.15 2.6 0.61 1.5
(Saghaan et al., 2003)
3-D DES 3.0 0.35 111 0.53 2.2 0.41 1.7
(Travin et al., 2000)
3-D DES + curvature 3.0 0.33 106 0.64 2.1 0.51 1.5
(Travin et al., 2000)
3-D URANS k 1.0 0.31 0.41 2.3 0.40 1.9
(Catalano et al., 2003)
3-D LES Smagorinsky 1.0 0.35 0.32 2.4 0.31 2.1
(Catalano et al., 2003)
2-D LES discretization 10.0 1.25 0.85
(Singh and Mittal, 2005)
Table 4.2: Comparison of the Reynolds number Re, the Strouhal number St, the
separation angle
s
b
p
, the minimum
pressure coecient C
min
p
d
and the adverse
pressure recovery APR with numerical results from the literature.
4.5 Turbulent air ow around a cylinder group
As cylinders are often placed in groups and the conguration of these groups
largely inuences the pressure distribution around the cylinders, an experimental
or numerical study of the wind ow should provide a more realistic estimation of
the pressure coecients and forces.
For the silo group in the port of Antwerp (gure 1.3), the forty silos are placed in
ve rows of eight silos with gaps of 30 cm between two neighbouring silos (gure
1.4). The spacing ratio of the distance P between the center of two cylinders to the
cylinder diameter is P/D = 5.8/5.5 = 1.05. In this section, the turbulent air ow
around cylinder groups with spacing ratio P/D = 1.05 is studied. No experimental
or numerical data are available for the ow around a group of cylinders with such
a small spacing ratio and at the present high Reynolds number. Therefore, the
incompressible turbulent wind ow around a group of 8 by 5 cylinders is simulated.
Before, the less comprehensive case of a group of 2 by 2 cylinders is modelled.
The ow around two cylinders in a staggered conguration has been thoroughly
studied experimentally, mostly at high subcritical Reynolds numbers. Gu and Sun
(1999) visualized the ow at Reynolds number 5.6 10
3
, measured time averaged
pressure distributions at Reynolds number 2.2 10
5
, and derived lift and drag
forces. Extensive ow visualization studies by Sumner et al. (2000) at Reynolds
number 850 Re 1900 identied nine dierent ow patterns. Alam et al.
(2005) measured time averaged and uctuating pressure distributions at Reynolds
number 5.5 10
4
for spacing ratios P/D from 1.1 to 6 and for incidence angles
between 10
to 75
. Closely spaced cylinder congurations behave similarly as

a single blu body: Karman vortex shedding occurs from the group as a whole,
and one Karman vortex street exists in the combined wake. The time averaged
aerodynamic forces undergo large changes with the incidence angle between the
wind direction and the longitudinal group axis (Sumner et al., 2005). Upstream
cylinders experience smaller uctuating forces than a single cylinder.
For small incidence angles < 1020
(nearly tandem conguration - gure

4.11a) the inner separated shear layer from the upstream cylinder reattaches
onto the outer side of the downstream cylinder, preventing ow through the
gap (shear layer reattachment (Sumner et al., 2000) or pattern I
B
(Gu and
Sun, 1999)). The pressure distribution of the downstream cylinder attains
a maximum where reattachment occurs and a minimum where the outer
boundary layer separates. The downstream cylinder experiences a negative
drag. Only one Strouhal number is measured in the wake, which is slightly
higher than the value for a single cylinder.
TURBULENT AIR FLOW AROUND A CYLINDER GROUP 135
At incidence angles 20
45
(gure 4.11b) the inner shear layer from

the upstream cylinder is directed through the gap and reattaches with its
high-speed side onto the inner side of the downstream cylinder. It induces
separation on the inner side of the downstream cylinder (induced separation
(Sumner et al., 2000) or pattern II
B
(Gu and Sun, 1999)). A large area of
suction is developed on the downstream cylinder in the gap region, resulting
in a peak in the lift coecient of the downstream cylinder. This corresponds
with a local maximum drag for the upstream and a local minimum drag for
the downstream cylinder. Vortex shedding occurs at dierent frequencies
from the outer shear layers of the upstream and downstream cylinder.
For large angles between 45
and 90
(gure 4.11c), uid enters through

the gap in the base region, causing a lengthening of the near-wake region
(base bleed (Sumner et al., 2000) or pattern III
B
(Gu and Sun, 1999)). This
gap ow between the shear layers, separated from the inner side of both
cylinders, is typically deected towards the upstream cylinder, although it
might also be instantaneously deected towards the downstream cylinder.
For large angles, the lift forces pull the cylinders away from each other. The
same low-frequency Strouhal number is measured behind both cylinders. It
decreases with increasing to 0.10 at = 90
.
(a)
(b)
(c)
Figure 4.11: Flow patterns around two closely spaced cylinders for (a) small, (b)
medium and (c) large incidence angles (Sumner et al., 2000).
Gu (1996) measured pressure distributions at supercritical Reynolds numbers 4.5
10
5
, which dier from those at subcritical Reynolds number. The downstream
cylinder experiences higher drag coecients than a single cylinder for angles >
45
due to an increased base pressure and an extended stagnation area.

Numerical simulations are mainly available at low Reynolds numbers (Re 1000).
Flow patterns computed using the vortex method for the laminar two-dimensional
ow at Reynolds number 800 (Akbari and Price, 2005) correspond well with
experimental data for a large number of congurations. Jester and Kallinderis
(2003) used the nite element method to compute time averaged drag coecients
for spacing ratios P/D 1.5.
Studies of the ow around arrays with more than two cylinders are quite rare and
mainly treat side-by-side or tandem congurations. Experiments on groups of four
cylinders in a square conguration were performed by Lam and Fang (1995). The
time averaged pressure distributions were measured at Reynolds number 12.810
3
for spacing ratios P/D from 1.26 to 5.80 in 15 steps and for angles between 0
to
45
in steps of 15
. Lift and drag forces were obtained by integrating the pressure

distributions. Flow patterns change rapidly as the spacing ratio is decreased and
highly depend on the incidence angle. Consequently, time averaged drag and lift
coecients of closely spaced arrays are very sensitive to changes in spacing ratio.
Unique characteristics were observed at the minimum spacing ratio. Lam et al.
(2003) measured the time averaged and uctuating drag and lift forces at the
subcritical Reynolds number 4.1 10
4
for spacing ratios P/D from 1.69 to 3.83
and for angles from 0
to 45
in steps of 15
. Strouhal numbers were derived

from the power spectral density of the forces. Dierent cylinders may have distinct
Strouhal numbers that change with the incidence angle, due to the existence of
wider and narrower wakes behind the cylinders. Farrant et al. (2000) computed
the laminar ow around an array of four cylinders using the cell boundary element
method at Reynolds number 200 for large spacing ratios P/D = 3 and 5 and for
angles = 0
and 45
.
A lot of experimental research focuses on the ow in tube bundles in heat
exchangers (Zdravkovich, 2003). Weaver et al. (1993) studied the periodic
excitation mechanism in a rotated square ( = 45
) arrangement with spacing

ratios 1.21 P/D 2.83. At the smallest spacing ratio, only one Strouhal
number was found, while at higher spacing ratios, two distinct numbers were
observed, which correspond to vortex shedding from the rst and the second
tube rows. A ow visualization in a rotated square arrangement with spacing
ratio P/D = 1.50 at Reynolds numbers 80 Re 1300 was carried out by
Price et al. (1995) and only revealed one Strouhal number. Using periodic
boundary conditions, the ow computation in rotated square arrangements forms
an attractive test for turbulence models due to the contractions and expansions of
the ow, the high turbulence intensities (35%) and streamline curvatures. RANS
models generally predict the velocity proles reasonably well, but Reynolds
stress proles are poor because the size of the larger eddies is comparable to
the cylinder radius and uctuating velocities are larger than the mean velocities
in some regions (Rollet-Miet et al., 1999). In the stagnation region in front of
the cylinders, the turbulent kinetic energy is decreasing which is not predicted
by RANS models. Benhamadouche and Laurence (2003) obtained satisfactory
results for both velocity and Reynolds stress proles with a LES and a Reynolds
stress model on a ne and a coarse three-dimensional mesh for a rotated square
array with spacing ratio P/D = 1.40 at Reynolds number 9000. The Reynolds
number is based on the bulk velocity in the gaps. Results computed with a
RSM on a two-dimensional mesh show strong vortex shedding and severely
overestimate the normal Reynolds stresses, because the ow structures are highly
three-dimensional. Beale and Spalding (1999) imposed spanwise perturbations
to compute a two-dimensional transient periodic laminar ow in in-line square
and rotated square arrays with spacing ratio P/D = 2 at Reynolds numbers
30 Re 3000. Derived Strouhal numbers are in agreement with experimental
data. Hassan and Barsamian (2004) performed a LES of the ow around a full
three-dimensional non-symmetric bundle consisting of ve rows of 2 cylinders
with a spacing ratio P/D = 2.07 and an angle = 45
at Reynolds number 21700.

Cylinders of the last row have larger wakes and highest pressures occur at the
side of the bundle. No distinct peaks are observed in the power spectral density
of the lift and drag force. Large scale three-dimensional eects are absent at the
midplane.
In this section, the two-dimensional ow around groups with a spacing ratio P/D =
1.05 are computed in the post-critical regime (Reynolds number 12.4 10
6
) using
the same turbulence model as in the previous section. This spacing ratio is smaller
than the ratios studied in the literature. Experiments and simulations mainly treat
the laminar and the subcritical regime. The RANS model is not ideal, but predicts
the velocity proles quite well and has a reasonable computational cost.
The ow around groups of cylinders is strongly three-dimensional with probably
horseshoe vortices, trailing vortices and arch vortices around the group as a whole.
In a three-dimensional computation a lot of the uid will pass left, right and
above the group, while in a two-dimensional computation the uid is only allowed
to pass left or right and a larger amount of uid is forced to pass through the
small gaps. Therefore, in the two-dimensional computations higher pressures will
probably be found in the gaps than in the case of a three-dimensional computation.
The velocity gradients of an atmospheric boundary layer cannot be considered in
a two-dimensional computation. The turbulent kinetic energy and the turbulent
energy dissipation rate at the inlet are equal to the low values used for the case of
a single cylinder. All these important assumptions complicate the extrapolation of
the results of the two-dimensional computations to groups of cylindrical structures
located in an atmospheric boundary layer.
Two congurations are studied: rst the less comprehensive group of 2 by 2
cylinders and then a group of 8 by 5 cylinders, as in the port of Antwerp.
The incidence angle (gure 4.12) between the wind ow direction and the
cylinder group is equal to 30
for both congurations, corresponding to the

predominant wind direction at the site in Antwerp. During the storm on October
27, 2002, the wind direction (west-southwest) corresponded to this incidence angle.
Computations for other incidence angles should be performed as well, but are
skipped as they are completely analogous and just a matter of computation time.
The lateral boundaries in the model are located at a distance of 9 times the width
Lsin(30
) + W cos(30
) of the group, the inlet at a distance of 9 times the depth

Lcos(30
) + W sin(30
) of the group and the outlet at a distance of 30 times the

depth of the group (gure 1.4). A smaller problem domain would be unlikely to
inuence the results near the cylinders, but nevertheless the number of elements
in the far eld is anyhow negligible with respect to the number of elements next
to the group.
4.5.1 Group of 2 by 2 cylinders
Figure 4.12 shows a detail of the mesh A near the group of 2 by 2 cylinders.
A structured mesh consisting of 280 elements around the circumference and 30
elements in the radial direction is used around each cylinder. Mesh A consists of
201184 elements and 243348 nodes. The four cylinders are numbered 1 to 4 as
shown in the gure. The angle is dened independently of the wind direction.
q
A
E
B
D
F
C
1
3
2
4
a
v
f
Figure 4.12: Detail of the mesh for the group of 2 by 2 cylinders.
The transient solution is integrated with a dimensionless time step v t/D = 0.029.
Maximum 10 iterations are performed within every time step in order to reduce
the RMS of the normalized residuals to 10
6
. 2965 time steps are computed, which
results in a time window of 14.820 s.
Figure 4.13a shows the dimensionless distance y
+
of the nodes next to the cylinder
wall as a function of the angle for the four cylinders at time t = 14.820 s . Most
points lie within the region where the logarithmic law is valid.
The ow is also computed on a ner mesh B, where the nodes next to the cylinder
wall lie within the viscous sublayer (gure 4.13b). The structured mesh around
each cylinder consists of 800 elements around the circumference and 90 elements
in the radial direction. Mesh B is made up of 511444 elements and 810088 nodes.
1933 time steps are computed, which results in a time window of 9.665 s. Figure
4.14 compares the time averaged pressure coecients on meshes A and B as a
0 90 180 270 360
0
50
100
150
200
Angle []
D
i
m
e
n
s
i
o
n
l
e
s
s

d
i
s
t
a
n
c
e

[
]
(a)
0 90 180 270 360
0
0.5
1
1.5
2
2.5
Angle []
D
i
m
e
n
s
i
o
n
l
e
s
s

d
i
s
t
a
n
c
e

[
]
(b)
Figure 4.13: Dimensionless distance y
+
of the nodes next to the cylinder wall as
a function of the angle for cylinder 1 (solid line), 2 (dashed line), 3
(dotted line) and 4 (dash-dotted line) of the group of 2 by 2 cylinders
in (a) mesh A and (b) mesh B.
0 90 180 270 360
3
2
1
0
1
Angle []
P
r
e
s
s
u
r
e

c
o
e
f
f
i
c
i
e
n
t

[
]
(a)
0 90 180 270 360
3
2
1
0
1
Angle []
P
r
e
s
s
u
r
e

c
o
e
f
f
i
c
i
e
n
t

[
]
(c)
0 90 180 270 360
3
2
1
0
1
Angle []
P
r
e
s
s
u
r
e

c
o
e
f
f
i
c
i
e
n
t

[
]
(b)
0 90 180 270 360
3
2
1
0
1
Angle []
P
r
e
s
s
u
r
e

c
o
e
f
f
i
c
i
e
n
t

[
]
(d)
Figure 4.14: Time average (thick line), maximum and minimum (thin lines) of
the pressure coecient C
p
for mesh A (black) and mesh B (grey)
as a function of the angle for (a) cylinder 1, (b) cylinder 2, (c)
cylinder 3 and (d) cylinder 4 of the group of 2 by 2 cylinders.
function of the angle for the four cylinders. The minima and maxima are
also indicated. For the mesh B, the time average is computed from t = 5.170 s
to t = 9.665 s. As only small dierences between both meshes are noticeable,
most pronounced at the small gaps and around the separation points, the results
obtained with mesh A will be discussed further.
Figures 4.15 and 4.16 show the time history and the frequency content of the
pressure between t = 7.385 s and t = 14.820 s at a small gap (B) and in the middle
of the group (E), as indicated in gure 4.12. The largest variation of the pressure
occurs at the small gaps and at the separation points. The dominant peak in the
frequency content at 0.67 Hz (gures 4.15b and 4.16b) corresponds to a Strouhal
number of 0.28, using the projected width 13.42 m of the group as a characteristic
length. In the middle of the group and in the wake, higher harmonics are clearly
present in the pressure time history. From the third peak in the frequency content
at 2.03 Hz (gure 4.16b), a Strouhal number of 0.35 is derived using the cylinder
diameter as a characteristic length. Time variations of the pressure are generally
larger at the leeward side of the cylinder group (gure 4.14). At the stagnation
points, the pressure is constant in time.
8 9 10 11 12 13 14
1900
1800
1700
1600
1500
Time [s]
P
r
e
s
s
u
r
e

[
P
a
]
(a)
0 1 2 3 4 5
0
50
100
150
200
Frequency [Hz]
P
r
e
s
s
u
r
e

[
P
a
/
H
z
]
(b)
cylinders surface in the point B of the group of 2 by 2 cylinders.
Figure 4.17 shows the streamlines at t = 14.075 s during the last vortex shedding
period. The outer boundary layer F1 of cylinder 1 separates at 320
322
,
but reattaches at 333
334
. At = 12
the boundary layer separates

again and reattaches to cylinder 4. The gap ow F1 separates from cylinder 2 at
= 78
and is deected towards the downstream cylinder 4. It nally separates

from cylinder 4 at 54
66
. This yields a very wide recirculation zone W1

behind cylinder 2, where vortices shed at a lower frequency alternately from this
deected gap ow F1 and the outer shear layer separated at 324
326
from
cylinder 2. This ow pattern upstream and between cylinders 1 and 2 is similar to
8 9 10 11 12 13 14
400
300
200
100
0
Time [s]
P
r
e
s
s
u
r
e

[
P
a
]
(a)
0 1 2 3 4 5
0
50
100
150
200
Frequency [Hz]
P
r
e
s
s
u
r
e

[
P
a
/
H
z
]
(b)
cylinders surface in the point E of the group of 2 by 2 cylinders.
F2
W1
W3
W4
F1
W2
1
2
4
3
q
Figure 4.17: Streamlines at t = 14.075 s from a transient computation around a
group of 2 by 2 cylinders.
the patterns described by Alam et al. (2005) for = 25
and P/D = 1.3 (gure

4.18a) and for = 45
and P/D = 1.1 (gure 4.18b). The wake behaviour is

clearly dierent due to the presence of the two other cylinders.
The inner boundary layer F2 of cylinder 1 separates at = 78
and reattaches
to cylinder 4. The gap ow F2 separates from cylinder 3 at = 12
and is
deected towards the downstream cylinder 4. It nally separates from cylinder 4
at 95
107
. The recirculation zone W2 behind cylinder 3 is constrained

between this gap ow F2 and the outer shear layer separated at 100
102
from cylinder 3.
Both deected gap ows F1 and F2 constrain the spatial development of the
recirculation zone W3 behind cylinder 4, yielding a very small recirculation zone.
(a)
(b)
Figure 4.18: Flow patterns around two closely spaced cylinders for (a) = 25
and P/D = 1.3 and (b) = 45
and P/D = 1.1 (Alam et al., 2005).

The recirculation zone W4 behind cylinder 1 is highly constrained between the
gap ows F1 and F2 attached to cylinders 2 and 3 and is biased away from the
ow axis. The ow pattern upstream and between cylinders 1 and 3 is similar to
the base bleed pattern (Sumner et al., 2000) and pattern III
B
(Gu and Sun, 1999).
1 3
2 4
Figure 4.19: Time averaged pressure coecients C
p
for the ow around a group of
2 by 2 cylinders (solid line) and for the ow around a single cylinder
(dashed line). The arrow indicates the incidence angle .
Figure 4.19 shows the time averaged pressure coecients for the ow around the
group compared with the pressure coecient for the ow around a single cylinder.
The time average is computed from t = 7.385 s to t = 14.820 s. The group
conguration drastically changes the pressure distribution around the cylinders.
The stagnation points of cylinders 1 and 3 are shifted towards each other, while
the stagnation point of cylinder 2 is shifted away from cylinder 1. Similar shifts are
measured by Lam and Fang (1995) at Reynolds number 12.810
3
for P/D = 1.26
and angle = 30
. The positive pressure area on cylinder 1 is largely extended.

Cylinder 4 is submerged in the wake of the three other cylinders, yielding suction
over the whole circumference. The reattachment of the gap ows to cylinder 4
produces the maxima in its pressure distribution. The largest suction occurs where
the outer boundary layers separate from cylinders 2 and 3. The large velocities
in the gaps produce distinct minima in the pressure distributions where the gap
area is minimum. The results by Lam and Fang (1995) for P/D = 1.26 also show
distinct minima at some gaps, but not at all gaps. The larger spacing ratio and the
lower Reynolds number of the experiments might explain these dierences. The
base pressure of cylinder 2, 3 and 4 is decreased, compared with a single cylinder.
Figure 4.20 shows the time averaged and uctuating drag and lift coecients for
the ow around the group. Cylinder 1, at the head of the group, experiences a
higher time averaged drag coecient C
d
in comparison with the ow around a
single cylinder (0.35 C
d
0.36) due to the extended area of positive pressure
and the large suction in the gaps. Cylinder 4 is completely immersed in the wake
of the three other cylinders and is pulled forward towards these cylinders by the
large suction in the gaps, resulting in a negative drag coecient. Negative drag
was measured for a group of 4 cylinders by Lam and Fang (1995) for P/D = 1.26
and angle = 0
. In two-cylinder congurations, Sumner et al. (2005) observed

negative drag coecients for P/D = 1.125 and angle 0
15
and Gu (1996)
did at supercritical Reynolds number for P/D = 1.2 and angle 10
45
.

0.98
0.18
0.51
0.18
0.01
0.08
0.06
0.02
0.61
0.56
0.68
0.27
0.04
0.07
0.10
0.19
(a) (b) (c) (d)
Figure 4.20: (a) Time averaged drag coecient C
d
, (b) uctuating drag coecient
C
d
l
and (d) uctuating lift
coecient C
l
for the ow around the group of 2 by 2 cylinders. The
arrow indicates the incidence angle .
Time averaged lift coecients C
l
are quite high (gure 4.20c). As the inner part of
cylinders 2 and 3 are immersed in the wake of cylinder 1 and large suction occurs
where the outer boundary layer separates from these cylinders, they are pulled
away from the group. For cylinder 1, the large suction around = 280
and the
extended area of positive pressure yield a negative lift coecient. At P/D = 1.26,
Lam and Fang (1995) report attractive forces between cylinders 2 and 3 for an
angle = 30
, but repulsive forces at = 15
. Sumner et al. (2005) measured

attractive forces between two staggered cylinders with P/D = 1.125 in the range
0
25
and repulsive forces in the range 60
90
. Fluctuations on
cylinder 1 are small, while cylinder 4 in the wake experiences a larger uctuating
lift coecient.
The time averaged pressure coecient is decomposed into a series of cosine
functions with circumferential wavenumber n, corresponding to the mode shapes
of an axisymmetric structure (gure 3.14):
C
p
=
n=0
C
n
p
cos(n +
n
) (4.60)
The eigenmodes with n = 3 or n = 4 have the highest contribution to the response
of the silos under wind loading (Dooms et al., 2003).
Figure 4.21 shows the amplitudes C
n
p
for the ow around the group for n =
0 10, which represent the major part of the variation of the pressure coecient,
compared with the amplitudes for the ow around a single cylinder. For all values
of n, except n = 2, larger amplitudes C
n
p
than for a single cylinder occur for one
or more cylinders in the group. For the cylinders on the transverse corners of the
group, the amplitudes C
n
p
for n = 3 and n = 4 are larger than the amplitudes for
a single cylinder.
0
0.5
1
1.5
n = 0
n = 1 n = 2 n = 3 n = 4 n = 5
n = 6 n = 7 n = 8 n = 9 n = 10
Figure 4.21: Decomposition of the time averaged pressure coecient C
p
into a
ow around the group of 2 by 2 cylinders (2 by 2 circles) and for the
ow around a single cylinder (single circle). The arrow indicates the
incidence angle .
4.5.2 Group of 8 by 5 cylinders
Figure 4.22 shows a detail of the mesh near the group of 8 by 5 cylinders. The
same structured mesh consisting of 280 elements around the circumference and 30
elements in the radial direction is used around each cylinder. The mesh consists
of 566866 elements and 921438 nodes. The cylinders are numbered as shown in
the gure.
1
2
3
4
5
6
7
8
16
15
14
13
12
11
10
9
24
23
22
21
20
19
18
17
32
31
30
29
28
27
26
25
33
34
35
36
37
38
39
40
v
f
a
E
F
D
A
C
B
Figure 4.22: Detail of the mesh for the group of 8 by 5 cylinders.
The transient solution is integrated with a dimensionless time step v t/D = 0.058.
A maximum of 10 iterations are performed within every time step in order to reduce
the RMS of the normalized residuals to 10
6
. 5170 time steps are computed, which
results in a time window of 51.70 s.
0 90 180 270 360
0
50
100
150
200
Angle[]
D
i
m
e
n
s
i
o
n
l
e
s
s

d
i
s
t
a
n
c
e

[
]
Figure 4.23: Dimensionless distance y
+
of the nodes next to the cylinder wall as
a function of the angle for all cylinders of the group of 8 by 5
cylinders.
Figure 4.23 shows for the forty cylinders the dimensionless distance y
+
of the nodes
next to the cylinder wall as a function of the angle at time t = 51.70 s. Most
points lie within the region where the logarithmic law is valid.
Figures 4.24 and 4.25 show the time history and the frequency content between t =
33.48 s and t = 51.70 s of the pressure at the cylinders surface at a separation point
(B) and at a small gap (E), as indicated in gure 4.22. At the stagnation points,
the pressure is constant in time. Generally, the pressure variation on cylinders
at the windward side primarily has components at approximately 0.165 Hz (gure
4.24b), which corresponds to a Strouhal number of 0.24, using the projected width
45.89 m of the group as a characteristic length. In the middle of the group and
in the wake, the pressure spectrum has also a signicant peak at 2.85 Hz (gure
4.25b). A vortex shedding frequency of 2.85 Hz corresponds to a Strouhal number
of 0.49 using the cylinder diameter as a characteristic length. This is consistent
with higher Strouhal numbers measured in closely spaced tube arrays (Blevins,
1990). Generally, time variations of the pressure are larger at the leeward side of
the cylinder group.
35 40 45 50
2350
2300
2250
2200
2150
2100
Time [s]
P
r
e
s
s
u
r
e

[
P
a
]
(a)
0 1 2 3 4 5
0
5
10
15
20
25
30
35
40
Frequency [Hz]
P
r
e
s
s
u
r
e

[
P
a
/
H
z
]
(b)
cylinders surface in the point B of the group of 8 by 5 cylinders.
35 40 45 50
850
800
750
700
650
600
Time [s]
P
r
e
s
s
u
r
e

[
P
a
]
(a)
0 1 2 3 4 5
0
5
10
15
20
25
30
35
40
Frequency [Hz]
P
r
e
s
s
u
r
e

[
P
a
/
H
z
]
(b)
cylinders surface in the point E of the group of 8 by 5 cylinders.
Figure 4.26 shows the streamlines at t = 50.94 s during the last vortex shedding
period. The outer boundary layers of cylinders 1-7 separate around = 313
and reattach around = 341
before entering the gaps. The gap ows between

the cylinders have a sinusoidal pattern which makes an angle of 15
with the ow
direction. When leaving the group, most of the gap ows are deected towards the
downstream cylinders and join up with other deected gap ows. These unied
gap ows nally separate from cylinder 40. The recirculation zone behind most of
the cylinders on the leeward borders of the group are strongly constrained between
the dierent gap ows and biased away from the ow axis. The outer shear layer
7
8
6
5
4
3
2
1
16
13
14
15
12
11
24
23
22
21
20
19
18
17
10
28
25
26
27
30
29
9
33
34
35
37
38
32
31
39
40
36
W2
W3
W1
Figure 4.26: Streamlines at t = 50.94 s from a transient computation around a
group of 8 by 5 cylinders.
separated from cylinder 8 yields a very wide recirculation zone W1 behind this
cylinder, where vortices shed at a low frequency alternately from this separated
boundary layer and the unied gap ow separated from cylinder 40. The gap ow
between cylinders 34 and 35 is deected towards the upstream cylinder, yielding a
second very large recirculation zone W2 behind cylinder 35. The outer shear layer
separated from cylinder 33 and the gap ow between cylinders 33 and 34 form a
third recirculation zone W3 behind cylinder 33. The recirculation zones behind
cylinders within the group are highly constrained between the gap ows attached
to neighbouring cylinders and extend up to the downstream cylinder.
Figure 4.27 shows the time averaged pressure coecients for the ow around
the group compared with the pressure coecient for the ow around a single
cylinder. The time average is computed from t = 33.48 s to t = 51.70 s. The group
conguration drastically changes the pressure distribution around the cylinders.
The stagnation points are shifted towards the outer side for cylinders 2 to 8.
The positive pressure area on these cylinders is reduced. For cylinders 9, 17,
25 and 33, the stagnation points are shifted towards the upstream cylinder. The
positive pressure area on cylinder 1, 9 and 17 is largely extended. The minimum
1 9 17 25 33
2 10 18 26 34
3 11 19 27 35
4 12 20 28 36
5 13 21 29 37
6 14 22 30 38
7 15 23 31 39
8 16 24 32 40
Figure 4.27: Time averaged pressure coecients C
p
for the ow around a group of
8 by 5 cylinders (solid line) and for the ow around a single cylinder
(dashed line). The arrow indicates the incidence angle .
pressure coecient upstream of the boundary layer separation increases from
cylinder 1 to cylinder 8. On cylinders 2 to 7, after separation, a zone of constant
pressure is formed up to the reattachment. On cylinder 33, large suction is
produced upstream of the boundary layer separation. The large velocities in the
gaps produce distinct minima in the pressure distributions where the gap area
is minimum. All cylinders submerged in the wake of other cylinders have two
maxima in the pressure coecient at the locations where the gap ows reattach.
The base pressure generally decreases for cylinders located further downstream.
For some of the upstream cylinders, positive base pressures were computed.

0.27
0.05
0.05
0.03
0.03
0.06
0.16
0.20
0.61
0.16
0.16
0.13
0.13
0.13
0.13
0.27
0.67
0.19
0.16
0.14
0.15
0.15
0.15
0.34
0.74
0.21
0.15
0.09
0.14
0.14
0.12
0.22
0.13
0.16
0.15
0.01
0.19
0.17
0.14
0.02
0.01
0.01
0.00
0.00
0.00
0.00
0.01
0.01
0.00
0.00
0.00
0.00
0.00
0.01
0.01
0.01
0.01
0.00
0.00
0.01
0.00
0.01
0.01
0.02
0.01
0.00
0.01
0.02
0.03
0.03
0.01
0.03
0.01
0.00
0.00
0.00
0.02
0.03
0.03
0.02
(a) (b)

0.50
0.06
0.09
0.14
0.20
0.27
0.36
0.84
0.22
0.14
0.08
0.08
0.05
0.02
0.16
0.69
0.13
0.23
0.21
0.16
0.14
0.10
0.04
0.19
0.01
0.37
0.31
0.25
0.16
0.13
0.09
0.09
1.01
0.22
0.22
0.52
0.37
0.30
0.23
0.00
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.00
0.00
0.01
0.01
0.01
0.01
0.01
0.01
0.00
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.00
0.01
0.01
0.02
0.02
0.03
0.04
0.02
0.01
0.01
0.01
0.01
0.02
0.03
0.03
0.05
(c) (d)
Figure 4.28: (a) Time averaged drag coecient C
d
, (b) uctuating drag coecient
C
d
l
and (d) uctuating lift
coecient C
l
for the ow around the group of 8 by 5 cylinders. The
arrow indicates the incidence angle .
Figure 4.28 shows the time averaged and uctuating drag and lift coecients for
the ow around the group. At the windward side of the group, cylinders 9, 17
and 25 experience a higher time averaged drag coecient C
d
in comparison with
the ow around a single cylinder (0.35 C
d
0.36), while in the middle of the
group the drag coecients are considerably lower. At the two transverse corners
(cylinders 8 and 33) the drag coecient is also low. Figure 4.28c shows the time
averaged lift coecients C
l
. The cylinders on the borders of the group are pulled
away from the group, which results in relatively high lift coecients for these
cylinders, especially for the two on the transverse corners (cylinders 8 and 33).
Fluctuating drag and lift coecients (gures 4.28b and 4.28d) are generally small.
The largest uctuations occur at the leeward side of the group.
As for the 2 by 2 group, the time averaged pressure coecient is decomposed into
a series of cosine functions with circumferential wavenumber n. Figure 4.29 shows
the amplitude C
n
p
for the ow around the group for 0 n 10, compared with
the amplitudes for the ow around a single cylinder. For all values of n, except
n = 2, larger amplitudes C
n
p
than for the ow around a single cylinder occur for
one or more cylinders in the group. The highest values are always situated on
the borders of the group. Eigenmodes with circumferential wavenumber n equal
to a multiple of 4 are heavily loaded in the group conguration due to the four
small gaps surrounding most of the cylinders. Especially for the cylinders on the
0
0.5
1
1.5
n = 0
n = 1 n = 2 n = 3 n = 4 n = 5
n = 6 n = 7 n = 8 n = 9 n = 10
Figure 4.29: Decomposition of the time averaged pressure coecient C
p
into a
ow around the group of 8 by 5 cylinders (8 by 5 circles) and for the
ow around a single cylinder (single circle). The arrow indicates the
incidence angle .
CONCLUSION 151
transverse corners of the group, eigenmodes with n = 3 or n = 4, which for silos
typically have the lowest eigenfrequencies, are strongly excited.
The large values for the time averaged lift coecients and large amplitudes C
n
p
for n = 3 and n = 4 for the cylinders on the transverse corners of the group are
consistent with the observations that that storm damage is mainly located on silos
at the corners of the groups.
4.6 Conclusion
In this chapter, some aspects of the computation of turbulent wind ows around
civil engineering structures which are relevant to the work in this thesis, have
been addressed. First the current state-of-the-art methods for the simulation
of turbulent ows are reviewed as a function of their performance and their
computational requirements. For wind ows around buildings, which have a very
high Reynolds number, DNS and LES are too demanding from a computational
point of view. The application of hybrid RANS/LES methods is very promising
as they are able to capture the large scale turbulent structures in the wake of blu
bodies at a more aordable computational cost. However, for the use in coupled
uid-structure problems, this computational cost today is still very high and for
the coupled problem in the next chapter URANS simulations will be performed.
In order to clarify the inuence of the RANS turbulence models, the near-wall mesh
renement and the unsteadiness, rst the steady state two-dimensional turbulent
air ow around a single cylinder in the post-critical regime at a Reynolds number
12.4 10
6
is computed. Within Flotran, near-wall modelling strongly inuences
the results, while the results obtained with CFX are quite insensitive. Results
computed with various eddy viscosity turbulence models strongly dier, but the
minimum pressure coecient is always underestimated, while the base pressure
coecient is overestimated. For the ow around a circular cylinder at these
Reynolds numbers, the Shih-Zhu-Lumley turbulence model in Flotran has the
best correspondence with experimental results. Within CFX, the shear stress
transport turbulence model produces the best correspondence. Dierent results
are obtained with Flotran and CFX using the same turbulence models. A transient
simulation using the SST turbulence model captures the regular vortex shedding
in the post-critical regime and is therefore preferred.
As an application, the interference eects are studied for a group of 8 by 5
cylinders, which is a simplication of the silo group in the port of Antwerp. A
two-dimensional computation is performed which neglects the three-dimensional
character of the ow around the group and the velocity gradients of the
atmospheric boundary layer. Vortices shed at 0.165 Hz from the group as a whole
and at 2.85 Hz from the individual cylinders. These vortex shedding frequencies
are lower than the lowest eigenfrequency of the structure (3.93 Hz). The large
velocities in the gaps produce distinct minima in the pressure distributions. The
two cylinders on the transverse corners of the group experience quite high lift
coecients. Eigenmodes with circumferential wavenumber n equal to a multiple
of 4 are more heavily loaded in the group conguration, where four small gaps
surround most of the cylinders, compared with the single cylinder conguration.
Especially for the cylinders on the transverse corners of the group, eigenmodes
with n = 3 or n = 4, which for silos typically have the lowest eigenfrequencies, are
strongly excited. This explains why storm damage is mainly located on silos at
the corners of the group.
Chapter 5
Coupled simulation of wind
loading on structures
5.1 Introduction
The behaviour under wind loading of civil engineering structures that are quite
exible, have a complex geometry or are located in the presence of neighbouring
structures is very dicult to determine. The aim of this chapter is to study coupled
numerical simulations of the uid and the structure to determine this behaviour.
First an extensive and coherent overviewof the state-of-the-art coupling algorithms
and load and motion transfer methods is given. As a practical application, the
interaction between the wind ow and a single silo as in the port of Antwerp is
computed.
In the case of large deformations of structures due to uid loads, the structural
response is primarily of interest. The most straightforward way would be to
eliminate the uid eld and only retain the structural degrees of freedom. Because
the Navier-Stokes equations which represent the uid eld are non-linear this
is impossible. If the structure undergoes large deformations and this non-linear
behaviour of the structure has to be incorporated, direct time integration methods
will be needed for both elds.
This non-linear transient aeroelastic problem is formulated as a three eld problem:
the uid, the structure and the deforming mesh. The governing equations of the
dierent elds were derived in the previous chapters. In the uid eld, the Navier-
Stokes equations (2.59- 2.76) are formulated in an Arbitrary Lagrangian Eulerian
153
154 COUPLED SIMULATION OF WIND LOADING ON STRUCTURES
description:
J
f
t
+

J (
f
c) = 0 (5.1)
J
f
v
t
+

J (
f
v c) =

J
f
+

J
f
b
f
(5.2)
The governing equations of the structure are formulated in a Lagrangian
description (2.79):
2
u
t
2
X
=
s
+
s
b
s
(5.3)
For the deformation of the uid mesh a diusion equation with a variable diusivity
is used (2.233):
(k u) = 0 (5.4)
The computation of the uid ow around a deforming structure is a surface coupled
problem: the uid and the structure only interact at the uid-structure interface
. At the interface the following boundary conditions should be satised: the
traction on the surface of the structure should be in equilibrium with the traction
on the surface of the uid:
t
s
+t
f
= 0 (5.5)
s
n
s
+
f
n
f
= 0 (5.6)
s
n
s
= p
f
n
f

f
n
f
(5.7)
The velocity of the structure u should be equal to the uid velocity v. At the
uid-structure interface a Lagrangian description is used in the uid domain, so
the uid velocity v is equal to the mesh velocity v:
u = v = v (5.8)
The displacements of the structure u are equal to the mesh displacements u:
u = u (5.9)
First the partitioned algorithms to solve the coupled problem by computing the
dierent elds sequentially or in parallel are discussed. Section 5.3 treats the
transfer of displacements and loads between non-matching discretizations of the
uid and the structure eld. In the last section the coupled computation of ovalling
oscillations of silos is studied.
COUPLING ALGORITHMS 155
5.2 Coupling algorithms
A number of techniques exist to solve coupled eld problems with direct time
integration. A short review with the advantages and disadvantages of all the
approaches will be given.
5.2.1 The monolithic and the partitioned approach
Generally two approaches exist to solve the coupled problem: the monolithic and
the partitioned approach. In the monolithic approach both elds are joined into
one fully coupled system and are simultaneously advanced in time. In partitioned
approaches (Felippa and Park, 1980; Felippa et al., 2001) the coupled system is
spatially decomposed into partitions which coincide with the physical elds (e.g.
uid and structure). The partitions are treated as isolated entities and separately
integrated in time. The interaction eects are applied on the individual partitions
as external boundary conditions which are exchanged from one partition to the
other.
The monolithic approach
As in the monolithic approach both elds are solved simultaneously, the coupling
conditions at the uid-structure interface are exactly satised. This improves
greatly the accuracy and stability of the computations in strongly coupled
problems, where the behaviour of the coupled system diers largely from the
behaviour of the individual elds.
A rst drawback is that the number of equations to be solved at the same time
is equal to the sum of the uid and the structural degrees of freedom. Moreover
a completely new implementation is required to assemble the global matrix and
to solve the resulting system. The global matrix is composed of submatrices with
a dierent topology and dierent numerical properties (e.g. for nite element
discretizations the stiness matrix of the structure is symmetric while the stiness
matrix of the uid is nonsymmetric) and might be ill-conditioned. An appropriate
and optimal iterative solver is dicult to nd. Although dierent time scales
might be present in the individual elds, the monolithic system is advanced with
a single time step which should be small enough to compute accurately and stably
both elds.
An inherent diculty is that the structural equations involve second order time-
derivatives, whereas the uid equations involve only rst order derivatives. Taking
both the structural velocities and displacements as degrees of freedom (Hron and
T urek, 2006), the second order dierential equations can be recast in rst order
form, but this doubles the number of degrees of freedom. Additional equations
impose the relation between the structural velocities and displacements:
u
t
= u (5.10)
The resulting coupled system of rst order dierential equations is integrated in
time with the same method and the same time step for both elds. Hron and
T urek (2006) used the Crank-Nicolson time integration method to compute the
coupled behaviour of an incompressible uid and an incompressible solid. The
incompressible solid is described with a mixed up formulation. Both elds have
displacement, velocity and pressure degrees of freedom and are discretized with
Q2P1 nite elements.
H ubner et al. (2004) used for the structure a mixed formulation with the structural
velocities and second Piola-Kirchho stresses as variables. This yields two
partial dierential equations with rst order time-derivatives. After the static
condensation on element level of these stresses, only the structural velocities
and the boundary tractions remain as degrees of freedom. The coupled system
is subsequently discretized with time-discontinuous stabilized space-time nite
elements for both elds.
Rugonyi and Bathe (2001) and Zhang et al. (2003) simply use the structural
displacements and the uid velocities and pressures as degrees of freedom. After
space and time discretization, the degrees of freedom of the structure are
subdivided into displacements u
at the uid-structure interface and the remaining

displacements u
in the interior of the domain. Similarly, the velocity degrees of

freedom of the uid are divided into v
at the interface and v
in the interior of
the domain. At the uid-structure interface the velocities v
are expressed as a
function of the displacements u
. If the movement of the interior nodes of the uid

is omitted for ease of notation, the discretized coupled system might be expressed
as:
_
_
K
f
K
f
p
K
f
0
K
f
p
K
f
pp
K
f
p
0
K
f
K
f
p
K
f
+K
s
K
s
0 0 K
s
K
s
_
_
_
v
p
u
_
=
_
_
f
f
f
fp
f
f
+f
s
f
s
_
_
(5.11)
This monolithic approach is implemented as well in the nite element program
Adina.
The partitioned approach
In partitioned approaches (Felippa and Park, 1980; Felippa et al., 2001) the coupled
system is spatially decomposed into partitions which coincide with the physical
elds (e.g. uid and structure). The partitions are treated as isolated entities and
separately integrated in time. The interaction eects are applied on the individual
partitions as external boundary conditions which are exchanged from one partition
to the other.
As an example the most straightforward partitioned algorithm for uid-structure
interaction is described. Suppose the solution of the coupled problem is known
at t
n
. First, the structural displacements u
n
of the uid-structure interface are

imposed as boundary conditions on the uid eld and the uid eld is advanced
to time level t
n+1
:
_
K
f
K
f
p
K
f
p
K
f
pp
_ _
v
n+1
p
n+1
_
=
_
f
f
K
f
u
n
f
fp
K
f
p
u
n
_
(5.12)
This yields the new solution of the uid eld v
n+1
and p
n+1
. The forces exerted
by the uid on the uid-structure interface are applied to the structure and the
structure is advanced to time level t
n+1
:
_
K
s
K
s
K
s
K
s
_ _
u
n+1
u
n+1
_
=
_
f
s
K
f
v
n+1
K
f
p
p
n+1
K
f
u
n
f
s
_
(5.13)
The structural displacements u
n+1
of the uid-structure interface and u

n+1
of the
rest of the structure are obtained. The systems (5.12) and (5.13) computed in this
example can be compared to the coupled system (5.11).
A rst advantage of partitioned approaches is that for each eld well-established
discretization techniques and solution algorithms can be used which are tailored to
the dierent behaviour of the individual elds. Dierent time integration schemes
account for the specic requirements of the dierent elds: an implicit or explicit
method might be chosen; possibly dierent time steps are used to advance the uid
and structure, what is called subcycling. Finally dierent solvers can be applied
which take the characteristics and the conditioning of the resulting matrices into
account.
Already available software for the individual elds can be reused. This software
is validated and highly adapted to the specic applications. A modular
implementation of the partitioned procedure should enable that the software used
for one of the elds is easily replaced by another program with dierent capabilities.
Adaptations to the models and techniques applied in one of the partitions can be
implemented and validated in the respective software independent of the coupling
procedure and the other elds.
The main disadvantage of partitioned approaches is that the coupling conditions
(5.8) might not exactly be fullled. In the example the structural displacements u
n
of the uid-structure interface applied to obtain the uid solution at the new time
level t
n+1
dier from the the structural displacements u
n+1
of the uid-structure
interface at this time level. Generally, the violation of the coupling conditions
might deteriorate the stability and the accuracy of the coupled computation. The
challenge is to design ecient partitioned procedures which do not degrade the
stability and the accuracy of the methods applied in the individual partitions.
The computation of the solutions in the individual partitions, which are of smaller
size than the coupled problem, requires less computational time and memory.
The eciency of the coupling procedure will determine if overall the partitioned
procedure is more ecient than the monolithic approach.
In the remainder of this section the partitioned algorithms are further elaborated.
First loosely coupled and then strongly coupled algorithms are treated.
5.2.2 Loosely coupled algorithms
In loosely coupled algorithms the coupling conditions (5.8) at the uid-structure
interface are not exactly fullled. Each eld is solved one or a few times in each
time step.
Method 1: conventional serial staggered algorithm
The conventional serial staggered algorithm (gure 5.1) was described in the
previous section. It consists of the following steps:
1. Transfer the structural displacements u
n
of the uid-structure interface to

the uid mesh
2. Compute the mesh deformation at time t
n+1
3. Transfer the mesh position x
n+1
to the uid
4. Advance the uid eld to time level t
n+1
5. Transfer the uid forces f
n+1
f
on the uid-structure interface to the structure
6. Advance the structure to time level t
n+1
In this algorithm each eld is solved once per time step step. The boundary
conditions are exchanged once per time step. Because the two elds are solved
sequentially, this is called a staggered method.
u
n
u
n+1
v
n
v
n+1
x
n
x
n+1
Mesh
Fluid
Structure
t
1
2
3
4
5
6
Figure 5.1: The serial staggered algorithm.
The displacements at the uid-structure interface at t
n
are used to compute
the uid ow at t
n+1
. This introduces explicitness in the coupling procedure,
irrespective of the fact that the individual elds are integrated in time with an
explicit or implicit scheme. Consequently, a restriction on the time step is expected,
even if both individual elds are integrated in time with an unconditionally stable
method.
As the displacements imposed on the uid eld at t
n+1
dier from the
displacements computed subsequently in the structure, the kinematical continuity
is not fullled. Consequently the conservation of mass, impulse and energy is
not guaranteed at the interface. This lack of conservation at the uid-structure
interface causes a loss in time-accuracy and numerical stability. Even if second-
order accurate time integration schemes are used for the uid and the structure,
the coupled computation is only rst-order time-accurate. As the objective is to
compute physical instabilities as utter, galloping and ovalling, the stability of
the coupled procedure is very important as weak numerical instabilities might be
confused with physical instabilities.
The above described approach is called Dirichlet-Neumann coupling: on one
partition Dirichlet boundary conditions are imposed, while on the other partition
Neumann boundary conditions are applied. Depending on the order in which the
partitions are solved and on which boundary condition is imposed to which eld,
four dierent coupling procedures exist.
Commonly, the uid eld is chosen as Dirichlet partition and has prescribed
velocities at the uid-structure interface, while the structure is the Neumann
partition with imposed traction on the interface. The uid is solved rst and
the structure secondly.
Method 2: enhanced serial staggered algorithms
A rst way to improve the conventional serial staggered algorithm is to transfer a
prediction u
n+1
p
of the structural displacements at t
n+1
to the uid mesh in the
rst step of the algorithm:
u
n+1
p
= u
n
+
0
t u
n
+
1
t( u
n
u
n1
) (5.14)
The trivial prediction, which is equal to the value computed at t
n
, is recovered
with
0
=
1
= 0. A rst order accurate prediction is obtained with
0
= 1 and
1
= 0 and a second order accurate with
0
= 1 and
1
=
1
2
.
Next, a corrected uid force f
n+1
fc
is transferred to the structure in the fth step
of the algorithm. This corrected value is for instance chosen so as to minimize the
amount of momentum and energy that is articially created at the uid-structure
interface (Piperno and Farhat, 2001) or as to obtain a global order of time-accuracy
(Farhat et al., 2006).
If the trapezoidal rule is used for the structure, the momentum articially created
by the staggering at the interface is (Piperno, 1997):
M
s
+ M
f
=
_
f
n
fc
+f
n+1
fc
2

f
n+1
f
_
t (5.15)
f
n+1
f
depends on the time integrator that is used for the uid eld. For the
forward Euler, the backward Euler and second order time-accurate methods

f
n+1
f
is respectively equal to:
f
n+1
f
= f
n
f
(5.16)
f
n+1
f
= f
n+1
f
(5.17)
f
n+1
f
f
n
f
+f
n+1
f
2
(5.18)
From equation (5.15) follows that conservation of momentum is satised at the
interface if:
f
n+1
fc
= 2
f
n+1
f
f
n
fc
(5.19)
If a second order time-accurate method (5.18) is used in the uid eld, this reduces
to:
f
n+1
fc
= f
n+1
f
(5.20)
Another choice for the corrected uid force f
n+1
fc
is to transfer the forces computed
at t
n
:
f
n+1
fc
= f
n
f
(5.21)
This enables to compute the uid and the structure in parallel. A last possibility
is to transfer the average of the forces computed at t
n
and t
n+1
:
f
n+1
fc
=
f
n+1
f
+f
n
f
2
(5.22)
The loss of numerical stability and accuracy of loosely coupled methods is related
to the conservation of energy at the uid-structure interface. If the trapezoidal rule
is used for the structure, the energy that is articially created by the staggering
at the uid-structure interface is (Piperno and Farhat, 2001):
E
s
+ E
f
= (u
n+1
u
n
)
T
f
n
sc
+f
n+1
sc
2
( u
n+1
p
u
n
p
)
T
f
n+1
f
(5.23)
where u
n+1
p
are the predicted displacements after transfer from the structure to the
possibly non-matching uid mesh and f
n
sc
are the corrected forces after transfer
from the uid to the possibly non-matching structure mesh. As the predicted
displacements u
n+1
p
dier from the computed displacements u
n+1
which are not

yet known at the moment the corrected force f
n+1
fc
is transferred, generally the
energy cannot be conserved at the uid-structure interface. The predictor and
corrector can be adjusted in order to reduce the imbalance as much as possible.
Therefore Piperno and Farhat (2001) studied the accuracy of the conservation
of energy at the uid-structure interface for the simplied case of a structure
vibrating at a certain frequency with a constant amplitude. For the conventional
serial staggered algorithm (u
n+1
p
= u
n
and f
n+1
fc
= f
n+1
f
), the conservation of
energy is rst order accurate independent of the time integration scheme which is
used in the uid eld. With the rst order predictor the conservation of energy
of several algorithms becomes second order accurate, among others all algorithms
where the corrected uid force is given by (5.19) and which thus conserve the
momentum at the uid-structure interface. With the second order predictor the
conservation of energy of these momentum-conserving algorithms becomes third-
order accurate.
Farhat et al. (2006) developed a loosely-coupled globally second order time-
accurate algorithm. Therefore, they combined the second order time-accurate
three-point backward dierence scheme (2.226) which satises the DGCL, with
the second order predictor (5.14) and the latest computed uid force (5.20) as
corrector. This approach conserves the momentum at the uid-structure interface
and the conservation of energy at the interface is third-order accurate. Finally the
mesh deformation algorithm should be adapted: the common approach which uses
the stiness matrix corresponding to the predicted displacements u
n
p
to compute
the predicted displacements u
n+1
p
, can be viewed as forward Euler time integration
of the mesh motion equations (2.230). In order to obtain a globally second order
time-accurate algorithm, the stiness matrices of the mesh at both the time levels
t
n
and t
n1
are needed to compute the predicted displacements u
n+1
p
.
Method 3: non-collocated algorithms
At the uid-structure interface the displacements (5.9) and the velocities (5.8) of
the uid and the structure should be equal. The above partitioned procedures
try to enforce displacement continuity by transferring the predicted structural
displacements to the uid mesh. Suppose that the displacement continuity is
fullled. In the uid the mesh velocity is computed from the mesh positions based
on equation (2.215) in order to satisfy the geometric conservation law (section
2.4.2). If the structure is advanced with a second order accurate time integration
procedure, the structural velocity u
s
diers from the mesh velocity and velocity
continuity will be violated. This violation results in an error in the exchange of
kinetic energy between the uid and the structure.
Therefore Lesoinne and Farhat (1996) and Farhat and Lesoinne (2000) proposed a
non-collocated algorithm (gure 5.2) which satises both displacement and velocity
continuity and the GCL. In the above algorithms the solutions for the uid and
the structure are computed at the same time levels. In the non-collocated (or
asynchronous) algorithm, the uid eld is computed at half time levels t
n+
1
2
,
while the structure is computed at full time levels t
n+1
: the uid and structure
computations are oset by half a time step. The structure is advanced using the
midpoint rule.
1. Predict the structural displacements u
n+
1
2
p
of the uid-structure interface
based on the displacements u
n
and transfer them to the uid mesh

n+
1
2
n+
1
2
and mesh velocities v
n
to the uid
n+
1
2
n+
1
2
f
on the uid-structure interface to the structure
n+1
The rst order accurate predicted structural displacements u
n+
1
2
p
are given by:
u
n+
1
2
p
= u
n
+
t
2
u
n
(5.24)
u
n
u
n+1
v
n1/2
v
n+1/2
x
n1/2
x
n+1/2
Mesh
Fluid
Structure
t
1
2
3
4
5
6
Figure 5.2: Non-collocated algorithm.
If the structure is advanced using the midpoint rule, this predictor automatically
implies the velocity of the uid nodes v
n
at the uid-structure interface

corresponds to the velocity of the structure u
n
. If the trapezoidal rule is used

for the structure, the same result is obtained by transferring a corrected uid
force f
n+1
fc
to the structure in step 5:
f
n+1
fc
= 2f
n+
1
2
f
f
n
fc
(5.25)
For the trapezoidal rule the energy that is articially created by the staggering at
the uid-structure interface is (Piperno and Farhat, 2001):
E
s
+ E
f
= (u
n+1
u
n
)
T
f
n
sc
+f
n+1
sc
2
(5.26)
1
2
( u
n+
1
2
sp
u
n
1
2
sp
)
T
f
n+
1
2
f
1
2
( u
n+
3
2
sp
u
n+
1
2
sp
)
T
f
n+
3
2
f
(5.27)
where

f
n+
1
2
f
depends on the time integrator that is used for the uid eld and is
analogous to equations (5.17) to (5.18).
As for the enhanced serial staggered algorithms, the accuracy of the conservation of
energy at the uid-structure interface was studied by Piperno and Farhat (2001) for
the simplied case of a structure vibrating at a certain frequency with a constant
amplitude. For the corrected uid force (5.25) alternatives are analysed, which
are given by the following expressions:
f
n+1
fc
=

f
n+
1
2
f
(5.28)
f
n+1
fc
= 2
f
n+
1
2
f
f
n
fc
(5.29)
For most algorithms the conservation of energy is rst order accurate, although
for some it is second order accurate. The combination of the corrected uid force
(5.25) with a second order accurate time integration in the uid eld yields the
only algorithm where the conservation of energy is third-order accurate.
Based on this algorithm Farhat et al. (2006) developed a loosely-coupled globally
second order time-accurate algorithm. It consists of the combination of the second
order time-accurate three-point backward dierence scheme (2.226) which satises
the DGCL for the uid, the midpoint rule for the structure, the latest computed
uid force f
n+
1
2
f
as corrector and a second order accurate predictor:
u
n+
1
2
p
= u
n
+
t
2
u
n
+
t
8
_
u
n
u
n1
_
(5.30)
The stiness matrix of the mesh at time level t
n
should be used to compute the
predicted displacements u
n+
1
2
p
. As the mesh is unknown at this time level, the
mesh deformation problem becomes non-linear. Farhat et al. (2006) linearized
these equations around t
n
1
2
.
For aeroelastic problems in compressible ows, these improved algorithms
(enhanced serial staggered and non-collocated algorithms) enable accurate and
stable computations with time steps that are 5 to 10 times larger than the time step
used in the conventional staggered procedure. These time steps are comparable to
ones used in monolithic algorithms with implicit time integration schemes in both
uid and structure.
Method 4: parallel loosely-coupled algorithms
The basic parallel algorithm (gure 5.3) sends the displacements u
n
to the uid
and the uid forces f
n
f
to the structure and then advances both to t
n+1
in parallel.
Only at the beginning of the time step the coupling conditions are exchanged. This
algorithm is rst order time accurate, but half as accurate as the conventional serial
staggered algorithm. It requires small time steps to be stable and suciently
accurate.
An enhanced parallel algorithm (gure 5.4) which has an increased time-accuracy
is proposed by (Piperno et al., 1995; Farhat and Lesoinne, 2000). In this algorithm
the boundary conditions are exchanged twice per time step: rst the displacements
u
n
are sent to the uid and the uid forces f

n
f
to the structure. The uid is
advanced to t
n+
1
2
and the structure to t
n+1
. The structural displacement u
n+1
is used to advance the uid from t

n+
1
2
to t
n+1
. The uid force f
n+
1
2
f
is used to
advance the structure again from t
n
to to t
n+1
. Each eld is solved twice per time
step.
u
n
u
n+1
v
n
v
n+1
x
n
x
n+1
Mesh
Fluid
Structure
t
1
1
2
3
4
4
Figure 5.3: The parallel algorithm.
u
n
u
n+1
v
n
v
n+1/2
v
n+1
x
n
x
n+1/2
x
n+1
Mesh
Fluid
Structure
t
1
1
2
3
4
4
5
5
6
7
8
8
Figure 5.4: The enhanced parallel algorithm.
Method 5: subcycling
The uid and structure are often characterized by dierent time scales. Usually
the uid requires smaller time steps than the structure. This is especially the case
if an explicit time integration procedure is used for the uid, while an implicit
method is applied to the structure. The subcycling factor m is dened as the ratio
of the time step t
s
in the structure and the time step t
f
in the uid:
m =
t
s
t
f
(5.31)
Piperno et al. (1995) and Piperno (1997) presented loosely coupled algorithms
with subcycling of the uid (gure 5.5). Starting from the second order accurate
predicted structural displacements at t
n+1
(5.14) the grid deformation x
n+1
is
computed. During all the uid time steps within one coupling time step, the mesh
velocity is assumed constant and equal to:
v =
x
n+1
x
n
t
s
(5.32)
For each uid time step the grid positions are linearly interpolated between x
n
and
x
n+1
. Analogously to equation (5.19) conservation of momentum at the interface
is satised if the corrected uid force is given by:
f
n+1
fc
=
2
t
s
_
t
n+1
t
n
f
f
dt f
n
fc
(5.33)
Piperno and Farhat (2001) showed by means of a simplied case that the
conservation of energy of this algorithm is third-order accurate.
u
n
u
n+1
v
k
v
k+5
x
n
x
n+1
Mesh
Fluid
Structure
t
s
t
f
1
2
3
4
5
6
Figure 5.5: Staggered algorithm with subcycling of the uid.
Added-mass eect
Farhat et al. (2006) proved the second-order time accuracy of two loosely coupled
algorithms. The numerical stability these algorithms is not formally proved, but in
the case of aeroelastic computations with compressible ows all numerical results
suggest a very stable behaviour.
However when dealing with incompressible ows, dierent authors report stability
problems. The more time accurate the coupling algorithm is, the sooner the
instability manifests itself. Surprisingly decreasing the time step does not suppress
or postpone the instability, on the contrary the instability occurs earlier.
These instabilities are related to the added-mass eect and were studied by
Le Tallec and Mouro (2001), Causin et al. (2005) and Forster et al. (2007). As the
instability usually occurs in the rst time steps, the non-linear (convective) term
in the uid equations can be neglected. Because the instability occurs even when
very small time steps are used, the uid stiness (viscosity) can be neglected as
well. The resulting uid equations only consist of an acceleration and a pressure-
gradient term. If these equations are condensed on the uid-structure interface,
the following interface equation is obtained:
f
n+1
f
= m
f
M
A
v
(5.34)
where m
f
is a characteristic uid mass, for instance the nodal mass of a lumped
mass matrix. The added mass operator M
A
is a purely geometrical quantity. For
stabilized nite element discretizations the stabilization makes the added mass
operator M
A
dependent on the time step: its largest eigenvalue max
i
increases
with decreasing time steps (Forster et al., 2007).
As the instability occurs even when very small time steps are used, the structural
stiness can be neglected. The uid force f
n+1
f
is imposed on the structure. The
uid acceleration v
at the interface is expressed as a function of the predicted

structural displacements. After the eigenvalue decomposition of the added mass
operator, an stability condition is derived from the structural equations:
m
f
m
s
max
i
< C
stab
(5.35)
where m
s
is the structural nodal mass of a lumped mass matrix. max
i
is the
largest eigenvalue of the added mass operator M
A
. The stability constant C
stab
depends on the accuracy of the coupling algorithm: the accuracy of the uid time
integration and the accuracy of the structural predictor. The constant decreases
with increasing time-accuracy. As the largest eigenvalue max
i
increases with
decreasing time steps, the instabilities cannot be suppressed by decreasing the
time steps.
If the uid acceleration v
n+1
depends on the uid acceleration v

n
at the previous
time level, as for one-step methods (2.174) with ,= 1 (e.g. the Crank-Nicolson
scheme), the stability condition is even more restrictive:
m
f
m
s
max
i
<
C
stab
n
(5.36)
where n is the number of the time step. The condition becomes every time step
more restrictive and instabilities will occur from a certain time step irrespectively
of the values of the other parameters.
The factor
m
f
m
s
shows that the mass density ratio
f
/
s
is a determining factor
in the stability of loosely coupled partitioned procedures. Every loosely coupled
partitioned procedure has a limiting mass density ratio beyond which it becomes
unstable.
Table 5.1 gives the density ratio
s
/
f
between some typical civil engineering
materials and air or water. These density ratios indicate that stable loosely coupled
simulations for civil engineering structures in contact with water are not possible.
The interaction between structures and air might be computed with a loosely
coupled algorithm for the materials with the highest density, although it depends
on the time accuracy of the simulations and it is not possible to precisely predict
the limiting density ratio.
air water
1.25 1000
steel 7850 6280 7.85
aluminium 2700 2160 2.70
concrete 2500 2000 2.50
wood 290-900 232-720 0.29-0.90
tensile 80 64 0.08
Table 5.1: Density ratios
s
/
f
between some typical civil engineering materials
and air or water.
For larger time steps, the uid viscosity and the structural stiness cannot be
neglected. Generally, increasing the viscosity destabilizes the computations, while
increasing the structural stiness slightly improves the stability. For very sti
structures, the density ratio does not inuence the stability anymore.
The above conclusions are valid in the case the uid partition is solved rst with
imposed Dirichlet boundary conditions and the structure is computed secondly
with Neumann boundary conditions. Matthies et al. (2006) studied the three other
combinations of boundary conditions and order in which the elds are solved. If
the boundary conditions are kept the same, but the structure is solved rst and
the uid secondly, still for stability the density ratio
f
/
s
should be small enough.
However if the structure is computed with imposed displacements at the uid-
structure interface and the uid with imposed stresses, the density ratio
s
/
f
should be small enough, no matter in which sequence they are solved.
The stability problems are related with the equations used for incompressible ows.
Farhat (2006) computed the numerical examples of Causin et al. (2005) using the
discretized equations for compressible ows and did not encounter any stability
problems in this case.
5.2.3 Strongly coupled algorithms
In strongly coupled algorithms each eld is solved multiple times per time step in
order to fulll within a certain tolerance the coupling conditions (5.8) at the uid-
structure interface. This improves the accuracy and the numerical stability of the
coupled computations. The solution obtained with a strongly coupled algorithm
corresponds to the solution obtained with a monolithic algorithm.
Method 1: Iteratively staggered algorithms
A rst approach is to introduce iterations between the two partitions within one
time step. The rst solution for the uid and the structure at time level t
n+1
is obtained by means of a loosely-coupled algorithm. The newly computed value
for the displacement is subsequently transferred to the uid domain in order to
compute the solution for the same time step again. The obtained uid forces
are transferred to the structure and the structure is computed again. This
procedure is repeated until the change of the displacements or the forces at
the uid-structure interface is smaller than a certain tolerance. These iterations
are called subiterations, (Dirichlet-Neumann) inner iterations, predictor-corrector
iterations or intereld iterations. If these subiterations converge, they converge to
the solution obtained with the monolithic approach. At convergence, the kinematic
and dynamic continuity at the uid-structure interface are almost exactly fullled.
Consequently the conservation of mass, momentum and energy at the interface is
satised.
This method is sometimes called the block Gauss-Seidel approach in analogy to
the Gauss-Seidel method used to solve iteratively systems of equations. It uses
new interface results as soon as they are available and is a serial algorithm. The
convergence depends on the order in which the elds are solved.
By analogy to iterative Dirichlet-Neumann substructure methods or non-
stationary Richardson iteration methods for solving systems of equations, a
relaxation parameter
n+1(i)
can be introduced for the interface displacements:
u
n+1(i+1)
=
n+1(i)
u
n+1(i+1)
+ (1
n+1(i)
)u
n+1(i)
(5.37)
= u
n+1(i)
+
n+1(i)
( u
n+1(i+1)
u
n+1(i)
) (5.38)
The interface displacements u
n+1(i+1)
that will be transferred in the next

iteration to the uid, are a linear combination of the newly computed interface
displacements u
n+1(i+1)
in the structure and the interface displacements u

n+1(i)
transferred to the uid at the beginning of the iteration. The relaxation of interface
displacements should guaranty and accelerate the convergence of the subiterations.
The iteratively staggered algorithm consists of the following steps:
1. Predict the structural displacements u
n+1(0)

at time level t
n+1
2. Set i = 0
3. Transfer the structural displacements u
n+1(i)

to the uid mesh
n+1
n+1(i)
to the uid
n+1
n+1(i+1)
f
on the uid-structure interface to the
structure
n+1
9. Relax
n+1(i)
the interface displacements u
n+1(i+1)
10. Check the convergence for the displacements and the forces. If the
convergence criteria are fullled, stop. Otherwise set i = i + 1 and go to
step 3.
Figure 5.6 shows these steps for an iteration i.
u
n
u
n+1(i)
v
n
v
n+1(i)
x
n
x
n+1(i)
Mesh
Fluid
Structure
t
7
8
3
4
5
6
Figure 5.6: The iteratively staggered algorithm.
The convergence of the subiterations is reached if:
|
n+1(i+1)

n+1(i)
|
L
2
|
n+1(i+1)
|
L
2
<
(5.39)
where
are the interface displacements or forces.

This iteratively staggered approach with a xed relaxation factor is available in
most of the commercial packages e.g. Flotran, CFX and Adina.
The choice of the value of the relaxation factor is crucial but problematic: it is
often obtained with trial and error or based on empirical formulas. An optimal
relaxation factor results in a minimal number of subiterations.
The added mass eect, which causes the stability problems of loosely coupled
algorithms, also inuences the convergence of the subiterations for incompressible
ows. Generally, problems that are more likely to become unstable using loosely
coupled algorithms require a lower relaxation factor which results in higher number
of subiterations (Mok, 2001; Causin et al., 2005; Heck et al., 2006): in order to
obtain convergence the relaxation factor
n+1(i)
should decrease if the density
ratio
f
/
s
is increased, the time step is decreased or the stiness of the structure
is reduced. In addition the problems that are more likely to become unstable using
loosely coupled algorithms, are more sensitive to the value of the relaxation factor
n+1(i)
: a small variation away from the optimal value of the relaxation factor
increases the number of subiterations drastically or even leads to divergence of the
intereld iterations.
Furthermore the relaxation factor which is optimal for the rst subiterations and
time steps, is not necessarily optimal during the following subiterations and time
steps due to the non-linear behaviour of the coupled problem.
A solution is to automatically calculate the relaxation factors. Le Tallec and Mouro
(2001) used the steepest descent method to determine the relaxation factors. The
structure and uid eld have to be computed once with as only loading the residual
interface displacements in order to determine the relaxation parameter. As the
computational cost is quite high, it is recommended to update the relaxation
parameters only after a few iterations or time steps.
Mok et al. (2001) and Mok (2001) applied the Aitkin method, originally employed
to accelerate modied Newton-Raphson methods for non-linear computations, to
the iteratively coupled FSI problem in order to obtain an optimal relaxation factor
n+1(i)
for each iteration i:
n+1(i)
=
n+1(i1)
_
_
_
1
_
u
n+1(i)
u
n+1(i+1)
_
T
u
n+1(i+1)
_
u
n+1(i)
u
n+1(i+1)
_
2
_
_
_
(5.40)
where
u
n+1(i)
= u
n+1(i1)
u
n+1(i)
(5.41)
For the rst iteration of the rst time step the relaxation factor
1(0)
has to be
supplied based on experience. The relaxation factors for the rst iteration of the
following time steps
n+1(0)
is equal to the last relaxation factor of the previous
time step
n(i
max
)
. The extra computational cost of the Aitkin method is almost
negligible.
As mentioned before, the convergence depends on the order in which the elds
are solved and on which boundary condition is imposed to which eld. H ubner
et al. (2007) computed the structure rst with imposed forces and achieved the
best convergence by relaxing the transferred forces instead of the displacements.
Causin et al. (2005) showed that if the structure is computed rst with imposed
displacements at the uid-structure interface and the uid next with imposed
stresses, the relaxation factor should go to zero if the mesh is rened, which limits
its practical use.
Often this Aitkin method accelerates the convergence very well and less
subiterations are needed. No trial and error or based on empirical formulas to
determine a relaxation factor which converges and needs a minimal number of
subiterations.
However, for strongly coupled problems, even with the Aitkin convergence
acceleration the required number of subiterations to obtain convergence may
become very high and make the computations inecient (Vierendeels, 2006).
Therefore alternative strongly coupled partitioned methods were developed.
Method 2: Newtons method
At time level t
n+1
the iterative solvers F and S compute the new value of the
variables for respectively the uid and the solid subproblems:
v
(i+1)
= F(v
(i)
, u) (5.42)
u
(i+1)
= S(u
(i)
, v) (5.43)
These solvers are in xed-point form. The residuals associated with each solver
are:
R
f
(v
(i)
, u
(i)
) = v
(i)
F(v
(i)
, u
(i)
) (5.44)
R
s
(u
(i)
, v
(i)
) = u
(i)
S(u
(i)
, v
(i)
) (5.45)
Upon convergence these residuals are equal to 0. The block Newton method tries
to nd the root of the residual:
_
D
v
R
f
(v
(i)
, u
(i)
) D
u
R
f
(v
(i)
, u
(i)
)
D
v
R
s
(u
(i)
, v
(i)
) D
u
R
s
(u
(i)
, v
(i)
)
_ _
v
(i)
u
(i)
_
=
_
R
f
(v
(i)
, u
(i)
)
R
s
(u
(i)
, v
(i)
)
_
(5.46)
where v
(i)
= v
(i+1)
v
(i)
and u
(i)
= u
(i+1)
u
(i)
. The rst matrix on the
left hand side consists of the Jacobians of the residuals with respect to the uid
and the structural degrees of freedom. The cross Jacobians D
u
R
f
and D
v
R
s
are a measure of the sensitivity of respectively the uid solution with respect to
structural motions and the structural solution with respect to changes in the uid
domain. Matthies and Steindorf (2002, 2003) eliminated v
(i)
from the system
of equations (5.46) and solved the resulting system with an iterative method
(e.g. GMRES, Bi-CGStab). The cross Jacobians are approximated by nite
dierences. Fernandez and Moubachir (2005) derived the exact expressions for
the cross Jacobians of the linearized subproblems.
If for simplicity the pressure degrees of freedomp are included in v
, the discretized
coupled system (5.11) can be simplied to:
_
_
K
f
K
f
0
K
f
K
f
+K
s
K
s
0 K
s
K
s
_
_
_
_
v
_
_
=
_
_
f
f
f
f
+f
s
f
s
_
_
(5.47)
If this system is condensed on the uid-structure interface, the following interface
equation is obtained:
(S
f
+S
s
)u
= f
f
+f
s
K
f
K
f
1
f
f
K
s
K
s
1
f
s
(5.48)
where S
f
and S
s
are respectively the Schur complement for the uid and the
structure eld:
S
f
= K
f
K
f
K
f
1
K
f
S
s
= K
s
K
s
K
s
1
K
s
(5.49)
Starting from the interface equation (5.48), the uid-structure interaction problem
can be formulated as a xed-point problem (or Picard iteration) for the interface
displacement u
:
u
(i+1)
= S
1
s
(S
f
u
(i)
+f
f
+f
s
K
f
K
f
1
f
f
K
s
K
s
1
f
s
) (5.50)
As an alternative to the above described Newtons method, the residual of the
structural degrees of freedom at the interface can be used:
R
(u
(i)
) = u
(i)
S
1
s
(S
f
(u
(i)
)+f
f
+f
s
K
f
K
f
1
f
f
K
s
K
s
1
f
s
) (5.51)
The application of Newtons method yields:
D
u
(u
(i)
)u
(i)
= R
(u
(i)
) (5.52)
where u
(i)
= u
(i+1)
u
(i)
. Gerbeau and Vidrascu (2003) solved this system

with the iterative GMRES method. In order to approximate the Jacobian a
simplied linear and inviscid model for the uid is used. Deparis et al. (2006)
applied domain decomposition techniques to the interface equation (5.48) which
enables the parallel computation of both elds.
Generally, these Newtons methods need less iterations per time step than the
xed-point methods, but the iterations are computationally more expensive. Only
for strongly coupled problems, which require lots of iterations even with Aitkin
acceleration technique, Newtons method will outperform the xed-point methods.
5.3 Load and motion transfer
When the uid and structure domain are discretized, two options exist. If the
discretizations of the uid and the structure are the same at the uid-structure
interface, the meshes are called matching or conforming. In this case the transfer
of displacements and forces becomes straightforward as the nodes of both meshes
coincide.
However, very often the mesh requirements at the uid-structure interface dier
for the uid and the structure. Typically the uid mesh has to be ner than the
structure mesh. The use of non-matching meshes oers the exibility to design
and rene independently the meshes of the uid and the structure. The elements
of the structure mesh might dier from those of the uid mesh in shape (e.g.
hexahedrons versus tetrahedrons) or in order.
The drawback is that two dierent discrete representations of the uid-structure
interface exist. If the uid-structure interface is not piecewise planar, the nodes
of one mesh might lay inside or outside the second mesh or accidentally on its
surface. This complicates the load and motion transfer, which will be treated in
the following sections.
If no-slip boundary conditions (5.7)-(5.9) are applied at the uid-structure
interface and the connectivity between the nodes and the elements of both the
meshes of the uid and the structure is kept during the computations, the matrices
for load and displacement transfer have to be computed only once at the beginning
of the calculation. When the elds are remeshed or when the meshes are adaptively
rened during the computations, the load and displacement transfer has to be
recalculated.
First the methods for associating points and surfaces of one mesh with points or
surfaces of the other mesh are reviewed.
Point to point mapping
A rst method to transfer data between two non-matching meshes, searches for
a point of the rst mesh the point of the second mesh closest to this point. The
points might be nodes or a Gauss integration points.
LOAD AND MOTION TRANSFER 175
A brute force search loops for each point of the rst mesh over all the points on the
interface of the other mesh in order to nd the best point to be mapped to. As a
brute force search is computationally inecient for problems with a large number
of points, more ecient search algorithms have been developed.
The bucket search method (Asano et al., 1985) divides the space around the uid-
structure interface into a number of buckets (Cartesian boxes). First all points of
the second mesh are located in one or more buckets. Next the point of the rst
mesh is located in a bucket. The loop over the candidate elements is now restricted
to the elements within that bucket. While the bucket search method uses equal
sized Cartesian boxes, the octree search method (Knuth, 1998) uses adaptively
rened Cartesian boxes and is able to account for large dierences in element size
within one of the meshes.
Point to element mapping
The second approach tries to nd for a point of the rst mesh the element of the
second mesh closest to this point. The best element should fulll two matching
criteria.
First, the projection of the point onto the element should lay within the element.
If the elements of the second mesh form a concave ridge, a point can be sometimes
mapped onto more than one element. The element which has the smallest normal
distance is selected. If the elements of the second mesh form a convex ridge or
if the edges of the uid-structure interface are misaligned, the projection of the
point may not lay within any element. In these cases the point should be mapped
onto the closest element edge.
Next, the normal distance between the point and the selected element should
be smaller than a specied tolerance. In Ansys (Ansys, 2005c) this tolerance is
a fraction (e.g. 10
6
) of the largest dimension of the Cartesian bounding box
around the specic uid-structure interface. L ohner et al. (1995) uses a tolerance
that varies for the dierent elements on the interface and is equal to ve percent
of the square root of the area of the element.
In this case, a brute force search loops for each point of the rst mesh over all the
elements on the interface of the other mesh in order to nd the best element to be
mapped onto. The more ecient bucket search and octree search methods can be
used as well.
L ohner (1995) developed for this specic case the vectorized advancing front
neighbour-to-neighbour algorithm. If for a point an element in the vicinity of this
point is already known and the neighbouring elements of this element are known,
it is possible to go, based on the matching criteria, from element to element until
the best element is found. If the best element is known for a point, it is very easy
to nd, using this technique, the best element for the neighbouring points . The
drawback is that a detailed knowledge of the topologies of both meshes is required.
By default the more ecient search algorithms should be used. If these fail, there
can be switched to more robust algorithms like the brute force search.
Element to element mapping
The third method projects all elements of the second mesh onto one element of
the rst mesh. The ratio of the overlapped area of the element of the rst mesh
and an element of the second mesh to the area of the element of the rst mesh
determines the weight used during the transfer.
In Ansys the element of the rst mesh is rst divided into as many faces as the
element has nodes. The overlapped areas are not computed exactly, but obtained
by converting all areas in pixel images which have a default resolution of 100 by
100 pixels.
5.3.1 Motion transfer or surface tracking
On the uid-structure interface equation (5.9) should be fullled. The
compatibility between the discretized displacement elds of uid and structure
on can generally be expressed as:
u
= Hu
(5.53)
where u
is the vector with the displacements of the nodes of the uid mesh
at the uid-structure interface and u
is the vector with the displacements of

the nodes of the structure mesh at the uid-structure interface. The matrix H
depends on the method used for transferring the displacements. Alternatively, the
uid displacements of the node j can be expressed as a function of the structural
displacements of the node i:
u
j
=
i
s
i=1
H
ji
u
i
(5.54)
Nearest neighbour interpolation
The nearest neighbour interpolation uses the point to point mapping. For each
node of the uid mesh the closest node of the structure mesh is determined. The
computed structural displacements at this node are transferred to the uid node.
This method results in a matrix H with a single one on each row. It only works
well if the nodes of both meshes are almost coincident.
Consistent interpolation method
The consistent interpolation method is based on a point to element mapping: each
node of the uid mesh is mapped onto one element of the mesh of the structure.
The value of the displacement at the projection of the uid node onto the structural
element is interpolated using the shape functions N
s
i
of the structure mesh.
u
j
= u(x
j
) =
i
s
i=1
N
s
i
(x
j
)u
i
(5.55)
where i
s
is the number of nodes of the structure at the uid-structure interface and
x
j
the coordinates of the projection of the uid node onto the structural element.
Comparison of equation 5.55 and 5.53 gives:
H
ji
= N
s
i
(x
j
) (5.56)
In this case the uid nodes, which receive the data, are projected onto the elements
of the structure which send the data.
Figure 5.7 shows the displacement proles in the uid that correspond to an
arbitrarily created displacement prole in the structure using the consistent
interpolation. If the mesh of the uid is ner than the mesh of the structure
(gure 5.7a), this method preserves the displacement prole. If the mesh of the
uid is coarser than the mesh of the structure (gure 5.7b), the number of nodes
of the uid mesh is insucient to accurately capture the displacement prole.
(a) (b)
Figure 5.7: Meshes and displacement proles at the uid-structure interface in the
structure ( ) and the uid partition ( ) obtained with the consistent
interpolation method when uid mesh is (a) ner and (b) coarser than
the solid mesh.
Quadratic interpolation
If the uid mesh is ner than the structure mesh and linear elements are used
for the structure, a consistent interpolation would yield a piecewise planar uid
domain. Using the displacements of the nodes as well as the normals dened in
the nodes of one element of the structure, a quadratic interpolation yields the
displacements in the uid nodes (L ohner et al., 1995). The normal dened in a
node is the average of the normals of the surrounding elements.
Incompatible dimensionality of the uid and the structure mesh
In several cases the uid mesh and the structure mesh have a dierent
dimensionality: usually the uid mesh has a higher dimensionality than the
structure mesh e.g. a plate model of a bridge, a beam model of a chimney or a shell
model of thick wall are coupled to a three-dimensional mesh for the ow around
it. In these cases the discretizations of the uid and the structure are separated in
space. The initial distances between the uid nodes and the structure mesh have to
be computed at the beginning of the computation. During the computation these
initial distance vectors should translated and rotated together with the structure
elements. The easiest approach is to translate and rotate the initial distance
vectors together with the element normals of the structure. A more advanced
approach uses for each uid grid point a normal that is interpolated between
the normals dened in the nodes of the element of the structure. This yields a
smoother deformed mesh (Cebral and L ohner, 1997b).
Element to element mapping
If an element to element mapping is used to transfer the displacements, the ratio
of the overlapped area of the uid element and an structural element to the area
of the uid element of the rst mesh determines the weight used to compute the
uid displacements.
Figure 5.8 shows the displacement proles in the uid that correspond to the
arbitrarily created displacement prole in the structure using the element to
element mapping.
If the mesh of the uid is ner than the mesh of the structure (gure 5.8a),
the method preserves the displacement prole. If the uid mesh is much ner
than the mesh of the structure, several neighbouring uid nodes may receive the
displacement from the same node of the structure which yields constant pieces in
the displacement prole of the uid.
If the mesh of the uid is coarser than the mesh of the structure (gure 5.8b),
the number of nodes of the uid mesh is insucient to accurately capture the
displacement prole. Any local oscillations present in the structural displacement
prole are smoothed by taking an area-weighted average of the displacements in
several nodes of the structure.
(a) (b)
Figure 5.8: Meshes and displacement proles at the uid-structure interface in the
to element mapping when uid mesh is (a) ner and (b) coarser than
the solid mesh.
5.3.2 Load transfer
At uid structure interface the traction on the surface of the structure should
be in equilibrium with the traction on the surface of the uid (equation 5.7).
The total force exerted by the uid over the uid-structure interface should be
equal to the total force imposed on the structure at this interface:
_
(
s
n)d =
_
(p
f
n +
f
n)d (5.57)
A conservative load transfer method fullls exactly the above equation. Not all
load transfer methods are conservative.
Consistent interpolation method
If the structure is discretized by the nite element method, the energy-equivalent
nodal forces f
s
i
in the element coordinate system can be computed from the
traction transferred from the uid to the structure:
f
s
i
=
_
s
N
s
T
i
df =
_
s
N
s
T
i
(p
f
n +
f
n)d
s
(5.58)
This integral can be approximated using a Gaussian quadrature:
f
s
i
=
n
g
g=1
w
g
N
s
T
i
(x
g
)(p
f
(x
g
)n(x
g
) +
f
(x
g
)n(x
g
)) det F
s
(5.59)
where w
g
is weight of the Gauss point x
g
, n
g
is the number of Gauss points used
for approximating the integral.
In order to determine the values of the uid pressure p
f
(x
g
) and stress
f
(x
g
)
at the Gauss points, the consistent interpolation method uses a point to element
mapping: each Gauss point (equation (5.59)) of the structure mesh is mapped
onto one element of the mesh of the uid. The values of p
f
(x
g
) and
f
(x
g
) at the
projection of the Gauss point onto the uid element are interpolated using the
discretization technique intrinsic to the ow solver.
The consistent interpolation transfers the traction from the uid to the structure
partition and integrates this traction over the surface of the structure. This means
that the method does not conserve exactly the total force (equation (5.57)): e.g. if
the uid mesh is ner than the structure mesh, the projections of the Gauss points
of the structure mesh are located in only a part of the uid elements. The traction
in the other uid elements are not used to compute the nodal forces. However
the number of Gauss points n
g
used for evaluating the nodal forces f
s
i
does not
have to be equal to the number of Gauss points selected for computing the element
stiness matrices (e.g. when a high pressure gradient is expected, a larger number
of Gauss points can be used).
Farhat et al. (1998b) shows that for aeroelastic simulations the consistent
interpolation method produces practically conservative results although it is not
exactly conservative.
Node to element conservative method
A conservative method is obtained if the traction is rst integrated over the surface
of the uid independently of the structure. The energy-equivalent nodal forces f
f
j
in the element coordinate system can be computed from the work performed in
one element at the uid-structure interface
f
:
W
f
=
_
f
n ud
f
=
j
f
j=1
_
f
(
f
n)
T
N
f
j
u
j
d
f
=
j
f
j=1
f
T
f
j
u
j
(5.60)
where the energy-equivalent nodal force f
f
j
is equal to:
f
f
j
=
_
f
N
f
T
j

f
nd
f
(5.61)
If the nite volume method is used to discretize the uid domain, the functions
N
f
j
are step functions. After transformation to the global coordinate system and
addition of the contributions of the neighbouring elements, the resulting nodal
force is transferred to the structure using a point to element mapping: each node
of the uid mesh is mapped onto one element of the mesh of the structure. The
transferred force is distributed over the nodes of this element according to the
value of the shape functions:
f
s
i
=
j
f
j=1
N
s
T
i
(x
j
)f
f
j
(5.62)
As the sum of the shape functions is equal to one the total force is conserved:
i
s
i=1
f
s
i
=
i
s
i=1
j
f
j=1
N
s
T
i
(x
j
)f
f
j
=
j
f
j=1
f
f
j
(5.63)
In this case the uid nodes, which send the data, are projected onto the elements
of the structure which receive the data. The relations between the nodal forces of
uid and structure on can generally be expressed as:
f
s
= Lf
f
(5.64)
Figure 5.9 shows the nodal forces on the structure that correspond to an arbitrarily
created pressure prole in the uid using the node to element conservative method.
If the mesh of the uid is ner than the mesh of the structure (gure 5.9a), the
load distribution is adequately captured. If the mesh of the uid is coarser than
the mesh of the structure (gure 5.9b), some nodes of the structure mesh may not
receive any loads. The total force is conserved, but the load distribution is not
accurate.
Element to element conservative method
In order to remove the disadvantage of the above conservative method that some
nodes of the structure mesh may not receive any loads if the mesh of the uid is
coarser than the mesh of the structure, an element to element mapping can be
used. The traction is rst integrated over the surface of the uid independently of
the structure. The uid element is divided into as many faces as the element has
nodes. All neighbouring elements of the structure are projected onto each of these
(a)
(b)
Figure 5.9: Meshes and nodal forces at the uid-structure interface in the
structure ( ) and the uid partition ( ) obtained with the
conservative method when uid mesh is (a) ner and (b) coarser than
the solid mesh.
faces. The ratio of the overlapped area of a face and an element of the structure
to the area of this face determines the weight used during the transfer.
Figure 5.10 shows the nodal forces on the structure that correspond to an
arbitrarily created pressure prole in the uid using the element to element
conservative method. The total force is conserved and the load distribution is
now adequately captured independent of the fact that the mesh of the uid is ner
(gure 5.10a) or coarser (gure 5.10b) than the mesh of the structure.
(a)
(b)
Figure 5.10: Meshes and nodal forces at the uid-structure interface in the
to element conservative method when uid mesh is (a) ner and (b)
coarser than the solid mesh.
Weighted residual formulation
At uid structure interface the traction on the surface of the structure should be
in equilibrium with this on the surface of the uid (equation (5.7)). To ensure this
equilibrium, a weighted residual formulation can be used with the shape functions
N
s
i
of the structure as weight functions (Cebral and L ohner, 1997a):
_
N
s
T
i
(
s
n)d =
_
N
s
T
i
(p
f
n +
f
n)d (5.65)
If the shape functions of the structure are used to impose the traction on the
uid-structure interface and the shape functions N
f
j
are used to approximate the
uid stress, the weighted residual formulation becomes:
_
s
N
s
T
i
i
s
i=1
N
s
i
(
s
i
n)d =
_
f
N
s
T
i
j
f
j=1
N
f
j
(p
f
j
n +
f
j
n)d (5.66)
The integral at the right hand side will be approximated using a Gaussian
quadrature. A rst possibility is to perform a loop over the Gauss points of the
mesh of the structure:
_
s
N
s
T
i
i
s
i=1
N
s
i
(
s
i
n)d =
n
e
s
e
s
=1
n
g
s
g
s
=1
w
g
s
N
s
T
i
(x
g
s
)N
f
j
(x
g
s
)(p
f
j
n+
f
j
n) det F
s
(5.67)
The right hand side corresponds to the nodal forces obtained with the consistent
interpolation method. As mentioned before, if the uid mesh is ner than the
structure mesh, the projections of the Gauss points of the structure mesh are
located in only a part of the uid elements. The traction in the other uid elements
are not used to compute the nodal forces. The second possibility is to perform a
loop over the Gauss points of the mesh of the uid:
_
s
N
s
T
i
i
s
i=1
N
s
i
(
s
i
n)d =
n
e
f
e
f
=1
n
g
f
g
f
=1
w
g
f
N
s
T
i
(x
g
f
)N
f
j
(x
g
f
)(p
f
j
n +
f
j
n) det F
f
(5.68)
The right hand side corresponds to the nodal forces obtained with the node to
element conservative method. The advantage is that the uid pressure of all
elements is taken into account. By rst integrating over the surface of the uid
the total force is conserved. If the mesh of the uid is coarser than the mesh of
the structure, the load distribution is not accurate. Therefore, Cebral and L ohner
(1997b) suggested an adaptive Gaussian quadrature: if the structure elements are
smaller than the uid elements, the number of Gauss points is increased by dividing
the uid elements into smaller elements. The ratio of the area of the uid element
to the areas of the structure elements that host its Gauss points determines the
number of subdivisions.
5.3.3 Conservation of energy
As the uid and the structure are closed system, the energy released or absorbed
by the structure (except for structural damping) should be equal to the energy
gained or released by the uid. Conservation of energy at the uid-structure
interface yields:
W
s
=
_
s
n ud
s
= W
f
=
_
f
n ud
f
(5.69)
After discretization of the displacement elds, the energy is expressed as a function
of the nodal displacements and the nodal forces (equations (5.58) and (5.61)):
u
T
f
s
= u
T
f
f
(5.70)
The combination of any motion transfer method (5.53) with the expression for
conservation of energy (5.70) determines the corresponding load transfer which
conserves the energy:
u
T
f
s
= u
T
H
T
f
f
(5.71)
f
s
= H
T
f
f
(5.72)
If the consistent interpolation method (5.56) is used for the displacement transfer,
the corresponding load transfer which conserves the energy becomes (Farhat et al.,
1998b):
f
s
i
=
j
f
j=1
N
s
T
i
(x
j
)f
f
j
(5.73)
This is the node to element load transfer method as described in equation (5.62)
and which conserves the total force.
As mentioned in section 5.2.2, the violation of the velocity continuity due to the
use of dierent time integration methods in the uid and the structure still results
in an error in the exchange of kinetic energy.
Similarly, the combination of any load transfer method (5.64) with the expression
for conservation of energy (5.70) determines the corresponding motion transfer
APPLICATION: OVALLING OF SILOS 185
which conserves the energy:
f
T
f
u
= f
T
f
L
T
u
(5.74)
u
= L
T
u
(5.75)
For the load transfer a method which conserves the total force can be selected.
According to L ohner (2006), this might lead to non-smooth deformations which
are not locally accurate. If you are dealing with transient aeroelastic problems, a
method which conserves the energy might be preferred, while in the case of shocks,
it is important to conserve the total force.
5.4 Application: ovalling of silos
In this section the wind induced ovalling oscillations of the silos located in the
port of Antwerp, are studied using uid-structure interaction. In order to limit
the computational cost, the computation is performed for one silo in a three-
dimensional ow.
In chapter 3 a three-dimensional nite element model with 18988 elements was
validated using experimental modal analysis. For the uid-structure interaction
computation, a coarser nite element model without the local renements near
the boundaries is used. The cylindrical part of the silo is modelled with 60 8-
node quadrilateral shell elements in the circumferential direction and 20 in the
axial direction. The total number of shell elements in this model is equal to 1756.
The boundary conditions are the same as in chapter 3. Table 5.2 compares the
eigenfrequencies of the coarser model with eigenfrequencies of the experimentally
validated model (table 3.2). For the lowest eigenfrequencies around 4 Hz, the
dierences between eigenfrequencies computed with the two models are smaller
than 2%.
To advance the nite element solution of the structure in time, the Newmark
method (3.106) with = 0.25 and = 0.5 is used. The time step t is equal
to 0.005 s, which is small enough to compute accurately the contributions of
eigenmodes up to 10 Hz. For the structure Rayleigh damping is added. The
damping matrix C is constructed as a linear combination of the mass and the
stiness matrix:
C = M+ K (5.76)
The modal damping ratios
k
at two dierent frequencies
k
determine the
multipliers and for respectively the mass and the stiness matrix by means of
the following system of equations:
2
k
k
= +
2
k
(5.77)
(m, n) Coarser mesh Validated mesh
(1,2) 7.90 8.80 7.75 8.48
(1,3) 4.00 4.00 3.93 3.93
(1,4) 3.93 4.05 3.93 4.04
(1,5) 5.37 5.37 5.25 5.25
(1,6) 7.37 7.37 7.37 7.37
(1,7) 9.72 9.72 9.72 9.72
(2,4) 8.71 8.97 8.71 8.94
(2,5) 5.93 5.93 5.56 5.56
(2,5) 8.08 8.08 8.01 8.01
(2,6) 9.29 9.49 9.29 9.39
Table 5.2: Eigenfrequencies (in Hz) computed with the coarser and the validated
three-dimensional nite element model.
A modal damping ratio
1
= 0.25% at f
1
= 3.93 Hz and
2
= 0.50% at f
2
= 20 Hz
corresponds to = 0.078 s
1
and = 0.75 10
4
s. The modal damping ratios are
estimated from the measured modal damping ratios for all eigenmodes between
3.93 Hz and 20 Hz (Dooms et al., 2003).
In chapter 4 the two-dimensional turbulent wind ow around a single silo and
around a group of 8 by 5 silos was computed. For the uid-structure interaction
computation, the three-dimensional wind ow around a single cylinder is modelled
with symmetric boundary conditions on top and bottom surfaces. The velocity
prole at the inlet is uniform. The turbulent kinetic energy and the turbulent
energy dissipation rate at the inlet are equal to the low values used for the case
of the two-dimensional ow. This simulation is not meant to resemble to an
atmospheric boundary layer ow around a free-standing cylinder as there are no
velocity gradients in the inlet prole and the turbulence intensity is low. An
unsteady RANS simulation using the Shear Stress Transport turbulence model is
performed.
The three-dimensional mesh is obtained by copying the mesh A
2
used for the
unsteady ow around a single silo (chapter 4) 12 times in the axial direction. The
number of elements in the axial direction is sucient to compute the ow around
a silo which deforms according to an eigenmode with m = 1. This number is
however too low to generate any variation in the vortex shedding pattern along
the axial direction. The total number of elements is equal to 642360.
The transient solution is integrated by the three-point backward dierence scheme
(2.176) with a time step t = 0.005 s, which is equal to the time step used for the
structure and small enough to take the vortex shedding into account. Within every
time step, maximum 5 iterations are performed to obtain a converged solution.
In order to obtain an initial solution for the uid-structure computation, the uid
eld is rst computed independently during 1080 time steps, which results in a
time window of 5.4 s. Figure 5.11 shows the time history and the frequency content
of the pressure between t = 3.55 s and t = 5.4 s at the cylinders surface at mid-
height for = 112
, = 174
and = 180
. The frequency resolution f = 0.54 is

quite low due to the short analysed time interval. The vortex shedding frequency
at 2.16 Hz and a higher harmonic at 4.31 Hz are clearly visible in the frequency
content. This vortex shedding frequency is slightly higher than the frequency
obtained in chapter 4 with a two-dimensional computation (2.13 Hz).
4 4.5 5
900
800
700
600
500
400
300
200
100
Time [s]
P
r
e
s
s
u
r
e

[
P
a
]
(a)
0 2 4 6 8 10
0
20
40
60
80
100
120
140
160
180
Frequency [Hz]
P
r
e
s
s
u
r
e

[
P
a
/
H
z
]
(b)
( ), = 174
( ) and
= 180
( ). The angle = 0
coincides with the stagnation point.

Next, the shell model of the structure is coupled with the three-dimensional
incompressible turbulent wind ow (gure 5.12). Between the structure and a
cylindrical surface with a diameter equal to twice the silo diameter, the uid
ow is computed on a deforming mesh, using the Arbitrary Lagrangian Eulerian
formulation (section 2.2.1). The grid point displacements of the uid mesh are
obtained by diusing the displacements of the structure through this domain
(equation (2.233)). As to preserve the quality of the mesh in rened regions, the
diusivity of a nite volume is equal to the inverse of its volume. The grid point
displacements are equal to zero at the inlet and the outlet. On the symmetry
planes, the grid point displacements are not specied.
A node to element conservative method (equations (5.61)-(5.62)) is used for the
load transfer between the non-matching grids, while the consistent interpolation
method (equation (5.55)) is used for the displacement transfer.
As initial conditions the undeformed structure and the transient solution of the
uid ow without interaction at t = 5.4 s are used. 500 time steps are computed,
which results in a time window of 2.5 s. The structure and the uid eld are
Figure 5.12: Model for the coupled simulation of the three-dimensional wind ow
around a cylinder and the response of the silo structure.
sequentially coupled using some of the available algorithms (table A.2). First
the conventional serial staggered algorithm (algorithm A - gure 5.1) is applied,
in which every eld is computed once at each time level t
n
. In a second
computation an iteratively staggered algorithm (algorithm B - gure 5.6) ensures
the equilibrium between the two elds at each time level t
n
. The transferred
interface displacements and forces are not relaxed. Maximum four subiterations
are needed to obtain a relative change of the transferred quantities smaller than
= 0.001 (equation (5.39)).

Figure 5.13a compares the time history of the radial displacements in three points
at mid-height of the silo computed with algorithm A and B. Within this short
time window both computations are stable. For a more rigourous evaluation of
the stability more time steps should be computed. Clearly, the results computed
with algorithm A show dierences that increase in time with the results of
algorithm B because the accuracy of the conventional serial staggered algorithm
is lower than the accuracy of the iteratively staggered algorithm. The accuracy
of the conventional serial staggered algorithm could be improved by the use of a
prediction for the structural displacements and a corrected uid force, but this
option is not available in the coupling between Ansys and CFX (table A.2). As
the staggered coupling algorithm is stable for this example, alternatively the time
step could be reduced in order to improve the accuracy. This might be cheaper
than the use of subiterations.
5.5 6 6.5 7 7.5
0.08
0.06
0.04
0.02
0
0.02
0.04
0.06
0.08
Time [s]
D
i
s
p
l
a
c
e
m
e
n
t

[
m
]
(a)
0 2 4 6 8 10
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
Frequency [Hz]
D
i
s
p
l
a
c
e
m
e
n
t

[
m
/
H
z
]
(b)
Figure 5.13: (a) Time history of the radial displacements at mid-height for =
66
( ), = 120
( ) and = 180
( ) computed with algorithm A

(dashed lines) and algorithm B (solid lines) and (b) frequency content
of the radial displacements computed with algorithm B. The angle
= 0
coincides with the stagnation point.

Figure 5.13b shows the frequency content of the radial displacements in the same
points computed with algorithm B. The frequency resolution f = 0.4 is quite low
due to the short time interval. The response of the silo is dominated by eigenmodes
(1,3) and (1,4) around 4 Hz. The peak around 2 Hz indicates the eects of vortex
shedding on the silo structure. The smaller peaks above 4 Hz are related to the
eigenmodes with higher frequencies.
Figure 5.14 shows the deformations (enlarged with a factor 5) of the structure
between a height of 11.25 m and 13.75 m at three dierent times. At all times an
antinode faces the free stream direction. At t = 5.905 s and t = 6.805 s the response
is dominated by eigenmodes (1,4) with respectively a negative and a positive radial
displacement at = 0
. At t = 7.250 s eigenmode (1,3) is dominant. The maximal

radial displacement is 0.105 m at t = 1.015 s and occurs at a height of 15 m and
= 0
.
Figure 5.15 shows the time history and frequency content of the pressure between
t = 5.4 s and t = 7.9 s at the cylinders surface at mid-height for = 112
, = 174
and = 180
. The comparison with gure 5.11 indicates that the structural

deformations inuence the pressure eld near the wall. While the pressure time
histories of the ow simulation around a rigid silo mainly showed contributions
at 2.16 and 4.31 Hz, in the coupled computation contributions are present as well
at higher frequencies. An important contribution is present around 4 Hz which
corresponds to the eigenfrequencies of eigenmodes (1,3) and (1,4). Due to the
(a) (b) (c)
Figure 5.14: Deformations (enlarged with a factor 5) of the structure between
11.25 m and 13.75 m high at (a) t = 5.905 s, (b) t = 6.805 s and (c)
t = 7.250 s. The wind ows from the left.
5.5 6 6.5 7 7.5
900
800
700
600
500
400
300
200
100
Time [s]
P
r
e
s
s
u
r
e

[
P
a
]
(a)
0 2 4 6 8 10
0
20
40
60
80
100
120
140
160
180
Frequency [Hz]
P
r
e
s
s
u
r
e

[
P
a
/
H
z
]
(b)
( ), = 174
( ) and
= 180
( ) using algorithm B. The angle = 0
coincides with the

stagnation point.
interaction the magnitude of the pressure uctuations around 2 Hz has clearly
increased, which indicates an amplication of the vortex shedding.
The pressure eld on the vertical plane through the cylinder axis parallel with the
inlet ow direction at t = 7.9 s is shown in gure 5.16. The pressure eld behind
the cylinder is clearly three-dimensional. Figure 5.17 shows the pressure at the
cylinders surface along the height at ve dierent times for three circumferential
angles. At the beginning of the simulation (at t = 5.4 s), the pressure is constant
along the height in the stagnation point ( = 0
) and for = 180
. At = 112
the pressure varies slightly with the height. During the coupled simulation the
largest variations along the height occur at = 112
, but also at = 0
and
= 180
considerable variations take place.

Figure 5.16: Pressure eld on the vertical plane through the cylinder axis parallel
with the inlet ow direction at t = 7.9 s.
400 600 800 1000
0
5
10
15
20
25
Pressure [Pa]
H
e
i
g
h
t

[
m
]
(a)
800 600 400 200
0
5
10
15
20
25
Pressure [Pa]
H
e
i
g
h
t

[
m
]
(b)
600 400 200 0
0
5
10
15
20
25
Pressure [Pa]
H
e
i
g
h
t

[
m
]
(c)
Figure 5.17: Pressure at the cylinders surface along the height for (a) = 0
,
(b) = 112
and (c) = 180
at t = 5.4 s ( ), t = 5.475 s ( ),
t = 5.925 s ( ), t = 6.35 s ( ) and t = 6.605 s ( ).
At every time the radial displacements along the circumference at mid-height are
decomposed into a Fourier series of modes with circumferential wavenumbers n.
For each mode, principle component analysis (Pearson, 1901) of the time series
yields the position of the rst and the second principle component with respect
to the silo. For all circumferential wavenumbers n the rst principle component
is positioned roughly with an antinode facing the ow and the second with a
node facing the ow. Figure 5.18 shows the time history and frequency content
of the rst and the second principle component corresponding to circumferential
wavenumbers n = 2, 3 or 4. The response of the silo mainly consists of modes
with circumferential wavenumber n = 3 and 4. Their contribution varies strongly
with time.
In order to evaluate the occurrence of ovalling oscillations, the response of the
structure should be computed during a much longer time interval (e.g. 40 s),
5.5 6 6.5 7 7.5
0.06
0.05
0.04
0.03
0.02
0.01
0
0.01
0.02
0.03
Time [s]
D
i
s
p
l
a
c
e
m
e
n
t

[
m
]
(a)
0 2 4 6 8 10
0
0.005
0.01
0.015
0.02
0.025
0.03
Frequency [Hz]
D
i
s
p
l
a
c
e
m
e
n
t

[
m
/
H
z
]
(b)
Figure 5.18: (a) Time history and (b) frequency content of the rst (solid line)
and second principal component (dashed line) of the displacements
at mid-height with circumferential wavenumber n = 2 ( ), n = 3
( ) and n = 4 ( ).
as the structure is suddenly loaded and the modal damping ratios are very low.
Nowadays, the computation times using a single processor are very high (312
hours for the coupled simulation of the wind ow around one silo during 2.5 s).
The major part (306 hours) is spent on the uid partition. In order to make
coupled simulations of ovalling feasible, more ecient coupling procedures are
required or the computation should be parallelized. A reduction of the dimensions
of the problem could decrease the computation times as well. The nite strip
method (appendix B) enables to build an approximate model of the structure in
two dimensions. This nite strip model of the silo could be coupled with a two
dimensional ow.
5.5 Conclusion
This chapter is concerned with some aspects of the coupled computation of uid
ows around deforming structures. First an overview is given of the existing
partitioned methods where the uid and the structure are separately integrated
in time and the interaction eects are applied as external boundary conditions.
Loosely coupled algorithms solve each eld one or a few times in each time step
and do not exactly fulll the coupling conditions at the uid-structure interface.
The accuracy of these algorithms can be increased by using a suitable predictor
for the transferred structural displacements at the interface and the corresponding
corrector for the transferred uid forces at the interface. The conservation of the
energy at the uid-structure interface is an interesting tool to evaluate the dierent
algorithms. For incompressible ows, loosely coupled algorithms are only stable for
CONCLUSION 193
loosely coupled problems, where the density of the structure is much higher than
the density of the uid. Strongly coupled algorithms solve each eld multiple times
per time step in order to fulll the coupling conditions within a certain tolerance.
Automatically determined relaxation factors for the subiterations between the uid
and the structure partition improve the convergence and reduce the number of
iterations for strongly-coupled problems.
Next, methods for load and motion transfer between non-matching meshes are
described. The ratio of the number of elements in the uid at the interface
to the number of elements in the structure at the interface strongly inuences
the results of the dierent methods. For the transfer of displacements mainly
the accuracy of the transferred displacement prole is of concern. During the
transfer of the uid forces to the structure the total load should be conserved. For
transient aeroelastic computations the conservation of energy at the interface is
very important. The combination of a consistent interpolation for motion transfer
with a node to element conservative method for load transfer conserves the energy
at the uid-structure interface and performs well for problems where the uid
mesh is ner than the structure mesh.
As an application, the shell model of the circular cylindrical silo is coupled with
the three-dimensional incompressible turbulent ow around a cylinder as to predict
ovalling oscillations. The inlet velocity is uniform along the height. The results
computed with the conventional serial staggered algorithm show dierences that
increase in time with the results computed using subiterations, which are more
accurate. The accuracy of the staggered algorithm could be improved using a
predictor for the structural displacements and the corresponding corrector for the
uid forces, but this option is not available in the coupling between Ansys and
CFX. The response of the silo is dominated by the lowest eigenmodes (1,3) and
(1,4) around 4 Hz. In order to evaluate the occurrence of ovalling oscillations, the
response of the structure should be computed during a much longer time interval.
Therefore, more ecient coupling procedures are required or the computation
should be parallelized. A reduction of the dimensions of the problem (e.g. the
nite strip method) could decrease the computation times as well.
Chapter 6
Conclusions and
recommendations for further
research
6.1 Conclusions
The behaviour of civil engineering structures under wind loads is an important
matter of concern. It is a trend in modern architecture to construct more slender
and lightweight structures with often complex geometries. The actual wind
distribution around these complex structures cannot be derived from the literature
and may be strongly inuenced by the presence of neighbouring structures. For
relatively rigid structures, the wind load can be determined for the undeformed
shape of the structure. This is no longer the case for exible structures, where
the pressures cause large structural displacements. The resulting change of shape
has an inuence on the wind pressure distribution around the structure. This may
give rise to structural instabilities, such as utter of bridges, galloping of cables,
ovalling of silos and vortex induced vibrations of tall buildings.
This thesis studies coupled numerical simulations of the uid and the structure to
determine the behaviour of civil engineering structures under wind loads. As an
example throughout the text, wind induced ovalling oscillations of a group of forty
silos are studied.
195
196 CONCLUSIONS AND RECOMMENDATIONS FOR FURTHER RESEARCH
For the numerical simulation of uid-structure interaction a partitioned approach
is used: the uid and the structure are treated as isolated elds and solved
separately. The interaction eects are applied on the individual elds as external
boundary conditions which are exchanged from one eld to the other. The
main advantage of this partitioned approach is that well-established discretization
techniques and solution algorithms (and by consequence existing software) can be
used in the individual elds which are tailored to the characteristic behaviour
of this eld. Due to the use of existing software (Ansys for the structural
computations and Flotran and CFX for the ow computations), the possible
choices regarding models and algorithms were restricted to those available in these
packages.
As the aim is to perform a coupled numerical simulation of the uid and the
structure, the position of the structure determines at least partially the uid
domain boundaries. If the structure undergoes large displacements, it is necessary
to perform the computations of the uid ow on a domain with deforming
boundaries. The uid mesh is aligned with this deforming boundary and should
follow its deformations. Therefore, the governing equations for the uid ow in
an arbitrary Lagrangian-Eulerian description are derived. The convective velocity
in these equations is the relative velocity between uid particles and the mesh.
The mesh velocity is arbitrary but should preferentially preserve the quality and
the renements of the mesh. The mesh deformation is computed by means of
a diusion equation with a variable diusivity. The uid domain is discretized
in space by means of the nite volume method. In the case of an arbitrary
Lagrangian-Eulerian description specic choices have to be made in order to
preserve the accuracy and the stability of the time integration. The implicit three-
point backward dierence method which respects the geometric conservation law,
is used.
The structure is discretized by means of the nite element method. The cylindrical
shell of the silos in the port of Antwerp is studied with Ansys. As for the prediction
of ovalling oscillations the exact value of the structural modal damping is very
important, an in situ experiment is performed in order to determine the modal
damping ratios, eigenfrequencies and eigenmodes of the silos. In order to support
the design of the in situ experiment, the mode shapes and eigenfrequencies of a
single silo are rst computed with a harmonic nite element model. The mode
shapes consist of a number of waves in the circumferential direction and a number
of waves in the axial direction. Radial accelerations at 10 points along the silo are
measured under ambient wind loading and modal parameters are extracted from
the output-only data using the stochastic subspace identication technique. The
eigenmode with the lowest eigenfrequency at 3.94 Hz has the largest contribution to
the measured response. All modes have a low damping ratio around 1%. A three-
dimensional nite element model of the silo is validated by means of the modal
parameters. The boundary conditions strongly inuence the eigenfrequencies and
CONCLUSIONS 197
mode shapes. The eigenmodes with three and four waves along the circumference
(n = 3 or 4) and half a wave along the height (m = 1) are both situated at 3.93 Hz.
Geometrically non-linear behaviour is taken into account through an updated
Lagrangian description and allows for large rotations and large displacements.
The Newton-Raphson technique solves the non-linear system of equations. The
inuence of geometrical non-linear behaviour on deformations of the silos of order
of magnitude of 0.1 m is not negligible: the magnitude of the displacements
increases where the curvature decreases and decreases where the curvature
increases.
Turbulent wind ows around civil engineering structures have a very high Reynolds
number. LES and hybrid RANS/LES approaches are well suited to capture the
large scale turbulent structures in the wake of blu bodies. However, as their
computational cost is today still too high for the use in coupled uid-structure
problems, URANS simulations are performed.
In order to clarify the inuence of the RANS turbulence models, the near-wall mesh
renement and the unsteadiness, rst the steady state two-dimensional turbulent
air ow around a single cylinder in the post-critical regime at Reynolds number
12.4 10
6
is computed with Flotran and CFX. Within Flotran, the near-wall
modelling strongly inuences the results, while the results obtained with CFX
are quite insensitive. Results computed with various eddy viscosity turbulence
models strongly dier, but with respect to experimental results reported in the
literature the minimum pressure coecient is always underestimated, while the
base pressure coecient is overestimated. For the ow around a circular cylinder
at these Reynolds numbers, the Shih-Zhu-Lumley turbulence model of Flotran
has the best correspondence with experimental results. Within CFX, the shear
stress transport turbulence model produces the best results. Dierent results are
obtained with Flotran and CFX using the same turbulence models. A transient
simulation model captures the regular vortex shedding in the post-critical regime
and is therefore preferred.
As an application the interference eects for the group of 8 by 5 silos in the port
of Antwerp with spacing ratio P/D = 1.05 are studied at an incidence angle of
30
. Vortices shed at 0.165 Hz from the group as a whole and at 2.85 Hz from the
individual cylinders. The large velocities in the gaps produce distinct minima in
the pressure distributions where the gap area is minimum. The two cylinders on
the side corners of the group experience quite high lift coecients. Eigenmodes
with circumferential wavenumber n equal to a multiple of 4 are more heavily loaded
in the group conguration, where four small gaps surround most of the cylinders,
compared with the single cylinder conguration. Especially for the cylinders on
the side corners of the group, eigenmodes with n = 3 or n = 4, which have the
lowest eigenfrequencies for these silos, are strongly excited. This explains why
storm damage is mainly located on silos on the corners of the group.
The coupled numerical simulation of the response of a structure under uid
loads is computed using a partitioned approach. As the uid and the structural
eld are non-linear, direct time integration methods are applied in both elds.
Two partitioned solution algorithms which are available in Ansys and CFX, are
employed: the rst, a loosely coupled serial staggered algorithm, solves each eld
once in each time step and transfers the last computed solution of one eld to
the other eld. It does not exactly fulll the coupling conditions at the uid-
structure interface, what may deteriorate the stability and the accuracy of the
coupled computation. For incompressible ows, loosely coupled algorithms are
only stable for loosely coupled problems, where the density of the structure is
much higher than the density of the uid. The second, a strongly coupled staggered
algorithm, iterates between the two elds in each time step and fullls the coupling
conditions within a certain tolerance.
The meshes of the uid and the structure are non-matching at the uid-structure
interface. During the displacement transfer, the consistent interpolation method
guarantees the accuracy of the transferred displacement prole, while for the load
transfer a node to element method conserves the total load. This combination
conserves the energy at the uid-structure interface and performs well for problems
where the uid mesh is ner than the structure mesh.
The interaction between the wind ow and a single silo is computed. A shell model
of a circular cylindrical silo is coupled with the three-dimensional incompressible
turbulent wind ow as to predict ovalling oscillations. The results computed with
the conventional serial staggered algorithm show dierences that increase in time
with the results computed using subiterations, which are more accurate. The
response of the silo is dominated by the lowest eigenmodes (1,3) and (1,4) around
4 Hz. In order to evaluate the occurrence of ovalling oscillations, the response of
the structure should be computed during a much longer time interval.
6.2 Recommendations for further research
In this text, the ow is computed on deforming meshes with the commercial
software CFX. The accuracy and the stability of ow computations on deforming
meshes could be evaluated. First it should be checked if the second order time
accuracy of the three-point backward dierence scheme is preserved in the ALE
description. For instance the two-dimensional ow around a cylinder with a
prescribed deformation with a given frequency, circumferential wavenumber and
amplitude could be computed. Based on computations with dierent time steps
the order of time accuracy can be derived.
Next, the implementation of the geometric conservation law could be checked
by computing a uniform ow and moving all the inner nodes of the mesh in
RECOMMENDATIONS FOR FURTHER RESEARCH 199
an arbitrary way. It is required that this solution can be computed exactly
independently of the mesh deformations and that the mass balance is fullled
for the computational domain.
In this thesis, the two-dimensional turbulent air ow around a group of 8 by 5
cylinders with spacing ratio P/D = 1.05 is studied at an incidence angle of 30
.
A signicant question that remains is how sensitive these computed results are to
the incidence angle and the spacing ratio. In order to quantify these eects, a
systematic study of the ow for a number of incidence angles and spacing ratios
can be performed.
The ow around one nite circular cylinder of aspect ratio h/D = 4.55 is strongly
three-dimensional with horseshoe vortices, trailing vortices and arch vortices. It is
expected that comparable three-dimensional vortex structures are found around
the group as a whole. The numerical simulation of the three-dimensional ow
around one nite cylinder could be calculated but is still challenging from a
computational point of view. Using multiple processors, the numerical simulation
of the three-dimensional ow around forty closely-spaced nite circular cylinders
might become feasible.
If the three-dimensional ow around one nite cylinder is computed the RANS
turbulence models could be replaced by a hybrid RANS/LES approach which is
able to capture the large scale turbulent structures in the wake of blu bodies.
In the last chapter the coupled simulation of three-dimensional incompressible
turbulent wind ow around a circular cylindrical silo is computed in order to
predict ovalling oscillations. As the computational cost is very high, more ecient
coupling procedures are required. The use of the commercially available coupling
between Ansys and CFX, limited the choice of employed coupling algorithms. A
new implementation of the coupling between the two commercial packages should
enable the use of more advanced coupling algorithms.
The accuracy of the conventional staggered algorithm could be improved using a
predictor for the structural displacements and the corresponding corrector for the
uid forces. If this enhanced staggered algorithm is still stable for loosely coupled
problems with incompressible ows, the use of subiterations could be avoided,
which substantially decreases the computational cost.
For strongly coupled problems, the algorithm with subiterations can be enhanced
using automatically determined relaxation factors for the subiterations between
the uid and the structure partition in order to improve the convergence and to
reduce the number of subiterations.
An interesting alternative is to construct reduced order models for the uid and
the structure during the subiterations (Vierendeels, 2006). For the uid in each
subiteration the pressure distribution at the uid-structure interface due to the
imposed displacements at this interface is computed. These sets of imposed
displacements and corresponding pressure distributions can be used to construct
a reduced order model for the uid. Analogously, for the structure in each
subiteration the displacements at the uid-structure interface which correspond
to the imposed pressures at this interface are computed. These sets can be used
to construct a reduced order model for the structure. The result of the coupled
computation using the two reduced order models can be used as an initial guess
for a subiteration. Each subiteration the reduced order models are improved.
This technique allows for the solution of strongly coupled problems with a small
number of subiterations. It can possibly be improved by reusing in the current
time step the reduced order models constructed during the previous time steps.
This might be very interesting for aeroelastic problems as utter, galloping and
ovalling because a lot of computations of the uid domain are required for nearly
identical geometries.
A global multigrid technique, which computes partitioned coupling between uid
and structure at dierent grid levels, might as well enhance the stability, eciency
and accuracy (Sch afer et al., 2006; van Zulen et al., 2007). The computation of
the coupled problem is rst performed on a coarse uid and a coarse structure mesh.
This solution can be used as a coarse grid prediction for a coupled computation
on a ner grid. The low frequency error in the solution obtained on a ne grid can
be corrected by computing a new solution on the coarse meshes.
The eciency, accuracy and stability of the dierent coupling algorithms can be
veried on a set of benchmark examples e.g. an elastic beam behind a square
(Wall and Ramm, 1998) or behind a circle (T urek and Hron, 2006). Bathe and
Ledezma (2007) give an overview of a series of benchmarks in order to evaluate
the coupling algorithms and load and motion transfer algorithms.
A reduction of the dimensions of the problem could decrease the computational
cost as well. For the example of the silo, the nite strip method (appendix B) may
be applied to make a representative two-dimensional model of the structure. The
structural displacements are supposed to vary as half a sine or cosine function in
the axial direction. The nite strip element is implemented as a combination of a
two-dimensional beam element and a user dened element. A drawback is that the
stiness matrix of the user dened elements is independent of the deformation, so
that geometrical non-linear eects are not included in these additional terms. This
nite strip model could be coupled to a two-dimensional ow computation. For
the transfer of the uid forces to the structure, an assumption for the variation of
the pressure along the height of the silo has to be made. As gure 5.17 shows, the
variations are neither constant, neither sinusoidal. As a validation, the results of a
coupled simulation with the nite strip model and the two-dimensional ow should
RECOMMENDATIONS FOR FURTHER RESEARCH 201
be compared to the results of a coupled simulation of the three-dimensional ow
around the shell model of the structure. The inuence of the group conguration
on the occurrence of ovalling oscillations can be subsequently studied by coupling
nite strip models for the silos with a two-dimensional ow around the group.
Obviously, the number of silos should be gradually increased, starting with two
closely-spaced silos.
While only the ovalling of silos is considered in the present work, the methodology
can also be applied to wind response of tensile structures, utter of bridges,
galloping of cables and vortex induced vibrations of marine riser pipes or cables.
Bibliography
Afgan, I., Moulinec, C., Prosser, R., Laurence, D., 2007. Large eddy simulation
of turbulent ow for wall mounted cantilever cylinders of aspect ratio 6 and 10.
International Journal of Heat and Fluid Flow 28 (4), 561574.
Akbari, M., Price, S., 2005. Numerical investigation of ow patterns for staggered
cylinder pairs in cross-ow. Journal of Fluids and Structures 20 (4), 533554.
Alam, M., Sakamoto, H., Zhou, Y., 2005. Determination of ow congurations
and uid forces acting on two staggered circular cylinders of equal diameter in
cross-ow. Journal of Fluids and Structures 21 (4), 363394.
Anderson, J., 1995. Computational uid dynamics : the basics with applications.
McGraw-Hill series in mechanical engineering. McGraw-Hill, New York.
Ansys, October 2005a. Ansys CFX-Solver Theory Guide Ansys CFX Release 10.0.
Ansys, Inc.
Ansys, July 2005b. Ansys Theory reference, Ansys Release 10.0. Ansys, Inc.
Ansys, August 2005c. Coupled Field Analysis Guide, ANSYS Release 10.0. ANSYS
Inc.
Arnold, R., Warburton, G., 1949. Flexural vibrations of the walls of thin cylindrical
shells having freely supported ends. Proceedings of the Royal Society of London
197A, 238256.
Arnold, R., Warburton, G., 1953. The exural vibrations of thin cylinders.
Proceedings of the Institution of Mechanical Engineers 167, 6280.
Asano, T., Edahiro, M., Imai, H., Iri, M., Murota, K., 1985. Practical use
of bucketing techniques in computational geometry. In: Toussaint, G. (Ed.),
Computational Geometry. Vol. 2. North-Holland, Amsterdam, pp. 153195.
Baaens, F., 2001. A ctitious domain/mortar element method for uid-structure
interaction. International Journal for Numerical Methods in Fluids 35 (7), 743
761.
203
204 BIBLIOGRAPHY
Babuska, I., 1971. Error-bounds for nite element method. Numerische
Mathematik 16 (4), 322333.
Bar-Yoseph, P., Mereu, S., Chippada, S., Kalro, V., 2001. Automatic monitoring
of element shape quality in 2-D and 3-D computational mesh dynamics.
Computational Mechanics 27 (5), 378395.
Bathe, K., 1996. Finite Element Procedures, 2nd Edition. Prentice-Hall,
Englewood Clis, NJ.
Bathe, K., Ledezma, G., 2007. Benchmark problems for incompressible uid ows
with structural interactions. Computers and Structures 85 (1114), 628644.
Bathe, K., Zhang, H., Ji, S., 1999. Finite element analysis of uid ows fully
coupled with structural interactions. Computers and Structures 72 (13), 116.
Batina, J., 1990. Unsteady euler airfoil solutions using unstructured dynamic
meshes. AIAA Journal 28 (8), 13811388.
Beale, S., Spalding, D., 1999. A numerical study of unsteady uid ow in in-line
and staggered tube banks. Journal of Fluids and Structures 13 (6), 723754.
Behr, M., Hastreiter, D., Mittal, S., Tezduyar, T., 1995. Incompressible ow past
a circular cylinder: Dependence of the computed ow eld on the location of the
lateral boundaries. Computer Methods in Applied Mechanics and Engineering
123, 309316.
Behr, M., Liou, J., Shih, R., Tezduyar, T., 1991. Vorticity streamfunction
formulation of unsteady incompressible ow past a cylinder : Sensitivity of
the computed ow eld to the location of the outow boundary. International
Journal for Numerical Methods in Fluids 12, 323342.
Benhamadouche, S., Laurence, D., 2003. LES, coarse LES, and transient RANS
comparisons on the ow across a tube bundle. International Journal of Heat and
Fluid Flow 24 (4), 470479.
Billington, D., 1965. Thin shell concrete structures. McGraw-Hill.
BIN, November 1995. NBN ENV 1991-2-4: Eurocode 1: Basis of design and actions
on structures - Part 2-4 : Actions on structures - Wind actions. Belgisch Instituut
voor Normalisatie.
Bletzinger, K., W uchner, R., Kupzok, A., 2006. Algorithmic treatment of shells
and free form-membranes in FSI. In: Bungartz, H.-J., Sch afer, M. (Eds.), Fluid-
Structure Interaction. Modelling, Simulation, Optimisation. Vol. 53 of Lecture
Notes in Computational Science and Engineering. Springer, pp. 336335.
Blevins, R., 1990. Flow-induced vibration. Van Nostrand Reinhold.
BIBLIOGRAPHY 205
Blocken, B., Stathopoulos, T., Carmeliet, J., 2007. CFD simulation of the
atmospheric boundary layer: wall function problems. Atmospheric Environment
41, 238252.
Blom, F., 2000. Considerations on the spring analogy. International Journal for
Numerical Methods in Fluids 32 (6), 647668.
Bo, D., Gastaldi, L., 2004. Stability and geometric conservation laws for
ALE formulations. Computer Methods in Applied Mechanics and Engineering
193 (4244), 47174739.
Bonet, J., Wood, R., 1997. Nonlinear continuum mechanics for nite element
analysis. Cambridge University Press.
Brooks, A., Hughes, T., 1982. Streamline Upwind Petrov-Galerkin formulations
for convection dominated ows with particular emphasis on the incompressible
Navier-Stokes equations. Computer Methods in Applied Mechanics and
Engineering 32 (13), 199259.
Catalano, P., Wang, M., Iaccarino, G., Moin, P., 2003. Numerical simulation of the
ow around a circular cylinder at high Reynolds numbers. International Journal
of Heat and Fluid Flow 24, 463469.
Causin, P., Gerbeau, J., Nobile, F., 2005. Added-mass eect in the design
of partitioned algorithms for uid-structure problems. Computer Methods in
Applied Mechanics and Engineering 194 (4244), 45064527.
Cebral, J., L ohner, R., 1997a. Conservative load projection and tracking for uid-
structure problems. AIAA Journal 35 (4), 687692.
Cebral, J., L ohner, R., January 1997b. Fluid-structure coupling: Extensions and
improvements. In: 35th Aerospace Sciences Meeting & Exhibit, AIAA-970858.
Reno, Nevada, US.
Celik, I., Karatekin, O., 1997. Numerical experiments on application of
Richardson extrapolation with nonuniform grids. Journal of Fluids Engineering,
Transactions of the ASME 119 (3), 584590.
Celik, I., Shaer, F. D., 1995. Long time-averaged solutions of turbulent-ow past
a circular-cylinder. Journal of Wind Engineering and Industrial Aerodynamics
56, 185212.
Champneys, A., July 23 2008. The Tacoma Narrows bridge disaster, november
1940. http://www.enm.bris.ac.uk/anm/tacoma/tacoma.html.
Chen, X., Kareem, A., 2001. Nonlinear response analysis of long-span
bridges under turbulent winds. Journal of Wind Engineering and Industrial
Aerodynamics 89 (1415), 13351350.
206 BIBLIOGRAPHY
Cheung, Y., 1976. Finite strip method in structural analysis. Pergamon Oxford.
Cook, R., Malkus, D., Plesha, M., Witt, R., 2002. Concepts and applications of
nite element analysis, 4th Edition. John Wiley and Sons.
Craft, T., Launder, B., Suga, K., 1996. Development and application of a cubic
eddy-viscosity model of turbulence. International Journal of Heat and Fluid
Flow 17 (2), 108115.
Criseld, M., 1991. Non-linear nite element analysis of solids and structures,
Volume 1. John Wiley and Sons.
Criseld, M., 1997. Non-linear nite element analysis of solids and structures,
Volume 2: advanced topics. John Wiley and Sons.
de Boer, A., van der Schoot, M., Bl, H., 2007. A three-dimensional torsional
spring analogy method for unstructured dynamic meshes. Computers and
Structures 85 (1114), 784795.
De Hart, J., Peters, G., Schreurs, P., Baaens, F., 2003. A three-dimensional
computational analysis of uid-structure interaction in the aortic valve. Journal
of Biomechanics 36 (1), 103112.
Degand, C., Farhat, C., 2002. A three-dimensional torsional spring analogy method
for unstructured dynamic meshes. Computers and Structures 80 (34), 305316.
Demirdzic, I., Peric, M., 1988. Space conservation law in nite volume calculations
of uid-ow. International Journal for Numerical Methods in Fluids 8 (9), 1037
1050.
Deparis, S., Discacciati, M., Fourestey, G., Quarteroni, A., 2006. Fluid-structure
algorithms based on Steklov-Poincare operators. Computer Methods in Applied
Mechanics and Engineering 195 (4143), 57975812.
Dockstader, E., Swiger, W., Ireland, E., 1956. Resonant vibration of steel stacks.
Transactions of the American Society of Civil Engineers 121, 10881112.
Donea, J., Guiliani, S., Halleux, J., 1982. An Arbitrary Lagrangian-Eulerian nite-
element method for transient dynamic uid structure interactions. Computer
Methods in Applied Mechanics and Engineering 33 (1-3), 689723.
Donea, J., Huerta, A., 2003. Finite Element Methods for ow problems. John
Wiley and Sons, Chichester, England.
Dooms, D., De Roeck, G., Degrande, G., September 2006a. Inuence of the
group positioning of cylinders on the wind pressure distribution in the post-
critical regime. In: Wesseling, P., O nate, E., Periaux, J. (Eds.), Proceedings
of European Conference on Computational Fluid Dynamics ECCOMAS CFD
2006. Egmond aan Zee, the Netherlands.
BIBLIOGRAPHY 207
Dooms, D., De Roeck, G., Degrande, G., May 2007. Fluid-structure interaction
applied to ovalling oscillations of silos. In: O nate, E., Papadrakakis, M.,
Schreer, B. (Eds.), Computational Methods for Coupled Problems in Science
and Engineering II. Santa Eulalia, Ibiza, Spain, pp. 586589.
Dooms, D., Degrande, G., De Roeck, G., Reynders, E., 2006b. Finite element
modelling of a silo based on experimental modal analysis. Engineering Structures
28 (4), 532542.
Dooms, D., Jacobs, S., Degrande, G., De Roeck, G., June 2003. Dynamische
analyse van een groep silos onder dynamische windbelasting: In situ metingen
op een lege hoeksilo. Report to Ellimetal BWM-2003-09, Department of Civil
Engineering, K.U.Leuven.
Dowell, E., Hall, K., 2001. Modeling of uid-structure interaction. Annual Review
of Fluid Mechanics 33, 445490.
Farhat, C., May 2006. On the three-eld formulation & solution of nonlinear uid-
structure interaction problems. In: Matthies, H., Ohayon, R. (Eds.), Course on
Advanced Computational Methods for Fluid-Structure Interaction. Ibiza, Spain.
Farhat, C., Degand, C., Koobus, B., Lesoinne, M., 1998a. Torsional springs for two-
dimensional dynamic unstructured uid meshes. Computer Methods in Applied
Farhat, C., Geuzaine, P., 2004. Design and analysis of robust ALE time-integrators
for the solution of unsteady ow problems on moving grids. Computer Methods
in Applied Mechanics and Engineering 193 (3941), 40734095.
Farhat, C., Geuzaine, P., Grandmont, C., 2001. The discrete geometric
conservation law and the nonlinear stability of ALE schemes for the solution
of ow problems on moving grids. Journal of Computational Physics 174 (2),
669694.
Farhat, C., Lesoinne, M., 2000. Two ecient staggered algorithms for the serial and
parallel solution of three-dimensional nonlinear transient aeroelastic problems.
Computer Methods in Applied Mechanics and Engineering 182 (34), 499515.
Farhat, C., Lesoinne, M., LeTallec, P., 1998b. Load and motion transfer algorithms
for uid/structure interaction problems with non-matching discrete interfaces:
Momentum and energy conservation, optimal discretization and application to
aeroelasticity. Computer Methods in Applied Mechanics and Engineering 157,
95114.
Farhat, C., van der Zee, K., Geuzaine, P., 2006. Provably second-order
time-accurate loosely-coupled solution algorithms for transient nonlinear
computational aeroelasticity. Computer Methods in Applied Mechanics and
Engineering 195 (1718), 19732001.
208 BIBLIOGRAPHY
Farrant, T., Tan, M., Price, W., 2000. A cell boundary element method applied
to laminar vortex-shedding from arrays of cylinders in various arrangements.
Journal of Fluids and Structures 14 (3), 375402.
Felippa, C., Park, K., 1980. Staggered transient analysis for coupled mechanical
systems. Computer Methods in Applied Mechanics and Engineering 24, 61111.
Felippa, C., Park, K., Farhat, C., 2001. Partitioned analysis of coupled mechanical
systems. Computer Methods in Applied Mechanics and Engineering 190, 3247
3270.
Fernandez, M., Moubachir, M., 2005. A Newton method using exact jacobians for
solving uid-structure coupling. Computers and Structures 83 (23), 127142.
Ferziger, J., Peric, M., 2002. Computational methods for uid dynamics, 3rd
Edition. Springer, Berlin.
Formaggia, L., Nobile, F., 2004. Stability analysis of second-order time
accurate schemes for ALE-FEM. Computer Methods in Applied Mechanics and
Engineering 193 (3941), 40974116.
Forsberg, K., 1964. Inuence of boundary conditions on the modal characteristics
of thin cylindrical shells. AIAA Journal 2 (12), 21502157.
Forster, C., Wall, W., Ramm, E., 2006. On the geometric conservation law in
transient ow calculations on deforming domains. International Journal for
Forster, C., Wall, W., Ramm, E., 2007. Articial added mass instabilities in
sequential staggered coupling of nonlinear structures and incompressible viscous
ows. Computer Methods in Applied Mechanics and Engineering 196 (7), 1278
1293.
Fourestey, G., Piperno, S., 2004. A second-order time-accurate ALE Lagrange-
Galerkin method applied to wind engineering and control of bridge proles.
Computer Methods in Applied Mechanics and Engineering 193 (3941), 4117
4137.
Frohlich, J., Rodi, W., 2004. LES of the ow around a circular cylinder of nite
height. International Journal of Heat and Fluid Flow 25 (3), 537548.
Gerbeau, J., Vidrascu, M., 2003. A quasi-Newton algorithm based on a
reduced model for uid-structure interaction problems in blood ows. ESAIM-
Mathematical Modelling and Numerical Analysis-Modelisation Mathematique
et Analyse Numerique 37 (4), 631647.
Gerstenberger, A., Wall, W., 2008. An eXtended Finite Element Method/Lagrange
multiplier based approach for uid-structure interaction. Computer Methods in
BIBLIOGRAPHY 209
Geuzaine, P., Grandmont, C., Farhat, C., 2003. Design and analysis of ALE
schemes with provable second-order time-accuracy for inviscid and viscous ow
simulations. Journal of Computational Physics 191 (1), 206227.
Glowinski, R., Pan, T., Periaux, J., 1994. A ctitious domain method for external
incompressible viscous-ow modeled by Navier-Stokes equations. Computer
Methods in Applied Mechanics and Engineering 112 (14), 133148.
Gl uck, M., Breuer, M., Durst, F., Halfmann, A., Rank, E., 2001. Computation
of uid-structure interaction on lightweight structures. Journal of Wind
Engineering and Industrial Aerodynamics 89 (1415), 13511368.
Gl uck, M., Breuer, M., Durst, F., Halfmann, A., Rank, E., 2003. Computation of
wind-induced vibrations of exible shells and membranous structures. Journal
of Fluids and Structures 17, 739765.
Gresho, P., Sani, R., 2000a. Incompressible Flow and the Finite Element Method,
Volume 1, Advection-Diusion and Isothermal Laminar Flow. John Wiley and
Sons, Chichester, England.
Gresho, P., Sani, R., 2000b. Incompressible Flow and the Finite Element Method,
Volume 2, Isothermal Laminar Flow. John Wiley and Sons, Chichester, England.
Gu, Z., 1996. On interference between two circular cylinders at supercritical
reynolds number. Journal of Wind Engineering and Industrial Aerodynamics
62 (23), 175190.
Gu, Z., Sun, T., 1999. On interference between two circular cylinders in staggered
arrangement at high subcritical Reynolds numbers. Journal of Wind Engineering
and Industrial Aerodynamics 80 (3), 287309.
Guillard, H., Farhat, C., 2000. On the signicance of the geometric conservation
law for ow computations on moving meshes. Computer Methods in Applied
Gunn, D., Malik, A., 1966. Wind forces and proximity of cooling towers to each
other. Nature 210, 11421143.
Hassan, Y., Barsamian, H., 2004. Tube bundle ows with the large eddy simulation
technique in curvilinear coordinates. International Journal of Heat and Mass
Transfer 47 (1416), 30573071.
Haug, E., de Kermel, P., Gelenne, P., September 2003. Numerical simulations for
wind load design of large umbrellas. In: Mollaert, M., Haase, J., Chilton, J.,
Moncrie, E., Dencher, M., Barnes, M. (Eds.), Tensinet Symposium: Designing
tensile architecture. Brussels, Belgium.
210 BIBLIOGRAPHY
Heck, M., Sternel, D., Sch afer, M., Yigit, S., September 2006. Inuence of
numerical and physical parameters on an implicit partitioned uid-structure
solver. In: Wesseling, P., O nate, E., Periaux, J. (Eds.), Proceedings of European
Conference on Computational Fluid Dynamics ECCOMAS CFD 2006. Egmond
aan Zee, the Netherlands.
Helenbrook, B., 2003. Mesh deformation using the biharmonic operator.
International Journal for Numerical Methods in Engineering 56 (7), 10071021.
Hirsch, C., 1995a. Numerical computation of internal and external ows ; 1
Fundamentals of numerical discretization. Wiley series in numerical methods
in engineering. John Wiley and Sons, Chichester, England.
Hirsch, C., 1995b. Numerical computation of internal and external ows ; 2
Computational methods for inviscid and viscous ows. Wiley series in numerical
methods in engineering. John Wiley and Sons, Chichester, England.
Hirt, C., Amsden, A., Cook, J., 1974. Arbitrary Lagrangian-Eulerian computing
method for all ow speeds. Journal of Computational Physics 14 (3), 227253.
Holloway, D., Walters, D., Leylek, J., 2004. Prediction of unsteady, separated
boundary layer over a blunt body for laminar, turbulent, and transitional ow.
International Journal for Numerical Methods in Fluids 45 (12), 12911315.
Hron, J., T urek, S., 2006. A monolithic FEM/multigrid solver for an ALE
formulation of uid-structure interaction with applications in biomechanics. In:
Bungartz, H.-J., Sch afer, M. (Eds.), Fluid-Structure Interaction. Modelling,
Simulation, Optimisation. Vol. 53 of Lecture Notes in Computational Science
and Engineering. Springer, pp. 146170.
H ubner, B., Seidel, U., Scherer, T., May 2007. Inuence of system properties
and solver parameters on the convergence behavior of partitioned solutions to
hydroelastic systems. In: O nate, E., Papadrakakis, M., Schreer, B. (Eds.),
Computational Methods for Coupled Problems in Science and Engineering II.
Santa Eulalia, Ibiza, Spain, pp. 598601.
H ubner, B., Walhorn, E., Dinkler, D., July 2002. Numerical investigations to
bridge aeroelasticity.
H ubner, B., Walhorn, E., Dinkler, D., 2004. A monolithic approach to uid-
structure interaction using space-time nite elements. Computer Methods in
Hughes, T., 1987. The nite element method: linear static and dynamic nite
element analysis. Prentice-Hall, Englewood Clis, New Jersey.
BIBLIOGRAPHY 211
Hughes, T., Franca, L., Balestra, M., 1986. A new nite element formulation for
computational uid dynamics: V. circumventing the Babuska-Brezzi condition:
a stable Petrov-Galerkin formulation of the Stokes problem accomodating equal-
order interpolations. Computer Methods in Applied Mechanics and Engineering
59 (1), 8599.
Hughes, T., Franca, L., Hulbert, G., 1989. A new nite-element formulation
for computational uid-dynamics .8. the Galerkin least-squares method for
advective-diusive equations. Computer Methods in Applied Mechanics and
Engineering 73 (2), 173189.
Hughes, T., Liu, W., Zimmermann, T., 1981. Lagrangian-Eulerian nite element
formulation for incompressible viscous ows. Computer Methods in Applied
Mechanics and Engineering 29, 329349.
Iaccarino, G., Ooi, A., Durbin, P., Behnia, M., 2003. Reynolds averaged simulation
of unsteady separated ow. International Journal of Heat and Fluid Flow 24 (2),
147156.
Idelsohn, S., Onate, E., Del pin, F., 2004. The particle nite element method: a
powerful tool to solve incompressible ows with free-surfaces and breaking waves.
International Journal for Numerical Methods in Engineering 61 (7), 964989.
Jester, W., Kallinderis, Y., 2003. Numerical study of incompressible ow about
xed cylinder pairs. Journal of Fluids and Structures 17 (4), 561577.
Johns, D., Sharma, C., August 1974. On the mechanism of wind-excited ovalling
vibrations of thin circular cylindrical shells. In: Naudascher, E. (Ed.), Flow-
induced structural vibrations. Springer-Verslag, Berlin, Karlsruhe, Germany, pp.
650662.
Johnson, A., Tezduyar, T., 1994. Mesh update strategies in parallel nite-
element computations of ow problems with moving boundaries and interfaces.
Computer Methods in Applied Mechanics and Engineering 119 (12), 7394.
Jones, W., Launder, B., 1972. Prediction of laminarization with a 2-equation model
of turbulence. International Journal of Heat and Mass Transfer 15 (2), 301314.
Katsura, S., 1985. Wind-excited ovalling vibration of a thin circular cylindrical-
shell. Journal of Sound and Vibration 100 (4), 527550.
Kjellgren, P., Hyvarinen, J., 1998. An arbitrary lagrangian-eulerian nite element
method. Computational Mechanics 21 (1), 8190.
Knuth, D., 1998. The art of computer programming, Volume 3: Sorting and
Searching, 2nd Edition. Addison-Wesley, Reading, Massachusetts.
212 BIBLIOGRAPHY
Koga, T., 1988. Eects of boundary conditions on the free vibrations of circular
cylindrical shells. AIAA Journal 26 (11), 13871394.
Koobus, B., Farhat, C., 1999. Second-order time-accurate and geometrically
conservative implicit schemes for ow computations on unstructured dynamic
meshes. Computer Methods in Applied Mechanics and Engineering 170 (12),
103129.
Kraus, H., 1967. Thin elastic shells: an introduction to the theoretical foundations
and the analysis of their static and dynamic behavior. John Wiley and Sons.
Lam, K., Fang, X., 1995. The eect of interference of 4 equispaced cylinders in
cross-ow on pressure and force coecients. Journal of Fluids and Structures
9 (2), 195214.
Lam, K., Li, J., So, R., 2003. Force coecients and Strouhal numbers of four
cylinders in cross ow. Journal of Fluids and Structures 18 (34), 305324.
Laneville, A., Mazouzi, A., 1995. Ovalling oscillations of cantilevered cylindrical
shells in cross-ow: New experimental data. Journal of Fluids and Structures
9 (7), 729745.
Laneville, A., Mazouzi, A., 1996. Wind-induced ovalling oscillations of cylindrical
shells: Critical onset velocity and mode prediction. Journal of Fluids and
Structures 10 (7), 691704.
Lazzari, M., Vitaliani, R., Majowiecki, M., Saetta, A., 2003. Dynamic behavior of a
tensegrity system subjected to follower wind loading. Computers and Structures
81 (2223), 21992217.
Le Tallec, P., Mouro, J., 2001. Fluid structure interaction with large structural
displacements. Computer Methods in Applied Mechanics and Engineering
190 (2425), 30393067.
Leissa, A., 1993. Vibration of shells. Acoustical Society of America - American
institute of physics.
Lesoinne, M., Farhat, C., 1996. Geometric conservation laws for ow problems
with moving boundaries and deformable meshes, and their impact on aeroelastic
computations. Computer Methods in Applied Mechanics and Engineering
134 (12), 7190.
Liu, W., Chang, H., Chen, J., Belytschko, T., 1988. Arbitrary Lagrangian-Eulerian
petrov-galerkin nite-elements for nonlinear continua. Computer Methods in
L ohner, R., 1995. Robust, vectorized search algorithms for interpolation on
unstructured grids. Journal of Computational Physics 118 (2), 380387.
BIBLIOGRAPHY 213
L ohner, R., 2001. Applied Computational Fluid Dynamics Techniques: An
Introduction Based on Finite Element Methods. John Wiley and Sons,
Chichester, England.
L ohner, R., May 2006. Finite element formulations and advanced applications. In:
Matthies, H., Ohayon, R. (Eds.), Course on Advanced Computational Methods
for Fluid-Structure Interaction. Ibiza, Spain.
L ohner, R., Cebral, J., May 1996. Fluid-structure interaction in industry: issues
and outlook. In: 3rd World Conference in Applied Computational Fluid
Dynamics. Germany.
Lohner, R., Yang, C., 1996. Improved ale mesh velocities for moving bodies.
Communications in Numerical Methods in Engineering 12 (10), 599608.
L ohner, R., Yang, C., Cebral, J., Baum, J., Luo, H., Pelessone, D., Charman, C.,
1995. Fluid-structure interaction using a loose coupling algorithm and adaptive
unstructured grids. In: Hafez, M., Oshima, K. (Eds.), Computational uid
dynamics review. John Wiley, pp. 755776.
Matthies, H., Niekamp, R., Steindorf, J., 2006. Algorithms for strong coupling
procedures. Computer Methods in Applied Mechanics and Engineering 195 (17
18), 20282049.
Matthies, H., Steindorf, J., 2002. Partitioned but strongly coupled interation
schemes for nonlinear uid-structure interaction. Computers and Structures
80 (2730), 19911999.
Matthies, H., Steindorf, J., 2003. Partitioned strong coupling algorithms for uid-
structure interaction. Computers and Structures 81 (811), 805812.
Mazouzi, A., Laneville, A., Vittecoq, P., 1991. An analytical model of the ovalling
oscillations of clamped-free and clamped-clamped cylindrical-shells in cross-ow.
Melchers, R., 1987. Structural reliability: analysis and prediction. Ellis Horwood.
Menter, F., 1994. Two-equation eddy-viscosity turbulence models for engineering
applications. AIAA Journal 32 (8), 15981605.
Meynen, S., Verma, H., Hagedorn, P., Sch afer, M., 2005. On the numerical
simulation of vortex-induced vibrations of oscillating conductors. Journal of
Fluids and Structures 21 (1), 4148.
Moin, P., Mahesh, K., 1998. Direct numerical simulation: A tool in turbulence
research. Annual Review of Fluid Mechanics 30, 539578.
Mok, D., 2001. Partitionierte L osunsans atze in der strukturdynamik und der Fluid-
Struktur-Interaction. Ph.D. thesis, Institut f ur Baustatik, Universitat Stuttgart.
214 BIBLIOGRAPHY
Mok, D., Wall, W., Ramm, E., 2001. Accelerated iterative substructuring schemes
for instationary uid-structure interaction. In: Bathe, K. (Ed.), Computational
Fluid and Solid Mechanics. Elsevier, pp. 13251328.
Niemann, H., Kopper, H., 1998. Inuence of adjacent buildings on wind eects on
cooling towers. Engineering Structures 20 (10), 874880.
Norfolk, M., July 23 2008. Knottingley and Ferrybridge local history: Tales and
events. http://www.knottingley.org/history/talesandevents.htm.
Nozu, T., Tamura, T., Okuda, Y., Sanada, S., 2008. LES of the ow and
building wall pressures in the center of Tokyo. Journal of Wind Engineering and
Industrial Aerodynamics 96 (1011), 17621773, 4th International Symposium
on Computational Wind Engineering (CWE 2006), Yokohama, Japan, July 16-
19, 2006.
Orlando, M., 2001. Wind-induced interference eects on two adjacent cooling
towers. Engineering Structures 23 (8), 979992.
Padoussis, M., Helleur, C., 1979. Ovalling oscillations of cylindrical-shells in cross-
ow. Journal of Sound and Vibration 63 (4), 527542.
Padoussis, M., Price, S., Ang, S., 1991. An improved theory for utter of
cylindrical-shells in cross-ow. Journal of Sound and Vibration 149 (2), 197
218.
Padoussis, M., Price, S., Fekete, G., Newman, B., 1983. Ovalling of chimneys:
Induced by vortex shedding or self-excited? Journal of Wind Engineering and
Industrial Aerodynamics 14 (13), 119128.
Padoussis, M., Price, S., Suen, H., 1982a. An analytical model for ovalling
oscillation of clamped-clamped cylindrical-shells in cross ow. Journal of Sound
and Vibration 83 (4), 555572.
Padoussis, M., Price, S., Suen, H., 1982b. Ovalling oscillations of cantilevered
and clamped-clamped cylindrical-shells in cross ow: An experimental-study.
Journal of Sound and Vibration 83 (4), 533553.
Panesar, A., Johns, D., 1985. Ovalling oscillations of thin circular cylindrical-shells
in cross ow - an experimental study. Journal of Sound and Vibration 103 (2),
201209.
Parkinson, G., Jandali, T., 1970. A wake source model for blu body potential
ow. Journal of Fluid Mechanics 40 (3), 577594.
Pattenden, R., Bresslo, N., Turnock, S., Zhang, X., 2007. Unsteady simulations
of the ow around a short surface-mounted cylinder. International Journal for
BIBLIOGRAPHY 215
Pattenden, R., Turnock, S., Zhang, X., 2005. Measurements of the ow over a low-
aspect-ratio cylinder mounted on a ground plane. Experiments in Fluids 39 (1),
10 21.
Pearson, K., 1901. On lines and planes of closest t to systems of points in space.
Philosophical Magazine 2 (6), 559572.
Peeters, B., De Roeck, G., 1999. Reference-based stochastic subspace identication
for output-only modal analysis. Mechanical Systems and Signal Processing
13 (6), 855878.
Peskin, C., 1972. Flow patterns around heart valves - numerical method. Journal
of Computational Physics 10 (2), 252271.
Peskin, C., Mcqueen, D., 1980. Modeling prosthetic heart-valves for numerical-
analysis of blood-ow in the heart. Journal of Computational Physics 37 (1),
113132.
Piperno, S., 1997. Explicit/implicit uid/structure staggered procedures with a
structural predictor and uid subcycling for 2d inviscid aeroelastic simulations.
International Journal for Numerical Methods in Fluids 25, 12071226.
Piperno, S., Farhat, C., 2001. Partitioned procedures for the transient solution
of coupled aeroelastic problems - part II: Energy transfer analysis and
three-dimensional applications. Computer Methods in Applied Mechanics and
Engineering 190 (2425), 31473170.
Piperno, S., Farhat, C., Larrouturou, B., 1995. Partitioned procedures for the
transient solution of coupled aroelastic problems. 1. model problem, theory
and 2-dimensional application. Computer Methods in Applied Mechanics and
Engineering 124 (12), 79112.
Pope, S., 2000. Turbulent ows. Cambridge University Press.
Portela, G., Godoy, L., 2007. Wind pressure and buckling of grouped steel tanks.
Wind and Structures 10 (1), 2344.
Price, S., Padoussis, M., Mark, B., 1995. Flow visualization of the interstitial
cross-ow through parallel triangular and rotated square arrays of cylinders.
Journal of Sound and Vibration 181 (1), 8598.
Robertson, I., Sherwin, S., 1999. Free-surface ow simulation using hp/spectral
elements. Journal of Computational Physics 155 (1), 2653.
Rollet-Miet, P., Laurence, D., Ferziger, J., 1999. LES and RANS of turbulent ow
in tube bundles. International Journal of Heat and Fluid Flow 20 (3), 241254.
216 BIBLIOGRAPHY
Rugonyi, S., Bathe, K., 2001. On nite element analysis of uid ows fully coupled
with structural interactions. Computer Modeling in Engineering & Sciences 2 (2),
195212.
Saghaan, M., Stansby, P. K., Saidi, M. S., Apsley, D. D., 2003. Simulation
of turbulent ows around a circular cylinder using nonlinear eddy-viscosity
modelling: steady and oscillatory ambient ows. Journal of Fluids and
Structures 17, 12131236.
Sch afer, M., Heck, M., Yigit, S., 2006. An implicit partitioned method for the
numerical simulation of uid-structure interaction. In: Bungartz, H.-J., Sch afer,
M. (Eds.), Fluid-Structure Interaction. Modelling, Simulation, Optimisation.
Vol. 53 of Lecture Notes in Computational Science and Engineering. Springer,
pp. 169194.
Shih, T., Liou, W., Shabbir, A., Yang, Z., Zhu, J., 1995a. A new k- eddy viscosity
model for high Reynolds number turbulent ows. Computers and Fluids 24 (3),
227238.
Shih, T., Zhu, J., Lumley, J., 1993. A realizable Reynolds stress algebraic equation
model. Tech. Rep. 105993, NASA Lewis Research Center.
Shih, T., Zhu, J., Lumley, J., 1995b. A new Reynolds stress algebraic equation
model. Computer Methods in Applied Mechanics and Engineering 125, 287302.
Simiu, E., Scanlan, R., 1986. Wind eects on structures. John Wiley and Sons.
Singh, S., Mittal, S., 2005. Flow past a cylinder: shear layer instability and drag
crisis. International Journal for Numerical Methods in Fluids 47 (1), 7598.
Spalart, P., 2000. Strategies for turbulence modelling and simulations.
International Journal of Heat and Fluid Flow 21 (3), 252263.
Spalart, P., Allmaras, S., 1994. A one-equation turbulence model for aerodynamic
ows. La Recherche Aerospatiale (1), 521.
Spalart, P., Jou, W.-H., Strelets, M., Allmaras, S., 1997. Comments on the
feasibility of LES for wings, and on a hybrid RANS/LES approach. In: Liu,
C., Liu, Z. (Eds.), Advances in DNS/LES. Columbus, Ohio, USA.
Speziale, C., 1998. Turbulence modeling for time-dependent RANS and VLES: A
review. AIAA Journal 36 (2), 173184.
Stathopoulos, T., 2002. The numerical wind tunnel for industrial aerodynamics:
Real or virtual in the new millennium? Wind and Structures 5 (2-4), 193208.
Steger, J., Dougherty, F., Benek, J., 1983. A Chimera grid scheme. In: Ghia, K.,
Ghia, U. (Eds.), Advances in Grid Generation. Vol. 5 of ASME FED. pp. 5969.
BIBLIOGRAPHY 217
Stein, K., Tezduyar, T., Benney, R., 2004. Automatic mesh update with the solid-
extension mesh moving technique. Computer Methods in Applied Mechanics
and Engineering 193 (21-22), 20192032.
Stern, F., Wilson, R., Coleman, H., Paterson, E., 2001. Comprehensive approach
to verication and validation of CFD simulations - Part 1: Methodology and
procedures. Journal of Fluids Engineering, Transactions of the ASME 123 (4),
793802.
Sumner, D., Price, S., Padoussis, M., 2000. Flow-pattern identication for two
staggered circular cylinders in cross-ow. Journal of Fluid Mechanics 411, 263
303.
Sumner, D., Richards, M., Akosile, O., 2005. Two staggered circular cylinders of
equal diameter in cross-ow. Journal of Fluids and Structures 20 (2), 255276.
Tamura, T., Ohta, I., Kuwahara, K., 1990. On the reliability of 2-dimensional
simulation for unsteady ows around a cylinder-type structure. Journal of Wind
Engineering and Industrial Aerodynamics 35 (1-3), 275298.
Travin, A., Shur, M., Strelets, M., Spalart, P., 2000. Detached-eddy simulations
past a circular cylinder. Flow, Turbulence and Combustion 63 (1-4), 293313.
T urek, S., Hron, J., 2006. Proposal for numerical benchmarking of uid-structure
interaction between an elastic object and laminar incompressible ow. In:
Bungartz, H.-J., Sch afer, M. (Eds.), Fluid-Structure Interaction. Modelling,
Simulation, Optimisation. Vol. 53 of Lecture Notes in Computational Science
and Engineering. Springer, pp. 371385.
Uematsu, Y., Uchiyama, K., 1985. An experimental investigation of wind-
induced ovalling oscillations of thin, circular cylindrical-shells. Journal of Wind
Engineering and Industrial Aerodynamics 18 (3), 229243.
Van Overschee, P., De Moor, B., 1996. Subspace identication for linear systems.
Kluwer Academic Publishers, Dordrecht, The Netherlands.
van Zulen, A., Bosscher, S., Bl, H., 2007. Two level algorithms for
partitioned uid-structure interaction computations. Computer Methods in
Versteeg, H., Malalasekera, W., 1995. An introduction to computational uid
dynamics : the nite volume method. Longman, London.
Vickery, B., Majowiecki, M., 1992. Wind induced response of a cable supported
stadium roof. Journal of Wind Engineering and Industrial Aerodynamics 42 (1-
3), 14471458.
218 BIBLIOGRAPHY
Vierendeels, J., 2006. Implicit coupling of partitioned uid-structure interaction
solvers using reduced-order models. In: Bungartz, H.-J., Sch afer, M. (Eds.),
Fluid-Structure Interaction. Modelling, Simulation, Optimisation. Vol. 53 of
Lecture Notes in Computational Science and Engineering. Springer, pp. 118.
Vierendeels, J., Riemslagh, K., Dick, E., Verdonck, P., 2000. Computer simulation
of intraventricular ow and pressure gradients during diastole. Journal of
Biomechanical Engineering, Transactions of the ASME 122 (6), 667674.
Wall, W., Gerstenberger, A., Gamnitzer, P., Forster, C., Ramm, E., 2006. Large
deformation uid-structure interaction - advances in ale methods and new
xed grid approaches. In: Bungartz, H.-J., Sch afer, M. (Eds.), Fluid-Structure
Interaction. Modelling, Simulation, Optimisation. Vol. 53 of Lecture Notes in
Computational Science and Engineering. Springer, pp. 195232.
Wall, W., Ramm, E., 1998. Fluid-structure interaction based upon stabilized
(ALE) nite element method. In: Idelsohn, S., O nate, E., Dvorkin, E. (Eds.),
Computational mechanics - New trends and applications (Proceedings of
WCCM IV). Barcelona, Spain.
Wang, G., 2001. A fast and robust variant of the SIMPLE algorithm for nite-
element simulations of incompressible ows. Computational Fluid and Solid
Mechanics 2, 10141016.
Weaver, D., Lian, H., Huang, X., 1993. Vortex shedding in rotated square arrays.
Wilcox, D., 1988. Reassessment of the scale-determining equation for advanced
turbulence models. AIAA Journal 26 (11), 12991310.
Wilcox, D., 1998. Turbulence modeling for CFD. DCW industries.
Willcox, K., Peraire, J., 2002. Balanced model reduction via the proper orthogonal
decomposition. AIAA Journal 40 (11), 23232330.
Willden, R., Graham, J., 2004. Multi-modal vortex-induced vibrations of a vertical
riser pipe subject to a uniform current prole. European Journal of Mechanics,
B/Fluids 23, 209218.
W uchner, R., Kupzok, A., Bletzinger, K., May 2005. Towards FSI for light-
weight structures subjected to wind. In: Papadrakakis, M., O nate, E., Schreer,
B. (Eds.), Computational Methods for Coupled Problems in Science and
Engineering. Santorini, Greece.
Younis, B., Przulj, V., 2006. Computation of turbulent vortex shedding.
Computational Mechanics 37 (5), 408425.
BIBLIOGRAPHY 219
Zahlten, W., Eusani, R., 2006. Numerical simulation of the aeroelastic response
of bridge structures including instabilities. Journal of Wind Engineering and
Industrial Aerodynamics 94 (11), 909922.
Zdravkovich, M., 1997. Flow around circular cylinders. Vol 1: Fundamentals.
Oxford University Press.
Zdravkovich, M., 2003. Flow around circular cylinders. Vol 2: Applications. Oxford
University Press.
Zhang, H., Zhang, X., Ji, S., Guo, Y., Ledezma, G., Elabbasi, N., deCougny,
H., 2003. Recent development of uid-structure interaction capabilities in the
ADINA system. Computers and Structures 81, 10711085.
Zienkiewicz, O., Taylor, R., 2005. The nite element method for solid and
structural mechanics, sixth Edition. Elsevier Butterworth-Heinemann.
Zienkiewicz, O., Taylor, R., Nithiarasu, P., 2005a. The nite element method for
uid dynamics, sixth Edition. Elsevier Butterworth-Heinemann.
Zienkiewicz, O., Taylor, R., Zhu, J., 2005b. The nite element method: its basis
and fundamentals, sixth Edition. Elsevier Butterworth-Heinemann.
Curriculum vitae
David Dooms
18 September 1978, Sint-Niklaas (Belgium)

Education
2001-2008
PhD in Engineering, Department of Civil Engineering, K.U.Leuven
1996-2001
MSc in Engineering: Architecture , K.U.Leuven
1992-1998
Secondary education, LatinMathematics, Sint-Jozef-Klein-Seminarie, Sint-
Niklaas
Work
2001-2007
Assistant, Department of Civil Engineering, K.U.Leuven
2007-2009
Research assistant, OI-project An interactive and adaptive application for the
static and dynamic analysis of structures, Department of Civil Engineering,
K.U.Leuven
221
222 CURRICULUM VITAE
Publications
International journal papers
D. Dooms, G. Degrande, G. De Roeck, and E. Reynders. Finite element modelling
of a silo based on experimental modal analysis. Engineering Structures, 28(4):532
542, 2006.
L. Karl, W. Haegeman, G. Degrande, and D. Dooms. Determination of the
material damping ratio with the bender element test. Journal of Geotechnical
and Geoenvironmental Engineering, Proceedings of the ASCE, 134(12):17431756,
2008.
International conference papers
D. Dooms, G. De Roeck, and G. Degrande. Fluid-structure interaction applied to
ovalling oscillations of a silo. In 8th World Congress on Computational Mechanics
and 5th European Congress on Computational Methods in Applied Sciences and
Engineering, Venice, Italy, July 2008.
D. Dooms, G. De Roeck, and G. Degrande. Wind induced ovalling oscillations of
thin-walled cylindrical structures. In International Workshop on Fluid-Structure
Interaction: Theory, Numerics and Applications, Herrsching am Ammersee,
Germany, September 2008.
D. Dooms, G. De Roeck, and G. Degrande. Fluid-structure interaction applied
to ovalling oscillations of a silo. In Fourth International Conference on Advanced
Computational Methods in Engineering, Liege, Belgium, May 2008.
S. Arnout, D. Dooms, and G. De Roeck. Shape and size optimization of
shell structures with variable thickness. In 6th International Conference on the
computation of shell & spatial structures, Ithaca, New York, USA, May 2008.
W. Figeys, S. Ignoul, D. Dooms, L. Schueremans, D. Van Gemert, and G. De Roeck.
Strengthening of an industrial cylindrical shell damaged by a collision. In
D. DAyala, editor, Proceedings of the Structural Analysis of Historic Construction,
volume 2, pages 10871094, Bath, United Kingdom, July 2008.
CURRICULUM VITAE 223
D. Dooms, G. De Roeck, and G. Degrande. Fluid-structure interaction applied
to ovalling oscillations of silos. In E. O nate, M. Papadrakakis, and B. Schreer,
editors, Computational Methods for Coupled Problems in Science and Engineering
II, pages 586589, Santa Eulalia, Ibiza, Spain, May 2007.
W. Figeys, S. Ignoul, D. Dooms, L. Schueremans, D. Van Gemert, and G. De Roeck.
Repair of impact damage in an industrial cylindrical shell. In Proceedings of 3nd
International b Congress, page 305, Dubrovnik, Croatia, May 2007.
D. Dooms, G. De Roeck, and G. Degrande. Inuence of the group positioning
of cylinders on the wind pressure distribution in the post-critical regime. In
P. Wesseling, E. O nate, and J. Periaux, editors, Proceedings of European
Conference on Computational Fluid Dynamics ECCOMAS CFD 2006, Egmond
aan Zee, the Netherlands, September 2006.
D. Dooms, G. De Roeck, and G. Degrande. Fluid-structure interaction applied to
ovalling oscillations of silos. In P. Sas and M. De Munck, editors, Proceedings of
ISMA2006 International Conference on Noise and Vibration Engineering, pages
45574571, Leuven, Belgium, September 2006.
D. Dooms, G. Degrande, G. De Roeck, and E. Reynders. Wind induced
oscillations of thin-walled silos. In M. Papadrakakis, E. O nate, and B. Schreer,
editors, Computational Methods for Coupled Problems in Science and Engineering,
Santorini, Greece, May 2005. CD-ROM.
D. Dooms, G. Degrande, G. De Roeck, and E. Reynders. Wind induced vibrations
of thin-walled cylindrical structures. In ISMA2004 International Conference on
Noise and Vibration Engineering, pages 781796, Leuven, Belgium, September
2004.
National conference papers
D. Dooms, G. De Roeck, and G. Degrande. Reynolds Averaged Navier Stokes
simulation of the post-critical ow around a single circular cylinder and groups of
cylinders. In Proceedings of the 7th National Congress on Theoretical and Applied
Mechanics, Mons, Belgium, May 2006. National Committee for Theoretical and
Applied Mechanics.
224 CURRICULUM VITAE
Internal reports
D. Dooms, G. Degrande, and G. De Roeck. Dynamic analysis of a group of silos
under dynamic wind loading: modelling of the wind ow around a group of silos.
Report to Ellimetal BWM-2006-06, Department of Civil Engineering, K.U.Leuven,
April 2006.
M. Schevenels, G. Degrande, G. Lombaert, and D. Dooms. Trillingsmetingen
in een woning aan de Loverstraat 37 te Sint-Baafs-Vve. Report to F. Dooms
BWM-2004-05, Department of Civil Engineering, K.U.Leuven, May 2004.
D. Dooms, G. Degrande, and G. De Roeck. Dynamische analyse van een groep
silos onder dynamische windbelasting: Eindige elementenmodellering van een
lege silo. Report to Ellimetal BWM-2003-08, Department of Civil Engineering,
K.U.Leuven, June 2003.
D. Dooms, S. Jacobs, G. Degrande, and G. De Roeck. Dynamische analyse van
een groep silos onder dynamische windbelasting: In situ metingen op een lege
hoeksilo. Report to Ellimetal BWM-2003-09, Department of Civil Engineering,
K.U.Leuven, June 2003.
D. Dooms, G. Degrande, G. De Roeck, and S. Jacobs. Dynamische analyse van
een dansvloer in de Bowling Factory te Braine lAlleud. Report to J. De Laere
BWM-2002-06, Department of Civil Engineering, K.U.Leuven, May 2002.
Appendix A
Possibilities of commercial
software packages Flotran and
CFX
225
226 POSSIBILITIES OF COMMERCIAL SOFTWARE PACKAGES FLOTRAN AND CFX
F
l
o
t
r
a
n
C
F
X
T
e
c
h
n
i
q
u
e
F
i
n
i
t
e
e
l
e
m
e
n
t
m
e
t
h
o
d
F
i
n
i
t
e
v
o
l
u
m
e
m
e
t
h
o
d
S
p
a
t
i
a
l
d
i
s
c
r
e
t
i
z
a
t
i
o
n
M
o
n
o
t
o
n
e
S
t
r
e
a
m
l
i
n
e
U
p
w
i
n
d
(
M
S
U
)
U
p
w
i
n
d
D
i
e
r
e
n
c
e
S
c
h
e
m
e
(
U
D
S
)
S
t
r
e
a
m
l
i
n
e
U
p
w
i
n
d
P
e
t
r
o
v
-
G
a
l
e
r
k
i
n
(
S
U
P
G
)
S
p
e
c
i
e
d
b
l
e
n
d
C
o
l
l
o
c
a
t
e
d
G
a
l
e
r
k
i
n
(
C
O
L
G
)
H
i
g
h
R
e
s
o
l
u
t
i
o
n
(
H
i
R
e
s
)
T
i
m
e
d
i
s
c
r
e
t
i
z
a
t
i
o
n
b
a
c
k
w
a
r
d
E
u
l
e
r
b
a
c
k
w
a
r
d
E
u
l
e
r
N
e
w
m
a
r
k
t
h
r
e
e
-
p
o
i
n
t
b
a
c
k
w
a
r
d
d
i
e
r
e
n
c
e
(
B
D
F
2
)
A
l
g
o
r
i
t
h
m
s
S
e
m
i
-
I
m
p
l
i
c
i
t
M
e
t
h
o
d
f
o
r
P
r
e
s
s
u
r
e
-
L
i
n
k
e
d
E
q
u
a
t
i
o
n
s
(
S
I
M
P
L
E
)
c
o
u
p
l
e
d
s
o
l
v
e
r
T
u
r
b
u
l
e
n
c
e
Z
e
r
o
E
q
u
a
t
i
o
n
Z
e
r
o
E
q
u
a
t
i
o
n
k
R
N
G
k
R
N
G
k
S
h
e
a
r
S
t
r
e
s
s
T
r
a
n
s
p
o
r
t
(
S
S
T
)
S
h
e
a
r
S
t
r
e
s
s
T
r
a
n
s
p
o
r
t
(
S
S
T
)
S
h
i
h
-
Z
h
u
-
L
u
m
l
e
y
(
S
Z
L
)
R
e
a
l
i
z
a
b
l
e
k
(
N
K
E
)
N
o
n
l
i
n
e
a
r
M
o
d
e
l
o
f
G
i
r
i
m
a
j
i
(
G
I
R
)
R
e
y
n
o
l
d
s
S
t
r
e
s
s
M
o
d
e
l
(
R
S
M
)
D
e
t
a
c
h
e
d
E
d
d
y
S
i
m
u
l
a
t
i
o
n
(
D
E
S
)
L
a
r
g
e
E
d
d
y
S
i
m
u
l
a
t
i
o
n
(
L
E
S
)
W
a
l
l
f
u
n
c
t
i
o
n
s
e
q
u
i
l
i
b
r
i
u
m
s
t
a
n
d
a
r
d
V
a
n
D
r
i
e
s
t
s
c
a
l
a
b
l
e
/
a
u
t
o
m
a
t
i
c
S
p
a
l
d
i
n
g
Table A.1: Overview of the dierent possibilities for CFD computations within
Flotran 10.0 and CFX 10.0.
POSSIBILITIES OF COMMERCIAL SOFTWARE PACKAGES FLOTRAN AND CFX 227
F
l
o
t
r
a
n
C
F
X
D
e
s
c
r
i
p
t
i
o
n
a
r
b
i
t
r
a
r
y
L
a
g
r
a
n
g
i
a
n
-
E
u
l
e
r
i
a
n
(
A
L
E
)
a
r
b
i
t
r
a
r
y
L
a
g
r
a
n
g
i
a
n
-
E
u
l
e
r
i
a
n
(
A
L
E
)
M
e
s
h
d
e
f
o
r
m
a
t
i
o
n
p
s
e
u
d
o
-
s
t
r
u
c
t
u
r
e
d
i
u
s
i
o
n
e
q
u
a
t
i
o
n
C
o
u
p
l
i
n
g
a
l
g
o
r
i
t
h
m
s
p
a
r
t
i
t
i
o
n
e
d
p
a
r
t
i
t
i
o
n
e
d
i
t
e
r
a
t
i
v
e
l
y
s
t
a
g
g
e
r
e
d
i
t
e
r
a
t
i
v
e
l
y
s
t
a
g
g
e
r
e
d
x
e
d
r
e
l
a
x
a
t
i
o
n
f
a
c
t
o
r
x
e
d
r
e
l
a
x
a
t
i
o
n
f
a
c
t
o
r
s
u
b
c
y
c
l
i
n
g
s
u
b
c
y
c
l
i
n
g
s
e
r
i
a
l
s
e
r
i
a
l
o
r
p
a
r
a
l
l
e
l
L
o
a
d
t
r
a
n
s
f
e
r
c
o
n
s
i
s
t
e
n
t
i
n
t
e
r
p
o
l
a
t
i
o
n
c
o
n
s
i
s
t
e
n
t
i
n
t
e
r
p
o
l
a
t
i
o
n
n
o
d
e
t
o
e
l
e
m
e
n
t
c
o
n
s
e
r
v
a
t
i
v
e
n
o
d
e
t
o
e
l
e
m
e
n
t
c
o
n
s
e
r
v
a
t
i
v
e
M
o
t
i
o
n
t
r
a
n
s
f
e
r
c
o
n
s
i
s
t
e
n
t
i
n
t
e
r
p
o
l
a
t
i
o
n
c
o
n
s
i
s
t
e
n
t
i
n
t
e
r
p
o
l
a
t
i
o
n
Table A.2: Overview of the dierent possibilities to couple Ansys 10.0 with
Flotran 10.0 and CFX 10.0.
Appendix B
Finite strip model of the silo
A uid-structure interaction analysis aims to predict the ovalling onset ow velocity
and to investigate the inuence of the distance between the silos. As the coupled
three-dimensional calculation of turbulent wind ow around a group of silos is too
demanding from the computational point of view, an approximate analysis will be
performed by reducing the structure to two dimensions and by coupling it with a
two dimensional ow.
Using a nite strip formulation (Cheung, 1976), the displacements of the three-
dimensional structure are decomposed into a series of orthogonal functions that
satisfy a priori the Dirichlet boundary conditions in the axial z-direction and a two-
dimensional displacement eld in the (r, )-plane. The use of orthogonal functions
results in a decoupled system of equations for every term in the series. The sine
functions reect that the radial and circumferential displacements are assumed to
be zero at both ends of the cylinder, while the cosine function allows for free axial
displacements at both ends:
_
_
u
r
(r, , z, t)
u
(r, , z, t)
u
z
(r, , z, t)
_
_
=
m=0
_
_
sin(
mz
h
) 0 0
0 sin(
mz
h
) 0
0 0 cos(
mz
h
)
_
_
_
_
u
rm
(r, , t)
u
m
(r, , t)
u
zm
(r, , t)
_
_
(B.1)
In reality, however, the cylinder is welded to the cones at both ends and bolted to
the supporting structure at the bottom in four points along the circumference.
As the experimental results indicate that the eigenmodes with the highest modal
contributions consists of half a wavelength along the height (m = 1), the series in
equation (B.1) is limited to the term m = 1.
229
230 FINITE STRIP MODEL OF THE SILO
Using equation (B.1), the in-plane stiness matrix for a nite strip membrane
element for m = 1 is obtained (Cheung, 1976):
K
IP
= t
_
_
E
p
h/2
L
+ G
L
2
6h
symmetric
E
p
4
G
4
E
p
L
2
6h
+ G
h
2L
E
p
h/2
L
+ G
L
2
12h
E
p
4
G
4
E
p
h/2
L
+ G
L
2
6h
E
1
2
4
+ G
4
E
p
L
2
12h
G
h
2L
E
p
4
+ G
4
E
p
L
2
6h
+ G
h
2L
_
_
(B.2)
where E
p
and G are:
E
p
=
E
1
2
(B.3)
G =
E
2(1 + )
(B.4)
The out-of-plane stiness matrix for a nite strip plate-bending element for m = 1
consists of:
[K
OOP
]
11
= [K
OOP
]
33
=
13L
4
D
70h
3
+
12
2
D
xy
5Lh
+
6
2
D
5Lh
+
12
h
2
D
L
3
(B.5)
[K
OOP
]
21
= [K
OOP
]
43
=
3
2
D
5h
+

2
D
xy
5h
+
11L
2
4
D
420h
3
+
6
h
2
D
L
2
(B.6)
[K
OOP
]
22
= [K
OOP
]
44
=
L
3
4
D
210h
3
+
4L
2
D
xy
15h
+
2L
2
D
15h
+
4
h
2
D
L
(B.7)
[K
OOP
]
31
=
9L
4
D
140h
3

12
2
D
xy
5Lh

6
2
D
5Lh

12
h
2
D
L
3
(B.8)
[K
OOP
]
32
= [K
OOP
]
41
=
13L
2
4
D
840h
3

2
D
xy
5h

2
D
10h

6
h
2
D
L
2
(B.9)
[K
OOP
]
42
=
3L
3
4
D
840h
3

L
2
D
xy
15h

L
2
D
30h
+
2
h
2
D
L
(B.10)
where the bending stiness D and D
xy
are equal to:
D =
Et
3
12(1
2
)
(B.11)
D
xy
=
Gt
3
12
(B.12)
FINITE STRIP MODEL OF THE SILO 231
The in-plane mass matrix for a nite strip membrane element for m = 1 is:
M
IP
= t
h
2
_
_
L
3
0
L
3
L
6
0
L
3
0
L
6
0
L
3
_
_
(B.13)
The out-of-plane mass matrix for a nite strip plate-bending element for m = 1 is:
M
OOP
= t
h
2
_
_
13L
35
11L
2
210
L
3
105
9L
70
13L
2
420
13L
35
13L
2
420

3L
1
140

11L
2
210
L
3
105
_
_
(B.14)
The stiness and mass matrix for a nite strip shell element are obtained by
combining the matrices for in-plane and out-of-plane behaviour. The resulting
2-node element has four degrees of freedom per node: the displacements in the
radial, circumferential and axial directions and the rotation around the z-axis.
All terms marked in grey in the stiness and mass matrices of equations (B.2)-
(B.14) are similar to terms of a two-dimensional beam element. The nite strip
element is implemented in the nite element program Ansys as a combination of a
two-dimensional BEAM3 (Ansys, 2005b) element and a user dened MATRIX27
element, which adds all the black terms. The terms marked in grey become
identical to those of a two-dimensional beam element if the area and the moment
of inertia are computed based on the thickness and half the height of the nite
strip:
A =
(h/2)t
1
2
(B.15)
I =
(h/2)t
3
12(1
2
)
(B.16)
The area and moment of inertia are divided by (1
2
) to incorporate the increase
in stiness caused by the Poisson eect. The density is multiplied by a factor
(1
2
) to obtain the correct mass:
= (1
2
) (B.17)
The stiness matrix of the user dened elements is independent of the deformation,
so that geometrical non-linear eects are not included in these additional terms.
Figure B.1 shows the relative error on the predicted eigenfrequencies as a function
of the number of elements along the circumference. A reference solution is obtained
based on a very ne mesh with 1152 elements in the circumferential direction. The
232 FINITE STRIP MODEL OF THE SILO
72 90 144 288
10
6
10
5
10
4
10
3
10
2
Number of elements []
R
e
l
a
t
i
v
e

e
r
r
o
r

[
]
(1,4)
(1,3)
(1,5)
(1,2)
(1,6)
(1,7)
(1,8)
(1,9)
(1,10)
Figure B.1: The relative error on the nine lowest eigenfrequencies predicted with
the nite strip model as a function of the number of elements in the
circumferential direction.
relative errors on the nine lowest eigenfrequencies, predicted with a nite strip
model with 90 elements along the circumference of the silo are lower than 10
3
,
so that this model can be used with condence.
In the nite strip method, a constant thickness of the aluminium shell along the
height of the silo is assumed, whereas, in reality, this thickness reduces with the
height (table 3.1). Furthermore, the radial modal displacement varies as a sine
function along the height. As the degree of clamping at the top and the bottom
of the silo increases with increasing circumferential wave number n, an equivalent
height can be derived that depends on n (Arnold and Warburton, 1953). A better
prediction accuracy could be pursued by formulating an optimization problem
where the objective function is dened as the weighted squared dierence between
the predicted and experimental eigenfrequencies and minimized with respect to
the (uniform) shell thickness t and the height h in the Fourier expansion (B.1) of
the displacements. Alternatively, a sensitivity analysis is performed here where
the eigenfrequencies predicted with three nite strip models with dierent shell
thickness and height are compared with the results of the three-dimensional nite
element model (table B.1). Every couple (m, n) corresponds to two eigenmodes,
that have the same eigenfrequencies in case of the nite strip model and may have
slightly dierent eigenfrequencies for the three-dimensional nite element model.
For a model with a height of 25 m and a thickness of 7 mm (gure B.2), the
eigenfrequencies for n = 3 and n = 4 correspond well with the results of the three-
dimensional calculation, while the dierences between both models increase for
higher values of n. The correspondence for n = 3 and n = 4 improves by changing
the height and the shell thickness to 25.34 m and 7.215 mm, respectively (table
B.1).
FINITE STRIP MODEL OF THE SILO 233
n 3D model Finite strip model
h = 24.70 m h = 25.00 m h = 25.34 m
t = 6.600 mm t = 7.000 mm t = 7.215 mm
2 7.75 8.48 7.67 7.50 7.31
3 3.93 3.93 4.03 4.00 3.93
4 3.93 4.04 3.80 3.93 3.99
5 5.25 5.25 5.25 5.54 5.69
6 7.37 7.37 7.49 7.93 8.17
7 9.72 9.72 10.2 10.8 11.2
8 12.5 12.5 13.4 14.2 14.7
9 15.7 15.7 17.1 18.1 18.7
10 19.4 19.4 21.1 21.2 23.1
Table B.1: Comparison of the eigenfrequencies computed with the three-
dimensional nite element model and the nite strip models for varying
n and m = 1.
a. (1, 4) b. (1, 3) c. (1, 5) d. (1, 2)
3.93 Hz 4.00 Hz 5.54 Hz 7.50 Hz
Figure B.2: Eigenmodes of a silo with a height h = 25 m and a thickness t = 7 mm,
computed with a nite strip model.
As the degree of clamping increases mainly at the bottom of the cylinder with
increasing circumferential wave number n and as the shell thickness decreases
with the height, the eigenfrequencies of the eigenmodes with high circumferential
wave number n are better approximated by nite strip models with a lower height
and a smaller thickness. For n = 5, for example, results obtained with a nite
strip model with a height h = 24.7 m and a thickness t = 6.6 mm agree well with
the results of the three-dimensional model (table B.1).
The inuence of the height h of the nite strip model decreases for increasing
values of the circumferential wave number n, while the shell thickness t strongly
aects the eigenfrequencies of the modes with a high circumferential wave number,
but has a minor inuence on the modes corresponding to low values of n.

Tese Silo

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Tese Silo

Uploaded by

Copyright:

Available Formats

KATHOLIEKE UNIVERSITEIT LEUVEN

bna verwaarloosbaar is (Forsberg, 1964; Koga, 1988; Leissa, 1993). Het

( )berekend met algoritme A

. Een pertinente vraag is hoe gevoelig de berekende resultaten

rst order partial time derivative of the variable with

variable in the referential domain

geometric stiness matrix

uctuating part of the variable

variable in the interior of the domain

variable on the uid-structure interface

discretized displacements of the uid mesh on the

discretized displacements in the structure on the

( )berekend met algoritme

valt samen met het stuwpunt. . . . . . . . . . . . . . . . . . xiv

, changing material domain

for a single cylinder using mesh B. . 126

and P/D = 1.3 and (b) = 45

and P/D = 1.1 (Alam et al., 2005). 142

coincides with the stagnation point. 187

( ) computed with algorithm

coincides with the stagnation point. . . . . . . . . . . 189

( ) using algorithm B. The angle = 0

and (c) = 180

is indicated by a dashed line. The mapping can be interpreted

: (X, t) = (X, t) (2.3)

, changing material domain

x(, t) = I which is the case when the mesh

in the referential domain to

in the referential domain:

. In the continuum equation (2.97), no time derivative occurs.

is constant, the equation is rewritten as:

, x) is equal to zero. The

, x) is equal to zero as well. From the

are equal to the square of the eigenvalues

of the right Cauchy-Green deformation tensor C and the stretch

. In contrast to the total

with the horizontal plane, respectively. An octagonal beam supports the

2. The approximated relative error

was not included in the non-linear computations.

was not included in the non-linear computations.

appears in the momentum equations.

are conventionally called the

y/. This dimensionless

G(r, x)v(x r)dr (4.9)

. If the specied lter width is equal

v v, and the product of the

can be split in its isotropic and deviatoric part:

the normal Reynolds stresses are always

(equation (4.17)) cannot be

are modelled as a function of

. The drag coecient

, which is explained by the assumption that the

for a single cylinder using mesh B.

2 ner than mesh A

2 coarser than mesh A

. Near the minimum,

. In the recirculation zone the largest dierences

are similar to those for 0

), the convergence is oscillatory or divergent. Near the

) the solution diverges as well. For mesh

the apparent order becomes very large.

and between 326

, as the pressure coecient is