o
ENSC2007/76
TH
`
ESE DE DOCTORAT
DE L’
´
ECOLE NORMALE SUP
´
ERIEURE DE CACHAN
Pr´esent´ee par
Denys DUTYKH
pour obtenir le grade de
DOCTEUR DE L’
´
ECOLE NORMALE SUP
´
ERIEURE DE
CACHAN
Domaine:
Math´ematiques Appliqu´ees
Sujet de la th`ese:
MOD
´
ELISATION MATH
´
EMATIQUE DES TSUNAMIS
Th`ese pr´esent´ee et soutenue `a Cachan le 3 d´ecembre 2007 devant le jury compos´e de:
M. JeanMichel GHIDAGLIA Examinateur
M. JeanClaude SAUT Rapporteur et Pr´esident
M. Didier BRESCH Rapporteur
M. Costas SYNOLAKIS Examinateur
M. Vassilios DOUGALIS Examinateur
M. Daniel BOUCHE Membre invit´e
M. Fr´ed´eric DIAS Directeur de th`ese
Centre de Math´ematiques et de Leurs Applications
ENS CACHAN/CNRS/UMR 8536
61, avenue du Pr´esident Wilson, 94235 CACHAN CEDEX (France)
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
To my parents, my sister and A.M.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
Remerciements
Je tiens `a remercier, en tout premier lieu, Fr´ed´eric Dias qui a encadr´e ce travail. Il
m’a beaucoup soutenu et encourag´e au cours de ces magniﬁques ann´ees de recherche. Je
lui en suis tr`es profond´ement reconnaissant.
Je remercie tout particuli`erement JeanMichel Ghidaglia qui m’a beaucoup guid´e dans
mes recherches. Je tiens `a souligner que par sa grande disponibilit´e, j’ai pu b´en´eﬁcier de
ses tr`es grandes comp´etences scientiﬁques tout au long de mes investigations. Enﬁn, c’est
un honneur pour moi qu’il ait accept´e de pr´esider le jury de cette th`ese.
J’exprime ma gratitude aux Professeurs JeanClaude Saut et Didier Bresch, qui ont
accept´e d’ˆetre les rapporteurs, ainsi qu’aux Professeurs Costas Synolakis et Vassilios Dou
galis pour leur int´erˆet qu’ils ont port´e `a mon travail en ayant accept´e de faire partie du jury.
Mes trois ann´ees de th`ese se sont d´eroul´ees dans un contexte scientiﬁque de grande
qualit´e au Centre de Math´ematiques et de Leurs Applications de l’ENS Cachan. Je re
mercie son directeur Laurent Desvillettes de m’avoir accueilli et d’avoir cr´ee d’excellentes
conditions de travail. Parmi tout le personnel de ce laboratoire je voudrais plus partic
uli`erement remercier Micheline Brunetti pour son aimabilit´e et eﬃcacit´e dans le traitement
des questions administratives.
Je remercie Fr´ed´eric Pascal pour son tutorat p´edagogique. Ses conseils m’ont beau
coup servis lors de trois ans d’enseignement du calcul scientiﬁque aux agr´egatifs de l’ENS
Cachan. J’en proﬁte aussi pour remercier tous mes ´el`eves de leur assiduit´e et intelligence.
J’exprime mes plus vifs remerciements `a Daniel Bouche de m’avoir accueilli au sein de
LRC CEA/CMLA. Ses remarques pertinentes et conseils avis´es, sa gentillesse m’ont beau
coup aid´e et motiv´e durant le DEA MN2MC et plus particuli`erement pendant la r´ealisation
de cette th`ese.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
ii REMERCIEMENTS
Parmi tous mes coll`egues de LRC, Christophe Fochesato, JeanMichel Rovarch et Denis
Gueyﬃer ont occup´e une place tr`es particuli`ere: en plus de leur amiti´e, ils m’ont oﬀert de
nombreux conseils amicaux et professionnels. Je vous en remercie tr`es sinc`erement.
Parmi les personnes qui ont contribu´e `a ma formation, je suis tout particuli`erement
reconnaissant envers Vladimir Lamzyuk qui m’a fait d´ecouvrir le monde de la mod´elisation
math´ematique et des math´ematiques appliqu´ees en g´en´eral. Il m’a soutenu constamment
durant mes ´etudes `a l’Universit´e Nationale de Dniprop´etrovsk.
Mes pens´ees vont ´egalement `a tous mes concurrents qui m’ont constamment pouss´e
dans mes recherches et m’ont donn´e les forces pour continuer ce travail.
Je dis grand merci `a Yann, Joachim, Lavinia, Serge, Nanthilde, Thomas, Alexandra
et aux autres doctorants du LMT pour leur amiti´e et tous les instants de d´etente partag´es
ensemble. Je tiens `a remercier particuli`erement Alexandra Pradeau de nos discussions pas
sionnantes ainsi qu’enrichissantes. Sans ces personnes cette th`ese n’aurait pas eu la mˆeme
saveur.
Je remercie ma famille et leur d´edie cette th`ese pour leurs encouragements permanents
et toutes autres raisons qui s’imposent.
Je conclurai en remerciant de tout coeur Anna Maria pour la patience et le soutien sans
limite dont elle fait preuve quotidiennement.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
R´esum´e
Cette th`ese est consacr´ee `a la mod´elisation des tsunamis. La vie de ces vagues peut
ˆetre conditionnellement divis´ee en trois parties: g´en´eration, propagation et inondation.
Dans un premier temps, nous nous int´eressons `a la g´en´eration de ces vagues extrˆemes.
Dans cette partie du m´emoire, nous examinons les diﬀ´erentes approches existantes pour
la mod´elisation, puis nous en proposons d’autres. La conclusion principale `a laquelle nous
sommes arriv´es est que le couplage entre la sismologie et l’hydrodynamique est actuellement
assez mal compris.
Le deuxi`eme chapitre est d´edi´e essentiellement aux ´equations de Boussinesq qui sont
souvent utilis´ees pour mod´eliser la propagation d’un tsunami. Certains auteurs les utilisent
mˆeme pour mod´eliser le processus d’inondation (le runup). Plus pr´ecisement, nous discu
tons de l’importance, de la nature et de l’inclusion des eﬀets dissipatifs dans les mod`eles
d’ondes longues.
Dans le troisi`eme chapitre, nous changeons de sujet et nous nous tournons vers les
´ecoulements diphasiques. Le but de ce chapitre est de proposer un mod`ele simple et
op´erationnel pour la mod´elisation de l’impact d’une vague sur les structures cˆoti`eres. En
suite, nous discutons de la discr´etisation num´erique de ces ´equations avec un sch´ema de
type volumes ﬁnis sur des maillages non structur´es.
Finalement, le m´emoire se termine par un sujet qui devrait ˆetre pr´esent dans tous les
manuels classiques d’hydrodynamique mais qui ne l’est pas. Nous parlons des ´ecoulements
viscopotentiels. Nous proposons une nouvelle approche simpliﬁ´ee pour les ´ecoulements
faiblement visqueux. Nous conservons la simplicit´e des ´ecoulements potentiels tout en
ajoutant la dissipation. Dans le cas de la profondeur ﬁnie nous incluons un terme correcteur
dˆ u `a la pr´esence de la couche limite au fond. Cette correction s’av`ere ˆetre non locale en
temps. Donc, la couche limite au fond apporte un certain eﬀet de m´emoire `a l’´ecoulement.
Mots cl´es: Ondes de surface, g´en´eration des tsunamis, ´equations de Boussi
nesq, ´ecoulements diphasiques, ´ecoulements viscopotentiels, volumes ﬁnis
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
Abstract
This thesis is devoted to tsunami wave modelling. The life of tsunami waves can be
conditionally divided into three parts: generation, propagation and inundation (or run
up). In the ﬁrst part of the manuscript we consider the generation process of such extreme
waves. We examine various existing approaches to its modelling. Then we propose a few
alternatives. The main conclusion is that the seismology/hydrodynamics coupling is poorly
understood at the present time.
The second chapter essentially deals with Boussinesq equations which are often used
to model tsunami propagation and sometimes even runup. More precisely, we discuss the
importance, nature and inclusion of dissipative eﬀects in long wave models.
In the third chapter we slightly change the subject and turn to twophase ﬂows. The
main purpose of this chapter is to propose an operational and simple set of equations in
order to model wave impacts on coastal structures. Another important application includes
wave sloshing in liquiﬁed natural gas carriers. Then, we discuss the numerical discretization
of governing equations in the ﬁnite volume framework on unstructured meshes.
Finally, this thesis deals with a topic which should be present in any textbook on hydro
dynamics but it is not. We mean viscopotential ﬂows. We propose a novel and suﬃciently
simple approach for weakly viscous ﬂow modelling. We succeeded in keeping the simplicity
of the classical potential ﬂow formulation with the addition of viscous eﬀects. In the case of
ﬁnite depth we derive a correction term due to the presence of the bottom boundary layer.
This term is nonlocal in time. Hence, the bottom boundary layer introduces a memory
eﬀect to the governing equations.
Keywords: Water waves, tsunami generation, Boussinesq equations, two
phase ﬂows, viscopotential ﬂows, ﬁnite volumes
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
Contents
Remerciements i
Contents vii
List of Figures xiii
List of Tables xv
Introduction xvii
Propagation of tsunamis . . . . . . . . . . . . . . . . . . . . . . . . . . . . xviii
Classical formulation . . . . . . . . . . . . . . . . . . . . . . . . . . xx
Dimensionless equations . . . . . . . . . . . . . . . . . . . . . . . . xxi
Shallowwater equations . . . . . . . . . . . . . . . . . . . . . . . . xxi
Boussinesq equations . . . . . . . . . . . . . . . . . . . . . . . . . . xxiii
Korteweg–de Vries equation . . . . . . . . . . . . . . . . . . . . . . xxvi
Energy of a tsunami . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxvi
Tsunami runup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxvii
1 Tsunami generation 1
1.1 Waves generated by a moving bottom . . . . . . . . . . . . . . . . . . . . . 2
1.1.1 Source model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.2 Volterra’s theory of dislocations . . . . . . . . . . . . . . . . . . . . 5
1.1.3 Dislocations in elastic halfspace . . . . . . . . . . . . . . . . . . . . 6
1.1.3.1 Finite rectangular source . . . . . . . . . . . . . . . . . . . 10
1.1.3.2 Curvilinear fault . . . . . . . . . . . . . . . . . . . . . . . 12
1.1.4 Solution in ﬂuid domain . . . . . . . . . . . . . . . . . . . . . . . . 16
1.1.5 Freesurface elevation . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.1.6 Velocity ﬁeld . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.1.6.1 Pressure on the bottom . . . . . . . . . . . . . . . . . . . 28
1.1.7 Asymptotic analysis of integral solutions . . . . . . . . . . . . . . . 29
vii
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
viii CONTENTS
1.1.8 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.2 Comparison of tsunami generation models . . . . . . . . . . . . . . . . . . 36
1.2.1 Physical problem description . . . . . . . . . . . . . . . . . . . . . . 37
1.2.2 Linear theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
1.2.3 Active generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
1.2.4 Passive generation . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
1.2.4.1 Numerical method for the linear problem . . . . . . . . . . 47
1.2.5 Nonlinear shallow water equations . . . . . . . . . . . . . . . . . . . 47
1.2.6 Mathematical model . . . . . . . . . . . . . . . . . . . . . . . . . . 48
1.2.7 Numerical method . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
1.2.8 Numerical method for the full equations . . . . . . . . . . . . . . . 50
1.2.9 Comparisons and discussion . . . . . . . . . . . . . . . . . . . . . . 51
1.2.10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
1.3 Tsunami generation by dynamic displacement of sea bed . . . . . . . . . . 64
1.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
1.3.2 Mathematical models . . . . . . . . . . . . . . . . . . . . . . . . . . 67
1.3.2.1 Dynamic fault model . . . . . . . . . . . . . . . . . . . . . 67
1.3.2.2 Fluid layer model . . . . . . . . . . . . . . . . . . . . . . . 69
1.3.3 Numerical methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
1.3.3.1 Discretization of the viscoelastodynamic equations . . . . 70
1.3.3.2 Finitevolume scheme . . . . . . . . . . . . . . . . . . . . 71
1.3.4 Validation of the numerical method . . . . . . . . . . . . . . . . . . 72
1.3.5 Results of the simulation . . . . . . . . . . . . . . . . . . . . . . . . 73
1.3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
2 Dissipative Boussinesq equations 81
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
2.2 Derivation of the Boussinesq equations . . . . . . . . . . . . . . . . . . . . 85
2.2.1 Asymptotic expansion . . . . . . . . . . . . . . . . . . . . . . . . . 89
2.3 Analysis of the linear dispersion relations . . . . . . . . . . . . . . . . . . . 93
2.3.1 Linearized potential ﬂow equations . . . . . . . . . . . . . . . . . . 93
2.3.2 Dissipative Boussinesq equations . . . . . . . . . . . . . . . . . . . 95
2.3.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
2.4 Alternative version of the Boussinesq equations . . . . . . . . . . . . . . . 99
2.4.1 Derivation of the equations . . . . . . . . . . . . . . . . . . . . . . . 100
2.5 Improvement of the linear dispersion relations . . . . . . . . . . . . . . . . 104
2.6 Regularization of Boussinesq equations . . . . . . . . . . . . . . . . . . . . 105
2.7 Bottom friction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
CONTENTS ix
2.8 Spectral Fourier method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
2.8.1 Validation of the numerical method . . . . . . . . . . . . . . . . . . 110
2.9 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
2.9.1 Construction of the initial condition . . . . . . . . . . . . . . . . . . 113
2.9.2 Comparison between the dissipative models . . . . . . . . . . . . . 115
2.10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
3 Two phase ﬂows 121
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
3.2 Mathematical model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
3.2.1 Sound speed in the mixture . . . . . . . . . . . . . . . . . . . . . . 126
3.2.2 Equation of state . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
3.3 Formal limit in barotropic case . . . . . . . . . . . . . . . . . . . . . . . . 129
3.4 Finite volume scheme on unstructured meshes . . . . . . . . . . . . . . . . 131
3.4.1 Sign matrix computation . . . . . . . . . . . . . . . . . . . . . . . . 136
3.4.2 Second order scheme . . . . . . . . . . . . . . . . . . . . . . . . . . 137
3.4.2.1 Historical remark . . . . . . . . . . . . . . . . . . . . . . . 137
3.4.3 TVD and MUSCL schemes . . . . . . . . . . . . . . . . . . . . . . . 138
3.4.3.1 GreenGauss gradient reconstruction . . . . . . . . . . . . 140
3.4.3.2 Leastsquares gradient reconstruction method . . . . . . . 141
3.4.3.3 Slope limiter . . . . . . . . . . . . . . . . . . . . . . . . . 143
3.4.4 Diﬀusive ﬂuxes computation . . . . . . . . . . . . . . . . . . . . . . 144
3.4.5 Solution interpolation to mesh nodes . . . . . . . . . . . . . . . . . 146
3.4.6 Time stepping methods . . . . . . . . . . . . . . . . . . . . . . . . . 148
3.4.7 Boundary conditions treatment . . . . . . . . . . . . . . . . . . . . 151
3.5 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
3.5.1 Convergence test . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
3.5.2 Falling water column . . . . . . . . . . . . . . . . . . . . . . . . . . 157
3.5.3 Water drop test case . . . . . . . . . . . . . . . . . . . . . . . . . . 157
3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
4 Viscous potential ﬂows 169
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
4.2 Anatomy of dissipation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
4.3 Derivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
4.3.1 Dissipative KdV equation . . . . . . . . . . . . . . . . . . . . . . . 179
4.4 Dispersion relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
4.4.1 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
x CONTENTS
4.5 Attenuation of linear progressive waves . . . . . . . . . . . . . . . . . . . . 182
4.6 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
4.6.1 Approximate solitary wave solution . . . . . . . . . . . . . . . . . . 187
4.6.2 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
4.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
Direction for future research 195
A VFFC scheme 197
A.1 Discretization in the ﬁnite volume framework . . . . . . . . . . . . . . . . 198
A.1.1 The one dimensional case . . . . . . . . . . . . . . . . . . . . . . . 198
A.1.2 Extension to the multidimensional case . . . . . . . . . . . . . . . . 201
A.1.3 On the discretization of source terms . . . . . . . . . . . . . . . . . 202
A.2 On the discretization of boundary conditions . . . . . . . . . . . . . . . . . 204
Bibliography 206
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
List of Figures
1.1 Coordinate system adopted in this study. . . . . . . . . . . . . . . . . . . . 7
1.2 Illustration for strike angle deﬁnition. . . . . . . . . . . . . . . . . . . . . . 8
1.3 Geometry of the source model. . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4 Dipslip fault . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.5 Strikeslip fault . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.6 Tensile fault . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.7 Geometry of a fault with elliptical shape . . . . . . . . . . . . . . . . . . . 15
1.8 Freesurface deformation due to curvilinear faulting . . . . . . . . . . . . . 17
1.9 Deﬁnition of the ﬂuid domain and coordinate system . . . . . . . . . . . . 17
1.10 Plot of the frequency ω(m) and its derivatives dω/dm, d
2
ω/dm
2
. . . . . . 21
1.11 Typical graphs of T
e
(t) and T
c
(t) . . . . . . . . . . . . . . . . . . . . . . . 24
1.12 Freesurface elevation at t = 0.01, 0.6, 3, 5 in dimensionless time . . . . . . 33
1.13 Velocity ﬁeld computed along the free surface at t = 0.01 . . . . . . . . . . 34
1.14 Bottom pressure at t = 0.01, 0.6, 3, 5 in dimensionless time . . . . . . . . . 35
1.15 Typical seaﬂoor deformation due to dipslip faulting . . . . . . . . . . . . . 39
1.16 Initial net volume V of the seaﬂoor displacement . . . . . . . . . . . . . . . 39
1.17 Wave proﬁles at diﬀerent times for the case of a normal fault . . . . . . . . 41
1.18 Comparisons of the freesurface elevation for S = 5 . . . . . . . . . . . . . 53
1.19 Comparisons of the freesurface elevation for S = 0.04 . . . . . . . . . . . . 54
1.20 Top view of the initial free surface deformation . . . . . . . . . . . . . . . . 55
1.21 Transient waves generated by an underwater earthquake . . . . . . . . . . 56
1.22 Comparisons of the freesurface elevation for S = 0.005 . . . . . . . . . . . 58
1.23 Same as Figure 1.22 for later times (t = 150 s, t = 200 s). . . . . . . . . . . 59
1.24 Transient waves generated by an underwater earthquake . . . . . . . . . . 60
1.25 Transient waves, the water depth is h = 500 m . . . . . . . . . . . . . . . . 61
1.26 Same as Figure 1.25, h = 1km . . . . . . . . . . . . . . . . . . . . . . . . . 62
1.27 Relative diﬀerence between the two solutions  I . . . . . . . . . . . . . . . 63
1.28 Relative diﬀerence between the two solutions  II . . . . . . . . . . . . . . 63
xi
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
xii LIST OF FIGURES
1.29 Geometry of the source model. . . . . . . . . . . . . . . . . . . . . . . . . . 66
1.30 Comparison between analytical and numerical solutions . . . . . . . . . . . 73
1.31 Static deformation due to the dislocation. . . . . . . . . . . . . . . . . . . 74
1.32 Water free surface at the beginning of the earthquake (t = 2s). . . . . . . . 76
1.33 Same as Figure 1.32 for times t = 4s and t = 6s. . . . . . . . . . . . . . . . 76
1.34 Same as Figure 1.32 for times t = 7s and t = 10s. . . . . . . . . . . . . . . 77
1.35 Same as Figure 1.32 for times t = 150s and t = 250s. . . . . . . . . . . . . 77
1.36 Same as Figure 1.35 (left plot) for two rupture velocities. . . . . . . . . . . 77
1.37 Same as Figure 1.36 at time t = 250s. . . . . . . . . . . . . . . . . . . . . . 78
1.38 Water free surface at t = 150s for a longer earthquake. . . . . . . . . . . . 78
2.1 Sketch of the ﬂuid domain . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
2.2 Dissipation model I. Real part of the phase velocity. . . . . . . . . . . . . . 96
2.3 Dissipation model I. Zoom on long waves . . . . . . . . . . . . . . . . . . . 96
2.4 Dissipation model I. Imaginary part of the frequency. . . . . . . . . . . . . 97
2.5 Dissipation model II. Real part of the phase velocity. . . . . . . . . . . . . 97
2.6 Dissipation model II. Imaginary part of the frequency. . . . . . . . . . . . . 98
2.7 Dissipation model II. Zoom on long waves . . . . . . . . . . . . . . . . . . 98
2.8 Regularized dispersion relation . . . . . . . . . . . . . . . . . . . . . . . . . 107
2.9 Error of the numerical method . . . . . . . . . . . . . . . . . . . . . . . . . 113
2.10 Solitary wave and a step interaction . . . . . . . . . . . . . . . . . . . . . . 116
2.11 Freesurface snapshot before the interaction . . . . . . . . . . . . . . . . . 116
2.12 Free surface just before the interaction . . . . . . . . . . . . . . . . . . . . 117
2.13 Beginning of the solitary wave deformation . . . . . . . . . . . . . . . . . . 117
2.14 Initiation of the reﬂected wave separation . . . . . . . . . . . . . . . . . . . 118
2.15 Separation of the reﬂected wave . . . . . . . . . . . . . . . . . . . . . . . . 118
2.16 Two separate waves moving in opposite directions . . . . . . . . . . . . . . 119
3.1 Example of water/air mixture after wave breaking. . . . . . . . . . . . . . 123
3.2 Sound speed c
s
in the water/air mixture at normal conditions. . . . . . . . 124
3.3 Sketch of the ﬂow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
3.4 An example of control volume K. . . . . . . . . . . . . . . . . . . . . . . . 132
3.5 Illustration for cellcentered ﬁnite volume method . . . . . . . . . . . . . . 133
3.6 Illustration for vertexcentered ﬁnite volume method . . . . . . . . . . . . 134
3.7 Illustration for GreenGauss gradient reconstruction . . . . . . . . . . . . . 140
3.8 Illustration for leastsquares gradient reconstruction. . . . . . . . . . . . . 141
3.9 Diamond cell constructed around face f. . . . . . . . . . . . . . . . . . . . 144
3.10 Triangles sharing the same vertex N. . . . . . . . . . . . . . . . . . . . . . 146
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
LIST OF FIGURES xiii
3.11 Linear absolute stability regions . . . . . . . . . . . . . . . . . . . . . . . . 150
3.12 Nonlinear absolute stability regions . . . . . . . . . . . . . . . . . . . . . . 150
3.13 Control volume sharing a face with boundary ∂Ω. . . . . . . . . . . . . . . 152
3.14 Shock tube of Sod. This plot represents the density. . . . . . . . . . . . . . 153
3.15 Numerical method error in L
∞
norm. . . . . . . . . . . . . . . . . . . . . . 155
3.16 CPU time for diﬀerent ﬁnite volume schemes. . . . . . . . . . . . . . . . . 156
3.17 Geometry and initial condition for falling water . . . . . . . . . . . . . . . 157
3.18 The beginning of column dropping process . . . . . . . . . . . . . . . . . . 158
3.19 Splash creation due to the interaction with step . . . . . . . . . . . . . . . 158
3.20 Water strikes the wall  I . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
3.21 Water strikes the wall  II . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
3.22 Splash is climbing the wall . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
3.23 Turbulent mixing process . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
3.24 Maximal pressure on the right wall. Heavy gas case. . . . . . . . . . . . . . 161
3.25 Maximal pressure on the right wall as a function of time. Light gas. . . . . 162
3.26 Geometry and initial condition for water drop test case . . . . . . . . . . . 163
3.27 Initial conﬁguration and the beginning of the fall . . . . . . . . . . . . . . 163
3.28 Drop approaching container bottom . . . . . . . . . . . . . . . . . . . . . . 164
3.29 Drop/bottom compressible interaction . . . . . . . . . . . . . . . . . . . . 164
3.30 Water drop test case. Vertical jets formation . . . . . . . . . . . . . . . . . 165
3.31 Water drop test case. Side jets crossing . . . . . . . . . . . . . . . . . . . . 165
3.32 Water drop test case. Side jets interﬂow at the center . . . . . . . . . . . . 166
3.33 Central jet reﬂection from the bottom . . . . . . . . . . . . . . . . . . . . . 166
3.34 Water drop test case. Maximal bottom pressure as a function of time. . . . 167
4.1 Conventional partition of the ﬂow region . . . . . . . . . . . . . . . . . . . 172
4.2 Real and imaginary part of dispersion curve at t = 0. . . . . . . . . . . . . 183
4.3 Phase velocity at t = 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
4.4 Phase velocity at t = 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
4.5 Phase velocity at t = 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
4.6 Real part of the phase velocity at t = 4. . . . . . . . . . . . . . . . . . . . 185
4.7 Amplitude of linear progressive waves . . . . . . . . . . . . . . . . . . . . . 188
4.8 Snapshots of the free surface at t = 0.5 . . . . . . . . . . . . . . . . . . . . 190
4.9 Snapshots of the free surface at t = 1.0 . . . . . . . . . . . . . . . . . . . . 190
4.10 Snapshots of the free surface at t = 2.0 . . . . . . . . . . . . . . . . . . . . 191
4.11 Zoom on the soliton crest at t = 2.0 . . . . . . . . . . . . . . . . . . . . . . 191
4.12 Amplitude of the solitary wave as function of time . . . . . . . . . . . . . . 192
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
xiv LIST OF FIGURES
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
List of Tables
1.1 Parameter set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.2 Parameter set used in Figure 1.8. . . . . . . . . . . . . . . . . . . . . . . . 16
1.3 Laplace transforms of diﬀerent timedepencies . . . . . . . . . . . . . . . . 24
1.4 Physical parameters used in the numerical computations . . . . . . . . . . 32
1.5 Parameters set for the source model . . . . . . . . . . . . . . . . . . . . . . 38
1.6 Typical physical parameters used in the numerical computations. . . . . . 74
2.1 Typical parameters in tsunami applications . . . . . . . . . . . . . . . . . . 89
2.2 Solitonstep interaction parameters . . . . . . . . . . . . . . . . . . . . . . 114
3.1 Parameters for an air/water mixture under normal conditions . . . . . . . 128
4.1 Parameters of the linear progressive waves . . . . . . . . . . . . . . . . . . 187
4.2 Parameters used in the numerical computations . . . . . . . . . . . . . . . 188
xv
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
xvi LIST OF TABLES
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
Introduction
An expert is a man who has made all the mistakes,
which can be made, in a very narrow ﬁeld.
Niels Henrik David Bohr (1885 – 1962)
Contents
Propagation of tsunamis . . . . . . . . . . . . . . . . . . . . . . xviii
Classical formulation . . . . . . . . . . . . . . . . . . . . . . . . . xx
Dimensionless equations . . . . . . . . . . . . . . . . . . . . . . . xxi
Shallowwater equations . . . . . . . . . . . . . . . . . . . . . . . xxi
Boussinesq equations . . . . . . . . . . . . . . . . . . . . . . . . . xxiii
Korteweg–de Vries equation . . . . . . . . . . . . . . . . . . . . . xxvi
Energy of a tsunami . . . . . . . . . . . . . . . . . . . . . . . . . xxvi
Tsunami runup . . . . . . . . . . . . . . . . . . . . . . . . . . . xxvii
Given the broadness of the topic of tsunamis [DD06], our purpose here is to recall
some of the basics of tsunami modeling and to emphasize some general aspects, which
are sometimes overlooked. The life of a tsunami is usually divided into three phases:
the generation (tsunami source), the propagation and the inundation. The third and most
diﬃcult phase of the dynamics of tsunami waves deals with their breaking as they approach
the shore. This phase depends greatly on the bottom bathymetry and on the coastline
type. The breaking can be progressive. Then the inundation process is relatively slow and
can last for several minutes. Structural damages are mainly caused by inundation. The
breaking can also be explosive and lead to the formation of a plunging jet. The impact on
the coast is then very rapid. In very shallow water, the amplitude of tsunami waves grows
to such an extent that typically an undulation appears on the long wave, which develops
into a progressive bore [Cha05]. This turbulent front, similar to the wave that occurs when
a dam breaks, can be quite high and travel onto the beach at great speed. Then the front
and the turbulent current behind it move onto the shore, past buildings and vegetation
xvii
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
xviii INTRODUCTION
until they are ﬁnally stopped by rising ground. The water level can rise rapidly, typically
from 0 to 3 meters in 90 seconds.
The trajectory of these currents and their velocity are quite unpredictable, especially
in the ﬁnal stages because they are sensitive to small changes in the topography, and
to the stochastic patterns of the collapse of buildings, and to the accumulation of debris
such as trees, cars, logs, furniture. The dynamics of this ﬁnal stage of tsunami waves is
somewhat similar to the dynamics of ﬂood waves caused by dam breaking, dyke breaking or
overtopping of dykes (cf. the recent tragedy of hurricane Katrina in August 2005). Hence
research on ﬂooding events and measures to deal with them may be able to contribute to
improved warning and damage reduction systems for tsunami waves in the areas of the
world where these waves are likely to occur as shallow surge waves (cf. the recent tragedy
of the Indian Ocean tsunami in December 2004).
Civil engineers who visited the damage area following the Boxing day tsunami came
up with several basic conclusions. Buildings that had been constructed to satisfy mod
ern safety standards oﬀered a satisfactory resistance, in particular those with reinforced
concrete beams properly integrated in the frame structure. These were able to withstand
pressure associated with the leading front of the order of 1 atmosphere (recall that an equiv
alent pressure p is obtained with a windspeed U of about 450 m/s, since p = ρ
air
U
2
/2).
By contrast brick buildings collapsed and were washed away. Highly porous or open struc
tures survived. Buildings further away from the beach survived the front in some cases,
but they were then destroyed by the erosion of the ground around the buildings by the
water currents [HB05].
Propagation of tsunamis
The problem of tsunami propagation is a special case of the general waterwave problem.
The study of water waves relies on several common assumptions. Some are obvious while
some others are questionable under certain circumstances. The water is assumed to be
incompressible. Dissipation is not often included. However there are three main sources of
dissipation for water waves: bottom friction, surface dissipation and body dissipation. For
tsunamis, bottom friction is the most important one, especially in the later stages, and is
sometimes included in the computations in an adhoc way. In most theoretical analyses,
it is not included. The question of dissipation in potential ﬂows and long waves equations
is thoroughly investigated in Chapter 4.
A brief description of the common mathematical model used to study water waves
follows. The horizontal coordinates are denoted by x and y, and the vertical coordinate
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
INTRODUCTION xix
by z. The horizontal gradient is denoted by
∇ :=
_
∂
∂x
,
∂
∂y
_
.
The horizontal velocity is denoted by
u(x, y, z, t) = (u, v)
and the vertical velocity by w(x, y, z, t). The threedimensional ﬂow of an inviscid and
incompressible ﬂuid is governed by the conservation of mass
∇ u +
∂w
∂z
= 0 (1)
and by the conservation of momentum
ρ
Du
Dt
= −∇p, ρ
Dw
Dt
= −ρg −
∂p
∂z
, (2)
where
Df
Dt
is the material derivative deﬁned as
Df
Dt
:=
∂f
∂t
+ u ∇f, u = (u, w) = (u, v, w).
In (2), ρ is the density of water (assumed to be constant throughout the ﬂuid domain), g
is the acceleration due to gravity and p(x, y, z, t) the pressure ﬁeld.
The assumption that the ﬂow is irrotational is commonly made to analyze surface
waves. Then there exists a scalar function φ(x, y, z, t) (the velocity potential) such that
u = ∇φ, w =
∂φ
∂z
.
The continuity equation (1) becomes
∇
2
φ +
∂
2
φ
∂z
2
= 0 . (3)
The equation of momentum conservation (2) can be integrated into Bernoulli’s equation
∂φ
∂t
+
1
2
[∇φ[
2
+
1
2
_
∂φ
∂z
_
2
+ gz +
p −p
0
ρ
= 0 , (4)
which is valid everywhere in the ﬂuid. The constant p
0
is a pressure of reference, for
example the atmospheric pressure. The eﬀects of surface tension are not important for
tsunami propagation.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
xx INTRODUCTION
Classical formulation
The surface wave problem consists in solving Laplace’s equation (3) in a domain Ω(t)
bounded above by a moving free surface (the interface between air and water) and below
by a ﬁxed solid boundary (the bottom).
1
The free surface is represented by F(x, y, z, t) :=
η(x, y, t) − z = 0. The shape of the bottom is given by z = −h(x, y). The main driving
force is gravity.
The free surface must be found as part of the solution. Two boundary conditions are
required. The ﬁrst one is the kinematic condition. It can be stated as DF/Dt = 0 (the
material derivative of F vanishes), which leads to
η
t
+∇φ ∇η −φ
z
= 0 at z = η(x, y, t) . (5)
The second boundary condition is the dynamic condition which states that the normal
stresses must be in balance at the free surface. The normal stress at the free surface is
given by the diﬀerence in pressure. Bernoulli’s equation (4) evaluated on the free surface
z = η gives
φ
t
+
1
2
[∇φ[
2
+
1
2
φ
2
z
+ gη = 0 at z = η(x, y, t) . (6)
Finally, the boundary condition at the bottom is
∇φ ∇h + φ
z
= 0 at z = −h(x, y) . (7)
To summarize, the goal is to solve the set of equations (3), (5), (6) and (7) for η(x, y, t)
and φ(x, y, z, t). When the initial value problem is integrated, the ﬁelds η(x, y, 0) and
φ(x, y, z, 0) must be speciﬁed at t = 0. The conservation of momentum equation (2) is not
required in the solution procedure; it is used a posteriori to ﬁnd the pressure p once η and
φ have been found.
In the following subsections, we will consider various approximations of the full water
wave equations. One is the system of Boussinesq equations, that retains nonlinearity and
dispersion up to a certain order. Another one is the system of nonlinear shallowwater
equations that retains nonlinearity but no dispersion. The simplest one is the system of
linear shallowwater equations. The concept of shallow water is based on the smallness
of the ratio between water depth and wavelength. In the case of tsunamis propagating
on the surface of deep oceans, one can consider that shallowwater theory is appropriate
because the water depth (typically several kilometers) is much smaller than the wavelength
(typically several hundred kilometers).
1
The surface wave problem can be easily extended to the case of a moving bottom. This extension may
be needed to model tsunami generation if the bottom deformation is relatively slow.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
INTRODUCTION xxi
Dimensionless equations
The derivation of shallowwater type equations is a classical topic. Two dimensionless
numbers, which are supposed to be small, are introduced:
α =
a
d
≪1, β =
d
2
ℓ
2
≪1, (8)
where d is a typical water depth, a a typical wave amplitude and ℓ a typical wavelength.
The assumptions on the smallness of these two numbers are satisﬁed for the Indian Ocean
tsunami. Indeed the satellite altimetry observations of the tsunami waves obtained by two
satellites that passed over the Indian Ocean a couple of hours after the rupture occurred
give an amplitude a of roughly 60 cm in the open ocean. The typical wavelength estimated
from the width of the segments that experienced slip is between 160 and 240 km [LKA
+
05].
The water depth ranges from 4 km towards the west of the rupture to 1 km towards the
east. These values give the following ranges for the two dimensionless numbers:
1.5 10
−4
< α < 6 10
−4
, 1.7 10
−5
< β < 6.25 10
−4
. (9)
The equations are more transparent when written in dimensionless variables. The new
independent variables are
x = ℓ˜ x, y = ℓ˜ y, z = d˜ z, t = ℓ
˜
t/c
0
, (10)
where c
0
=
√
gd, the famous speed of propagation of tsunamis in the open ocean ranging
from 356 km/h for a 1 km water depth to 712 km/h for a 4 km water depth. The new
dependent variables are
η = a˜ η, h = d
˜
h, φ = gaℓ
˜
φ/c
0
. (11)
In dimensionless form, and after dropping the tildes, the equations become
β∇
2
φ + φ
zz
= 0, (12)
β∇φ ∇h + φ
z
= 0 at z = −h(x, y), (13)
βη
t
+ αβ∇φ ∇η = φ
z
at z = αη(x, y, t), (14)
βφ
t
+
1
2
αβ[∇φ[
2
+
1
2
αφ
2
z
+ βη = 0 at z = αη(x, y, t). (15)
So far, no approximation has been made. In particular, we have not used the fact that the
numbers α and β are small.
Shallowwater equations
When β is small, the water is considered to be shallow. The linearized theory of water
waves is recovered by letting α go to zero. For the shallow waterwave theory, one assumes
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
xxii INTRODUCTION
that β is small and expand φ in terms of β:
φ = φ
0
+ βφ
1
+ β
2
φ
2
+ .
This expansion is substituted into the governing equation and the boundary conditions.
The lowestorder term in Laplace’s equation is
φ
0zz
= 0. (16)
The boundary conditions imply that φ
0
= φ
0
(x, y, t). Thus the vertical velocity component
is zero and the horizontal velocity components are independent of the vertical coordinate
z at lowest order. Let φ
0x
= u(x, y, t) and φ
0y
= v(x, y, t). Assume now for simplicity that
the water depth is constant (h = 1). Solving Laplace’s equation and taking into account
the bottom kinematic condition yields the following expressions for φ
1
and φ
2
:
φ
1
(x, y, z, t) = −
1
2
(1 + z)
2
(u
x
+ v
y
), (17)
φ
2
(x, y, z, t) =
1
24
(1 + z)
4
[(∇
2
u)
x
+ (∇
2
v)
y
]. (18)
The next step consists in retaining terms of requested order in the freesurface boundary
conditions. Powers of α will appear when expanding in Taylor series the freesurface
conditions around z = 0. For example, if one keeps terms of order αβ and β
2
in the dynamic
boundary condition (15) and in the kinematic boundary condition (14), one obtains
βφ
0t
−
1
2
β
2
(u
tx
+ v
ty
) + βη +
1
2
αβ(u
2
+ v
2
) = 0, (19)
β[η
t
+ α(uη
x
+ vη
y
) + (1 + αη)(u
x
+ v
y
)] =
1
6
β
2
[(∇
2
u)
x
+ (∇
2
v)
y
]. (20)
Diﬀerentiating (19) ﬁrst with respect to x and then to respect to y gives a set of two
equations:
u
t
+ α(uu
x
+ vv
x
) + η
x
−
1
2
β(u
txx
+ v
txy
) = 0, (21)
v
t
+ α(uu
y
+ vv
y
) + η
y
−
1
2
β(u
txy
+ v
tyy
) = 0. (22)
The kinematic condition (20) can be rewritten as
η
t
+ [u(1 + αη)]
x
+ [v(1 + αη)]
y
=
1
6
β[(∇
2
u)
x
+ (∇
2
v)
y
]. (23)
Equations (21)–(23) contain in fact various shallowwater models. The socalled funda
mental shallowwater equations are obtained by neglecting the terms of order β:
u
t
+ α(uu
x
+ vu
y
) + η
x
= 0, (24)
v
t
+ α(uv
x
+ vv
y
) + η
y
= 0, (25)
η
t
+ [u(1 + αη)]
x
+ [v(1 + αη)]
y
= 0. (26)
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
INTRODUCTION xxiii
Recall that we assumed h to be constant for the derivation. Going back to an arbitrary
water depth and to dimensional variables, the system of nonlinear shallow water equations
reads
u
t
+ uu
x
+ vu
y
+ gη
x
= 0, (27)
v
t
+ uv
x
+ vv
y
+ gη
y
= 0, (28)
η
t
+ [u(h + η)]
x
+ [v(h + η)]
y
= 0. (29)
This system of equations has been used for example by Titov and Synolakis for the nu
merical computation of tidal wave runup [TS98]. Note that this model does not include
any bottom friction terms. To solve the problem of tsunami generation caused by bottom
displacement, the motion of the seaﬂoor obtained from seismological models [Oka85] can
be prescribed during a time t
0
. Usually t
0
is assumed to be small, so that the bottom
displacement is considered as an instantaneous vertical displacement. This assumption
may not be appropriate for slow events.
The satellite altimetry observations of the Indian Ocean tsunami clearly show dispersive
eﬀects. The question of dispersive eﬀects in tsunamis is open. Most propagation codes
ignore dispersion. A few propagation codes that include dispersion have been developed
[DGK06]. A wellknown code is FUNWAVE, developed at the University of Delaware over
the past ten years [KWC
+
98]. Dispersive shallow waterwave models are presented next.
Boussinesq equations
An additional dimensionless number, sometimes called the Stokes number (in other
references it is called Ursell number as well), is introduced:
S =
α
β
. (30)
For the Indian Ocean tsunami, one ﬁnds
0.24 < S < 46. (31)
Therefore the additional assumption that S ≈ 1 may be realistic.
In this subsection, we provide the guidelines to derive Boussinesqtype systems of equa
tions [BCS02]. Of course, the variation of bathymetry is essential for the propagation of
tsunamis, but for the derivation the water depth will be assumed to be constant. Some
notation is introduced. The potential evaluated along the free surface is denoted by
Φ(x, y, t) := φ(x, y, η, t). The derivatives of the velocity potential evaluated on the free
surface are denoted by Φ
(∗)
(x, y, t) := φ
∗
(x, y, η, t), where the star stands for x, y, z or t.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
xxiv INTRODUCTION
Consequently, Φ
∗
(deﬁned for ∗ ,= z) and Φ
(∗)
have diﬀerent meanings. They are however
related since
Φ
∗
= Φ
(∗)
+ Φ
(z)
η
∗
.
The vertical velocity at the free surface is denoted by W(x, y, t) := φ
z
(x, y, η, t).
The boundary conditions on the free surface (5) and (6) become
η
t
+∇Φ ∇η −W(1 +∇η ∇η) = 0, (32)
Φ
t
+ gη +
1
2
[∇Φ[
2
−
1
2
W
2
(1 +∇η ∇η) = 0. (33)
These two nonlinear equations provide timestepping for η and Φ. In addition, Laplace’s
equation as well as the kinematic condition on the bottom must be satisﬁed. In order
to relate the freesurface variables with the bottom variables, one must solve Laplace’s
equation in the whole water column. In Boussinesqtype models, the velocity potential is
represented as a formal expansion,
φ(x, y, z, t) =
∞
n=0
φ
(n)
(x, y, t) z
n
. (34)
Here the expansion is about z = 0, which is the location of the free surface at rest.
Demanding that φ formally satisfy Laplace’s equation leads to a recurrence relation between
φ
(n)
and φ
(n+2)
. Let φ
o
denote the velocity potential at z = 0, u
o
the horizontal velocity
at z = 0, and w
o
the vertical velocity at z = 0. Note that φ
o
and w
o
are nothing else than
φ
(0)
and φ
(1)
. The potential φ can be expressed in terms of φ
o
and w
o
only. Finally, one
obtains the velocity ﬁeld in the whole water column (−h ≤ z ≤ η) [MBS03]:
u(x, y, z, t) = cos(z∇)u
o
+ sin(z∇)w
o
, (35)
w(x, y, z, t) = cos(z∇)w
o
−sin(z∇)u
o
. (36)
Here the cosine and sine operators are inﬁnite Taylor series operators deﬁned by
cos(z∇) =
∞
n=0
(−1)
n
z
2n
(2n)!
∇
2n
, sin(z∇) =
∞
n=0
(−1)
n
z
2n+1
(2n + 1)!
∇
2n+1
.
Then one can substitute the representation (35)(36) into the kinematic bottom condi
tion and use successive approximations to obtain an explicit recursive expression for w
o
in
terms of u
o
to inﬁnite order in h∇.
A wide variety of Boussinesq systems can been derived [MBS03]. One can generalize
the expansions to an arbitrary z−level, instead of the z = 0 level. The Taylor series for the
cosine and sine operators can be truncated, Pad´e approximants can be used in operators
at z = −h and/or at z = 0.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
INTRODUCTION xxv
The classical Boussinesq equations are more transparent when written in the dimen
sionless variables used in the previous subsection. We further assume that h is constant,
drop the tildes, and write the equations for one spatial dimension (x). Performing the
expansion about z = 0 leads to the vanishing of the odd terms in the velocity potential.
Substituting the expression for φ into the freesurface boundary conditions evaluated at
z = 1 + αη(x, t) leads to two equations in η and φ
o
with terms of various order in α and
β. The small parameters α and β are of the same order, while η and φ
o
as well as their
partial derivatives are of order one.
Classical Boussinesq equations
The classical Boussinesq equations are obtained by keeping all terms that are at most
linear in α or β. In the derivation of the fundamental nonlinear shallowwater equations
(24)–(26), the terms in β were neglected. It is therefore implicitly assumed that the Stokes
number is large. Since the cube of the water depth appears in the denominator of the
Stokes number (S = α/β = aℓ
2
/d
3
), it means that the Stokes number is 64 times larger
in a 1 km depth than in a 4 km depth! Based on these arguments, dispersion is more
important to the west of the rupture. Considering the Stokes number to be of order one
leads to the following system in dimensional form
2
:
u
t
+ uu
x
+ gη
x
−
1
2
h
2
u
txx
= 0, (37)
η
t
+ [u(h + η)]
x
−
1
6
h
3
u
xxx
= 0. (38)
The classical Boussinesq equations are in fact slightly diﬀerent. They are obtained by
replacing u with the depth averaged velocity
1
h
_
η
−h
udz.
They read
u
t
+ uu
x
+ gη
x
−
1
3
h
2
u
txx
= 0, (39)
η
t
+ [u(h + η)]
x
= 0. (40)
A number of variants of the classical Boussinesq system were studied by Bona et al.
[BCS02], who in particular showed that depending on the modeling of dispersion the lin
earization about the rest state may or may not be wellposed.
2
Equations (37) and (38) could have been obtained from equations (21) and (23).
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
xxvi INTRODUCTION
Korteweg–de Vries equation
The previous system allows the propagation of waves in both the positive and negative
x−directions. Seeking solutions travelling in only one direction, for example the positive
x−direction, leads to a single equation for η, the Korteweg–de Vries equation:
η
t
+ c
0
_
1 +
3η
2d
_
η
x
+
1
6
c
0
d
2
η
xxx
= 0, (41)
where d is the water depth. It admits solitary wave solutions travelling at speed V in the
form
η(x, t) = a sech
2
_
_
3a
4d
3
(x −V t)
_
, with V = c
0
_
1 +
a
2d
_
.
The solitary wave solutions of the Korteweg–de Vries equation are of elevation (a > 0) and
travel faster than c
0
. Their speed increases with amplitude. Note that a natural length
scale appears:
ℓ =
_
4d
3
3a
.
For the Indian Ocean tsunami, it gives roughly ℓ = 377 km. It is of the order of magnitude
of the wavelength estimated from the width of the segments that experienced slip.
Solitary waves exist for more general nonlinear dispersive equations such as KdVKdV,
and BonaSmith systems. Their existence was recently studied in [FDG05, DM04]. Nu
merical methods (based on Galerkin discretization) for the approximation of solutions to
KdVKdV systems are discussed in [BDM07].
Energy of a tsunami
The energy of the earthquake is measured via the strain energy released by the faulting.
The part of the energy transmitted to the tsunami wave is less than one percent Lay et al.
[LKA
+
05]. They estimate the tsunami energy to be 4.2 10
15
J. They do not give details
on how they obtained this estimate. However, a simple calculation based on considering
the tsunami as a soliton
η(x) = a sech
2
_
x
ℓ
_
, u(x) = αc
0
sech
2
_
x
ℓ
_
,
gives for the energy
E =
1
√
3
α
3/2
ρd
2
(c
2
0
+ gd)
_
∞
−∞
sech
4
xdx + O(α
2
).
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
INTRODUCTION xxvii
The value for the integral is 4/3. The numerical estimate for E is close to that of Lay
et al. (2005) [LKA
+
05]. Incidently, at this level of approximation, there is equipartition
between kinetic and potential energy. It is also important to point out that a tsunami
being a shallow water wave, the whole water column is moving as the wave propagates.
For the parameter values used so far, the maximum horizontal current is 3 cm/s. However,
as the water depth decreases, the current increases and becomes important when the depth
becomes less than 500 m. Additional properties of solitary waves can be found for example
in [LH74].
Tsunami runup
The last phase of a tsunami is its runup and inundation. Although in some cases it may
be important to consider the coupling between ﬂuid and structures, we restrict ourselves
to the description of the ﬂuid ﬂow. The problem of waves climbing a beach is a classical
one [CG58]. The transformations used by Carrier and Greenspan are still used nowadays.
The basis of their analysis is the onedimensional counterpart of the system (27)–(29). In
addition, they assume the depth to be of uniform slope: h = −xtan θ. Introduce the
following dimensionless quantities, where ℓ is a characteristic length
3
:
x = ℓ˜ x, η = ℓ˜ η, u =
_
gℓ ˜ u, t =
_
ℓ/g
˜
t, c
2
= (h + η)/ℓ.
After dropping the tildes, the dimensionless system of equations (27)(29) becomes
u
t
+ uu
x
+ η
x
= 0,
η
t
+ [u(−xtan θ + η)]
x
= 0.
In terms of the variable c, these equations become
u
t
+ uu
x
+ 2cc
x
+ tan θ = 0,
2c
t
+ cu
x
+ 2uc
x
= 0.
The equations written in characteristic form are
_
∂
∂t
+ (u + c)
∂
∂x
_
(u + 2c + t tan θ) = 0,
_
∂
∂t
+ (u −c)
∂
∂x
_
(u −2c + t tan θ) = 0.
3
In fact there is no obvious characteristic length in this idealized problem. Some authors simply say at
this point that ℓ is speciﬁc to the problem under consideration.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
xxviii INTRODUCTION
The characteristic curves C
+
and C
−
as well as the Riemann invariants are
C
+
:
dx
dt
= u + c, u + 2c + t tan θ = r,
C
−
:
dx
dt
= u −c, u −2c + t tan θ = s.
Next one can rewrite the hyperbolic equations in terms of the new variables λ and σ deﬁned
as follows:
λ
2
=
1
2
(r + s) = u + t tan θ,
σ
4
=
1
4
(r −s) = c.
One obtains
x
s
−
_
1
4
(3r + s) −t tan θ
_
t
s
= 0,
x
r
−
_
1
4
(r + 3s) −t tan θ
_
t
r
= 0.
The elimination of x results in the linear secondorder equation for t
σ(t
λλ
−t
σσ
) −3t
σ
= 0. (42)
Since u + t tan θ = λ/2, u must also satisfy (42). Introducing the potential φ(σ, λ) such
that
u =
φ
σ
σ
,
one obtains the equation
(σφ
σ
)
σ
−σφ
λλ
= 0
after integrating once. Two major simpliﬁcations have been obtained. The nonlinear set of
equations have been reduced to a linear equation for u or φ and the free boundary is now
the ﬁxed line σ = 0 in the (σ, λ)plane. The free boundary is the instantaneous shoreline
c = 0, which moves as a wave climbs a beach.
The above formulation has been used by several authors to study the runup of vari
ous types of waves on sloping beaches [Syn87, TS94, CWY03, TT05]. Synolakis [Syn87]
validated analytical results by comparing them with experiments on the runup of a soli
tary wave on a plane beach. He used CarrierGreenspan transformation and linear theory
solutions to draw several important conclusions:
• A law predicting the maximum runup of nonbreaking waves was given;
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
INTRODUCTION xxix
• The author found diﬀerent runup variations for breaking and nonbreaking solitary
waves;
• Nonlinear theory models adequately describe surface proﬁles of nonbreaking waves;
• Diﬀerent criteria were considered whether a solitary wave with given height/depth
ratio will break or not as it climbs up a sloping beach.
Later, it was shown that leading depression Nwaves runup higher than leading eleva
tion Nwaves, suggesting that perhaps the solitary wave model may not be adequate for
predicting an upper limit for the runup of nearshore generated tsunamis. In [KS06] it
was shown that there is a diﬀerence in the maximum runup by a factor 2 in shoreline
motions with and without initial velocity. This diﬀerence suggests strong implications for
the runup predictions and tsunami warning system [GBM
+
05] if the appropriate initial
velocity is not speciﬁed.
There is a rule of thumb that says that the runup does not usually exceed twice the
fault slip. Since runups of 30 meters were observed in Sumatra during the Boxing Day
tsunami, the slip might have been of 15 meters or even more.
Analytical models are useful, especially to perform parametric studies. However, the
breaking of tsunami waves as well as the subsequent ﬂoodings must be studied numerically.
The most natural methods that can be used are the free surface capturing methods based
on a ﬁnite volume discretisation, such as the Volume Of Fluid (VOF) or the Level Set
methods, and the family of Smoothed Particle Hydrodynamics methods (SPH), applied to
freesurface ﬂow problems [Mon94, GGD04, GGCCD05]. Such methods allow a study of
ﬂood wave dynamics, of wave breaking on the land behind beaches, and of the ﬂow over
rising ground with and without the presence of obstacles. This task is an essential part
of tsunami modelling, since it allows the determination of the level of risk due to major
ﬂooding, the prediction of the resulting water levels in the ﬂooded areas, the determination
of security zones. It also provides some help in the conception and validation of protection
systems in the most exposed areas.
The present manuscript is organized as follows. In Chapter 1 we extensively investi
gate the tsunami generation process (see Sections 1.1 and 1.3) and the ﬁrst minutes of
propagation (Section 1.2) using diﬀerent mathematical models. Several important conclu
sions about the physics of tsunami waves were drew out. Chapter 2 discusses diﬀerent
possibilities of introducing dissipative eﬀects into Boussinesq equations. Pros and cons of
two models under consideration are revealed. In Chapter 3 our attention is attracted by
twophase ﬂows and wave impact problem. An eﬃcient numerical method based on a ﬁnite
volume scheme is described and a few numerical results are presented. In the last Chapter
4 we return to the investigation of dissipative eﬀects onto water waves. Novel bottom
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
xxx INTRODUCTION
boundary layer correction is derived and the new set of governing equations is thoroughly
analysed.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
Chapter 1
Tsunami generation modelling
From a drop of water a logician could predict an Atlantic or a Niagara.
Sir Arthur Conan Doyle
Contents
1.1 Waves generated by a moving bottom . . . . . . . . . . . . . . 2
1.1.1 Source model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.2 Volterra’s theory of dislocations . . . . . . . . . . . . . . . . . . . 5
1.1.3 Dislocations in elastic halfspace . . . . . . . . . . . . . . . . . . 6
1.1.4 Solution in ﬂuid domain . . . . . . . . . . . . . . . . . . . . . . . 16
1.1.5 Freesurface elevation . . . . . . . . . . . . . . . . . . . . . . . . 21
1.1.6 Velocity ﬁeld . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.1.7 Asymptotic analysis of integral solutions . . . . . . . . . . . . . . 29
1.1.8 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.2 Comparison of tsunami generation models . . . . . . . . . . . 36
1.2.1 Physical problem description . . . . . . . . . . . . . . . . . . . . 37
1.2.2 Linear theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
1.2.3 Active generation . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
1.2.4 Passive generation . . . . . . . . . . . . . . . . . . . . . . . . . . 44
1.2.5 Nonlinear shallow water equations . . . . . . . . . . . . . . . . . 47
1.2.6 Mathematical model . . . . . . . . . . . . . . . . . . . . . . . . . 48
1.2.7 Numerical method . . . . . . . . . . . . . . . . . . . . . . . . . . 50
1.2.8 Numerical method for the full equations . . . . . . . . . . . . . . 50
1
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
2 Tsunami generation
1.2.9 Comparisons and discussion . . . . . . . . . . . . . . . . . . . . . 51
1.2.10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
1.3 Tsunami generation by dynamic displacement of sea bed . . 64
1.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
1.3.2 Mathematical models . . . . . . . . . . . . . . . . . . . . . . . . 67
1.3.3 Numerical methods . . . . . . . . . . . . . . . . . . . . . . . . . . 70
1.3.4 Validation of the numerical method . . . . . . . . . . . . . . . . 72
1.3.5 Results of the simulation . . . . . . . . . . . . . . . . . . . . . . 73
1.3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
1.1 Waves generated by a moving bottom
Waves at the surface of a liquid can be generated by various mechanisms: wind blowing
on the free surface, wavemaker, moving disturbance on the bottom or the surface, or even
inside the liquid, fall of an object into the liquid, liquid inside a moving container, etc. In
this chapter, we concentrate on the case where the waves are created by a given motion of
the bottom. One example is the generation of tsunamis by a sudden seaﬂoor deformation.
There are diﬀerent natural phenomena that can lead to a tsunami. For example, one
can mention submarine slumps, slides, volcanic explosions, etc. In this chapter we use a
submarine faulting generation mechanism as tsunami source. The resulting waves have
some wellknown features. For example, characteristic wavelengths are large and wave
amplitudes are small compared with water depth.
Two factors are usually necessary for an accurate modelling of tsunamis: information
on the magnitude and distribution of the displacements caused by the earthquake, and
a model of surface gravity waves generation resulting from this motion of the seaﬂoor.
Most studies of tsunami generation assume that the initial freesurface deformation is
equal to the vertical displacement of the ocean bottom. The details of wave motion are
neglected during the time that the source operates. While this is often justiﬁed because the
earthquake rupture occurs very rapidly, there are some speciﬁc cases where the time scale
of the bottom deformation may become an important factor. This was emphasized for
example by Trifunac and Todorovska [TT01], who considered the generation of tsunamis
by a slowly spreading uplift of the seaﬂoor and were able to explain some observations.
During the 26 December 2004 SumatraAndaman event, there was in the northern extent
of the source a relatively slow faulting motion that led to signiﬁcant vertical bottom motion
but left little record in the seismic data. It is interesting to point out that it is the inversion
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.1 Waves generated by a moving bottom 3
of tidegauge data from Paradip, the northernmost of the Indian eastcoast stations, that
led Neetu et al. [NSS
+
05] to conclude that the source length was greater by roughly 30%
than the initial estimate of Lay et al. [LKA
+
05]. Incidentally, the generation time is also
longer for landslide tsunamis.
Our study is restricted to the water region where the incompressible Euler equations
for potential ﬂow can be linearized. The wave propagation away from the source can be
investigated by shallow water models which may or may not take into account nonlinear
eﬀects and frequency dispersion. Such models include the Kortewegde Vries equation
[KdV95] for unidirectional propagation, nonlinear shallowwater equations and Boussinesq
type models [Bou71b, Per66, BBM72].
Several authors have modeled the incompressible ﬂuid layer as a special case of an
elastic medium [Pod68, Kaj63, Gus72, AG73, Gus76]. In our opinion it may be convenient
to model the liquid by an elastic material from a mathematical point of view, but it is
questionable from a physical point of view. The crust was modeled as an elastic isotropic
halfspace. This assumption will also be adopted in the present study.
The problem of tsunami generation has been considered by a number of authors: see for
example [Car71, vdD72, BvdDP73]. The models discussed in these papers lack ﬂexibility
in terms of modelling the source due to the earthquake. The present study provides some
extensions. A good review on the subject is [Sab86].
Here we essentially follow the framework proposed by Hammack [Ham73] and others.
The tsunami generation problem is reduced to a CauchyPoisson boundary value problem
in a region of constant depth. The main extensions given in the present work consist in
threedimensional modelling and more realistic source models. This approach was followed
recently in [TT01, THT02], where the mathematical model was the same as in [Ham73]
but the source was diﬀerent.
Most analytical studies of linearized wave motion use integral transform methods. The
complexity of the integral solutions forced many authors [Kaj63, Kel63] to use asymptotic
methods such as the method of stationary phase to estimate the farﬁeld behaviour of the
solutions. In the present study we have also obtained asymptotic formulas for integral
solutions. They are useful from a qualitative point of view, but in practice it is better
to use numerical integration formulas [Fil28] that take into account the oscillatory nature
of the integrals. All the numerical results presented in this section were obtained in this
manner.
One should use asymptotic solutions with caution since they approximate exact solu
tions of the linearized problem. The relative importance of linear and nonlinear eﬀects can
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
4 Tsunami generation
be measured by the Stokes (or Ursell) number [Urs53]:
U :=
a/h
(kh)
2
=
a
k
2
h
3
,
where k is a wave number, a a typical wave amplitude and h the water depth. For U ≫1,
the nonlinear eﬀects control wave propagation and only nonlinear models are applicable.
Ursell [Urs53] proved that near the wave front U behaves like
U ∼ t
1
3
.
Hence, regardless of how small nonlinear eﬀects are initially, they will become important.
Recently, the methodology used in this thesis for tsunami generation was applied in the
framework of the Boussinesq equations [Mit07] over uneven bottom. These equations have
the advantage of being twodimensional while the physical problem and, consequently, the
potential ﬂow formulation are 3D. On the other hand, Boussinesq equations represent a
long wave approximation to the complete waterwave problem.
1.1.1 Source model
The inversion of seismic wave data allows the reconstruction of permanent deformations
of the sea bottom following earthquakes. In spite of the complexity of the seismic source
and of the internal structure of the earth, scientists have been relatively successful in using
simple models for the source. One of these models is Okada’s model [Oka85]. Its description
follows.
The fracture zones, along which the foci of earthquakes are to be found, have been
described in various papers. For example, it has been suggested that Volterra’s theory of
dislocations might be the proper tool for a quantitative description of these fracture zones
[Ste58]. This suggestion was made for the following reason. If the mechanism involved
in earthquakes and the fracture zones is indeed one of fracture, discontinuities in the
displacement components across the fractured surface will exist. As dislocation theory
may be described as that part of the theory of elasticity dealing with surfaces across which
the displacement ﬁeld is discontinuous, the suggestion makes sense.
As is often done in mathematical physics, it is necessary for simplicity’s sake to make
some assumptions. Here we neglect the curvature of the earth, its gravity, temperature,
magnetism, nonhomogeneity, and consider a semiinﬁnite medium, which is homogeneous
and isotropic. We further assume that the laws of classical linear elasticity theory hold.
Several studies showed that the eﬀect of earth curvature is negligible for shallow events
at distances of less than 20
◦
[BMSS69, BMSS70, SM71]. The sensitivity to earth topogra
phy, homogeneity, isotropy and halfspace assumptions was studied and discussed recently
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.1 Waves generated by a moving bottom 5
[Mas03]. A commercially available code, ABACUS, which is based on a ﬁnite element
model (FEM), was used. Six FEMs were constructed to test the sensitivity of deformation
predictions to each assumption. The author came to the conclusion that the vertical layer
ing of lateral inhomogeneity can sometimes cause considerable eﬀects on the deformation
ﬁelds.
The usual boundary conditions for dealing with earth problems require that the surface
of the elastic medium (the earth) shall be free from forces. The resulting mixed boundary
value problem was solved a century ago [Vol07]. Later, Steketee proposed an alternative
method to solve this problem using Green’s functions [Ste58].
1.1.2 Volterra’s theory of dislocations
In order to introduce the concept of dislocation and for simplicity’s sake, this section is
devoted to the case of an entire elastic space, as was done in the original paper by Volterra
[Vol07].
Let O be the origin of a Cartesian coordinate system in an inﬁnite elastic medium, x
i
the Cartesian coordinates (i = 1, 2, 3), and e
i
a unit vector in the positive x
i
−direction. A
force F = Fe
k
at O generates a displacement ﬁeld u
k
i
(P, O) at point P, which is determined
by the wellknown Somigliana tensor
u
k
i
(P, O) =
F
8πµ
(δ
ik
r
, nn
−αr
, ik
), with α =
λ + µ
λ + 2µ
. (1.1)
In this relation δ
ik
is the Kronecker delta, λ and µ are Lam´e’s constants, and r is the
distance from P to O. The coeﬃcient α can be rewritten as α = 1/2(1 − ν), where ν is
Poisson’s ratio. Later we will also use Young’s modulus E, which is deﬁned as
E =
µ(3λ + 2µ)
λ + µ
.
The notation r
, i
means ∂r/∂x
i
and the summation convention applies.
The stresses due to the displacement ﬁeld (1.1) are easily computed from Hooke’s law:
σ
ij
= λδ
ij
u
k,k
+ µ(u
i,j
+ u
j,i
). (1.2)
One ﬁnds
σ
k
ij
(P, O) = −
αF
4π
_
3x
i
x
j
x
k
r
5
+
µ
λ + µ
δ
ki
x
j
+ δ
kj
x
i
−δ
ij
x
k
r
3
_
.
The components of the force per unit area on a surface element are denoted as follows:
T
k
i
= σ
k
ij
ν
j
,
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
6 Tsunami generation
where the ν
j
’s are the components of the normal to the surface element. A Volterra dislo
cation is deﬁned as a surface Σ in the elastic medium across which there is a discontinuity
∆u
i
in the displacement ﬁelds of the type
∆u
i
= u
+
i
−u
−
i
= U
i
+ Ω
ij
x
j
, (1.3)
Ω
ij
= −Ω
ji
. (1.4)
Equation (1.3) in which U
i
and Ω
ij
are constants is the wellknown Weingarten relation
which states that the discontinuity ∆u
i
should be of the type of a rigid body displacement,
thereby maintaining continuity of the components of stress and strain across Σ.
The displacement ﬁeld in an inﬁnite elastic medium due to the dislocation is then
determined by Volterra’s formula [Vol07]
u
k
(Q) =
1
F
__
Σ
∆u
i
T
k
i
dS. (1.5)
Once the surface Σ is given, the dislocation is essentially determined by the six constants
U
i
and Ω
ij
. Therefore we also write
u
k
(Q) =
U
i
F
__
Σ
σ
k
ij
(P, Q)ν
j
dS +
Ω
ij
F
__
Σ
¦x
j
σ
k
il
(P, Q) −x
i
σ
k
jl
(P, Q)¦ν
l
dS, (1.6)
where Ω
ij
takes only the values Ω
12
, Ω
23
, Ω
31
. Following Volterra [Vol07] and Love [Lov44]
we call each of the six integrals in (1.6) an elementary dislocation.
It is clear from (1.5) and (1.6) that the computation of the displacement ﬁeld u
k
(Q)
is performed as follows. A force Fe
k
is applied at Q, and the stresses σ
k
ij
(P, Q) that this
force generates are computed at the points P(x
i
) on Σ. In particular the components of
the force on Σ are computed. After multiplication with prescribed weights of magnitude
∆u
i
these forces are integrated over Σ to give the displacement component in Q due to the
dislocation on Σ.
1.1.3 Dislocations in elastic halfspace
When the case of an elastic halfspace is considered, equation (1.5) remains valid, but
we have to replace σ
k
ij
in T
k
i
by another tensor ω
k
ij
. This can be explained by the fact that
the elementary solutions for a halfspace are diﬀerent from Somigliana solution (1.1).
The ω
k
ij
can be obtained from the displacements corresponding to nuclei of strain in a
halfspace through relation (1.2). Steketee showed a method of obtaining the six ω
k
ij
ﬁelds
by using a Green’s function and derived ω
k
12
, which is relevant to a vertical strikeslip fault
(see below). Maruyama derived the remaining ﬁve functions [Mar64].
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.1 Waves generated by a moving bottom 7
It is interesting to mention here that historically these solutions were ﬁrst derived in
a straightforward manner by Mindlin [Min36, MC50], who gave explicit expressions of the
displacement and stress ﬁelds for halfspace nuclei of strain consisting of single forces with
and without moment. It is only necessary to write the single force results since the other
forms can be obtained by taking appropriate derivatives. The method consists in ﬁnding
the displacement ﬁeld in Westergaard’s form of the Galerkin vector [Wes35]. This vector is
then determined by taking a linear combination of some biharmonic elementary solutions.
The coeﬃcients are chosen to satisfy boundary and equilibrium conditions. These solutions
were also derived by Press in a slightly diﬀerent manner [Pre65].
x
y
z
W
L
D
δ
U
1
U
2
U
3
Figure 1.1: Coordinate system adopted in this study and geometry of the source model.
Here, we take the Cartesian coordinate system shown in Figure 1.1. The elastic medium
occupies the region x
3
≤ 0 and the x
1
−axis is taken to be parallel to the strike direction
of the fault. In this coordinate system, u
j
i
(x
1
, x
2
, x
3
; ξ
1
, ξ
2
, ξ
3
) is the ith component of
the displacement at (x
1
, x
2
, x
3
) due to the jth direction point force of magnitude F at
(ξ
1
, ξ
2
, ξ
3
). It can be expressed as follows [Oka85, Min36, Pre65, Oka92]:
u
j
i
(x
1
, x
2
, x
3
) = u
j
iA
(x
1
, x
2
, −x
3
) −u
j
iA
(x
1
, x
2
, x
3
) (1.7)
+u
j
iB
(x
1
, x
2
, x
3
) + x
3
u
j
iC
(x
1
, x
2
, x
3
),
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
8 Tsunami generation
x
y
φ
Figure 1.2: Illustration for strike angle deﬁnition.
where
u
j
iA
=
F
8πµ
_
(2 −α)
δ
ij
R
+ α
R
i
R
j
R
3
_
,
u
j
iB
=
F
4πµ
_
δ
ij
R
+
R
i
R
j
R
3
+
1 −α
α
_
δ
ij
R + R
3
+
+
R
i
δ
j3
−R
j
δ
i3
(1 −δ
j3
)
R(R + R
3
)
−
R
i
R
j
R(R + R
3
)
2
(1 −δ
i3
)(1 −δ
j3
)
_
_
,
u
j
iC
=
F
4πµ
(1 −2δ
i3
)
_
(2 −α)
R
i
δ
j3
−R
j
δ
i3
R
3
+ αξ
3
_
δ
ij
R
3
−3
R
i
R
j
R
5
__
.
In these expressions R
1
= x
1
−ξ
1
, R
2
= x
2
−ξ
2
, R
3
= −x
3
−ξ
3
and R
2
= R
2
1
+ R
2
2
+ R
2
3
.
The ﬁrst term in equation (1.7), u
j
iA
(x
1
, x
2
, −x
3
), is the wellknown Somigliana tensor,
which represents the displacement ﬁeld due to a single force placed at (ξ
1
, ξ
2
, ξ
3
) in an
inﬁnite medium [Lov44]. The second term also looks like a Somigliana tensor. This term
corresponds to a contribution from an image source of the given point force placed at
(ξ
1
, ξ
2
, −ξ
3
) in the inﬁnite medium. The third term, u
j
iB
(x
1
, x
2
, x
3
), and u
j
iC
(x
1
, x
2
, x
3
) in
the fourth term are naturally depth dependent. When x
3
is set equal to zero in equation
(1.7), the ﬁrst and the second terms cancel each other, and the fourth term vanishes. The
remaining term, u
j
iB
(x
1
, x
2
, 0), reduces to the formula for the surface displacement ﬁeld
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.1 Waves generated by a moving bottom 9
due to a point force in a halfspace [Oka85]:
_
¸
¸
¸
_
¸
¸
¸
_
u
1
1
=
F
4πµ
_
1
R
+
(x
1
−ξ
1
)
2
R
3
+
µ
λ+µ
_
1
R−ξ
3
−
(x
1
−ξ
1
)
2
R(R−ξ
3
)
2
__
,
u
1
2
=
F
4πµ
(x
1
−ξ
1
)(x
2
−ξ
2
)
_
1
R
3
−
µ
λ+µ
1
R(R−ξ
3
)
2
_
,
u
1
3
=
F
4πµ
(x
1
−ξ
1
)
_
−
ξ
3
R
3
−
µ
λ+µ
1
R(R−ξ
3
)
_
,
_
¸
¸
¸
_
¸
¸
¸
_
u
2
1
=
F
4πµ
(x
1
−ξ
1
)(x
2
−ξ
2
)
_
1
R
3
−
µ
λ+µ
1
R(R−ξ
3
)
2
_
,
u
2
2
=
F
4πµ
_
1
R
+
(x
2
−ξ
2
)
2
R
3
+
µ
λ+µ
_
1
R−ξ
3
−
(x
2
−ξ
2
)
2
R(R−ξ
3
)
2
__
,
u
2
3
=
F
4πµ
(x
2
−ξ
2
)
_
−
ξ
3
R
3
−
µ
λ+µ
1
R(R−ξ
3
)
_
,
_
¸
¸
¸
_
¸
¸
¸
_
u
3
1
=
F
4πµ
(x
1
−ξ
1
)
_
−
ξ
3
R
3
+
µ
λ+µ
1
R(R−ξ
3
)
_
,
u
3
2
=
F
4πµ
(x
2
−ξ
2
)
_
−
ξ
3
R
3
+
µ
λ+µ
1
R(R−ξ
3
)
_
,
u
3
3
=
F
4πµ
_
1
R
+
ξ
2
3
R
3
+
µ
λ+µ
1
R
_
.
In these formulas R
2
= (x
1
−ξ
1
)
2
+ (x
2
−ξ
2
)
2
+ ξ
2
3
.
In order to obtain the displacements due to the dislocation we need to calculate the
corresponding ξ
k
derivatives of the point force solution (1.7) and to insert them in Volterra’s
formula (1.5)
u
i
=
1
F
__
Σ
∆u
j
_
λδ
jk
∂u
n
i
∂ξ
n
+ µ
_
∂u
j
i
∂ξ
k
+
∂u
k
i
∂ξ
j
__
ν
k
dS.
The ξ
k
derivatives are expressed as follows:
∂u
j
i
∂ξ
k
(x
1
, x
2
, x
3
) =
∂u
j
iA
∂ξ
k
(x
1
, x
2
, −x
3
) −
∂u
j
iA
∂ξ
k
(x
1
, x
2
, x
3
) +
+
∂u
j
iB
∂ξ
k
(x
1
, x
2
, x
3
) + x
3
∂u
j
iC
∂ξ
k
(x
1
, x
2
, x
3
),
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
10 Tsunami generation
with
∂u
j
iA
∂ξ
k
=
F
8πµ
_
(2 −α)
R
k
R
3
δ
ij
−α
R
i
δ
jk
+ R
j
δ
ik
R
3
+ 3α
R
i
R
j
R
k
R
5
_
,
∂u
j
iB
∂ξ
k
=
F
4πµ
_
−
R
i
δ
jk
+ R
j
δ
ik
−R
k
δ
ij
R
3
+ 3
R
i
R
j
R
k
R
5
+
+
1 −α
α
_
δ
3k
R + R
k
R(R + R
3
)
2
δ
ij
−
δ
ik
δ
j3
−δ
jk
δ
i3
(1 −δ
j3
)
R(R + R
3
)
+
+
_
R
i
δ
j3
−R
j
δ
i3
(1 −δ
j3
)
_
δ
3k
R
2
+ R
k
(2R + R
3
)
R
3
(R + R
3
)
2
+
+(1 −δ
i3
)(1 −δ
j3
)
_
R
i
δ
jk
+ R
j
δ
ik
R(R + R
3
)
2
−R
i
R
j
2δ
3k
R
2
+ R
k
(3R + R
3
)
R
3
(R + R
3
)
3
_
_
_
,
∂u
j
iC
∂ξ
k
=
F
4πµ
(1 −2δ
i3
)
_
(2 −α)
_
δ
jk
δ
i3
−δ
ik
δ
j3
R
3
+
3R
k
(R
i
δ
j3
−R
j
δ
i3
)
R
5
_
+
+αδ
3k
_
δ
ij
R
3
−
3R
i
R
j
R
5
_
+ 3αξ
3
_
R
i
δ
jk
+ R
j
δ
ik
+ R
k
δ
ij
R
5
−
5R
i
R
j
R
k
R
7
_
_
.
1.1.3.1 Finite rectangular source
Let us now consider a more practical problem. We deﬁne the elementary dislocations
U
1
, U
2
and U
3
, corresponding to the strikeslip, dipslip and tensile components of an
arbitrary dislocation. In Figure 1.1 each vector represents the direction of the elementary
faults. The vector D is the socalled Burger’s vector, which shows how both sides of the
fault are spread out: D = u
+
−u
−
.
A general dislocation can be determined by three angles: the dip angle δ of the fault
(0 ≤ δ ≤ π), the slip or rake angle θ (0 ≤ θ ≤ π), and the angle φ between the fault plane
and Burger’s vector D. When dealing with a geophysical application, an additional angle,
the azimuth or strike (see Figure 1.2 for strike angle deﬁnition), is introduced in order
to provide an orientation of the fault. The general situation is schematically described in
Figure 1.3.
For a ﬁnite rectangular fault with length L and width W occurring at depth d (Fig
ure 1.3), the deformation ﬁeld can be evaluated analytically by a change of variables and
by integrating over the rectangle. This was done by several authors [Oka85, Oka92, Chi63,
SM74, IS79]. Here we give the results of their computations. The ﬁnal results are repre
sented below in compact form, using Chinnery’s notation  to represent the substitution
f(ξ, η) = f(x, p) −f(x, p −W) −f(x −L, p) + f(x −L, p −W),
where p = y cos δ + d sin δ. Next we introduce the notation
q = y sin δ −d cos δ, ˜ y = η cos δ + q sin δ,
˜
d = η sin δ −q cos δ
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.1 Waves generated by a moving bottom 11
x
y
z
O
D W
δ
φ
θ
L
x
′
Figure 1.3: Geometry of the source model (dip angle δ, depth d
f
, length L, width W) and
orientation of Burger’s vector D (rake angle θ, angle φ between the fault plane and Burger’s
vector).
and
R
2
= ξ
2
+ η
2
+ q
2
= ξ
2
+ ˜ y
2
+
˜
d
2
, X
2
= ξ
2
+ q
2
.
The quantities U
1
, U
2
and U
3
are linked to Burger’s vector through the identities
U
1
= [D[ cos φcos θ, U
2
= [D[ cos φsin θ, U
3
= [D[ sin φ.
For a strikeslip dislocation, one has
u
1
= −
U
1
2π
_
ξq
R(R + η)
+ arctan
ξη
qR
+ I
1
sin δ
__
_
_
_
,
u
2
= −
U
1
2π
_
˜ yq
R(R + η)
+
q cos δ
R + η
+ I
2
sin δ
__
_
_
_
,
u
3
= −
U
1
2π
_
˜
dq
R(R + η)
+
q sin δ
R + η
+ I
4
sin δ
__
_
_
_
_
.
For a dipslip dislocation, one has
u
1
= −
U
2
2π
_
q
R
−I
3
sin δ cos δ
__
_
_ ,
u
2
= −
U
2
2π
_
˜ yq
R(R + ξ)
+ cos δ arctan
ξη
qR
−I
1
sin δ cos δ
__
_
_
_
,
u
3
= −
U
2
2π
_
˜
dq
R(R + ξ)
+ sin δ arctan
ξη
qR
−I
5
sin δ cos δ
__
_
_
_
_
.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
12 Tsunami generation
For a tensile fault dislocation, one has
u
1
=
U
3
2π
_
q
2
R(R + η)
−I
3
sin
2
δ
__
_
_
_
,
u
2
=
U
3
2π
_
−
˜
dq
R(R + ξ)
−sin δ
_
ξq
R(R + η)
−arctan
ξη
qR
_
−I
1
sin
2
δ
__
_
_
_
_
,
u
3
=
U
3
2π
_
˜ yq
R(R + ξ)
+ cos δ
_
ξq
R(R + η)
−arctan
ξη
qR
_
−I
5
sin
2
δ
__
_
_
_
.
The terms I
1
, . . . , I
5
are given by
I
1
= −
µ
λ + µ
ξ
(R +
˜
d) cos δ
−tan δI
5
,
I
2
= −
µ
λ + µ
log(R + η) −I
3
,
I
3
=
µ
λ + µ
_
1
cos δ
˜ y
R +
˜
d
−log(R + η)
_
+ tan δI
4
,
I
4
=
µ
µ + λ
1
cos δ
_
log(R +
˜
d) −sin δ log(R + η)
_
,
I
5
=
µ
λ + µ
2
cos δ
arctan
η(X + q cos δ) + X(R + X) sin δ
ξ(R + X) cos δ
,
and if cos δ = 0,
I
1
= −
µ
2(λ + µ)
ξq
(R +
˜
d)
2
,
I
3
=
µ
2(λ + µ)
_
η
R +
˜
d
+
˜ yq
(R +
˜
d)
2
−log(R + η)
_
,
I
4
= −
µ
λ + µ
q
R +
˜
d
,
I
5
= −
µ
λ + µ
ξ sin δ
R +
˜
d
.
Figures 1.4, 1.5, and 1.6 show the freesurface deformation due to the three elementary
dislocations. The values of the parameters are given in Table 1.1.
1.1.3.2 Curvilinear fault
In the previous subsection analytical formulas for the freesurface deformation in the
special case of a rectangular fault were given. In fact, Volterra’s formula (1.5) allows to
evaluate the displacement ﬁeld that accompanies fault events with much more general
geometry. The shape of the fault and Burger’s vector are suggested by seismologists and
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.1 Waves generated by a moving bottom 13
parameter value
Dip angle δ 13
◦
Fault depth d, km 25
Fault length L, km 220
Fault width W, km 90
U
i
, m 15
Young modulus E, GPa 9.5
Poisson’s ratio ν 0.23
Table 1.1: Parameter set used in Figures 1.4, 1.5, and 1.6.
Figure 1.4: Dimensionless freesurface deformation z/a due to dipslip faulting: φ = 0, θ = π/2,
D = (0, U
2
, 0). Here a is [D[ (15 m in the present application). The horizontal distances x and y
are expressed in kilometers.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
14 Tsunami generation
Figure 1.5: Dimensionless freesurface deformation z/a due to strikeslip faulting: φ = 0, θ = 0,
D = (U
1
, 0, 0). Here a is [D[ (15 m in the present application). The horizontal distances x and y
are expressed in kilometers.
Figure 1.6: Dimensionless freesurface deformation z/a due to tensile faulting: φ = π/2, D =
(0, 0, U
3
). Here a is [D[. The horizontal distances x and y are expressed in kilometers.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.1 Waves generated by a moving bottom 15
after numerical integration one can obtain the deformation of the seaﬂoor for more general
types of events as well.
Here we will consider the case of a fault whose geometry is described by an elliptical
arc (see Figure 1.7).
x
y
z
O
D n
d
b
a
Figure 1.7: Geometry of a fault with elliptical shape
The parametric equations of this surface are given by
x(ξ, η) = ξ, 0 ≤ ξ ≤ a, y(ξ, η) = η, −
c
2
≤ η ≤
c
2
,
z(ξ, η) = −(b + d) +
b
a
_
a
2
−ξ
2
.
Then the unit normal to this surface can be easily calculated:
n =
_
bξ
_
a
4
+ (b
2
−a
2
)ξ
2
, 0,
a
_
a
2
−ξ
2
_
a
4
+ (b
2
−a
2
)ξ
2
_
.
We also need to compute the coeﬃcients of the ﬁrst fundamental form in order to reduce
the surface integral in (1.5) to a double Riemann integral. These coeﬃcients are
E =
a
4
+ ξ
2
(b
2
−a
2
)
a
2
(a
2
−ξ
2
)
, F = 0, G = 1
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
16 Tsunami generation
parameter value
Depth event d, km 20
Ellipse semiminor axis a, km 17
Ellipse semimajor axis b, km 6
Fault width c, km 15
Young modulus E, GPa 9.5
Poisson’s ratio ν 0.23
Table 1.2: Parameter set used in Figure 1.8.
and the surface element dS is
dS =
√
EG−F
2
dξdη =
1
a
_
a
4
+ ξ
2
(b
2
−a
2
)
_
a
2
−ξ
2
dξdη.
Since in the crust the hydrostatic pressure is very large, it is natural to impose the
condition that D n = 0. The physical meaning of this condition is that both sides of
the fault slide and do not detach. This condition is obviously satisﬁed if we take Burger’s
vector as
D = D
_
a
_
a
2
−ξ
2
_
a
4
+ ξ
2
(b
2
−a
2
)
, 0, −
bξ
_
a
4
+ ξ
2
(b
2
−a
2
)
_
.
It is evident that D = [D[.
The numerical integration was performed using a 9point twodimensional Gausstype
integration formula. The result is presented on Figure 1.8. The parameter values are given
in Table 1.2.
The example considered in this subsection may not be physically relevant. However
it shows how Okada’s solution can be extended. For a more precise modeling of the
faulting event we need to have more information about the earthquake source and its
related parameters.
After having reviewed the description of the source, we now switch to the deformation
of the ocean surface following a submarine earthquake. The traditional approach for hy
drodynamic modelers is to use elastic models similar to the model we just described with
the seismic parameters as input in order to evaluate the details of the seaﬂoor deformation.
Then this deformation is translated to the free surface of the ocean and serves as initial
condition of the evolution problem described in the next section.
1.1.4 Solution in ﬂuid domain
The ﬂuid domain is supposed to represent the ocean above the fault area. Let us
consider the ﬂuid domain Ω shown in Figure 1.9. It is bounded above by the free surface of
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.1 Waves generated by a moving bottom 17
Figure 1.8: Freesurface deformation due to curvilinear faulting. The horizontal distances x
and y are expressed in kilometers.
x
y
z
O
h Ω
η(x, y, t)
ζ(x, y, t)
Figure 1.9: Deﬁnition of the ﬂuid domain and coordinate system
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
18 Tsunami generation
the ocean and below by the rigid ocean ﬂoor. The domain Ω is unbounded in the horizontal
directions x and y, and can be written as
Ω = R
2
[−h + ζ(x, y, t), η(x, y, t)] .
Initially the ﬂuid is assumed to be at rest and the sea bottom to be horizontal. Thus,
at time t = 0, the free surface and the sea bottom are deﬁned by z = 0 and z = −h,
respectively. For time t > 0 the bottom boundary moves in a prescribed manner which is
given by
z = −h + ζ(x, y, t).
The displacement of the sea bottom is assumed to have all the properties required to com
pute its Fourier transform in x, y and its Laplace transform in t. The resulting deformation
of the free surface z = η(x, y, t) must be found. It is also assumed that the ﬂuid is in
compressible and the ﬂow is irrotational. The latter implies the existence of a velocity
potential φ(x, y, z, t) which completely describes this ﬂow. By deﬁnition of φ, the ﬂuid
velocity vector can be expressed as q = ∇φ. Thus, the continuity equation becomes
∇ q = ∆φ = 0, (x, y, z) ∈ Ω. (1.8)
The potential φ(x, y, z, t) must also satisfy the following kinematic boundary conditions on
the freesurface and the solid boundary, respectively:
∂φ
∂z
=
∂η
∂t
+
∂φ
∂x
∂η
∂x
+
∂φ
∂y
∂η
∂y
, z = η(x, y, t), (1.9)
∂φ
∂z
=
∂ζ
∂t
+
∂φ
∂x
∂ζ
∂x
+
∂φ
∂y
∂ζ
∂y
, z = −h + ζ(x, y, t). (1.10)
Assuming that viscous eﬀects as well as capillary eﬀects can be neglected, the dynamic
condition to be satisﬁed on the free surface reads
∂φ
∂t
+
1
2
[∇φ[
2
+ gη = 0, z = η(x, y, t). (1.11)
As described above, the initial conditions are given by
η(x, y, 0) = 0 and ζ(x, y, 0) = 0. (1.12)
The signiﬁcance of the various terms in the equations is more transparent when the
equations are written in dimensionless variables. The new independent variables are
¯ x = κx, ¯ y = κy, ¯ z = κz,
¯
t = σt,
where κ is a wavenumber and σ is a typical frequency. Note that here the same unit length
is used in the horizontal and vertical directions, as opposed to shallowwater theory.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.1 Waves generated by a moving bottom 19
The new dependent variables are
¯ η =
η
a
,
¯
ζ =
ζ
a
,
¯
φ =
κ
aσ
φ,
where a is a characteristic wave amplitude. A dimensionless water depth is also introduced:
¯
h = κh.
In dimensionless form, and after dropping the tildes, equations (1.8–1.11) become
∆φ = 0, (x, y, z) ∈ Ω,
∂φ
∂z
=
∂η
∂t
+ κa
_
∂φ
∂x
∂η
∂x
+
∂φ
∂y
∂η
∂y
_
, z = κa η(x, y, t),
∂φ
∂z
=
∂ζ
∂t
+ κa
_
∂φ
∂x
∂ζ
∂x
+
∂φ
∂y
∂ζ
∂y
_
, z = −h + κa ζ(x, y, t),
∂φ
∂t
+
1
2
κa[∇φ[
2
+
gκ
σ
2
η = 0, z = κa η(x, y, t).
Finding the solution to this problem is quite a diﬃcult task due to the nonlinearities
and the a priori unknown free surface. In this study we linearize the equations and the
boundary conditions by taking the limit as κa →0. In fact, the linearized problem can be
found by expanding the unknown functions as power series of a small parameter ε := κa.
Collecting the lowest order terms in ε yields the linear approximation. For the sake of
convenience, we now switch back to the physical variables. The linearized problem in
dimensional variables reads
∆φ = 0, (x, y, z) ∈ R
2
[−h, 0], (1.13)
∂φ
∂z
=
∂η
∂t
, z = 0, (1.14)
∂φ
∂z
=
∂ζ
∂t
, z = −h, (1.15)
∂φ
∂t
+ gη = 0, z = 0. (1.16)
Combining equations (1.14) and (1.16) yields the single freesurface condition
∂
2
φ
∂t
2
+ g
∂φ
∂z
= 0, z = 0. (1.17)
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
20 Tsunami generation
This problem will be solved by using the method of integral transforms. We apply the
Fourier transform in (x, y):
F[f] =
´
f(k, ℓ) =
_
R
2
f(x, y)e
−i(kx+ℓy)
dxdy,
F
−1
[
´
f] = f(x, y) =
1
(2π)
2
_
R
2
´
f(k, ℓ)e
i(kx+ℓy)
dkdℓ,
and the Laplace transform in time t:
L[g] = g(s) =
+∞
_
0
g(t)e
−st
dt.
For the combined Fourier and Laplace transforms, the following notation is introduced:
FL[F(x, y, t)] = F(k, ℓ, s) =
_
R
2
e
−i(kx+ℓy)
dxdy
+∞
_
0
F(x, y, t)e
−st
dt.
After applying the transforms, equations (1.13), (1.15) and (1.17) become
d
2
φ
dz
2
−(k
2
+ ℓ
2
)φ = 0, (1.18)
dφ
dz
(k, ℓ, −h, s) = sζ(k, ℓ, s), (1.19)
s
2
φ(k, ℓ, 0, s) + g
dφ
dz
(k, ℓ, 0, s) = 0. (1.20)
The transformed freesurface elevation can be obtained from (1.16):
η(k, ℓ, s) = −
s
g
φ(k, ℓ, 0, s). (1.21)
A general solution of equation (1.18) is given by
φ(k, ℓ, z, s) = A(k, ℓ, s) cosh(mz) + B(k, ℓ, s) sinh(mz), (1.22)
where m =
√
k
2
+ ℓ
2
. The functions A(k, ℓ, s) and B(k, ℓ, s) can be easily found from the
boundary conditions (1.19) and (1.20):
A(k, ℓ, s) = −
gsζ(k, ℓ, s)
cosh(mh)[s
2
+ gmtanh(mh)]
,
B(k, ℓ, s) =
s
3
ζ(k, ℓ, s)
mcosh(mh)[s
2
+ gmtanh(mh)]
.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.1 Waves generated by a moving bottom 21
0 1 2 3 4 5 6 7 8 9 10
−0.5
0
0.5
1
1.5
2
2.5
3
3.5
m
ω
ω(m)
ω’(m)
ω’’(m)
Figure 1.10: Plot of the frequency ω(m) =
_
gmtanh(mh) and its derivatives dω/dm, d
2
ω/dm
2
.
The acceleration due to gravity g and the water depth h have been set equal to 1.
From now on, the notation
ω =
_
gmtanh(mh) (1.23)
will be used. The graphs of ω(m), ω
′
(m) and ω
′′
(m) are shown in Figure 1.10.
Substituting the expressions for the functions A, B in (1.22) yields
φ(k, ℓ, z, s) = −
gsζ(k, ℓ, s)
cosh(mh)(s
2
+ ω
2
)
_
cosh(mz) −
s
2
gm
sinh(mz)
_
. (1.24)
1.1.5 Freesurface elevation
From (1.21), the freesurface elevation becomes
η(k, ℓ, s) =
s
2
ζ(k, ℓ, s)
cosh(mh)(s
2
+ ω
2
)
.
Inverting the Laplace and Fourier transforms provides the general integral solution
η(x, y, t) =
1
(2π)
2
__
R
2
e
i(kx+ℓy)
cosh(mh)
1
2πi
µ+i∞
_
µ−i∞
s
2
ζ(k, ℓ, s)
s
2
+ ω
2
e
st
ds dkdℓ. (1.25)
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
22 Tsunami generation
One can evaluate the Laplace integral in (1.25) using the convolution theorem:
L[f
1
(t) ∗ f
2
(t)] = f
1
(s)f
2
(s).
It yields
η(x, y, t) =
1
(2π)
2
__
R
2
e
i(kx+ℓy)
cosh(mh)
t
_
0
(1 −ω sin ωτ)ζ(k, ℓ, t −τ)dτ dkdℓ.
This general solution contains as a special case the solution for an axisymmetric prob
lem, which we now describe in detail. Assume that the initial solid boundary deformation
is axisymmetric:
ζ(x, y) = ζ(r), r =
_
x
2
+ y
2
.
The Fourier transform F[ζ(x, y)] =
´
ζ(k, ℓ) of an axisymmetric function is also axisymmetric
with respect to transformation parameters, i.e.
´
ζ(k, ℓ) =
´
ζ(m), m :=
√
k
2
+ ℓ
2
.
In the following calculation, we use the notation ψ = arctan(ℓ/k). One has
´
ζ(k, ℓ) =
__
R
2
ζ(r)e
−i(kx+ℓy)
dxdy =
2π
_
0
dφ
∞
_
0
ζ(r)e
−ir(k cos φ+ℓ sin φ)
rdr =
=
2π
_
0
dφ
∞
_
0
rζ(r)e
−irmcos(φ−ψ)
dr =
∞
_
0
rζ(r)dr
π
_
0
(e
−irmcos φ
+ e
irmcos φ
)dφ.
Using an integral representation of Bessel functions [GR00] ﬁnally yields
´
ζ(k, ℓ) = 2π
∞
_
0
rζ(r)J
0
(mr)dr ≡
´
ζ(m).
It follows that
η(r, t) =
1
(2π)
2
2π
_
0
dψ
+∞
_
0
me
imr cos(φ−ψ)
cosh(mh)
dm
t
_
0
(1 −ω sin ωτ)ζ(m, t −τ) dτ
=
1
2π
+∞
_
0
m
J
0
(mr)
cosh(mh)
dm
t
_
0
(1 −ω sin ωτ)ζ(m, t −τ)dτ.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.1 Waves generated by a moving bottom 23
The last equation gives the general integral solution of the problem in the case of an
axisymmetric seabed deformation. Below we no longer make this assumption since Okada’s
solution does not have this property.
In the present study we consider seabed deformations with the following structure:
ζ(x, y, t) := ζ(x, y)T(t). (1.26)
Mathematically we separate the time dependence from the spatial coordinates. There are
two main reasons for doing this. First of all we want to be able to invert analytically
the Laplace transform. The second reason is more fundamental. In fact, dynamic source
models are not easily available. Okada’s solution, which was described in the previous
section, provides the static seabed deformation ζ
0
(x, y) and we will consider diﬀerent
time dependencies T(t) to model the time evolution of the source. Four scenarios will be
considered:
1. Instantaneous: T
i
(t) = H(t), where H(t) denotes the Heaviside step function,
2. Exponential:
T
e
(t) =
_
0, t < 0,
1 −e
−αt
, t ≥ 0,
with α > 0,
3. Trigonometric: T
c
(t) = H(t −t
0
) +
1
2
[1 −cos(πt/t
0
)]H(t
0
−t),
4. Linear:
T
l
(t) =
_
¸
_
¸
_
0, t < 0,
t/t
0
, 0 ≤ t ≤ t
0
,
1, t > t
0
.
The typical graphs of T
c
(t) and T
e
(t) are shown in Figure 1.11. Inserting (1.26) into (1.25)
yields
η(x, y, t) =
1
(2π)
2
__
R
2
´
ζ(k, ℓ)e
i(kx+ℓy)
cosh(mh)
1
2πi
µ+i∞
_
µ−i∞
s
2
T(s)
s
2
+ ω
2
e
st
ds dkdℓ. (1.27)
Clearly, η(x, y, t) depends continuously on the source ζ(x, y). Physically it means that
small variations of ζ (in a reasonable space of functions such as L
2
) yield small variations
of η. Mathematically this problem is said to be wellposed, and this property is essential
for modelling the physical processes, since it means that small modiﬁcations of the ground
motion (for example, the error in measurements) do not induce huge modiﬁcations of the
wave patterns.
Using the special representation (1.26) of seabed deformation and prescribed time
dependencies, one can compute analytically the Laplace integral in (1.27). To perform this
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
24 Tsunami generation
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.2
0.4
0.6
0.8
1
t
T
(
t
)
T
e
(t)
T
c
(t)
Figure 1.11: Typical graphs of T
e
(t) and T
c
(t). Here we have set α = 6.2, t
0
= 0.7.
integration, we ﬁrst have to compute the Laplace transform of T
i,e,c,l
(t). The results of
these computaions are given in Table 1.3.
scenario Laplace transform
Instantaneous L[T
i
] =
1
s
Exponential L[T
e
] =
α
s(α+s)
Trigonometric L[T
c
] =
γ
2
2s(s
2
+γ
2
)
(1 + e
−st
0
)
with γ =
π
t
0
Linear L[T
l
] =
1−e
−st
0
t
0
s
2
Table 1.3: Laplace transforms of diﬀerent timedepencies for prescribed bottom motion.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.1 Waves generated by a moving bottom 25
Inserting these formulas into the inverse Laplace integral yields
1
2πi
µ+i∞
_
µ−i∞
e
st
s
2
T
i
(s)
s
2
+ ω
2
ds = cos ωt,
1
2πi
µ+i∞
_
µ−i∞
e
st
s
2
T
e
(s)
s
2
+ ω
2
ds = −
α
2
α
2
+ ω
2
_
e
−αt
−cos ωt −
ω
α
sin ωt
_
,
1
2πi
µ+i∞
_
µ−i∞
e
st
s
2
T
c
(s)
s
2
+ ω
2
ds =
γ
2
2(γ
2
−ω
2
)
(cos ωt −cos γt + H(t −t
0
)[cos ω(t −t
0
) + cos γt]) ,
1
2πi
µ+i∞
_
µ−i∞
e
st
s
2
T
l
(s)
s
2
+ ω
2
ds =
sin ωt −H(t −t
0
) sin ω(t −t
0
)
ωt
0
.
The ﬁnal integral formulas for the freesurface elevations with diﬀerent time dependen
cies are as follows:
η
i
(x, y, t) =
1
(2π)
2
__
R
2
´
ζ(k, ℓ)e
i(kx+ℓy)
cosh(mh)
cos ωt dkdℓ,
η
e
(x, y, t) =
−α
2
(2π)
2
__
R
2
´
ζ(k, ℓ)e
i(kx+ℓy)
cosh(mh)
_
e
−αt
−cos ωt −
ω
α
sin ωt
α
2
+ ω
2
_
dkdℓ,
η
c
(x, y, t) =
γ
2
(2π)
2
__
R
2
´
ζ(k, ℓ)e
i(kx+ℓy)
2(γ
2
−ω
2
) cosh(mh)
(cos ωt −cos γt + H(t −t
0
)[cos ω(t −t
0
) + cos γt]) dkdℓ,
η
l
(x, y, t) =
1
(2π)
2
__
R
2
´
ζ(k, ℓ)e
i(kx+ℓy)
cosh(mh)
_
sin ωt −H(t −t
0
) sin ω(t −t
0
)
ωt
0
_
dkdℓ.
1.1.6 Velocity ﬁeld
In some applications it is important to know not only the freesurface elevation but also
the velocity ﬁeld in the ﬂuid domain. One of the goals of this work is to provide an initial
condition for tsunami propagation codes. For the time being, tsunami modelers take initial
seabed deformations and translate them directly to the free surface in order to obtain the
initial condition η(x, y, 0). Since a priori there is no information on the ﬂow velocities,
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
26 Tsunami generation
they take a zero velocity ﬁeld as initial condition for the velocity: ∇φ(x, y, z, 0) = 0. The
present computations show that it is indeed a very good approximation if the generation
time is short.
In equation (1.24), we obtained the Fourier transform of the velocity potential φ(x, y, z, t):
φ(k, ℓ, z, s) = −
gs
´
ζ(k, ℓ)T(s)
cosh(mh)(s
2
+ ω
2
)
_
cosh(mz) −
s
2
gm
sinh(mz)
_
. (1.28)
Let us evaluate the velocity ﬁeld at an arbitrary level z = βh with −1 ≤ β ≤ 0. In
the linear approximation the value β = 0 corresponds to the free surface while β = −1
corresponds to the bottom. Next we introduce some notation. The horizontal velocities
are denoted by u. The horizontal gradient (∂/∂x, ∂/∂y) is denoted by ∇
h
. The vertical
velocity component is simply w. The Fourier transform parameters are denoted k = (k, ℓ).
Taking the Fourier and Laplace transforms of
u(x, y, t) = ∇
h
φ(x, y, z, t)[
z=βh
yields
u(k, ℓ, s) = −iφ(k, ℓ, βh, s)k
= i
gs
´
ζ(k, ℓ)T(s)
cosh(mh)(s
2
+ ω
2
)
_
cosh(βmh) −
s
2
gm
sinh(βmh)
_
k.
Inverting the Fourier and Laplace transforms gives the general formula for the horizontal
velocities:
u(x, y, t) =
ig
4π
2
__
R
2
k
´
ζ(k, ℓ) cosh(mβh)e
i(kx+ℓy)
cosh(mh)
1
2πi
µ+i∞
_
µ−i∞
sT(s)e
st
s
2
+ ω
2
ds dk
−
i
4π
2
__
R
2
k
´
ζ(k, ℓ) sinh(mβh)e
i(kx+ℓy)
mcosh(mh)
1
2πi
µ+i∞
_
µ−i∞
s
3
T(s)e
st
s
2
+ ω
2
ds dk.
After a few computations, one ﬁnds the formulas for the time dependencies T
i
, T
e
and
T
l
. For simplicity we only give the velocities along the free surface (β = 0):
u
i
(x, y, t) =
ig
4π
2
__
R
2
k
´
ζ(k, ℓ)e
i(kx+ℓy)
cosh(mh)
sin ωt
ω
dk,
u
e
(x, y, t) =
igα
4π
2
__
R
2
k
´
ζ(k, ℓ)e
i(kx+ℓy)
(α
2
+ ω
2
) cosh(mh)
_
e
−αt
−cos ωt +
α
ω
sin ωt
_
dk,
u
l
(x, y, t) =
ig
4t
0
π
2
__
R
2
k
´
ζ(k, ℓ)e
i(kx+ℓy)
ω
2
cosh(mh)
(1 −cos ωt −H(t −t
0
)[1 −cos ω(t −t
0
)]) dk.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.1 Waves generated by a moving bottom 27
Next we determine the vertical component of the velocity w(x, y, z, t). It is easy to
obtain the Fourier–Laplace transform w(k, ℓ, z, s) by diﬀerentiating (1.28):
w(k, ℓ, z, s) =
∂φ
∂z
=
sg
´
ζ(k, ℓ)T(s)
cosh(mh)(s
2
+ ω
2
)
_
s
2
g
cosh(mz) −msinh(mz)
_
.
Inverting this transform yields
w(x, y, z, t) =
1
4π
2
__
R
2
cosh(mz)
´
ζ(k, ℓ)
cosh(mh)
e
i(kx+ℓy)
1
2πi
µ+i∞
_
µ−i∞
s
3
T(s)e
st
s
2
+ ω
2
ds dk
−
g
4π
2
__
R
2
msinh(mz)
´
ζ(k, ℓ)
cosh(mh)
e
i(kx+ℓy)
1
2πi
µ+i∞
_
µ−i∞
sT(s)e
st
s
2
+ ω
2
ds dk,
for −h < z ≤ 0. One can easily obtain the expression of the vertical velocity at a given
vertical level by substituting z = βh in the expression for w.
The easiest way to compute the vertical velocity w along the free surface is to use
the boundary condition (1.14). Indeed, the expression for w can be simply derived by
diﬀerentiating the known formula for η
i,e,c,l
(x, y, t). Note that formally the derivative gives
the distributions δ(t) and δ(t − t
0
) under the integral sign. It is a consequence of the
idealized time behaviour (such as the instantaneous scenario) and it is a disadvantage of
the Laplace transform method. In order to avoid these distributions we can consider the
solutions only for t > 0 and t ,= t
0
. From a practical point of view there is no restriction
since for any ε > 0 we can set t = ε or t = t
0
+ε. For small values of ε this will give a very
good approximation of the solution behaviour at these “critical” instants of time. Under
this assumption we give the distributionfree expressions for the vertical velocity along the
free surface:
w
i
(x, y, t) = −
1
4π
2
__
R
2
´
ζ(k, ℓ)e
i(kx+ℓy)
cosh(mh)
ω sin ωt dk,
w
e
(x, y, t) =
α
3
4π
2
__
R
2
´
ζ(k, ℓ)e
i(kx+ℓy)
(α
2
+ ω
2
) cosh(mh)
_
e
−αt
+
ω
2
α
2
cos ωt −
ω
α
sin ωt
_
dk,
w
c
(x, y, t) = −
γ
2
4π
2
__
R
2
´
ζ(k, ℓ)e
i(kx+ℓy)
2(γ
2
−ω
2
) cosh(mh)
_
ω sin ωt −γ sin γt
+H(t −t
0
)[ω sin ω(t −t
0
) + γ sin γt]
_
dk,
w
l
(x, y, t) =
1
4t
0
π
2
__
R
2
´
ζ(k, ℓ)e
i(kx+ℓy)
cosh(mh)
[cos ωt −H(t −t
0
) cos ω(t −t
0
)] dk.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
28 Tsunami generation
1.1.6.1 Pressure on the bottom
Since tsunameters have one component that measures the pressure at the bottom (bot
tom pressure recorder or simply BPR [GBM
+
05]), it is interesting to provide as well the
expression p
b
(x, y, t) for the pressure at the bottom. The pressure p(x, y, z, t) can be
obtained from Bernoulli’s equation, which was written explicitly for the free surface in
equation (1.11), but is valid everywhere in the ﬂuid:
∂φ
∂t
+
1
2
[∇φ[
2
+ gz +
p
ρ
= 0. (1.29)
After linearization, equation (1.29) becomes
∂φ
∂t
+ gz +
p
ρ
= 0. (1.30)
Along the bottom, it reduces to
∂φ
∂t
+ g(−h + ζ) +
p
b
ρ
= 0, z = −h. (1.31)
The timederivative of the velocity potential is readily available in Fourier space. Inverting
the Fourier and Laplace transforms and evaluating the resulting expression at z = −h
gives for the four time scenarios, respectively,
∂φ
i
∂t
= −
g
(2π)
2
__
R
2
´
ζ(k, ℓ)e
i(kx+ℓy)
cosh
2
(mh)
cos ωt dk,
∂φ
e
∂t
=
gα
2
(2π)
2
__
R
2
´
ζ(k, ℓ)e
i(kx+ℓy)
α
2
+ ω
2
_
e
−αt
−cos ωt −
ω
α
sin ωt
_
dk +
α
4
(2π)
2
__
R
2
´
ζ(k, ℓ) tanh(mh)e
i(kx+ℓy)
m(α
2
+ ω
2
)
_
e
−αt
+
_
ω
α
_
2
cos ωt +
_
ω
α
_
3
sin ωt
_
dk,
∂φ
l
∂t
= −
g
t
0
(2π)
2
__
R
2
´
ζ(k, ℓ)e
i(kx+ℓy)
ω cosh
2
(mh)
[sin ωt −H(t −t
0
) sin ω(t −t
0
)] dk.
The bottom pressure deviation from the hydrostatic pressure is then given by
p
b
(x, y, t) = − ρ
∂φ
∂t
¸
¸
¸
¸
z=−h
−ρgζ.
Plots of the bottom pressure will be given in Section 1.1.8.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.1 Waves generated by a moving bottom 29
1.1.7 Asymptotic analysis of integral solutions
In this subsection, we apply the method of stationary phase in order to estimate the far
ﬁeld behaviour of the solutions. There is a lot of literature on this topic (see for example
[Erd56, Mur92, PL71, BH86, ES94]). This method is a classical method in asymptotic
analysis. To our knowledge, the stationary phase method was ﬁrst used by Kelvin [Kel87]
in the context of linear waterwave theory.
The motivation to obtain asymptotic formulas for integral solutions was mainly due to
numerical diﬃculties to calculate the solutions for large values of x and y. From equation
(1.25), it is clear that the integrand is highly oscillatory. In order to be able to resolve these
oscillations, several discretization points are needed per period. This becomes extremely
expensive as r =
_
x
2
+ y
2
→ ∞. The numerical method used in the present study is
based on a Filontype quadrature formula [Fil28] and has been adapted to double integrals
with exp[i(kx + ℓy)] oscillations. The idea of this method consists in interpolating only
the amplitude of the integrand at discretization points by some kind of polynomial or
spline and then performing exact integration for the oscillating part of the integrand. This
method seems to be quite eﬃcient.
Let us ﬁrst obtain an asymptotic representation for integral solutions of the general
form
η(x, y, t) =
1
4π
2
__
R
2
´
ζ(k, ℓ)e
i(kx+ℓy)
cosh(mh)
T(m, t) dkdℓ, m =
√
k
2
+ ℓ
2
. (1.32)
Comparing with equation (1.27) shows that T(m, t) is in fact
T(m, t) =
1
2πi
µ+i∞
_
µ−i∞
s
2
T(s)
s
2
+ ω
2
e
st
ds.
For example, we showed above that for an instantaneous seabed deformation T(m, t) =
cos ωt, where ω
2
= gmtanh mh. For the time being, we do not specify the time behaviour
T(s).
In equation (1.32), we switch to polar coordinates m and ψ = arctan(ℓ/k):
η(x, y, t) =
1
4π
2
∞
_
0
2π
_
0
´
ζ(m, ψ)e
imr cos(ϕ−ψ)
cosh(mh)
T(m, t)m dψdm
=
1
4π
2
∞
_
0
mT(m, t)
cosh(mh)
dm
2π
_
0
´
ζ(m, ψ)e
imr cos(ϕ−ψ)
dψ,
where (r, ϕ) are the polar coordinates of (x, y). In the last expression, the phase function
is Φ = mr cos(ϕ − ψ). Stationary phase points satisfy the condition ∂Φ/∂ψ = 0, which
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
30 Tsunami generation
yields two phases: ψ
1
= ϕ and ψ
2
= ϕ + π. An approximation to equation (1.32) is then
obtained by applying the method of stationary phase to the integral over ψ:
η(r, φ, t) ≃
1
√
8π
3
r
∞
_
0
√
mT(m, t)
cosh(mh)
_
´
ζ(m, ϕ)e
i(
π
4
−mr)
+
´
ζ(m, ϕ + π)e
i(mr−
π
4
)
_
dm.
This expression cannot be simpliﬁed if we do not make any further hypotheses on the
function T(m, t).
Since we are looking for the far ﬁeld solution behaviour, the details of wave formation are
not important. Thus we will assume that the initial seabed deformation is instantaneous:
T(m, t) = cos ωt =
e
iωt
+ e
−iωt
2
.
Inserting this particular function T(m, t) in equation (1.32) yields
η(r, ϕ, t) =
1
8π
2
_
I
1
+ I
2
_
,
where
I
1
=
∞
_
0
m
´
ζ(m, ψ)
cosh(mh)
2π
_
0
e
i(ωt+mr cos(ϕ−ψ))
dψdm,
I
2
=
∞
_
0
m
´
ζ(m, ψ)
cosh(mh)
2π
_
0
e
i(−ωt+mr cos(ϕ−ψ))
dψdm.
The stationary phase function in these integrals is
Φ(m, ψ) = mr cos(ϕ −ψ) ±ωt, ω
2
(m) = gmtanh mh.
The points of stationary phase are then obtained from the conditions
∂Φ
∂ψ
= 0,
∂Φ
∂m
= 0.
The ﬁrst equation gives two points, ψ
1
= ϕ and ψ
2
= ϕ+π, as before. The second condition
yields
r
t
cos(ϕ −ψ
1,2
) = ∓
dω
dm
. (1.33)
Since dω/dm decreases from
√
gh to 0 as m goes from 0 to ∞ (see Figure 1.10), this
equation has a unique solution for m if [r/t[ ≤
√
gh. This unique solution will be denoted
by m
∗
.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.1 Waves generated by a moving bottom 31
For [r[ > t
√
gh, there is no stationary phase. It means physically that the wave has
not yet reached this region. So we can approximately set I
1
≈ 0 and I
2
≈ 0. From the
positivity of the function dω/dm one can deduce that ψ
1
= ϕ is a stationary phase point
only for the integral I
2
. Similarly, ψ
2
= ϕ+π is a stationary point only for the integral I
1
.
Let us obtain an asymptotic formula for the ﬁrst integral:
I
1
≈
∞
_
0
m
cosh(mh)
_
_
2π
mr
´
ζ(m, ϕ + π)e
i(ωt−mr)
e
i
π
4
_
dm
=
_
2π
r
e
i
π
4
∞
_
0
´
ζ(m, ϕ + π)
cosh(mh)
√
me
i(ωt−mr)
dm
≈
_
2π
r
e
i
π
4
_¸
2πm
∗
[ω
′′
(m
∗
)[ t
´
ζ(m
∗
, ϕ + π)
cosh(m
∗
h)
e
i(ω(m
∗
)t−m
∗
r)
e
−i
π
4
_
=
2π
t
_
m
∗
−ω
′′
ω
′
´
ζ(m
∗
, ϕ + π)
cosh(m
∗
h)
e
i(ω(m
∗
)t−m
∗
r)
.
In this estimate we have used equation (1.33) evaluated at the stationary phase point
(m
∗
, ψ
2
):
r = t
dω
dm
¸
¸
¸
¸
m=m
∗
. (1.34)
Similarly one can obtain an estimate for the integral I
2
:
I
2
≈
2π
t
_
m
∗
−ω
′′
ω
′
´
ζ(m
∗
, ϕ)
cosh(m
∗
h)
e
−i(ω(m
∗
)t−m
∗
r)
.
Asymptotic values have been obtained for the integrals. As is easily observed from the
expressions for I
1
and I
2
, the wave train decays as 1/t, or 1/r, which is equivalent since r
and t are connected by relation (1.34).
1.1.8 Numerical results
A lot of numerical computations based on the analytical formulas obtained in the
previous sections have been performed. Because of the lack of information about the
real dynamical characteristics of tsunami sources, we cannot really conclude which time
dependence gives the best description of tsunami generation. At this stage it is still very
diﬃcult or even impossible.
Numerical experiments showed that the largest wave amplitudes with the time depen
dence T
c
(t) were obtained for relatively small values of the characteristic time t
0
. The
exponential dependence has shown higher amplitudes for relatively longer characteristic
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
32 Tsunami generation
Parameter Value
Young modulus, E, GPa 9.5
Poisson ratio, ν 0.27
Fault depth, d, km 20
Dip angle, δ,
◦
13
Strike angle, θ,
◦
90
Normal angle, φ,
◦
0
Fault length, L, km 60
Fault width, W, km 40
Burger’s vector length, [D[, m 15
Water depth, h, km 4
Acceleration due to gravity, g, m/s
2
9.8
Wave number, k, 1/m 10
−4
Angular frequency, ω, Hz 10
−2
Table 1.4: Physical parameters used in the numerical computations
times. The instantaneous scenario T
i
gives at the free surface the initial seabed defor
mation with a slightly lower amplitude (the factor that we obtained was typically about
0.8 ∼ 0.94). The water has a highpass ﬁlter eﬀect on the initial solid boundary deforma
tion. The linear time dependence T
l
(t) showed a linear growth of wave amplitude from 0
to also ≈ 0.9ζ
0
, where ζ
0
= max
(x,y)∈R
2
[ζ(x, y)[.
In this section we provide several plots (Figure 1.12) of the freesurface deformation.
For illustration purposes, we have chosen the instantaneous seabed deformation since it is
the most widely used. The values of the parameters used in the computations are given in
Table 1.4. We also give plots of the velocity components on the free surface a few seconds
(physical) after the instantaneous deformation (Figure 1.13). Finally, plots of the bottom
dynamic pressure are given in Figure 1.14.
From Figure 1.13 it is clear that the velocity ﬁeld is really negligible in the beginning
of wave formation. Numerical computations showed that this situation does not change if
one takes other timedependences.
The main focus of the present section is the generation of waves by a moving bottom.
The asymptotic behaviour of various sets of initial data propagating in a ﬂuid of uniform
depth has been studied in detail by Hammack and Segur [HS74, HS78]. In particular, they
showed that the behaviours for an initial elevation wave and for an initial depression wave
are diﬀerent.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.1 Waves generated by a moving bottom 33
Figure 1.12: Freesurface elevation at t = 0.01, 0.6, 3, 5 in dimensionless time. In physical time
it corresponds to one second, one minute, ﬁve minutes and eight minutes and a half after the
initial seabed deformation.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
34 Tsunami generation
Figure 1.13: Components u, v and w of the velocity ﬁeld computed along the free surface at
t = 0.01, that is one second after the initial seabed deformation.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.1 Waves generated by a moving bottom 35
Figure 1.14: Bottom pressure at t = 0.01, 0.6, 3, 5 in dimensionless time. In physical time it
corresponds to one second, one minute, ﬁve minutes and eight minutes and a half after the initial
seabed deformation.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
36 Tsunami generation
1.2 Comparison between threedimensional linear and non
linear tsunami generation models
Tsunami wave modeling is a challenging task. In particular, it is essential to understand
the ﬁrst minutes of a tsunami, its propagation and ﬁnally the resulting inundation and
impact on structures. The focus of the present work is on the generation process. There
are diﬀerent natural phenomena that can lead to a tsunami. For example, one can mention
submarine mass failures, slides, volcanic eruptions, falls of asteroids, etc. We refer to the
recent review on tsunami science [SB06] for a complete bibliography on the topic. The
present work focuses on tsunami generation by earthquakes.
Two steps in modeling are necessary for an accurate description of tsunami genera
tion: a model for the earthquake fed by the various seismic parameters, and a model for
the formation of surface gravity waves resulting from the deformation of the seaﬂoor. In
the absence of sophisticated source models, one often uses analytical solutions based on
dislocation theory in an elastic halfspace for the seaﬂoor displacement [Oka85]. For the
resulting water motion, the standard practice is to transfer the inferred seaﬂoor displace
ment to the free surface of the ocean. In this study, we will call this approach the passive
generation approach.
1
This approach leads to a wellposed initial value problem with zero
velocity. An open question for tsunami forecasting modelers is the validity of neglecting
the initial velocity. In a recent note, Dutykh et al. [DDK06] used linear theory to show
that indeed diﬀerences may exist between the standard passive generation and the active
generation that takes into account the dynamics of seaﬂoor displacement. The transient
wave generation due to the coupling between the seaﬂoor motion and the free surface has
been considered by a few authors only. One of the reasons is that it is commonly as
sumed that the source details are not important.
2
BenMenahem and Rosenman [BMR72]
calculated the twodimensional radiation pattern from a moving source (linear theory).
Tuck and Hwang [Tuc74] solved the linear longwave equation in the presence of a moving
bottom and a uniformly sloping beach. Hammack [Ham73] generated waves experimen
tally by raising or lowering a box at one end of a channel. According to Synolakis and
Bernard [SB06], Houston and Garcia [HG74] were the ﬁrst to use more geophysically real
1
In the pioneering paper [Kaj63], Kajiura analyzed the applicability of the passive approach using
Green’s functions. In the tsunami literature, this approach is sometimes called the piston model of tsunami
generation.
2
As pointed out by Geist et al. [GTS06], the 2004 Indian Ocean tsunami shed some doubts about
this belief. The measurements from land based stations that use the Global Positioning System to track
ground movements revealed that the fault continued to slip after it stopped releasing seismic energy. Even
though this slip was relatively slow, it contributed to the tsunami and may explain the surprising tsunami
heights.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.2 Comparison of tsunami generation models 37
istic initial conditions. For obvious reasons, the quantitative diﬀerences in the distribution
of seaﬂoor displacement due to underwater earthquakes compared with more conventional
earthquakes are still poorly known. Villeneuve and Savage [VS93] derived model equa
tions which combine the linear eﬀect of frequency dispersion and the nonlinear eﬀect of
amplitude dispersion, and included the eﬀects of a moving bed. Todorovska and Trifunac
[TT01] considered the generation of tsunamis by a slowly spreading uplift of the seaﬂoor.
In this work, we mostly follow the standard passive generation approach. Several
tsunami generation models and numerical methods suited for these models are presented
and compared. The focus of our work is on modelling the ﬂuid motion. It is assumed that
the seabed deformation satisﬁes all the necessary hypotheses required to apply Okada’s
solution. The main objective is to conﬁrm or inﬁrm the lack of importance of nonlinear
eﬀects and/or frequency dispersion in tsunami generation. This result may have impli
cations in terms of computational cost. The goal is to optimize the ratio between the
complexity of the model and the accuracy of the results. Government agencies need to
compute accurately tsunami propagation in real time in order to know where to evacuate
people. Therefore any saving in computational time is crucial (see for example the code
MOST used by the National Oceanic and Atmospheric Administration in the US [TS98] or
the code TUNAMI developed by the Disaster Control Research Center in Japan). Liu and
Liggett [LL83] already performed comparisons between linear and nonlinear water waves
but their study was restricted to simple bottom deformations, namely the generation of
transient waves by an upthrust of a rectangular block, and the nonlinear computations
were restricted to twodimensional ﬂows. Bona et al. [BPS81] assessed how well a model
equation with weak nonlinearity and dispersion describes the propagation of surface wa
ter waves generated at one end of a long channel. In their experiments, they found that
the inclusion of a dissipative term was more important than the inclusion of nonlinear
ity, although the inclusion of nonlinearity was undoubtedly beneﬁcial in describing the
observations. The importance of dispersive eﬀects in tsunami propagation is not directly
addressed in the present study. Indeed these eﬀects cannot be measured without taking
into account the duration (or distance) of tsunami propagation [Tuc79].
1.2.1 Physical problem description
In the whole chapter, the vertical coordinate is denoted by z, while the two horizontal
coordinates are denoted by x and y, respectively. The sea bottom deformation follow
ing an underwater earthquake is a complex phenomenon. This is why, for theoretical or
experimental studies, researchers have often used simpliﬁed bottom motions such as the
vertical motion of a box. In order to determine the deformations of the sea bottom due
to an earthquake, we use the analytical solution obtained for a dislocation in an elastic
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
38 Tsunami generation
parameter value
Dip angle δ 13
◦
Fault depth d
f
, km 3
Fault length L, km 6
Fault width W, km 4
Magnitude of Burger’s vector [D[, m 1
Young’s modulus E, GPa 9.5
Poisson ratio ν 0.23
Table 1.5: Typical parameter set for the source used to model the seaﬂoor deformation due to an
earthquake in the present study. The dip angle, Young’s modulus and Poisson ratio correspond
roughly to those of the 2004 Sumatra event. The fault depth, length and width, as well as the
magnitude of Burger’s vector, have been reduced for computation purposes.
halfspace [Oka85]. This solution, which at present time is used by the majority of tsunami
wave modelers to produce an initial condition for tsunami propagation simulations, pro
vides an explicit expression of the bottom surface deformation that depends on a dozen
of source parameters such as the dip angle δ, fault depth d
f
, fault dimensions (length and
width), Burger’s vector D, Young’s modulus, Poisson ratio, etc. Some of these parameters
are shown in Figure 1.3. More details can be found in [DD07c] for example.
A value of 90
◦
for the dip angle corresponds to a vertical fault. Varying the fault slip
[D[ does not change the coseismic deformation pattern, only its magnitude. The values of
the parameters used in the present study are given in Table 1.5. A typical dipslip solution
is shown in Figure 1.15 (the angle φ is equal to 0, while the rake angle θ is equal to π/2).
Let z = ζ(x, y, t) denote the deformation of the sea bottom. Hammack and Segur [HS74]
suggested that there are two main kinds of behaviour for the generated waves depending
on whether the net volume V of the initial bottom surface deformation
V =
_
R
2
ζ(x, y, 0) dxdy
is positive or not.
3
A positive V is achieved for example for a “reverse fault”, i.e. when
the dip angle δ satisﬁes 0 ≤ δ ≤ π/2 or −π ≤ δ ≤ −π/2, as shown in Figure 1.16. A
negative V is achieved for a “normal fault”, i.e. when the dip angle δ satisﬁes π/2 ≤ δ ≤ π
or −π/2 ≤ δ ≤ 0.
The conclusions of [HS74] are based on the Korteweg–de Vries (KdV) equation and were
3
However it should be noted that the analysis of [HS74] is restricted to onedimensional unidirectional
waves. We assume here that their conclusions can be extended to twodimensional bidirectional waves.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.2 Comparison of tsunami generation models 39
Figure 1.15: Typical seaﬂoor deformation due to dipslip faulting. The parameters are those of
Table 1.5. The distances along the horizontal axes x and y are expressed in kilometers.
−150 −100 −50 0 50 100 150
−8
−6
−4
−2
0
2
4
6
8
δ
V
Reverse
fault
Normal
fault
Reverse
fault
Normal
fault
Figure 1.16: Initial net volume V (in km
3
) of the seaﬂoor displacement as a function of the dip
angle δ (in
◦
). All the other parameters, which are given in Table 1.5, are kept constant.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
40 Tsunami generation
in part conﬁrmed by their experiments. If V is positive, waves of stable form (solitons)
evolve and are followed by a dispersive train of oscillatory waves, regardless of the exact
structure of ζ(x, y, 0). If V is negative, and if the initial data is nonpositive everywhere,
no solitons evolve. But, if V is negative and there is a region of elevation in the initial data
(which corresponds to a typical Okada solution for a normal fault), solitons can evolve
and we have checked this last result using the FNPF equations (see Figure 1.17). In this
study, we focus on the case where V is positive with a dip angle δ equal to 13
◦
, according
to the seismic data of the 26 December 2004 SumatraAndaman event (see for example
[LKA
+
05]). However, the sea bottom deformation often has an N−shape, with subsidence
on one side of the fault and uplift on the other side as shown in Figure 1.15. In that case,
one may expect the positive V behaviour on one side and the negative V behaviour on the
other side. Recall that the experiments of Hammack and Segur [HS74] were performed in
the presence of a vertical wall next to the moving bottom and their analysis was based on
the unidirectional KdV wave equation.
We now consider the ﬂuid domain. A sketch is shown in Figure 1.9. The ﬂuid domain
Ω is bounded above by the free surface and below by the rigid ocean ﬂoor. It is unbounded
in the horizontal x− and y− directions. So, one can write
Ω = R
2
[−h(x, y) + ζ(x, y, t), η(x, y, t)].
Before the earthquake the ﬂuid is assumed to be at rest, thus the free surface and the
solid boundary are deﬁned by z = 0 and z = −h(x, y), respectively. For simplicity h(x, y)
is assumed to be a constant. Of course, in real situations, this is never the case but for
our purpose the bottom bathymetry is not important. Starting at time t = 0, the solid
boundary moves in a prescribed manner which is given by
z = −h + ζ(x, y, t), t ≥ 0.
The deformation of the sea bottom is assumed to have all the necessary properties
needed to compute its Fourier transform in x, y and its Laplace transform in t. The resulting
deformation of the free surface z = η(x, y, t) is to be found as part of the solution. It is
also assumed that the ﬂuid is incompressible and the ﬂow irrotational. The latter implies
the existence of a velocity potential φ(x, y, z, t) which completely describes the ﬂow. By
deﬁnition of φ the ﬂuid velocity vector can be expressed as q = ∇φ. Thus, the continuity
equation becomes
∇ q = ∆φ = 0, (x, y, z) ∈ Ω. (1.35)
The potential φ(x, y, z, t) must satisfy the following kinematic boundary conditions on the
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.2 Comparison of tsunami generation models 41
Figure 1.17: Wave proﬁles at diﬀerent times for the case of a normal fault (δ = 167
◦
). The
seaﬂoor deformation occurs instantaneously at t = 0. The water depth h(x, y) is assumed to be
constant.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
42 Tsunami generation
free surface and the solid boundary, respectively:
∂φ
∂z
=
∂η
∂t
+
∂φ
∂x
∂η
∂x
+
∂φ
∂y
∂η
∂y
, z = η(x, y, t),
∂φ
∂z
=
∂ζ
∂t
+
∂φ
∂x
∂ζ
∂x
+
∂φ
∂y
∂ζ
∂y
, z = −h + ζ(x, y, t).
Further assuming the ﬂow to be inviscid and neglecting surface tension eﬀects, one can
write the dynamic condition to be satisﬁed on the free surface as
∂φ
∂t
+
1
2
[∇φ[
2
+ gη = 0, z = η(x, y, t), (1.36)
where g is the acceleration due to gravity. The atmospheric pressure has been chosen as
reference pressure.
The equations are more transparent when written in dimensionless variables. However
the choice of the reference lengths and speeds is subtle. Diﬀerent choices lead to diﬀerent
models. Let the new independent variables be
˜ x = x/λ, ˜ y = y/λ, ˜ z = z/d,
˜
t = c
0
t/λ,
where λ is the horizontal scale of the motion and d a typical water depth. The speed c
0
is
the long wave speed based on the depth d (c
0
=
√
gd). Let the new dependent variables
be
˜ η =
η
a
,
˜
ζ =
ζ
a
,
˜
φ =
c
0
agλ
φ,
where a is a characteristic wave amplitude.
In dimensionless form, and after dropping the tildes, the equations become
∂
2
φ
∂z
2
+ µ
2
_
∂
2
φ
∂x
2
+
∂
2
φ
∂y
2
_
= 0, (x, y, z) ∈ Ω, (1.37)
∂φ
∂z
= µ
2
∂η
∂t
+ εµ
2
_
∂φ
∂x
∂η
∂x
+
∂φ
∂y
∂η
∂y
_
, z = εη(x, y, t), (1.38)
∂φ
∂z
= µ
2
∂ζ
∂t
+ εµ
2
_
∂φ
∂x
∂ζ
∂x
+
∂φ
∂y
∂ζ
∂y
_
, z = −
h
d
+ εζ(x, y, t), (1.39)
µ
2
∂φ
∂t
+
1
2
ε
_
µ
2
_
∂φ
∂x
_
2
+ µ
2
_
∂φ
∂y
_
2
+
_
∂φ
∂z
_
2
_
+ µ
2
η = 0, z = εη(x, y, t), (1.40)
where two dimensionless numbers have been introduced:
ε = a/d, µ = d/λ. (1.41)
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.2 Comparison of tsunami generation models 43
For the propagation of tsunamis, both numbers ε and µ are small. Indeed the satellite
altimetry observations of the 2004 Boxing Day tsunami waves obtained by two satellites
that passed over the Indian Ocean a couple of hours after the rupture process occurred gave
an amplitude a of roughly 60 cm in the open ocean. The typical wavelength estimated from
the width of the segments that experienced slip is between 160 and 240 km [LKA
+
05]. The
water depth ranges from 4 km towards the west of the rupture to 1 km towards the east.
Therefore average values for ε and µ in the open ocean are ε ≈ 2 10
−4
and µ ≈ 2 10
−2
.
A more precise range for these two dimensionless numbers is
1.5 10
−4
< ε < 6 10
−4
, 4 10
−3
< µ < 2.5 10
−2
. (1.42)
The waterwave problem, either in the form of an initial value problem (IVP) or in the
form of a boundary value problem (BVP), is diﬃcult to solve because of the nonlinearities
in the boundary conditions and the unknown computational domain.
1.2.2 Linear theory
First we perform the linearization of the above equations and boundary conditions. It
is equivalent to taking the limit of (1.37)–(1.40) as ε →0. The linearized problem can also
be obtained by expanding the unknown functions as power series of the small parameter
ε. Collecting terms of the lowest order in ε yields the linear approximation. For the sake
of convenience, we now switch back to the physical variables. The linearized problem in
dimensional variables reads
∆φ = 0, (x, y, z) ∈ R
2
[−h, 0], (1.43)
∂φ
∂z
=
∂η
∂t
, z = 0, (1.44)
∂φ
∂z
=
∂ζ
∂t
, z = −h, (1.45)
∂φ
∂t
+ gη = 0, z = 0. (1.46)
The bottom motion appears in equation (1.45). Combining equations (1.44) and (1.46)
yields the single freesurface condition
∂
2
φ
∂t
2
+ g
∂φ
∂z
= 0, z = 0. (1.47)
Most studies of tsunami generation assume that the initial freesurface deformation is
equal to the vertical displacement of the ocean bottom and take a zero velocity ﬁeld as
initial condition. The details of wave motion are completely neglected during the time
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
44 Tsunami generation
that the source operates. While tsunami modelers often justify this assumption by the
fact that the earthquake rupture occurs very rapidly, there are some speciﬁc cases where
the time scale and/or the horizontal extent of the bottom deformation may become an im
portant factor. This was emphasized for example by Todorovska and Trifunac [TT01] and
Todorovska et al. [THT02], who considered the generation of tsunamis by a slowly spread
ing uplift of the seaﬂoor in order to explain some observations related to past tsunamis.
However they did not use realistic source models.
Our claim is that it is important to make a distinction between two mechanisms of
generation: an active mechanism in which the bottom moves according to a given time law
and a passive mechanism in which the seaﬂoor deformation is simply translated to the free
surface. Recently Dutykh et al. [DDK06] showed that even in the case of an instantaneous
seaﬂoor deformation, there may be diﬀerences between these two generation processes.
1.2.3 Active generation
This case was extensively studied in Section 1.1. So, we do not repeat the computations
here and give only ﬁnal expressions (on the free surface and for t > 0) which correspond
to instantaneous bottom motion:
η(x, y, t) =
1
(2π)
2
__
R
2
´
ζ
0
(k, ℓ)e
i(kx+ℓy)
cosh(mh)
cos ωt dkdℓ, (1.48)
u(x, y, t; 0) =
ig
4π
2
__
R
2
k
´
ζ
0
(k, ℓ)e
i(kx+ℓy)
cosh(mh)
sin ωt
ω
dk, (1.49)
w(x, y, t; 0) = −
1
4π
2
__
R
2
´
ζ
0
(k, ℓ)e
i(kx+ℓy)
cosh(mh)
ω sin ωt dk. (1.50)
1.2.4 Passive generation
In this case equation (1.45) becomes
∂φ
∂z
= 0, z = −h, (1.51)
and the initial condition for η now reads
η(x, y, 0) = ζ
0
(x, y),
where ζ
0
(x, y) is the seaﬂoor deformation. Initial velocities are assumed to be zero.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.2 Comparison of tsunami generation models 45
Again we apply the Fourier transform in the horizontal coordinates (x, y). The Laplace
transform is not applied since there is no substantial dynamics in the problem. Equations
(1.43), (1.51) and (1.47) become
d
2
´
φ
dz
2
−(k
2
+ ℓ
2
)
´
φ = 0, (1.52)
d
´
φ
dz
(k, ℓ, −h, t) = 0, (1.53)
∂
2
´
φ
∂t
2
(k, ℓ, 0, t) + g
∂
´
φ
∂z
(k, ℓ, 0, t) = 0. (1.54)
A general solution to Laplace’s equation (1.52) is again given by
´
φ(k, ℓ, z, t) = A(k, ℓ, t) cosh(mz) + B(k, ℓ, t) sinh(mz), (1.55)
where m =
√
k
2
+ ℓ
2
. The relationship between the functions A(k, ℓ, t) and B(k, ℓ, t) can
be easily found from the boundary condition (1.53):
B(k, ℓ, t) = A(k, ℓ, t) tanh(mh). (1.56)
From equation (1.54) and the initial conditions one ﬁnds
A(k, ℓ, t) = −
g
ω
´
ζ
0
(k, ℓ) sin ωt. (1.57)
Substituting the expressions for the functions A and B in (1.55) yields
´
φ(k, ℓ, z, t) = −
g
ω
´
ζ
0
(k, ℓ) sin ωt
_
cosh(mz) + tanh(mh) sinh(mz)
_
. (1.58)
From (1.46), the freesurface elevation becomes
´ η(k, ℓ, t) =
´
ζ
0
(k, ℓ) cos ωt.
Inverting the Fourier transform provides the general integral solution
η(x, y, t) =
1
(2π)
2
__
R
2
´
ζ
0
(k, ℓ) cos ωt e
i(kx+ℓy)
dkdℓ. (1.59)
Let us now evaluate the velocity ﬁeld in the ﬂuid domain. Equation (1.58) gives the
Fourier transform of the velocity potential φ(x, y, z, t). Taking the Fourier transform of
u(x, y, t; β) = ∇
h
φ(x, y, z, t)[
z=βh
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
46 Tsunami generation
yields
´ u(k, ℓ, t; β) = −i
´
φ(k, ℓ, βh, t)k
= i
g
ω
´
ζ
0
(k, ℓ) sin ωt
_
cosh(βmh) + tanh(mh) sinh(βmh)
_
k.
Inverting the Fourier transform gives the general formula for the horizontal velocities
u(x, y, t; β) =
ig
4π
2
__
R
2
k
´
ζ
0
(k, ℓ)
sin ωt
ω
_
cosh(βmh) + tanh(mh) sinh(βmh)
_
e
i(kx+ℓy)
dk.
Along the free surface β = 0, the horizontal velocity vector becomes
u(x, y, t; 0) =
ig
4π
2
__
R
2
k
´
ζ
0
(k, ℓ)
sin ωt
ω
e
i(kx+ℓy)
dk. (1.60)
Next we determine the vertical component of the velocity w(x, y, t; β) at a given vertical
level z = βh. It is easy to obtain the Fourier transform ´ w(k, ℓ, t; β) by diﬀerentiating (1.58):
´ w(k, ℓ, t; β) =
∂
´
φ
∂z
¸
¸
¸
¸
¸
z=βh
= −mg
sin ωt
ω
´
ζ
0
(k, ℓ)
_
sinh(βmh) + tanh(mh) cosh(βmh)
_
.
Inverting this transform yields
w(x, y, t; β) = −
g
4π
2
__
R
2
msin ωt
ω
´
ζ
0
(k, ℓ)
_
sinh(βmh)+
tanh(mh) cosh(βmh)
_
e
i(kx+ℓy)
dk
for −1 ≤ β ≤ 0. Using the dispersion relation, one can write the vertical component of
the velocity along the free surface (β = 0) as
w(x, y, t; 0) = −
1
4π
2
__
R
2
ω sin ωt
´
ζ
0
(k, ℓ)e
i(kx+ℓy)
dk. (1.61)
All the formulas obtained in this section are valid only if the integrals converge.
Again, one can compute the bottom pressure. At z = −h, one has
∂φ
∂t
¸
¸
¸
¸
z=−h
= −
g
(2π)
2
__
R
2
´
ζ
0
(k, ℓ)e
i(kx+ℓy)
cosh(mh)
cos ωt dk.
The bottom pressure deviation from the hydrostatic pressure is then given by
p
b
(x, y, t) = − ρ
∂φ
∂t
¸
¸
¸
¸
z=−h
−ρgζ.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.2 Comparison of tsunami generation models 47
Again, away from the deformed seabed, p
b
reduces to − ρφ
t
[
z=−h
. The only diﬀerence
between p
b
and ρgη is the presence of an additional cosh(mh) term in the denominator of
p
b
.
The main diﬀerences between passive and active generation processes are that (i) the
wave amplitudes and velocities obtained with the instantly moving bottom are lower than
those generated by initial translation of the bottom motion and that (ii) the water column
plays the role of a lowpass ﬁlter (compare equations (1.48)–(1.50) with equations (1.59)–
(1.61)). High frequencies are attenuated in the moving bottom solution. Ward [War01],
who studied landslide tsunamis, also commented on the 1/ cosh(mh) term, which lowpass
ﬁlters the source spectrum. So the ﬁlter favors long waves. In the discussion section, we
will come back to the diﬀerences between passive generation and active generation.
1.2.4.1 Numerical method for the linear problem
All the expressions derived from linear theory are explicit but they must be computed
numerically. It is not a trivial task because of the oscillatory behaviour of the integrand
functions. All integrals were computed with Filon type numerical integration formulas
[Fil28], which explicitly take into account this oscillatory behaviour. Numerical results
will be shown in Section 1.2.9.
1.2.5 Nonlinear shallow water equations
Synolakis and Bernard [SB06] introduced a clear distinction between the various shallow
water models. At the lowest order of approximation, one obtains the linear shallow water
wave equation. The next level of approximation provides the nondispersive nonlinear shal
low water equations (NSW). In the next level, dispersive terms are added and the resulting
equations constitute the Boussinesq equations. Since there are many diﬀerent ways to go
to this level of approximation, there are a lot of diﬀerent types of Boussinesq equations.
The NSW equations are the most commonly used equations for tsunami propagation (see
in particular the code MOST developed by the National Oceanic and Atmospheric Ad
ministration in the US [TS98] or the code TUNAMI developed by the Disaster Control
Research Center in Japan). They are also used for generation and runup/inundation. For
wave runup, the eﬀects of bottom friction become important and must be included in the
codes. Our analysis will focus on the NSW equations. For simplicity, we assume below
that h is constant. Therefore one can take h as reference depth, so that the seaﬂoor is
given by z = −1 + εζ.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
48 Tsunami generation
1.2.6 Mathematical model
In this subsection, partial derivatives are denoted by subscripts. When µ
2
is a small
parameter, the water is considered to be shallow. For the shallow water theory, one formally
expands the potential φ in powers of µ
2
:
φ = φ
0
+ µ
2
φ
1
+ µ
4
φ
2
+ .
This expansion is substituted into the governing equation and the boundary conditions.
The lowestorder term in Laplace’s equation is
φ
0zz
= 0.
The boundary conditions imply that φ
0
= φ
0
(x, y, t). Thus the vertical velocity component
is zero and the horizontal velocity components are independent of the vertical coordinate
z at lowest order. Let us introduce the notation u := φ
0x
(x, y, t) and v := φ
0y
(x, y, t).
Solving Laplace’s equation and taking into account the bottom kinematic condition yield
the following expressions for φ
1
and φ
2
:
φ
1
(x, y, z, t) = −
1
2
Z
2
(u
x
+ v
y
) + z [ζ
t
+ ε(uζ
x
+ vζ
y
)] ,
φ
2
(x, y, z, t) =
1
24
Z
4
(∆u
x
+ ∆v
y
) + ε
_
ε
z
2
2
[∇ζ[
2
−
1
6
Z
3
∆ζ
_
(u
x
+ v
y
)
−
ε
3
Z
3
∇ζ ∇(u
x
+ v
y
) −
z
3
6
_
∆ζ
t
+ ε∆(uζ
x
+ vζ
y
)
_
+
z(−1 + εζ)
_
ε∇ζ ∇
_
ζ
t
+ ε(uζ
x
+ vζ
y
)
_
−ε
2
[∇ζ[
2
(u
x
+ v
y
)−
1
2
(−1 + εζ)
_
∆ζ
t
+ ε∆(uζ
x
+ vζ
y
)
_
_
,
where
Z = 1 + z −εζ.
The next step consists in retaining terms of requested order in the freesurface bound
ary conditions. Powers of ε will appear when expanding in Taylor series the freesurface
conditions around z = 0. For example, if one keeps terms of order εµ
2
and µ
4
in the
dynamic boundary condition (1.40) and in the kinematic boundary condition (1.38), one
obtains
µ
2
φ
0t
−
1
2
µ
4
(u
tx
+ v
ty
) + µ
2
η +
1
2
εµ
2
(u
2
+ v
2
) = 0, (1.62)
µ
2
[η
t
+ ε(uη
x
+ vη
y
) +
_
1 + ε(η −ζ)
_
(u
x
+ v
y
) −ζ
t
−ε(uζ
x
+ vζ
y
)] =
1
6
µ
4
(∆u
x
+ ∆v
y
). (1.63)
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.2 Comparison of tsunami generation models 49
Diﬀerentiating (1.62) ﬁrst with respect to x and then with respect to y gives a set of two
equations:
u
t
+ ε(uu
x
+ vv
x
) + η
x
−
1
2
µ
2
(u
txx
+ v
txy
) = 0, (1.64)
v
t
+ ε(uu
y
+ vv
y
) + η
y
−
1
2
µ
2
(u
txy
+ v
tyy
) = 0. (1.65)
The kinematic condition (1.63) becomes
(η −ζ)
t
+ [u(1 + ε(η −ζ))]
x
+ [v(1 + ε(η −ζ))]
y
=
1
6
µ
2
(∆u
x
+ ∆v
y
). (1.66)
Equations (1.64),(1.65) and (1.66) contain in fact various shallowwater models. The so
called fundamental NSW equations which contain no dispersive eﬀects are obtained by
neglecting the terms of order µ
2
:
u
t
+ ε(uu
x
+ vu
y
) + η
x
= 0, (1.67)
v
t
+ ε(uv
x
+ vv
y
) + η
y
= 0, (1.68)
η
t
+ [u(1 + ε(η −ζ))]
x
+ [v(1 + ε(η −ζ))]
y
= ζ
t
. (1.69)
Going back to a bathymetry h
∗
(x, y, t) equal to 1−εζ(x, y, t) and using the fact that (u, v)
is the horizontal gradient of φ
0
, one can rewrite the system of NSW equations as
u
t
+
ε
2
(u
2
+ v
2
)
x
+ η
x
= 0, (1.70)
v
t
+
ε
2
(u
2
+ v
2
)
y
+ η
y
= 0, (1.71)
η
t
+ [u(h
∗
+ εη)]
x
+ [v(h
∗
+ εη)]
y
= −
1
ε
h
∗
t
. (1.72)
The system of equations (1.70)–(1.72) has been used for example by Titov and Synolakis
[TS98] for the numerical computation of tidal wave runup. Note that this model does not
include any bottom friction terms.
The NSW equations with dispersion (1.64)–(1.66), also known as the Boussinesq equa
tions, can be written in the following form:
u
t
+
ε
2
(u
2
+ v
2
)
x
+ η
x
−
1
2
µ
2
∆u
t
= 0, (1.73)
v
t
+
ε
2
(u
2
+ v
2
)
y
+ η
y
−
1
2
µ
2
∆v
t
= 0, (1.74)
η
t
+ [u(h
∗
+ εη)]
x
+ [v(h
∗
+ εη)]
y
−
1
6
µ
2
(∆u
x
+ ∆v
y
) = −
1
ε
h
∗
t
. (1.75)
Kulikov et al. [KML05] have argued that the satellite altimetry observations of the Indian
Ocean tsunami show some dispersive eﬀects. However the steepness is so small that the
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
50 Tsunami generation
origin of these eﬀects is questionable. Guesmia et al. [GHM98] compared Boussinesq and
shallowwater models and came to the conclusion that the eﬀects of frequency dispersion
are minor. As pointed out in [KS06], dispersive eﬀects are necessary only when examining
steep gravity waves, which are not encountered in the context of tsunami hydrodynamics
in deep water. However they can be encountered in experiments such as those of Hammack
[Ham73] because the parameter µ is much bigger.
1.2.7 Numerical method
In order to solve the NSW equations, a ﬁnitevolume approach is used. For example
LeVeque [LeV98] used a highorder ﬁnite volume scheme to solve a system of NSW equa
tions. Here the ﬂux scheme we use is the characteristic ﬂux scheme, which was introduced
by Ghidaglia et al. [GKC96]. This numerical method satisﬁes the conservative properties
at the discrete level. The detailed description of this scheme is given in Appendix A.
1.2.8 Numerical method for the full equations
The fully nonlinear potential ﬂow (FNPF) equations (1.37)–(1.40) are solved by using
a numerical model based on the Boundary Element Method (BEM). An accurate code
was developed by Grilli et al. [GGD01]. It uses a highorder threedimensional bound
ary element method combined with mixed Eulerian–Lagrangian time updating, based on
secondorder explicit Taylor expansions with adaptive time steps. The eﬃciency of the
code was recently greatly improved by introducing a more eﬃcient spatial solver, based
on the fast multipole algorithm [FD06]. By replacing every matrix–vector product of the
iterative solver and avoiding the building of the inﬂuence matrix, this algorithm reduces
the computing complexity from O(N
2
) to nearly O(N) up to logarithms, where N is the
number of nodes on the boundary.
By using Green’s second identity, Laplace’s equation (1.35) is transformed into the
boundary integral equation
α(x
l
)φ(x
l
) =
_
Γ
_
∂φ
∂n
(x)G(x, x
l
) −φ(x)
∂G
∂n
(x, x
l
)
_
dΓ, (1.76)
where G is the threedimensional free space Green’s function. The notation ∂G/∂n means
the normal derivative, that is ∂G/∂n = ∇G n, with n the unit outward normal vec
tor. The vectors x = (x, y, z) and x
l
= (x
l
, y
l
, z
l
) are position vectors for points on the
boundary, and α(x
l
) = θ
l
/(4π) is a geometric coeﬃcient, with θ
l
the exterior solid angle
made by the boundary at point x
l
. The boundary Γ is divided into various parts with
diﬀerent boundary conditions. On the free surface, one rewrites the nonlinear kinematic
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.2 Comparison of tsunami generation models 51
and dynamic boundary conditions in a mixed EulerianLagrangian form,
DR
Dt
= ∇φ, (1.77)
Dφ
Dt
= −gz +
1
2
∇φ ∇φ, (1.78)
with R the position vector of a freesurface ﬂuid particle. The material derivative is deﬁned
as
D
Dt
=
∂
∂t
+q ∇. (1.79)
For time integration, secondorder explicit Taylor series expansions are used to ﬁnd
the new position and the potential on the free surface at time t + ∆t. This time stepping
scheme presents the advantage of being explicit, and the use of spatial derivatives along
the free surface provides a better stability of the computed solution.
The integral equations are solved by BEM. The boundary is discretized into N col
location nodes and M highorder elements are used to interpolate between these nodes.
Within each element, the boundary geometry and the ﬁeld variables are discretized using
polynomial shape functions. The integrals on the boundary are converted into a sum on
the elements, each one being calculated on the reference element. The matrices are built
with the numerical computation of the integrals on the reference element. The linear sys
tems resulting from the two boundary integral equations (one for the pair (φ, ∂φ/∂n) and
one for the pair (∂φ/∂t, ∂
2
φ/∂t∂n)) are full and non symmetric. Assembling the matrix as
well as performing the integrations accurately are time consuming tasks. They are done
only once at each time step, since the same matrix is used for both systems. Solving the
linear system is another time consuming task. Even with the GMRES algorithm with pre
conditioning, the computational complexity is O(N
2
), which is the same as the complexity
of the assembling phase. The introduction of the fast multipole algorithm reduces consid
erably the complexity of the problem. The matrix is no longer built. Far away nodes are
placed in groups, so less time is spent in numerical integrations and memory requirements
are reduced. The hierarchical structure involved in the algorithm gives automatically the
distance criteria for adaptive integrations.
Grilli et al. [GVW02] used the earlier version of the code to study tsunami generation
by underwater landslides. They included the bottom motion due to the landslide. For
the comparisons shown below, we only used the passive approach: we did not include the
dynamics of the bottom motion.
1.2.9 Comparisons and discussion
The passive generation approach is followed for the numerical comparisons between the
three models: (i) linear equations, (ii) NSW equations and (iii) fully nonlinear equations.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
52 Tsunami generation
As shown in Section 1.2.2 (one can have a look at [DDK06]), this generation process gives
the largest transientwave amplitudes for a given permanent deformation of the seaﬂoor.
Therefore it is in some sense a worst case scenario.
The small dimensionless numbers ε and µ
2
introduced in (1.41) represent the magnitude
of the nonlinear terms and dispersive terms in the governing equations, respectively. Hence,
the relative importance of the nonlinear and the dispersive eﬀects is given by the parameter
S =
nonlinear terms
dispersive terms
=
ε
µ
2
=
aλ
2
d
3
, (1.80)
which is called the Stokes (or Ursell) number [Urs53].
4
An important assumption in the
derivation of the Boussinesq system (1.73)–(1.75) is that the Ursell number is O(1). Here,
the symbol O() is used informally in the way that is common in the construction and
formal analysis of model equations for physical phenomena. We are concerned with the
limits ε →0 and µ →0. Thus, S = O(1) means that, as ε →0 and µ →0, S takes values
that are neither very large nor very small. We emphasize here that the Ursell number
does not convey any information by itself about the separate negligibility of nonlinear and
frequency dispersion eﬀects. Another important aspect of models is the time scale of their
validity. In the NSW equations, terms of order O(ε
2
) and O(µ
2
) have been neglected.
Therefore one expects these terms to make an orderone relative contribution on a time
scale of order min(ε
−2
, µ
−2
).
All the ﬁgures shown below are twodimensional plots for convenience but we recall that
all computations for the three models are threedimensional. Figure 1.18 shows proﬁles
of the freesurface elevation along the main direction of propagation (y−axis) of transient
waves generated by a permanent seaﬂoor deformation corresponding to the parameters
given in Table 1.5. This deformation, which has been plotted in Figure 1.15, has been
translated to the free surface. The water depth is 100 m. The small dimensionless numbers
are roughly ε = 5 10
−4
and µ = 10
−2
, with a corresponding Ursell number equal to 5.
One can see that the front system splits in two and propagates in both directions, with
a leading wave of depression to the left and a leading wave of elevation to the right, in
qualitative agreement with the satellite and tide gauge measurements of the 2004 Sumatra
event. When tsunamis are generated along subduction zones, they usually split in two; one
moves quickly inland while the second heads toward the open ocean. The three models are
4
One ﬁnds sometimes in the literature a subtle diﬀerence between the Stokes and Ursell numbers. Both
involve a wave amplitude multiplied by the square of a wavelength divided by the cube of a water depth.
The Stokes number is deﬁned speciﬁcally for the excitation of a closed basin while the Ursell number
is used in a more general context to describe the evolution of a long wave system. Therefore only the
characteristic length is diﬀerent. For the Stokes number the length is the usual wavelength λ related to the
frequency ω by λ ≈ 2π
√
gd/ω. In the Ursell number, the length refers to the local wave shape independent
of the exciting conditions.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.2 Comparison of tsunami generation models 53
Figure 1.18: Comparisons of the freesurface elevation at x = 0 resulting from the integration
of the linear equations ( ), NSW equations (−−) and nonlinear equations (−) at diﬀerent times
of the propagation of transient waves generated by an earthquake (t = 0 s, t = 95 s, t = 143 s,
t = 191 s). The parameters for the earthquake are those given in Table 1.5. The water depth is
h = 100 m. One has the following estimates: ε = 5 10
−4
, µ
2
= 10
−4
and consequently S = 5.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
54 Tsunami generation
Figure 1.19: Comparisons of the freesurface elevation at x = 0 resulting from the integration
of the linear equations ( ), NSW equations (−−) and nonlinear equations (−) at diﬀerent times
of the propagation of transient waves generated by an earthquake (t = 52 s, t = 104 s, t = 157
s). The parameters for the earthquake are those given in Table 1.5. The water depth is h = 500
m. One has the following estimates: ε = 10
−4
, µ
2
= 2.5 10
−3
and consequently S = 0.04.
almost undistinguishable at all times: the waves propagate with the same speed and the
same proﬁle. Nonlinear eﬀects and dispersive eﬀects are clearly negligible during the ﬁrst
moments of transient waves generated by a moving bottom, at least for these particular
choices of ε and µ.
Let us now decrease the Ursell number by increasing the water depth. Figure 1.19
illustrates the evolution of transient water waves computed with the three models for the
same parameters as those of Figure 1.18, except for the water depth now equal to 500
m. The small dimensionless numbers are roughly ε = 10
−4
and µ = 5 10
−2
, with a
corresponding Ursell number equal to 0.04. The linear and nonlinear proﬁles cannot be
distinguished within graphical accuracy. Only the NSW proﬁle is slightly diﬀerent.
Let us introduce several sensors (tide gauges) at selected locations which are represen
tative of the initial deformation of the free surface (see Figure 1.20). One can study the
evolution of the surface elevation during the generation time at each gauge. Figure 1.21
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.2 Comparison of tsunami generation models 55
Figure 1.20: Top view of the initial free surface deformation showing the location of six selected
gauges, with the following coordinates (in km): (1) 0,0 ; (2) 0,3 ; (3) 0,−3 ; (4) 10,5; (5) −2,5 ; (6)
1,10. The lower oval area represents the initial subsidence while the upper oval area represents
the initial uplift.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
56 Tsunami generation
0 20 40 60 80 100 120 140 160
−0.06
−0.04
−0.02
0
0.02
0.04
0.06
t, s
z
,
m
0 20 40 60 80 100 120 140 160
−0.05
0
0.05
0.1
0.15
t, s
z
,
m
0 20 40 60 80 100 120 140 160
−0.05
0
0.05
t, s
z
,
m
0 20 40 60 80 100 120 140 160
−0.04
−0.02
0
0.02
0.04
t, s
z
,
m
0 20 40 60 80 100 120 140 160
−0.05
0
0.05
t, s
z
,
m
0 20 40 60 80 100 120 140 160
−0.1
−0.05
0
0.05
0.1
t, s
z
,
m
linear
NSWE
Tide gauge 1
Tide gauge 3
Tide gauge 2
Tide gauge 4
Tide gauge 5
Tide gauge 6
Figure 1.21: Transient waves generated by an underwater earthquake. Comparisons of the
freesurface elevation as a function of time at the selected gauges shown in Figure 1.20: −, linear
model ; −− nonlinear shallow water model. The time t is expressed in seconds. The physical
parameters are those of Figure 1.19. Since the fully nonlinear results cannot be distinguished
from the linear ones, they are not shown.
shows freesurface elevations corresponding to the linear and nonlinear shallow water mod
els. They are plotted on the same graph for comparison purposes. Again there is a slight
diﬀerence between the linear and the NSW models, but dispersion eﬀects are still small.
Let us decrease the Ursell number even further by increasing the water depth. Figures
1.22 and 1.23 illustrate the evolution of transient water waves computed with the three
models for the same parameters as those of Figure 1.18, except for the water depth now
equal to 1 km. The small dimensionless numbers are roughly ε = 5 10
−5
and µ =
0.1, with a corresponding Ursell number equal to 0.005. On one hand, linear and fully
nonlinear models are essentially undistinguishable at all times: the waves propagate with
the same speed and the same proﬁle. Nonlinear eﬀects are clearly negligible during the ﬁrst
moments of transient waves generated by a moving bottom, at least in this context. On the
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.2 Comparison of tsunami generation models 57
other hand, the numerical solution obtained with the NSW model gives slightly diﬀerent
results. Waves computed with this model do not propagate with the same speed and have
diﬀerent amplitudes compared to those obtained with the linear and fully nonlinear models.
Dispersive eﬀects come into the picture essentially because the waves are shorter compared
to the water depth. As shown in the previous examples, dispersive eﬀects do not play a
role for long enough waves.
Figure 1.24 shows the transient waves at the gauges selected in Figure 1.20. One can
see that the elevations obtained with the linear and fully nonlinear models are very close
within graphical accuracy. On the contrary, the nonlinear shallow water model leads to a
higher speed and the diﬀerence is obvious for the points away from the generation zone.
These results show that one cannot neglect the dispersive eﬀects any longer. The NSW
equations, which contain no dispersive eﬀects, lead to diﬀerent speed and amplitudes.
Moreover, the oscillatory behaviour just behind the two front waves is no longer present.
This oscillatory behaviour has been observed for the water waves computed with the linear
and fully nonlinear models and is due to the presence of frequency dispersion. So, one
should replace the NSW equations with Boussinesq models which combine the two funda
mentals eﬀects of nonlinearity and dispersion. Wei et al. [WKGS95] provided comparisons
for twodimensional waves resulting from the integration of a Boussinesq model and the
twodimensional version of the FNPF model described above. In fact they used a fully
nonlinear variant of the Boussinesq model, which predicts wave heights, phase speeds and
particle kinematics more accurately than the standard weakly nonlinear approximation
ﬁrst derived by Peregrine [Per67] and improved by Nwogu’s modiﬁed Boussinesq model
[Nwo93]. We refer to the review [Kir03] on Boussinesq models and their applications for a
complete description of modern Boussinesq theory.
From a physical point of view, we emphasize that the wavelength of the tsunami waves
is directly related to the mechanism of generation and to the dimensions of the source
event. And so is the dimensionless number µ which determines the importance of the
dispersive eﬀects. In general it will remain small.
Adapting the discussion by Bona et al. [BCL05], one can expect the solutions to
the long wave models to be good approximations of the solutions to the full waterwave
equations on a time scale of the order min(ε
−1
, µ
−2
) and also the neglected eﬀects to make
an orderone relative contribution on a time scale of order min(ε
−2
, µ
−4
, ε
−1
µ
−2
). Even
though we have not computed precisely the constant in front of these estimates, the results
shown in this study are in agreement with these estimates. Considering the 2004 Boxing
Day tsunami, it is clear that dispersive and nonlinear eﬀects did not have suﬃcient time to
develop during the ﬁrst hours due to the extreme smallness of ε and µ
2
, except of course
when the tsunami waves approached the coast.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
58 Tsunami generation
Figure 1.22: Comparisons of the freesurface elevation at x = 0 resulting from the integration
of the linear equations (− −), NSW equations (−−) and FNPF equations (−) at diﬀerent times
of the propagation of transient waves generated by an earthquake (t = 50 s, t = 100 s). The
parameters for the earthquake are those given in Table 1.5. The water depth is 1 km. One has
the following estimates: ε = 5 10
−5
, µ
2
= 10
−2
and consequently S = 0.005.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.2 Comparison of tsunami generation models 59
Figure 1.23: Same as Figure 1.22 for later times (t = 150 s, t = 200 s).
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
60 Tsunami generation
0 20 40 60 80 100 120 140 160
−0.06
−0.04
−0.02
0
0.02
0.04
0.06
t, s
z
,
m
0 20 40 60 80 100 120 140 160
−0.05
0
0.05
0.1
0.15
t, s
z
,
m
0 20 40 60 80 100 120 140 160
−0.05
0
0.05
t, s
z
,
m
0 20 40 60 80 100 120 140 160
−0.04
−0.02
0
0.02
0.04
t, s
z
,
m
0 20 40 60 80 100 120 140 160
−0.05
0
0.05
t, s
z
,
m
0 20 40 60 80 100 120 140 160
−0.1
−0.05
0
0.05
0.1
t, s
z
,
m
Tide gauge 1 Tide gauge 2
Tide gauge 3
Tide gauge 4
Tide gauge 5
Tide gauge 6
Figure 1.24: Transient waves generated by an underwater earthquake. The physical parameters
are those of Figures 1.22 and 1.23. Comparisons of the freesurface elevation as a function of time
at the selected gauges shown in Figure 1.20: −, linear model ; −− nonlinear shallow water model.
The time t is expressed in seconds. The FNPF results cannot be distinguished from the linear
results.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.2 Comparison of tsunami generation models 61
0 20 40 60 80 100 120 140 160 180 200
−0.06
−0.04
−0.02
0
0.02
0.04
0.06
0.08
time (s)
z
,
m
Active generation
Passive generation
0 20 40 60 80 100 120 140 160 180 200
−0.1
−0.05
0
0.05
0.1
0.15
0.2
0.25
time (s)
z
,
m
0 20 40 60 80 100 120 140 160 180 200
−0.06
−0.04
−0.02
0
0.02
0.04
0.06
0.08
time (s)
z
,
m
0 20 40 60 80 100 120 140 160 180 200
−0.03
−0.02
−0.01
0
0.01
0.02
0.03
0.04
time (s)
z
,
m
Tide gauge 1
Tide gauge 3 Tide gauge 4
Tide gauge at (2,3)
Figure 1.25: Transient waves generated by an underwater earthquake. The computations are
based on linear wave theory. Comparisons of the freesurface elevation as a function of time at
selected gauges for active and passive generation processes. The time t is expressed in seconds.
The physical parameters are those of Figure 1.19. In particular, the water depth is h = 500 m.
Let us conclude this section with a discussion on the generation methods, which extends
the results given in [DDK06]
5
. We show the major diﬀerences between the classical passive
approach and the active approach of wave generation by a moving bottom. Recall that the
classical approach consists in translating the sea bed deformation to the free surface and
letting it propagate. Results are presented for waves computed with the linear model.
Figure 1.25 shows the waves measured at several artiﬁcial gauges. The parameters
are those of Table 1.5, and the water depth is h = 500 m. The solid line represents the
solution with an instantaneous bottom deformation while the dashed line represents the
passive wave generation scenario. Both scenarios give roughly the same wave proﬁles. Let
us now consider a slightly diﬀerent set of parameters: the only diﬀerence is the water depth
which is now h = 1 km. As shown in Figure 1.26, the two generation models diﬀer. The
passive mechanism gives higher wave amplitudes.
Let us quantify this diﬀerence by considering the relative diﬀerence between the two
5
In Figures 1 and 2 of [DDK06], a mistake was introduced in the time scale. All times must be multiplied
by a factor
√
1000.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
62 Tsunami generation
0 20 40 60 80 100 120 140 160 180 200
−0.06
−0.04
−0.02
0
0.02
0.04
0.06
0.08
time (s)
z
,
m
Active generation
Passive generation
0 20 40 60 80 100 120 140 160 180 200
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
0.2
0.25
time (s)
z
,
m
0 20 40 60 80 100 120 140 160 180 200
−0.06
−0.04
−0.02
0
0.02
0.04
0.06
0.08
time (s)
z
,
m
0 20 40 60 80 100 120 140 160 180 200
−0.03
−0.02
−0.01
0
0.01
0.02
0.03
time (s)
z
,
m
Tide gauge 1
Tide gauge 3
Tide gauge 4
Tide gauge at (2, 3)
Figure 1.26: Same as Figure 1.25, except for the water depth, which is equal to 1 km.
mechanisms deﬁned by
r(x, y, t) =
[η
active
(x, y, t) −η
passive
(x, y, t)[
[[η
active
[[
∞
.
Intuitively this quantity represents the deviation of the passive solution from the active
one with a moving bottom in units of the maximum amplitude of η
active
(x, y, t).
Results are presented on Figures (1.27) and (1.28). The diﬀerences can be easily ex
plained by looking at the analytical formulas (1.48) and (1.59) of Section 1.2.2. These
diﬀerences, which can be crucial for accurate tsunami modelling, are twofold.
First of all, the wave amplitudes obtained with the instantly moving bottom are lower
than those generated by the passive approach (this statement follows from the inequality
cosh mh ≥ 1). The numerical experiments show that this diﬀerence is about 6% in the
ﬁrst case and 20% in the second case.
The second feature is more subtle. The water column has an eﬀect of a lowpass ﬁlter.
In other words, if the initial deformation contains high frequencies, they will be attenuated
in the moving bottom solution because of the presence of the hyperbolic cosine cosh(mh)
in the denominator which grows exponentially with m. Incidently, in the framework of the
NSW equations, there is no diﬀerence between the passive and the active approach for an
instantaneous seabed deformation [Tuc79, Tuc74].
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.2 Comparison of tsunami generation models 63
0 50 100 150 200
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
time (s)

η
a
c
t
iv
e
−
η
p
a
s
s
iv
e

/

η
a
c
t
iv
e

0 50 100 150 200
0
0.01
0.02
0.03
0.04
0.05
0.06
time (s)

η
a
c
t
iv
e
−
η
p
a
s
s
iv
e

/

η
a
c
t
iv
e

0 50 100 150 200
0
0.01
0.02
0.03
0.04
0.05
0.06
time (s)

η
a
c
t
iv
e
−
η
p
a
s
s
iv
e

/

η
a
c
t
iv
e

0 50 100 150 200
0
0.01
0.02
0.03
0.04
0.05
0.06
time (s)

η
a
c
t
iv
e
−
η
p
a
s
s
iv
e

/

η
a
c
t
iv
e

Tide gauge 1
Tide gauge 3 Tide gauge 4
Tide gauge at (2,3)
Figure 1.27: Relative diﬀerence between the two solutions shown in Figure 1.25. The time t is
expressed in seconds.
0 50 100 150 200
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
time (s)

η
a
c
t
iv
e
−
η
p
a
s
s
iv
e

/

η
a
c
t
iv
e

0 50 100 150 200
0
0.05
0.1
0.15
0.2
0.25
time (s)

η
a
c
t
iv
e
−
η
p
a
s
s
iv
e

/

η
a
c
t
iv
e

0 50 100 150 200
0
0.05
0.1
0.15
0.2
time (s)

η
a
c
t
iv
e
−
η
p
a
s
s
iv
e

/

η
a
c
t
iv
e

0 50 100 150 200
0
0.05
0.1
0.15
0.2
time (s)

η
a
c
t
iv
e
−
η
p
a
s
s
iv
e

/

η
a
c
t
iv
e

Tide gauge 1
Tide gauge at
(2,3)
Tide gauge 3
Tide gauge 4
Figure 1.28: Relative diﬀerence between the two solutions shown in Figure 1.26.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
64 Tsunami generation
If we prescribe a more realistic bottom motion as in [DD07c] for example, the results
will depend on the characteristic time of the seabed deformation. When the characteristic
time of the bottom motion decreases, the linearized solution tends to the instantaneous
wave generation scenario. So, in the framework of linear water wave equations, one cannot
exceed the passive generation amplitude with an active process. However, during slow
events, Todorovska and Trifunac [TT01] have shown that ampliﬁcation of one order of
magnitude may occur when the sea ﬂoor uplift spreads with velocity similar to the long
wave tsunami velocity.
1.2.10 Conclusions
Comparisons between linear and nonlinear models for tsunami generation by an under
water earthquake have been presented. There are two main conclusions that are of great
importance for modelling the ﬁrst instants of a tsunami and for providing an eﬃcient initial
condition to propagation models. To begin with, a very good agreement is observed from
the superposition of plots of wave proﬁles computed with the linear and fully nonlinear
models. Secondly, the nonlinear shallow water model was not suﬃcient to model some of
the waves generated by a moving bottom because of the presence of frequency dispersion.
However classical tsunami waves are much longer, compared to the water depth, than the
waves considered in the present work, so that the NSW model is also suﬃcient to describe
tsunami generation by a moving bottom. Comparisons between the NSW equations and
the FNPF equations for modeling tsunami runup are left for future work. Another as
pect which deserves attention is the consideration of Earth rotation and the derivation of
Boussinesq models in spherical coordinates.
1.3 Tsunami generation by dynamic displacement of sea
bed due to dipslip faulting
1.3.1 Introduction
The accuracy of the computation of the whole life of a tsunami, from generation to
inundation, obviously depends on the construction of the initial condition. Moreover, the
error made on the initial condition cannot be corrected by the numerical method used to
propagate the tsunami. This is why the process of tsunami generation must be modelled as
accurately as possible. Even though the constraint of being able to predict tsunami arrival
time, height and location as fast as possible must deﬁnitely be taken into account (in other
words, a tradeoﬀ must be found between the precision and the speed of computation of
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.3 Tsunami generation by dynamic displacement of sea bed 65
the initial condition), we believe that so far the scientiﬁc community has not payed enough
attention to the crucial subject of tsunami generation.
After the pioneer work of Kajiura [Kaj70] it has become a common practice in the
tsunami community to translate the static sea bed deformation generated by an underwater
earthquake onto the free surface and let it propagate. We will refer to this method as passive
approach. The validity of this technique was already discussed in [OTM01, DDK06]. Three
dimensional analytical expressions derived from Volterra’s formula applied to the general
study of dislocations [MS71, Oka85] are used to construct the static initial deformation.
Similar analytical expressions for twodimensional problems were also derived by Freund &
Barnett [FB76], who used the theory of analytic functions of a complex variable. Obviously,
the popularity of these analytic solutions can be explained by their relatively simple explicit
form. Thus, their computation is easy and inexpensive. A nice feature of the solution
of Freund & Barnett is that nonuniform slip distributions can be easily considered. In
particular, slip distributions which remove the singular behavior of the internal stresses
at the ends of the slip zone can be dealt with, simply by imposing the socalled smooth
closure condition on the slip: the slip is zero at the ends.
When simplifying hypotheses such as homogeneity or isotropy are removed, analytical
solutions are no longer available and the governing equations must be solved numerically.
Static deformations caused by slip along a fault have been extensively simulated by Mas
terlark [Mas03], who used several dislocation models based on the ﬁniteelement method
(FEM) to estimate the importance of diﬀerent physical hypotheses. Anisotropy and hetero
geneity turned out to be the most important factors in this type of modelling. Megna et al.
[MBS05] also used the FEM to compare numerical results with analytical solutions. How
ever neither in [Mas03] nor in [MBS05] were the dynamical aspects and the coupling with
hydrodynamics considered. Moreover the consequences for the resulting tsunami waves
were not pointed out.
When one uses as initial condition a static seismic source together with the translation
of the sea bed deformation onto the free surface, one neglects the rupture velocity and
the rise time. Several studies have already been performed to understand wave formation
due to diﬀerent prescribed bottom motions, either by introducing some type of rise time,
or by introducing some type of rupture velocity. For example, Todorovska & Trifunac
[TT01] studied the generation of waves by a slowly spreading uplift of the bottom. The
studies of Hammack [Ham73] and Dutykh & Dias [DD07c] take into account the rise
time. Hammack [Ham73] generated waves experimentally by raising or lowering a box at
one end of a channel, and considered various laws for the rise or the fall of the box. In
their review paper, Dutykh & Dias [DD07c] generated waves theoretically by multiplying
the static deformations caused by slip along a fault by various time laws: instantaneous,
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
66 Tsunami generation
exponential, trigonometric, linear. Haskell [Has69] was one of the ﬁrst to take into account
the rupture velocity. In fact he considered both rise time and rupture velocity. Let us
consider the source shown in Figure 1.29. The two horizontal coordinates x and y, and the
vertical coordinate z are denoted by x = (x, y, z).
x
y
z
O
W
−d −
L
2
L
2
δ
Figure 1.29: Geometry of the source model. The fault has width W, length L, depth d and dip
angle δ.
Let
b(x, t) denote the fault displacement function and
b
0
(x) the ﬁnal displacement. The
following form for
b(x, t) was considered by Haskell:
b(x, t) =
_
¸
_
¸
_
0 t −ζ/V < 0
(
b
0
/T)(t −ζ/V ) 0 < t −ζ/V < T
b
0
t −ζ/V > T
(1.81)
where T is the rise time and V the rupture velocity. The coordinate ζ is a coordinate
along the fault. Equation (1.81) implies that at t = 0 a fracture front is established
instantaneously over a width W of the y−axis at depth d. The front propagates unilaterally
at constant velocity V over a length Lcos δ of the x−axis. At any ﬁxed point on the fault
plane the relative displacement increases at constant velocity from 0 at t = ζ/V to a
constant ﬁnal value
b
0
at t = T + ζ/V . More recently, Okumura & Kawata [OK07] used
Haskell’s approach to investigate the eﬀects of rise time and rupture velocity on tsunami
generation. They considered two cases of sea bottom motion: (i) with only rise time and (ii)
with both rise time and rupture velocity. They found that the eﬀects of rupture velocity are
much smaller than those of rise time when the rise time is assumed to be long (over 10 min).
Ohmachi et al. [OTM01] also considered rise time and rupture velocity but unfortunately
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.3 Tsunami generation by dynamic displacement of sea bed 67
the dynamics is not clearly explained in their paper. Apparently they did not solve the
elastodynamic equations with the secondorder time derivative (see next section). Another
interesting attempt to understand dynamical eﬀects is that of Madariaga [Mad03]. He
considers a dipslip dislocation propagating in a halfspace. He solves the elastodynamic
equations by using the double Laplace transform. The solution is elegant but it is relatively
complex. Unfortunately Madariaga does not provide any plots of the deformation of the
free surface and does not consider the coupling with the water layer. The present study
can be considered as an attempt to understand the coupling between seismic faulting and
hydrodynamics by integrating numerically the timedependent elasticity equations as well
as the timedependent ﬂuid equations. The authors have already adressed the problem of
tsunami generation in [DDK06, DD07c]. The main feature of the present study is the use
of a more realistic earthquake source model.
The work is organised as follows. In Section 1.3.2 we brieﬂy describe the mathematical
models, both for solid and ﬂuid motions, which are used in the present study. Section 1.3.3
contains details on the numerical methods used to solve the governing equations. The
numerical method for the solid motion is validated in Section 1.3.4. The last Section 1.3.5,
provides a comparison between the traditional approach to tsunami generation (in which
the static sea bed deformation is translated onto the free surface) and the more realistic
approach of dynamic generation (in which the wavetrain is generated by the motion of the
bottom). We reveal numerically that the dynamical aspects of tsunami generation can for
example lead to a leading depression wave when one expects a leading elevation wave.
1.3.2 Mathematical models
Even though the numerical results shown in this study are for twodimensional con
ﬁgurations, the modeling is performed for threedimensional problems. The horizontal
coordinates are denoted by x and y, while the vertical coordinate is denoted by z. The
displacements are denoted by u
x
, u
y
and u
z
. We use diﬀerent origins along the vertical axis
for the solid and ﬂuid motions. In the earth domain, z = 0 denotes the sea bed at rest
(assumed to be ﬂat). In the ﬂuid domain, z = 0 denotes the sea surface at rest.
1.3.2.1 Dynamic fault model
We assume that the fault is inside a geological viscous medium. Earth’s crust is assumed
to be a viscoelastic material of density ρ. We choose the KelvinVoigt viscosity model
[Qui97] which consists in using complex elastic coeﬃcients (with negative imaginary parts
in order to dissipate wave energy). For isotropic media it means that the Lam´e coeﬃcients
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
68 Tsunami generation
have a nonpositive imaginary part:
λ
∗
= λ
r
−iλ
i
, µ
∗
= µ
r
−iµ
i
,
where λ
r
, µ
r
> 0 and λ
i
, µ
i
≥ 0. The classical elasticity equations are obtained by choosing
λ
i
≡ 0 and µ
i
≡ 0. Notice that on the timescales relevant to our problem, elasticity is
suﬃcient and the assumption of a KelvinVoigt viscous material is unnecessary. But we
keep it for the sake of completeness.
Let c
P
and c
S
be the classical velocities for the propagation of P and S waves in a
medium of density ρ:
c
P
=
¸
λ
r
+ 2µ
r
ρ
, c
S
=
_
µ
r
ρ
.
Complex Lam´e coeﬃcients yield complex velocities for wave propagation,
c
∗
P
= c
P
¸
1 +
i
Q
P
, c
∗
S
= c
S
¸
1 +
i
Q
S
,
where the coeﬃcients Q
P
and Q
S
are deﬁned as follows:
Q
P
= −
λ
r
+ 2µ
r
λ
i
+ 2µ
i
, Q
S
= −
µ
r
µ
i
.
The factors Q
P
and Q
S
measure the viscosity of the geological medium. In this study we
restrict our attention to the weakly viscous case. Mathematically it means that 1/Q
P
≪1
and 1/Q
S
≪1.
Let σ represent the stress tensor. The displacement ﬁeld u(x, y, z, t) = (u
x
, u
y
, u
z
)
satisﬁes the classical elastodynamic equations from continuum mechanics:
∇ σ = ρ
∂
2
u
∂t
2
. (1.82)
It is common in seismology to assume that the stress tensor σ is determined by Hooke’s
law through the strain tensor ε =
1
2
_
∇u +∇
t
u
_
. Therefore
σ = λ
∗
(∇ u)I + 2µ
∗
ε. (1.83)
Thus, we come to the following linear viscoelastodynamic problem
6
:
∇
_
λ
∗
(∇ u)I + µ
∗
(∇u +∇
t
u)
_
= ρ
∂
2
u
∂t
2
. (1.84)
6
We use the preﬁx “visco” due to the presence of the imaginary part in the Lam´e coeﬃcients, which
is responsible for small wave damping.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.3 Tsunami generation by dynamic displacement of sea bed 69
Recall that the mechanical characteristics ρ, λ
∗
and µ
∗
can possibly depend on the spa
tial coordinates (x, y, z). However we will assume that they are constant in the numerical
applications.
The fault is modeled as a dislocation inside a viscoelastic material. This type of model is
widely used for the interpretation of seismic motion. A dislocation is considered as a surface
(in threedimensional problems) or a line (in twodimensional problems) in a continuous
medium where the displacement ﬁeld is discontinuous. The displacement vector is increased
by the amount of the Burgers vector
b along any contour C enclosing the dislocation surface
(or line), i.e.
_
C
du =
b. (1.85)
We let a dislocation run at speed V along a fault inclined at an angle δ with respect to the
horizontal. Rupture starts at position x = 0 and z = −d (it is supposed to be inﬁnitely
long in the transverse y−direction) and propagates with constant rupture speed V for a
ﬁnite time L/V in the direction δ stopping at a distance L. Let ζ be a coordinate along
the dislocation line. On the fault located in the interval 0 < ζ < L slip is assumed to be
constant. The rise time is assumed to be 0.
1.3.2.2 Fluid layer model
Since the main purpose is to model tsunami generation processes and since tsunamis
are long waves, it is natural to choose the nonlinear shallow water equations (NSWE) as
hydrodynamic model. These equations are widely used in tsunami modelling, especially in
codes for operational use [TS98, SB06]. The validity of the NSWE model and the question
of the importance of dispersive eﬀects have already been addressed by the authors in the
previous Section (one can have a look at our publication [KDD07]).
Let η denote the freesurface elevation with respect to the still water level z = 0, v =
(v
x
, v
y
) the horizontal velocity vector, g the acceleration due to gravity and z = −h(x, y, t)
the bathymetry. The NSWE in dimensional form read
∂η
∂t
+∇ ((h + η)v) = −
∂h
∂t
,
∂v
∂t
+
1
2
∇[v[
2
+ g∇η = 0.
The eﬀect of the moving bottom appears in the source term −∂h/∂t in the ﬁrst equation.
The unknowns η and v are functions of time and of the horizontal coordinates x and
y. Since the NSWE are essentially obtained from depthaveraging the Euler equations, the
dependence on the vertical coordinate z disappears from the equations.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
70 Tsunami generation
The coupling between the earth and ﬂuid models is made through the function h(x, y, t)
which describes the moving sea bottom bathymetry.
1.3.3 Numerical methods
In the present study we made two natural choices. The solid mechanics equations
of the model are solved using the FEM with fully implicit time integration, while for
the hydrodynamic part we take advantage of the hyperbolic structure of the governing
equations and use a solver based on the ﬁnitevolume (FV) scheme (see for example [BQ06,
KDD07]).
1.3.3.1 Discretization of the viscoelastodynamic equations
In order to apply the FEM one ﬁrst rewrites the governing equation (1.84) in variational
form. The timederivative operator is discretized through a classical secondorder ﬁnite
diﬀerence scheme. The method we use is fully implicit and has the advantage of being
free of any CourantFriedrichsLewytype condition. In such problems implicit schemes
become advantageous since the velocity of propagation of seismic waves is of the order of
3 to 4 km/s. We apply the P2 ﬁniteelement discretization procedure. For the numerical
computations, the freely available code FreeFem++ [HPHO] is used.
Let us say a few words about the boundary conditions and the treatment of the dislo
cation in the program. Concerning the boundary conditions, we assume that the sea bed
is a free surface, that is σ n =
0 at z = 0. The other boundaries are assumed to be ﬁxed
or, in other words, Dirichlet type boundary conditions u =
0 are applied. The authors are
aware of the reﬂective properties of this type of boundary conditions. In order to avoid
the reﬂection of seismic waves along the boundaries during the simulation time, we take
a computational domain which is suﬃciently large. This approach is not computationally
expensive since we use adaptive mesh algorithms [HPHO] and in the regions far away from
the fault, element sizes are considerably bigger than in the fault vicinity.
Next we discuss the implementation of the dislocation surface. Across the fault, the
displacement ﬁeld is discontinuous and satisﬁes the following relation:
u
+
(x, t) −u
−
(x, t) =
b(x, t), (1.86)
where the signs ± denote the upper and lower boundary of the dislocation surface, respec
tively. The propagation of Burger’s vector along the fault is given by
b(x, t) =
b
0
H(t −ζ/V ), (1.87)
where V is the rupture velocity, H the Heaviside unit step function and ζ a coordinate
along the dislocation line.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.3 Tsunami generation by dynamic displacement of sea bed 71
Remark 1 Due to the presence of huge hydrostatic pressures in the crust, the two sides
of the fault cannot detach physically. In any case this situation does not occur in nature.
Mathematically it means that the vector
b is tangent to the dislocation surface at each point.
1.3.3.2 Finitevolume scheme
In this study we adopt the characteristic ﬂux (CF) approach proposed in [GKC96,
GKC01] and described in details in Appendix A. This approach is a particular scheme of
FV methods. There are other schemes that have been proposed for the NSWE (see [BQ06]
for example).
We chose this method for the following reasons. First of all, the CF scheme is easy to
implement. Then it is not based on the solution to the Riemann problem and consequently
we do not impose a onedimensional wave structure which is no longer relevant in two
dimensions. It is a multidimensional scheme and we do not need to split operators in order
to treat separately each dimension. And ﬁnally, the characteristic ﬂux is more versatile
than Roe’s scheme in the sense that it does not rely on an algebraic property of the ﬂux.
In this section we consider an application of the characteristic ﬂuxes scheme to NSWE.
In this case one can explicitly ﬁnd eigenvalues, as well as left and right eigenvectors.
This property together with the use of an approximate Riemann solver makes all the
computations very eﬃcient.
Let v = (v
x
, v
y
). The ﬂux of the NSWE system is given by
F(V) =
_
_
_
(h + η)v
x
(h + η)v
y
1
2
[v[
2
+ gη 0
0
1
2
[v[
2
+ gη
_
_
_
, (1.88)
where V = (η, u, v) is the vector of conservative variables
7
. The ﬂux F(V) projected on
the normal n becomes
F n =
_
(h + η)(v n)
_
1
2
[v[
2
+ gη
_
n
_
.
It follows that the Jacobian matrix is given by
A
n
:=
∂
_
F(V) n
_
∂V
=
_
v n (h + η)n
gn v ⊗n
_
. (1.89)
It has three distinct eigenvalues
λ
1,3
= v n ±c
s
, λ
2
= 0,
7
In the present study the physical variables coincide with the conservative ones.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
72 Tsunami generation
where c
s
=
_
g(h + η) is the velocity of long gravity waves.
The right eigenvectors constitute the columns of the matrix R:
R =
_
_
_
−c
s
(c
2
s
/g)v t c
s
gn
x
n
y
c
2
s
−v
y
(v n) gn
x
gn
y
−n
x
c
2
s
+ v
x
(v n) gn
y
_
_
_
, (1.90)
where t is the tangent vector. The matrix of left eigenvectors can also be computed
analytically through R
−1
.
Using these expressions it is straightforward to compute the sign matrix deﬁned as
sign(A
n
) := Rsign(Λ)R
−1
,
and the numerical scheme is thus completely deﬁned.
1.3.4 Validation of the numerical method
In this section we consider an analytic solution to the line dislocation problem in the
static case. Use is made of the wellknown result described for example by Freund &
Barnett [FB76] or Okada [Oka85]. In order to simplify the expressions, we only consider
the twodimensional case (in other words, the fault is inﬁnite in the y−direction). In fact
the most appropriate expression is that given by equation (24) in [Mad03]. We checked
that it is in full agreement with the limit of Okada’s solution as the width becomes inﬁnite.
The sketch of the domain is given in Figure 1.29. The fault has inﬁnite width (W →∞).
Its length is L, its depth d and its dip angle δ.
In the present work we only give the vertical displacement component u
z
along the free
surface, since it plays the most important role in tsunami formation. It can be expressed
as the diﬀerence between two contributions, that from a ﬁrst dislocation located at the
beginning of the fault and that from a second dislocation located at the end of the fault.
Let d
L
= d −Lsin δ. One has
u
z
= [
b
0
[
_
U
z
_
x
d
, δ
_
−U
z
_
x −Lcos δ
d
L
, δ
__
, (1.91)
where
U
z
_
x
d
, δ
_
=
1
π
_
sin δ arctan
x
d
−
d(d cos δ −xsin δ)
x
2
+ d
2
_
. (1.92)
For the validation of our numerical method we chose the most violent fault which
corresponds to the dip angle δ = π/2. The values of the other parameters are given in
Table 1.6. This problem was solved by FEM after neglecting the dynamic terms. The
results of the comparison with solution (1.92) are given in Figure 1.30. Good qualitative
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.3 Tsunami generation by dynamic displacement of sea bed 73
Figure 1.30: Comparison between analytical and numerical solutions for a static two
dimensional fault with a dip angle equal to π/2.
and quantitative agreement can be seen. Megna et al. [MBS05], who also considered static
displacement due to uniform slip across a normal fault, compared the twodimensional FEM
results with the analytical solution in the case of a normal fault. In their conclusion, they
state that it is for the vertical component of the surface displacement that the discrepancies
are the largest.
1.3.5 Results of the simulation
In this section, we use the set of physical parameters given in Table 1.6. The static
sea bed deformation obtained with the analytical solution is depicted in Figure 1.31. Note
that the only diﬀerence between Figures 1.30 and 1.31 is the value of the dip angle.
In order to illustrate the numerical computations we chose several test cases of ac
tive/passive tsunami generation. The passive generation approach was introduced in
[Kaj70]. It consists in translating the static sea bed deformation onto the free surface
and letting it propagate under gravity. On the other hand, the active approach uses the
bottom motion for wave generation. We proceed by computing the ﬁrst eight or ﬁfteen
seconds of the earthquake dynamics. Then the bottom conﬁguration is assumed to remain
frozen during the rest of the simulation. Concerning the dynamical aspects of rupture
propagation, we consider the Heavisidetype approach (1.87) where the dislocation prop
agates along the fault with rupture velocity V . One could use instead a dislocation for
which Burger’s vector
b
0
is also spacedependent. But the main goal of the present study
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
74 Tsunami generation
Figure 1.31: Static deformation due to the dislocation corresponding to the parameters given
in Table 1.6.
Parameter Value
Young modulus, E, GPa 9.5
Poisson ratio, ν 0.27
Damping coeﬃcient, λ
i
500
Damping coeﬃcient, µ
i
200
Fault depth, d, km 4
Fault length, L, km 2
Dip angle, δ,
◦
13
Burger’s vector length, [
b
0
[, m 10
Water depth (uniform), h
0
, m 400
Acceleration due to gravity, g, m/s
2
9.8
Table 1.6: Typical physical parameters used in the numerical computations. The water depth
as well as the spatial extent in the main direction of propagation were chosen so that dispersive
eﬀects can be neglected.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.3 Tsunami generation by dynamic displacement of sea bed 75
is to make an attempt to include the dynamic displacement of the sea bed. In the dynam
ical approach, we consider three cases: the limiting case where the rupture velocity V is
inﬁnite, a fast event with V = 2.5 km/s and a slower event with V = 1 km/s.
We show below the diﬀerences between the passive and the dynamic approaches. This
question has already been addressed by the authors [DDK06] in the framework of the
linearized potential ﬂow equations and of a simpliﬁed model for bottom deformation.
In the ﬁrst comparison we use a strong coupling between the dynamic displacement of
the sea bed and the ﬂuid layer equations and compare it with the passive approach, in
which the static solution shown in Figure 1.31 is translated onto the free surface as initial
condition. The rupture velocity V is assumed to be inﬁnite. Moreover the earthquake
dynamics is computed during the ﬁrst eight seconds. The free surface at the beginning of
the tsunami generation process is shown on Figure 1.32. Further steps of this process are
given in Figures 1.331.34. The reader may have the impression that the passive solution
does not evolve. In fact, the explanation lies in the presence of two diﬀerent time scales
in this problem. The fast time scale is provided by the earthquake (P− and S−waves)
and the slow one by water gravity waves. Since the active generation solution is directly
coupled to the bottom dynamics, it evolves with the fast time scale. It is interesting to
compare Figures 1.33 and 1.34. One can see that the active approach gives at the beginning
an amplitude which is almost twice larger but the amplitudes become comparable a few
seconds later.
The freesurface elevations are computed until the wave enters the purely propagation
stage. This corresponds to Figure 1.35. One notices that the resulting wave amplitude
and velocity are almost the same. Of course the waveform is diﬀerent. One can see as well
that the location of the elevation wave is the same, while the depression wave is slightly
shifted. It can be explained by the larger extent of the dynamic solution. Thus, we can
conclude from this ﬁrst comparison that if one is only interested in tsunami travel time or
even in rough inundation zone estimation, the passive approach can be used.
The second comparison focuses on the inﬂuence of the rupture velocity at two separate
times (Figures 1.36 and 1.37). The diﬀerences between the fast and the relatively slow
rupture velocities are small.
The most interesting comparison is the third one, which focuses on the duration of
the earthquake. Recall that our somewhat artiﬁcial deﬁnition of earthquake duration is
the time at which we stop the bottom motion. After that time, the sea bottom remains
frozen. Figure 1.38 shows the eﬀect of a longer earthquake. One sees that the shapes of the
wave train obtained with the dynamic analysis look more complicated than that obtained
by the passive analysis. In particular, the distinction between leading elevation wave and
leading depression wave is not as clear when using the dynamical analysis. It could be an
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
76 Tsunami generation
Figure 1.32: Water free surface at the beginning of the earthquake (t = 2s) according to two
approaches of tsunami generation: passive versus active (with inﬁnite rupture velocity).
Figure 1.33: Same as Figure 1.32 for times t = 4s and t = 6s.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.3 Tsunami generation by dynamic displacement of sea bed 77
Figure 1.34: Same as Figure 1.32 for times t = 7s and t = 10s.
Figure 1.35: Same as Figure 1.32 for times t = 150s and t = 250s. The wave is leaving the
generation zone (left plot) and starting to propagate (right plot).
Figure 1.36: Same as Figure 1.35 (left plot) for two rupture velocities: V = 1 km/s (left) and
V = 2.5 km/s (right).
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
78 Tsunami generation
Figure 1.37: Same as Figure 1.36 at time t = 250s.
Figure 1.38: Water free surface at t = 150s for a longer earthquake. The rupture motion stops
after 15s. The left plot compares the inﬁnite rupture velocity solution with the passive approach.
The right plot compares the case of a slow rupture (V = 1 km/s) with the passive case.
explanation for the discrepancies between modeled and recorded time series of water levels
at various locations along the California coast for the 1960 Chilean tsunami [BUST06].
1.3.6 Conclusions
An approach to model the dynamical character of sea bed deformations during an
underwater earthquake was presented. The governing elastodynamic equations were solved
by a ﬁniteelement method. The principal novelty of the present study is the coupling of
the resulting displacement ﬁeld with the hydrodynamic model.
Two methods for tsunami generation have been compared: static versus dynamic. The
computational results speak by themselves. One can say that the dynamic approach leads
to higher water levels in the nearfault area. These signiﬁcant diﬀerences only occur during
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
1.3 Tsunami generation by dynamic displacement of sea bed 79
the ﬁrst instants of the surface deformation and level oﬀ later on. However it was also
observed that the shape of the wave train can be altered by dynamical eﬀects. Consequently
the distinction between leading elevation wave and leading depression wave may not be as
clear as anticipated. Of course the present method is computationally more expensive but
there is an overall gain in accuracy. Not surprisingly more accurate tsunami computations
require ﬁner initial conditions such as those obtained by the active generation methodology
used in the present study.
In future work we intend to extend this modeling to three space dimensions since it is
evident that the twodimensional computations presented here have little interest beyond
academics.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
80 Tsunami generation
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
Chapter 2
Dissipative Boussinesq equations
We are usually convinced more easily by reasons we have found ourselves
than by those which have occurred to others.
Blaise Pascal (1623 – 1662)
Contents
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
2.2 Derivation of the Boussinesq equations . . . . . . . . . . . . . 85
2.2.1 Asymptotic expansion . . . . . . . . . . . . . . . . . . . . . . . . 89
2.3 Analysis of the linear dispersion relations . . . . . . . . . . . . 93
2.3.1 Linearized potential ﬂow equations . . . . . . . . . . . . . . . . . 93
2.3.2 Dissipative Boussinesq equations . . . . . . . . . . . . . . . . . . 95
2.3.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
2.4 Alternative version of the Boussinesq equations . . . . . . . . 99
2.4.1 Derivation of the equations . . . . . . . . . . . . . . . . . . . . . 100
2.5 Improvement of the linear dispersion relations . . . . . . . . . 104
2.6 Regularization of Boussinesq equations . . . . . . . . . . . . . 105
2.7 Bottom friction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
2.8 Spectral Fourier method . . . . . . . . . . . . . . . . . . . . . . 108
2.8.1 Validation of the numerical method . . . . . . . . . . . . . . . . 110
2.9 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . 113
2.9.1 Construction of the initial condition . . . . . . . . . . . . . . . . 113
2.9.2 Comparison between the dissipative models . . . . . . . . . . . . 115
81
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
82 Dissipative Boussinesq equations
2.10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
2.1 Introduction
Boussinesq equations are widely used in coastal and ocean engineering. One example
among others is tsunami wave modelling. These equations can also be used to model tidal
oscillations. Of course, these types of wave motion are perfectly described by the Navier
Stokes equations, but currently it is impossible to solve fully threedimensional (3D) models
in any signiﬁcant domain. Thus, approximate models such as the Boussinesq equations
must be used.
The years 1871 and 1872 were particularly important in the development of the Boussi
nesq equations. It is in 1871 that Valentin Joseph Boussinesq received the Poncelet prize
from the Academy of Sciences for his work. In the Volumes 72 and 73 of the “Comptes
Rendus Hebdomadaires des S´eances de l’Acad´emie des Sciences”, which cover respectively
the sixmonth periods January–June 1871 and July–December 1871, there are several con
tributions of Boussinesq. On June 19, 1871, Boussinesq presents the now famous note on
the solitary wave entitled “Th´eorie de l’intumescence liquide appel´ee onde solitaire ou de
translation, se propageant dans un canal rectangulaire” (72, pp. 755–759), which will be
extended later in the note entitled “Th´eorie g´en´erale des mouvements qui sont propag´es
dans un canal rectangulaire horizontal” (73, pp. 256–260). SaintVenant presents a couple
of notes of Boussinesq entitled “Sur le mouvement permanent vari´e de l’eau dans les tuyaux
de conduite et dans les canaux d´ecouverts” (73, pp. 34–38 and pp. 101–105). SaintVenant
himself publishes a couple of notes entitled “Th´eorie du mouvement non permanent des
eaux, avec application aux crues des rivi`eres et `a l’introduction des mar´ees dans leur lit”
(73, pp. 147–154 and pp. 237–240). All these notes deal with shallowwater theory. On
November 13, 1871, Boussinesq submits a paper entitled “Th´eorie des ondes et des remous
qui se propagent le long d’un canal rectangulaire horizontal, en communiquant au liquide
contenu dans ce canal des vitesses sensiblement pareilles de la surface au fond”, which
will be published in 1872 in the Journal de Math´ematiques Pures et Appliqu´ees (17, pp.
55–108).
Articles [Bou71a, Bou72] included dispersive eﬀects for the ﬁrst time in the SaintVenant
equations [dSV71]. One should mention that Boussinesq’s derivation was restricted to 1+1
dimensions (x and t) and a horizontal bottom. Boussinesq equations contain more physics
than the SaintVenant equations but at the same time they are more complicated from
the mathematical and numerical points of views. These equations possess a hyperbolic
structure (the same as in the nonlinear shallowwater equations) combined with highorder
derivatives to model wave dispersion. There have been a lot of further developments of
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
2.1 Introduction 83
these equations like in [Per67, Nwo93, WKGS95, MS98].
Let us outline the physical assumptions. The Boussinesq equations are intended to
describe the irrotational motion of an incompressible homogeneous inviscid ﬂuid in the
long wave limit. The goal of this type of modelling is to reduce 3D problems to two
dimensional (2D) ones. This is done by assuming a polynomial (usually linear) vertical
distribution of the ﬂow ﬁeld, while taking into account nonhydrostatic eﬀects. This is the
principal physical diﬀerence with the nonlinear shallowwater (NSW) equations.
There are a lot of forms of the Boussinesq equations. This diversity is due to diﬀerent
possibilities in the choice of the velocity variable. In most cases one chooses the velocity
at an arbitrary water level or the depthaveraged velocity vector. The resulting model
performance is highly sensitive to linear dispersion properties. The right choice of the
velocity variable can signiﬁcantly improve the propagation of moderately long waves. A
good review is given by [Kir03]. There is another technique used by [BCS02]. Formally, one
can transform higherorder terms by invoking lowerorder asymptotic relations. It provides
an elegant way to improve the properties of the linear dispersion relation and it gives a
quite general mathematical framework to study these systems.
The main purpose of this work is to include dissipative eﬀects in the Boussinesq equa
tions. It is wellknown that the eﬀect of viscosity on free oscillatory waves on deep water
was studied by [Lam32]. What is less known is that Boussinesq himself studied this eﬀect
as well. Boussinesq wrote three related papers in 1895 in the “Comptes Rendus Hebdo
madaires des S´eances de l’Acad´emie des Sciences”: (i) “Sur l’extinction graduelle de la
houle de mer aux grandes distances de son lieu de production : formation des ´equations du
probl`eme” (120, pp. 13811386), (ii) “Lois de l’extinction de la houle en haute mer” (121,
pp. 1520), (iii) “Sur la mani`ere dont se r´egularise au loin, en s’y r´eduisant `a une houle
simple, toute agitation confuse mais p´eriodique des ﬂots” (121, pp. 8588). It should be
pointed out that the famous treatise on hydrodynamics by Lamb has six editions. The
paragraphs on wave damping are not present in the ﬁrst edition (1879) while they are
present in the third edition (1906). The authors did not have access to the second edition
(1895), so it is possible that Boussinesq and Lamb published similar results at the same
time. Indeed Lamb derived the decay rate of the linear wave amplitude in two diﬀerent
ways: in paragraph 348 of the sixth edition by a dissipation calculation (this is also what
[Bou95] did) and in paragraph 349 by a direct calculation based on the linearized Navier
Stokes equations. Let α denote the wave amplitude, ν the kinematic viscosity of the ﬂuid
and k the wavenumber of the decaying wave. Boussinesq (see Eq. (12) in [Bou95]) and
Lamb both showed that
dα
dt
= −2νk
2
α(t). (2.1)
Equation (2.1) leads to the classical law for viscous decay of waves of amplitude α, namely
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
84 Dissipative Boussinesq equations
α ∼ exp(−2νk
2
t) (see Eq. (13) in [Bou95] after a few calculations).
In the present study, we use two diﬀerent models for dissipation and derive correspond
ing systems of longwave equations. There are several methods to derive the Boussinesq
equations but the resulting equations are not the same. So one expects the solutions to be
diﬀerent. We will investigate numerically whether corresponding solutions remain close or
not.
One may ask why dissipation is needed in Boussinesq equations. First of all, real world
liquids are viscous. This physical eﬀect is “translated” in the language of partial diﬀerential
equations by dissipative terms (e.g. the Laplacian in the NavierStokes equations). So, it
is natural to have analogous terms in the long wave limit. In other words, a nondissipative
model means that there is no energy loss, which is not pertinent from a physical point of
view, since any ﬂow is accompanied by energy dissipation.
Let us mention an earlier numerical and experimental study by [BPS81]. They pointed
out the importance of dissipative eﬀects for accurate long wave modelling. In the “R´esum´e”
section one can read
[...] it was found that the inclusion of a dissipative term was much more im
portant than the inclusion of the nonlinear term, although the inclusion of the
nonlinear term was undoubtedly beneﬁcial in describing the observations [...].
The complexity of the mathematical equations due to the inclusion of this term is negligible
compared to the beneﬁt of a better physical description.
Let us consider the incompressible NavierStokes (NS) equations for a Newtonian ﬂuid:
∇ u
∗
= 0,
∂u
∗
∂t
∗
+u
∗
∇u
∗
= −
∇p
∗
ρ
+ ν∆u
∗
+
F
∗
ρ
,
where u
∗
(x, y, z, t) = (u
∗
, v
∗
, w
∗
)(x
∗
, y
∗
, z
∗
, t
∗
) is the ﬂuid velocity vector, p
∗
the pressure,
F
∗
the body force vector, ρ the constant ﬂuid density and ν the kinematic viscosity.
Switching to dimensionless variables by introducing a characteristic velocity U, a char
acteristic length L and a characteristic pressure ρU
2
, neglecting body forces
1
in this dis
cussion, the NS equations become
∇ u = 0,
∂u
∂t
+u ∇u = −∇p +
1
Re
∆u,
1
The presence or absence of body forces is not important for discussing viscous eﬀects.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
2.2 Derivation of the Boussinesq equations 85
where Re is the wellknown dimensionless parameter known as the Reynolds number and
deﬁned as
Re =
F
inertia
F
viscous
=
UL
ν
.
From a physical point of view the Reynolds number is a measure of the relative im
portance of inertial forces compared to viscous eﬀects. For typical tsunami propagation
applications the characteristic particle velocity U is about 5 cm/s and the characteristic
wave amplitude, which we use here as characteristic length scale, is about 1 m. The kine
matic viscosity ν depends on the temperature but its order of magnitude for water is 10
−6
m
2
/s. Considering that as the tsunami approaches the coast both the particle velocity and
the wave amplitude increase, one can write that the corresponding Reynolds number is of
the order of 10
5
or 10
6
. This simple estimate clearly shows that the ﬂow is turbulent (as
many other ﬂows in nature).
It is a common practice in ﬂuid dynamics (addition of an “eddy viscosity” into the
governing equations for Large Eddy Simulations) to ignore the smallscale vortices when
one is only interested in largescale motion. It can signiﬁcantly simplify computational and
modelling aspects. So the inclusion of dissipation can be viewed as the simplest way to
take into account the turbulence.
There are several authors [Tuc74, LH92, SVBM02, SKP96, DDZ07, RFF91] who in
cluded dissipation due to viscosity in potential ﬂow solutions and there are also authors
[KCKD00, Zel91, HH70] who already included in Boussinesq models adhoc dissipative
terms into momentum conservation equations in order to model wave breaking. Modelling
this eﬀect is not the primary goal of the present work, since the ﬂow is no longer irrota
tional after wave breaking. Strictly speaking the Boussinesq equations can no longer be
valid at this stage. Nevertheless scientists and engineers continue to use these equations
even to model the runup on the beach. In our approach a suitable choice of the eddy
viscosity, which is a function of both space and time, can model wave breaking at least as
well as in the articles cited above.
2.2 Derivation of the Boussinesq equations
In order to derive the Boussinesq equations, we begin with the full waterwave problem.
A Cartesian coordinate system (x
′
, y
′
, z
′
) is used, with the x
′
− and y
′
−axis along the still
water level and the z
′
− axis pointing vertically upwards. Let Ω
t
be the ﬂuid domain in R
3
which is occupied by an inviscid and incompressible ﬂuid. The subscript t underlines the
fact that the domain varies with time and is not known a priori. The domain Ω
t
is bounded
below by the seabed z
′
= −h
′
(x
′
, y
′
, t
′
) and above by the free surface z
′
= η
′
(x
′
, y
′
, t
′
). In
this section we choose the domain Ω
t
to be unbounded in the horizontal directions in order
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
86 Dissipative Boussinesq equations
−4
−2
0
2
4
−4
−2
0
2
4
−2
−1.5
−1
−0.5
0
0.5
x’
y’
z
’
x’
y’
z’
O
a
η(x’,y’,t’)
h
0
−h’(x’,y’,t’)
L
Figure 2.1: Sketch of the ﬂuid domain
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
2.2 Derivation of the Boussinesq equations 87
to avoid the discussion on lateral boundary conditions. The reason is twofold. First of
all, the choice of the boundary value problem (BVP) (e.g. generating and/or absorbing
boundary, wall, runup on a beach) depends on the application under consideration and
secondly, the question of the wellposedness of the BVP for the Boussinesq equations is
essentially open. Primes stand for dimensional variables. A typical sketch of the domain
Ω
t
is given in Figure 2.1. If the ﬂow is assumed to be irrotational one can introduce the
velocity potential φ
′
deﬁned by
u
′
= ∇
′
φ
′
, ∇
′
:=
_
∂
∂x
′
,
∂
∂y
′
,
∂
∂z
′
_
T
,
where u
′
denotes the velocity ﬁeld. Then we write down the following system of equations
for potential ﬂow theory in the presence of a free surface:
∆
′
φ
′
= 0, (x
′
, y
′
, z
′
) ∈ Ω
t
= R
2
[−h
′
, η
′
],
φ
′
z
′ = η
′
t
′ +∇
′
φ
′
∇
′
η
′
, z
′
= η
′
,
φ
′
t
′ +
1
2
[∇
′
φ
′
[
2
+ gη
′
= 0, z
′
= η
′
, (2.2)
φ
′
z
′ + h
′
t
′ +∇
′
φ
′
∇
′
h
′
= 0, z
′
= −h
′
,
where g denotes the acceleration due to gravity (surface tension eﬀects are usually neglected
for longwave applications). It has been assumed implicitly that the free surface is a graph
and that the pressure is constant on the free surface (no forcing). Moreover we assume
that the total water depth remains positive, i.e. η
′
+ h
′
> 0 (there is no dry zone).
As written, this system of equations does not contain any dissipation. Thus, we com
plete the freesurface dynamic boundary condition (2.2) by adding a dissipative term to
account for the viscous eﬀects
2
:
φ
′
t
′ +
1
2
[∇
′
φ
′
[
2
+ gη
′
+ D
′
φ
′ = 0, z
′
= η
′
.
In this work we investigate two models for the dissipative term D
′
φ
′ . For simplicity,one
can choose a constant dissipation model (referred hereafter as Model I) which is often used
(e.g. [JTPS96]):
Model I: D
′
φ
′ := δ
1
φ
′
. (2.3)
There is a more physically realistic dissipation model which is obtained upon balancing of
normal stress on the free surface (e.g. [RFF91, ZV97, DDZ07]):
Model II: D
′
φ
′ := δ
2
φ
′
z
′
z
′ . (2.4)
2
[DDZ07] pointed out that a viscous correction should also be added to the kinematic boundary con
dition if one takes into account the vortical component of the velocity. This correction will be added in
future work.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
88 Dissipative Boussinesq equations
The derivation of Boussinesq equations is more transparent when one works with scaled
variables. Let us introduce the following independent and dependent nondimensional
variables:
x =
x
′
ℓ
, y =
y
′
ℓ
, z =
z
′
h
0
, t =
√
gh
0
ℓ
t
′
,
h =
h
′
h
0
, η =
η
′
a
0
, φ =
√
gh
0
ga
0
ℓ
φ
′
,
where h
0
, ℓ and a
0
denote a characteristic water depth, wavelength and wave amplitude,
respectively.
After this change of variables, the set of equations becomes
µ
2
(φ
xx
+ φ
yy
) + φ
zz
= 0, (x, y, z) ∈ Ω
t
, (2.5)
φ
z
= µ
2
η
t
+ εµ
2
∇φ ∇η, z = εη, (2.6)
µ
2
φ
t
+
1
2
εµ
2
[∇φ[
2
+
1
2
εφ
2
z
+ µ
2
η + εD
φ
= 0, z = εη (2.7)
φ
z
+
µ
2
ε
h
t
+ εµ
2
∇φ ∇h = 0, z = −h, (2.8)
where ε and µ are the classical nonlinearity and frequency dispersion parameters deﬁned
by
ε :=
a
0
h
0
, µ :=
h
0
ℓ
. (2.9)
In these equations and hereafter the symbol ∇ denotes the horizontal gradient:
∇ :=
_
∂
∂x
,
∂
∂y
_
T
.
The dissipative term D
φ
is given by the chosen model (2.3) or (2.4):
Model I: D
φ
=
1
R
1
φ, Model II: D
φ
=
1
R
2
φ
zz
,
where the following dimensionless numbers have been introduced:
R
1
:=
1
δ
1
_
ga
0
ℓ
h
2
0
√
gh
0
_
, R
2
:=
1
δ
2
_
ga
0
ℓ
√
gh
0
_
.
From this dimensional analysis, one can conclude that the dimension of the coeﬃcient δ
1
is [s
−1
] and that of δ
2
is [m
2
s
−1
]. Thus, it is natural to call the ﬁrst coeﬃcient viscous fre
quency (since it has the dimensions of a frequency) and the second one kinematic viscosity
(by analogy with the NS equations).
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
2.2 Derivation of the Boussinesq equations 89
parameter value
Acceleration due to gravity g, m/s
2
10
Amplitude a
0
, m 1
Wave length ℓ, km 100
Water depth h
0
, km 4
Kinematic viscosity δ, m
2
/s 10
−6
Table 2.1: Typical values of characteristic parameters in tsunami applications
It is interesting to estimate R
2
, since we know how to relate the value of δ
2
to the
kinematic viscosity of water. Typical parameters which are used in tsunami wave modelling
are given in Table 2.1. For these parameters R
2
= 5 10
9
and µ
2
= 1.6 10
−3
. The ratio
between inertial forces and viscous forces is
1
2
εµ
2
[∇φ[
2
/ε[D
φ
[. Its order of magnitude is
µ
2
R
2
, that is 8 10
6
. It clearly shows that the ﬂow is turbulent and eddyviscosity type
approaches should be used. It means that, at zerothorder approximation, the main eﬀect
of turbulence is energy dissipation. Thus, one needs to increase the importance of viscous
terms in the governing equations in order to account for turbulent dissipation.
As an example, we refer one more time to the work by [BPS81]. They modeled long
wave propagation by using a modiﬁed dissipative Korteweg–de Vries equation:
η
t
+ η
x
+
3
2
ηη
x
−µη
xx
−
1
6
η
xxt
= 0. (2.10)
In numerical computations the authors took the coeﬃcient µ = 0.014. This value gave
good agreement with laboratory data.
From now on, we will use the notation ν
i
:= 1/R
i
. This will allow us to unify the
physical origin of the numbers R
i
with the eddyviscosity approach. In other words, for
the sake of convenience, we will “forget” about the origin of these coeﬃcients, because
their values can be given by other physical considerations.
2.2.1 Asymptotic expansion
Consider a formal asymptotic expansion of the velocity potential φ in powers of the
small parameter µ
2
:
φ = φ
0
+ µ
2
φ
1
+ µ
4
φ
2
+ . . . . (2.11)
Then substitute this expansion into the continuity equation (2.5) and the boundary
conditions. After substitution, the Laplace equation becomes
µ
2
(∇
2
φ
0
+ µ
2
∇
2
φ
1
+ µ
4
∇
2
φ
2
+ . . .) + φ
0zz
+ µ
2
φ
1zz
+ µ
4
φ
2zz
+ . . . = 0.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
90 Dissipative Boussinesq equations
Collecting the same order terms yields the following equations in the domain Ω
t
:
µ
0
: φ
0zz
= 0, (2.12)
µ
2
: φ
1zz
+∇
2
φ
0
= 0, (2.13)
µ
4
: φ
2zz
+∇
2
φ
1
= 0. (2.14)
Performing the same computation for the bottom boundary condition yields the following
relations at z = −h:
µ
0
: φ
0z
= 0, (2.15)
µ
2
: φ
1z
+
1
ε
h
t
+∇φ
0
∇h = 0, (2.16)
µ
4
: φ
2z
+∇φ
1
∇h = 0. (2.17)
From equation (2.12) and the boundary condition (2.15) one immediately concludes that
φ
0
= φ
0
(x, y, t).
Let us deﬁne the horizontal velocity vector
u(x, y, t) := ∇φ
0
, u = (u, v)
T
.
The expansion of Laplace equation in powers of µ
2
gives recurrence relations between
φ
0
, φ
1
, φ
2
, etc. Using (2.13) one can express φ
1
in terms of the derivatives of φ
0
:
φ
1zz
= −∇ u.
Integrating once with respect to z yields
φ
1z
= −z∇ u + C
1
(x, y, t).
The unknown function C
1
(x, y, t) can be found by using condition (2.16):
φ
1z
= −(z + h)∇ u −
1
ε
h
t
−u ∇h,
and integrating one more time with respect to z gives the expression for φ
1
:
φ
1
= −
1
2
(z + h)
2
∇ u −z
_
1
ε
h
t
+u ∇h
_
. (2.18)
Now we will determine φ
2
. For this purpose we use equation (2.14):
φ
2zz
=
1
2
(z + h)
2
∇
2
(∇ u) +
_
(h + z)∇
2
h +[∇h[
2
_
∇ u+
+ 2(h + z)∇h ∇(∇ u) + z
_
1
ε
∇
2
h
t
+∇
2
(u ∇h)
_
. (2.19)
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
2.2 Derivation of the Boussinesq equations 91
Integrating twice with respect to z and using the bottom boundary condition (2.17)
yields the following expression for φ
2
:
φ
2
=
1
24
(h + z)
4
∇
2
(∇ u) +
_
1
6
(z + h)
3
∇
2
h +
1
2
z
2
[∇h[
2
_
∇ u
+
1
3
(z + h)
3
∇h ∇(∇ u) +
z
3
6
_
1
ε
∇
2
h
t
+∇
2
(u ∇h)
_
−zh
_
h
2
∇
2
_
1
ε
h
t
+u ∇h
_
+∇h ∇
_
1
ε
h
t
+u ∇h
_
−[∇h[
2
∇ u
_
. (2.20)
Remark 2 In these equations one ﬁnds the term (1/ε)h
t
due to the moving bathymetry.
We would like to emphasize that this term is O(1), since in problems of wave generation by
a moving bottom the bathymetry h(x, y, t) has the following special form in dimensionless
variables:
h(x, y, t) := h
0
(x, y) −εζ(x, y, t), (2.21)
where h
0
(x, y) is the static seabed and ζ(x, y, t) is the dynamic component due to a seis
mic event or a landslide (see for example [DD07c] for a practical algorithm constructing
ζ(x, y, t) in the absence of a dynamic source model). The amplitude of the bottom motion
has to be of the same order of magnitude as the resulting waves, since we assume the ﬂuid
to be inviscid and incompressible. Thus (1/ε)h
t
= −ζ
t
= O(1).
In the present study we restrict our attention to dispersion terms up to order O(µ
2
).
We will also assume that the UrsellStokes number [Urs53] is O(1):
S :=
ε
µ
2
= O(1).
This assumption implies that terms of order O(ε
2
) and O(εµ
2
) must be neglected, since
ε
2
= S
2
µ
4
= O(µ
4
), εµ
2
= Sµ
4
= O(µ
4
).
Of course, it is possible to obtain highorder Boussinesq equations. We decided not to
take this research direction. For highorder asymptotic expansions we refer to [WKGS95,
MS98]. Recently, [Ben06] performed a comparative study between fullynonlinear equa
tions [WKGS95] and Boussinesq equations with optimized dispersion relation [Nwo93]. No
substantial diﬀerence was revealed.
Now, we are ready to derive dissipative Boussinesq equations in their simplest form.
First of all, we substitute the asymptotic expansion (2.11) into the kinematic freesurface
boundary condition (2.6):
φ
0z
+ µ
2
φ
1z
+ µ
4
φ
2z
= µ
2
η
t
+ εµ
2
∇φ
0
∇η + O(ε
2
+ εµ
4
+ µ
6
), z = εη. (2.22)
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
92 Dissipative Boussinesq equations
The ﬁrst term on the left hand side is equal to zero because of Eq. (2.15).
Using expressions (2.18) and (2.20) one can evaluate φ
1z
and φ
2z
on the free surface:
φ
1z
[
z=εη
= −(h + εη)∇ u −
1
ε
h
t
−u ∇h,
φ
2z
[
z=εη
=
h
3
6
∇
2
(∇ u) + h
2
∇h ∇(∇ u) + h
_
h
2
∇
2
h +[∇h[
2
_
∇ u
−
h
2
2
1
ε
∇
2
h
t
−h
1
ε
∇h
t
∇h + O(ε).
Substituting these expressions into (2.22) and retaining only terms of order O(ε+µ
2
) yields
the freesurface elevation equation:
η
t
+∇
_
(h + εη)u
_
= −
_
1 +
µ
2
2
h
2
∇
2
+ µ
2
h∇h ∇
_
1
ε
h
t
+ µ
2
h
3
6
∇
2
(∇ u)
+ µ
2
h
_
h∇h ∇(∇ u) +
_
h
2
∇
2
h +[∇h[
2
_
∇ u
_
.
The equation for the evolution of the velocity ﬁeld is derived similarly from the dynamic
boundary condition (2.7). This derivation will depend on the selected dissipation model.
For both models one has to evaluate φ
1
, φ
1t
and φ
1zz
along the free surface z = εη and
then substitute the expressions into the asymptotic form of (2.7):
µ
2
φ
0t
+ µ
4
φ
1t
+
1
2
εµ
2
[∇φ
0
[
2
+ µ
2
η + εν
2
µ
2
φ
1zz
= O(ε
2
+ εµ
4
+ µ
6
),
where, as an example, dissipative terms are given according to the second model. After
performing all these operations one can write down the following equations:
Model I: φ
0t
+
ε
2
u
2
+ η + ν
1
ε
µ
2
φ
0
−
ν
1
ε
2
h
2
∇ u −
µ
2
2
h
2
∇ u
t
= 0,
Model II: φ
0t
+
ε
2
u
2
+ η −ν
2
ε∇ u −
µ
2
2
h
2
∇ u
t
= 0.
The last step consists in diﬀerentiating the above equations with respect to the horizontal
coordinates in order to obtain equations for the evolution of the velocity. We also perform
some minor transformations using the fact that the vector u is a gradient by deﬁnition, so
we have the obvious relation
∂u
∂y
=
∂v
∂x
.
The resulting Boussinesq equations for the ﬁrst and second dissipation models, respec
tively, are given below:
η
t
+∇
_
(h + εη)u
_
= −
_
1 +
µ
2
2
h
2
∇
2
+ µ
2
h∇h ∇
_
1
ε
h
t
+ µ
2
h
3
6
∇
2
(∇ u)
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
2.3 Analysis of the linear dispersion relations 93
+ µ
2
h
_
h∇h ∇(∇ u) +
_
h
2
∇
2
h +[∇h[
2
_
∇ u
_
, (2.23)
Model I: u
t
+
1
2
ε∇u
2
+∇η + ν
1
Su =
1
2
εν
1
∇(h
2
∇ u) +
1
2
µ
2
∇(h
2
∇ u
t
), (2.24)
Model II: u
t
+
1
2
ε∇u
2
+∇η = εν
2
∇
2
u +
1
2
µ
2
∇(h
2
∇ u
t
). (2.25)
2.3 Analysis of the linear dispersion relations
For simplicity, we will consider in this section only 2D problems. The generalization to
3D problems is straightforward and does not change the analysis.
2.3.1 Linearization of the full potential ﬂow equations with dissipation
First we write down the linearization of the full potential ﬂow equations in dimensional
form, after dropping the primes:
∆φ = 0, (x, z) ∈ R [−h, 0], (2.26)
φ
z
= η
t
, z = 0, (2.27)
φ
t
+ gη + D
φ
= 0, z = 0, (2.28)
φ
z
= 0, z = −h. (2.29)
Remark 3 In this section the water layer is assumed to be of uniform depth, so h = const.
As above the term D
φ
depends on the selected dissipation model and is equal to δ
1
φ or
δ
2
φ
zz
. The next step consists in choosing a special form of solutions:
φ(x, z, t) = ϕ
0
e
i(kx−ωt)
ϕ(z), η(x, t) = η
0
e
i(kx−ωt)
, (2.30)
where ϕ
0
and η
0
are constants. Substituting this form of solutions into equations (2.26),
(2.27) and (2.29) yields the following boundary value problem for an ordinary diﬀerential
equation:
ϕ
′′
(z) −k
2
ϕ(z) = 0, z ∈ [−h, 0],
ϕ
′
(0) =
η
0
ϕ
0
(−iω), ϕ
′
(−h) = 0.
Straightforward computations give the solution to this problem:
ϕ(z) = −i
η
0
ϕ
0
_
e
k(2h+z)
+ e
−kz
e
2kh
−1
_
ω
k
.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
94 Dissipative Boussinesq equations
The dispersion relation can be thought as a necessary condition for solutions of the
form (2.30) to exist. The problem is that ω and k cannot be arbitrary. We obtain the
required relation ω = ω(k), which is called the dispersion relation, after substituting this
solution into (2.28).
When the dissipative term is chosen according to model I (2.3), D
φ
= δ
1
φ and the
dispersion relation is given implicitly by
ω
2
+ iδ
1
ω −gk tanh(kh) = 0,
or in explicit form by
ω = ±
_
gk tanh(kh) −
δ
2
1
4
−
iδ
1
2
. (2.31)
For the second dissipation model (2.4) one obtains the following relation:
ω
2
+ iδ
2
ωk
2
−gk tanh(kh) = 0.
One can easily solve this quadratic equation for ω as a function of k:
ω = ±
¸
gk tanh(kh) −
_
δ
2
k
2
2
_
2
−
iδ
2
2
k
2
. (2.32)
If δ
1,2
≡ 0 one easily recognizes the dispersion relation of the classical waterwave
problem:
ω = ±
_
gk tanh(kh). (2.33)
Remark 4 It is important to have the property Imω(k) ≤ 0, ∀k in order to avoid the
exponential growth of certain wavelengths, since
e
i(kx−ω(k)t)
= e
Imω(k)t
e
i(kx−Re ω(k)t)
.
For our analysis it is more interesting to look at the phase speed which is deﬁned as
c
p
(k) :=
ω(k)
k
.
The phase velocity is directly connected to the speed of wave propagation and is extremely
important for accurate tsunami modelling since tsunami arrival time obviously depends
on the propagation speed. The expressions for the phase velocity are obtained from the
corresponding dispersion relations (2.31) and (2.32):
c
(1)
p
(k) = ±
_
gh
tanh(kh)
kh
−
_
δ
1
2k
_
2
−
iδ
1
2k
, (2.34)
c
(2)
p
(k) = ±
_
gh
tanh(kh)
kh
−
_
δ
2
k
2
_
2
−
iδ
2
2
k. (2.35)
It can be shown that in order to keep the phase velocity unchanged by the addition of
dissipation, similar dissipative terms must be included in both the kinematic and the
dynamic boundary conditions [DDZ07].
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
2.3 Analysis of the linear dispersion relations 95
2.3.2 Dissipative Boussinesq equations
The analysis of the dispersion relation is even more straightforward for Boussinesq
equations. In order to be coherent with the previous subsection, we switch to dimensional
variables. As usual we begin with the (1 + 1)D linearized equations:
η
t
+ hu
x
=
h
3
6
u
xxx
,
Model I: u
t
+ gη
x
+ δ
1
u =
1
2
δ
1
h
2
u
xx
+
1
2
h
2
u
xxt
,
Model II: u
t
+ gη
x
= δ
2
u
xx
+
1
2
h
2
u
xxt
.
Now we substitute a special ansatz in these equations:
η = η
0
e
i(kx−ωt)
, u = u
0
e
i(kx−ωt)
,
where η
0
and u
0
are constants. In the case of the ﬁrst model, one obtains the following
homogeneous system of linear equations:
(−iω)η
0
+ ikh
_
1 +
1
6
(kh)
2
_
u
0
= 0,
gikη
0
+
_
−iω + δ
1
+
δ
1
2
(kh)
2
−
iω
2
(kh)
2
_
u
0
= 0.
This system admits nontrivial solutions if its determinant is equal to zero. It gives the
required dispersion relation:
ω
2
+ iωδ
1
−ghk
2
_
1 +
1
6
(kh)
2
1 +
1
2
(kh)
2
_
= 0.
A similar relation is found for the second model:
ω
2
+
iωδ
2
1 +
1
2
(kh)
2
k
2
−ghk
2
_
1 +
1
6
(kh)
2
1 +
1
2
(kh)
2
_
= 0.
The corresponding phase velocities are given by
c
(1)
pb
=
¸
gh
_
1 +
1
6
(kh)
2
1 +
1
2
(kh)
2
_
−
_
δ
1
2k
_
2
−
iδ
1
2k
, (2.36)
c
(2)
pb
=
¸
gh
_
1 +
1
6
(kh)
2
1 +
1
2
(kh)
2
_
−
_
δ
2
k
2 + (kh)
2
_
2
−
iδ
2
k
2 + (kh)
2
. (2.37)
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
96 Dissipative Boussinesq equations
0 2 4 6 8 10 12 14
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
kh
R
e
c
p
/
(
g
h
)
1
/
2
Phase speeds. Real part
Full, no dissipation
Full equations with dissipation−1
Dissipative Boussinesq−1
Figure 2.2: Dissipation model I. Real part of the phase velocity.
0 0.1 0.2 0.3 0.4 0.5 0.6
0
0.2
0.4
0.6
0.8
1
kh
R
e
c
p
/
(
g
h
)
1
/
2
Phase speeds. Real part. Zoom
Full, no dissipation
Full equations with dissipation−1
Dissipative Boussinesq−1
Figure 2.3: Dissipation model I. Same as Figure 2.2 with a zoom on long waves.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
2.3 Analysis of the linear dispersion relations 97
0 0.05 0.1 0.15 0.2
−0.14
−0.12
−0.1
−0.08
−0.06
−0.04
−0.02
0
kh
I
m
ω
(
k
)
Dispersion relation. Imaginary part
Full equations with dissipation−1. Branch −
Full equations with dissipation−1. Branch +
Dissipative Boussinesq−1
Figure 2.4: Dissipation model I. Imaginary part of the frequency.
0 2 4 6 8 10 12 14
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
kh
R
e
c
p
/
(
g
h
)
1
/
2
Phase speeds. Real part
Full, no dissipation
Full equations with dissipation−2
Dissipative Boussinesq−2
Figure 2.5: Dissipation model II. Real part of the phase velocity.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
98 Dissipative Boussinesq equations
0 2 4 6 8 10 12 14
−20
−18
−16
−14
−12
−10
−8
−6
−4
−2
0
kh
I
m
ω
(
k
)
Dispersion relations. Imaginary part
Full equations with dissipation−2. Branch −
Full equations with dissipation−2. Branch +
Dissipative Boussinesq−2
Figure 2.6: Dissipation model II. Imaginary part of the frequency.
0 0.5 1 1.5 2 2.5 3 3.5 4
−1
−0.9
−0.8
−0.7
−0.6
−0.5
−0.4
−0.3
−0.2
−0.1
0
kh
I
m
ω
(
k
)
Dispersion relations. Imaginary part
Full equations with dissipation−2. Branch +
Dissipative Boussinesq−2
Figure 2.7: Dissipation model II. Same as Figure 2.6 with a zoom on long waves.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
2.4 Alternative version of the Boussinesq equations 99
2.3.3 Discussion
Let us now provide a discussion on the dispersion relations. The real and imaginary
parts of the phase velocities (2.34) – (2.37) for the full and long wave linearized equations
are shown graphically on Figures 2.2 – 2.7. In this example the parameters are given by
δ
1
= 0.14, δ
2
= 0.14. Together with the dissipative models we also plotted for comparison
the wellknown phase velocity corresponding to the full conservative (linearized) water
wave problem:
c
p
(k) =
_
gh
tanh(kh)
kh
.
First of all, one can see that dissipation is very selective, as is often the case in physics.
Clearly, the ﬁrst dissipation model prefers very long waves, while the second model dis
sipates essentially short waves. Moreover one can see from the expressions (2.34), (2.36)
that the phase velocity has a 1/k singular behaviour in the vicinity of kh = 0 (in the long
wave limit). Furthermore, it can be clearly seen in Figure 2.3 that very long linear waves
are not advected in the ﬁrst dissipation model, since the real part of their phase velocity
is identically equal to zero.
That is why we suggest to make use of the second model in applications involving very
long waves such as tsunamis.
On the other hand we would like to point out that the second model admits a critical
wavenumber k
c
such that the phase velocity (2.35) becomes purely imaginary with negative
imaginary part. From a physical point of view it means that the waves shorter than k
c
are
not advected, but only dissipated. When one switches to the Boussinesq approximation,
this property disappears for physically realistic values of the parameters g, h and δ
2
(see
Table 2.1).
Let us clarify this situation. The qualitative behaviour of the phase velocity c
(2)
pb
(see
equation (2.37)) depends on the roots of the following polynomial equation:
(kh)
4
+
_
8 −
3δ
2
2
gh
3
_
(kh)
2
+ 12 = 0.
This equation does not have real roots since
3δ
2
2
gh
3
≪1.
2.4 Alternative version of the Boussinesq equations
In this section we give an alternative derivation of Boussinesq equations. We use another
classical method for deriving Boussinesqtype equations [Whi99, Ben74, Per72], which
provides slightly diﬀerent governing equations. Namely, the hyperbolic structure is the
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
100 Dissipative Boussinesq equations
same, but the dispersive terms diﬀer. In numerical simulations we suggest to use this
system of equations.
The derivation follows closely the paper by [MS98]. The main diﬀerences are that we
neglect the terms of order O(µ
4
), take into account a moving bathymetry and, of course,
dissipative eﬀects which are modelled this time according to model II (2.4) because, in our
opinion, this model is more appropriate for long wave applications. Anyhow, the derivation
process can be performed in a similar fashion for model I (2.3).
2.4.1 Derivation of the equations
The starting point is the same: equations (2.5), (2.6), (2.7) and (2.8). This time the
procedure begins with representing the velocity potential φ(x, y, z, t) as a formal expansion
in powers of z rather than of µ
2
:
φ(x, y, z, t) =
∞
n=0
z
n
φ
n
(x, y, t). (2.38)
We would like to emphasize that this expansion is only formal and no convergence result is
needed. In other words, it is just convenient to use this notation in asymptotic expansions
but in practice, seldom more than four terms are used. It is not necessary to justify the
convergence of the sum with three or four terms.
When we substitute the expansion (2.38) into Laplace equation (2.5), we have an inﬁnite
polynomial in z. Requiring that φ formally satisﬁes Laplace equation implies that the
coeﬃcients of each power of z vanish (since the righthand side is identically zero). This
leads to the classical recurrence relation
φ
n+2
(x, y, t) = −
µ
2
(n + 1)(n + 2)
∇
2
φ
n
(x, y, t), n = 0, 1, 2, . . .
Using this relation one can eliminate all but two unknown functions in (2.38):
φ(x, y, z, t) =
∞
n=0
(−1)
n
µ
2n
_
z
2n
(2n)!
∇
2n
φ
0
+
z
2n+1
(2n + 1)!
∇
2n
φ
1
_
.
The following notation is introduced:
u
0
:= u(x, y, 0, t), w
0
:=
1
µ
2
w(x, y, 0, t).
It is straightforward to ﬁnd the relations between u
0
, w
0
and φ
0
, φ
1
if one remembers that
(u, w) = (∇,
∂
∂z
)φ:
u
0
= ∇φ
0
, w
0
=
1
µ
2
φ
1
.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
2.4 Alternative version of the Boussinesq equations 101
Using the deﬁnition of the velocity potential φ one can express the velocity ﬁeld in terms
of u
0
, w
0
:
u =
∞
n=0
(−1)
n
_
z
2n
(2n)!
µ
2n
∇
_
∇
2n−2
(∇ u
0
)
_
+
z
2n+1
(2n + 1)!
µ
2n+2
∇
_
∇
2n
w
0
_
_
,
w =
∞
n=0
(−1)
n
_
−
z
2n+1
(2n + 1)!
µ
2n+2
∇
2n
(∇ u
0
) +
z
2n
(2n)!
µ
2n+2
∇
2n
w
0
_
.
These formulas are exact but not practical. In the present work we neglect the terms of
order O(µ
4
) and higher. In this asymptotic framework the above formulas become much
simpler:
φ = φ
0
+ zφ
1
−
µ
2
z
2
2
_
∇
2
φ
0
+
z
3
∇
2
φ
1
_
+ O(µ
4
), (2.39)
u = u
0
+ zµ
2
∇w
0
−
µ
2
z
2
2
∇(∇ u
0
) + O(µ
4
), (2.40)
w = µ
2
w
0
−zµ
2
∇ u
0
+ O(µ
4
). (2.41)
In order to establish the relation between w
0
and u
0
one uses the bottom kinematic
boundary condition (2.8), which has the following form after substituting the asymptotic
expansions (2.39), (2.40), (2.41) in it:
h
t
+ ε∇h
_
u
0
−hµ
2
∇w
0
−
µ
2
h
2
2
∇(∇ u
0
)
_
+ ε
_
w
0
−
h
3
6
µ
2
∇
2
(∇ u
0
) −
µ
2
h
2
2
∇
2
w
0
_
+ O(µ
4
) = 0. (2.42)
In order to obtain the expression of w
0
in terms of u
0
one introduces one more expansion:
w
0
(x, y, t) = w
(0)
0
(x, y, t) + µ
2
w
(1)
0
(x, y, t) + . . . . (2.43)
We insert this expansion into the asymptotic bottom boundary condition (2.42). This
leads to the following explicit expressions for w
(0)
0
and w
(1)
0
:
w
(0)
0
= −
1
ε
h
t
−∇ (hu
0
),
w
(1)
0
=
h
2
2
_
∇h ∇(∇ u
0
) +
h
3
∇
2
(∇ u
0
)
_
−h
_
∇h
1
ε
∇h
t
+∇h ∇
_
∇ (hu
0
)
_
+
h
2
_
1
ε
∇
2
h
t
+∇
2
(∇ (hu
0
))
_
_
.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
102 Dissipative Boussinesq equations
Substituting these expansions into (2.43) and performing some simpliﬁcations yields the
required relation between u
0
and w
0
:
w
0
= −
1
ε
h
t
−∇ (hu
0
)
−µ
2
∇
_
h
2
2ε
∇h
t
+
h
2
2
∇
_
∇ (hu
0
)
_
−
h
3
6
∇(∇ u
0
)
_
+ O(µ
4
). (2.44)
Now one can eliminate the vertical velocity w
0
since one has its expression (2.44) in terms
of u
0
. Equations (2.40)(2.41) become
u = u
0
−z
µ
2
ε
∇h
t
−µ
2
_
z∇
_
∇ (hu
0
)
_
+
z
2
2
∇(∇ u
0
)
_
+ O(µ
4
), (2.45)
w = −
µ
2
ε
h
t
−µ
2
_
∇ (hu
0
) + z∇ u
0
_
+ O(µ
4
). (2.46)
In this work we apply a trick due to [Nwo93]. Namely, we introduce a new velocity
variable u
α
deﬁned at an arbitrary water level z
α
= −αh. Technically this change of
variables is done as follows. First we evaluate (2.45) at z = z
α
, which gives the connection
between u
0
and u
α
:
u
α
= u
0
−z
α
µ
2
ε
∇h
t
−µ
2
_
z
α
∇
_
∇ (hu
0
)
_
+
z
2
α
2
∇(∇ u
0
)
_
+ O(µ
4
).
Using the standard techniques of inversion one can rewrite the last expression as an asymp
totic formula for u
0
in terms of u
α
:
u
0
= u
α
+ z
α
µ
2
ε
h
t
+ µ
2
_
z
α
∇
_
∇ (hu
α
)
_
+
z
2
α
2
∇(∇ u
α
)
_
+ O(µ
4
). (2.47)
Remark 5 Behind this change of variables there is one subtlety which is generally hushed
up in the literature. In fact, the wave motion is assumed to be irrotational since we use
the potential ﬂow formulation (2.5), (2.6), (2.7), (2.8) of the waterwave problem. By
construction rot(u, w) = O when u and w are computed according to (2.45), (2.46) or,
in other words, in terms of the variable u
0
. When one turns to the velocity variable u
α
deﬁned at an arbitrary level, one can improve the linear dispersion relation and this is
important for wave modelling. But on the other hand, one loses the property that the ﬂow
is irrotational. That is to say, a direct computation shows that rot(u, w) ,= O when u and
w are expressed in terms of the variable u
α
. The purpose of this remark is simply to inform
the reader about the price to be paid while improving the dispersion relation properties. It
seems that this point is not clearly mentioned in the literature on this topic.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
2.4 Alternative version of the Boussinesq equations 103
Let us now derive the Boussinesq equations. There are two diﬀerent methods to obtain
the freesurface elevation equation. The ﬁrst method consists in integrating the continuity
equation (2.5) over the depth and then use the kinematic freesurface and bottom boundary
conditions. The second way is more straightforward. It consists in using directly the
kinematic freesurface boundary condition (2.6):
η
t
+ ε∇φ ∇η −
1
µ
2
φ
z
= 0.
Then one can substitute (2.39) into (2.6) and perform several simpliﬁcations. Neglecting
all terms of order O(ε
2
+ εµ
2
+ µ
4
) yields the following equation
3
:
η
t
+∇
_
(h + εη)u
0
_
+
µ
2
2
∇
_
h
2
∇
_
∇ (hu
0
)
_
−
h
3
3
∇(∇ u
0
)
_
=
= ζ
t
+
µ
2
2
∇ (h
2
∇ζ
t
).
Recall that ζ(x, y, t) is deﬁned according to (2.21). When the bathymetry is static, ζ ≡ 0.
We prefer to introduce this function in order to eliminate the division by ε in the source
terms since this division can give the impression that stiﬀ source terms are present in our
problem, which is not the case.
In order to be able to optimize the dispersion relation properties, we switch to the
variable u
α
. Technically it is done by using the relation (2.47) between u
0
and u
α
. The
result is given below:
η
t
+∇
_
(h + εη)u
α
_
+ µ
2
∇
_
h
_
z
α
+
h
2
_
∇
_
∇ (hu
α
)
_
+
h
2
_
z
2
α
−
h
2
3
_
∇(∇ u
α
)
_
=
= ζ
t
+ µ
2
∇
_
h
_
z
α
+
h
2
_
∇ζ
t
_
. (2.48)
As above, the equation for the horizontal velocity ﬁeld is derived from the dynamic
freesurface boundary condition (2.7). It is done exactly as in section 2.2 and we do not
insist on this point:
u
0t
+
ε
2
∇[u
0
[
2
+∇η −εδ∇
2
u
0
= O.
Switching to the variable u
α
yields the following governing equation:
u
αt
+
ε
2
∇[u
α
[
2
+∇η + µ
2
_
z
α
∇
_
∇ (hu
α
)
_
+
z
2
α
2
∇(∇ u
α
)
_
t
=
= εδ∆u
α
+ µ
2
(z
α
∇ζ
t
)
t
. (2.49)
3
We already discussed this point on page 91. In this section we also assume that the StokesUrsell
number S is of order O(1).
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
104 Dissipative Boussinesq equations
In several numerical methods it can be advantageous to rewrite the system (2.48), (2.49)
in vector form:
U
t
+ µ
2
L(U)
t
+∇ F(U) + µ
2
∇ P(U) = S(x, y, t) + εδ∇ (D∇U)),
where
U :=
_
_
_
η
u
α
v
α
_
_
_
, ∇ F :=
∂F
∂x
+
∂G
∂y
,
F :=
_
_
_
(h + εη)u
α
ε
2
[u
α
[
2
+ η
0
_
_
_
, G :=
_
_
_
(h + εη)v
α
0
ε
2
[u
α
[
2
+ η
_
_
_
,
L :=
_
0
z
α
∇
_
∇ (hu
α
)
_
+
z
2
α
2
∇(∇ u
α
)
_
,
P :=
_
h
_
z
α
+
h
2
_
∇
_
∇ (hu
α
)
_
+
h
2
_
z
2
α
−
h
2
3
_
∇(∇ u
α
)
O
_
,
S :=
_
ζ
t
+ µ
2
∇
_
h
_
z
α
+
h
2
_
∇ζ
t
_
µ
2
(z
α
∇ζ
t
)
t
_
,
D :=
_
_
_
0 0 0
0 1 0
0 0 1
_
_
_
.
2.5 Improvement of the linear dispersion relations
As said above, the idea of using one free parameter α ∈ [0, 1] to optimize the linear
dispersion relation properties appears to have been proposed ﬁrst by [Nwo93].
The idea of manipulating the dispersion relation was wellknown before 1993. See for
example [Mur89, MMS91]. But these authors started with a desired dispersion relation and
artiﬁcially added extra terms to the momentum equation in order to produce the desired
characteristics. We prefer to follow the ideas of [Nwo93].
Remark 6 When one plays with the dispersion relation it is important to remember that
the resulting problem must be wellposed, at least linearly. We refer to [BCS02] as a general
reference on this topic. Usually Boussinesqtype models with good dispersion characteristics
are linearly wellposed as well.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
2.6 Regularization of Boussinesq equations 105
In order to look for an optimal value of α we will drop dissipative terms. Indeed we
want to concentrate our attention on the propagation properties which are more important.
The choice for the parameter α depends on the optimization criterion. In the present
work we choose α by comparing the coeﬃcients in the Taylor expansions of the phase
velocity in the vicinity of kh = 0, which corresponds to the longwave limit. Another
possibility is to match the dispersion relation of the full linearized equations (2.33) in the
least square sense. One can also use Pad´e approximants [Wit84] since rational functions
have better approximation properties than polynomials.
We brieﬂy describe the procedure. First of all one has to obtain the phase velocity of
the linearized, nonviscous, Boussinesq equations (2.48)(2.49). The result is
c
2
b
(k)
gh
=
1 −
_
α
2
2
−α +
1
3
_
(kh)
2
1 −α
_
α
2
−1
_
(kh)
2
= 1 −
1
3
(kh)
2
+
α(2 −α)
6
(kh)
4
+ O
_
(kh)
6
_
. (2.50)
On the other hand one can write down the phase velocity of the full linearized equations
(2.33):
c
2
(k)
gh
=
tanh(kh)
kh
= 1 −
1
3
(kh)
2
+
2
15
(kh)
4
+ O
_
(kh)
6
_
. (2.51)
If one insists on the dispersion relation (2.50) to be exact up to order O((kh)
4
) one imme
diately obtains an equation for α
opt
:
α
opt
(2 −α
opt
)
6
=
2
15
⇒α
opt
= 1 −
√
5
5
≈ 0.55.
We suggest using this value of α in numerical computations.
2.6 Regularization of Boussinesq equations
In this section we are going to modify further just obtained Boussinesq equations (2.48),
(2.49). For simplicity we assume the bottom to be static h = h(x, y) and we drop out
viscous terms. This operation will allow us to obtain very
gentle
system of equations
from the numerical point of view. The inﬂuence of this modiﬁcation on dispersion relation
properties will be discussed below.
The main idea behind is very simple. It was already extensively exploited in [BCS02]
and other publications. We are going to use lower order asymptotic relations in order to
play
with dispersive terms. Namely, from (2.48), (2.49) one can deduce two following
relations:
η
t
= −∇ (hu) +O(ε + µ
2
), (2.52)
u
t
= −∇η +O(ε + µ
2
).
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
106 Dissipative Boussinesq equations
In the present work we will only use the relation (2.52). With its help we can modify the
ﬁrst dispersive term in the equation (2.48):
µ
2
∇
_
h
_
z
α
+
h
2
_
∇
_
∇ (hu
α
)
_
_
= −µ
2
∇
_
h
_
z
α
+
h
2
_
∇η
_
t
+O(εµ
2
+ µ
4
).
Thus, the regularized system of equations reads:
η
t
+∇
_
(h+εη)u
α
_
−µ
2
∇
_
h
_
z
α
+
h
2
_
∇η
_
t
+µ
2
∇
_
h
2
_
z
2
α
−
h
2
3
_
∇(∇ u
α
)
_
= 0, (2.53)
u
αt
+
ε
2
∇[u
α
[
2
+∇η + µ
2
_
z
α
∇
_
∇ (hu
α
)
_
+
z
2
α
2
∇(∇ u
α
)
_
t
=
0. (2.54)
Just obtained system of equations has the same relation with classical Boussinesq equa
tions [Per67] as KdV equation u
t
+u
x
+uu
x
+u
xxx
= 0 is related to BBM model [BBM72]
u
t
+ u
x
+ uu
x
− u
xxt
= 0. The main purpose of this little modiﬁcation is to decrease the
order of dispersive operator which can be stiﬀ from the numerical point of view:
(1 −∂
xx
)u
t
+ u
x
+ uu
x
= 0 →u
t
+ (1 −∂
xx
)
−1
(u
x
+ uu
x
) = 0.
In practice, it can drastically change the CFL condition from ∆t = O(∆x
2
) for KdV (see
[Tre00] for example) to ∆t = O(∆x) for BBM equation (or even unconditionally stable
schemes if one considers an equivalent integral representation of solutions [BPS81, BPS85]).
We have to say that this operation does not change the formal order of approximation.
Moreover, the experiments show that the regularized equations (such as BBM) completed
by dissipative terms seem to be excellent in description of long water waves [BPS81].
The dispersion relation of just obtained equations (2.53), (2.54) is the following:
c
2
r
(k)
gh
:=
1 +
1
2
_
1
3
−α
2
_
_
1 +
_
1
2
−α
_
(kh)
2
__
1 + α
_
1 −
α
2
_
(kh)
2
_ =
= 1 −
1
3
(kh)
2
−
_
1
2
α
3
−
13
12
α
2
+
1
2
α −
1
6
_
(kh)
4
+O
_
(kh)
6
_
(2.55)
Here again we can choose the parameter α so that the dispersion relation (2.55) be exact
up to terms of order O
_
(kh)
4
_
. In this way we get the following equation for α:
1
2
α
3
−
13
12
α
2
+
1
2
α −
1
30
= 0.
This equation has only one root which has physical sense and lead to linearly wellposed
problem α
opt
:= 0.0800 (let us recall that α = 0 corresponds to the free surface and α = 1
to the bottom level). We suggest to use this value in numerical simulations. In Figure 2.8
we plot dispersion relations (2.50), (2.51) and (2.55) with optimal values of α.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
2.7 Bottom friction 107
0 1 2 3 4 5 6 7 8
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Phase velocities with optimal value of α
kh
c
p
(
k
)
/
(
g
h
)
1
/
2
Euler equations
Boussinesq−Nwogu model
Regularized equations
Figure 2.8: Dispersion relations of complete water wave problem, Nwogu version of Boussinesq
equations and its regularized version
2.7 Bottom friction
In this section, one switches back to dimensional variables. It is a common practice in
hydraulics engineering to take into account the eﬀect of bottom friction or bottom rugosity.
In the Boussinesq and nonlinear shallow water equations there is also a possibility to include
some kind of empirical terms to model these physical eﬀects. From the mathematical and
especially numerical viewpoints these terms do not add any complexity, since they have the
form of source terms that do not involve diﬀerential operators. So it is highly recommended
to introduce these source terms in numerical models.
There is no unique bottom friction law. Most frequently, Ch´ezy and DarcyWeisbach
laws are used. Both laws have similar structures. We give here these models in dimensional
form. The following terms have to be added to the source terms of Boussinesq equations
when one wants to include bottom friction modelling.
• Ch´ezy law:
S
f
= −C
f
g
u[u[
h + η
,
where C
f
is the Ch´ezy coeﬃcient.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
108 Dissipative Boussinesq equations
• DarcyWeisbach law:
S
f
= −
λu[u[
8(h + η)
,
where λ is the socalled resistance value. This parameter is determined according to
the simpliﬁed form of the ColebrookWhite relation:
1
√
λ
= −2.03 log
_
k
s
14.84(h + η)
_
.
Here k
s
denotes the friction parameter, which depends on the composition of the
bottom. Typically k
s
can vary from 1mm for concrete to 300mm for bottom with
dense vegetation.
• ManningStrickler law:
S
f
= −k
2
g
u[u[
(h + η)
4
3
,
where k is the Manning roughness coeﬃcient.
Remark 7 In chapter 4 we develop an alternative approach to bottom friction modelling.
Our analysis is based on formal asymptotic expansion of NavierStokes equations in ﬂuid
and in the bottom boundary layer. The resulting viscouspotential formulation contains
local and nonlocal dissipative terms. Then we derive corresponding long wave equations
which take the following form in the absence of dispersive terms:
η
t
+∇ ((h + η)u) = 2ν∆η +
_
ν
π
t
_
0
∇ u
√
t −τ
dτ,
u
t
+
1
2
∇[u[
2
+ g∇η = 2ν∆u.
The principal diﬀerence with just presented
engineer’s
approach consists in at least
formal justiﬁcation of the governing equations.
2.8 Spectral Fourier method
In this study we adopted a wellknown and widely used spectral Fourier method. The
main idea consists in discretizing the spatial derivatives using Fourier transforms. The
eﬀectiveness of this method is explained by two main reasons. First, the diﬀerentiation
operation in Fourier transform space is extremely simple due to the following property of
Fourier transforms: f
′
= ikf. Secondly, there are very powerful tools for the fast and
accurate computation of discrete Fourier transforms (DFT). So, spatial derivatives are
computed with the following algorithm:
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
2.8 Spectral Fourier method 109
1: f ←ﬀt (f)
2: v ←ikf
3: f
′
←iﬀt (v)
where k is the wavenumber.
This approach, which is extremely eﬃcient, has the drawbacks of almost all spectral
methods. The ﬁrst drawback consists in imposing periodic boundary conditions since we
use DFT. The second drawback is that we can only handle simple geometries, namely,
Cartesian products of 1D intervals. For the purpose of academic research, this type of
method is appropriate.
Let us now consider the discretization of the dissipative Boussinesq equations. We
show in detail how the discretization is performed on equations (2.23), (2.25). The other
systems are discretized in the same way. We chose equations (2.23), (2.25) in order to
avoid cumbersome expressions and make the description as clear as possible.
Let us apply the Fourier transform to both sides of equations (2.23), (2.25):
η
t
= −ik (h + εη)u −
1
ε
h
t
−
µ
2
2ε
h
2
∇
2
h
t
−
µ
2
ε
h∇h ∇h
t
+
µ
2
6
h
3
∇
2
∇ u
+
µ
2
2
h
2
∇
2
h∇ u + µ
2
h[∇h[
2
∇ u + µ
2
h
2
∇h ∇(∇ u), (2.56)
u
t
+
1
2
εik[u[
2
+ ikη + εν
2
[k[
2
u −
1
2
µ
2
ikh
2
∇ u
t
= 0, (2.57)
where k = (k
x
, k
y
) denotes the Fourier transform parameters.
Equations (2.56) and (2.57) constitute a system of ordinary diﬀerential equations to
be integrated numerically. In the present study we use the classical explicit fourthorder
RungeKutta method.
k
1
= ∆tf(t
n
, y
n
),
k
2
= ∆tf
_
t
n
+
∆t
2
, y
n
+
1
2
k
1
_
,
k
3
= ∆tf
_
t
n
+
∆t
2
, y
n
+
1
2
k
2
_
,
k
4
= ∆tf (t
n
+ ∆t, y
n
+ k
3
) ,
y
n+1
= y
n
+
1
6
k
1
+
1
3
k
2
+
1
3
k
3
+
1
6
k
4
+ O(∆t
6
).
Remark 8 A lot of researchers who integrated numerically the KdV equation noticed that
the stability criterion has the form
∆t =
λ
N
2
,
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
110 Dissipative Boussinesq equations
where λ is the CourantFriedrichsLewy (CFL) number and N the number of points of
discretization. In order to increase the time integration step ∆t they solved exactly the
linear part of the partial diﬀerential equation since the linear term is the one involving
high frequencies and constraining the stability. This method, which is usually called the
integrating factor method, allows an increase of the CFL number up to a factor ten, but it
cannot ﬁx the dependence on 1/N
2
.
We do not have this diﬃculty because we use regularized dispersive terms. The regular
ization eﬀect can be seen from equation (2.57). The same idea was exploited by [BPS81],
who used the modiﬁed KdV equation (2.10).
Let us brieﬂy explain how we treat the nonlinear terms. Since the time integration
scheme is explicit, one can easily handle nonlinearities. For example the term (h + εη)u is
computed as follows:
(h + εη)u = ﬀt
_
(h + ε Re iﬀt (η)) Re iﬀt (u)
_
.
The other nonlinear terms are computed in the same way.
2.8.1 Validation of the numerical method
One way to validate a numerical scheme is to compare the numerical results with
analytical solutions. Unfortunately, the authors did not succeed in deriving analytical
solutions to the (1 + 1)D dissipative Boussinesq equations over a ﬂat bottom. But for
validation purposes, one can neglect the viscous term. With this simpliﬁcation several
solitary wave solutions can be obtained. We follow closely the work of [Che98]. In (1+1)D
in the presence of a ﬂat bottom, the Boussinesq system without dissipation becomes
η
t
+ u
x
+ ε(uη)
x
−
µ
2
6
u
xxx
= 0, (2.58)
u
t
+ η
x
+ εuu
x
−
µ
2
2
u
xxt
= 0. (2.59)
We look for solitarywave solutions travelling to the left in the form
η(x, t) = η(ξ) = η(x
0
+ x + ct), u(x, t) = Bη(ξ),
where we introduced the new variable ξ = x
0
+x+ct and B, c, x
0
are constants. From the
physical point of view this change of variables is nothing else than Galilean transformation.
In other words we choose a new frame of reference which moves with the same celerity as
the solitary wave. Since c is constant (there is no acceleration), the observer moving with
the wave will see a steady picture.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
2.8 Spectral Fourier method 111
In the following primes denote derivation with respect to ξ. Substituting this special
form into the governing equations (2.58)(2.59) gives
cη
′
+ u
′
+ ε(uη)
′
−
µ
2
6
u
′′′
= 0,
cu
′
+ η
′
+ εuu
′
−c
µ
2
2
u
′′′
= 0.
One can decrease the order of derivatives by integrating once:
cη + u + εuη −
µ
2
6
u
′′
= 0,
cu + η +
ε
2
u
2
−c
µ
2
2
u
′′
= 0.
The solution is integrable on R and there are no integration constants, since a priori the
solution behaviour at inﬁnity is known: the solitary wave is exponentially small at large
distances from the crest. Mathematically it can be expressed as
lim
ξ→±∞
η(x, t) = lim
ξ→±∞
u(x, t) = 0.
Now we use the relation u(ξ) = Bη(ξ) to eliminate the variable u from the system:
(c + B)η −B
µ
2
6
η
′′
= −εBη
2
, (2.60)
(1 + cB)η −cB
µ
2
2
η
′′
= −
ε
2
B
2
η
2
. (2.61)
In order to have nontrivial solutions both equations must be compatible. Compatibility
conditions are obtained by comparing the coeﬃcients of corresponding terms in equations
(2.60)(2.61):
1
2
B
2
−
1
2
Bc = 1,
1
6
B
2
−Bc = 0.
These relations can be thought as a system of linear equations with respect to B
2
and Bc.
The unique solution of those equations is
B
2
=
12
5
, c =
B
6
.
Choosing B > 0 so that c > 0 leads to
B =
6
√
15
, c =
1
√
15
.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
112 Dissipative Boussinesq equations
These constants determine the amplitude and the propagation speed of the solitary wave.
In order to ﬁnd the shape of the wave, one diﬀerentiates once equation (2.61):
7η
′
−µ
2
η
′′′
= −12εηη
′
. (2.62)
The solution to this equation is wellknown (see for example [New77, Che98]):
Lemma 1 Let α, β be real constants; the equation
αη
′
(ξ) −βη
′′′
(ξ) = η(ξ)η
′
(ξ)
has a solitarywave solution if αβ > 0. Moreover, the solitarywave solution is
η(ξ) = 3αsech
2
_
1
2
_
α
β
(ξ + ξ
0
)
_
where ξ
0
is an arbitrary constant.
Applying this lemma to equation (2.62) yields the following solution:
η(x, t) = −
7
4ε
sech
2
_
√
7
2µ
(x + ct + x
0
)
_
, (2.63)
u(x, t) = −
7
√
15
10ε
sech
2
_
√
7
2µ
(x + ct + x
0
)
_
.
Note that this exact solitary wave solution is not physical. Indeed the velocity is
negative whereas one expects it to be positive for a depression wave propagating to the
left. In any case, the goal here is to validate the numerical computations by comparing
with an exact solution. The methodology is simple. We choose a solitary wave as initial
condition and let it propagate during a certain time T with the spectral method. At the end
of the computations one computes the L
∞
norm of the diﬀerence between the analytical
solution (2.63) and the numerical one ˜ η(x, T):
ǫ
N
:= max
1≤i≤N
[η(x
i
, T) − ˜ η(x
i
, T)[ ,
where ¦x
i
¦
1≤i≤N
are the discretization points.
Figure 2.9 shows the graph of ǫ
N
as a function of N. This result shows an excellent
performance of this spectral method with an exponential convergence rate. In general, the
error ǫ
N
is bounded below by the maximum between the error due to the time integration
algorithm and ﬂoating point arithmetic precision.
The exponential convergence rate to the exact solution is one of the features of spectral
methods. It explains the success of these methods in several domains such as direct nu
merical simulation (DNS) of turbulence. One of the main drawbacks of spectral methods
consists in the diﬃculties in handling complex geometries and various types of boundary
conditions.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
2.9 Numerical results 113
100 150 200 250 300 350 400 450 500 550
10
−12
10
−10
10
−8
10
−6
10
−4
10
−2
10
0
Error between exact and numerical solutions
N − degrees of freedom
ε
n
Figure 2.9: Error on the numerical computation of a solitary wave solution. Here T = 1.
2.9 Numerical results
In this section we perform comparisons between the two dissipation models (2.24) and
(2.25). Even though the computations we show deal with a 1D wave propagating in the
negative x−direction, they have been performed with the 2D version of the code. The
bathymetry z = −h(x, y) is chosen to be a regularized step function which is translated in
the y−direction. A typical function h(x, y) is given by
h(x) =
_
¸
_
¸
_
h
l
, x ≤ x
0
,
h
l
+
1
2
(h
r
−h
l
)
_
1 + sin
_
π
∆x
(x −x
0
−
1
2
∆x)
__
, x
0
< x < x
0
+ ∆x,
h
r
, x ≥ x
0
+ ∆x.
(2.64)
This test case is interesting from a practical point of view since it clearly illustrates the
phenomena of long wave reﬂection by bottom topography. The parameters used in this
computation are given in Table 2.2. All values are given in nondimensional form.
2.9.1 Construction of the initial condition
We propagate on the free surface a socalled approximate soliton. Its classical con
struction is as follows. We begin with the nondissipative Boussinesq equations on a ﬂat
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
114 Dissipative Boussinesq equations
parameter h
l
h
r
x
0
∆x ε µ ν
1
, ν
2
value 0.5 1.0 −0.5 0.3 0.005 0.06 0.14
Table 2.2: Typical values of the parameters used in the numerical computations
bottom:
η
t
+
_
(1 + εη)u
_
x
−
µ
2
6
u
xxx
= 0,
u
t
+ η
x
+
ε
2
(u
2
)
x
−
µ
2
2
u
xxt
= 0, (2.65)
and look for u in the following form:
u = −η + εP + µ
2
Q + O(ε
2
+ εµ
2
+ µ
4
). (2.66)
It is precisely at this step that one makes an approximation. One substitutes this asymp
totic expansion into the governing equations and retains only the terms of order O(ε+µ
2
):
η
t
−η
x
+ εP
x
+ µ
2
Q
x
−2εηη
x
+
µ
2
6
η
xxx
= O(ε
2
+ εµ
2
+ µ
4
), (2.67)
−η
t
+ η
x
+ εP
t
+ µ
2
Q
t
+ εηη
x
+
µ
2
2
η
xxt
= O(ε
2
+ εµ
2
+ µ
4
).
Add these two equations and set the coeﬃcients of ε and µ
2
equal to 0:
ε : P
x
+ P
t
−ηη
x
= 0, (2.68)
µ
2
: Q
x
+ Q
t
+
1
6
η
xxx
+
1
2
η
xxt
= 0. (2.69)
Since the water depth is h = 1 + εη = 1 + O(ε), the approximate solitary wave should
travel to the left with a celerity c = 1+O(ε) and depend on the variable x+ct = x+t+O(ε).
Consequently one has the following relations:
∂f
∂t
=
∂f
∂x
+ O(ε + µ
2
), f ∈ ¦η, P, Q¦.
Replacing time derivatives by spatial ones in (2.68)(2.69) yields
P
x
=
1
2
ηη
x
, Q
x
= −
1
3
η
xxx
.
By integration (using the fact that solitary waves tend to zero at inﬁnity), one obtains
P =
1
4
η
2
, Q = −
1
3
η
xx
(2.70)
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
2.9 Numerical results 115
and the relation (2.66) connecting η and u becomes
u = −η +
ε
4
η
2
−
µ
2
3
η
xx
+ . . . . (2.71)
Substituting this expression for u into (2.67) yields a classical KdV equation for η:
η
t
−
_
1 +
3
2
εη
_
η
x
−
µ
2
6
η
xxx
= 0, (2.72)
which admits solitary wave solutions of the form η = η(x + ct):
η(x, t) =
2(c −1)
ε
sech
2
_
1
2µ
_
6(c −1)(x + ct)
_
,
where c > 1. The velocity u is obtained from (2.71) by simple substitution. This approxi
mate soliton is used in the numerical computations.
2.9.2 Comparison between the dissipative models
The snapshot of the function η(x, y, t
0
) (divided by 10 for clarity’s sake) during and
just after reﬂection by the step is given on Figure 2.10.
Then we compare the two sets of equations (2.23), (2.24) and (2.23), (2.25). To do so
we look at the section of the free surface at y = 0 along the propagation direction.
Figure 2.11 shows that even at the beginning of the computations the two models give
slightly diﬀerent results. The amplitude of the pulse obtained with model I is smaller.
It can be explained by the presence of the term ν
1
Su which is bigger in magnitude than
εν
2
∇
2
u. Within graphical accuracy, there is almost no diﬀerence between the conservative
case and model II.
In Figure 2.12 one can see that diﬀerences between the two solitons continue to grow. In
particular we see an important drawback of the dissipation model I: just after the wave crest
the free surface has some kind of residual deformation which is clearly nonphysical. Our
numerical experiments show that the amplitude of this residue depends almost linearly on
the parameter ν
1
. We could hardly predict this eﬀect directly from the equations without
numerical experiments.
We would like to point out several soliton transformations in Figure 2.13 due to the
interaction with bathymetry. First of all, since the depth decreases, the wave amplitude
grows. Quantitatively speaking, the wave amplitude before the interaction is equal exactly
to 8 (without including dissipation) and over the step it becomes roughly 9.4. On the
other hand the soliton becomes less symmetric which is also expected. Because of periodic
numerical boundary conditions we also observe the residue of the freesurface deformation
coming through the left boundary.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
116 Dissipative Boussinesq equations
(a) t = 2.1s (b) t = 2.5s
Figure 2.10: Interaction between a leftrunning solitary wave and a step.
−3 −2 −1 0 1 2 3
−1
0
1
2
3
4
5
6
7
8
9
10
t = 0.400
x
η
(
x
,
0
)
No dissipation
Model 1
Model 2
−4 −3 −2 −1 0 1 2 3 4
−0.08
−0.06
−0.04
−0.02
0
0.02
0.04
0.06
0.08
0.1
x
η
2
−
η
1
Difference between model 2 and 1 at t = 0.4
Figure 2.11: Freesurface snapshot before the interaction with the step: (left) the curves corre
sponding to the three models are almost superimposed; (right) diﬀerence between model II and
model I.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
2.9 Numerical results 117
−3 −2 −1 0 1 2 3
−1
0
1
2
3
4
5
6
7
8
9
10
t = 1.100
x
η
(
x
,
0
)
No dissipation
Model 1
Model 2
Figure 2.12: Free surface just before the interaction with the step
−3 −2 −1 0 1 2 3
−1
0
1
2
3
4
5
6
7
8
9
10
t = 2.100
x
η
(
x
,
0
)
No dissipation
Model 1
Model 2
Figure 2.13: Beginning of the solitary wave deformation under the change in bathymetry
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
118 Dissipative Boussinesq equations
−3 −2 −1 0 1 2 3
−1
0
1
2
3
4
5
6
7
8
9
10
t = 2.250
x
η
(
x
,
0
)
No dissipation
Model 1
Model 2
Figure 2.14: Initiation of the reﬂected wave separation
−3 −2 −1 0 1 2 3
−1
0
1
2
3
4
5
6
7
8
9
10
t = 2.800
x
η
(
x
,
0
)
No dissipation
Model 1
Model 2
Figure 2.15: Separation of the reﬂected wave
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
2.10 Conclusions 119
−3 −2 −1 0 1 2 3
−1
0
1
2
3
4
5
6
7
8
9
10
t = 3.200
x
η
(
x
,
0
)
No dissipation
Model 1
Model 2
Figure 2.16: Two separate waves moving in opposite directions
Figures 2.14, 2.15 and 2.16 show the process of wave reﬂection from the step at the
bottom. The reﬂected wave clearly moves in the opposite direction.
2.10 Conclusions
Comparisons have been made between two dissipation models. Model II, in which the
decay is proportional to the second derivative of the velocity, appears to be better. At this
stage we cannot show comparisons with laboratory experiments in order to demonstrate
the performance of model II. Nevertheless, there is an indirect evidence. We refer one
more time to the theoretical as well as experimental work of [BPS81]. In order to model
wave trains, they added to the Korteweg–de Vries equation an adhoc dissipative term
in the form of the Laplacian (but in 1D). This term coincides with the results of our
derivation if we model dissipation in the equations according to the second model. Their
work shows excellent agreement between experiments and numerical solutions to dissipative
KdV equation. Moreover our dissipative Boussinesq equations are in the same relationship
with the classical Boussinesq equations [Per67] as Euler and NavierStokes equations. This
is a second argument towards the physical pertinency of the results obtained with model
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
120 Dissipative Boussinesq equations
II.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
Chapter 3
Two phase approach to free surface
compressible ﬂows
I think that there is a moral to this story, namely that it is more important to have beauty in
one’s equations that to have them ﬁt experiment. [...] It seems that if one is working from the
point of view of getting beauty in one’s equations, and if one has really a sound insight, one is
on a sure line of progress. If there is not complete agreement between the results of one’s work
and experiment, one should not allow oneself to be too discouraged, because the discrepancy
may well be due to minor features that are not properly taken into account and that will get
cleared up with further development of the theory.
Paul Adrien Maurice Dirac (1902 – 1984)
Contents
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
3.2 Mathematical model . . . . . . . . . . . . . . . . . . . . . . . . . 125
3.2.1 Sound speed in the mixture . . . . . . . . . . . . . . . . . . . . . 126
3.2.2 Equation of state . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
3.3 Formal limit in barotropic case . . . . . . . . . . . . . . . . . . 129
3.4 Finite volume scheme on unstructured meshes . . . . . . . . . 131
3.4.1 Sign matrix computation . . . . . . . . . . . . . . . . . . . . . . 136
3.4.2 Second order scheme . . . . . . . . . . . . . . . . . . . . . . . . . 137
3.4.3 TVD and MUSCL schemes . . . . . . . . . . . . . . . . . . . . . 138
3.4.4 Diﬀusive ﬂuxes computation . . . . . . . . . . . . . . . . . . . . . 144
3.4.5 Solution interpolation to mesh nodes . . . . . . . . . . . . . . . . 146
121
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
122 Two phase ﬂows
3.4.6 Time stepping methods . . . . . . . . . . . . . . . . . . . . . . . 148
3.4.7 Boundary conditions treatment . . . . . . . . . . . . . . . . . . . 151
3.5 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . 153
3.5.1 Convergence test . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
3.5.2 Falling water column . . . . . . . . . . . . . . . . . . . . . . . . . 157
3.5.3 Water drop test case . . . . . . . . . . . . . . . . . . . . . . . . . 157
3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
3.1 Introduction
Twophase or even multiphase ﬂows are very common in nature. They occur in a
system containing gas and liquid
1
with a meniscus separating the two phases. Probably,
the most important applications of twophase ﬂow are in largescale power systems. The
design of power stations requires a detailed understanding of twophase ﬂow heattransfer
and pressure drop behaviour, which is signiﬁcantly diﬀerent from the singlephase case.
Even more critically, nuclear reactors use water to remove heat from the reactor core using
twophase ﬂow. Another important application of twophase ﬂows is pump cavitation.
Twophase ﬂows in nature include many interesting examples such as clouds, groundwater
ﬂow, bubbles, rain, waves on the sea, foam, fountains and gas/oil slicks.
Several features make twophase ﬂow an interesting and challenging branch of ﬂuid
mechanics. First of all, surface tension complicates considerably all dynamical problems.
In the present study we do not take into account this eﬀect since in our applications it is
unimportant. Then, in the case of air and water at standard temperature and pressure, the
density of the two phases diﬀers by a factor of about 1000. Similar diﬀerences are typical
of water liquid/water vapor densities and can make the computation very stiﬀ. The sound
speed changes dramatically for materials undergoing phase change, and can be orders of
magnitude diﬀerent (see Figure 3.2). This introduces strong compressible eﬀects into the
problem. Finally, the phase changes are not instantaneous, and the liquid vapor system
will not necessarily be in phase equilibrium.
In the present chapter we are going to present a mathematical model which is issued
from twophase ﬂows ﬂuid mechanics. Namely, we are going to use the homogeneous model
(with equal velocity, pressure and temperature for both phases). Basically, our contribution
consists in suggesting to use this model for the numerical simulation of freesurface aerated
ﬂows in various containers. The most important applications are issued from petroleum
1
They can occur even in a system containing two diﬀerent liquids as well.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.1 Introduction 123
Figure 3.1: Example of homogeneous water/air mixture just after the wave breaking. This
image is taken from [BOPB07].
industry and simulation of wave impacts onto a wall or breakwater. Figure 3.1 shows an
image of a broken wave. It gives an example of homogeneous water/air mixture and shows
the physical situation which we aim to model.
Accurate predictions of wave loading are crucial to the design of coastal structures.
When a wave breaks directly onto a vertical or inclined wall, impact pressures can be
extremely large in comparison to the pressures exerted by nonbreaking waves. Despite all
this, many engineers disregard wave impact loading in the belief that the duration of the
impulse is too short for the structure to respond. Presently, the loading due to breaking
waves is diﬃcult to predict and is poorly understood in general. We refer to [Per03] as
general excellent review of waterwave impact problem.
Interesting experimental results are obtained in [BOPB07]. The authors report that
even for series of supposedly regular waves, there is great variability in the violence of
the impacts. The greatest impact pressures are highly localized in space and time. They
are produced when the front of the breaker is almost parallel to the wall at the instant
of impact. If the wave overturns as it hits the wall, it can trap an air pocket. The
compressibility of the trapped or entrained air will aﬀect the dynamics and is often thought
to reduce the maximum pressure due to a cushioning eﬀect. However, an air pocket will
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
124 Two phase ﬂows
0 0.2 0.4 0.6 0.8 1
0
500
1000
1500
Sound velocity in the water/air mixture
α, gas void fraction
c
s
,
m
/
s
Figure 3.2: Sound speed c
s
in the water/air mixture at normal conditions. It is plotted as a
function of gas void fraction α.
also tend to distribute the impact pressures more widely so that the overall force on the
wall may not be reduced.
Wave impacts are classiﬁed as lowaeration when the measurements indicate that the
water adjacent to the wall contains relatively little air (typically a voids ratio < 5%). On
the other hand, an impact is called a highaeration impact if it contains a higher level
of entrained air, sometimes accompanied by clear evidence of entrapment. Bullock et
al. [BOPB07] showed that lowaeration impacts have temporally and spatially localised
pressure spike of short rise time and duration (≈ 80 to 200 ms) while in highaeration
situation the rise time, fall time and duration are longer (≈ 100 to 400 ms) and the
pressure spike is less localised.
Even when the pressures during a highaeration impact are lower, the fact that the
impact is generally less spatially localised than a lowaeration impact reduces the chance
of the resultant force being lower. When this is combined with the longer duration, the
impulse associated with the impact may well be higher. It suggests that the greater per
sistence of air in seawater waves may not help to protect structures from impact damage
as is often assumed.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.2 Mathematical model 125
3.2 Mathematical model
We would like to begin this section by explaining our choice of a compressible model
for the liquid phase. If the liquid is pure (water, for example), the importance of its
compressibility is questionable
2
. In the present work we aim to model the mixtures of
gas and liquid which typically occur in oil containers and broken waves (see Figure 3.1).
In order to show the importance of compressibility eﬀects, we depict on Figure 3.2 the
dependence of sound velocity in the mixture of water and air as a function of gas volume
fraction at normal conditions. The reader can see that the sound velocity drops down very
quickly (to about 50
m
s
) when we add a small quantity of air and the Mach number is not
so small anymore. Recall that according to [BOPB07] lowaerated impacts contain about
5% of air, thus dropping down signiﬁcantly the sound speed.
In this section we present the equations which govern the motion of two phase mixtures
in a computational domain Ω. First of all, we need to introduce the notation which will
be used throughout this chapter. We use superscripts ± to denote any quantity which is
related to liquid and gas respectively. For example, α
+
and α
−
denote the volume fraction
of liquid and gas and obviously satisfy the condition α
+
+ α
−
= 1. Then, we have the
following classical quantities: ρ
±
, u, p, e, E, g which denote the density of each phase, the
velocity ﬁeld vector, the pressure, the internal and total energy and the acceleration due
to gravity correspondingly.
Conservation of mass (one equation for each phase), momentum and energy lead to the
four following equations:
(α
±
ρ
±
)
t
+∇ (α
±
ρ
±
u) = 0, (3.1)
(ρu)
t
+∇
_
ρu ⊗u + pI
_
= ρg, (3.2)
_
ρE
_
t
+∇
_
ρHu
_
= ρg u, (3.3)
where ρ := α
+
ρ
+
+ α
−
ρ
−
(the total density), H := E +
p
ρ
(the speciﬁc enthalpy), E =
e +
1
2
[u[
2
(the total energy). This system can be seen as the single energy and inﬁnite drag
limit of the more conventional six equations model [Ish75].
The above system contains ﬁve unknowns α
±
ρ
±
, u, p and E and only four governing
equations (3.1)  (3.3). In order to close the system, we need to provide the socalled
equation of state (EOS) p = p
±
(ρ
±
, e). The construction of the EOS will be discussed
below (see Section 3.2.2).
It is possible to rewrite these equations as a system of balance laws
∂w
∂t
+∇ T(w) = o(w),
2
More precisely, if the problem under consideration does not require explicitly the compressibility (like
in acoustics, for example), this eﬀect can usually be neglected since it complicates considerably the study.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
126 Two phase ﬂows
where the conservative variables in the 2D case are deﬁned as follows:
w = (w
i
)
5
i=1
:= (α
+
ρ
+
, α
−
ρ
−
, ρu, ρv, ρE). (3.4)
The ﬂux projection on the normal direction n = (n
1
, n
2
) can be expressed in physical and
conservative variables
T n = (α
+
ρ
+
u
n
, α
−
ρ
−
u
n
, ρuu
n
+ pn
1
, ρvu
n
+ pn
2
, ρHu
n
) =
_
w
1
w
3
n
1
+ w
4
n
2
w
1
+ w
2
, w
2
w
3
n
1
+ w
4
n
2
w
1
+ w
2
, w
3
w
3
n
1
+ w
4
n
2
w
1
+ w
2
+ pn
1
,
w
4
w
3
n
1
+ w
4
n
2
w
1
+ w
2
+ pn
2
, (w
5
+ p)
w
3
n
1
+ w
4
n
2
w
1
+ w
2
_
(3.5)
where u
n
:= u n = un
1
+ vn
2
is the velocity projection on the normal direction n. The
jacobian matrix A
n
(w) :=
∂(F·n)(w)
∂w
can be easily computed. Its expression in the physical
variables is
A
n
=
_
_
_
_
_
_
_
u
n
c
−
−u
n
c
+
c
+
n
1
c
+
n
2
0
−u
n
c
−
u
n
c
+
c
−
n
1
c
−
n
2
0
−uu
n
+
∂p
∂w
1
n
1
−uu
n
+
∂p
∂w
2
n
1
u
n
+ un
1
+
∂p
∂w
3
n
1
un
2
+
∂p
∂w
4
n
1
∂p
∂w
5
n
1
−vu
n
+
∂p
∂w
1
n
2
−vu
n
+
∂p
∂w
2
n
2
vn
1
+
∂p
∂w
3
n
2
u
n
+ vn
2
+
∂p
∂w
4
n
2
∂p
∂w
5
n
2
u
n
_
∂p
∂w
1
−H
_
u
n
_
∂p
∂w
2
−H
_
u
n
∂p
∂w
3
+ Hn
1
u
n
∂p
∂w
4
+ Hn
2
u
n
_
1 +
∂p
∂w
5
_
_
_
_
_
_
_
_
This matrix has three distinct eigenvalues:
λ
1
= u
n
−c
s
, λ
2,3,4
= u
n
, λ
5
= u
n
+ c
s
,
where c
s
is deﬁned in the next section. One can conclude that the system (3.1) – (3.3)
is hyperbolic. This hyperbolicity represents the major advantage of this model. The
computation of the eigenvectors is trickier but can still be performed analytically. We do
not give here the ﬁnal expressions since they are cumbersome.
3.2.1 Sound speed in the mixture
In [Ghi08] it is shown by direct computation of the eigenvalues of the Jacobian A
n
(w)
that the square of the sound speed in the mixture is given by
c
2
s
:=
ρ
+
ρ
−
(c
+
s
)
2
(c
−
s
)
2
ρ(α
+
ρ
−
(c
−
s
)
2
+ α
−
ρ
+
(c
+
s
)
2
)
,
where c
±
s
are the sound velocities in each phase. In the present setting they are deﬁned by
formulas (3.8).
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.2 Mathematical model 127
3.2.2 Equation of state
In the present work we assume that the light ﬂuid is described by an ideal gas type law
p
−
= (γ −1)ρ
−
e
−
, e
−
= c
−
v
T
−
, (3.6)
while the heavy ﬂuid is modeled by Tate’s law
3
[GZI
+
79, HA71]
p
+
+ π
0
= (^ −1)ρ
+
e
+
, e
+
= c
+
v
T
+
+
π
0
^ρ
+
, (3.7)
where the quantities γ, c
±
v
, ^, π
0
are constants. For example, pure water is well described
when we take ^ = 7 and π
0
= 2.1 10
9
Pa.
Remark 9 In practice, the constants c
±
v
can be calculated after simple algebraic manipula
tions of equations (3.6), (3.7) and matching with experimental values at normal conditions:
c
−
v
≡
p
0
(γ −1)ρ
−
0
T
0
,
c
+
v
≡
^p
0
+ π
0
(^ −1)^ρ
+
0
T
0
.
For example, for an air/water mixture under normal conditions we have the values given
in Table 3.1.
The sound velocities in each phase are given by the following formulas:
(c
−
s
)
2
=
γp
−
ρ
−
, (c
+
s
)
2
=
^p
+
+ π
0
ρ
+
. (3.8)
In order to construct an equation of state for the mixture, we make the additional
assumption that the two phases are in thermodynamic equilibrium:
p
+
= p
−
, T
+
= T
−
. (3.9)
Below values of the common pressure and common temperature will be denoted by p and
T respectively.
Now, let us construct a system of nonlinear equations which will help us obtain the
pressure in the mixture. To do this, we ﬁrst introduce the following notation for the mass
fractions of each phase in the mixture
c
±
:=
α
±
ρ
±
ρ
.
3
In the literature Tate’s law is sometimes called the stiﬀened gas law.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
128 Two phase ﬂows
parameter value
p
0
10
5
Pa
ρ
+
0
10
3
kg/m
3
ρ
−
0
1.29 kg/m
3
T
0
300 K
γ 1.4
^ 7
π
0
2.1 10
9
Pa
c
+
v
166.72
J
kg·K
c
−
v
646.0
J
kg·K
g 100
m
/
s
2
Table 3.1: Values of the parameters for an air/water mixture under normal conditions. We
would like to comment on the rather high value of the acceleration due to gravity. Obviously,
it does not correspond to any physical situation. Nevertheless, we choose this value in order to
accelerate all dynamic processes in our test cases.
We would like to reduce the number of unknowns. To do this, we express the internal
energies e
±
of each phase as a function of the other thermodynamic variables and constants.
Let us write the two following relations:
c
+
e
+
+ c
−
e
−
= e, T
+
(ρ
+
, e
+
) = T
−
(ρ
−
, e
−
) .
After substituting the expressions (3.6), (3.7) for T
±
, our system becomes
c
+
e
+
+ c
−
e
−
= e,
e
+
c
+
v
−
e
−
c
−
v
=
π
0
^c
+
v
ρ
+
.
This linear system can be easily solved to give us the expressions for e
±
e
+
=
1
T
_
e
c
−
v
+
π
0
c
−
^c
+
v
ρ
+
_
, (3.10)
e
−
=
1
Tc
+
v
_
e −
π
0
c
+
^ρ
+
_
, (3.11)
where T :=
c
+
c
−
v
+
c
−
c
+
v
.
Let us analyse the obtained results. From (3.9) we have an additional relation
p ≡ p
+
(ρ
+
, e
+
) = p
−
(ρ
−
, e
−
).
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.3 Formal limit in barotropic case 129
It means that the pressure p is completely determined if we know (ρ
+
, e
+
) or (ρ
−
, e
−
). Since
internal energies e
±
are expressed in terms of (α
±
, ρ
±
) by expressions (3.10), (3.11), it is
suﬃcient to determine these four quantities (α
±
, ρ
±
). It is done by solving the following
system of nonlinear equations:
(^ −1)ρ
+
_
e
c
−
v
+
π
0
c
−
^c
+
v
ρ
+
_
−(γ −1)
ρ
−
c
+
v
_
e −
π
0
c
+
^ρ
+
_
= Tπ
0
,
α
+
+ α
−
= 1,
α
+
ρ
+
= w
1
,
α
−
ρ
−
= w
2
.
The last two equations can be easily eliminated to reduce the system to two equations
(^ −1)ρ
+
_
e
c
−
v
+
π
0
c
−
^c
+
v
ρ
+
_
−(γ −1)
ρ
−
c
+
v
_
e −
π
0
c
+
^ρ
+
_
= Tπ
0
,
w
1
ρ
+
+
w
2
ρ
−
= 1.
Finally, using the second equation, one can express ρ
−
in terms of ρ
+
and the conservative
variables w
1,2
as
ρ
−
=
ρ
+
w
2
ρ
+
−w
1
.
This expression is then substituted in the pressure equality condition to give the following
single equation with respect to ρ
+
:
(^ −1)
e
c
−
v
ρ
+
(ρ
+
−w
1
) +
(^ −1)π
0
c
−
^c
+
v
(ρ
+
−w
1
)
−(γ −1)
e
c
+
v
ρ
+
w
2
+ (γ −1)
π
0
c
+
^c
+
v
w
2
= Tπ
0
(ρ
+
−w
1
).
This equation can be solved analytically (we do not give here the solution’s expression
since it is cumbersome) or by a Newtontype numerical method. Once ρ
+
is determined,
we can ﬁnd e
+
from (3.10) and, ﬁnally, the pressure
p = (^ −1)ρ
+
e
+
−π
0
.
3.3 Formal limit in barotropic case
The propagation of waves at the interface between a heavy compressible ﬂuid of density
ρ
+
and a light compressible ﬂuid of density ρ
−
is investigated. The superscript + is used to
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
130 Two phase ﬂows
denote the heavy ﬂuid while the superscript − is used to denote the light ﬂuid. A typical
geometry is shown in Figure 3.3.
x, y
z
z = η(x, t)
0
−: light ﬂuid
+: heavy ﬂuid
Figure 3.3: Sketch of the ﬂow.
The governing equations are
(α
±
ρ
±
)
t
+∇ (α
±
ρ
±
u) = 0 , (3.12)
(ρu)
t
+∇ (ρu ⊗u) +∇p = ρg , (3.13)
where u = (u, v, w) is the common velocity, α
±
are the volume fractions of each ﬂuid, ρ
±
the densities of each ﬂuid, with the relations α
+
+ α
−
= 1, ρ := α
+
ρ
+
+ α
−
ρ
−
and the
equation of state reads:
p ≡ p
+
(ρ
+
) = p
−
(ρ
−
) . (3.14)
The speeds of sound are given by
(c
±
s
)
2
=
∂p
∂ρ
±
.
Let
γ
±
=
ρ
±
(c
±
s
)
2
p
, γ = α
+
γ
−
+ α
−
γ
+
.
Let also
δ
±
γ = γ
±
−γ
∓
.
Then one can show that
α
±
t
+ (u ∇)α
±
+ α
+
α
−
δ
±
γ
γ
∇ u = 0 , (3.15)
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.4 Finite volume scheme on unstructured meshes 131
p
t
+ (u ∇)p + p
γ
+
γ
−
γ
∇ u = 0 . (3.16)
Consider the case where α
−
is either 0 or 1. More precisely let
α
−
:= H(z −η(x, t)) , (3.17)
where H is the Heaviside step function. Physically this substitution means that we consider
two pure phases separated by an interface. It follows that
α
+
α
−
= 0 .
Substituting the expression (3.17) into the equation (3.15) for α
−
gives
η
t
+ (u, v) ∇
x
η = w.
Equations (3.12) and (3.13) become
ρ
t
+ (u ∇)ρ + ρ∇ u = 0 ,
u
t
+ (u ∇)u +
∇p
ρ
= g .
This system of equations is nothing else than the system of a discontinuous twoﬂuid system
with an interface at z = η(x, t).
This simple computation shows an interesting property of our model. For the important
special case of barotropic equations, we can show that our model automatically degenerates
into a discontinuous twoﬂuid system where two pure phases are separated by an interface.
3.4 Finite volume scheme on unstructured meshes
Finite volume methods are a class of discretization schemes that have proven highly
successful in solving numerically of a wide class of conservation law systems. These systems
often come from compressible ﬂuid dynamics. In electromagnetism, for example, Discontin
uous Galerkin methods have proven to be more eﬃcient [CLS04]. When compared to other
discretization methods such as ﬁnite elements or ﬁnite diﬀerences, the primary interests of
ﬁnite volume methods are robustness, applicability on very general unstructured meshes,
and the intrinsic local conservation properties. Hence, with this type of discretization, we
conserve “exactly” the mass, momentum and total energy
4
.
In order to solve numerically the system of balance laws (3.1) – (3.3) we rewrite the
governing equations in the following form
∂w
∂t
+∇ T(w) = ∇ (T∇w) +o(w), (3.18)
4
This statement is true in the absence of source terms and appropriate boundary conditions.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
132 Two phase ﬂows
where w(x, t) : R
d
R
+
→ R
m
is the vector of conservative variables (in the present
study d = 2 or 3 and m = 5), T(w) is the advective ﬂux function, T the diﬀusion
matrix and o(w) the source term. We must say that the governing equations (3.1) – (3.3)
considered in this chapter do not possess viscous terms. However, we prefer to include
in the discussion their discretization since it can be useful for NavierStokes or viscous
SaintVenant and Boussinesq equations [DD07a]. The system (3.18) should be provided
with initial condition
w(x, 0) = w
0
(x) (3.19)
and appropriate boundary conditions.
O
K
∂K
n
KL
L
Figure 3.4: An example of control volume K with barycenter O. The normal pointing from K
to L is denoted by n
KL
.
The computational domain Ω ⊂ R
d
is triangulated into a set of non overlapping control
volumes that completely cover the domain. Let T denote a tesselation of the domain Ω
with control volume K such that
∪
K∈T
¯
K =
¯
Ω,
¯
K := K ∪ ∂K.
For two distinct control volumes K and L in T , the intersection is either an edge (2D) or
face (3D) with oriented normal n
KL
or else a set of measure at most d −2 (in 2D it is just
a vertex, in 3D it can also be a segment, for example). We need to introduce the following
notation for the neighbourhood of K
^(K) := ¦L ∈ T : area(K ∩ L) ,= 0¦ ,
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.4 Finite volume scheme on unstructured meshes 133
a set of all control volumes L which share a face (or an edge in 2D) with the given volume
K. In this chapter, we denote by vol() and area() the d and d −1 dimensional measures
5
respectively.
The choice of control volume tesselation is ﬂexible in the ﬁnite volume method. In the
present study we retained a socalled cellcentered approach (see Figure 3.5), which means
that degrees of freedom are associated to cell barycenters. There exists an alternative
vertexcentered method [BJ89, BO04] (see Figure 3.6) which necessitates the construction
of dual mesh even for ﬁrstorder schemes. In the cellcentered ﬁnite volume scheme, the
triangles themselves serve as control volumes with solution unknowns attributed to trian
gles barycenters. In the vertexcentered ﬁnite volume scheme, control volumes are formed
as a geometric dual to the triangle complex and solution unknowns stored on vertex basis.
Remark 10 Except for the construction of dual mesh in vertexcentered approach, these
two methods are almost equivalent in the interior of the computational domain Ω. However,
the boundary conditions treatment is diﬀerent and it is harder (or less natural in author’s
opinion) when data is stored at vertices. This is one more reason why we retained the cell
centers to store the solution’s information.
Storage location
Control volume
Figure 3.5: Illustration for cellcentered ﬁnite volume method
The ﬁrst steps in Finite Volume (FV) methods are classical. We start by integrating
equation (3.18) on the control volume K (see Figure 3.4 for illustration) and we apply
GaussOstrogradsky theorem for advective and diﬀusive ﬂuxes. Then, in each control
5
In other words, in 3D the notation area() and vol() are very natural and mean area and volume
respectively, while in 2D they refer to the area and the length.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
134 Two phase ﬂows
Storage location
Control volume
Figure 3.6: Illustration for vertexcentered ﬁnite volume method
volume, an integral conservation law statement is imposed:
d
dt
_
K
w dΩ +
_
∂K
T(w) n
KL
dσ =
_
∂K
(T∇w) n
KL
dσ +
_
K
o(w) dΩ . (3.20)
Physically an integral conservation law asserts that the rate of change of the total amount
of a substance (for example: mass, momentum, total energy, etc) with density w in a ﬁxed
control volume K is balanced by the ﬂux T, diﬀusion T of the substance through the
boundary ∂K and the production of this quantity o inside the control volume.
The next step consists in introducing the socalled control volume cell average for each
K ∈ T
w
K
(t) :=
1
vol(K)
_
K
w(x, t) dΩ .
After the averaging step, the ﬁnite volume method can be interpreted as producing a
system of evolution equations for cell averages, since
d
dt
_
K
w(x, t) dΩ = vol(K)
dw
K
dt
.
Godunov was the ﬁrst [God59] who pursued and applied these ideas to the discretization
of the gas dynamics equations.
However, the averaging process implies piecewise constant solution representation in
each control volume with value equal to the cell average. The use of such representation
renders the numerical solution multivalued at control volume interfaces. Thereby the calcu
lation of the ﬂuxes
_
∂K
(T(w)n
KL
) dσ at these interfaces is ambiguous. Next fundamental
aspect of ﬁnite volume methods is the idea of substituting the true ﬂux at interfaces by a
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.4 Finite volume scheme on unstructured meshes 135
numerical ﬂux function
_
T(w) n
_¸
¸
∂K∩∂L
←− Φ(w
K
, w
L
; n
KL
) : R
m
R
m
→R
m
,
a Lipschitz continuous function of the two interface states w
K
and w
L
. The heart of the
matter in ﬁnite volume method consists in the choice of the numerical ﬂux function Φ. In
general this function is calculated as an exact or even better approximate local solution
of the Riemann problem posed at these interfaces. In the present study we decided to
choose the numerical ﬂux function according to VFFC scheme extensively described in
Appendix A.
The numerical ﬂux is assumed to satisfy the properties:
Conservation. This property ensures that ﬂuxes from adjacent control volumes sharing
an interface exactly cancel when summed. This is achieved if the numerical ﬂux
function satisﬁes the identity
Φ(w
K
, w
L
; n
KL
) = −Φ(w
L
, w
K
; n
LK
).
Consistency. The consistency is obtained when the numerical ﬂux with identical state
arguments (in other words it means that the solution is continuous through an inter
face) reduces to the true ﬂux of the same state, i.e.
Φ(w, w; n) = (T n)(w).
After introducing the cell averages w
K
and numerical ﬂuxes into (3.20), the integral
conservation law statement becomes
dw
K
dt
+
L∈N(K)
area(L ∩ K)
vol(K)
Φ(w
K
, w
L
; n
KL
) =
=
1
vol(K)
L∈N(K)
_
K∩L
(T∇w) n
KL
dσ +
1
vol(K)
_
K
o(w) dΩ .
The discretization of diﬀusive ﬂuxes will be discussed in Section 3.4.4. We denote by o
K
the approximation of the following quantity
1
vol(K)
_
K
o(w) dΩ. In practice, the source
term discretization is discussed in Section A.1.3. Thus, the following system of ordinary
diﬀerential equations (ODE) is called a semidiscrete ﬁnite volume method:
dw
K
dt
+
L∈N(K)
area(L ∩ K)
vol(K)
_
Φ(w
K
, w
L
; n
KL
) −(T∇w) n
KL
_
= o
K
, ∀K ∈ T . (3.21)
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
136 Two phase ﬂows
The initial condition for this system is given by projecting (3.19) onto the space of piecewise
constant functions
w
K
(0) =
1
vol K
_
K
w
0
(x) dΩ .
This system of ODE should also be discretized. There is a variety of explicit and implicit
time integration methods. Let w
n
K
denote a numerical approximation of the cell average
solution in the control volume K at time t
n
= n∆t. The simplest time integration method
is the forward Euler scheme
dw
K
dt
∼
=
w
n+1
K
−w
n
K
∆t
.
When applied to (3.21) it produces the fullydiscrete ﬁnite volume scheme. The time
integration approach used in this study is detailed in Section 3.4.6.
3.4.1 Sign matrix computation
In the context of the VFFC scheme (see Appendix A for more details), we need to
compute the socalled sign matrix which is deﬁned in the following way
U
n
:= sign(A
n
) = Rsign(Λ)L,
where R, L are matrices composed of right and left eigenvectors correspondingly, and
Λ = diag(λ
1
, . . . , λ
m
) is the diagonal matrix of eigenvalues of the Jacobian.
This deﬁnition gives the ﬁrst “direct” method of sign matrix computation. Since the
advection operator is relatively simple, after a few tricks, we can succeed in computing
analytically the matrices R and L. For more complicated twophase models it is almost
impossible to perform this computation in closed analytical form. In this case, one has
to apply numerical techniques for eigensystem computation. It turns out to be costly and
quite approximative. In the present work we use physical information about the model in
numerical computations.
There is another way which is less expensive. The main idea is to construct a kind of
interpolation polynomial which takes the following values
P(u
n
±c
s
) = sign(u
n
±c
s
), P(u
n
) = sign(u
n
).
These three conditions allow us to construct a second degree interpolation polynomial.
Obviously, when P(λ) is evaluated with λ = A
n
we obtain the sign matrix U
n
as a result.
The construction of Lagrange interpolation polynomial P(λ) is simple and we give only its
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.4 Finite volume scheme on unstructured meshes 137
expression
P(λ) =
_
sign(u
n
+ c
s
) −2 sign(u
n
) + sign(u
n
−c
s
)
2c
2
s
(λ −u
n
)+
sign(u
n
) −sign(u
n
−c
s
)
c
s
_
(λ −u
n
+ c
s
) + sign(u
n
−c
s
).
In our research code we have implemented both methods. Our experience shows that
the interpolation method is quicker and gives correct results in most test cases. However,
when we approach pure phase states, it shows a rather bad numerical behaviour. It can
lead to instabilities and diminish overall code robustness. Thus, whenever possible we
suggest to use the computation of the Jacobian eigenstructure.
3.4.2 Second order scheme
If we analyze the above scheme, we understand that in fact, we have only one degree
of freedom per data storage location. Hence, it seems that we can expect to be ﬁrst
order accurate at most. In the numerical community ﬁrst order schemes are generally
considered to be too inaccurate for most quantitative calculations. Of course, we can
always make the mesh spacing extremely small but it cannot be a solution since it makes
the scheme ineﬃcient. From the theoretical point of view the situation is even worse since
an O(h
1
2
) L
1
norm error bound for the monotone and Eﬂux schemes [Osh84] is known
to be sharp [Pet91], although an O(h) solution error is routinely observed in numerical
experiments. On the other hand, Godunov has shown [God59] that all linear schemes that
preserve solution monotonicity are at most ﬁrst order accurate. This rather negative result
suggests that a higher order accurate scheme has to be essentially nonlinear in order to
attain simultaneously a monotone resolution of discontinuities and high order accuracy in
continuous regions.
A signiﬁcant breakthrough in the generalization of ﬁnite volume methods to higher
order accuracy is due to N.E. Kolgan [Kol72, Kol75] and van Leer [vL79]. They proposed
a kind of posttreatment procedure currently known as solution reconstruction or MUSCL
6
scheme. In the above papers the authors used linear reconstruction (it will be retained in
this study as well) but we must say that this method was already extended to quadratic
approximations in each cell [BF90].
3.4.2.1 Historical remark
In general, when we read numerical articles which use the MUSCL scheme, the authors
often cite the paper by van Leer [vL79]. It is commonly believed in the scientiﬁc commu
6
MUSCL stands for monotone upstreamcentered scheme for conservation laws.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
138 Two phase ﬂows
nity that B. van Leer was ﬁrst to propose the gradient reconstruction and slope limiting
ideas. Unfortunately, because of political reasons, the works of N.E. Kolgan [Kol72, Kol75]
remained unknown for a long time. We would like to underline the fact that the ﬁrst
publication of Kolgan came out seven years before van Leer’s paper. Van Leer seems to be
aware of this situation since in his recent review paper [vL06] one can ﬁnd “A historical
injustice” section:
“It has been pointed out to me by Dr. Vladimir Sabelnikov, formerly of TsAGI,
the Central Aerodynamical National Laboratory near Moscow, that a scheme
closely resembling MUSCL (including limiting) was developed in this laboratory
by V. P. Kolgan (1972). Kolgan died young; his work apparently received little
notice outside TsAGI.”
Thus, it is desirable to correctly attribute this result in future publications.
3.4.3 TVD and MUSCL schemes
There is a property of scalar nonlinear conservation laws, which was probably observed
for the ﬁrst time by Peter Lax [Lax73]:
The total increasing and decreasing variations of a diﬀerentiable solution be
tween any pair of characteristics are conserved.
In the presence of shock waves, information is lost and the total variation decreases. For
compactly supported or periodic solutions, one can establish the following inequality
+∞
_
−∞
[du(x, t
2
)[ ≤
+∞
_
−∞
[du(x, t
1
)[ , t
2
≥ t
1
. (3.22)
This motivated Harten [Har83] to introduce the notion of discrete total variation of nu
merical solution u
h
:= ¦u
j
¦
TV (u
h
) :=
j
[u
j+1
−u
j
[ ,
and the discrete counterpart to (3.22)
TV (u
n+1
h
) ≤ TV (u
n
h
).
If this property is fulﬁlled we say that a ﬁnite volume scheme is total variation diminishing
(TVD). The following theorem was proved in [Har83]:
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.4 Finite volume scheme on unstructured meshes 139
Theorem 1 (i): Monotone schemes are TVD
(ii): TVD schemes are monotonicity preserving, i.e. the number of solution extrema is
preserved in time.
Remark 11 From the mathematical point of view it would be more correct to say “the
total variation nonincreasing (TVNI) scheme” but the “wrong” term TVD is generally
accepted in the scientiﬁc literature.
In one space dimension the construction of TVD schemes is not a problem anymore.
Let us recall that in this study we are rather interested in two space dimensions (or even
three in future work). In these cases the situation is considerably more complicated. Even
if we consider the simplest case of structured cartesian meshes and apply a 1D TVD scheme
on a dimensionbydimension basis, a result of Goodman and Leveque shows [GV85] that
TVD schemes in two or more space dimensions are only ﬁrst order accurate. Motivated by
this negative result, weaker conditions yielding solution monotonicity preservation should
be developed.
In this chapter we will describe the construction and practical implementation of a
secondorder nonlinear scheme on unstructured (possibly highly distorted) meshes. The
main idea is to ﬁnd our solution as a piecewise aﬃne function on each cell. This kind of
linear reconstruction operators on simplicial control volumes often exploit the fact that the
cell average is also a pointwise value of any valid (conservative) linear reconstruction eval
uated at the gravity center of a simplex. This reduces the reconstruction problem to that
of gradient estimation given cell averaged data. In this case, we express the reconstruction
in the form
w
K
(x) = ¯ w
K
+ (∇w)
K
(x −x
0
), K ∈ T , (3.23)
where ¯ w
K
is the cell averaged value given by ﬁnite volume method, (∇w)
K
is the solution
gradient estimation (to be determined) on the cell K, x ∈ K and the point x
0
is chosen to
be the gravity center for the simplex K.
It is very important to note that with this type of representation (3.23) we remain
absolutely conservative, i.e.
1
vol(K)
_
K
w
K
(x) dΩ ≡ ¯ w
K
due to the choice of the point x
0
. This point is crucial for ﬁnite volumes because of intrinsic
conservative properties of this method.
In next sections we describe brieﬂy two common techniques: GreenGauss integration
and least squares methods for solution gradient estimation on each cell. There are other
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
140 Two phase ﬂows
available techniques. We can mention here an implicit gradient reconstruction method pro
posed in [MG96] and reused later in [AMS04], for example. We decided not to implement
this approach in our research code since this procedure is computationally expensive
7
.
3.4.3.1 GreenGauss gradient reconstruction
O
K
∂K
n
e
N
1
N
2
Figure 3.7: Illustration for GreenGauss gradient reconstruction. Control volume K with
barycenter O and exterior normal n.
This gradient reconstruction technique can be easily implemented on simplicial meshes.
It is based on two very simple ideas: the mean value approximation and GreenGauss
Ostrogradsky formula.
Consider a control volume K with barycenter O. The exterior normal to an edge e ∈ ∂K
is denoted by n
e
. This conﬁguration is depicted on Figure 3.7. In order to estimate the
solution gradient on K (or in other words, to estimate its value at gravity center O) we
make the following mean value approximation
(∇w)
K
= (∇w)[
O
∼
=
1
vol(K)
_
K
∇w dΩ,
and if we apply GreenGaussOstrogradsky formula
(∇w)
K
∼
=
1
vol(K)
_
∂K
w⊗n dσ =
1
vol(K)
e∈∂K
_
e
w⊗n
e
dσ
∼
=
e∈∂K
area(e)
vol(K)
w[
e/2
⊗n
e
,
7
In order to reconstruct the solution gradient we have to solve a linear system of equations. Recall that
the gradient is estimated at each time step on each control volume. This factor slows down considerably
explicit time discretizations.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.4 Finite volume scheme on unstructured meshes 141
where w[
e/2
denote the solution value at the face (or edge in 2D) centroid. The face value
needed to compute the reconstruction gradient can be obtained from a weighted average
of the values at the vertices on the face [HC89]. In 2D it simply becomes
w[
e/2
=
w
N
1
+w
N
2
2
.
This approximation yields the following formula for gradient estimation:
(∇w)
K
∼
=
e∈∂K
area(e)
vol(K)
(w
N
1
+w
N
2
)
2
⊗n
e
.
The gradient calculation is exact whenever the numerical solution varies linearly over the
support of the reconstruction.
This procedure requires the knowledge of the solution values at the mesh nodes ¦N
i
¦.
Recall that a cell centered ﬁnite volume scheme provides us with data located at cell gravity
centers. Thus, an interpolation technique is needed. We have to say that the quality of
GreenGauss gradient reconstruction greatly depends on the chosen interpolation method.
The method retained in this study is explained in Section 3.4.5.
3.4.3.2 Leastsquares gradient reconstruction method
K
T
1
T
2
T
3
O
O
1
O
2
O
3
Figure 3.8: Illustration for leastsquares gradient reconstruction. We depict a triangle control
volume with three adjacent neighbors.
In this section we consider a triangle
8
control volume K with three adjacent neigh
bors T
1
, T
2
and T
3
. Their barycenters are denoted by O(x
0
), O
1
(x
1
), O
2
(x
2
) and O
3
(x
3
)
8
Generalization to other simplicial control volumes is straightforward.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
142 Two phase ﬂows
respectively. In the following we denote by w
i
the solution value at gravity centers O
i
:
w
i
:= w(x
i
), w
0
:= w(x
0
).
Our purpose here is to estimate ∇w = (∂
x
w, ∂
y
w) on the cell K. Using Taylor formula,
we can write down the three following relations:
w
1
−w
0
= (∇w)
K
(x
1
−x
0
) +O(h
2
), (3.24)
w
2
−w
0
= (∇w)
K
(x
2
−x
0
) +O(h
2
), (3.25)
w
3
−w
0
= (∇w)
K
(x
3
−x
0
) +O(h
2
). (3.26)
If we drop higher order terms O(h
2
), these relations can be viewed as a linear system of
three equations for two unknowns
9
(∂
x
w, ∂
y
w). This situation is due to the fact that the
number of edges incident to a simplex mesh in R
d
is greater or equal (in this case see
Remark 12) to d thereby producing linear constraint equations (3.24) – (3.26) which will
be solved analytically here in a least squares sense.
First of all, each constraint (3.24) – (3.26) is multiplied by a weight ω
i
∈ (0, 1) which
will be chosen below to account for distorted meshes. In matrix form our nonsquare
system becomes
_
_
_
ω
1
∆x
1
ω
1
∆y
1
ω
2
∆x
2
ω
2
∆y
2
ω
3
∆x
3
ω
3
∆y
3
_
_
_
(∇w)
K
=
_
_
_
ω
1
(w
1
−w
0
)
ω
2
(w
2
−w
0
)
ω
3
(w
3
−w
0
)
_
_
_
,
where ∆x
i
= x
i
− x
0
, ∆y
i
= y
i
− y
0
. For further developments it is convenient to rewrite
our constraints in abstract form
[
L
1
,
L
2
] (∇w)
K
=
f. (3.27)
We use a normal equation technique in order to solve symbolically this abstract form in a
least squares sense. Multiplying on the left both sides of (3.27) by [
L
1
L
2
]
t
yields
G(∇w)
K
=
b, G = (l
ij
)
1≤i,j≤2
=
_
(
L
1
L
1
) (
L
1
L
2
)
(
L
2
L
1
) (
L
2
L
2
)
_
(3.28)
where G is the Gram matrix of vectors
_
L
1
,
L
2
_
and
b =
_
(
L
1
f)
(
L
2
f)
_
. The socalled normal
equation (3.28) is easily solved by Cramer’s rule to give the following result
(∇w)
K
=
1
l
11
l
22
−l
2
12
_
l
22
(
L
1
f) −l
12
(
L
2
f)
l
11
(
L
2
f) −l
12
(
L
1
f)
_
.
9
This simple estimation is done for scalar case only w = (w). For more general vector problems the
numbers of equations and unknowns have to be changed depending on the dimension of vector w.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.4 Finite volume scheme on unstructured meshes 143
The form of this solution suggests that the least squares linear reconstruction can be
eﬃciently computed without the need for storing a nonsquare matrix.
Now we have to discuss the choice of weight coeﬃcients ¦ω
i
¦
3
i=1
. The basic idea is
to attribute bigger weights to cells barycenters closer to the node N under consideration.
One of the possible choices consists in taking a harmonic mean of respective distances
r
i
= [[x
i
−x
N
[[. This purely metric argument takes the following mathematical form:
ω
i
=
[[x
i
−x
N
[[
−k
3
j=1
[[x
j
−x
N
[[
−k
,
where k in practice is taken to be one or two (in our numerical code we choose k = 1).
Remark 12 When we have a triangle sharing an edge with boundary ∂Ω (see Figure 3.13
for illustration), the gradient reconstruction procedure becomes even simpler, since the num
ber of constraints is equal to d and linear system (3.24) – (3.26) becomes completely deter
mined:
w
1
−w
0
= (∇w)
K
(x
1
−x
0
) +O(h
2
),
w
2
−w
0
= (∇w)
K
(x
2
−x
0
) +O(h
2
),
or in componentwise form it reads
_
x
1
−x
0
y
1
−y
0
x
2
−x
0
y
2
−y
0
_
(∇w)
K
=
_
w
1
−w
0
w
2
−w
0
_
.
The unique solution to this linear system is given again by Cramer’s rule
(∇w)
K
=
_
(y
2
−y
0
)(w
1
−w
0
) −(y
1
−y
0
)(w
2
−w
0
)
(x
1
−x
0
)(w
2
−w
0
) −(x
2
−x
0
)(w
1
−w
0
)
_
(x
1
−x
0
)(y
2
−y
0
) −(x
2
−x
0
)(y
1
−y
0
)
.
3.4.3.3 Slope limiter
The idea of incorporating limiter functions to obtain nonoscillatory resolution of dis
continuities and steep gradients dates back to Boris and Book [BB73]. When the limiter
is identically equal to 1, we have the unlimited form of the linear interpolation. In the
1D case one can easily ﬁnd in the literature about 15 diﬀerent limiter functions such as
CHARM, minmod, superbee, van Albada and many others. On unstructured meshes the
situation is quite diﬀerent. In the present study we decided to retain the BarthJespersen
limiter proposed in [BJ89]. Here we do not discuss its construction and properties but just
give the ﬁnal formula. We need to introduce the following notation
w
min
K
:= min
L∈N(K)
w
L
, w
max
K
:= max
L∈N(K)
w
L
.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
144 Two phase ﬂows
The limited version of (3.23) is given by the following modiﬁed reconstruction operator
w
K
(x) = ¯ w
K
+ α
K
(∇w)
K
(x −x
0
), K ∈ T ,
where it is assumed that α
K
∈ [0, 1]. Obviously, the choice α
K
= 0 corresponds to the ﬁrst
order scheme while α
K
= 1 is the unlimited form. Barth and Jespersen [BJ89] propose the
following choice of α
K
:
α
BJ
K
:= min
∀f∈∂K
_
¸
_
¸
_
w
max
K
−¯ w
K
w
K
(x
f
)−¯ w
K
if w
K
(x
f
) > w
max
K
,
w
min
K
−¯ w
K
w
K
(x
f
)−¯ w
K
if w
K
(x
f
) < w
min
K
,
1 otherwise.
where x
f
denotes the face f centroid.
Although this limiter function does not fulﬁll all the requirements of ﬁnite volume
maximum principle on unstructured meshes [BO04], it can be shown that it yields ﬁnite
volume schemes possessing a global extremum diminishing property. Also this limiter
produces the least amount of slope reduction which can be advantageous for accuracy.
Note that in practical implementation little modiﬁcations are required to prevent near
zero division for almost constant solution data.
3.4.4 Diﬀusive ﬂuxes computation
O
1
O
2
K
L
f
N
1
N
2
Figure 3.9: Diamond cell constructed around face f.
In this section we brieﬂy describe a method for diﬀusive ﬂuxes discretization. It was
proposed for nonconforming cartesian meshes in [Coi94] and rigorous justiﬁcation was
given later in [CVV99].
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.4 Finite volume scheme on unstructured meshes 145
When we look at the integral conservation law (3.20) for control volume K, the diﬀusion
term has the following form:
1
vol(K)
f∈∂K
_
f
(T∇w) n
f
dσ,
where the integration is performed along all the faces of the cell K. The quantity (T∇w)
is called the diﬀusion ﬂux and should be properly discretized in this section. The diﬀusion
matrix T is given by physical properties, while n
KL
is determined by mesh geometry. The
last integral can be further simpliﬁed by applying a second order Gauss quadrature
10
:
f∈∂K
area(f)
vol(K)
T (∇w[
f/2
) n
f
.
Thus, in order to compute the diﬀusive ﬂux, we have just to estimate the solution gradient
∇w but this time on the face centroid denoted by
f
/
2
. In order to express it in terms of
cell averages ¦w
L
¦
L∈T
we apply once again GreenGauss technique.
We have to start by constructing a dual mesh T
∗
. It is composed of diamond cells
created around each interior face (diﬀusive ﬂux computation through a boundary face is
another topic which is brieﬂy discussed in Remark 13). Consider two simplicial control
volumes K and L with barycenters O
1
and O
2
respectively. These volumes share a common
face (or an edge in 2D) f with nodes N
1
, N
2
, . . . We denote by f
⋄
the diamond cell associated
to the face f. This situation is schematically depicted on Figure 3.9. For the sake of clarity
we construct the diamond cell in the 2D case. In the present study we choose f
⋄
as the
closed polygon formed by nodes N
1
O
1
N
2
O
2
N
1
(see Figure 3.9). Once the diamond cell is
deﬁned, we can apply to its boundary the GreenGauss method (see Section 3.4.3.1):
(∇w[
f/2
)
∼
=
1
vol(f
⋄
)
_
f
⋄
∇w dΩ =
1
vol(f
⋄
)
ℓ∈∂f
⋄
_
f
⋄
w⊗n
ℓ
dσ
∼
=
ℓ∈∂f
⋄
area(ℓ)
vol(f
⋄
)
(∇w[
f/2
) ⊗n
ℓ
∼
=
ℓ∈∂f
⋄
area(ℓ)
vol(f
⋄
)
(w
P
1
+w
P
2
)
2
⊗n
ℓ
,
where ℓ ∈ ¦N
1
O
1
, O
1
N
2
, N
2
O
2
, O
2
N
1
¦ and P
1,2
∈ ¦N
1
, O
1
, N
2
, O
2
¦.
Remark 13 There is a little question omitted in this section. We have not discussed
the implementation of boundary conditions for diﬀusive ﬂuxes. Actually it is rather a
complicated topic. We assume that in our physical problems viscosity plays a secondary
role. Consider a control volume L having a face ℓ on the boundary ∂Ω. For the time being
10
In 1D it simply degenerates to a very well known middle point rule
1
_
−1
g(x) dx
∼
= 2g(0).
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
146 Two phase ﬂows
we impose a Neumann type condition on diﬀusive ﬂuxes. After discretization this condition
takes the following form:
(T∇w
ℓ
) n := (T∇w
L
) n.
In other words it means that the gradient on the boundary is replaced by its value on the
cell L.
3.4.5 Solution interpolation to mesh nodes
We have seen above that several gradient reconstruction procedures (in particular gra
dient estimation on the faces) require the knowledge of the solution at mesh nodes (or
vertices). This information is not directly given by the ﬁnite volume method since we
retained the cellcentered approach.
O
i+1
O
i
O
i−1
N
Figure 3.10: Triangles with their barycenters O
i
sharing the same vertex N.
Let us consider a node N(x
n
, y
n
) of the tesselation T and a control volume K
i
with
barycenter O
i
(x
i
, y
i
) having this node as a vertex (see Figure 3.10 for illustration). The
MUSCL procedure provides us solution gradient on each cell. Thus, using the Taylor
formula or, equivalently, representation (3.23) we can estimate the solution value at the
node N
w
N
= ¯ w
K
i
+ (∇w)
K
i
(x
N
−x
i
). (3.29)
The problem is that we will have d(N) diﬀerent values of the solution in the same point
depending on the control volume under consideration. Here d(N) is the degree of vertex
N in the sense of graph theory. One of the possible ways to overcome this contradiction is
the averaging. One interesting technique was proposed in [HC89] and further improved in
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.4 Finite volume scheme on unstructured meshes 147
[KMC03]. In our turn, we slightly modiﬁed this method. The algorithm implemented in
our code is brieﬂy described here.
First of all, let us look for the vertex value ¯ w
N
as a weighted sum of the values w
N
i
computed by formula (3.29) from each surrounding cell
¯ w
N
=
d(N)
i=1
ω
i
w
N
i
d(N)
i=1
ω
i
.
The weighting factors ¦ω
i
¦
d(N)
i=1
are made to satisfy the condition of zero pseudoLaplacian
L(x
n
) ≡
d(N)
i=1
ω
i
(x
i
−x
n
), L(y
n
) ≡
d(N)
i=1
ω
i
(y
i
−y
n
) . (3.30)
These conditions have a very simple interpretation. They are imposed so that the method
be exact for aﬃne data over the stencil.
As in the original formulation by Holmes and Connell [HC89], the weighting factor ω
i
is written as
ω
i
= 1 + ∆ω
i
.
The weights ¦ω
i
¦ are determined by solving an optimization problem in which the
costfunction to be minimized is deﬁned as
1
2
d(N)
i=1
_
r
i
∆ω
i
_
2
→min (3.31)
with two constraints given by (3.30). It should be noted that the cost function is slightly
diﬀerent from the original formulation. The diﬀerence lies in the factor of
r
2
i
≡ [[
ON −
OO
i
[[
2
which was introduced in [KMC03]. This modiﬁcation eﬀectively allows larger values of
weight ∆ω
i
for those cells closer to the node in question.
Employing the method of Lagrange multipliers, the original optimization problem,
which was to minimize the cost function given by (3.31) with the constraints (3.30) is
equivalent to minimizing the function L deﬁned by
L =
1
2
d(N)
i=1
_
r
i
∆ω
i
_
2
−λ
d(N)
i=1
ω
i
(x
i
−x
n
) −µ
d(N)
i=1
ω
i
(y
i
−y
n
) →min
which leads to
∆ω
i
=
λ(x
i
−x
n
) + µ(y
i
−y
n
)
r
2
i
.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
148 Two phase ﬂows
The two Lagrangian multipliers, λ and µ, are obtained from
λ =
r
y
I
xy
−r
x
I
yy
I
xx
I
yy
−I
2
xy
, µ =
r
x
I
xy
−r
y
I
xx
I
xx
I
yy
−I
2
xy
,
where
r
x
=
d(N)
i=1
(x
i
−x
n
), r
y
=
d(N)
i=1
(y
i
−y
n
).
I
xx
=
d(N)
i=1
(x
i
−x
n
)
2
r
2
i
, I
yy
=
d(N)
i=1
(y
i
−y
n
)
2
r
2
i
, I
xy
=
d(N)
i=1
(x
i
−x
n
)(y
i
−y
n
)
r
2
i
.
The last step consists in renormalizing the weights ¦ω
i
¦
d(N)
i=1
to the range [0, 1].
Remark 14 The above algorithm is not computationally expensive since the weights ¦ω
i
¦
d(N)
i=1
depend only on the tesselation T geometry. It means that they can be computed and stored
before the main loop in time and reused during the computations later.
Remark 15 Even if we suggest to use the above method (since it gives slightly better
results), we would like to give another idea (based on a purely metrics argument) of how
to construct the weights ¦ω
i
¦
d(N)
i=1
. This approach is considerably simpler than optimization
problem solving and it was already used in the least squares gradient reconstruction method
(see Section 3.4.3.2). In fact, in order to calculate ω
i
one can simply take the harmonic
mean of distances between the node N under the question and respective cell barycenter O
i
(see Figure 3.10):
ω
i
:=
[[x
N
−x
i
[[
−k
d(N)
j=1
[[x
N
−x
i
[[
−k
,
where k in practice is taken 1 or 2.
We have to say that this choice does not guarantee exact interpolation of globally linear
functions.
3.4.6 Time stepping methods
In previous sections we considered the spatial discretization procedure with a ﬁnite
volume scheme. It is a common practice in solving timedependent PDEs to ﬁrst discretize
the spatial variables. This approach is called a method of lines:
u
t
+ ∂
x
f(u) = S(u)
FV
=⇒u
t
= L(u) (3.32)
In order to obtain a fully discrete scheme, we have to discretize the time evolution operator.
In the present work we decided to retain the socalled Strong StabilityPreserving (SSP)
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.4 Finite volume scheme on unstructured meshes 149
time discretization methods described in [Shu88, GST01, SR02]. Historically these methods
were initially called Total Variation Diminishing (TVD) time discretizations.
The main idea behind SSP methods is to assume that the ﬁrst order forward Euler
method is strongly stable (see the deﬁnition below) under a certain norm for our method
of lines ODE (3.32). Then, we try to ﬁnd a higher order scheme. Usually the relevant
norm is the total variation
11
norm:
TV(u
n
) :=
j
¸
¸
u
n
j
−u
n
j−1
¸
¸
and TVD discretizations have the property TV(u
n+1
) ≤ TV(u
n
).
Remark 16 Special approaches are needed for hyperbolic PDEs since they contain dis
continuous solutions and the usual linear stability analysis is inadequate. Thus a stronger
measure of stability is usually required:
Deﬁnition 1 A sequence ¦u
n
¦ is said to be strongly stable in a given norm [[[[ provided
that [[u
n+1
[[ ≤ [[u
n
[[ for all n ≥ 0.
A general mstage RungeKutta method for (3.32) can be written in the form
u
(0)
= u
n
, (3.33)
u
(i)
=
i−1
k=0
_
α
i,k
u
(k)
+ ∆tβ
i,k
L(u
(k)
)
_
, α
i,k
≥ 0, i = 1, . . . , m, (3.34)
u
n+1
= u
(m)
. (3.35)
In [SO88] the following result is proved
Theorem 2 If the forward Euler method is strongly stable under the CFL restriction ∆t ≤
∆t
FE
[[u
n
+ ∆tL(u
n
)[[ ≤ [[u
n
[[ ,
then the RungeKutta method (3.33) – (3.35) with β
i,k
≥ 0 is SSP, [[u
n+1
[[ ≤ [[u
n
[[,
provided the following CFL restriction is fulﬁlled:
∆t ≤ c∆t
FE
, c = min
i,k
α
i,k
β
i,k
.
Here we give a few examples of SSP schemes which are commonly used in applications:
11
We have to say that the notion of total variation is used essentially for onedimensional discrete
solutions.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
150 Two phase ﬂows
Absolute stability regions
−6 −5 −4 −3 −2 −1 0 1 2
−3
−2
−1
0
1
2
3
SSPRK(2,2) & RK2
SSPRK(3,3) & RK3
SSPRK(3,4)
RK4
Figure 3.11: Linear absolute stability regions for diﬀerent time stepping methods
Figure 3.12: Nonlinear absolute stability regions
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.4 Finite volume scheme on unstructured meshes 151
• Optimal
12
second order twostage SSPRK(2,2) scheme with CFL = 1:
u
(1)
= u
(n)
+ ∆tL(u
(n)
),
u
(n+1)
=
1
2
u
(n)
+
1
2
u
(1)
+
1
2
∆tL(u
(1)
);
• Optimal third order threestage SSPRK(3,3) scheme with CFL = 1:
u
(1)
= u
(n)
+ ∆tL(u
(n)
),
u
(2)
=
3
4
u
(n)
+
1
4
u
(1)
+
1
4
∆tL(u
(1)
),
u
(n+1)
=
1
3
u
(n)
+
2
3
u
(2)
+
2
3
∆tL(u
(2)
);
• Third order fourstage SSPRK(3,4) scheme with CFL = 2:
u
(1)
= u
(n)
+
1
2
∆tL(u
(n)
),
u
(2)
= u
(1)
+
1
2
∆tL(u
(1)
),
u
(3)
=
2
3
u
(n)
+
1
3
u
(2)
+
1
6
∆tL(u
(n)
),
u
(n+1)
= u
(3)
+
1
2
∆tL(u
(3)
).
We depicted on Figures 3.11 and 3.12 the linear and nonlinear absolute stability regions of
these schemes, and compared them with the classical RungeKutta methods respectively.
Note that the linear absolute stability region for corresponding RK and SSPRK schemes
is the same. It is a consequence of linearization since these schemes are clearly diﬀerent
for nonlinear problems. That is why we decided to look also at nonlinear absolute stability
[CP92].
We tested these diﬀerent schemes in our numerical code and we decided to adopt SSP
RK(3,4) due to its accuracy and wide stability region. In our opinion this scheme represents
a very good tradeoﬀ between precision and robustness.
3.4.7 Boundary conditions treatment
In this section we show how to implement the wall boundary condition in the ﬁnite
volumes framework. The ﬂavor of boundary conditions treatment for hyperbolic systems
is given in Section A.2 and we refer to [GP05] for a general discussion.
12
Optimality in the sense of CFL condition.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
152 Two phase ﬂows
From the numerical point of view, we need to employ the boundary conditions when
we have a control volume sharing a face (or an edge in 2D) with the computational domain
boundary (see Figure 3.13 for illustration). In this situation, the numerical method must
be provided by an additional information (such as the value of normal advective or diﬀusive
ﬂuxes on ∂K). Physical conditions completed if necessary by numerical ones allow us to
reconstruct the normal ﬂux T n through the boundary face. Consider the case of a rigid
wall boundary
u(x, t) n = 0, x ∈ ∂Ω, (3.36)
and the hyperbolic system (3.1) – (3.3).
K
∂Ω
Figure 3.13: Control volume sharing a face with boundary ∂Ω.
The advective ﬂux
13
through an oriented boundary face with exterior normal n is given
by (3.5)
T n = (α
+
ρ
+
u
n
, α
−
ρ
−
u
n
, ρuu
n
+ pn
1
, ρvu
n
+ pn
2
, ρHu
n
), u
n
≡ u n,
and if we take into account physical condition (3.36) it becomes
(T n)[
x∈∂Ω
= (0, 0, p
b
n
1
, p
b
n
2
, 0), p
b
:= p[
x∈∂Ω
The last identity means that we need only to know the pressure value on the wall in order
to reconstruct the advective ﬂux (T n)[
x∈∂Ω
.
13
A few words about numerical boundary conditions for diﬀusive ﬂuxes are given in Section 3.4.4
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.5 Numerical results 153
Figure 3.14: Shock tube of Sod. This plot represents the density.
After some straightforward computations (for more details see [GP05]) one establishes
the following formula for the boundary pressure
p
b
= p + ρu
n
c
s
,
where the righthand side is evaluated in the boundary cell center.
Remark 17 The last formula has a very interesting physical interpretation. Actually,
the boundary pressure p
b
measures the force “exerted” by the wall in order to annihilate
completely the normal velocity u
n
.
3.5 Numerical results
In this section we present a few test cases which show our research code perfomance.
In order to reduce the document size, we decided not to put here classical test cases such
as the shock tube of Sod [Sod78] etc. . The reader can believe us that our numerical code
passed successfully these classical tests (see Figure 3.14).
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
154 Two phase ﬂows
3.5.1 Convergence test
We begin the presentation of numerical tests by the simplest one – convergence test.
We would like to show the accuracy of the MUSCL scheme implementation. To do it, we
solve numerically the following scalar linear advection equation
∂v
∂t
+u
0
∇v = 0, u
0
∈ R
2
with smooth
14
initial conditions. Moreover, it has almost compact support in order to
reduce the inﬂuence of boundary conditions. It is obvious that this equation will just
translate the initial form in the direction u
0
. So, we have an analytical solution which
can be used to quantify the numerical method error. On the other hand, to measure the
convergence rate, we constructed a sequence of reﬁned meshes.
On Figure 3.15 we depict the error of numerical method in L
∞
norm as a function of
the mesh characteristic size. The slope of these curves represents an approximation to the
theoretical convergence rate. On this plot, the blue curve corresponds to the ﬁrst order
upwind scheme while the other two (red and black) correspond to the MUSCL scheme
with leastsquares (see Section 3.4.3.2) and GreenGauss (see Section 3.4.3.1) gradient
reconstruction procedures respectively. One can see that the blue curve slope is equal
approximatively to 0.97 which means ﬁrst order convergence. The other two curves have
almost the same slope equal to 1.90 indicating a second order convergence rate for the
MUSCL scheme. We remark that in our implementation of the secondorder scheme the
leastsquares reconstruction seems to give slightly more accurate results than the Green
Gauss procedure.
The next ﬁgure represents the measured CPU time in seconds again as a function of
the mesh size. Obviously, this kind of data is extremely computer dependent but the
qualitative behaviour is the same on all systems. On Figure 3.16 one can see that the
“fastest” curve is the blue one (ﬁrst order upwind scheme). Then we have two almost
superimposed (black and red) curves referring to the secondorder gradient reconstruction
on variables. Here again one can notice that the leastsquares method is slightly faster than
the GreenGauss procedure. On this ﬁgure we represented one more curve (the highest
one) which corresponds to GreenGauss gradient reconstruction on ﬂuxes (it seems to be
very natural in the context of VFFC scheme). Our numerical tests show that this method
is quite expensive from the computational point of view and we decided not to retain it.
14
We intentionally choose a smooth initial condition since the discontinuities can decrease the overall
accuracy of the scheme.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.5 Numerical results 155
10
−2
10
−1
10
0
10
1
10
−4
10
−3
10
−2
10
−1
10
0
h − average edge length
L
∞
e
r
r
o
r
Convergence rate in L
∞
norm
Upwind. Conv. rate = 0.97731
LeastSq. Conv. rate = 1.9051
GGO on Variables. Conv. rate = 1.9012
Figure 3.15: Numerical method error in L
∞
norm.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
156 Two phase ﬂows
10
−2
10
−1
10
0
10
1
10
−3
10
−2
10
−1
10
0
10
1
10
2
10
3
CPU time in seconds
h
T
Upwind 1st order
Green−Gauss
Least Sqs
GG on fluxes
Figure 3.16: CPU time for diﬀerent ﬁnite volume schemes.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.5 Numerical results 157
3.5.2 Falling water column
The geometry and initial condition for this test case are depicted on Figure 3.17. Ini
tially the velocity ﬁeld is taken to be zero. The values of the other parameters are given
in Table 3.1. The mesh used in this computation contained about 108000 control volumes
(in this case it was triangles). The results of this simulation are presented on Figures 3.18
– 3.23.
On the other hand, we depicted on Figure 3.24 the maximal pressure on the right wall
as a function of time:
t −→ max
(x,y)∈1×[0,1]
p(x, y, t).
We performed another computation when gas was modelled as α
+
= 0.05, α
−
= 0.95
mixture. The pressure is recorded as well and this result is ploted on Figure 3.25. One can
see that the peak value is higher and the impact is more localised in time.
α
+
= 0.9
α
−
= 0.1
α
+
= 0.1
α
−
= 0.9
0 0.3 0.65 0.7
0.05
1
1
0.9
g
Figure 3.17: Falling water column test case. Geometry and initial condition.
3.5.3 Water drop test case
The geometry and initial condition for this test case are depicted on Figure 3.26. Ini
tially the velocity ﬁeld is taken to be zero. The values of the other parameters are given
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
158 Two phase ﬂows
(a) t = 0.005 (b) t = 0.06
Figure 3.18: Falling water column test case. Initial condition and the beginning of column
dropping process.
(a) t = 0.1 (b) t = 0.125
Figure 3.19: Falling water column test case. Splash creation due to the interaction with step.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.5 Numerical results 159
(a) t = 0.15 (b) t = 0.175
Figure 3.20: Falling water column test case. Water strikes the wall  I.
(a) t = 0.2 (b) t = 0.225
Figure 3.21: Falling water column test case. Water strikes the wall  II.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
160 Two phase ﬂows
(a) t = 0.3 (b) t = 0.4
Figure 3.22: Falling water column test case. Splash is climbing the wall.
(a) t = 0.5 (b) t = 0.675
Figure 3.23: Falling water column test case. Turbulent mixing process.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.5 Numerical results 161
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0.9
1
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
t, time
p
m
a
x
/
p
0
Maximal pressure on the right wall
Figure 3.24: Maximal pressure on the right wall. Heavy gas case.
in Table 3.1. The mesh used in this computation contained about 92000 control volumes
(in this case it was triangles). The results of this simulation with our code are presented
on Figures 3.27 – 3.33. On Figure 3.34 we plot the maximal pressure on the bottom as a
function of time:
t −→ max
(x,y)∈[0,1]×0
p(x, y, t).
From this ﬁgure, it is clear to see that the pressure exerted on the bottom reaches 2.5p
0
due to the drop impact at t ≈ 0.16.
Remark 18 Beginning with Figure 3.32 one can see some asymmetry in the solution.
It is not expected since the initial condition, computational domain and forcing term are
fully symmetric with respect to the line x = 0.5. This discrepancy is explained by the use
of unstructured meshes in the computation. The arbitrariness of the orientation of the
triangles introduces small perturbations which are suﬃcient to break the symmetry at the
discrete level. At the present time, symmetry preserving schemes on unstructured meshes
represent a big challenge in numerical analysis. The author is not aware of any work in
this direction.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
162 Two phase ﬂows
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0.8
1
1.2
1.4
1.6
1.8
2
2.2
t, time
p
m
a
x
/
p
0
Maximal pressure on the right wall
Figure 3.25: Maximal pressure on the right wall as a function of time. Light gas.
3.6 Conclusions
In this chapter we presented a simple mathematical model for simulating water waves
impacts. The validation of this approach is the subject of future publications. Namely, we
are going to perform qualitative and quantitative comparison with the more general six
equations model [Ish75].
On the other hand, we presented a numerical approach for discretizing the govern
ing equations. It is a second order ﬁnite volume scheme on unstructured meshes. This
method was implemented in our research code. By construction, our code has excellent
mass, momentum and total energy conservation properties. Diﬀerent numerical tests from
Section 3.5 validate the numerical method.
In the future we would like to carry out a parametric study with our solver. The
inﬂuence of aeration, gas properties and other factors on the impact pressures is very
important for industrial applications.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.6 Conclusions 163
α
+
= 0.1
α
−
= 0.9
α
+
= 0.9
α
−
= 0.1
0 0.5
0.7
1
1
g
R = 0.15
Figure 3.26: Geometry and initial condition for water drop test case.
(a) t = 0.005 (b) t = 0.075
Figure 3.27: Water drop test case. Initial conﬁguration and the beginning of the fall.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
164 Two phase ﬂows
(a) t = 0.1 (b) t = 0.125
Figure 3.28: Water drop test case. Drop approaching container bottom.
(a) t = 0.135 (b) t = 0.15
Figure 3.29: Water drop test case. Drop/bottom compressible interaction.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.6 Conclusions 165
(a) t = 0.175 (b) t = 0.2
Figure 3.30: Water drop test case. Vertical jets formation.
(a) t = 0.225 (b) t = 0.275
Figure 3.31: Water drop test case. Side jets crossing.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
166 Two phase ﬂows
(a) t = 0.325 (b) t = 0.35
Figure 3.32: Water drop test case. Side jets interﬂow at the center.
(a) t = 0.4 (b) t = 0.45
Figure 3.33: Water drop test case. Central jet reﬂection from the bottom.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
3.6 Conclusions 167
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
1
1.5
2
2.5
t, time
p
m
a
x
/
p
0
Maximal pressure on the bottom
Figure 3.34: Water drop test case. Maximal bottom pressure as a function of time.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
168 Two phase ﬂows
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
Chapter 4
Weakly damped free surface ﬂows
Everything should be made as simple as possible, but not simpler.
Albert Einstein (1879 – 1955)
Contents
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
4.2 Anatomy of dissipation . . . . . . . . . . . . . . . . . . . . . . . 172
4.3 Derivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
4.3.1 Dissipative KdV equation . . . . . . . . . . . . . . . . . . . . . . 179
4.4 Dispersion relation . . . . . . . . . . . . . . . . . . . . . . . . . . 180
4.4.1 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
4.5 Attenuation of linear progressive waves . . . . . . . . . . . . . 182
4.6 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . 187
4.6.1 Approximate solitary wave solution . . . . . . . . . . . . . . . . . 187
4.6.2 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
4.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
4.1 Introduction
Even though the irrotational theory of freesurface ﬂows can predict successfully many
observed wave phenomena, viscous eﬀects cannot be neglected under certain circumstances.
Indeed the question of dissipation in potential ﬂows of ﬂuid with a free surface is a very im
portant one. As stated by [LH92], it would be convenient to have equations and boundary
169
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
170 Viscous potential ﬂows
conditions of comparable simplicity as for undamped freesurface ﬂows. The peculiarity
here lies in the fact that the viscous term in the Navier–Stokes (NS) equations is identically
equal to zero for a velocity deriving from a potential.
The eﬀects of viscosity on gravity waves have been addressed since the end of the
nineteenth century in the context of the linearized Navier–Stokes (NS) equations. It is
wellknown that Lamb [Lam32] studied this question in the case of oscillatory waves on
deep water. What is less known is that Boussinesq studied this eﬀect as well [Bou95]. In
this particular case they both showed that
dα
dt
= −2νk
2
α, (4.1)
where α denotes the wave amplitude, ν the kinematic viscosity of the ﬂuid and k the
wavenumber of the decaying wave. This equation leads to the classical law for viscous
decay, namely
α(t) = α
0
e
−2νk
2
t
. (4.2)
The importance of viscous eﬀects for water waves has been realized by numerous ex
perimental studies for at least thirty years. Here we are going to quote a few of them. For
example, in [ZG71] one can ﬁnd
[...] However, the amplitude disagrees somewhat, and we suppose that this
might be due to the viscous dissipation [...]
Another author [Wu81] points out this defect of the classical water waves theory:
[...] the peak amplitudes observed in the experiments are slightly smaller than
those predicted by the theory. This discrepancy can be ascribed to the neglect
of the viscous eﬀects in the theory [...]
And ﬁnally in the classical article [BPS81] one ﬁnds the following important conclusion:
[...] it was found that the inclusion of a dissipative term was much more im
portant than the inclusion of the nonlinear term, although the inclusion of the
nonlinear term was undoubtedly beneﬁcial in describing the observations [...]
Water wave energy can be dissipated by diﬀerent physical mechanisms. The research
community agrees at least on one point: the molecular viscosity is unimportant. Now let
us discuss more debatable statements. For example if we take a tsunami wave and estimate
its Reynolds number, we ﬁnd Re ≈ 10
6
. So, the ﬂow is clearly turbulent and in practice it
can be modelled by various eddy viscosity models. On the other hand, in laboratory ex
periments the Reynolds number is much more moderate and sometimes we can neglect this
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
4.1 Introduction 171
eﬀect. Then one can think about boundary layers. In this chapter we will deal essentially
with the bottom boundary layer. We brieﬂy discuss the free surface boundary layer and
explain why we do not take it into account in this study. We have to say that the bound
ary layer mechanism is especially important for long waves since they “feel” the bottom.
Finally, the most important (and the most challenging) mechanism of energy dissipation
is wave breaking. This process is extremely diﬃcult from the mathematical but also the
physical and numerical points of view since we have to deal with multivalued functions,
topological changes in the ﬂow and complex turbulent mixing processes. Nowadays the
practitioners can only be happy to model roughly this process by adding adhoc diﬀusive
terms when the wave becomes steep enough.
In this work we keep the features of undamped freesurface ﬂows while adding dissipative
eﬀects. The classical theory of viscous potential ﬂows [JW04] is based on pressure and
boundary conditions corrections due to the presence of viscous stresses. We present here
a novel approach.
Currently, potential ﬂows with adhoc dissipative terms are used for example in direct
numerical simulations of weak turbulence of gravity waves [DKZ03, DKZ04, ZKPD05].
There have also been several attempts to introduce dissipative eﬀects into long wave mod
elling [Mei94, DD07a, CG07].
The present chapter is a direct continuation of the recent study [DD07b, DDZ07]. In
that work the authors considered twodimensional (2D) periodic waves in inﬁnite depth,
while in the present study we remove these two hypotheses and all the computations are
done in 3D. This point is important since the vorticity structure is more complicated in
3D. In other words we consider a general wavetrain on the free surface of a ﬂuid layer of
ﬁnite depth. As a result we obtain a qualitatively diﬀerent formulation which contains a
nonlocal term in the bottom kinematic condition. The inclusion of this term is natural
since it represents the correction to potential ﬂow due to the presence of a boundary layer.
Moreover, this term is predominant since its magnitude scales with O(
√
ν), while other
terms in the freesurface boundary conditions are of order O(ν). Other researchers have
obtained nonlocal corrections but they diﬀer from ours [KM75]. The diﬀerences will be
explained below.
Recently we discovered that the nonlocal term in exactly the same form was already
derived in [LO04]. Unfortunately we made this discovery after publishing our ﬁrst paper
[DD07b] on the subject and, consequently, we could not cite the original work of P. Liu.
Even though we try to read the literature on a regular basis, it is impossible to be aware
of all papers. Consequently, we sometimes rederive what others did before. It happened
to us and we ensure that it was done completely independently.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
172 Viscous potential ﬂows
4.2 Anatomy of dissipation
In this section we are going to discuss the contribution of diﬀerent ﬂow regions into
water wave energy dissipation. We conventionally [Mei94] divide the ﬂow into three regions
illustrated on Figure 4.1. On this ﬁgure S
f
and S
b
stand for free surface and bottom
respectively. Then, R
i
, R
f
and R
b
denote the interior region, free surface and bottom
boundary layers.
O
x
z
S
f
S
b
R
f
R
b
R
i
Figure 4.1: Conventional partition of the ﬂow region into interior region and free surface, bottom
boundary layers.
In order to make some estimates we introduce the notation which will be used in
this section: µ is the dynamic viscosity, δ = O(
√
µ) is the boundary layer thickness
characterization, t
0
is the characteristic time, a
0
is the characteristic wave amplitude and
ℓ is the wavelength or basin length (in laboratory experiments, for example).
We assume that the ﬂow is governed by the incompressible NavierStokes equations:
∇ u = 0,
∂u
∂t
+u ∇u +
1
ρ
∇p = g +
1
ρ
∇ τ,
where τ is the viscous stress tensor
τ
ij
= 2µε
ij
, ε
ij
=
1
2
_
∂u
i
∂x
j
+
∂u
j
∂x
i
_
.
We multiply the second equation by u and integrate over the domain Ω to get the
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
4.2 Anatomy of dissipation 173
following energy balance equation:
1
2
_
Ω
∂
∂t
_
ρ[u[
2
_
dΩ +
1
2
_
∂Ω
ρ[u[
2
u n dσ =
=
_
∂Ω
_
−pI + τ
_
n u dσ +
_
Ω
ρg u dΩ −
1
2µ
_
Ω
τ : τ dΩ
. ¸¸ .
T
.
In this identity each term has a precise physical meaning. The lefthand side is the total
rate of energy change in Ω. The second term is the ﬂux of energy across the boundary.
On the righthand side, the ﬁrst integral represents the rate of work by surface stresses
acting on the boundary. The second integral is the rate of work done by the gravity force
throughout the volume, and the third integral T is the rate of viscous dissipation. We will
focus our attention on the last one T . We estimate the order of magnitude of the rate of
dissipation in various regions of the ﬂuid.
We start by the interior region R
i
. Outside the boundary layers, it is reasonable to
expect that the rate of strain is dominated by the irrotational part whose scale is
a
0
t
0
and
the length scale is the wavelength ℓ. The energy dissipation rate is then
O
_
T
R
i
_
∼
1
µ
_
µ
a
t
0
ℓ
_
2
ℓ
3
= µ
_
a
t
0
_
2
ℓ ∼ O(µ) .
Inside the bottom boundary layer the normal gradient of the solenoidal part of u dom
inates the strain rate, so that
O
_
T
R
b
_
∼
1
µ
_
µ
a
t
0
δ
_
2
δℓ
2
=
µ
δ
_
aℓ
t
0
_
2
∼ O(µ
1
2
) .
A free surface boundary layer also exists. Its importance depends on the free surface
conditions. Consider ﬁrst the classical case of a clean surface. The stress is mainly con
trolled by the potential velocity ﬁeld which is of the same order as in the main body of
the ﬂuid. Because of the small volume O(δℓ
2
) the rate of dissipation in the free surface
boundary layer is only
O
_
T
R
f
_
∼
1
µ
_
µ
a
t
0
ℓ
_
2
δℓ
2
= µδ
_
a
t
0
_
2
∼ O(µ
3
2
) .
From the physical point of view it is weaker, since only the zero shear stress condition on
the free surface is required.
Another extreme case is when the free surface is heavily contaminated, for example,
by oil slicks. The stress in the free surface boundary layer can then be as great as in
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
174 Viscous potential ﬂows
the boundary layer near a solid wall. In the present study we do not treat such extreme
situations and the surface contamination is assumed to be absent.
The previous scalings suggest the following diagram which represents the hierarchy of
dissipative terms:
O
_
µ
1
2
_
. ¸¸ .
R
b
֒→O(µ)
. ¸¸ .
R
i
֒→O
_
µ
3
2
_
. ¸¸ .
R
f
֒→. . .
It is clear that the largest energy dissipation takes place inside the wall boundary layer.
We take into account only the two ﬁrst phenomena from this diagram. Consequently, all
dissipative terms of order O(µ
3
2
) and higher will be neglected.
4.3 Derivation
Consider the linearized 3D incompressible NS equations describing freesurface ﬂows in
a ﬂuid layer of uniform depth h:
∂v
∂t
= −
1
ρ
∇p + ν∆v +g, ∇ v = 0, (4.3)
with v the velocity vector, p the pressure, ρ the ﬂuid density and g the acceleration due to
gravity. We represent v = (u, v, w) in the form of the Helmholtz–Leray decomposition:
v = ∇φ +∇
ψ,
ψ = (ψ
1
, ψ
2
, ψ
3
). (4.4)
After substitution of the decomposition (4.4) into (4.3), one notices that the equations are
veriﬁed provided that the functions φ and
ψ satisfy the following equations:
∆φ = 0, φ
t
+
p −p
0
ρ
+ gz = 0,
∂
ψ
∂t
= ν∆
ψ.
Next we discuss the boundary conditions. We assume that the velocity ﬁeld satisﬁes
the conventional noslip condition at the bottom v[
z=−h
=
0, while at the free surface we
have the usual kinematic condition, which can be stated as
η
t
+v ∇η = w.
After linearization it becomes simply η
t
= w.
Dynamic condition states that the forces must be equal on both sides of the free surface:
[σ n] ≡ −(p −p
0
)n + τ n = 0 at z = η(x, t) ,
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
4.3 Derivation 175
where σ is the stress tensor, [f] denotes the jump of a function f across the free surface, n
is the normal to the free surface and τ the viscous part of the stress tensor σ. The explicit
expressions (in two dimensions for simplicity) of τ and n are
τ = ρν
_
2
∂u
∂x
∂u
∂z
+
∂w
∂x
∂u
∂z
+
∂w
∂x
2
∂w
∂z
_
, n =
1
_
1 + (
∂η
∂x
)
2
_
−
∂η
∂x
1
_
,
After linearization the normal vector n equals to (0, 0, 1).
Using Fourier–Laplace transforms, which we denote by L
F
≡ L ◦ T,
f(x, t)
L
F
−→
ˆ
f(k, s), k = (k
x
, k
y
)
we can determine the structure of the unknown functions φ,
ψ in the transform space. We
assume that all the functions involved in the present computation satisfy the necessary
regularity requirements and have suﬃcient decay at inﬁnity so that the integral transforms
can be applied. The solution for φ is obtained from the transformed continuity equation
∆φ = 0
L
F
−→
ˆ
φ
zz
−[k[
2
ˆ
φ = 0
and
ψ from the corresponding transformed equation
ψ
t
= ν∆
ψ
L
F
−→s
ˆ
ψ = ν
_
ˆ
ψ
zz
−[k[
2 ˆ
ψ
_
ˆ
φ = ˆ ϕ
+
0
(k, s)e
kz
+ ˆ ϕ
−
0
(k, s)e
−kz
,
ˆ
ψ
i
=
ˆ
ψ
i0
(k, s)
_
e
mz
+ C
i
(k, s)e
−mz
_
,
where m
2
:= [k[
2
+
s
ν
and ˆ ϕ
+
0
, ˆ ϕ
−
0
,
ˆ
ψ
0
,
C := (C
1
, C
2
, C
3
) are unknown functions of the
transform parameters (k, s), determined by the initial and appropriate boundary condi
tions.
There are three dynamic conditions on the free surface. Let us use ﬁrst those related
to the tangential stresses (the third one will be used later), where µ = ρν:
σ
xz
= µ
_
∂w
∂x
+
∂u
∂z
_
= 0, σ
yz
= µ
_
∂w
∂y
+
∂v
∂z
_
= 0, at z = 0.
Substituting decomposition (4.4) into these two identities yields
2
∂
2
φ
∂x∂z
+
∂
2
ψ
2
∂x
2
−
∂
2
ψ
1
∂x∂y
+
∂
2
ψ
3
∂y∂z
−
∂
2
ψ
2
∂z
2
= 0, z = 0, (4.5)
2
∂
2
φ
∂y∂z
+
∂
2
ψ
2
∂x∂y
−
∂
2
ψ
1
∂y
2
+
∂
2
ψ
1
∂z
2
−
∂
2
ψ
3
∂x∂z
= 0, z = 0. (4.6)
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
176 Viscous potential ﬂows
The next step consists in taking the Fourier–Laplace transform to these relations. We
do not give here the explicit expressions since this operation is straightforward. The
combination (−ik
x
)
(4.5) + (−ik
y
)
(4.6) gives the important relation
ik
y
ˆ
ψ
10
(1 + C
1
) −ik
x
ˆ
ψ
20
(1 + C
2
) = −
2 [k[
3
m
2
+[k[
2
( ˆ ϕ
+
0
− ˆ ϕ
−
0
). (4.7)
Let us turn to the freesurface kinematic condition
∂η
∂t
= w ≡
∂φ
∂z
+
∂ψ
2
∂x
−
∂ψ
1
∂y
, z = 0.
In transform space it becomes
sˆ η = [k[ ( ˆ ϕ
+
0
− ˆ ϕ
−
0
) + ik
y
ˆ
ψ
10
(1 + C
1
) −ik
x
ˆ
ψ
20
(1 + C
2
). (4.8)
Equations (4.7) and (4.8) can be rewritten as
[k[ ( ˆ ϕ
+
0
− ˆ ϕ
−
0
)
ν(m
2
+[k[
2
)
= ˆ η, (4.9)
ik
y
ˆ
ψ
10
(1 + C
1
) −ik
x
ˆ
ψ
20
(1 + C
2
) = −2ν [k[
2
ˆ η. (4.10)
Using (4.10) one can replace the rotational part in the kinematic freesurface condition:
η
t
= φ
z
+L
−1
F
_
−2ν [k[
2
ˆ η
¸
= φ
z
+ 2ν∆η. (4.11)
In order to account for the presence of viscous stresses, we have to modify the dynamic
freesurface condition as well. This is done using the balance of normal stresses at the free
surface:
σ
zz
= 0 at z = 0 ⇒p −p
0
= 2ρν
∂w
∂z
≡ 2ρν
_
∂
2
φ
∂z
2
+
∂
2
ψ
2
∂x∂z
−
∂
2
ψ
1
∂y∂z
_
.
Using (4.10) one can show that
∂
2
ψ
2
∂x∂z
−
∂
2
ψ
1
∂y∂z
= O(ν
1
2
), so Bernoulli’s equation becomes
φ
t
+ gη + 2νφ
zz
+O(ν
3
2
) = 0. (4.12)
Since we only consider weak dissipation (ν ∼ 10
−6
−10
−3
m
2
/s), we neglect terms of order
o(ν).
The second step in our derivation consists in introducing a boundary layer correction
at the bottom. Obviously, this was not done in the previous study [DDZ07], since the
derivation dealt with the inﬁnite depth case. In order to include this modiﬁcation, we
consider a semiinﬁnite ﬂuid layer as it is usually done in boundary layer theory. The ﬂuid
occupies the domain z > −h. In this derivation we use the pure Leray decomposition
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
4.3 Derivation 177
of the velocity ﬁeld v = ∇φ + u together with the divergencefree constraint ∇ u = 0.
Expecting that the rotational part u varies rapidly in a distance δ =
√
ν,
1
we introduce
the boundarylayer coordinate ζ ≡ (z + h)/δ, so that u = u(x, ζ) =
_
u
x
, u
z
_
(x, ζ). The
solid boundary is given by ζ = 0, and the potential part of the ﬂow is not subject to this
change of variables. With the new scaling, the divergencefree condition becomes
∂u
z
∂ζ
+ δ∇
x
u
x
= 0. (4.13)
As done in [Mei94], we expand the unknown functions in powers of the small parameter δ:
φ = φ
0
(x, z, t) + δφ
1
(x, z, t) + . . . , u = q
0
(x, ζ, t) + δq
1
(x, ζ, t) + . . . .
Substituting the expansion for u into (4.13) gives the following relations:
δ
0
:
∂q
0
z
∂ζ
= 0, δ
1
:
∂q
1
z
∂ζ
= −∇
x
q
0
x
, (4.14)
where q
0
x
denotes the ﬁrst two components of the vector q
0
corresponding to the horizon
tal coordinates x. Recall that we would like to determine the correction to the bottom
boundary condition
φ
z
= − u
z
[
ζ=0
= − (q
0
z
+ δq
1
z
)[
ζ=0
+ o(δ). (4.15)
So we only need to compute q
0
z
and q
1
z
at the bottom ζ = 0.
Using the same asymptotic considerations as above, we can write down the following
sequence of problems:
∆φ
0
= 0,
∂φ
0
∂z
¸
¸
¸
¸
z=−h
= 0,
∂q
0
∂t
=
∂
2
q
0
∂ζ
2
, q
0
= −∇φ
0
[
ζ=0
,
∆φ
1
= 0,
∂φ
1
∂z
¸
¸
¸
¸
z=−h
= −q
1
z
,
∂q
1
∂t
=
∂
2
q
1
∂ζ
2
, q
1
= −∇φ
1
[
ζ=0
,
together with the radiation condition q →
0 as ζ →∞.
This sequence of linear problems can be solved using Fourier transforms. In Fourier
space one ﬁnds immediately that
ˆ
φ
0
(t, z, k) = ˆ ϕ
0
(t, k)
_
e
kz
+ e
−kz
_
. Since we know
ˆ
φ
0
,
we can determine the rotational component
ˆ
q
0
.
Analytical solutions to the equation
∂
ˆ
q
0
x
∂t
=
∂
2 ˆ
q
0
x
∂ζ
2
(4.16)
1
Of course we should nondimensionalize all quantities in order to deﬁne small numbers. One would
ﬁnd that δ is in fact equivalent to
√
Re
−1
L, where Re is the Reynolds number and L a typical length.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
178 Viscous potential ﬂows
are wellknown. If we assume that initially the ﬂow is potential and the boundary condition
is
ˆ
q
0
x
= ik
ˆ
φ
0
(z = −h; k), the solution is
ˆ
q
0
x
=
1
2
√
π
t
_
0
ζ
(t −τ)
3
2
e
−
ζ
2
4(t−τ)
ik
ˆ
φ
0
(τ, z = −h, k) dτ.
Let us now integrate the second equation in (4.14) from 0 to ∞, using the appropriate
decay at inﬁnity:
ˆ q
1
z
[
ζ=0
= −
∞
_
0
ik
ˆ
q
0
x
dζ =
1
2
√
π
∞
_
0
t
_
0
ζ
(t −τ)
3
2
e
−
ζ
2
4(t−τ)
[k[
2
ˆ
φ
0
(τ, z = −h, k) dτ dζ.
One can interchange integral signs and evaluate the inner integral on ζ to obtain:
ˆ q
1
z
[
ζ=0
=
1
√
π
t
_
0
[k[
2
ˆ
φ
0
(τ, z = −h, k)
√
t −τ
dτ.
Hence, the bottom boundary condition becomes, at order δ,
∂φ
∂z
¸
¸
¸
¸
z=−h
= −
_
ν
π
t
_
0
T
−1
_
[k[
2
ˆ
φ
0
(−h, k)
_
√
t −τ
dτ =
=
_
ν
π
t
_
0
∇
2
x
φ
0
[
z=−h
√
t −τ
dτ = −
_
ν
π
t
_
0
φ
0zz
[
z=−h
√
t −τ
dτ. (4.17)
One recognizes on the righthand side a halforder integral operator. Summarizing the
developments made above and generalizing our equations by including nonlinear terms, we
obtain a new set of viscous potential freesurface ﬂow equations:
∆φ = 0, (x, z) ∈ Ω = R
2
[−h, η] (4.18)
η
t
+∇η ∇φ = φ
z
+ 2ν∆η, z = η (4.19)
φ
t
+
1
2
[∇φ[
2
+ gη = −2νφ
zz
, z = η (4.20)
φ
z
= −
_
ν
π
t
_
0
φ
zz
√
t −τ
dτ, z = −h. (4.21)
At the present stage, the addition of nonlinear terms is rather a conjecture. However, a
recent study by Liu et al. [LPC07] suggests that this conjecture is rather true. The authors
investigated the importance of nonlinearity in boundary layer equation (4.16) in the case
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
4.3 Derivation 179
of a solitary wave solution. They came to the conclusion that “the nonlinear eﬀects are
not very signiﬁcant”.
Using this weakly damped potential ﬂow formulation and the procedure of Boussinesq
equations derivation described in previous chapters (see Chapter 2, for example), one can
derive the following system of equations with horizontal velocity u
θ
deﬁned at the depth
z
θ
= −θh, 0 ≤ θ ≤ 1:
η
t
+∇ ((h + η)u
θ
) + h
3
_
θ
2
2
−θ +
1
3
_
∇
2
(∇ u
θ
) = 2ν∆η +
_
ν
π
t
_
0
∇ u
θ
√
t −τ
dτ, (4.22)
u
θt
+
1
2
∇[u
θ
[
2
+ g∇η −h
2
θ
_
1 −
θ
2
_
∇(∇ u
θt
) = 2ν∆u
θ
. (4.23)
4.3.1 Dissipative KdV equation
In this section we derive a viscous Kortewegde Vries (KdV) equation from just obtained
Boussinesq equations (4.22), (4.23). Since KdVtype equations model only unidirectional
wave propagation, our attention is naturally restricted to 1D case. In order to perform
asymptotic computations, all the equations have to be switched to nondimensional variables
as it is explained in Section 2.2. We ﬁnd the velocity variable u in this form:
u
θ
= η + εP + µ
2
Q + . . .
where ε and µ are deﬁned in (2.9), P and Q are unknown at the present moment. Using
the methods similar to those used in Section 2.9.1, one can easily show that
P = −
1
4
η
2
, Q =
_
θ −
1
6
−
θ
2
2
_
η
xx
.
This result immediately yields the following asymptotic representation of the velocity ﬁeld
u
θ
= η −
1
4
εη
2
+ µ
2
_
θ −
1
6
−
θ
2
2
_
η
xx
+ . . . (4.24)
Substituting the last formula (4.24) into equation (4.22) and switching again to dimensional
variables, one obtains this viscous KdVtype equation:
η
t
+
_
g
h
_
(h +
3
2
η)η
x
+
1
6
h
3
η
xxx
−
_
ν
π
t
_
0
η
x
√
t −τ
dτ
_
= 2νη
xx
. (4.25)
This equation will be used in Section 4.5 to study the damping of linear progressive waves.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
180 Viscous potential ﬂows
4.4 Dispersion relation
Interesting information about the governing equations can be obtained from the linear
dispersion relation analysis. In this section we are going to analyse the new set of equations
(4.18)–(4.21) for the complete water wave problem and the corresponding long wave limit
(Boussinesq equations) (4.22)–(4.23).
To simplify the computations, we consider the twodimensional problem. The general
ization to higher dimensions is straightforward. Traditionally the governing equations are
linearized and the bottom is assumed to be ﬂat. After all these simpliﬁcations the new set
of equations becomes
φ
xx
+ φ
zz
= 0, (x, z) ∈ R [−h, 0], (4.26)
η
t
= φ
z
+ 2νη
xx
, z = 0, (4.27)
φ
t
+ gη + 2νφ
zz
= 0, z = 0, (4.28)
φ
z
+
_
ν
π
t
_
0
φ
zz
√
t −τ
dτ = 0, z = −h. (4.29)
The next classical step consists in ﬁnding solutions of the special form
φ(x, z, t) = ϕ(z)e
i(kx−ωt)
, η(x, t) = η
0
e
i(kx−ωt)
. (4.30)
From continuity equation (4.26) we can determine the structure of the function ϕ(z):
ϕ(z) = C
1
e
kz
+ C
2
e
−kz
.
Altogether we have three unknown constants
C = (C
1
, C
2
, η
0
) and three boundary condi
tions (4.27)–(4.29) which can be viewed as a linear system with respect to
C:
M
C =
0. (4.31)
The matrix M has the following elements
M =
_
_
_
k −k iω −2νk
2
iω −2νk
2
iω −2νk
2
−g
e
−kh
_
1 + kF(t, ω)
_
e
kh
_
−1 + kF(t, ω)
_
0
_
_
_
where the function F(t, ω) can be expressed in terms of the error function of complex
argument:
F(t, ω) :=
_
ν
π
t
_
0
e
iω(t−τ)
√
t −τ
dτ =
_
ν
−iω
Erf(
√
−iωt).
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
4.4 Dispersion relation 181
In order to have nontrivial solutions
2
of (4.26)–(4.29), the determinant of the system (4.31)
has to be equal to zero det M = 0. It gives us a relation between ω and wavenumber k.
This relation is called the linear dispersion relation:
D(ω, k) := (iω−2νk
2
)
2
+gk tanh(kh) −
_
iν
ω
Erf(
√
−iωt)
_
(iω−2νk
2
) tanh(kh) +gk
_
≡ 0.
A similar procedure can be followed for Boussinesq equations (4.22), (4.23). We do not
give here the details of the computations but only the ﬁnal result:
D
b
(ω, k) := (iω−2νk
2
)
2
+b(kh)
2
iω(iω−2νk
2
)+ghk
2
(1−a(kh)
2
)−gk
2
_
iν
ω
Erf(
√
−iωt) ≡ 0,
where we introduced the following notation: a :=
θ
2
2
−θ +
1
3
, b := θ
_
1 −
θ
2
_
.
Remark 19 Contrary to the classical water wave problem and, by consequence, standard
Boussinesq equations
3
where the dispersion relation does not depend on time
ω
2
−gk tanh(kh) ≡ 0, (4.32)
here we have additionally the dependence of ω
t
(k) on time t as a parameter. It is a con
sequence of the presence of the nonlocal term in time in the bottom boundary condition
(4.21). Physically it means that the boundary layer “remembers” the ﬂow history.
Unfortunately, the relations D(ω, k) ≡ 0 and D
b
(ω, k) ≡ 0 cannot be solved analytically
to give an explicit dependence of ω on k. That is why we applied Newtontype method
to solve numerically these equations with respect to ω. Iterations were initialized with the
classical dispersion relation (4.32). The Jacobians can be computed exactly:
dD(ω)
dω
= 2i(iω −2νk
2
) −
k
ω
_
iν
ω
_
(iω −2νk
2
)
2
tanh(kh) + gk
__
1
√
π
e
iωt
√
−iωt
−
1
2
Erf(
√
−iωt)
_
−2i
_
iν
ω
k tanh(kh)(iω −2νk
2
) Erf(
√
−iωt),
dD
b
(ω)
dω
= 2i(iω −2νk
2
) + ib(kh)
2
+
gk
2
2ω
_
iν
ω
Erf(
√
−iωt) −
gk
2
ω
_
νt
π
e
iωt
.
2
Obviously of the special form given by equation (4.30).
3
Their dispersion relation is given by (2.50)
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
182 Viscous potential ﬂows
4.4.1 Discussion
Numerical snapshots of the nonclassical dispersion relation
4
at diﬀerent times for com
plete and Boussinesq equations are given on Figures (4.2)–(4.6). We will try to make
several comments on the results we obtained.
Just at the beginning (when t = 0), there is no eﬀect of the nonlocal term. This is
why on Figure 4.2 new and classical curves are superimposed. With no surprise, the phase
velocity of Boussinesq equations represents well only long waves limit (let us say up to
kh ≈ 2). When time evolves, we can see that the main eﬀect of nonlocal term consists
in slowing down long waves (see Figures 4.3–4.5). Namely, in the vicinity of kh = 0 the
real part of the phase velocity is slightly smaller with respect to the classical formulation.
From physical point of view this situation is comprehensible since only long waves “feel” the
bottom and, by consequence, are aﬀected by bottom boundary layer. On the other hand,
the imaginary part of the phase velocity is responsible for the wave amplitude attenuation.
The minimum of Imc
p
(k) in the region of long waves indicates that there is a “preferred”
wavelength which is attenuated the most. In the range of short waves the imaginary
part is monotonically decreasing. In practice it means that highfrequency components
are damped by the model. This property can be advantageous in numerics, for example.
On Figure 4.6 we depicted the real part of c
p
(k) with zoom made on long and moderate
waves. The reader can see that nonlocal complete and Boussinesq equations have similar
behaviour in the vicinity of kh = 0.
4.5 Attenuation of linear progressive waves
In this Section we investigate the damping rate of linear progressive waves. Thus, the
ﬁrst step consists in linearizing dissipative KdV equation (4.25):
η
t
+
_
g
h
_
hη
x
+
1
6
h
3
η
xxx
−
_
ν
π
t
_
0
η
x
√
t −τ
dτ
_
= 2νη
xx
(4.33)
In other words, we can say that we restrict our attention only to small amplitude waves.
Equation (4.33) can be called nonlocal dissipative Airy equation.
Now we make the next assumption. We look for a particular form of the solutions:
η(x, t) = /(t)e
ikξ
, ξ = x −
_
ght. (4.34)
where k is the wavenumber and /(t) is called the complex amplitude, since [η(x, t)[ =
[/(t)[. Integrodiﬀerential equation governing the temporal evolution of /(t) can be easily
4
To be precise, we plot the phase velocity which is deﬁned as c
p
(k) :=
ω(k)
k
.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
4.5 Attenuation of linear progressive waves 183
0 2 4 6 8 10 12 14
0.2
0.4
0.6
0.8
1
kh
R
e
c
p
(
k
)
/
(
g
h
)
1
/
2
Phase velocity at t = 0.00
Nonlocal equations
Classical equations
Nonlocal Boussinesq
Classical Boussinesq
0 2 4 6 8 10 12 14
−0.03
−0.02
−0.01
0
kh
I
m
c
p
(
k
)
/
(
g
h
)
1
/
2
Nonlocal equations
Nonlocal Boussinesq
Figure 4.2: Real and imaginary part of dispersion curve at t = 0. At the beginning, the
nonlocal term has no eﬀect. Thus, the real parts of the classical and new set of equations are
exactly superimposed on this ﬁgure. The imaginary part represents only local dissipation at this
stage.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
184 Viscous potential ﬂows
0 2 4 6 8 10 12 14
0.2
0.4
0.6
0.8
1
kh
R
e
c
p
(
k
)
/
(
g
h
)
1
/
2
Phase velocity at t = 1.00
Nonlocal equations
Classical equations
Nonlocal Boussinesq
Classical Boussinesq
0 2 4 6 8 10 12 14
−0.03
−0.02
−0.01
0
kh
I
m
c
p
(
k
)
/
(
g
h
)
1
/
2
Nonlocal equations
Nonlocal Boussinesq
Figure 4.3: Phase velocity at t = 1. Boundary layer eﬀects start to be visible: the real part
of the velocity slightly drops down and the straight lines of the imaginary part are deformed by
the nonlocal term. Within graphical accuracy, the classical and their nonlocal counterparts are
superimposed.
0 2 4 6 8 10 12 14
0.2
0.4
0.6
0.8
1
kh
R
e
c
p
(
k
)
/
(
g
h
)
1
/
2
Phase velocity at t = 2.00
Nonlocal equations
Classical equations
Nonlocal Boussinesq
Classical Boussinesq
0 2 4 6 8 10 12 14
−0.03
−0.02
−0.01
0
kh
I
m
c
p
(
k
)
/
(
g
h
)
1
/
2
Nonlocal equations
Nonlocal Boussinesq
Figure 4.4: Phase velocity at t = 2. Nonlocal term slows down long waves since the real part
of the phase velocity decreases.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
4.5 Attenuation of linear progressive waves 185
0 2 4 6 8 10 12 14
0.2
0.4
0.6
0.8
1
kh
R
e
c
p
(
k
)
/
(
g
h
)
1
/
2
Phase velocity at t = 4.00
Nonlocal equations
Classical equations
Nonlocal Boussinesq
Classical Boussinesq
0 2 4 6 8 10 12 14
−0.03
−0.02
−0.01
0
kh
I
m
c
p
(
k
)
/
(
g
h
)
1
/
2
Nonlocal equations
Nonlocal Boussinesq
Figure 4.5: Phase velocity at t = 4. On Figure 4.6 we plot a zoom on long and moderate waves.
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
0.7
0.75
0.8
0.85
0.9
0.95
1
kh
R
e
c
p
(
k
)
/
(
g
h
)
1
/
2
Phase velocity at t = 4.00
Nonlocal equations
Classical equations
Nonlocal Boussinesq
Classical Boussinesq
(a) Zoom on long waves. Classical and correspond
ing Boussinesq equations are almost superimposed
in this region.
2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9
0.58
0.6
0.62
0.64
0.66
0.68
0.7
kh
R
e
c
p
(
k
)
/
(
g
h
)
1
/
2
Phase velocity at t = 4.00
Nonlocal equations
Classical equations
Nonlocal Boussinesq
Classical Boussinesq
(b) Zoom on moderate wavelengths. In this region
classical and nonlocal complete equations are almost
superimposed. In the same time one can notice a
little diﬀerence in Boussinesq models.
Figure 4.6: Real part of the phase velocity at t = 4.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
186 Viscous potential ﬂows
derived by substituting the special representation (4.34) into linearized KdV equation
(4.33):
d/
dt
−
i
6
_
g
h
(kh)
3
/(t) + 2νk
2
/(t) −ik
_
gν
πh
t
_
0
/(τ)
√
t −τ
dτ = 0. (4.35)
In our applications we are rather interested in temporal evolution of the absolute value
[/(t)[. It is straightforward to derive the equation for [/(t)[
2
if we recall that
[/(t)[
2
= /(t)
¯
/(t) =⇒
d[/(t)[
2
dt
=
d/(t)
dt
¯
/(t) +
d
¯
/(t)
dt
/(t), (4.36)
where
¯
/(t) means the complex conjugate of /(t). The last missing ingredient is the time
derivative
d
¯
A
dt
. This term is easily computed by taking the complex conjugate of (4.35):
d
¯
/
dt
+
i
6
_
g
h
(kh)
3
¯
/(t) + 2νk
2
¯
/(t) + ik
_
gν
πh
t
_
0
¯
/(τ)
√
t −τ
dτ = 0.
By combining just obtained equations according to (4.36), we get required evolution equa
tion for [/[
2
:
d[/[
2
dt
+ 4νk
2
[/(t)[
2
−ik
_
gν
πh
t
_
0
¯
/(t)/(τ) −/(t)
¯
/(τ)
√
t −τ
dτ = 0.
If we denote by /
r
(t) and /
i
(t) real and imaginary parts of /(t) respectively, the last
equation can be further simpliﬁed:
d[/[
2
dt
+ 4νk
2
[/(t)[
2
+ 2k
_
gν
πh
t
_
0
/
r
(t)/
i
(τ) −/
i
(t)/
r
(τ)
√
t −τ
dτ = 0. (4.37)
Just derived integrodiﬀerential equation represents a generalisation to the classical equa
tion (4.1) by Boussinesq [Bou95] and Lamb [Lam32] for the wave amplitude evolution in
a viscous ﬂuid. We recall that novel integral term is a direct consequence of the bottom
boundary layer modelling.
Unfortunately, equation (4.37) cannot be used directly for numerical computations
since we need to know the following combination of real and imaginary parts /
r
(t)/
i
(τ) −
/
i
(t)/
r
(τ) for τ ∈ [0, t]. It represents a new and nonclassical aspect of the present theory.
In numerical computations it is advantageous to integrate exactly local terms in equa
tion (4.35). It is done by making the following change of variables:
/(t) = e
−2νk
2
t
e
i
6
√
g
h
(kh)
3
t
˜
/(t).
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
4.6 Numerical results 187
parameter deﬁnition value
ν eddy viscosity 10
−3 m
2
s
g gravity acceleration 9.8
m
s
2
h water depth 3600 m
ℓ wavelength 50 km
k wavenumber =
2π
ℓ
m
−1
Table 4.1: Values of the parameters used in the numerical computations of the linear progressive
waves amplitude. These values correspond to a typical Indian Ocean tsunami.
One can easily show that new function
˜
A(t) satisﬁes the following equation:
d
˜
/
dt
= ik
_
gν
πh
t
_
0
e
2νk
2
(t−τ)
e
−
i
6
√
g
h
(kh)
3
(t−τ)
√
t −τ
˜
/(τ) dτ
On Figure 4.7 we plot a solution of integrodiﬀerential equation (4.35). All parameters
related to this case are given in Table 4.1. These values were chosen to simulate a typical
tsunami in Indian Ocean [DD06]. We have to say that the wave amplitude damping is
entirely due to the dissipation in boundary layer since local terms are unimportant for
suﬃciently long waves.
4.6 Numerical results
In this section we would like to show the eﬀect of nonlocal term on the solitary wave
attenuation. For simplicity, we will consider wave propagation in a 1D channel.
For numerical computations we use the same Fouriertype spectral method that was
described in Section 2.8. Obviously this method has to be slightly adapted because of the
presence of nonlocal in time term. We have to say that this term necessitates the storage
of ∇ u
(n)
at previous time steps. Hence, long computations can be memory consuming.
The values of all parameters are given in Table 4.2.
4.6.1 Approximate solitary wave solution
In order to provide an initial condition for equations (4.22), (4.23), we are going to ob
tain an approximate solitary wave solution for nondissipative 1D version of these equations
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
188 Viscous potential ﬂows
0 1 2 3 4 5 6
0.9
0.92
0.94
0.96
0.98
1
1.02
t, hours

A
(
t
)

/

A
(
0
)

Amplitude of linear progressive waves
Figure 4.7: Amplitude of linear progressive waves as a function of time. Values of all parameters
are given in Table 4.1.
parameter deﬁnition value
ε nonlinearity 0.02
µ dispersion 0.06
ν eddy viscosity 0.001
c soliton velocity 1.02
θ z
θ
= −θh 1 −
√
5/5
x
0
soliton center at t = 0 −1.5
Table 4.2: Values of the parameters used in the numerical computations
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
4.6 Numerical results 189
over the ﬂat bottom:
η
t
+
_
(1 + εη)u
_
x
+ µ
2
_
θ
2
2
−θ +
1
3
_
u
xxx
= 0,
u
t
+ η
x
+
ε
2
(u
2
)
x
−µ
2
θ
_
1 −
θ
2
_
u
xxt
= 0.
Then, we apply the same approach that in Section 2.9.1. We do not provide the
computations here since they are simple and can be done without any diﬃculties. The
ﬁnal result is the following:
η(x, t) =
2(c −1)
ε
sech
2
_
_
6(c −1)
2µ
(x + x
0
−ct)
_
and the velocity is given by this formula:
u = η −
1
4
εη
2
+ µ
2
_
θ −
θ
2
2
−
1
6
_
η
xx
+O(ε
2
+ εµ
2
+ µ
4
).
In the numerical results presented here, we use η(x, 0) and u(x, 0) as initial conditions.
4.6.2 Discussion
On Figures 4.8–4.10 we present three curves. They depict the free surface elevation
according to three diﬀerent formulations. The ﬁrst corresponds to classical Boussinesq
equations without dissipation. The second one to dissipative system with diﬀerential terms
(for example ν∆u in momentum conservation equation) and the third curve corresponds
to equations (4.22), (4.23). On Figure 4.11 we made a zoom on the soliton crest.
It can be seen that Boussinesq equations with nonlocal term provide stronger atten
uation of the amplitude. In the same time, as it was shown in the previous section, this
nonlocal term slightly slows down the solitary wave.
In order to show explicitly the rate of amplitude attenuation, we plot on Figure 4.12 the
graph of the following application t → sup
−π<x<π
[η(x, t)[. One can see on this plot little
oscillations which are of numerical nature. Our numerical experiments show that their
amplitude decreases when N →∞. This result shows one more time that nonlocal model
provides stronger damping properties. One can have the impression that the amplitude
decays linearly but it is only an impression because of (4.2). This behaviour for moderate
t can be explained by simple Taylor expansion which is valid when νk
2
t ≪1:
α(t) = α
0
e
−2νk
2
t
= α
0
(1 −2νk
2
t) +O
_
ν
2
k
4
t
2
_
.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
190 Viscous potential ﬂows
−3 −2 −1 0 1 2 3
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
t = 0.500
x
η
No dissipation
Local
Nonlocal
Figure 4.8: Comparison among two dissipative and nondissipative Boussinesq equations. Snap
shots of the free surface at t = 0.5
−3 −2 −1 0 1 2 3
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
t = 1.000
x
η
No dissipation
Local
Nonlocal
Figure 4.9: Comparison among two dissipative and nondissipative Boussinesq equations. Snap
shots of the free surface at t = 1.0
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
4.6 Numerical results 191
−3 −2 −1 0 1 2 3
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
t = 2.000
x
η
No dissipation
Local
Nonlocal
Figure 4.10: Comparison among two dissipative and nondissipative Boussinesq equations. Snap
shots of the free surface at t = 2.0
0 0.2 0.4 0.6 0.8 1 1.2
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
t = 2.000
x
η
No dissipation
Local
Nonlocal
Figure 4.11: Comparison among two dissipative and nondissipative Boussinesq equations. Zoom
on the soliton crest at t = 2.0
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
192 Viscous potential ﬂows
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
1.93
1.94
1.95
1.96
1.97
1.98
1.99
2
2.01
t
m
a
x
η
(
x
,
t
)
Maximum wave amplitude as function of time
Nonviscous model
Nonlocal equations
Local terms
Figure 4.12: Amplitude of the solitary wave as function of time
4.7 Conclusion
In the present chapter we have shown how to express the rotational component of
the velocity ﬁeld in terms of the potential part of Helmholtz–Leray decomposition. This
expression contains diﬀerential and integral operators. Obviously, this analysis is only
linear but we give nonlinear equations. In future work we will try to extend the present
derivation to the nonlinear case. A long wave approximation was derived from this new
potential ﬂow formulation.
Our results are diﬀerent from [KM75]. This discrepancy can be explained by a diﬀer
ent scaling chosen by Kakutani & Matsuuchi in the boundary layer. Consequently, their
governing equations contain a nonlocal term in space. The performance of the present non
local term (4.17) was studied in [LSVO06]. The authors carried out in a wave tank a set
of experiments, analyzing the damping and shoaling of solitary waves. It is shown that the
viscous damping due to the bottom boundary layer is well represented. Their numerical
results ﬁt very well with the experiments. The numerical model not only properly predicts
the wave height at a given point but also provides a good representation of the changes on
the shape and celerity of the soliton. We can conclude that the experimental study by P.
Liu [LSVO06] completely validates our theory.
It is interesting to note that local dissipative terms of this form have been used to verify
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
4.7 Conclusion 193
the theory of weak turbulence of surface gravity waves in deep water
5
[DKZ03, DKZ04,
ZKPD05]. They were added without justiﬁcation to model dissipation at small scales.
Note that a good qualitative agreement was obtained between the Kolmogorov spectrum
predicted by weak turbulence theory and the results of DNS. Hence, the present work can
be considered as an attempt to justify the inclusion of these terms.
Our ﬁnal remark concerns the nonlocal term in the kinematic bottom boundary condi
tion. This term can be also considered as a boundary layer correction at the bottom. In
modelling viscous eﬀects this term plays the main role, since its magnitude is O(
√
ν). Of
course, the numerical implementation of this term is another matter. It is interesting to
note that the boundary layer eﬀect is not instantaneous but rather cumulative. The ﬂow
history is weighted by (t −τ)
−
1
2
in favour of the current time t. As pointed out in [LO04],
this nonlocal term is essential to have an accurate estimation of the bottom shear stress
based on the calculated wave ﬁeld above the bed. Then, this information can be used for
calculating sedimentbedload transport ﬂuxes and, in turn, morphological changes.
We would like to mention here a paper of N. Sugimoto [Sug91]. He considered initial
value problems for the Burgers equation with the inclusion of a hereditary integral known
as the fractional derivative of order
1
/
2
. The form of this term was not justiﬁed in that
work. Note, that from fractional calculus point of view our nonlocal term is a halforder
integral.
The dispersion relation associated with a new nonlocal formulation was described. Due
to the presence of special functions, we cannot obtain a simple analytical dependence of
the frequency ω on the wavenumber k as in classical equations. Consequently the disper
sion curve was obtained numerically by a Newtontype method. We made a comparison
between the phase velocity of the complete viscopotential problem and the corresponding
Boussinesq equations. The eﬀect of the nonlocal term on the solitary wave attenuation
was investigated numerically with a Fouriertype spectral method.
What is the value of ν to be taken in numerical simulations? There is surprisingly little
published information of this subject. What is clear is that the molecular diﬀusion is too
small to model true viscous damping and one should rather consider the eddy viscosity
parameter.
5
Since the water is considered to be of inﬁnite depth, they do not have the bottom boundary layer
phenomenon. Hence, the local viscous terms are very natural choice to model dissipation at Kolmogorov
scale.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
194 Viscous potential ﬂows
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
Direction for future research
Progress imposes not only new possibilities
for the future but new restrictions.
Norbert Wiener (1894 – 1964)
A useful direction for future research in the dynamics of tsunami waves is the three
dimensional (3D) simulation of tsunami breaking along a coast. For this purpose, diﬀerent
validation steps are necessary. First more simulations of a twodimensional (2D) tsunami
interacting with a sloping beach ought to be performed. Then these simulations should be
extended to the case of a 2D tsunami interacting with a sloping beach in the presence of
obstacles. An important output of these computations will be the hydrodynamic loading
on obstacles. The nonlinear inelastic behaviour of the obstacles may be accounted for using
damage or plasticity models. The development of Boussinesq type models coupled with
structure interactions is also a promising task. Finally there is a need for 3D numerical
simulations of a tsunami interacting with a beach of complex bathymetry, with or without
obstacles. These simulations will hopefully demonstrate the usefulness of numerical simu
lations for the deﬁnition of protecting devices or security zones. An important challenge
in that respect is to make the numerical methods capable of handling interaction problems
involving diﬀerent scales: the ﬁne scale needed for representing the damage of a ﬂexible
obstacle and a coarse scale needed to quantify the tsunami propagation.
195
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
196 Direction for future research
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
Appendix A
Finite Volumes: Characteristic ﬂux
approach
Most of the fundamental ideas of science are essentially simple, and may, as a rule, be expressed
in a language comprehensible to everyone.
Albert Einstein (1879 – 1955)
Contents
A.1 Discretization in the ﬁnite volume framework . . . . . . . . . 198
A.1.1 The one dimensional case . . . . . . . . . . . . . . . . . . . . . . 198
A.1.2 Extension to the multidimensional case . . . . . . . . . . . . . . 201
A.1.3 On the discretization of source terms . . . . . . . . . . . . . . . . 202
A.2 On the discretization of boundary conditions . . . . . . . . . . 204
The VFFC
1
approach is based on a discretization of the continuous models throught
a cellcentered ﬁnite volume method. More precisely, we discretize space derivatives using
ﬁnite volumes and time derivatives by classical ﬁnite diﬀerences. So at a ﬁxed time step,
the solution is represented by a collection of numbers which intend to approximate the
mean values of the physical solution on the control volumes. As it is well known, the heart
of the matter consists in computing from these degrees of freedom the values of the physical
ﬂuxes on the boundaries of the control volumes.
These ﬂuxes are of two diﬀerent nature: convective ﬂuxes and viscous (or diﬀusive)
ﬂuxes. We know from many instances steaming from single phase ﬂuid ﬂows that the
discretization of these ﬂuxes must be adapted to their physical origin. More precisely on
1
Volumes Finis `a Flux Caract´eristiques
197
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
198 VFFC scheme
the one hand, discretization of convective ﬂuxes must take into account the wave phenom
ena associated with convection, that is upwind based methods. On the other hand the
discretization of viscous ﬂuxes can rely on centered methods.
In this chapter we are going only to deal with convective ﬂuxes. Moreover we keep
the timevariable continuous in order to clearly separate the features of space and time
discretizations.
A.1 Discretization in the ﬁnite volume framework
In this section we present a method [GKC96, GKC01] for the discretization of hyper
bolic systems in the one dimensional case. The extension to the multidimensional case is
discussed in Section A.1.2.
A.1.1 The one dimensional case
Let us consider general systems that can be written as follows:
∂v
∂t
+
∂f(v)
∂x
= 0 , (A.1)
where v ∈ R
m
and f : R
m
→ R
m
. We denote by A(v) the jacobian matrix
∂f(v)
∂v
and we
deal ﬁrst with the case where (A.1) is smoothly hyperbolic that is to say: for every v there
exists a smooth basis (r
1
(v), . . . , r
m
(v)) of R
m
consisting of eigenvectors of A(v) : ∃λ
k
(v) ∈
R such that A(v)r
k
(v) = λ
k
(v)r
k
(v). It is then possible to construct (l
1
(v), . . . , l
m
(v)) such
that
t
A(v)l
k
(v) = λ
k
(v)l
k
(v) and l
k
(v) r
p
(v) = δ
k,p
.
Let R = ∪
j∈Z
[x
j−1/2
, x
j+1/2
] be a one dimensional mesh. Our goal is to discretize (A.1)
by a ﬁnite volume method. We set ∆x
j
≡ x
j+1/2
− x
j−1/2
, ∆t
n
≡ t
n+1
− t
n
(we also have
R
+
= ∪
n∈N
[t
n
, t
n+1
]) and
˜ v
n
j
≡
1
∆x
j
_
x
j+1/2
x
j−1/2
v(x, t
n
) dx,
˜
f
n
j+1/2
≡
1
∆t
n
_
t
n+1
t
n
f(v(x
j+1/2
, t)) dt .
With these notations, we deduce from (A.1) the exact relation:
˜ v
n+1
j
= ˜ v
n
j
−
∆t
n
∆x
j
_
˜
f
n
j+1/2
−
˜
f
n
j−1/2
_
. (A.2)
Since the (
˜
f
n
j+1/2
)
j∈Z
cannot be expressed in terms of the (˜ v
n
j
)
j∈Z
, one has to make an
approximation. In order to keep a compact stencil, it is more eﬃcient to use a three points
scheme: the physical ﬂux
˜
f
n
j+1/2
is approximated by a numerical ﬂux g
n
j
(v
n
j
, v
n
j+1
). Let us
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
A.1 Discretization in the ﬁnite volume framework 199
show how we construct this ﬂux here. We observe that since A(v)
∂v
∂t
=
∂f(v)
∂t
then according
to (A.1),
∂f(v)
∂t
+ A(v)
∂f(v)
∂x
= 0 . (A.3)
This shows that the ﬂux f(v) is advected by A(v) (like v since we also have
∂v
∂t
+A(v)
∂v
∂x
=
0). The numerical ﬂux g
n
j
(v
n
j
, v
n
j+1
) represents the ﬂux at an interface. Using a mean value
µ
n
j+1/2
of v at this interface, we replace (A.3) by the linearization:
∂f(v)
∂t
+ A(µ
n
j+1/2
)
∂f(v)
∂x
= 0 . (A.4)
It follows that, deﬁning the kth characteristic ﬂux component to be f
k
(v) ≡ l
k
(µ
n
j+1/2
)
f(v), we have
∂f
k
(v)
∂t
+ λ
k
(µ
n
j+1/2
)
∂f
k
(v)
∂x
= 0 . (A.5)
This linear equation can be solved explicitly now and we have:
f
k
(v)(x, t) = f
k
(v)(x −λ
k
(µ
n
j+1/2
)(t −t
n
), t
n
) . (A.6)
From this equation it is then natural to introduce the following deﬁnition.
Deﬁnition 2 For the conservative system (A.1), at the interface between the two cells
[x
j−1/2
, x
j+1/2
] and [x
j+1/2
, x
j+3/2
], the characteristic ﬂux g
CF
is deﬁned by the following
formula
_
we take µ
n
j+1/2
≡
_
∆x
j
v
n
j
+ ∆x
j+1
v
n
j+1
_
/
_
∆x
j
+ ∆x
j+1
_
_
: for k ∈ ¦1, . . . , m¦,
l
k
(µ
n
j+1/2
) g
CF,n
j
(v
n
j
, v
n
j+1
) = l
k
(µ
n
j+1/2
) f(v
n
j
) , when λ
k
(µ
n
j+1/2
) > 0 ,
l
k
(µ
n
j+1/2
) g
CF,n
j
(v
n
j
, v
n
j+1
) = l
k
(µ
n
j+1/2
) f(v
n
j+1
) , when λ
k
(µ
n
j+1/2
) < 0 , (A.7)
l
k
(µ
n
j+1/2
) g
CF,n
j
(v
n
j
, v
n
j+1
) = l
k
(µ
n
j+1/2
)
_
f(v
n
j+1
) + f(v
n
j
)
2
_
,
when λ
k
(µ
n
j+1/2
) = 0.
Remark 20 At ﬁrst glance, the derivation of (A.3) from (A.1), is only valid for continuous
solutions since A(v)
∂f(v)
∂x
is a non conservative product. In fact equation (A.3) can be
justiﬁed even in the case of shocks as is proved in [Ghi98]. Let us brieﬂy recall here the
key point. Assuming that the solution undergoes a discontinuity along a family of disjoint
curves, we can focus on one of these curves that we parameterize by the time variable t.
Hence, locally, on each side of this curve, v(x, t) is smooth and jumps across the curve
x = Σ(t). The RankineHugoniot condition implies that f(v(x, t)) − σ(t)v(x, t), where
σ(t) ≡
dΣ(t)
dt
, is smooth across the discontinuity curve and therefore A(v)
∂f(v)
∂x
can be deﬁned
as A(v)
∂f(v)
∂x
≡ A(v)
∂(f(v)−σv)
∂x
+ σ
∂f(v)
∂x
.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
200 VFFC scheme
Proposition 1 Formula (A.7) can be written as follows: g
CF,n
j
(v
n
j
, v
n
j+1
) = g
CF
(µ
n
j
; v
n
j
, v
n
j+1
)
where
g
CF
(µ; v, w) ≡
λ
k
(µ)<0
(l
k
(µ) f(w))r
k
(µ) +
λ
k
(µ)=0
_
l
k
(µ)
f(v) + f(w)
2
_
r
k
(µ)+
+
λ
k
(µ)>0
(l
k
(µ) f(v))r
k
(µ) . (A.8)
Proof. This comes from the useful identity valid for all vectors Φ and µ in R
m
:
Φ =
k=m
k=1
(l
k
(µ) Φ)r
k
(µ). We also observe that (A.8) can be written under the following
condensed form:
g
CF
(µ; v, w) =
f(v) + f(w)
2
−U(µ; v, w)
f(w) −f(v)
2
, (A.9)
where U(µ; v, w) is the sign of the matrix A(µ) which is deﬁned by
sign(A(µ))Φ =
m
k=1
sign(λ
k
)(l
k
(µ) Φ)r
k
(µ).
The form (A.9) refers to what we have called a numerical ﬂux leading to a ﬂux scheme
([Ghi98]).
Remark 21 Let us discuss the relation, in the conservative case, between the characteristic
numerical ﬂux g
CF
and the numerical ﬂux leading to Roe’s scheme [Roe81]. This later
scheme relies on an algebraic property of the continuous ﬂux f(v) which is as follows. It
is assumed that for all admissible states v and w, there exists a mm matrix A
ROE
(v, w)
such that f(v) − f(w) = A
ROE
(v, w)(v − w) (Roe’s identity). Then the numerical ﬂux
leading to Roe’s scheme is given by:
g
ROE
(v, w) =
f(v) + f(w)
2
−[A
ROE
(v, w)[
w −v
2
. (A.10)
But using Roe’s identity, we obtain that
g
ROE
(v, w) =
f(v) + f(w)
2
−sign(A
ROE
(v, w))
f(w) −f(v)
2
, (A.11)
which is of the form (A.9): Roe’s scheme is also a ﬂux scheme. The characteristic ﬂux
proposed in this paper is more versatile than Roe’s scheme in the sense that it does not rely
on an algebraic property of the ﬂux. Hence for complex systems (like those encountered
in the context of two phase ﬂows) this scheme appears like an eﬃcient generalization of
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
A.1 Discretization in the ﬁnite volume framework 201
Roe’s scheme. Moreover, as we shall see below, this scheme has a natural generalization
to arbitrary non conservative systems. Finally, the fact that the numerical ﬂux is a linear
combination of the two ﬂuxes induces a quite weak dependance on the state µ which appears
in formula (A.8), see [CG00].
Combining (A.2) and (A.7), we arrive to the explicit scheme:
v
n+1
j
= v
n
j
−
∆t
n
∆x
j
_
g
CF,n
j
(v
n
j
, v
n
j+1
) −g
CF,n
j
(v
n
j−1
, v
n
j
)
_
. (A.12)
A.1.2 Extension to the multidimensional case
Let us consider a system of m balance equations: (v = (v
1
, . . . , v
m
) ∈ R
m
))
∂v
∂t
+∇ F(v) = 0, (A.13)
here ∇ F(v) =
nd
j=1
∂F
j
(v)
∂x
j
, where F
j
maps G into R
m
, where G is an open subset of
R
m
corresponding to the physically admissible states. This equation is posed in a nd
dimensional domain Ω (nd = 1, 2 or 3 in practice).
We assume that the computational domain Ω is decomposed in smaller volumes (the
socalled control volumes) K : Ω = ∪
K∈T
K and consider ﬁrst the case where Ω = ∪
K∈T
K
is “conformal” i.e. that it is a ﬁnite element triangulation of Ω. In practice one can use
triangles for nd = 2 and tetrahedrons for nd = 3. The cellcentered ﬁnite volume approach
for solving (A.13) consists in approximating the means
v
K
(t) ≡
1
vol(K)
_
K
v(x, t) dx, (A.14)
where vol(K) denotes the nddimensional volume of K and area(A) stands for the (nd−1)
dimensional volume of an hypersurface A. Integrating (A.13) on K makes the normal
ﬂuxes, F
ν
∂K
, appear
F
ν
∂K
(t) =
_
∂K
F(v(σ, t)).ν(σ) dσ,
where ∂K is the boundary of K, ν(σ) the unit external normal on ∂K and dσ denotes the
(nd1)volume element on this hypersurface.
The heart of the matter in ﬁnite volume methods consists in providing a formula for the
normal ﬂuxes F
ν
∂K
in terms of the ¦v
L
¦
L∈T
. Assuming that the control volumes K are poly
hedra, as is most often the case, the boundary ∂K is the union of hypersurfaces K∩L where
L belongs to the set ^(K), the set of L ∈ T , L ,= K, such that K∩L has positive (nd−1)
measure. We can therefore decompose the normal ﬂux as a sum: F
ν
∂K
=
L∈N(K)
F
K,L
, where
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
202 VFFC scheme
(ν
K,L
points into L): F
K,L
=
_
K∩L
F(v(σ, t)).ν
K,L
dσ . Inspired by the 1Dcase, we take an
approximation of F
K,L
in terms of v
K
and v
L
: F
K,L
≈ area(K ∩L) Φ(v
K
, v
L
; K, L) , where
Φ is the numerical ﬂux that we construct by the following formula.
Deﬁnition 3 The numerical ﬂux of the Finite Volume method with characteristic ﬂux is
obtained by the formula
Φ(v, w; K, L) =
F(v) + F(w)
2
ν
K,L
−U(v, w; K, L)
F(w) −F(v)
2
ν
K,L
, (A.15)
when we take:
U(v, w; K, L) = sign(A
ν
K,L
(µ(v, w; K, L))) , (A.16)
where µ(v, w; K, L) is a mean between v
K
and v
L
which only depends on the geometry of
K and L:
µ(v, w; K, L) =
vol(K)v + vol(L)w
vol(K) + vol(L)
, (A.17)
and A
ν
(v) ≡
∂F(v)·ν
∂v
.
This allows us to generalize the explicit scheme (A.12) to the multidimensional case as
follows.
Deﬁnition 4 The explicit multidimensional characteristic ﬂux for the approximation of
equation (A.13) reads as follows:
v
n+1
K
= v
n
K
−
∆t
n
vol(K)
L∈N(K)
area(K ∩ L)Φ(v
n
K
, v
n
L
; K, L). (A.18)
A.1.3 On the discretization of source terms
Let us return to the 1D setting. Instead of (A.1), we want to solve
∂v
∂t
+
∂f(v)
∂x
= S, (A.19)
by the numerical scheme:
v
n+1
j
= v
n
j
−
∆t
n
∆x
j
_
f
n
j+1/2
−f
n
j−1/2
_
+ ∆t
n
Σ
n
j
. (A.20)
Equation (A.20) can be obtained by integration of (A.19) on the rectangle [t
n
, t
n+1
]
_
x
j−1/2
, x
j+1/2
¸
and this leads to the following form for Σ
n
j
:
Σ
n
j
= Σ
j
=
1
∆x
j
_
x
j+1/2
x
j−1/2
S(x)dx. (A.21)
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
A.1 Discretization in the ﬁnite volume framework 203
Actually, this formula is not suitable when the source term is large and one must modify
(A.21) according to the expression of the numerical ﬂux g
j
where f
n
j+1/2
= g
j
(v
n
j+1
, v
n
j
) and
g(v, w) =
f(v) + f(w)
2
−U(v, w)
f(w) −f(v)
2
, (A.22)
where U(, ) is a matrix to be speciﬁed.
Due to the fact that the numerical ﬂux is not centered (upwind bias) one can observe
large errors on the permanent solution under investigation. Let us introduce the notion of
enhanced consistency as follows. Denoting by
φ = φ
conv.
(x) ≡
_
x
0
S(y)dy , (A.23)
we say that Σ
n
j
satisﬁes, with respect to the numerical ﬂux g, the enhanced consistency
property when we have: if at some timestep, v
n
j
is such that
f(v
n
j
) =
1
∆x
j
_
x
j+1/2
x
j−1/2
φ
conv.
(x)dx, ∀j ∈ Z, (A.24)
then v
n+1
j
given by (A.20) must be equal to v
n
j
. This condition can be equivalently formu
lated as
Σ
n
j
=
f
n
j+1/2
−f
n
j−1/2
∆x
j
, (A.25)
if v
n
j
satisﬁes (A.24).
In [AGT99], the following results are shown.
Theorem 3 Let g be given in the form (A.22) and denote by U
n
j
= U(v
n
j
, v
n
j+1
). The
enhanced consistency will be satisﬁed if we discretize the forcing term S according to the
following formula
Σ
n
j
=
1
2∆x
j
_
_
x
j+1
x
j−1
S(y)dy −U
n
j
_
x
j+1
x
j
S(y)dy + U
n
j−1
_
x
j
x
j−1
S(y)dy
_
. (A.26)
Corollary 1 Assuming that S is given by a piecewise constant function: S(x) = S
j
for
x ∈
¸
x
j−1/2
, x
j+1/2
_
, the formula (A.26) reads
Σ
n
j
=
I + U
n
j−1
4
∆x
j−1
∆x
j
S
j−1
+
I −U
n
j
+ U
n
j−1
2
S
j
+
I −U
n
j
4
∆x
j+1
∆x
j
S
j+1
. (A.27)
In order to illustrate this formula, let us show what it means in the case of a single
linear equation v
t
+ c v
x
= S where e.g. c > 0 and v ∈ R. Here the usual upwind scheme
amounts to take U
n
j
= 1, i.e. g(v, w) = c v and (A.27) reads as:
Σ
n
j
=
1
2
∆x
j−1
∆x
j
S
j−1
+
1
2
S
j
. (A.28)
These results were one of the key points in the solution to the simulation of a boiling
tube (see [TF98]).
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
204 VFFC scheme
A.2 On the discretization of boundary conditions
So far we have not discussed the implementation of boundary conditions. This is a
very important topic since they actually determine the solution. Let us consider the space
discretization of the system (A.1) by our cell centered ﬁnite volume method. For instance
for the time explicit discretization we have the scheme (A.18). Of course this formula
is not valid when K meets the boundary of Ω. When this occurs, we have to ﬁnd the
numerical ﬂux Φ(v
n
K
, K, ∂Ω). In practice, this ﬂux is not given by the physical boundary
conditions and moreover, in general, (A.1) is an illposed problem if we try to impose
either v or F(v) ν on ∂Ω. This can be understood in a simple way by using the following
linearization of this system:
∂v
∂t
+ A
ν
∂v
∂ν
= 0 , (A.29)
where ν represents the direction of the external normal on K ∩ ∂Ω, A
ν
is the advection
matrix:
A
ν
≡
∂F(v) ν
∂v
[
v=v
, (A.30)
and v is the state around which the linearization is performed. When (A.1) is hyperbolic,
the matrix A
ν
is diagonalizable on R and by a change of coordinates, this system becomes
an uncoupled set of m advection equations:
∂ξ
k
∂t
+ c
k
∂ξ
k
∂ν
, k = 1, . . . , m. (A.31)
Here the c
k
are the eigenvalues of A
ν
and according to the sign of these numbers, waves
are going either into the domain Ω (c
k
< 0) or out of the domain Ω (c
k
> 0). Hence
we expect that it is only possible to impose p conditions on K ∩ ∂Ω where p ≡ ♯¦k ∈
¦1, . . . , m¦ such that c
k
< 0¦.
Let us consider now a control volume K which meets the boundary ∂Ω. We take v = v
n
K
and write the previous linearization. We denote by x the coordinate along the outer normal
so that (A.29) reads:
∂v
∂t
+ A
ν
∂v
∂x
= 0 , (A.32)
which happens to be the linearization of the 1D (i.e. when nd = 1) system. First we label
the eigenvalues c
k
(v) of A
ν
by increasing order:
c
1
(v) ≤ c
2
(v) ≤ . . . ≤ c
p
(v) < 0 ≤ c
p+1
(v) . . . ≤ c
m
(v) . (A.33)
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
A.2 On the discretization of boundary conditions 205
(i) The case p = 0. In this case information comes from inside Ω and therefore we take:
Φ(v
n
K
, K, ∂Ω) = F(v
n
K
) ν
K
. (A.34)
In the Computational Fluid Dynamics literature this is known as the “supersonic
outﬂow” case.
(ii) The case p = m. In this case information come from outside Ω and therefore we take:
Φ(v
n
K
, K, ∂Ω) = Φ
given
, (A.35)
where Φ
given
are the given physical boundary conditions. In the Computational Fluid
Dynamics literature this is known as the “supersonic inﬂow” case.
(iii) The case 1 ≤ p ≤ m−1. As already discussed, we need p scalar information coming
from outside of Ω. Hence we assume that we have on physical ground p relations on
the boundary:
g
l
(v) = 0 , l = 1, . . . , p. (A.36)
Remark 22 The notation g
l
(v) = 0 means that we have a relation between the components
of v. However, in general, the function g
l
is not given explicitly in terms of v. For example
g
l
(v) could be the pressure which is not, in general, one of the components of v.
Since we have to determine the m components of Φ(v
n
K
, K, ∂Ω), we need m − p
supplementary scalar conditions. Let us write them as
h
l
(v) = 0 , l = p + 1, . . . , m. (A.37)
In general (A.36) are named as “physical boundary conditions” while (A.37) are
named as “numerical boundary conditions”.
Then we take:
Φ(v
n
K
, K, ∂Ω) = F(v) ν
K
, (A.38)
where v is solution to (A.36)(A.37) (see however Remark 25 and (A.44)).
Remark 23 The system (A.36)(A.37) for the m unknownsv ∈ G is a m m nonlinear
system of equations. We are going to study its solvability, see Theorem 4.
Let us ﬁrst discuss the numerical boundary conditions (A.37). By analogy with what
we did on an interface between two control volumes K and L, we take (recall that v = v
n
K
):
˜
l
k
(v) (F(v) ν
K
) =
˜
l
k
(v) (F(v
n
K
) ν
K
) , k = p + 1, . . . , m. (A.39)
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
206 VFFC scheme
In other words, we set h
k
(v) ≡
˜
l
k
(v
n
K
) (F(v) ν
K
) −
˜
l
k
(v
n
K
) (F(v
n
K
) ν
K
). We have
denoted by (
˜
l
1
(v), . . . ,
˜
l
m
(v)) a set of left eigenvectors of
˜
A
ν
:
t
˜
A
ν
l
k
(v) = c
k
l
k
(v) and by
(r
1
(v), . . . , r
m
(v)) a set of right eigenvectors of
˜
A
ν
:
˜
A
ν
r
k
(v) = c
k
r
k
(v). Moreover the
following normalization is taken:
˜
l
k
(v) ˜ r
p
(v) = δ
k,p
.
According to [GP05] we have the following result on the solvability of (A.36)(A.37).
Theorem 4 In the case 1 ≤ p ≤ m−1, assume that c
p+1
(v) > 0, and
det
1≤k,l≤p
_
m
i=1
r
i
k
(v)
∂g
l
∂v
i
(v)
_
,= 0 . (A.40)
With the choice (A.39) the nonlinear system (A.36)(A.37) has one and only one solution
v, for v −v and g
l
(v) suﬃciently small.
Remark 24 In this result we exclude the case where the boundary is characteristic i.e. the
case where one of the c
k
is equal to 0. This case cannot be dealt at this level of generality.
On the other hand, wall boundary conditions belong to this category. They can be discussed
and handled directly on the physical system under consideration, see [GP05].
Remark 25 In practice, (A.36)(A.37) are written in a parametric way. We have a set of
m physical variables w (e.g. pressure, densities, velocities,. . . ) and we look for w satisfying:
g
l
(w) = 0 , l = 1, . . . , p , (A.41)
˜
l
k
(v) Φ =
˜
l
k
(v) (F(v
n
K
) ν
K
) , (A.42)
Φ = F(w) ν
K
, (A.43)
and then we take:
Φ(v
n
K
, K, ∂Ω) = Φ. (A.44)
The system (A.41)(A.42)(A.43) is then solved by Newton’s method.
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
Bibliography
[AG73] A.S. Alekseev and V.K. Gusyakov. Theory of diﬀraction and wave propaga
tion, volume 2, chapter Numerical modelling of tsunami and seismoacoustic
waves generation by submarine earthquakes, pages 194–197. MoscowErevan,
1973. in Russian. 3
[AGT99] F. Alouges, J.M. Ghiadaglia, and M. Tajchman. On the interaction between
upwinding and forcing for hyperbolic systems of conservation laws. Technical
report, Centre de Math´ematiques et Leurs Applications, 1999. 203
[AMS04] F. Archambeau, N. Mehitoua, and M. Sakiz. Code Saturne: A ﬁnite volume
code for turbulent ﬂows. International Journal On Finite Volumes, 1, 2004.
140
[BB73] J.P. Boris and D.L. Book. Flux corrected transport: Shasta, a ﬂuid transport
algorithm that works. J. Comp. Phys., 11:38–69, 1973. 143
[BBM72] T.B. Benjamin, J.L. Bona, and J.J. Mahony. Model equations for long waves
in nonlinear dispersive systems. Philos. Trans. Royal Soc. London Ser. A,
272:47–78, 1972. 3, 106
[BCL05] J.L. Bona, T. Colin, and D. Lannes. Long wave approximations for water
waves. Arch. Rational Mech. Anal., 178:373–410, 2005. 57
[BCS02] J.L. Bona, M. Chen, and J.C. Saut. Boussinesq equations and other systems
for smallamplitude long waves in nonlinear dispersive media. i: Derivation
and linear theory. Journal of Nonlinear Science, 12:283–318, 2002. xxiii, xxv,
83, 104, 105
[BDM07] J.L. Bona, V.A. Dougalis, and D.E. Mitsotakis. Numerical solution of KdV
KdV systems of Boussinesq equations: I. The numerical scheme and general
ized solitary waves. Mat. Comp. Simul., 74:214–228, 2007. xxvi
207
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
208 BIBLIOGRAPHY
[Ben74] T. B. Benjamin. Lectures in Appl. Math., volume 15, chapter Lectures on
nonlinear wave motion, pages 3–47. Amer. Math. Soc., Providence, RI, 1974.
99
[Ben06] M. Benoit. Contribution `a l’´etude des ´etats de mer et des vagues, depuis
l’oc´ean jusqu’aux ouvrages cotiers. 2006. M´emoire d’habilitation `a diriger des
recherches. 91
[BF90] T.J. Barth and P.O. Frederickson. Higher order solution of the Euler equa
tions on unstructured grids using quadratic reconstruction. AIAA, 900013,
1990. 137
[BH86] N. Bleistein and R.A. Handelsman. Asymptotic Expansions of Integrals.
Courier Dover Publications, 1986. 29
[BJ89] T.J. Barth and D.C. Jespersen. The design and application of upwind schemes
on unstructured meshes. AIAA, 0366, 1989. 133, 143, 144
[BMR72] A. BenMenahem and M. Rosenman. Amplitude patterns of tsunami waves
from submarine earthquakes. J. Geophys. Res., 77:3097–3128, 1972. 36
[BMSS69] A. BenMenahem, S. J. Singh, and F. Solomon. Static deformation of a
spherical earth model by internal dislocations. Bull. Seism. Soc. Am., 59:813–
853, 1969. 4
[BMSS70] A. BenMehanem, S. J. Singh, and F. Solomon. Deformation of an homoge
neous earth model ﬁnite by dislocations. Rev. Geophys. Space Phys., 8:591–
632, 1970. 4
[BO04] T.J. Barth and M. Ohlberger. Encyclopedia of Computational Mechanics,
Volume 1, Fundamentals, chapter Finite Volume Methods: Foundation and
Analysis. John Wiley and Sons, Ltd, 2004. 133, 144
[BOPB07] G. N. Bullock, C. Obhrai, D. H. Peregrine, and H. Bredmose. Violent breaking
wave impacts. part 1: Results from largescale regular wave tests on vertical
and sloping walls. Coastal Engineering, 54:602–617, 2007. 123, 124, 125
[Bou71a] J. V. Boussinesq. Th´eorie g´en´erale des mouvements qui sont propag´es dans
un canal rectangulaire horizontal. C. R. Acad. Sc. Paris, 73:256–260, 1871.
82
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
BIBLIOGRAPHY 209
[Bou71b] M.J. Boussinesq. Th´eorie de l’intumescence liquide appel´ee onde solitaire ou
de translation se propageant dans un canal rectangulaire. C.R. Acad. Sci.
Paris S´er. AB, 72:755–759, 1871. 3
[Bou72] J. Boussinesq. Th´eorie des ondes et des remous qui se propagent le long d’un
canal rectangulaire horizontal, en communiquant au liquide contenu dans ce
canal des vitesses sensiblement pareilles de la surface au fond. J. Math. Pures
Appl., 17:55–108, 1872. 82
[Bou95] J. Boussinesq. Lois de l’extinction de la houle en haute mer. C. R. Acad. Sci.
Paris, 121:15–20, 1895. 83, 84, 170, 186
[BPS81] J.L. Bona, W.G. Pritchard, and L.R. Scott. An evaluation of a model equation
for water waves. Phil. Trans. R. Soc. Lond. A, 302:457–510, 1981. 37, 84, 89,
106, 110, 119, 170
[BPS85] J. L. Bona, W.G. Pritchard, and L.R. Scott. Numerical schemes for a model
for nonlinear dispersive waves. Journal of Computational Physics, 60:167–
186, 1985. 106
[BQ06] F. Benkhaldoun and L. Quivy. A non homogeneous riemann solver for shallow
water and two phase ﬂows. Flow, Turbulence and Combustion, 76:391–402,
2006. 70, 71
[BUST06] J. Borrero, B. Uslu, C. Synolakis, and V.V. Titov. Modeling far ﬁeld tsunamis
for california ports and harbors. Coastal Engineering, pages 1566–1578, 2006.
78
[BvdDP73] R.D. Braddock, P. van den Driessche, and G.W. Peady. Tsunamis generation.
J. Fluid Mech., 59(4):817–828, 1973. 3
[Car71] G.F. Carrier. The dynamics of tsunamis. Mathematical Problems in the
Geophysical Sciences. Lectures in Applied Mathematics, 13:157–187, 1971. 3
[CG58] G. F. Carrier and H. P. Greenspan. Water waves of ﬁnite amplitude on a
sloping beach. Journal of Fluid Mechanics, 2:97–109, 1958. xxvii
[CG00] J. Cortes and J.M. Ghidaglia. Upwinding at low cost for complex models and
ﬂux schemes. In Trends in Numerical and Physical Modeling for Industrial
Multiphase Flows, 2000. 201
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
210 BIBLIOGRAPHY
[CG07] M. Chen and O. Goubet. Longtime asymptotic behaviour of dissipative
Boussinesq systems. Discrete and Continuous Dynamical Systems, 17:61–80,
2007. 171
[Cha05] H. Chanson. Le tsunami du 26 d´ecembre 2004: un ph´enom`ene hydraulique
d’ampleur internationale. premiers constats. La Houille Blanche, 2:25–32,
2005. xvii
[Che98] M. Chen. Exact travelingwave solutions to bidirectional wave equations.
International Journal of Theoretical Physics, 37:1547–1567, 1998. 110, 112
[Chi63] M. A. Chinnery. The stress changes that accompany strikeslip faulting. Bull.
Seism. Soc. Am., 53:921–932, 1963. 10
[CLS04] B. Cockburn, F. Li, and C.W. Shu. Locally divergencefree discontinu
ous Galerkin methods for the Maxwell equations. Journal of Computational
Physics, 194:588–610, 2004. 131
[Coi94] W.J. Coirier. An AdaptativelyReﬁned, Cartesian, Cellbased Scheme for the
Euler and NavierStokes Equations. PhD thesis, Michigan Univ., 1994. 144
[CP92] J. H. E. Cartwright and O. Piro. The dynamics of Runge–Kutta methods.
Int. J. Bifurcation and Chaos, 2:427–449, 1992. 151
[CVV99] Y. Coudiere, J.P. Vila, and P. Villedieu. Convergence rate of a ﬁnite vol
ume scheme for a two dimensionnal convection diﬀusion problem. M2AN, 33
(3):493–516, 1999. 144
[CWY03] G. F. Carrier, T. T. Wu, and H. Yeh. Tsunami runup and drawdown on a
plane beach. Journal of Fluid Mechanics, 475:79–99, 2003. xxviii
[DD06] F. Dias and D. Dutykh. Extreme ManMade and Natural Hazards in Dynam
ics of Structures, chapter Dynamics of tsunami waves, pages 35–60. Springer,
2006. xvii, 187
[DD07a] D. Dutykh and F. Dias. Dissipative Boussinesq equations. C. R. Mecanique,
335:559–583, 2007. 132, 171
[DD07b] D. Dutykh and F. Dias. Viscous potential freesurface ﬂows in a ﬂuid layer
of ﬁnite depth. C. R. Acad. Sci. Paris, Ser. I, 345:113–118, 2007. 171
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
BIBLIOGRAPHY 211
[DD07c] D. Dutykh and F. Dias. Water waves generated by a moving bottom. In
Anjan Kundu, editor, Tsunami and Nonlinear waves. Springer Verlag (Geo
Sc.), 2007. 38, 64, 65, 67, 91
[DDK06] D. Dutykh, F. Dias, and Y. Kervella. Linear theory of wave generation by a
moving bottom. C. R. Acad. Sci. Paris, Ser. I, 343:499–504, 2006. 36, 44,
52, 61, 65, 67, 75
[DDZ07] F. Dias, A.I. Dyachenko, and V.E. Zakharov. Theory of weakly damped free
surface ﬂows: a new formulation based on potential ﬂow solutions. Physics
Letters A, to appear, 2007. 85, 87, 94, 171, 176
[DGK06] R. A. Dalrymple, S. T. Grilli, and J. T. Kirby. Tsunamis and challenges for
accurate modeling. Oceanography, 19:142–151, 2006. xxiii
[DKZ03] A.I. Dyachenko, A.O. Korotkevich, and V.E. Zakharov. Weak turbulence of
gravity waves. JETP Lett., 77:546–550, 2003. 171, 193
[DKZ04] A.I. Dyachenko, A.O. Korotkevich, and V.E. Zakharov. Weak turbulent Kol
mogorov spectrum for surface gravity waves. Phys. Rev. Lett., 92:134501,
2004. 171, 193
[DM04] V.A. Dougalis and D.E. Mitsotakis. Advances in scattering theory and
biomedical engineering, chapter Solitary waves of the BonaSmith system,
pages 286–294. World Scientiﬁc, New Jersey, 2004. xxvi
[dSV71] A.J.C. de SaintVenant. Th´eorie du mouvement nonpermanent des eaux,
avec application aux crues des rivi`eres et `a l’introduction des mar´ees dans
leur lit. C. R. Acad. Sc. Paris, 73:147–154, 1871. 82
[Erd56] A. Erd´elyi. Asymptotic expansions. Dover publivations, 1956. 29
[ES94] Yu.V. Egorov and M.A. Shubin. Partial Diﬀerential Equations, volume 2
of Encyclopaedia of Mathematical Sciences, chapter Elements of the Modern
Theory. Equations with Constant Coeﬃcients. Springer, 1994. 29
[FB76] L. B. Freund and D. M. Barnett. A twodimensional analysis of surface
deformation due to dipslip faulting. Bull. Seism. Soc. Am., 66:667–675,
1976. 65, 72
[FD06] C. Fochesato and F Dias. A fast method for nonlinear threedimensional
freesurface waves. Proc. R. Soc. A, 462(2073):2715–2735, September 2006.
50
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
212 BIBLIOGRAPHY
[FDG05] C. Fochesato, F. Dias, and R. Grimshaw. Generalized solitary waves and
fronts in coupled Korteweg–de Vries systems. Physica D: Nonlinear Phe
nomena, 210:96–117, October 2005. xxvi
[Fil28] L.N.G. Filon. On a quadrature formula for trigonometric integrals. Proceed
ings of the Royal Society of Edinburgh, 49:38–47, 1928. 3, 29, 47
[GBM
+
05] F.I. Gonz´alez, E.N. Bernard, C. Meinig, M.C. Eble, H.O. Mofjeld, and
S. Stalin. The NTHMP tsunameter network. Natural Hazards, 35:25–39,
2005. xxix, 28
[GGCCD05] M. GomezGesteira, D. Cerqueiro, C. Crespo, and R. A. Dalrymple. Green
water overtopping analyzed with a SPH model. Ocean Engineering, 32:223–
238, 2005. xxix
[GGD01] S. Grilli, P. Guyenne, and F. Dias. A fully nonlinear model for three
dimensional overturning waves over an arbitrary bottom. Int. J. Numer.
Meth. Fluids, 35:829–867, 2001. 50
[GGD04] M. GomezGesteira and R. A. Dalrymple. Using SPH for wave impact on a
tall structure. Journal of Waterways, Port, Coastal, and Ocean Engineering,
130:63–69, 2004. xxix
[Ghi98] J.M. Ghidaglia. Flux schemes for solving nonlinear systems of conservation
laws. In J.J. Chattot and M. Hafez, editors, Proceedings of the meeting in
honor of P.L. Roe, Arcachon, July 1998. 199, 200
[Ghi08] J.M. Ghidaglia. Finite volumes methods and complex multiﬂuid ﬂows. In
preparation, 2008. 126
[GHM98] M. Guesmia, P.H. Heinrich, and C. Mariotti. Numerical simulation of the 1969
Portuguese tsunami by a ﬁnite element method. Natural Hazards, 17:31–46,
1998. 50
[GKC96] J.M. Ghidaglia, A. Kumbaro, and G. Le Coq. Une m´ethode volumesﬁnis `a
ﬂux caract´eristiques pour la r´esolution num´erique des syst`emes hyperboliques
de lois de conservation. C. R. Acad. Sci. I, 322:981–988, 1996. 50, 71, 198
[GKC01] J.M. Ghidaglia, A. Kumbaro, and G. Le Coq. On the numerical solution
to two ﬂuid models via cell centered ﬁnite volume method. Eur. J. Mech.
B/Fluids, 20:841–867, 2001. 71, 198
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
BIBLIOGRAPHY 213
[God59] S.K. Godunov. A ﬁnite diﬀerence method for the numerical computation of
discontinuous solutions of the equations of ﬂuid dynamics. Mat. Sb., 47:271–
290, 1959. 134, 137
[GP05] J.M. Ghidaglia and F. Pascal. The normal ﬂux method at the boundary for
multidimensional ﬁnite volume approximations in cfd. European Journal of
Mechanics B/Fluids, 24:1–17, 2005. 151, 153, 206
[GR00] I. S. Gradshteyn and M. Ryzhik. Tables of Integrals, Series, and Products.
Academic Press, Orlando, Florida, 6 edition, 2000. 22
[GST01] S. Gottlieb, C.W. Shu, and E. Tadmor. Strong stabilitypreserving high
order time discretization methods. SIAM Review, 43:89–112, 2001. 149
[GTS06] E.L. Geist, V.V. Titov, and C.E. Synolakis. Tsunami: wave of change. Sci
entiﬁc American, 294:56–63, 2006. 36
[Gus72] V.K. Gusyakov. Mathematical problems of geophysics, volume 3, chapter
Generation of tsunami waves and ocean Rayleigh waves by submarine earth
quakes, pages 250–272. Novosibirsk, VZ SO AN SSSR, 1972. in Russian.
3
[Gus76] V.K. Gusyakov. Illposed problems of mathematical physics and problems
of interpretation of geophysical observations, chapter Estimation of tsunami
energy, pages 46–64. Novosibirsk, VZ SO AN SSSR, 1976. in Russian. 3
[GV85] J.D. Goodman and R.J. Le Veque. On the accuracy of stable schemes for 2D
conservation laws. Math. Comp., 45(171):15–21, 1985. 139
[GVW02] S. Grilli, S. Vogelmann, and P. Watts. Development of a 3d numerical wave
tank for modeling tsunami generation by underwater landslides. Engng Anal.
Bound. Elem., 26:301–313, 2002. 51
[GZI
+
79] S.K. Godunov, A. Zabrodine, M. Ivanov, A. Kraiko, and G. Prokopov.
R´esolution num´erique des probl`emes multidimensionnels de la dynamique des
gaz. Editions Mir, Moscow, 1979. 127
[HA71] F. Harlow and A. Amsden. Fluid dynamics. LANL Monograph LA4700,
1971. 127
[Ham73] J. Hammack. A note on tsunamis: their generation and propagation in an
ocean of uniform depth. Journal of Fluid Mechanics, 60:769–799, 1973. 3,
36, 50, 65
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
214 BIBLIOGRAPHY
[Har83] A. Harten. High resolution schemes for hyperbolic conservation laws. J.
Comp. Phys., 49:357–393, 1983. 138
[Has69] N. A. Haskell. Elastic displacements in the nearﬁeld of a propagating fault.
Bull. Seism. Soc. Am., 59:865–908, 1969. 66
[HB05] J. Hunt and J. M. Burgers. Tsunami waves and coastal ﬂooding. Mathematics
TODAY, pages 144–146, October 2005. xviii
[HC89] D.G. Holmes and S.D. Connel. Solution of the 2d NavierStokes equations
on unstructured adaptive grids. In AIAA 9th Computational Fluid Dynamics
Conference, volume 891932CP, June 1989. 141, 146, 147
[HG74] J.R. Houston and A.W. Garcia. Type 16 ﬂood insurance study. Technical
Report Report No. H743, USACE WES Report No. H743, 1974. 36
[HH70] K.L. Heitner and G.W. Housner. Numerical model for tsunami runup. J.
Waterway, Port, Coastal and Ocean Engineering, 96:701–719, 1970. 85
[HPHO] F. Hecht, O. Pironneau, A. Le Hyaric, and K. Ohtsuka. FreeFem++. Labo
ratoire JL Lions, University of Paris VI, France. 70
[HS74] J.L. Hammack and H. Segur. The Korteweg–de Vries equation and water
waves. Part 2. Comparison with experiments. J Fluid Mech, 65:289–314,
1974. 32, 38, 40
[HS78] J.L. Hammack and H. Segur. The Korteweg–de Vries equation and water
waves. Part 3. Oscillatory waves. J Fluid Mech, 84:337–358, 1978. 32
[IS79] T. Iwasaki and R. Sato. Strain ﬁeld in a semiinﬁnite medium due to an
inclined rectangular fault. J. Phys. Earth, 27:285–314, 1979. 10
[Ish75] M. Ishii. ThermoFluid Dynamic Theory of TwoPhase Flow. Eyrolles, Paris,
1975. 125, 162
[JTPS96] L. Jiang, C.L. Ting, M. Perlin, and W.W. Schultz. Moderate and steep
Faraday waves: instabilities, modulation and temporal asymmetries. J. Fluid
Mech., 329:275–307, 1996. 87
[JW04] D.D. Joseph and J. Wang. The dissipation approximation and viscous po
tential ﬂow. J. Fluid Mech., 505:365–377, 2004. 171
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
BIBLIOGRAPHY 215
[Kaj63] K. Kajiura. The leading wave of tsunami. Bull. Earthquake Res. Inst., Tokyo
Univ., 41:535–571, 1963. 3, 36
[Kaj70] K. Kajiura. Tsunami source, energy and the directivity of wave radiation.
Bull. Earthquake Research Institute, 48:835–869, 1970. 65, 73
[KCKD00] A.B. Kennedy, Q. Chen, J.T. Kirby, and R.A. Dalrymple. Boussinesq mod
elling of wave transformation, breaking, and runup. J. Waterway, Port,
Coastal and Ocean Engineering, 126:39–47, 2000. 85
[KDD07] Y. Kervella, D. Dutykh, and F. Dias. Comparison between threedimensional
linear and nonlinear tsunami generation models. Theor. Comput. Fluid Dyn.,
21:245–269, 2007. 69, 70
[KdV95] D.J. Korteweg and G. de Vries. On the change of form of long waves advancing
in a rectangular canal, and on a new type of long stationary waves. Phil. Mag.,
39(5):422–443, 1895. 3
[Kel87] Lord (W. Thomson) Kelvin. Phil. Mag., 23(5):252–257, 1887. 29
[Kel63] J.B. Keller. Tsunamis: water waves produced by earthquakes. Int. Un.
Geodesy & Geophys. Monograph, 24:154–166, 1963. 3
[Kir03] J.T. Kirby. Advances in Coastal Modeling, V. C. Lakhan (ed), chapter Boussi
nesq models and applications to nearshore wave propagation, surfzone pro
cesses and waveinduced currents, pages 1–41. Elsevier, 2003. 57, 83
[KM75] T. Kakutani and K. Matsuuchi. Eﬀect of viscosity on long gravity waves. J.
Phys. Soc. Japan, 39:237–246, 1975. 171, 192
[KMC03] S.E. Kim, B. Makarov, and D. Caraeni. A multidimensional linear recon
struction scheme for arbitrary unstructured grids. Technical report, Fluent
Inc., 2003. 147
[KML05] E.A. Kulikov, P.P. Medvedev, and S.S. Lappo. Satellite recording of the
Indian Ocean tsunami on December 26, 2004. Doklady Earth Sciences A,
401:444–448, 2005. 49
[Kol72] N.E. Kolgan. Application of the minimumderivative principle in the con
struction of ﬁnitediﬀerence schemes for numerical analysis of discontinuous
solutions in gas dynamics. Uchenye Zapiski TsaGI [Sci. Notes Central Inst.
Aerodyn], 3(6):68–77, 1972. 137, 138
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
216 BIBLIOGRAPHY
[Kol75] N.E. Kolgan. Finitediﬀerence schemes for computation of three dimensional
solutions of gas dynamics and calculation of a ﬂow over a body under an
angle of attack. Uchenye Zapiski TsaGI [Sci. Notes Central Inst. Aerodyn],
6(2):1–6, 1975. 137, 138
[KS06] K. Kanoglu and C. Synolakis. Initial value problem solution of nonlinear
shallow waterwave equations. Phys. Rev. Lett., 97:148501, 2006. xxix, 50
[KWC
+
98] J. T. Kirby, G. Wei, Q. Chen, A. B. Kennedy, and R. A. Dalrymple. Funwave
1.0, fully nonlinear boussinesq wave model documentation and user’s manual.
Research Report No. CACR9806, 1998. xxiii
[Lam32] H. Lamb. Hydrodynamics. Cambridge University Press, 1932. 83, 170, 186
[Lax73] P.D. Lax. Hyperbolic Systems of Conservation Laws and the Mathematical
Theory of Shock Waves. SIAM, Philadelphia, Penn., 1973. 138
[LeV98] R.J. LeVeque. Balancing source terms and ﬂux gradients in highresolution
Godunov methods: the quasisteady wavepropagation algorithm. J. Com
putational Phys., 146:346–365, 1998. 50
[LH74] M. S. LonguetHiggins. On the mass, momentum, energy and circulation of
a solitary wave. Proc. R. Soc. Lond. A, 337:1–13, 1974. xxvii
[LH92] M.S. LonguetHiggins. Theory of weakly damped stokes waves: a new for
mulation and its physical interpretation. J. Fluid Mech., 235:319–324, 1992.
85, 169
[LKA
+
05] T. Lay, H. Kanamori, C. J. Ammon, M. Nettles, S. N. Ward, R. C. Aster, S. L.
Beck, S. L. Bilek, M. R. Brudzinski, R. Butler, H. R. DeShon, G. Ekstrom,
K. Satake, and S. Sipkin. The great SumatraAndaman earthquake of 26
December 2004. Science, 308:1127–1133, 2005. xxi, xxvi, xxvii, 3, 40, 43
[LL83] P.L.F. Liu and J.A. Liggett. Applications of boundary element methods to
problems of water waves, chapter Chapter 3, pages 37–67. 1983. 37
[LO04] P.L.F. Liu and A. Orﬁla. Viscous eﬀects on transient longwave propagation.
J. Fluid Mech., 520:83–92, 2004. 171, 193
[Lov44] A. E. H. Love. A treatise on the mathematical theory of elasticity. Dover
Publications, New York, 1944. 6, 8
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
BIBLIOGRAPHY 217
[LPC07] P.L.F. Liu, Y.S. Park, and E.A. Cowen. Boundary layer ﬂow and bed shear
stress under a solitary wave. J. Fluid Mech., 574:449–463, 2007. 178
[LSVO06] P.L.F. Liu, G. Simarro, J. Vandever, and A. Orﬁla. Experimental and nu
merical investigation of viscous eﬀects on solitary wave propagation in a wave
tank. Coastal Engineering, 53:181–190, 2006. 192
[Mad03] R. Madariaga. Radiation from a ﬁnite reverse fault in half space. Pure Appl.
Geophys., 160:555–577, 2003. 67, 72
[Mar64] T. Maruyama. Statical elastic dislocations in an inﬁnite and semiinﬁnite
medium. Bull. Earthquake Res. Inst., Tokyo Univ., 42:289–368, 1964. 6
[Mas03] T. Masterlark. Finite element model predictions of static deformation
from dislocation sources in a subduction zone: Sensivities to homogeneous,
isotropic, poissonsolid, and halfspace assumptions. J. Geophys. Res.,
108(B11):2540, 2003. 5, 65
[MBS03] P. A. Madsen, H. B. Bingham, and H. A. Schaﬀer. Boussinesqtype formu
lations for fully nonlinear and extremely dispersive water waves: derivation
and analysis. Proc. R. Soc. Lond. A, 459:1075–1104, 2003. xxiv
[MBS05] A. Megna, S. Barba, and S. Santini. Normalfault stress and displacement
through ﬁniteelement analysis. Annals Geophys., 48:1009–1016, 2005. 65, 73
[MC50] R. D. Mindlin and D. H. Cheng. Nuclei of strain in the semiinﬁnite solid. J.
Appl. Phys., 21:926–930, 1950. 7
[Mei94] C.C. Mei. The applied dynamics of ocean surface waves. World Scientiﬁc,
1994. 171, 172, 177
[MG96] S. Musaferija and D. Gosman. Finitevolume CFD procedure and adap
tive error control strategy for grids of arbitrary topology. J. Comp. Phys.,
138:766–787, 1996. 140
[Min36] R. D. Mindlin. Force at a point in the interior of a semiinﬁnite medium.
Physics, 7:195–202, 1936. 7
[Mit07] D.E. Mitsotakis. Boussinesq systems in two space dimensions over a variable
bottom for the generation and propagation of tsunami waves. Mathematics
and Computers in Simulation, submitted, 2007. 4
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
218 BIBLIOGRAPHY
[MMS91] P. A. Madsen, R. Murray, and O. R. Sorensen. A new form of the Boussinesq
equations with improved linear dispersion characteristics. Coastal Engineer
ing, 15:371–388, 1991. 104
[Mon94] J. J. Monaghan. Simulating free surface ﬂows with SPH. Physica D, 110:399–
406, 1994. xxix
[MS71] L. Mansinha and D. E. Smylie. The displacement ﬁelds of inclined faults.
Bull. Seism. Soc. Am., 61:1433–1440, 1971. 65
[MS98] P. A. Madsen and H. A. Schaﬀer. Higherorder Boussinesqtype equations for
surface gravity waves: Derivation and analysis. Phil. Trans. R. Soc. Lond.
A, 356:3123–3184, 1998. 83, 91, 100
[Mur89] R. J. Murray. Short wave modelling using new equations of Boussinesq type.
In Proc., 9th Australian Conf. on Coast. and Oc. Engrg., 1989. 104
[Mur92] J.D. Murray. Asymptotic Analysis. Springer, 1992. 29
[New77] A. C. Newell. Finite amplitude instabilities of partial diﬀerence equations.
SIAM Journal of Applied Mathematics, 33:133–160, 1977. 112
[NSS
+
05] S. Neetu, I. Suresh, R. Shankar, D. Shankar, S.S.C. Shenoi, S.R. Shetye,
D. Sundar, and B. Nagarajan. Comment on “The Great SumatraAndaman
Earthquake of 26 December 2004”. Science, 310:1431a–1431b, 2005. 3
[Nwo93] O. Nwogu. Alternative form of Boussinesq equations for nearshore wave
propagation. J. Waterway, Port, Coastal and Ocean Engineering, 119:618–
638, 1993. 57, 83, 91, 102, 104
[OK07] Y. Okumura and Y. Kawata. Eﬀects of rise time and rupture velocity on
tsunami. Proceedings of the Seventeenth International Oﬀshore and Polar
Engineering Conference, III, 2007. 66
[Oka85] Y. Okada. Surface deformation due to shear and tensile faults in a halfspace.
Bull. Seism. Soc. Am., 75:1135–1154, 1985. xxiii, 4, 7, 9, 10, 36, 38, 65, 72
[Oka92] Y. Okada. Internal deformation due to shear and tensile faults in a halfspace.
Bull. Seism. Soc. Am., 82:1018–1040, 1992. 7, 10
[Osh84] S. Osher. Riemann solvers, the entropy condition, and diﬀerence approxima
tions. SIAM J. Numer. Anal., 21(2):217–235, 1984. 137
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
BIBLIOGRAPHY 219
[OTM01] T. Ohmachi, H. Tsukiyama, and H. Matsumoto. Simulation of tsunami in
duced by dynamic displacement of seabed due to seismic faulting. Bull. Seism.
Soc. Am., 91:1898–1909, 2001. 65, 66
[Per66] D.H. Peregrine. Calculations of the development of an undual bore. J. Fluid
Mech., 25:321–330, 1966. 3
[Per67] D. H. Peregrine. Long waves on a beach. J. Fluid Mech., 27:815–827, 1967.
57, 83, 106, 119
[Per72] D. H. Peregrine. Waves on Beaches and Resulting Sediment Transport, chap
ter Equations for water waves and the approximation behind them, pages
95–121. Academic Press, New York, 1972. 99
[Per03] D. H. Peregrine. Waterwave impact on walls. Annu. Rev. Fluid Mech.,
35:23–43, 2003. 123
[Pet91] T. Peterson. A note on the convergence of the discontinuous Galerkin method
for a scalar hyperbolic equation. SIAM J. Numer. Anal., 28(1):133–140, 1991.
137
[PL71] G.I. Petrashen’ and K.P. Latyshev. Asymptotic Methods and Stochastic Mod
els in Problems of Wave Propagation. American Mathematical Society, 1971.
29
[Pod68] G.S. Podyapolsky. The generation of linear gravitational waves in the ocean
by seismic sources in the crust. Izvestiya, Earth Physics, Akademia Nauk
SSSR, 1:4–12, 1968. 3
[Pre65] F. Press. Displacements, strains and tilts at teleseismic distances. J. Geo
phys. Res., 70:2395–2412, 1965. 7
[Qui97] J. Quiblier. Propagation des ondes en g´eophysique et en g´eotechnique.
Mod´elisation par m´ethodes de Fourier. Paris: Editions Technip, 1997. 67
[RFF91] K.D. Ruvinsky, F.I. Feldstein, and G.I. Freidman. Numerical simulations of
the quasistationary stage of ripple excitation by steepcapillary waves. J.
Fluid Mech., 230:339–353, 1991. 85, 87
[Roe81] P. L. Roe. Approximate riemann solvers, parameter vectors and diﬀerence
schemes. J. Comput. Phys., 43:357–372, 1981. 200
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
220 BIBLIOGRAPHY
[Sab86] P. Sabatier. Encyclopedia of ﬂuid mechanics, chapter Formation of waves by
ground motion, pages 723–759. Gulf Publishing Company, 1986. 3
[SB06] C.E. Synolakis and E.N. Bernard. Tsunami science before and beyond Boxing
Day 2004. Phil. Trans. R. Soc. A, 364:2231–2265, 2006. 36, 47, 69
[Shu88] C.W. Shu. Totalvariationdiminishing time discretizations. SIAM J. Sci.
Statist. Comput., 9:1073–1084, 1988. 149
[SKP96] C. Skandrani, C. Kharif, and J. Poitevin. Nonlinear evolution of water surface
waves: the frequency downshift phenomenon. Contemp. Math., 200:157–171,
1996. 85
[SM71] D. E. Smylie and L. Mansinha. The elasticity theory of dislocations in real
earth models and changes in the rotation of the earth. Geophys. J.R. Astr.
Soc., 23:329–354, 1971. 4
[SM74] R. Sato and M. Matsu’ura. Strains and tilts on the surface of a semiinﬁnite
medium. J. Phys. Earth, 22:213–221, 1974. 10
[SO88] C.W. Shu and S. Osher. Eﬃcient implementation of essentially non
oscillatory shockcapturing schemes. J. Comput. Phys., 77:439–471, 1988.
149
[Sod78] G. A. Sod. A survey of several ﬁnite diﬀerence methods for systems of non
linear hyperbolic conservation laws. J. Comput. Phys., 43:1–31, 1978. 153
[SR02] R. J. Spiteri and S. J. Ruuth. A new class of optimal highorder strong
stabilitypreserving time discretization methods. SIAM Journal on Numerical
Analysis, 40:469–491, 2002. 149
[Ste58] J. A. Steketee. On Volterra’s dislocation in a semiinﬁnite elastic medium.
Can. J. Phys., 36:192–205, 1958. 4, 5
[Sug91] N. Sugimoto. Burgers equation with a fractional derivative; hereditary eﬀects
on nonlinear acoustic waves. J. Fluid Mech., 225:631–653, 1991. 193
[SVBM02] B. Spivak, J.M. VandenBroeck, and T. Miloh. Freesurface wave damping
due to viscosity and surfactants. European Journal of Mechanics BFluids,
21:207–224, 2002. 85
[Syn87] C. Synolakis. The runup of solitary waves. J. Fluid Mech., 185:523–545, 1987.
xxviii
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
BIBLIOGRAPHY 221
[TF98] M. Tajchman and P. Freydier. Sch´ema VFFC : application `a l’´etude d’un cas
test d’´ebullition en tuyau droit repr´esentant le fonctionnement en bouillotte
d’un coeur rep. Technical report, Note EDF HT33/98/033/A, 1998. 203
[THT02] M.I. Todorovska, A. Hayir, and M.D. Trifunac. A note on tsunami ampli
tudes above submarine slides and slumps. Soil Dynamics and Earthquake
Engineering, 22:129–141, 2002. 3, 44
[Tre00] Lloyd N. Trefethen. Spectral methods in MatLab. Society for Industrial and
Applied Mathematics, Philadelphia, PA, USA, 2000. 106
[TS94] S. Tadepalli and C. E. Synolakis. The runup of nwaves on sloping beaches.
Proc. R. Soc. Lond. A, 445:99–112, 1994. xxviii
[TS98] V. V. Titov and C. E. Synolakis. Numerical modeling of tidal wave runup. J.
Waterway, Port, Coastal, and Ocean Engineering, 124:157–171, 1998. xxiii,
37, 47, 49, 69
[TT01] M. I. Todorovska and M. D. Trifunac. Generation of tsunamis by a slowly
spreading uplift of the seaﬂoor. Soil Dynamics and Earthquake Engineering,
21:151–167, 2001. 2, 3, 37, 44, 64, 65
[TT05] S. Tinti and R. Tonini. Analytical evolution of tsunamis induced by near
shore earthquakes on a constantslope ocean. Journal of Fluid Mechanics,
535:33–64, 2005. xxviii
[Tuc74] E.O. Tuck. The eﬀect of a surface layer of viscous ﬂuid on the wave resistance
of a thin ship. J. Ship Research, 18:265–271, 1974. 36, 62, 85
[Tuc79] E.O. Tuck. Models for predicting tsunami propagation. In L.S. Hwang and
Tetra Tech Inc. Y.K. Lee, editors, NSF Workshop on Tsunamis, California,
pages 43–109, 1979. 37, 62
[Urs53] F. Ursell. The longwave paradox in the theory of gravity waves. Proc. Camb.
Phil. Soc., 49:685–694, 1953. 4, 52, 91
[vdD72] R.D. van den Driessche, P. & Braddock. On the elliptic generating region of
a tsunami. J. Mar. Res., 30:217–226, 1972. 3
[vL79] B. van Leer. Towards the ultimate conservative diﬀerence scheme v: a second
order sequel to Godunov’ method. J. Comput. Phys., 32:101–136, 1979. 137
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
222 BIBLIOGRAPHY
[vL06] B. van Leer. Upwind and highresolution methods for compressible ﬂow:
From donor cell to residualdistribution schemes. Communications in Com
putational Physics, 1:192–206, 2006. 138
[Vol07] V. Volterra. Sur l’´equilibre des corps ´elastiques multiplement connexes. An
nales Scientiﬁques de l’Ecole Normale Sup´erieure, 24(3):401–517, 1907. 5,
6
[VS93] M. Villeneuve and S.B. Savage. Nonlinear, dispersive, shallowwater waves
developed by a moving bed. J. Hydraulic Res., 31:249–266, 1993. 37
[War01] S.N. Ward. Landslide tsunami. J. Geophysical Res., 106:11201–11215, 2001.
47
[Wes35] H. M. Westergaard. General solution of the problem of elastostatics of an
ndimensional homogeneous isotropic solid in an ndimensional space. Bull.
Amer. Math. Soc., 41:695, 1935. 7
[Whi99] G.B. Whitham. Linear and nonlinear waves. John Wiley & Sons Inc., New
York, 1999. 99
[Wit84] J. M. Witting. A uniﬁed model for the evolution of nonlinear water waves.
J. Comput. Phys., 56:203–236, 1984. 105
[WKGS95] G. Wei, J. T. Kirby, S. T. Grilli, and R. Subramanya. A fully nonlinear
Boussinesq model for surface waves. Part 1. Highly nonlinear unsteady waves.
J. Fluid Mech., 294:71–92, 1995. 57, 83, 91
[Wu81] T.Y. Wu. Long waves in ocean and coastal waters. Journal of Engineering
Mechanics, 107:501–522, 1981. 170
[Zel91] J.A. Zelt. The runup of nonbreaking and breaking solitary waves. Coastal
Engineering, 15:205–246, 1991. 85
[ZG71] N.J. Zabusky and C.C.J. Galvin. Shallow water waves, the Kortewegde Vries
equation and solitons. J. Fluid Mech., 47:811–824, 1971. 170
[ZKPD05] V.E. Zakharov, A.O. Korotkevich, A.N. Pushkarev, and A.I. Dyachenko.
Mesoscopic wave turbulence. JETP Lett., 82:487–491, 2005. 171, 193
[ZV97] W. Zhang and J. Vinals. Pattern formation in weakly damped parametric
surface waves. J. Fluid Mech., 336:301–330, 1997. 87
t
e
l

0
0
1
9
4
7
6
3
,
v
e
r
s
i
o
n
2

2
1
J
a
n
2
0
0
8
tel00194763, version 2  21 Jan 2008
tel00194763, version 2  21 Jan 2008
To my parents, my sister and A.M.
tel00194763, version 2  21 Jan 2008
Remerciements
Je tiens ` remercier, en tout premier lieu, Fr´d´ric Dias qui a encadr´ ce travail. Il a e e e m’a beaucoup soutenu et encourag´ au cours de ces magniﬁques ann´es de recherche. Je e e lui en suis tr`s profond´ment reconnaissant. e e
tel00194763, version 2  21 Jan 2008
Je remercie tout particuli`rement JeanMichel Ghidaglia qui m’a beaucoup guid´ dans e e mes recherches. Je tiens ` souligner que par sa grande disponibilit´, j’ai pu b´n´ﬁcier de a e e e ses tr`s grandes comp´tences scientiﬁques tout au long de mes investigations. Enﬁn, c’est e e un honneur pour moi qu’il ait accept´ de pr´sider le jury de cette th`se. e e e J’exprime ma gratitude aux Professeurs JeanClaude Saut et Didier Bresch, qui ont accept´ d’ˆtre les rapporteurs, ainsi qu’aux Professeurs Costas Synolakis et Vassilios Doue e galis pour leur int´rˆt qu’ils ont port´ ` mon travail en ayant accept´ de faire partie du jury. e e ea e Mes trois ann´es de th`se se sont d´roul´es dans un contexte scientiﬁque de grande e e e e qualit´ au Centre de Math´matiques et de Leurs Applications de l’ENS Cachan. Je ree e mercie son directeur Laurent Desvillettes de m’avoir accueilli et d’avoir cr´e d’excellentes e conditions de travail. Parmi tout le personnel de ce laboratoire je voudrais plus particuli`rement remercier Micheline Brunetti pour son aimabilit´ et eﬃcacit´ dans le traitement e e e des questions administratives. Je remercie Fr´d´ric Pascal pour son tutorat p´dagogique. Ses conseils m’ont beaue e e coup servis lors de trois ans d’enseignement du calcul scientiﬁque aux agr´gatifs de l’ENS e Cachan. J’en proﬁte aussi pour remercier tous mes ´l`ves de leur assiduit´ et intelligence. ee e J’exprime mes plus vifs remerciements ` Daniel Bouche de m’avoir accueilli au sein de a LRC CEA/CMLA. Ses remarques pertinentes et conseils avis´s, sa gentillesse m’ont beaue coup aid´ et motiv´ durant le DEA MN2MC et plus particuli`rement pendant la r´alisation e e e e de cette th`se. e
e a e e Mes pens´es vont ´galement ` tous mes concurrents qui m’ont constamment pouss´ e e a e dans mes recherches et m’ont donn´ les forces pour continuer ce travail. Serge. Je vous en remercie tr`s sinc`rement. Sans ces personnes cette th`se n’aurait pas eu la mˆme e e saveur.ii REMERCIEMENTS Parmi tous mes coll`gues de LRC. Je remercie ma famille et leur d´die cette th`se pour leurs encouragements permanents e e et toutes autres raisons qui s’imposent. version 2 . Thomas.21 Jan 2008 . Nanthilde. Je conclurai en remerciant de tout coeur Anna Maria pour la patience et le soutien sans limite dont elle fait preuve quotidiennement. Je tiens ` remercier particuli`rement Alexandra Pradeau de nos discussions pasa e sionnantes ainsi qu’enrichissantes. e e Parmi les personnes qui ont contribu´ ` ma formation. e Je dis grand merci ` Yann. Lavinia. tel00194763. ils m’ont oﬀert de e e e e nombreux conseils amicaux et professionnels. Joachim. Alexandra a et aux autres doctorants du LMT pour leur amiti´ et tous les instants de d´tente partag´s e e e ensemble. Christophe Fochesato. je suis tout particuli`rement e a e reconnaissant envers Vladimir Lamzyuk qui m’a fait d´couvrir le monde de la mod´lisation e e math´matique et des math´matiques appliqu´es en g´n´ral. JeanMichel Rovarch et Denis e Gueyﬃer ont occup´ une place tr`s particuli`re: en plus de leur amiti´. Il m’a soutenu constamment e e e e e durant mes ´tudes ` l’Universit´ Nationale de Dniprop´trovsk.
e Finalement. ´coulements diphasiques. nous discue e e tons de l’importance. Nous parlons des ´coulements e viscopotentiels.R´sum´ e e Cette th`se est consacr´e ` la mod´lisation des tsunamis. Dans le troisi`me chapitre. g´n´ration des tsunamis. nous nous int´ressons ` la g´n´ration de ces vagues extrˆmes. de la nature et de l’inclusion des eﬀets dissipatifs dans les mod`les e d’ondes longues. La vie de ces vagues peut e e a e ˆtre conditionnellement divis´e en trois parties: g´n´ration. Le deuxi`me chapitre est d´di´ essentiellement aux ´quations de Boussinesq qui sont e e e e souvent utilis´es pour mod´liser la propagation d’un tsunami. nous discutons de la discr´tisation num´rique de ces ´quations avec un sch´ma de e e e e type volumes ﬁnis sur des maillages non structur´s. La conclusion principale ` laquelle nous e a sommes arriv´s est que le couplage entre la sismologie et l’hydrodynamique est actuellement e assez mal compris. Donc. nous examinons les diﬀ´rentes approches existantes pour e e la mod´lisation. Nous proposons une nouvelle approche simpliﬁ´e pour les ´coulements e e faiblement visqueux. Ene e o e suite. e a e e e Dans cette partie du m´moire. Le but de ce chapitre est de proposer un mod`le simple et e e op´rationnel pour la mod´lisation de l’impact d’une vague sur les structures cˆti`res. propagation et inondation. Plus pr´cisement. Nous conservons la simplicit´ des ´coulements potentiels tout en e e ajoutant la dissipation. e a e tel00194763. puis nous en proposons d’autres. version 2 . Cette correction s’av`re ˆtre non locale en ua e e e temps. ´coulements viscopotentiels. nous changeons de sujet et nous nous tournons vers les e ´coulements diphasiques. la couche limite au fond apporte un certain eﬀet de m´moire ` l’´coulement.21 Jan 2008 Mots cl´s: Ondes de surface. e e e e Dans un premier temps. volumes ﬁnis e e . Certains auteurs les utilisent e e mˆme pour mod´liser le processus d’inondation (le runup). ´quations de Boussie e e e nesq. le m´moire se termine par un sujet qui devrait ˆtre pr´sent dans tous les e e e manuels classiques d’hydrodynamique mais qui ne l’est pas. Dans le cas de la profondeur ﬁnie nous incluons un terme correcteur dˆ ` la pr´sence de la couche limite au fond.
tel00194763.21 Jan 2008 . version 2 .
propagation and inundation (or runup). twophase ﬂows. Finally.21 Jan 2008 Keywords: Water waves. tel00194763. the bottom boundary layer introduces a memory eﬀect to the governing equations. This term is nonlocal in time. Another important application includes wave sloshing in liquiﬁed natural gas carriers. The life of tsunami waves can be conditionally divided into three parts: generation. The second chapter essentially deals with Boussinesq equations which are often used to model tsunami propagation and sometimes even runup. In the case of ﬁnite depth we derive a correction term due to the presence of the bottom boundary layer. we discuss the importance. Hence. We examine various existing approaches to its modelling. We propose a novel and suﬃciently simple approach for weakly viscous ﬂow modelling. In the third chapter we slightly change the subject and turn to twophase ﬂows. We succeeded in keeping the simplicity of the classical potential ﬂow formulation with the addition of viscous eﬀects. We mean viscopotential ﬂows. this thesis deals with a topic which should be present in any textbook on hydrodynamics but it is not.Abstract This thesis is devoted to tsunami wave modelling. Boussinesq equations. The main conclusion is that the seismology/hydrodynamics coupling is poorly understood at the present time. tsunami generation. viscopotential ﬂows. nature and inclusion of dissipative eﬀects in long wave models. Then we propose a few alternatives. Then. More precisely. ﬁnite volumes . In the ﬁrst part of the manuscript we consider the generation process of such extreme waves. version 2 . The main purpose of this chapter is to propose an operational and simple set of equations in order to model wave impacts on coastal structures. we discuss the numerical discretization of governing equations in the ﬁnite volume framework on unstructured meshes.
version 2 .tel00194763.21 Jan 2008 .
. . .1. . .1 Pressure on the bottom . . . . . . . . . . . . . . . 1. . . 1.1. . . . . . . . . 1. . . . . .4 Solution in ﬂuid domain . . . . . . . . . . .1. . . . .6 Velocity ﬁeld . . . . . . . . . . . . . .1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Waves generated by a moving bottom . . . . . . . . . . . . . . . .2 Volterra’s theory of dislocations . . . . . . . . . . . . . . . . . . . . . . . . . . . Tsunami runup . . . . . . . . . . . . . Boussinesq equations . . . . . . . . .7 Asymptotic analysis of integral solutions vii . . . . . . . 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . version 2 . . . . . .1. . . . . . . . . . . Shallowwater equations . . . . . . . . Korteweg–de Vries equation Energy of a tsunami .6. . . . . . . . . . . . . . . . . . . . .1 Source model . . . . . . . 1 Tsunami generation 1. . Classical formulation .1.5 Freesurface elevation . . . . . . . . .Contents Remerciements Contents i vii xiii xv xvii xviii xx xxi xxi xxiii xxvi xxvi xxvii 1 2 4 5 6 10 12 16 21 25 28 29 tel00194763. . . . . . .1. 1. . . . . . . . . . . . . . . . . . . 1. . . . . . . . . . .3 Dislocations in elastic halfspace . . . . . . . .3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1. . . . . . . . .1. . . . . Dimensionless equations . . . . . . . .21 Jan 2008 List of Figures List of Tables Introduction Propagation of tsunamis . . . . . 1. .1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Finite rectangular source . . . 1. . . .2 Curvilinear fault .1. . . . . . . . . . . . . .3. . . . . . . .
. . . . . . . . . . . . .1 Discretization of the viscoelastodynamic equations 1. . . . . . . . . . .3 Numerical methods . . . . . . . . . . . . .2. . .3. 1.3. . . . . . . . . . . . . . . 2. . . . . . .2. . . . . . . . . . .2 Derivation of the Boussinesq equations . . . 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Derivation of the equations . .2 tel00194763. . . . . . . . . . . . . . . . . .2. . . . .1 Physical problem description . . . . . . . .2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tsunami generation by dynamic displacement of sea bed . . . . . . . . . .2. . .8 Numerical results . . . . . . . . . . . . . . . . . .4 Passive generation . . Comparison of tsunami generation models . . . . . . . . . . . . . . . . . . . . . . . . .1 Introduction . .2.3 Analysis of the linear dispersion relations . . . . . . . . .3. . . 2.3.1 Linearized potential ﬂow equations . . . . . . . 1.1 Introduction . .4 Validation of the numerical method . . . .2. . . . . 2.3. . . . . . . . . . . . . . . . . . . version 2 . 31 36 37 43 44 44 47 47 48 50 50 51 64 64 64 67 67 69 70 70 71 72 73 78 81 82 85 89 93 93 95 99 99 100 104 105 107 1. . . . . . . .3. . . . . . . . . . . . . . . . . . . . . . .3 Discussion . . . . .7 Numerical method . . . . . . . . . . . . . . . . . . . . .3 2 Dissipative Boussinesq equations 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2. . . . 1. . . . . . . . . . 2. . . . . . .5 Results of the simulation . . . . . .2 Dissipative Boussinesq equations . . . . .viii CONTENTS 1. . . . . . . .2. . . . . . . .10 Conclusions . . . . . . . . .2 Finitevolume scheme . . . . . . . . . . . . . .3. 1. . 1. . . . . . . . . . . . . . . . . . . . 1.3. . . . . . . . . . . . . . . . . . . . . 2. . . .3. . . . . . . . . .6 Conclusions . . . . . . 1. . . . . 1. . . . . . .2 Fluid layer model . . . . . . . . . . . . . . . . . . . 1. . . 1. . 2. . . . . . . . .21 Jan 2008 1.1 Numerical method for the linear problem .5 Nonlinear shallow water equations . . . . .7 Bottom friction . .2. .9 Comparisons and discussion . . . . .3. .1 Asymptotic expansion . . . . . . . . . . . . . . . . 1. . . . . . . . . 1. . . . . .3. . .2 Mathematical models . .3. . . . . . . . . 1. .1 Dynamic fault model . . .5 Improvement of the linear dispersion relations . . . . 1. . .6 Mathematical model . . . . . . .1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3. . . . . . .2 Linear theory . . . . . . .6 Regularization of Boussinesq equations . . . . 1. . . . . . . 2. . .3 Active generation . . . . . . . . . . 1. . . . . . . . . . . . 1. . . . 1. . . . . . . . . . . . . . .4 Alternative version of the Boussinesq equations 2.3. . . . . . . . 1. .4. . . . . . . . .2. . . . .8 Numerical method for the full equations . . . . . . . . . . . . 2. . . . . . . . . . . .2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3. . . . . . . . . . . . . . .2. . . . . . . . . . . . . . . . .4. . .2. . .2. . . . . . . . . . . . . . 2. . .
. . . . .1 Sound speed in the mixture . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Slope limiter . . . . . . . .1 Introduction . . . . . . . 3. .1 Historical remark . . . . . . . . . .5. . . . . . . . .4. .4. . .4. . . . . . . . . . . . . 3. . . . . . . .2.2 Equation of state . . . . version 2 . . . . . . . . . . . . . .2 Mathematical model .2. . .3. . . . .6 Time stepping methods . .9 Numerical results . . . . .3 Water drop test case . . . . . .4 Finite volume scheme on unstructured meshes .8 Spectral Fourier method . . . . . . .6 Conclusions . . . . . .CONTENTS 2. . . . . . . method . .2. . . . .3 Derivation . . . . . . . . . . 4. . . . . . . . . . . . . . . . . . . . . . . . .2 Falling water column . . . 4. . . . . . . . . . . . . . . . .4.2 Leastsquares gradient reconstruction 3. 3. . . . . . 3. . . . . . . . . . . 3. . 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4. . .1 Convergence test . . . . . . . . . . . . . . .4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2. . . . . .4. . . 2. . . .5. . . 3.1 Discussion . . . . 3. . . . . . . . . . . . . . . . . . . . 3. . . . . . . . . . . .8. . . . . . . . . .5. . . . . . . . . . . . . . . . . . . . . . . . . .5 Solution interpolation to mesh nodes . . . . . . . . . .4. . .1 Sign matrix computation . . . . . .1 Introduction . . . . . . . . . . . . . .4. . . . . . 3.2 Comparison between the dissipative models 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2. . . . . . . . .1 Construction of the initial condition . . . . . 4. . . 2. . . . . . . . . . . . 3. . . . . . . . . . . . . . . . . . . . . .1 Dissipative KdV equation 4. . . . . . . . .4 Dispersion relation . . . . . . . . . . .3. . . . . . . . . . . 4 Viscous potential ﬂows 4. . . . . . 4. . . . . . . . . . . . . .4. 3. . . . . . . . . . . . . .3. . . . . . . . . . . . . . . .9. . . . . . . . . . .4 Diﬀusive ﬂuxes computation . . . . . .2 Anatomy of dissipation . . . . . .1 Validation of the numerical method . . . . . . . . . .10 Conclusions . . . . . . . . . . . . . . . . . . . . 3. . . . . . . . . . . . . . .4. . . . . . .1 GreenGauss gradient reconstruction 3. . . . . . . . . . . . . . . . . . 3. . . 3. . . . . . ix 108 110 113 113 115 119 121 122 125 126 127 129 131 136 137 137 138 140 141 143 144 146 148 151 153 154 157 157 162 169 169 172 174 179 180 182 3 Two phase ﬂows 3. tel00194763. . . . .3 Formal limit in barotropic case . . . . . 3. . . . . . . . . . .7 Boundary conditions treatment . . . . . . 3. . . . . . . . . . . .2 Second order scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 TVD and MUSCL schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4. . . . .9. . . . . .3. . . .21 Jan 2008 . . . . . . . . . . . . . .5 Numerical results . . 3. . . 3. . .
. . . . . . . . . . . . . 182 187 187 189 192 195 197 198 198 201 202 204 206 4. . . . . . . . .6. . A. . . . . . . . . . . . . . . . . .6 Attenuation of linear progressive waves . . . . . . . . . . Bibliography tel00194763. . . .1. . . . . . . . . .7 Direction for future research A VFFC scheme A. . . .1 Discretization in the ﬁnite volume framework A. . . . . . . . . . . . . . Numerical results . . . . . . . . . . . . . . . . .3 On the discretization of source terms . . . . . . . . . . . .x 4. . . . version 2 .6. .1. . . . . . . . . . . . . . . . . . . . . . .5 4. .2 Discussion . Conclusion . . . . . . . .1. . . . CONTENTS . . . . . . . . . . . . . . . . . . . . . . . 4. . . . . . . .1 The one dimensional case . A. . . . . . . . . . . . . . . . . . . . . . . . . . .2 On the discretization of boundary conditions . .21 Jan 2008 . . . . . . . . . .1 Approximate solitary wave solution 4. . .2 Extension to the multidimensional case A. . . . . . . . . . . . .
. . . . .II . . . . . . . . .16 1.2 1. . . . . . . . .6. . . . . . . .28 Coordinate system adopted in this study.23 1.4 1. . . . . . . . . . . . . . . Deﬁnition of the ﬂuid domain and coordinate system . . . .22 1. . . . . . . . . . .9 1. h = 1km .01. . . . . . Bottom pressure at t = 0. . . Freesurface elevation at t = 0. . . Illustration for strike angle deﬁnition. . . . . . . . . .15 1. . . . . . . .3 1. .11 1. . . . . . . . . . . . . . . . .26 1. . . . Same as Figure 1. . . . . .List of Figures 1. . . . . . . Typical seaﬂoor deformation due to dipslip faulting . . Geometry of the source model. . . . . 0. .5 1. . . . Comparisons of the freesurface elevation for S = 0. 0. . the water depth is h = 500 m . . . . .01. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Transient waves generated by an underwater earthquake . . . . . . . . . . . Dipslip fault .1 1. . . . . . . . . . .01 . . . . . . . . . . . . . . . Top view of the initial free surface deformation . . . . . . . . . . .6. . . . . . . Same as Figure 1. . . . . . . . . . . . . .24 1. . . .19 1. . . Relative diﬀerence between the two solutions . Tensile fault . . . .005 . . . . . . . . .17 1. . . . . . . Comparisons of the freesurface elevation for S = 5 . . .I . . . . . . . .6 1. . . . . . .25. . . . . . . . . . . . . . 5 in dimensionless time Velocity ﬁeld computed along the free surface at t = 0. . . . . . . . Wave proﬁles at diﬀerent times for the case of a normal fault . . . . . . . . . . . d2 ω/dm2 Typical graphs of Te (t) and Tc (t) . . . . . Relative diﬀerence between the two solutions . . . 3. . . . . . . . . . .21 1. . . . Transient waves. . . . . .27 1.13 1. . . . . . . . . . . . .12 1. . . . . . . .8 1. . 7 8 11 13 14 14 15 17 17 21 24 33 34 35 39 39 41 53 54 55 56 58 59 60 61 62 63 63 tel00194763. . . .7 1. . . . . . . . . . .22 for later times (t = 150 s. . . . . . . . Comparisons of the freesurface elevation for S = 0. . . . . . . .21 Jan 2008 . . . . . . . . . xi . Initial net volume V of the seaﬂoor displacement . . . . . . Freesurface deformation due to curvilinear faulting . . .25 1. . . . t = 200 s). . .18 1. . . . . . Strikeslip fault . . . . . . . . . . . . . 5 in dimensionless time . . .20 1. . . . . . . 3.04 . . . . . . .10 1. . Transient waves generated by an underwater earthquake . . . . . Geometry of a fault with elliptical shape .14 1. . . . . . . . . . . . Plot of the frequency ω(m) and its derivatives dω/dm. version 2 . . . . . .
. . . . . . . Same as Figure 1. . . . . . . . . . Error of the numerical method . . . .9 2. . . . . . . . . . . . . . Illustration for GreenGauss gradient reconstruction . .8 3. . . . . . . . . . . . . .4 2. . . . . . . . . Two separate waves moving in opposite directions .5 2. .33 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dissipation model II. . Beginning of the solitary wave deformation . . . . . . . . . . . . . . Dissipation model I. . . . . . . Water free surface at t = 150s for a longer earthquake. . . Illustration for cellcentered ﬁnite volume method .12 2. .21 Jan 2008 Example of water/air mixture after wave breaking. . . .32 for times t = 4s and t = 6s.3 3. . . Dissipation model I. . . Separation of the reﬂected wave . . . . . . . . . . Dissipation model II. . . . . . . . . . . . .2 3. . . . . . . . . . . . . . .8 2. . . . Regularized dispersion relation . . . . . . . . . . Comparison between analytical and numerical solutions . . . . . .29 1. . Free surface just before the interaction . .3 2. . . . . Triangles sharing the same vertex N . Diamond cell constructed around face f . . . . . .36 at time t = 250s. . . . . . . . . Solitary wave and a step interaction . . . . . . .32 for times t = 7s and t = 10s. . . . . . . . . . . . Illustration for leastsquares gradient reconstruction. . Water free surface at the beginning of the earthquake (t = 2s). . Zoom on long waves . Real part of the phase velocity. . . . . . . . . . . . . Illustration for vertexcentered ﬁnite volume method . . . . . . . . . .34 1. . . .32 for times t = 150s and t = 250s. . . . . . . . . . . . . . . . . . . . . . . . . . . Imaginary part of the frequency. .10 2. . . Zoom on long waves . . . . . . . . .1 2. . . .36 1. . . . . . . .32 1. . . . . . . . . . . . Freesurface snapshot before the interaction . . . . . . . . .11 2. . . . . . . Sound speed cs in the water/air mixture at normal conditions. . . . . . . . . .2 2. .1 3. . . . . . . . . . . . . .13 2. . . . . . . . . . . . .6 3. . . . . . . . . . . . . Sketch of the ﬂuid domain .4 3. .35 1. . . . Same as Figure 1. . . . . . .15 2.7 3.14 2.37 1. . Initiation of the reﬂected wave separation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .16 3.30 1.38 2. . .5 3. . . . . . . . . . . . .31 1. . version 2 . . . .6 2. . . . . . . . . . . . . Dissipation model I. . . . . . . . . An example of control volume K. . . . . . . . . . . . . . . . . . . . .35 (left plot) for two rupture velocities. . . . . Sketch of the ﬂow . . . . . Same as Figure 1. . . . . . . . . . . . . . . . . . . . . . . Real part of the phase velocity. . . . . . . . . . . . . . Dissipation model II. . . . . . . . . . . . 66 73 74 76 76 77 77 77 78 78 86 96 96 97 97 98 98 107 113 116 116 117 117 118 118 119 123 124 130 132 133 134 140 141 144 146 tel00194763. . . . . . . . . .10 LIST OF FIGURES Geometry of the source model. . . . . . . . . . . . . . . . .9 3. . . . . Same as Figure 1. . . . . . . . . . . . . . . . . . . . .xii 1. . . . . . . . . . . . . . .7 2. . Imaginary part of the frequency. . . . . . Same as Figure 1. . . . . . . . Static deformation due to the dislocation. .
. . . . . . . . .32 3. . . . . . . . . . . . . . . . . . . . . . xiii 150 150 152 153 155 156 157 158 158 159 159 160 160 161 162 163 163 164 164 165 165 166 166 167 172 183 184 184 185 185 188 190 190 191 191 192 tel00194763. . . . . . . . . . . . . . . . .27 3. . . . .0 . . . . . . . . . . . . . . . . . . . . .3 4. . . . . Turbulent mixing process . . . . . . . . . . . . . . . . .13 3. . Vertical jets formation . . . . . . . . . . . . . . . . . . . . . Geometry and initial condition for falling water . . .0 . . . . . . . . . . . . . . . Maximal bottom pressure as a function of time. . . . . . Nonlinear absolute stability regions . . . . . . Phase velocity at t = 1. . . . . . . . . . . . . . . . . .2 4. . . . . . . . . . . . . . . . . . . Splash creation due to the interaction with step . . . . . . . . . . .24 3. . CPU time for diﬀerent ﬁnite volume schemes. . . . Snapshots of the free surface at t = 0. . . . . . . . . . . . . . . . . . . . . . . . . . . . Side jets interﬂow at the center .23 3.7 4. . . . . . . . . . . . Maximal pressure on the right wall as a function of time. . . . . . . . . . . . . . . . . . .18 3. . . . . . . . . . . . . . . . . . . . Water strikes the wall . .5 . . . . . . . . . . . . . . . . . . . . . . . . Water drop test case. . . . . . .0 . . . . . . . Shock tube of Sod. . . . . . . . . . . . . Drop/bottom compressible interaction . . . . . . . Maximal pressure on the right wall. . . . . . . .33 3. . . . . . . . . . . Water strikes the wall .14 3. . Central jet reﬂection from the bottom . . .21 3. . . . . . Geometry and initial condition for water drop test case . version 2 . . . . . . . . . . . . . . . . . . . . .6 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Side jets crossing . . . Numerical method error in L∞ norm. . . . . . . . . . Amplitude of the solitary wave as function of time . . . . . . . . .19 3. . . . .29 3. . . . . . . . .22 3. . . . . . . . . .1 4. .17 3. .20 3. .31 3. . . . . . . .15 3. . . . . . . . . Control volume sharing a face with boundary ∂Ω. . . Water drop test case. . .21 Jan 2008 . . .28 3. . . . . . . . . . . . .8 4. . . This plot represents the density. . . . . . .30 3. . . . . . . . Phase velocity at t = 4. . . . . . . . . . . . . . . . . .16 3. . . . . . The beginning of column dropping process . . . . .LIST OF FIGURES 3. . .9 4. . . . . . Heavy gas case. . . Snapshots of the free surface at t = 2. Conventional partition of the ﬂow region . . . . Real and imaginary part of dispersion curve at t = 0. . . . Water drop test case. . . . . . . . . . . . . . .12 Linear absolute stability regions . . . .12 3. . . .5 4. . . Water drop test case. . . . . . . . Light gas. . . . . Phase velocity at t = 2. . . . .26 3. . . Zoom on the soliton crest at t = 2. . . . . . . . . . . . . . . . . Real part of the phase velocity at t = 4.11 3. .10 4. . . Splash is climbing the wall . . . . . . . . . Drop approaching container bottom . . . . . . . . .11 4. . . . . Snapshots of the free surface at t = 1. . .I .4 4. . . . . . Initial conﬁguration and the beginning of the fall .25 3. . . . . . . . .34 4. . . . . . . . .II . . . . . Amplitude of linear progressive waves . . . . . . . . . . . . . . . .
tel00194763.21 Jan 2008 xiv LIST OF FIGURES . version 2 .
. . . . . . . Parameters set for the source model . . . . . . . . . . .4 1. . . . . . Laplace transforms of diﬀerent timedepencies .6 2. . . . . . . . . 187 Parameters used in the numerical computations .2 1. . . . . . . . . . .1 4. . . . . 114 Parameters for an air/water mixture under normal conditions . . . . . . . . .3 1. . . . . . . . . . . . . . Physical parameters used in the numerical computations . . . . . . . . . . . . 13 16 24 32 38 74 tel00194763. . 128 Parameters of the linear progressive waves . . . . . . . . . . . . . . . . .21 Jan 2008 Typical parameters in tsunami applications . version 2 . . . . . . . . . .2 3.5 1. . . . . . .1 1. . Parameter set used in Figure 1. . . . . . . . . . . . . . . . .2 Parameter set . . . . . . . . . . . Typical physical parameters used in the numerical computations.1 2. . .1 4. . . . . . .List of Tables 1. 89 Solitonstep interaction parameters . . . . . . .8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 xv . . .
version 2 .tel00194763.21 Jan 2008 xvi LIST OF TABLES .
which can be made. . . . The life of a tsunami is usually divided into three phases: the generation (tsunami source). . . . The impact on the coast is then very rapid. . . . . . . . . . . . . . . . xx Dimensionless equations .21 Jan 2008 Contents Propagation of tsunamis . . The third and most diﬃcult phase of the dynamics of tsunami waves deals with their breaking as they approach the shore. The breaking can be progressive. Then the front and the turbulent current behind it move onto the shore. . . . . . .Introduction An expert is a man who has made all the mistakes. in a very narrow ﬁeld. . . . . . which are sometimes overlooked. This phase depends greatly on the bottom bathymetry and on the coastline type. which develops into a progressive bore [Cha05]. . . . . . . xxi Shallowwater equations . . . The breaking can also be explosive and lead to the formation of a plunging jet. . . . . . xxiii Korteweg–de Vries equation . . . . . . . . . . . . . . . . . . . . . . . . . . . xxi Boussinesq equations . . . . . . . the propagation and the inundation. . our purpose here is to recall some of the basics of tsunami modeling and to emphasize some general aspects. xxvi Tsunami runup . . . . . . . . . . . . . . . . . . . . . . . . . . . . similar to the wave that occurs when a dam breaks. . . . . . Niels Henrik David Bohr (1885 – 1962) tel00194763. . . . . . . . . . . . past buildings and vegetation xvii . This turbulent front. . . . can be quite high and travel onto the beach at great speed. . . . xxvi Energy of a tsunami . . . . . . . . . . . . . xviii Classical formulation . the amplitude of tsunami waves grows to such an extent that typically an undulation appears on the long wave. . . . . Structural damages are mainly caused by inundation. . . . . . . . . . . . . . In very shallow water. . . . . . . . . version 2 . . . . . . . . . . . . xxvii Given the broadness of the topic of tsunamis [DD06]. . Then the inundation process is relatively slow and can last for several minutes. . . . . . .
Buildings further away from the beach survived the front in some cases. However there are three main sources of dissipation for water waves: bottom friction. in particular those with reinforced concrete beams properly integrated in the frame structure. typically from 0 to 3 meters in 90 seconds. The horizontal coordinates are denoted by x and y. The water is assumed to be incompressible. Highly porous or open structures survived.xviii INTRODUCTION until they are ﬁnally stopped by rising ground. bottom friction is the most important one. especially in the later stages. The question of dissipation in potential ﬂows and long waves equations is thoroughly investigated in Chapter 4. Some are obvious while some others are questionable under certain circumstances. logs. Dissipation is not often included. and the vertical coordinate . Buildings that had been constructed to satisfy modern safety standards oﬀered a satisfactory resistance. and to the accumulation of debris such as trees. Propagation of tsunamis The problem of tsunami propagation is a special case of the general waterwave problem. cars. since p = ρair U 2 /2). and is sometimes included in the computations in an adhoc way.21 Jan 2008 Civil engineers who visited the damage area following the Boxing day tsunami came up with several basic conclusions. The water level can rise rapidly. Hence research on ﬂooding events and measures to deal with them may be able to contribute to improved warning and damage reduction systems for tsunami waves in the areas of the world where these waves are likely to occur as shallow surge waves (cf. dyke breaking or overtopping of dykes (cf. surface dissipation and body dissipation. version 2 . especially in the ﬁnal stages because they are sensitive to small changes in the topography. tel00194763. A brief description of the common mathematical model used to study water waves follows. it is not included. but they were then destroyed by the erosion of the ground around the buildings by the water currents [HB05]. and to the stochastic patterns of the collapse of buildings. The dynamics of this ﬁnal stage of tsunami waves is somewhat similar to the dynamics of ﬂood waves caused by dam breaking. furniture. In most theoretical analyses. The trajectory of these currents and their velocity are quite unpredictable. the recent tragedy of hurricane Katrina in August 2005). By contrast brick buildings collapsed and were washed away. These were able to withstand pressure associated with the leading front of the order of 1 atmosphere (recall that an equivalent pressure p is obtained with a windspeed U of about 450 m/s. The study of water waves relies on several common assumptions. the recent tragedy of the Indian Ocean tsunami in December 2004). For tsunamis.
t). Dt Dt ∂t In (2). . z. y. xix and the vertical velocity by w(x. t) the pressure ﬁeld. The threedimensional ﬂow of an inviscid and incompressible ﬂuid is governed by the conservation of mass ∇·u+ ∂w =0 ∂z (1) tel00194763. version 2 . t) (the velocity potential) such that u = ∇φ.21 Jan 2008 and by the conservation of momentum ρ Du = −∇p. u = (u. ∂x ∂y . g is the acceleration due to gravity and p(x. Dt ∂z (2) where Df is the material derivative deﬁned as Df := ∂f + u · ∇f . y.INTRODUCTION by z. y. The horizontal gradient is denoted by ∇ := The horizontal velocity is denoted by u(x. ∂z The equation of momentum conservation (2) can be integrated into Bernoulli’s equation 1 ∂φ 1 + ∇φ2 + ∂t 2 2 ∂φ ∂z 2 + gz + p − p0 = 0. ρ (4) which is valid everywhere in the ﬂuid. y. for example the atmospheric pressure. The assumption that the ﬂow is irrotational is commonly made to analyze surface waves. w) = (u. The continuity equation (1) becomes ∇2 φ + ∂ 2φ = 0. The eﬀects of surface tension are not important for tsunami propagation. z. z. The constant p0 is a pressure of reference. v. ρ is the density of water (assumed to be constant throughout the ﬂuid domain). Then there exists a scalar function φ(x. Dt ρ Dw ∂p = −ρg − . ∂z 2 (3) w= ∂φ . w). t) = (u. v) ∂ ∂ . z.
t) := η(x. The main driving force is gravity. When the initial value problem is integrated. version 2 . the boundary condition at the bottom is ∇φ · ∇h + φz = 0 at z = −h(x. y. y. z. y. one can consider that shallowwater theory is appropriate because the water depth (typically several kilometers) is much smaller than the wavelength (typically several hundred kilometers). The conservation of momentum equation (2) is not required in the solution procedure. Another one is the system of nonlinear shallowwater equations that retains nonlinearity but no dispersion. One is the system of Boussinesq equations. In the case of tsunamis propagating on the surface of deep oceans. 0) must be speciﬁed at t = 0.21 Jan 2008 The second boundary condition is the dynamic condition which states that the normal stresses must be in balance at the free surface. it is used a posteriori to ﬁnd the pressure p once η and φ have been found. we will consider various approximations of the full waterwave equations. t) − z = 0. Bernoulli’s equation (4) evaluated on the free surface z = η gives φt + 1 ∇φ2 + 1 φ2 + gη = 0 at z = η(x. (5) tel00194763. the ﬁelds η(x. The concept of shallow water is based on the smallness of the ratio between water depth and wavelength. In the following subsections. t).xx INTRODUCTION Classical formulation The surface wave problem consists in solving Laplace’s equation (3) in a domain Ω(t) bounded above by a moving free surface (the interface between air and water) and below by a ﬁxed solid boundary (the bottom). 2 2 z Finally. y) . The simplest one is the system of linear shallowwater equations. y. y.1 The free surface is represented by F (x. y. The surface wave problem can be easily extended to the case of a moving bottom. z. (7) (6) To summarize. that retains nonlinearity and dispersion up to a certain order. which leads to ηt + ∇φ · ∇η − φz = 0 at z = η(x. The free surface must be found as part of the solution. This extension may be needed to model tsunami generation if the bottom deformation is relatively slow. t) . It can be stated as DF/Dt = 0 (the material derivative of F vanishes). 1 . t) and φ(x. The normal stress at the free surface is given by the diﬀerence in pressure. y. y). the goal is to solve the set of equations (3). 0) and φ(x. z. (5). Two boundary conditions are required. The ﬁrst one is the kinematic condition. t) . (6) and (7) for η(x. The shape of the bottom is given by z = −h(x. y.
y. t). t). we have not used the fact that the numbers α and β are small.25 × 10−4 . at z = αη(x. 1. ℓ2 (8) tel00194763. which are supposed to be small. z ˜ t = ℓt/c0 . y. x y = ℓ˜.21 Jan 2008 where d is a typical water depth. These values give the following ranges for the two dimensionless numbers: 1. φ = gaℓφ/c0 . The assumptions on the smallness of these two numbers are satisﬁed for the Indian Ocean tsunami. Two dimensionless numbers. h = dh. one assumes .INTRODUCTION xxi Dimensionless equations The derivation of shallowwater type equations is a classical topic. the water is considered to be shallow. Shallowwater equations When β is small. the famous speed of propagation of tsunamis in the open ocean ranging from 356 km/h for a 1 km water depth to 712 km/h for a 4 km water depth.5 × 10−4 < α < 6 × 10−4 . are introduced: α= a ≪ 1. y). (10) √ where c0 = gd. a a typical wave amplitude and ℓ a typical wavelength. Indeed the satellite altimetry observations of the tsunami waves obtained by two satellites that passed over the Indian Ocean a couple of hours after the rupture occurred give an amplitude a of roughly 60 cm in the open ocean. βηt + αβ∇φ · ∇η = φz β∇φ · ∇h + φz = 0 at z = −h(x. (9) The equations are more transparent when written in dimensionless variables. For the shallow waterwave theory.7 × 10−5 < β < 6. y z = d˜. d β= d2 ≪ 1. and after dropping the tildes. In particular. The linearized theory of water waves is recovered by letting α go to zero. The new independent variables are x = ℓ˜. η (11) In dimensionless form. z 2 So far. The typical wavelength estimated from the width of the segments that experienced slip is between 160 and 240 km [LKA+ 05]. version 2 . (12) (13) (14) (15) 1 βφt + 1 αβ∇φ2 + 2 αφ2 + βη = 0 at z = αη(x. the equations become β∇2 φ + φzz = 0. no approximation has been made. The water depth ranges from 4 km towards the west of the rupture to 1 km towards the east. The new dependent variables are ˜ ˜ η = a˜.
21 Jan 2008 φ1 (x. The lowestorder term in Laplace’s equation is φ0zz = 0. Powers of α will appear when expanding in Taylor series the freesurface conditions around z = 0. z. t) = 1 (1 24 4 2 (17) 2 + z) [(∇ u)x + (∇ v)y ]. t). 2 vt + α(uuy + vvy ) + ηy − 1 β(utxy 2 (21) (22) + vtyy ) = 0. z. 2 (19) + (∇2 v)y ]. The kinematic condition (20) can be rewritten as ηt + [u(1 + αη)]x + [v(1 + αη)]y = 1 β[(∇2 u)x + (∇2 v)y ]. INTRODUCTION This expansion is substituted into the governing equation and the boundary conditions. t) = − 1 (1 + z)2 (ux + vy ). t) and φ0y = v(x. one obtains 1 βφ0t − 2 β 2 (utx + vty ) + βη + 1 αβ(u2 + v 2 ) = 0. y. y. Solving Laplace’s equation and taking into account the bottom kinematic condition yields the following expressions for φ1 and φ2 : tel00194763. Assume now for simplicity that the water depth is constant (h = 1). (24) (25) (26) . (20) β[ηt + α(uηx + vηy ) + (1 + αη)(ux + vy )] = 1 2 β [(∇2 u)x 6 Diﬀerentiating (19) ﬁrst with respect to x and then to respect to y gives a set of two equations: ut + α(uux + vvx ) + ηx − 1 β(utxx + vtxy ) = 0. (18) The next step consists in retaining terms of requested order in the freesurface boundary conditions. Let φ0x = u(x. For example. (16) The boundary conditions imply that φ0 = φ0 (x. y.xxii that β is small and expand φ in terms of β: φ = φ0 + βφ1 + β 2 φ2 + · · · . vt + α(uvx + vvy ) + ηy = 0. version 2 . y. 6 (23) Equations (21)–(23) contain in fact various shallowwater models. ηt + [u(1 + αη)]x + [v(1 + αη)]y = 0. 2 φ2 (x. y. t). The socalled fundamental shallowwater equations are obtained by neglecting the terms of order β: ut + α(uux + vuy ) + ηx = 0. if one keeps terms of order αβ and β 2 in the dynamic boundary condition (15) and in the kinematic boundary condition (14). Thus the vertical velocity component is zero and the horizontal velocity components are independent of the vertical coordinate z at lowest order.
where the star stands for x. The question of dispersive eﬀects in tsunamis is open. Dispersive shallow waterwave models are presented next. y. (31) α . To solve the problem of tsunami generation caused by bottom displacement. η. t). y. A wellknown code is FUNWAVE. t) := φ∗ (x. The derivatives of the velocity potential evaluated on the free surface are denoted by Φ(∗) (x. is introduced: S= For the Indian Ocean tsunami.24 < S < 46. version 2 . . y. but for the derivation the water depth will be assumed to be constant. sometimes called the Stokes number (in other references it is called Ursell number as well). Note that this model does not include any bottom friction terms. z or t. ηt + [u(h + η)]x + [v(h + η)]y = 0.INTRODUCTION xxiii Recall that we assumed h to be constant for the derivation. t). we provide the guidelines to derive Boussinesqtype systems of equations [BCS02]. Of course. A few propagation codes that include dispersion have been developed [DGK06]. vt + uvx + vvy + gηy = 0. so that the bottom displacement is considered as an instantaneous vertical displacement. one ﬁnds 0. the system of nonlinear shallow water equations reads ut + uux + vuy + gηx = 0. The satellite altimetry observations of the Indian Ocean tsunami clearly show dispersive eﬀects. t) := φ(x. (27) (28) (29) tel00194763. y. the motion of the seaﬂoor obtained from seismological models [Oka85] can be prescribed during a time t0 . developed at the University of Delaware over the past ten years [KWC+ 98]. Most propagation codes ignore dispersion. Boussinesq equations An additional dimensionless number. η. Usually t0 is assumed to be small. In this subsection. This assumption may not be appropriate for slow events. The potential evaluated along the free surface is denoted by Φ(x. the variation of bathymetry is essential for the propagation of tsunamis. Going back to an arbitrary water depth and to dimensional variables. y. Some notation is introduced. β (30) Therefore the additional assumption that S ≈ 1 may be realistic.21 Jan 2008 This system of equations has been used for example by Titov and Synolakis for the numerical computation of tidal wave runup [TS98].
Laplace’s equation as well as the kinematic condition on the bottom must be satisﬁed. The Taylor series for the cosine and sine operators can be truncated. In Boussinesqtype models. (32) (33) 1 1 Φt + gη + 2 ∇Φ2 − 2 W 2 (1 + ∇η · ∇η) = 0. instead of the z = 0 level. y. y. t). The potential φ can be expressed in terms of φo and wo only. z. Demanding that φ formally satisfy Laplace’s equation leads to a recurrence relation between φ(n) and φ(n+2) . In addition. Pad´ approximants can be used in operators e at z = −h and/or at z = 0. one must solve Laplace’s equation in the whole water column. the velocity potential is represented as a formal expansion. Note that φo and wo are nothing else than φ(0) and φ(1) . z. one obtains the velocity ﬁeld in the whole water column (−h ≤ z ≤ η) [MBS03]: u(x. t) z n . One can generalize the expansions to an arbitrary z−level. t) = cos(z∇)uo + sin(z∇)wo . t) := φz (x. The vertical velocity at the free surface is denoted by W (x. and wo the vertical velocity at z = 0. y. t) = cos(z∇)wo − sin(z∇)uo . A wide variety of Boussinesq systems can been derived [MBS03]. y. z. (34) Here the expansion is about z = 0. tel00194763. The boundary conditions on the free surface (5) and (6) become ηt + ∇Φ · ∇η − W (1 + ∇η · ∇η) = 0. Φ∗ (deﬁned for ∗ = z) and Φ(∗) have diﬀerent meanings. They are however related since Φ∗ = Φ(∗) + Φ(z) η∗ .21 Jan 2008 These two nonlinear equations provide timestepping for η and Φ. Let φo denote the velocity potential at z = 0. Here the cosine and sine operators are inﬁnite Taylor series operators deﬁned by z 2n 2n ∇ . ∞ φ(x. (2n + 1)! Then one can substitute the representation (35)(36) into the kinematic bottom condition and use successive approximations to obtain an explicit recursive expression for wo in terms of uo to inﬁnite order in h∇. Finally. y. In order to relate the freesurface variables with the bottom variables. uo the horizontal velocity at z = 0. y. which is the location of the free surface at rest.xxiv INTRODUCTION Consequently. version 2 . . t) = n=0 φ(n) (x. η. w(x. cos(z∇) = (−1) (2n)! n=0 n ∞ ∞ (35) (36) sin(z∇) = n=0 (−1)n z 2n+1 ∇2n+1 .
Considering the Stokes number to be of order one leads to the following system in dimensional form2 : ut + uux + gηx − 1 h2 utxx = 0. It is therefore implicitly assumed that the Stokes number is large. version 2 . and write the equations for one spatial dimension (x). They are obtained by replacing u with the depth averaged velocity 1 h They read ut + uux + gηx − 1 h2 utxx = 0.21 Jan 2008 ηt + [u(h + η)]x − 1 h3 uxxx = 0. it means that the Stokes number is 64 times larger in a 1 km depth than in a 4 km depth! Based on these arguments. A number of variants of the classical Boussinesq system were studied by Bona et al. 2 (37) (38) tel00194763. while η and φo as well as their partial derivatives are of order one. . In the derivation of the fundamental nonlinear shallowwater equations (24)–(26). who in particular showed that depending on the modeling of dispersion the linearization about the rest state may or may not be wellposed. −h ηt + [u(h + η)]x = 0.INTRODUCTION xxv The classical Boussinesq equations are more transparent when written in the dimensionless variables used in the previous subsection. the terms in β were neglected. Classical Boussinesq equations The classical Boussinesq equations are obtained by keeping all terms that are at most linear in α or β. Substituting the expression for φ into the freesurface boundary conditions evaluated at z = 1 + αη(x. 3 (39) (40) η u dz. We further assume that h is constant. Performing the expansion about z = 0 leads to the vanishing of the odd terms in the velocity potential. [BCS02]. 2 Equations (37) and (38) could have been obtained from equations (21) and (23). Since the cube of the water depth appears in the denominator of the Stokes number (S = α/β = aℓ2 /d3 ). 6 The classical Boussinesq equations are in fact slightly diﬀerent. dispersion is more important to the west of the rupture. t) leads to two equations in η and φo with terms of various order in α and β. drop the tildes. The small parameters α and β are of the same order.
21 Jan 2008 The solitary wave solutions of the Korteweg–de Vries equation are of elevation (a > 0) and travel faster than c0 . Note that a natural length scale appears: 4d3 . it gives roughly ℓ = 377 km. Energy of a tsunami The energy of the earthquake is measured via the strain energy released by the faulting. DM04]. and BonaSmith systems. leads to a single equation for η. for example the positive x−direction. However. version 2 .2 × 1015 J. 6 (41) where d is the water depth. It admits solitary wave solutions travelling at speed V in the form 3a a η(x. ℓ= 3a For the Indian Ocean tsunami. It is of the order of magnitude of the wavelength estimated from the width of the segments that experienced slip. Seeking solutions travelling in only one direction. [LKA+ 05].xxvi INTRODUCTION Korteweg–de Vries equation The previous system allows the propagation of waves in both the positive and negative x−directions. Solitary waves exist for more general nonlinear dispersive equations such as KdVKdV. ℓ sech 4 x dx + O(α2 ). 3 4d 2d tel00194763. with V = c0 1 + . The part of the energy transmitted to the tsunami wave is less than one percent Lay et al. Their speed increases with amplitude. They do not give details on how they obtained this estimate. Numerical methods (based on Galerkin discretization) for the approximation of solutions to KdVKdV systems are discussed in [BDM07]. Their existence was recently studied in [FDG05. a simple calculation based on considering the tsunami as a soliton η(x) = a sech 2 gives for the energy 1 E = √ α3/2 ρd2 (c2 + gd) 0 3 ∞ x . the Korteweg–de Vries equation: ηt + c 0 1 + 3η 2d 1 ηx + c0 d2 ηxxx = 0. They estimate the tsunami energy to be 4. −∞ . ℓ u(x) = αc0 sech 2 x . t) = a sech 2 (x − V t) .
The numerical estimate for E is close to that of Lay et al. + (u − c) ∂t ∂x In fact there is no obvious characteristic length in this idealized problem. In terms of the variable c. we restrict ourselves to the description of the ﬂuid ﬂow. the maximum horizontal current is 3 cm/s. version 2 . at this level of approximation. It is also important to point out that a tsunami being a shallow water wave. (2005) [LKA+ 05]. the whole water column is moving as the wave propagates. there is equipartition between kinetic and potential energy. Some authors simply say at this point that ℓ is speciﬁc to the problem under consideration. where ℓ is a characteristic length3 : x = ℓ˜. In addition. ηt + [u(−x tan θ + η)]x = 0. However. 2ct + cux + 2ucx = 0. 3 . The basis of their analysis is the onedimensional counterpart of the system (27)–(29). η = ℓ˜. the current increases and becomes important when the depth becomes less than 500 m.INTRODUCTION xxvii The value for the integral is 4/3. Additional properties of solitary waves can be found for example in [LH74]. For the parameter values used so far. After dropping the tildes. The equations written in characteristic form are ∂ ∂ + (u + c) (u + 2c + t tan θ) = 0. the dimensionless system of equations (27)(29) becomes ut + uux + ηx = 0. Incidently. The problem of waves climbing a beach is a classical one [CG58]. as the water depth decreases. The transformations used by Carrier and Greenspan are still used nowadays. u = x η gℓ u. Although in some cases it may be important to consider the coupling between ﬂuid and structures. c2 = (h + η)/ℓ.21 Jan 2008 The last phase of a tsunami is its runup and inundation. they assume the depth to be of uniform slope: h = −x tan θ. these equations become ut + uux + 2ccx + tan θ = 0. Tsunami runup tel00194763. ∂t ∂x ∂ ∂ (u − 2c + t tan θ) = 0. t = ˜ ˜ ℓ/g t. Introduce the following dimensionless quantities.
The free boundary is the instantaneous shoreline c = 0. The above formulation has been used by several authors to study the runup of various types of waves on sloping beaches [Syn87. λ) such that φσ u= . λ)plane. The nonlinear set of equations have been reduced to a linear equation for u or φ and the free boundary is now the ﬁxed line σ = 0 in the (σ. TT05]. (42) Since u + t tan θ = λ/2. TS94.21 Jan 2008 1 (3r + s) − t tan θ ts = 0. Two major simpliﬁcations have been obtained. CWY03. which moves as a wave climbs a beach. dt dx : = u − c. Introducing the potential φ(σ. 2 2 σ 1 = (r − s) = c. σ one obtains the equation (σφσ )σ − σφλλ = 0 after integrating once. 4 xs − The elimination of x results in the linear secondorder equation for t σ(tλλ − tσσ ) − 3tσ = 0. 4 4 One obtains tel00194763. dt u + 2c + t tan θ = r. version 2 . u must also satisfy (42). Synolakis [Syn87] validated analytical results by comparing them with experiments on the runup of a solitary wave on a plane beach. Next one can rewrite the hyperbolic equations in terms of the new variables λ and σ deﬁned as follows: λ 1 = (r + s) = u + t tan θ.xxviii INTRODUCTION The characteristic curves C + and C − as well as the Riemann invariants are C+ : C− dx = u + c. . 4 1 xr − (r + 3s) − t tan θ tr = 0. He used CarrierGreenspan transformation and linear theory solutions to draw several important conclusions: • A law predicting the maximum runup of nonbreaking waves was given. u − 2c + t tan θ = s.
GGCCD05]. since it allows the determination of the level of risk due to major ﬂooding. version 2 . the determination of security zones. The most natural methods that can be used are the free surface capturing methods based on a ﬁnite volume discretisation. the breaking of tsunami waves as well as the subsequent ﬂoodings must be studied numerically. In [KS06] it was shown that there is a diﬀerence in the maximum runup by a factor 2 in shoreline motions with and without initial velocity. suggesting that perhaps the solitary wave model may not be adequate for predicting an upper limit for the runup of nearshore generated tsunamis. Several important conclusions about the physics of tsunami waves were drew out. of wave breaking on the land behind beaches. applied to freesurface ﬂow problems [Mon94. Chapter 2 discusses diﬀerent possibilities of introducing dissipative eﬀects into Boussinesq equations. • Nonlinear theory models adequately describe surface proﬁles of nonbreaking waves. However. it was shown that leading depression N waves runup higher than leading elevation N waves. It also provides some help in the conception and validation of protection systems in the most exposed areas.INTRODUCTION xxix • The author found diﬀerent runup variations for breaking and nonbreaking solitary waves. and the family of Smoothed Particle Hydrodynamics methods (SPH). Since runups of 30 meters were observed in Sumatra during the Boxing Day tsunami. GGD04.2) using diﬀerent mathematical models. Analytical models are useful. such as the Volume Of Fluid (VOF) or the Level Set methods. especially to perform parametric studies. An eﬃcient numerical method based on a ﬁnite volume scheme is described and a few numerical results are presented. The present manuscript is organized as follows. Pros and cons of two models under consideration are revealed.3) and the ﬁrst minutes of propagation (Section 1.21 Jan 2008 . and of the ﬂow over rising ground with and without the presence of obstacles. In Chapter 1 we extensively investigate the tsunami generation process (see Sections 1. the slip might have been of 15 meters or even more. the prediction of the resulting water levels in the ﬂooded areas. There is a rule of thumb that says that the runup does not usually exceed twice the fault slip. Such methods allow a study of ﬂood wave dynamics. This task is an essential part of tsunami modelling. Later.1 and 1. • Diﬀerent criteria were considered whether a solitary wave with given height/depth ratio will break or not as it climbs up a sloping beach. In the last Chapter 4 we return to the investigation of dissipative eﬀects onto water waves. Novel bottom tel00194763. This diﬀerence suggests strong implications for the runup predictions and tsunami warning system [GBM+ 05] if the appropriate initial velocity is not speciﬁed. In Chapter 3 our attention is attracted by twophase ﬂows and wave impact problem.
xxx INTRODUCTION boundary layer correction is derived and the new set of governing equations is thoroughly analysed.21 Jan 2008 . tel00194763. version 2 .
. . . . . . . . . . . . . .2. . . . . . . . Numerical method . . . . .8 1. . Velocity ﬁeld . . . . tel00194763. . . . . . .1. . . . . . . . . . . .5 1. . .2 1. . . . Solution in ﬂuid domain . . . . . . . . . . . . . .1.2. .1. . Mathematical model . . . . .1 1. . . . . . . . .7 1. . . . . . .6 1. . . . . Passive generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4 1. . .2. . . . . . . . . . . . . . . .1. . . . . .1. . . . . 1 . Active generation . . . . . . . . . . . . Nonlinear shallow water equations . . . . Dislocations in elastic halfspace . . . . .7 1. . . . . . . . . . version 2 . Linear theory . . .2. . . . . Volterra’s theory of dislocations . . 1.2.3 1. . . . . . . . . .5 1. . . . . . . . . . . . . . . . . . . . . . . . . . .3 1. . . . . . . . . . . . . . . . . . .1. .2 1. . Freesurface elevation . . . .2 Source model . . . .2. . . . . . .1 1. . . 2 4 5 6 16 21 25 29 31 36 37 43 44 44 47 48 50 50 Comparison of tsunami generation models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1. . . . . .1 Waves generated by a moving bottom . . .2. . . .6 1. .8 Physical problem description . . . . . . . . .1. . . . . Asymptotic analysis of integral solutions . . . .2.21 Jan 2008 Sir Arthur Conan Doyle Contents 1. . . . . . . . . . . . . Numerical results . . . . . . . . . . . . . . . . . . . . . 1. Numerical method for the full equations . . . .4 1. . . .Chapter 1 Tsunami generation modelling From a drop of water a logician could predict an Atlantic or a Niagara. . . . . . . . . . .
. In this chapter. . . . . . . . there was in the northern extent of the source a relatively slow faulting motion that led to signiﬁcant vertical bottom motion but left little record in the seismic data. . . . . Numerical methods . . . . 1. . .3. . . . . . . . . . . . fall of an object into the liquid. . . . . . . . . . . . . This was emphasized for example by Trifunac and Todorovska [TT01]. .3. liquid inside a moving container. . . . . . . . . . . . It is interesting to point out that it is the inversion . . etc. . . . . moving disturbance on the bottom or the surface.4 1. . . . . . . . . One example is the generation of tsunamis by a sudden seaﬂoor deformation. . . . and a model of surface gravity waves generation resulting from this motion of the seaﬂoor. . . The resulting waves have some wellknown features. . characteristic wavelengths are large and wave amplitudes are small compared with water depth. . . . . . . or even inside the liquid. . . . . slides.1 Waves generated by a moving bottom Waves at the surface of a liquid can be generated by various mechanisms: wind blowing on the free surface. . . .6 Introduction . . . there are some speciﬁc cases where the time scale of the bottom deformation may become an important factor. . . .3 1. Validation of the numerical method . . . The details of wave motion are neglected during the time that the source operates. . . . . . .9 Tsunami generation Comparisons and discussion .1 1.3. . Mathematical models . . . . . wavemaker. . . . . etc. 51 64 64 64 67 70 72 73 78 1. . . . . . . .3. . . . . . . . . For example. . . . . .2 1. . While this is often justiﬁed because the earthquake rupture occurs very rapidly. . . For example.2. . . volcanic explosions. . .2 1. . .10 Conclusions . 1. During the 26 December 2004 SumatraAndaman event. . . . . .3. . . Results of the simulation . . . . one can mention submarine slumps. . Conclusions . . . .3 Tsunami generation by dynamic displacement of sea bed . . Two factors are usually necessary for an accurate modelling of tsunamis: information on the magnitude and distribution of the displacements caused by the earthquake. .5 1. . .2. . . . . . . Most studies of tsunami generation assume that the initial freesurface deformation is equal to the vertical displacement of the ocean bottom. . . tel00194763. . . .21 Jan 2008 1. . . . . . we concentrate on the case where the waves are created by a given motion of the bottom. . . . . . who considered the generation of tsunamis by a slowly spreading uplift of the seaﬂoor and were able to explain some observations. . . . .3. . . . . . . In this chapter we use a submarine faulting generation mechanism as tsunami source. . version 2 . . . . . There are diﬀerent natural phenomena that can lead to a tsunami. . .
the northernmost of the Indian eastcoast stations. Kel63] to use asymptotic methods such as the method of stationary phase to estimate the farﬁeld behaviour of the solutions. They are useful from a qualitative point of view.21 Jan 2008 . Several authors have modeled the incompressible ﬂuid layer as a special case of an elastic medium [Pod68. Gus72.1 Waves generated by a moving bottom 3 of tidegauge data from Paradip. The models discussed in these papers lack ﬂexibility in terms of modelling the source due to the earthquake. BvdDP73]. the generation time is also longer for landslide tsunamis. Kaj63. The wave propagation away from the source can be investigated by shallow water models which may or may not take into account nonlinear eﬀects and frequency dispersion. Our study is restricted to the water region where the incompressible Euler equations for potential ﬂow can be linearized. The problem of tsunami generation has been considered by a number of authors: see for example [Car71. A good review on the subject is [Sab86]. but it is questionable from a physical point of view. [LKA+ 05]. The crust was modeled as an elastic isotropic halfspace. This assumption will also be adopted in the present study. This approach was followed recently in [TT01. vdD72. Such models include the Kortewegde Vries equation [KdV95] for unidirectional propagation. THT02].1. Most analytical studies of linearized wave motion use integral transform methods. The main extensions given in the present work consist in threedimensional modelling and more realistic source models. BBM72]. The present study provides some extensions. The complexity of the integral solutions forced many authors [Kaj63. In the present study we have also obtained asymptotic formulas for integral solutions. Per66. The tsunami generation problem is reduced to a CauchyPoisson boundary value problem in a region of constant depth. One should use asymptotic solutions with caution since they approximate exact solutions of the linearized problem. Gus76]. In our opinion it may be convenient to model the liquid by an elastic material from a mathematical point of view. version 2 . All the numerical results presented in this section were obtained in this manner. Here we essentially follow the framework proposed by Hammack [Ham73] and others. [NSS+ 05] to conclude that the source length was greater by roughly 30% than the initial estimate of Lay et al. but in practice it is better to use numerical integration formulas [Fil28] that take into account the oscillatory nature of the integrals. The relative importance of linear and nonlinear eﬀects can tel00194763. Incidentally. that led Neetu et al. AG73. where the mathematical model was the same as in [Ham73] but the source was diﬀerent. nonlinear shallowwater equations and Boussinesqtype models [Bou71b.
scientists have been relatively successful in using simple models for the source. have been described in various papers. homogeneity. Its description follows. they will become important. the potential ﬂow formulation are 3D. The fracture zones.4 be measured by the Stokes (or Ursell) number [Urs53]: U := a a/h = 2 3. One of these models is Okada’s model [Oka85]. BMSS70. Recently. and consider a semiinﬁnite medium. nonhomogeneity. 1 tel00194763. Hence. These equations have the advantage of being twodimensional while the physical problem and. regardless of how small nonlinear eﬀects are initially. the suggestion makes sense. it is necessary for simplicity’s sake to make some assumptions. the nonlinear eﬀects control wave propagation and only nonlinear models are applicable. If the mechanism involved in earthquakes and the fracture zones is indeed one of fracture.1 Source model The inversion of seismic wave data allows the reconstruction of permanent deformations of the sea bottom following earthquakes. temperature. Several studies showed that the eﬀect of earth curvature is negligible for shallow events at distances of less than 20◦ [BMSS69. consequently. Here we neglect the curvature of the earth. As dislocation theory may be described as that part of the theory of elasticity dealing with surfaces across which the displacement ﬁeld is discontinuous. a a typical wave amplitude and h the water depth.1. The sensitivity to earth topography. the methodology used in this thesis for tsunami generation was applied in the framework of the Boussinesq equations [Mit07] over uneven bottom. For example. magnetism. For U ≫ 1. isotropy and halfspace assumptions was studied and discussed recently . On the other hand. As is often done in mathematical physics. Ursell [Urs53] proved that near the wave front U behaves like U ∼ t3 . which is homogeneous and isotropic. version 2 . along which the foci of earthquakes are to be found.21 Jan 2008 1. Boussinesq equations represent a long wave approximation to the complete waterwave problem. SM71]. This suggestion was made for the following reason. In spite of the complexity of the seismic source and of the internal structure of the earth. 2 (kh) k h Tsunami generation where k is a wave number. discontinuities in the displacement components across the fractured surface will exist. it has been suggested that Volterra’s theory of dislocations might be the proper tool for a quantitative description of these fracture zones [Ste58]. We further assume that the laws of classical linear elasticity theory hold. its gravity.
8πµ with α = λ+µ . The coeﬃcient α can be rewritten as α = 1/2(1 − ν).2 tel00194763.1. O) = i F (δik r. version 2 . this section is devoted to the case of an entire elastic space. . O) = − (1. The stresses due to the displacement ﬁeld (1. Let O be the origin of a Cartesian coordinate system in an inﬁnite elastic medium. 3). xi the Cartesian coordinates (i = 1. λ+µ The notation r.1. as was done in the original paper by Volterra [Vol07]. which is determined i by the wellknown Somigliana tensor uk (P.1) In this relation δik is the Kronecker delta. O) at point P .1) are easily computed from Hooke’s law: σij = λδij uk. ik ). A force F = F ek at O generates a displacement ﬁeld uk (P. which is based on a ﬁnite element model (FEM). was used.1 Waves generated by a moving bottom 5 [Mas03]. The resulting mixed boundaryvalue problem was solved a century ago [Vol07]. and ei a unit vector in the positive xi −direction.2) αF 4π µ δki xj + δkj xi − δij xk 3xi xj xk + 5 r λ+µ r3 . A commercially available code. and r is the e distance from P to O. where ν is Poisson’s ratio. Steketee proposed an alternative method to solve this problem using Green’s functions [Ste58]. i means ∂r/∂xi and the summation convention applies. Six FEMs were constructed to test the sensitivity of deformation predictions to each assumption. 2. The author came to the conclusion that the vertical layering of lateral inhomogeneity can sometimes cause considerable eﬀects on the deformation ﬁelds.i ). which is deﬁned as E= µ (3λ + 2µ) . Later we will also use Young’s modulus E. λ and µ are Lam´’s constants. The components of the force per unit area on a surface element are denoted as follows: k Tik = σij νj .j + uj. nn − αr. The usual boundary conditions for dealing with earth problems require that the surface of the elastic medium (the earth) shall be free from forces. ABACUS. 1.k + µ(ui. λ + 2µ (1.21 Jan 2008 Volterra’s theory of dislocations In order to introduce the concept of dislocation and for simplicity’s sake. One ﬁnds k σij (P. Later.
i i Ωij = −Ωji .5) and (1. It is clear from (1.21 Jan 2008 uk (Q) = 1 F Σ ∆ui Tik dS. The displacement ﬁeld in an inﬁnite elastic medium due to the dislocation is then determined by Volterra’s formula [Vol07] tel00194763. Maruyama derived the remaining ﬁve functions [Mar64]. the dislocation is essentially determined by the six constants Ui and Ωij . and the stresses σij (P.5) remains valid. (1.5) Once the surface Σ is given. (1.6 Tsunami generation where the νj ’s are the components of the normal to the surface element. This can be explained by the fact that the elementary solutions for a halfspace are diﬀerent from Somigliana solution (1. (1. . A force F ek is applied at Q. Ω31 .2).1).3) in which Ui and Ωij are constants is the wellknown Weingarten relation which states that the discontinuity ∆ui should be of the type of a rigid body displacement. but k k we have to replace σij in Tik by another tensor ωij . In particular the components of the force on Σ are computed.6) where Ωij takes only the values Ω12 . Steketee showed a method of obtaining the six ωij ﬁelds k by using a Green’s function and derived ω12 . thereby maintaining continuity of the components of stress and strain across Σ. Q) that this force generates are computed at the points P (xi ) on Σ. Therefore we also write uk (Q) = Ui F Σ k σij (P.6) an elementary dislocation. A Volterra dislocation is deﬁned as a surface Σ in the elastic medium across which there is a discontinuity ∆ui in the displacement ﬁelds of the type ∆ui = u+ − u− = Ui + Ωij xj . After multiplication with prescribed weights of magnitude ∆ui these forces are integrated over Σ to give the displacement component in Q due to the dislocation on Σ.6) that the computation of the displacement ﬁeld uk (Q) k is performed as follows. version 2 . Q)}νl dS. 1.3) (1.3 Dislocations in elastic halfspace When the case of an elastic halfspace is considered. Following Volterra [Vol07] and Love [Lov44] we call each of the six integrals in (1. Q)νj dS + Ωij F Σ k k {xj σil (P.4) Equation (1. k The ωij can be obtained from the displacements corresponding to nuclei of strain in a k halfspace through relation (1.1. equation (1. which is relevant to a vertical strikeslip fault (see below). Ω23 . Q) − xi σjl (P.
It is only necessary to write the single force results since the other forms can be obtained by taking appropriate derivatives. MC50].21 Jan 2008 y D W U3 U2 δ U1 L x Figure 1. In this coordinate system.7) x3 uj (x1 . z tel00194763. uj (x1 . Min36. x3 ) i iA iA +uj (x1 . The coeﬃcients are chosen to satisfy boundary and equilibrium conditions. ξ3 ). x3 ). we take the Cartesian coordinate system shown in Figure 1. Pre65. ξ3 ) is the ith component of i the displacement at (x1 . x3 ) iB + (1. version 2 . x3 ) = uj (x1 . x2 . x3 . x2 . Here.1.1: Coordinate system adopted in this study and geometry of the source model. x2 . ξ2 . These solutions were also derived by Press in a slightly diﬀerent manner [Pre65]. Oka92]: uj (x1 . x2 . This vector is then determined by taking a linear combination of some biharmonic elementary solutions. x2 . −x3 ) − uj (x1 . who gave explicit expressions of the displacement and stress ﬁelds for halfspace nuclei of strain consisting of single forces with and without moment. It can be expressed as follows [Oka85. x3 ) due to the jth direction point force of magnitude F at (ξ1 . ξ2 . iC .1 Waves generated by a moving bottom 7 It is interesting to mention here that historically these solutions were ﬁrst derived in a straightforward manner by Mindlin [Min36.1. x2 . ξ1 . The method consists in ﬁnding the displacement ﬁeld in Westergaard’s form of the Galerkin vector [Wes35]. x2 . The elastic medium occupies the region x3 ≤ 0 and the x1 −axis is taken to be parallel to the strike direction of the fault.
8 Tsunami generation y φ x tel00194763. uj = iC F Ri δj3 − Rj δi3 Ri Rj δij (1 − 2δi3 ) (2 − α) + αξ3 −3 5 4πµ R3 R3 R The ﬁrst term in equation (1. x3 ). The remaining term. .2: Illustration for strike angle deﬁnition. . reduces to the formula for the surface displacement ﬁeld iB 2 2 2 In these expressions R1 = x1 − ξ1 . −x3 ). ξ2 . R2 = x2 − ξ2 . version 2 . the ﬁrst and the second terms cancel each other. x2 . x3 ) in iB iC the fourth term are naturally depth dependent. uj (x1 . uj (x1 . iA which represents the displacement ﬁeld due to a single force placed at (ξ1 . x2 . is the wellknown Somigliana tensor. ξ3 ) in an inﬁnite medium [Lov44]. The third term. This term corresponds to a contribution from an image source of the given point force placed at (ξ1 . R3 = −x3 − ξ3 and R2 = R1 + R2 + R3 . where F 8πµ δij Ri Rj +α 3 R R uj = iA uj = iB (2 − α) . When x3 is set equal to zero in equation (1. ξ2 . −ξ3 ) in the inﬁnite medium. and uj (x1 .21 Jan 2008 Figure 1. x2 . F δij Ri Rj 1 − α δij + + + 4πµ R R3 α R + R3 + Ri Rj Ri δj3 − Rj δi3 (1 − δj3 ) − (1 − δi3 )(1 − δj3 ) R(R + R3 ) R(R + R3 )2 .7). The second term also looks like a Somigliana tensor. x2 . uj (x1 . 0).7). and the fourth term vanishes.
1. . u3 = 1 u3 = 2 3 u3 = F (x1 4πµ F (x2 4πµ F 4πµ 1 R ξ − ξ1 ) − R33 + ξ − ξ2 ) − R33 + 2 ξ3 R3 µ 1 λ+µ R(R−ξ3 ) µ 1 λ+µ R(R−ξ3 ) . x3 ) + ∂ξk ∂ξk ∂ξk j ∂uiB ∂uj + (x1 . . x3 ) + x3 iC (x1 . . F (x1 4πµ F 4πµ 1 R − ξ1 )(x2 − ξ2 ) + (x2 −ξ2 R3 )2 1 R3 + µ λ+µ tel00194763. The ξk derivatives are expressed as follows: ∂uj ∂uj ∂uj i iA (x1 . In order to obtain the displacements due to the dislocation we need to calculate the corresponding ξk derivatives of the point force solution (1.21 Jan 2008 F (x2 4πµ ξ − ξ2 ) − R33 − µ 1 λ+µ R(R−ξ3 ) µ 1 λ+µ R(R−ξ3 )2 (x2 −ξ 2 1 − R(R−ξ23))2 R−ξ3 − . x2 . .5) 1 ui = F Σ ∆uj ∂un λδjk i + µ ∂ξn ∂uj ∂uk i + i ∂ξk ∂ξj νk dS. 2 In these formulas R2 = (x1 − ξ1 )2 + (x2 − ξ2 )2 + ξ3 . x3 ) = (x1 . x2 . version 2 . x2 .7) and to insert them in Volterra’s formula (1.1 Waves generated by a moving bottom due to a point force in a halfspace [Oka85]: u1 = 1 u1 = 2 1 u3 = u2 = 1 u2 = 2 2 u3 = 9 F 4πµ 1 R + (x1 −ξ1 )2 R3 + µ λ+µ 1 R3 1 R−ξ3 F (x1 4πµ F (x1 4πµ − ξ1 )(x2 − ξ2 ) ξ − ξ1 ) − R33 − µ 1 λ+µ R(R−ξ3 ) − µ 1 λ+µ R(R−ξ3 )2 − (x1 −ξ1 )2 R(R−ξ3 )2 . −x3 ) − iA (x1 . x2 . ∂ξk ∂ξk . x2 . + + µ 1 λ+µ R . . x3 ).
Ri δjk + Rj δik − Rk δij Ri Rj Rk F − +3 + 3 4πµ R R5 1 − α δ3k R + Rk δik δj3 − δjk δi3 (1 − δj3 ) + δ − + 2 ij α R(R + R3 ) R(R + R3 ) δ3k R2 + Rk (2R + R3 ) + Ri δj3 − Rj δi3 (1 − δj3 ) + R3 (R + R3 )2 2δ3k R2 + Rk (3R + R3 ) Ri δjk + Rj δik − Ri Rj +(1 − δi3 )(1 − δj3 ) R(R + R3 )2 R3 (R + R3 )3 δjk δi3 − δik δj3 3Rk (Ri δj3 − Rj δi3 ) F + (1 − 2δi3 ) (2 − α) + 4πµ R3 R5 δij Ri δjk + Rj δik + Rk δij 3Ri Rj 5Ri Rj Rk +αδ3k 3 − + 3αξ3 − 5 5 R R R R7 . using Chinnery’s notation to represent the substitution f (ξ. Oka92. dipslip and tensile components of an arbitrary dislocation. ˜ ˜ d = η sin δ − q cos δ . Chi63. p − W ). is introduced in order to provide an orientation of the fault. y = η cos δ + q sin δ.3. where p = y cos δ + d sin δ. Here we give the results of their computations. When dealing with a geophysical application. η) = f (x. the slip or rake angle θ (0 ≤ θ ≤ π).1 each vector represents the direction of the elementary faults. We deﬁne the elementary dislocations U1 .21 Jan 2008 ∂uj iC ∂ξk = .1.3). an additional angle. IS79]. corresponding to the strikeslip. Next we introduce the notation q = y sin δ − d cos δ. 1. tel00194763. The ﬁnal results are represented below in compact form.2 for strike angle deﬁnition). p) + f (x − L.1 Finite rectangular source Let us now consider a more practical problem. For a ﬁnite rectangular fault with length L and width W occurring at depth d (Figure 1.10 with F ∂uj iA = ∂ξk 8πµ ∂uj iB ∂ξk = (2 − α) Rk Ri δjk + Rj δik Ri Rj Rk δ −α + 3α 3 ij 3 R R R5 Tsunami generation . This was done by several authors [Oka85. the azimuth or strike (see Figure 1. A general dislocation can be determined by three angles: the dip angle δ of the fault (0 ≤ δ ≤ π). The vector D is the socalled Burger’s vector.3. In Figure 1. The general situation is schematically described in Figure 1. the deformation ﬁeld can be evaluated analytically by a change of variables and by integrating over the rectangle. and the angle φ between the fault plane and Burger’s vector D. U2 and U3 . p) − f (x. which shows how both sides of the fault are spread out: D = u+ − u− . version 2 . p − W ) − f (x − L. SM74.
u3 = − U2 2π U2 = − 2π U2 2π For a dipslip dislocation. + + I4 sin δ R(R + η) R + η . U3 = D sin φ. U2 = D cos φ sin θ.3: Geometry of the source model (dip angle δ. R(R + η) R + η ˜ dq q sin δ . width W ) and orientation of Burger’s vector D (rake angle θ. and ˜ R2 = ξ 2 + η 2 + q 2 = ξ 2 + y 2 + d2 . length L. version 2 . depth df . one has u1 = − u2 q − I3 sin δ cos δ . angle φ between the fault plane and Burger’s vector).1 Waves generated by a moving bottom z 11 y O x D φ θ L x′ W δ tel00194763. u3 = − .21 Jan 2008 Figure 1. .1. The quantities U1 . R yq ˜ ξη + cos δ arctan − I1 sin δ cos δ R(R + ξ) qR ˜ dq ξη + sin δ arctan − I5 sin δ cos δ R(R + ξ) qR . ˜ X 2 = ξ 2 + q2. For a strikeslip dislocation. U2 and U3 are linked to Burger’s vector through the identities U1 = D cos φ cos θ. one has u1 = − u2 U1 2π U1 = − 2π U1 2π ξq ξη + arctan + I1 sin δ R(R + η) qR yq ˜ q cos δ + + I2 sin δ .
and 1. I1 = − I3 I4 I5 µ ξq .1.3. ˜ λ+µR+d Figures 1. . . ˜ λ + µ cos δ R + d 1 µ ˜ log(R + d) − sin δ log(R + η) . . ˜ λ+µR+d µ ξ sin δ = − . . = ˜ (R + d)2 ˜ 2(λ + µ) R + d µ q = − .1.6 show the freesurface deformation due to the three elementary dislocations. version 2 . 1. 1.2 Curvilinear fault In the previous subsection analytical formulas for the freesurface deformation in the special case of a rectangular fault were given. λ+µ µ 1 y ˜ − log(R + η) + tan δI4 . µ + λ cos δ µ η(X + q cos δ) + X(R + X) sin δ 2 arctan . In fact. λ + µ cos δ ξ(R + X) cos δ tel00194763. .12 For a tensile fault dislocation. The shape of the fault and Burger’s vector are suggested by seismologists and . The values of the parameters are given in Table 1. The terms I1 . R(R + η) ˜ −dq ξη ξq − sin δ − arctan − I1 sin2 δ R(R + ξ) R(R + η) qR ξη ξq yq ˜ + cos δ − arctan − I5 sin2 δ R(R + ξ) R(R + η) qR . one has u1 = u2 = u3 = U3 2π U3 2π U3 2π Tsunami generation q2 − I3 sin2 δ .21 Jan 2008 I2 = I3 = I4 = I5 = and if cos δ = 0.4.5. I5 are given by I1 = − ξ µ − tan δI5 . Volterra’s formula (1. ˜ λ + µ (R + d) cos δ µ − log(R + η) − I3 . ˜ 2(λ + µ) (R + d)2 yq ˜ η µ + − log(R + η) .5) allows to evaluate the displacement ﬁeld that accompanies fault events with much more general geometry.
The horizontal distances x and y are expressed in kilometers. tel00194763. Here a is D (15 m in the present application).1. version 2 . 1. .5.1 Waves generated by a moving bottom 13 parameter Dip angle δ Fault depth d.21 Jan 2008 Figure 1. km Fault width W . m Young modulus E.5 0.23 Table 1. GPa Poisson’s ratio ν value 13◦ 25 220 90 15 9. 0).4. and 1.1: Parameter set used in Figures 1.6. U2 . D = (0. km Ui . km Fault length L.4: Dimensionless freesurface deformation z/a due to dipslip faulting: φ = 0. θ = π/2.
D = (0. D = (U1 . Here a is D. The horizontal distances x and y are expressed in kilometers. Figure 1. θ = 0.5: Dimensionless freesurface deformation z/a due to strikeslip faulting: φ = 0.14 Tsunami generation tel00194763. U3 ). . The horizontal distances x and y are expressed in kilometers. 0.6: Dimensionless freesurface deformation z/a due to tensile faulting: φ = π/2. 0). version 2 .21 Jan 2008 Figure 1. Here a is D (15 m in the present application). 0.
21 Jan 2008 d D n b a Figure 1. η) = η. 0 ≤ ξ ≤ a. a2 (a2 − ξ 2 ) F = 0. z y O x tel00194763. Here we will consider the case of a fault whose geometry is described by an elliptical arc (see Figure 1. c c − ≤η≤ .5) to a double Riemann integral. y(ξ.1. a4 + (b2 − a2 )ξ 2 We also need to compute the coeﬃcients of the ﬁrst fundamental form in order to reduce the surface integral in (1. η) = ξ.7). a a2 − ξ 2 . a Then the unit normal to this surface can be easily calculated: z(ξ. η) = −(b + d) + n= bξ a4 + (b2 − a2 )ξ 2 . 2 2 b a2 − ξ 2 . These coeﬃcients are E= a4 + ξ 2 (b2 − a2 ) . 0. version 2 .7: Geometry of a fault with elliptical shape The parametric equations of this surface are given by x(ξ. G=1 .1 Waves generated by a moving bottom 15 after numerical integration one can obtain the deformation of the seaﬂoor for more general types of events as well.
tel00194763. The example considered in this subsection may not be physically relevant. It is bounded above by the free surface of . km Ellipse semimajor axis b. km Ellipse semiminor axis a. After having reviewed the description of the source. The result is presented on Figure 1. it is natural to impose the condition that D · n = 0.9. km Young modulus E. we now switch to the deformation of the ocean surface following a submarine earthquake.8. Since in the crust the hydrostatic pressure is very large.23 Tsunami generation Table 1. The parameter values are given in Table 1. and the surface element dS is √ 1 dS = EG − F 2 dξdη = a a4 + ξ 2 (b2 − a2 ) a2 − ξ 2 dξdη.16 parameter Depth event d. This condition is obviously satisﬁed if we take Burger’s vector as a a2 − ξ 2 bξ D=D .2: Parameter set used in Figure 1. GPa Poisson’s ratio ν value 20 17 6 15 9. Let us consider the ﬂuid domain Ω shown in Figure 1. However it shows how Okada’s solution can be extended.2.1.4 Solution in ﬂuid domain The ﬂuid domain is supposed to represent the ocean above the fault area. Then this deformation is translated to the free surface of the ocean and serves as initial condition of the evolution problem described in the next section. 0.8. The numerical integration was performed using a 9point twodimensional Gausstype integration formula. The traditional approach for hydrodynamic modelers is to use elastic models similar to the model we just described with the seismic parameters as input in order to evaluate the details of the seaﬂoor deformation. km Fault width c.21 Jan 2008 It is evident that D = D. version 2 . The physical meaning of this condition is that both sides of the fault slide and do not detach.5 0. 4 + ξ 2 (b2 − a2 ) 4 + ξ 2 (b2 − a2 ) a a 1. For a more precise modeling of the faulting event we need to have more information about the earthquake source and its related parameters. − .
The horizontal distances x and y are expressed in kilometers. z η(x. y. t) h Figure 1. t) y x O Ω ζ(x. y.21 Jan 2008 Figure 1.1.9: Deﬁnition of the ﬂuid domain and coordinate system .1 Waves generated by a moving bottom 17 tel00194763. version 2 .8: Freesurface deformation due to curvilinear faulting.
(1. y.18 Tsunami generation the ocean and below by the rigid ocean ﬂoor.12) The signiﬁcance of the various terms in the equations is more transparent when the equations are written in dimensionless variables. . respectively. the dynamic condition to be satisﬁed on the free surface reads ∂φ 1 + ∇φ2 + gη = 0.10) Assuming that viscous eﬀects as well as capillary eﬀects can be neglected. 0) = 0 and ζ(x. z = −h + ζ(x.11) As described above. y. y = κy. By deﬁnition of φ.8) tel00194763. Thus. (x. For time t > 0 the bottom boundary moves in a prescribed manner which is given by z = −h + ζ(x. t). version 2 . η(x. ∂t 2 z = η(x. y. the ﬂuid velocity vector can be expressed as q = ∇φ. Thus. y. t). y. y. The domain Ω is unbounded in the horizontal directions x and y. (1. z. y. z = κz.9) (1. at time t = 0. t)] . y. y.21 Jan 2008 The potential φ(x. t). the initial conditions are given by η(x. (1. y. Note that here the same unit length is used in the horizontal and vertical directions. t = σt. respectively: ∂φ ∂η ∂φ ∂η ∂φ ∂η = + + . t). ∂z ∂t ∂x ∂x ∂y ∂y ∂φ ∂ζ ∂φ ∂ζ ∂φ ∂ζ = + + . The new independent variables are x = κx. It is also assumed that the ﬂuid is incompressible and the ﬂow is irrotational. 0) = 0. y. Initially the ﬂuid is assumed to be at rest and the sea bottom to be horizontal. as opposed to shallowwater theory. z) ∈ Ω. The resulting deformation of the free surface z = η(x. t). the free surface and the sea bottom are deﬁned by z = 0 and z = −h. z. (1. t) must also satisfy the following kinematic boundary conditions on the freesurface and the solid boundary. t) which completely describes this ﬂow. the continuity equation becomes ∇ · q = ∆φ = 0. and can be written as Ω = R2 × [−h + ζ(x. The displacement of the sea bottom is assumed to have all the properties required to compute its Fourier transform in x. y. y and its Laplace transform in t. t) must be found. ∂z ∂t ∂x ∂x ∂y ∂y z = η(x. where κ is a wavenumber and σ is a typical frequency. The latter implies the existence of a velocity potential φ(x.
(1. y. y.1 Waves generated by a moving bottom The new dependent variables are η η= . tel00194763. The linearized problem in dimensional variables reads ∆φ = 0. version 2 . aσ 19 where a is a characteristic wave amplitude. equations (1. a ζ ζ= . and after dropping the tildes. z) ∈ Ω. (1. t). (x. In dimensionless form. A dimensionless water depth is also introduced: h = κh.16) ∂φ ∂η = . +g 2 ∂t ∂z z = 0. t). we now switch back to the physical variables. y. z = 0. For the sake of convenience. ∂z ∂t ∂φ ∂ζ = .14) and (1. the linearized problem can be found by expanding the unknown functions as power series of a small parameter ε := κa. Collecting the lowest order terms in ε yields the linear approximation.1.17) . .13) (1. In fact. .15) (1.14) (1.16) yields the single freesurface condition ∂2φ ∂φ = 0. a φ= κ φ.8–1. (x. y. z = −h + κa ζ(x. z) ∈ R2 × [−h. y. In this study we linearize the equations and the boundary conditions by taking the limit as κa → 0. 0].11) become ∆φ = 0. z = κa η(x. ∂t 2 σ Finding the solution to this problem is quite a diﬃcult task due to the nonlinearities and the a priori unknown free surface.21 Jan 2008 ∂η ∂φ ∂η ∂φ ∂η ∂φ = + κa + ∂z ∂t ∂x ∂x ∂y ∂y ∂φ ∂ζ ∂φ ∂ζ ∂φ ∂ζ = + κa + ∂z ∂t ∂x ∂x ∂y ∂y gκ ∂φ 1 + κa∇φ2 + 2 η = 0. z = −h. ∂z ∂t ∂φ + gη = 0. z = 0. z = κa η(x. ∂t Combining equations (1. t).
ℓ. 0. ℓ.18) is given by φ(k. The functions A(k. ℓ. y.22) (1.21) √ where m = k 2 + ℓ2 . s) = sζ(k. s) can be easily found from the boundary conditions (1. s) and B(k. s) = m cosh(mh)[s2 + gm tanh(mh)] A(k. cosh(mh)[s2 + gm tanh(mh)] s3 ζ(k. s) . s) = − . 0. equations (1.13). We apply the Fourier transform in (x. 0. dz dφ s2 φ(k. t)] = F (k. dz The transformed freesurface elevation can be obtained from (1. s) . s). +∞ L[g] = g(s) = 0 g(t)e−st dt. ℓ. y) = and the Laplace transform in time t: f (k. B(k. ℓ. the following notation is introduced: +∞ FL[F (x. ℓ) = R2 f (x. s) = 0. s) = − φ(k. t)e−st dt. version 2 .15) and (1. ℓ.18) (1. s) + g (k. s) = A(k. dz 2 dφ (k. (1.21 Jan 2008 For the combined Fourier and Laplace transforms. (1. ℓ)ei(kx+ℓy) dkdℓ. z. −h. y): F[f ] = f (k. ℓ. ℓ.20): gsζ(k. y. s) cosh(mz) + B(k. ℓ.17) become d2 φ − (k 2 + ℓ2 )φ = 0. ℓ.19) (1. tel00194763. ℓ.20) (1. ℓ. s) sinh(mz). g A general solution of equation (1. ℓ. s). After applying the transforms. 1 (2π)2 R2 F−1 [f ] = f (x. s) = R2 e−i(kx+ℓy) dxdy 0 F (x.20 Tsunami generation This problem will be solved by using the method of integral transforms. ℓ. ℓ. ℓ. y)e−i(kx+ℓy) dxdy.16): s η(k.19) and (1.
10.10: Plot of the frequency ω(m) = gm tanh(mh) and its derivatives dω/dm. z. ℓ. s) .5 Freesurface elevation From (1.22) yields φ(k. s) cosh(mh)(s2 + ω 2 ) cosh(mz) − s2 sinh(mz) . s) = − gsζ(k. ℓ. the freesurface elevation becomes η(k. the notation ω= gm tanh(mh) (1. version 2 . ℓ. gm (1. The acceleration due to gravity g and the water depth h have been set equal to 1. s) = s2 ζ(k.5 0 1 2 3 4 5 m 6 7 8 9 10 Figure 1.23) will be used.1. y.5 2 1. ω ′ (m) and ω ′′ (m) are shown in Figure 1.21). s) st e ds dkdℓ. The graphs of ω(m).25) µ−i∞ .21 Jan 2008 −0. B in (1.5 1 0. ℓ. d2 ω/dm2 .5 3 2. s2 + ω 2 (1.5 0 ω’’(m) ω ω(m) ω’(m) tel00194763. t) = (2π)2 R2 ei(kx+ℓy) 1 cosh(mh) 2πi µ+i∞ s2 ζ(k.1. ℓ. Substituting the expressions for the functions A.1 Waves generated by a moving bottom 21 3. From now on. cosh(mh)(s2 + ω 2 ) Inverting the Laplace and Fourier transforms provides the general integral solution 1 η(x.24) 1.
21 Jan 2008 The Fourier transform F[ζ(x. Using an integral representation of Bessel functions [GR00] ﬁnally yields ∞ ζ(k. ℓ. y.25) using the convolution theorem: L[f1 (t) ∗ f2 (t)] = f1 (s)f2 (s). ℓ) = 2π 0 rζ(r)J0 (mr)dr ≡ ζ(m). t) = (2π)2 R2 ei(kx+ℓy) cosh(mh) 0 t (1 − ω sin ωτ )ζ(k. tel00194763. which we now describe in detail. ζ(k. y)] = ζ(k. ℓ) = R2 ζ(r)e −i(kx+ℓy) dxdy = 0 dφ 0 ζ(r)e−ir(k cos φ+ℓ sin φ) rdr = ∞ π 2π ∞ = 0 dφ 0 rζ(r)e −irm cos(φ−ψ) dr = 0 rζ(r)dr 0 (e−irm cos φ + eirm cos φ )dφ. t − τ )dτ. .e. t − τ )dτ dkdℓ.22 Tsunami generation One can evaluate the Laplace integral in (1. m := √ k 2 + ℓ2 . i. One has 2π ∞ ζ(k. t − τ ) dτ 0 +∞ m 0 J0 (mr) dm cosh(mh) 0 (1 − ω sin ωτ )ζ(m. ℓ) = ζ(m). In the following calculation. y) = ζ(r). we use the notation ψ = arctan(ℓ/k). It yields 1 η(x. t) = (2π)2 = 1 2π 2π +∞ dψ 0 meimr cos(φ−ψ) dm cosh(mh) 0 t t (1 − ω sin ωτ )ζ(m. version 2 . r = x2 + y 2 . Assume that the initial solid boundary deformation is axisymmetric: ζ(x. ℓ) of an axisymmetric function is also axisymmetric with respect to transformation parameters. This general solution contains as a special case the solution for an axisymmetric problem. It follows that 1 η(r.
ζ(k. dynamic source models are not easily available. t > t0 . version 2 . 1. Instantaneous: Ti (t) = H(t).27) µ−i∞ Clearly. t < 0. In fact. Using the special representation (1. the error in measurements) do not induce huge modiﬁcations of the wave patterns. y) and we will consider diﬀerent time dependencies T (t) to model the time evolution of the source.26) tel00194763. provides the static seabed deformation ζ0 (x. y)T (t).26) of seabed deformation and prescribed timedependencies. with α > 0. y. 2. 2 4. t) depends continuously on the source ζ(x. t ≥ 0. There are two main reasons for doing this. s2 + ω 2 (1. which was described in the previous section. y. ℓ)ei(kx+ℓy) 1 cosh(mh) 2πi R2 µ+i∞ The typical graphs of Tc (t) and Te (t) are shown in Figure 1. t) := ζ(x. Physically it means that small variations of ζ (in a reasonable space of functions such as L2 ) yield small variations of η. Trigonometric: Tc (t) = H(t − t0 ) + 1 [1 − cos(πt/t0 )]H(t0 − t). Mathematically this problem is said to be wellposed. Okada’s solution. since it means that small modiﬁcations of the ground motion (for example. Tl (t) = t/t0 .27). y. η(x.21 Jan 2008 Mathematically we separate the time dependence from the spatial coordinates. Linear: 0.1. and this property is essential for modelling the physical processes.25) yields 1 η(x. 0 ≤ t ≤ t0 .26) into (1. Four scenarios will be considered: 1. The second reason is more fundamental. Exponential: Te (t) = 0. where H(t) denotes the Heaviside step function. t) = (2π)2 s2 T(s) st e ds dkdℓ. y). Inserting (1. First of all we want to be able to invert analytically the Laplace transform. −αt 1 − e . one can compute analytically the Laplace integral in (1. (1. 3.11.1 Waves generated by a moving bottom 23 The last equation gives the general integral solution of the problem in the case of an axisymmetric seabed deformation. Below we no longer make this assumption since Okada’s solution does not have this property. t < 0. To perform this . In the present study we consider seabed deformations with the following structure: ζ(x.
scenario Instantaneous Exponential Laplace transform L[Ti ] = 1 s α L[Te ] = s(α+s) γ2 (1 + 2s(s2 +γ 2 ) with γ = tπ 0 −st L[Tl ] = 1−e s2 0 t0 Trigonometric L[Tc ] = Linear e−st0 ) Table 1.21 Jan 2008 0 0 0.6 0.6 0.8 T(t) 0.4 0.7.c.11: Typical graphs of Te (t) and Tc (t).2 tel00194763.24 Tsunami generation 1 Te(t) 0. integration. Here we have set α = 6.l (t). we ﬁrst have to compute the Laplace transform of Ti. t0 = 0.1 0.3 0.8 0. version 2 .3.2.4 Tc(t) 0. The results of these computaions are given in Table 1.9 1 Figure 1.5 t 0. .7 0.2 0.3: Laplace transforms of diﬀerent timedepencies for prescribed bottom motion.e.
cosh(mh) ζ(k. Since a priori there is no information on the ﬂow velocities.1 Waves generated by a moving bottom Inserting these formulas into the inverse Laplace integral yields µ+i∞ 25 1 2πi µ−i∞ µ+i∞ est s2 Ti (s) ds = cos ωt. R2 1. ℓ)ei(kx+ℓy) cos ωt dkdℓ. sin ωt − H(t − t0 ) sin ω(t − t0 ) est s2 Tl (s) ds = . ℓ)ei(kx+ℓy) 2(γ 2 − ω 2 ) cosh(mh) ζ(k. dkdℓ. ℓ)ei(kx+ℓy) cosh(mh) R2 −α2 ηe (x. y. t) = 1 (2π)2 (cos ωt − cos γt + H(t − t0 )[cos ω(t − t0 ) + cos γt]) dkdℓ. ℓ)ei(kx+ℓy) cosh(mh) sin ωt − H(t − t0 ) sin ω(t − t0 ) ωt0 ηl (x. 0).1. ηc (x. t) = 1 (2π)2 R2 ζ(k. y. s2 + ω 2 1 2πi µ−i∞ µ+i∞ ω α2 est s2 Te (s) e−αt − cos ωt − sin ωt . y. t) = γ2 (2π)2 R2 ζ(k.21 Jan 2008 1 2πi µ−i∞ The ﬁnal integral formulas for the freesurface elevations with diﬀerent time dependencies are as follows: ηi (x.6 Velocity ﬁeld In some applications it is important to know not only the freesurface elevation but also the velocity ﬁeld in the ﬂuid domain. ds = − 2 2 + ω2 2 s α +ω α γ2 est s2 Tc (s) ds = s2 + ω 2 2(γ 2 − ω 2 ) (cos ωt − cos γt + H(t − t0 )[cos ω(t − t0 ) + cos γt]) . 2 + ω2 s ωt0 1 2πi µ−i∞ µ+i∞ tel00194763. One of the goals of this work is to provide an initial condition for tsunami propagation codes. For the time being.1. . version 2 . y. y. tsunami modelers take initial seabed deformations and translate them directly to the free surface in order to obtain the initial condition η(x. t) = (2π)2 e−αt − cos ωt − ω sin ωt α α2 + ω 2 dkdℓ.
26 Tsunami generation they take a zero velocity ﬁeld as initial condition for the velocity: ∇φ(x. t) = kζ(k. s2 + ω 2 µ−i∞ i − 2 4π R2 kζ(k.21 Jan 2008 u(x. y. ℓ. ul (x. s) = − gsζ(k. version 2 . y. t) = ig 4π 2 R2 kζ(k. y. y. In the linear approximation the value β = 0 corresponds to the free surface while β = −1 corresponds to the bottom. z. z. t) = 4π 2 R2 kζ(k. y. z. Next we introduce some notation.24). In equation (1. For simplicity we only give the velocities along the free surface (β = 0): ui (x. ℓ. y. one ﬁnds the formulas for the time dependencies Ti . ℓ)T(s) cosh(mh)(s2 + ω 2 ) cosh(mz) − s2 sinh(mz) . ℓ)T(s) cosh(mh)(s2 + ω 2 ) s2 sinh(βmh) k. ℓ) cosh(mβh)ei(kx+ℓy) 1 cosh(mh) 2πi µ+i∞ sT(s)est ds dk s2 + ω 2 s3 T(s)est ds dk. ℓ) sinh(mβh)ei(kx+ℓy) 1 m cosh(mh) 2πi µ+i∞ µ−i∞ After a few computations. ℓ)ei(kx+ℓy) α e−αt − cos ωt + sin ωt 2 + ω 2 ) cosh(mh) (α ω R2 ue (x. y. The horizontal gradient (∂/∂x. Te and Tl . s) = −iφ(k. y. t) = igα 4π 2 ig 4t0 π 2 dk. The horizontal velocities are denoted by u. cosh(mh) ω kζ(k. ℓ)ei(kx+ℓy) sin ωt dk. t)z=βh yields u(k. z. ℓ. . Taking the Fourier and Laplace transforms of tel00194763. s)k = i gsζ(k. 0) = 0. The Fourier transform parameters are denoted k = (k. we obtained the Fourier transform of the velocity potential φ(x. ℓ).28) Let us evaluate the velocity ﬁeld at an arbitrary level z = βh with −1 ≤ β ≤ 0. ∂/∂y) is denoted by ∇h . gm (1. gm cosh(βmh) − Inverting the Fourier and Laplace transforms gives the general formula for the horizontal velocities: ig u(x. The vertical velocity component is simply w. t): φ(k. t) = ∇h φ(x. βh. ℓ)ei(kx+ℓy) ω 2 cosh(mh) R2 (1 − cos ωt − H(t − t0 )[1 − cos ω(t − t0 )]) dk. The present computations show that it is indeed a very good approximation if the generation time is short.
1. ℓ)ei(kx+ℓy) ω sin ωt − γ sin γt 2(γ 2 − ω 2 ) cosh(mh) ζ(k. Note that formally the derivative gives the distributions δ(t) and δ(t − t0 ) under the integral sign. t) = − 1 4π 2 R2 3 tel00194763. the expression for w can be simply derived by diﬀerentiating the known formula for ηi. y. y. cosh(mh) +H(t − t0 )[ω sin ω(t − t0 ) + γ sin γt] dk.c. Indeed. Under this assumption we give the distributionfree expressions for the vertical velocity along the free surface: wi (x.l (x. ℓ. t). s) by diﬀerentiating (1. t) = − γ 4π 2 R2 2 ζ(k. ℓ)ei(kx+ℓy) [cos ωt − H(t − t0 ) cos ω(t − t0 )] dk. t). ℓ) i(kx+ℓy) 1 e cosh(mh) 2πi µ−i∞ for −h < z ≤ 0. wl (x. t) = 4π 2 R2 cosh(mz)ζ(k. y.14).21 Jan 2008 ζ(k. s) = sg ζ(k. cosh(mh) ζ(k.e. The easiest way to compute the vertical velocity w along the free surface is to use the boundary condition (1. y. t) = 1 4t0 π 2 R2 .28): w(k. It is easy to obtain the Fourier–Laplace transform w(k. ℓ) i(kx+ℓy) 1 e cosh(mh) 2πi µ−i∞ s3 T(s)est ds dk s2 + ω 2 sT(s)est ds dk.1 Waves generated by a moving bottom 27 Next we determine the vertical component of the velocity w(x. z. g Inverting this transform yields µ+i∞ 1 w(x. y. It is a consequence of the idealized time behaviour (such as the instantaneous scenario) and it is a disadvantage of the Laplace transform method. z. One can easily obtain the expression of the vertical velocity at a given vertical level by substituting z = βh in the expression for w. In order to avoid these distributions we can consider the solutions only for t > 0 and t = t0 . s2 + ω 2 µ+i∞ g − 2 4π R2 m sinh(mz)ζ(k. For small values of ε this will give a very good approximation of the solution behaviour at these “critical” instants of time. From a practical point of view there is no restriction since for any ε > 0 we can set t = ε or t = t0 + ε. ℓ. y. wc (x. ℓ)ei(kx+ℓy) ω sin ωt dk. ℓ)T(s) ∂φ = ∂z cosh(mh)(s2 + ω 2 ) s2 cosh(mz) − m sinh(mz) . version 2 . ℓ)ei(kx+ℓy) (α2 + ω 2 ) cosh(mh) e −αt α we (x. z. y. t) = 4π 2 R2 ω2 ω + 2 cos ωt − sin ωt α α dk. z.
y.1. cosh2 (mh) ω ζ(k. which was written explicitly for the free surface in equation (1.31) The timederivative of the velocity potential is readily available in Fourier space. y. t) for the pressure at the bottom. ℓ)ei(kx+ℓy) [sin ωt − H(t − t0 ) sin ω(t − t0 )] dk. ∂t ρ (1. ℓ)ei(kx+ℓy) −αt e − cos ωt − sin ωt 2 + ω2 α α dk + ω α 3 gα2 ∂φe = ∂t (2π)2 R2 α4 (2π)2 dk. respectively. ∂φi g = − ∂t (2π)2 R2 ζ(k. ℓ) tanh(mh)ei(kx+ℓy) m(α2 + ω 2 ) R2 e−αt + ω α 2 cos ωt + sin ωt g ∂φl = − ∂t t0 (2π)2 R2 ζ(k.1.30) (1.11). z. version 2 . ζ(k. but is valid everywhere in the ﬂuid: ∂φ 1 p + ∇φ2 + gz + = 0.6. it is interesting to provide as well the expression pb (x. . equation (1.1 Pressure on the bottom Tsunami generation Since tsunameters have one component that measures the pressure at the bottom (bottom pressure recorder or simply BPR [GBM+ 05]).29) tel00194763. ℓ)ei(kx+ℓy) cos ωt dk. z=−h Plots of the bottom pressure will be given in Section 1. t) = − ρ ∂φ ∂t − ρgζ.21 Jan 2008 Along the bottom.8.28 1. it reduces to pb ∂φ + g(−h + ζ) + = 0. y. (1.29) becomes ∂φ p + gz + = 0. Inverting the Fourier and Laplace transforms and evaluating the resulting expression at z = −h gives for the four time scenarios. ω cosh2 (mh) The bottom pressure deviation from the hydrostatic pressure is then given by pb (x. The pressure p(x. ∂t 2 ρ After linearization. ∂t ρ z = −h. t) can be obtained from Bernoulli’s equation.
where ω 2 = gm tanh mh.7 Asymptotic analysis of integral solutions tel00194763. Mur92. which .1 Waves generated by a moving bottom 29 1.21 Jan 2008 In this subsection. From equation (1. s2 + ω 2 For example. y). m = k 2 + ℓ2 . Let us ﬁrst obtain an asymptotic representation for integral solutions of the general form √ 1 ζ(k. t) = 4π 2 1 = 4π 2 ∞ 2π ζ(m. In order to be able to resolve these oscillations.27) shows that T (m. the stationary phase method was ﬁrst used by Kelvin [Kel87] in the context of linear waterwave theory. This method seems to be quite eﬃcient. To our knowledge. t) is in fact µ+i∞ 1 T (m.32). ϕ) are the polar coordinates of (x. we switch to polar coordinates m and ψ = arctan(ℓ/k): 1 η(x. In the last expression. several discretization points are needed per period. ℓ)ei(kx+ℓy) η(x. BH86. t) dm cosh(mh) ζ(m. t)m dψdm cosh(mh) 2π 0 0 ∞ mT (m. Stationary phase points satisfy the condition ∂Φ/∂ψ = 0. (1. we do not specify the time behaviour T(s). t) = 2πi µ−i∞ s2 T(s) st e ds.25). y. t) = 2 T (m. PL71. This method is a classical method in asymptotic analysis. The idea of this method consists in interpolating only the amplitude of the integrand at discretization points by some kind of polynomial or spline and then performing exact integration for the oscillating part of the integrand. y. The motivation to obtain asymptotic formulas for integral solutions was mainly due to numerical diﬃculties to calculate the solutions for large values of x and y.1. There is a lot of literature on this topic (see for example [Erd56. version 2 . In equation (1. the phase function is Φ = mr cos(ϕ − ψ).1. t) = cos ωt. 0 0 where (r. ψ)eimr cos(ϕ−ψ) T (m. ψ)eimr cos(ϕ−ψ) dψ. it is clear that the integrand is highly oscillatory.32) 4π cosh(mh) R2 Comparing with equation (1. For the time being. t) dkdℓ. we apply the method of stationary phase in order to estimate the farﬁeld behaviour of the solutions. ES94]). we showed above that for an instantaneous seabed deformation T (m. The numerical method used in the present study is based on a Filontype quadrature formula [Fil28] and has been adapted to double integrals with exp[i(kx + ℓy)] oscillations. This becomes extremely expensive as r = x2 + y 2 → ∞.
ϕ. ϕ)ei( 4 −mr) + ζ(m. . the details of wave formation are not important. t) = where ∞ 1 I1 + I2 . version 2 . ψ) = mr cos(ϕ − ψ) ± ωt.33) t dm √ Since dω/dm decreases from gh to 0 as m goes from 0 to ∞ (see Figure 1. An approximation to equation (1. ψ) cosh(mh) mζ(m. this √ equation has a unique solution for m if r/t ≤ gh.32) is then obtained by applying the method of stationary phase to the integral over ψ: η(r. ϕ + π)ei(mr− 4 ) dm. t) in equation (1. This unique solution will be denoted by m∗ .10). ψ1 = ϕ and ψ2 = ϕ+π.32) yields η(r. t) = cos ωt = eiωt + e−iωt . ∂ψ ∂Φ = 0. t). ω 2 (m) = gm tanh mh.21 Jan 2008 Inserting this particular function T (m. Since we are looking for the far ﬁeld solution behaviour. as before. (1. The second condition yields r dω cos(ϕ − ψ1. ∂m The ﬁrst equation gives two points.2 ) = ∓ . cosh(mh) This expression cannot be simpliﬁed if we do not make any further hypotheses on the function T (m. 0 The stationary phase function in these integrals is Φ(m. φ. Thus we will assume that the initial seabed deformation is instantaneous: T (m. 0 2π I2 = 0 ei(−ωt+mr cos(ϕ−ψ)) dψdm. ψ) cosh(mh) 2π ei(ωt+mr cos(ϕ−ψ)) dψdm. 2 tel00194763. t) ≃ √ 1 8π 3 r ∞√ 0 π π mT (m. The points of stationary phase are then obtained from the conditions ∂Φ = 0. 8π 2 I1 = 0 ∞ mζ(m. t) ζ(m.30 Tsunami generation yields two phases: ψ1 = ϕ and ψ2 = ϕ + π.
1. ϕ + π) √ i(ωt−mr) me dm cosh(mh) 2πm∗ ζ(m∗ . Similarly. Because of the lack of information about the real dynamical characteristics of tsunami sources. It means physically that the wave has not yet reached this region.1. the wave train decays as 1/t. or 1/r.21 Jan 2008 ≈ = 2π t m∗ ζ(m∗ . version 2 . there is no stationary phase.8 Numerical results A lot of numerical computations based on the analytical formulas obtained in the previous sections have been performed. As is easily observed from the expressions for I1 and I2 . ϕ) −i(ω(m∗ )t−m∗ r) m∗ e .34). So we can approximately set I1 ≈ 0 and I2 ≈ 0.1 Waves generated by a moving bottom 31 √ For r > t gh. ϕ + π)ei(ωt−mr) ei 4 mr dm = ζ(m. −ω ′′ ω ′ cosh(m∗ h) Asymptotic values have been obtained for the integrals. Let us obtain an asymptotic formula for the ﬁrst integral: ∞ I1 ≈ 0 m cosh(mh) 2π i π e4 r 2π i π e4 r ∞ π 2π ζ(m. which is equivalent since r and t are connected by relation (1. The exponential dependence has shown higher amplitudes for relatively longer characteristic . From the positivity of the function dω/dm one can deduce that ψ1 = ϕ is a stationary phase point only for the integral I2 . At this stage it is still very diﬃcult or even impossible.34) dm m=m∗ Similarly one can obtain an estimate for the integral I2 : I2 ≈ 2π t ζ(m∗ . ψ2 = ϕ + π is a stationary point only for the integral I1 . (1. ϕ + π) i(ω(m∗ )t−m∗ r) −i π e e 4 ω ′′ (m∗ ) t cosh(m∗ h) 0 tel00194763.33) evaluated at the stationary phase point (m∗ . ϕ + π) i(ω(m∗ )t−m∗ r) e . Numerical experiments showed that the largest wave amplitudes with the time dependence Tc (t) were obtained for relatively small values of the characteristic time t0 . we cannot really conclude which time dependence gives the best description of tsunami generation. 1. −ω ′′ ω ′ cosh(m∗ h) In this estimate we have used equation (1. ψ2 ): dω r=t .
y). For illustration purposes.8 ∼ 0. .27 20 13 90 0 60 40 15 4 9. δ. Numerical computations showed that this situation does not change if one takes other timedependences. d. ◦ Strike angle. we have chosen the instantaneous seabed deformation since it is the most widely used.y)∈R In this section we provide several plots (Figure 1.21 Jan 2008 Table 1.9ζ0 . g. ν Fault depth. Hz Tsunami generation Value 9.4. km Acceleration due to gravity. L. The main focus of the present section is the generation of waves by a moving bottom. where ζ0 = max 2 ζ(x.12) of the freesurface deformation. W . m Water depth. θ. m/s2 Wave number. In particular. The asymptotic behaviour of various sets of initial data propagating in a ﬂuid of uniform depth has been studied in detail by Hammack and Segur [HS74. km Dip angle.94).13). GPa Poisson ratio. km Burger’s vector length. E. h.8 10−4 10−2 tel00194763. The values of the parameters used in the computations are given in Table 1. Finally. ◦ Fault length. The linear time dependence Tl (t) showed a linear growth of wave amplitude from 0 to also ≈ 0. they showed that the behaviours for an initial elevation wave and for an initial depression wave are diﬀerent. ω. k. D. φ. plots of the bottom dynamic pressure are given in Figure 1. km Fault width. version 2 . (x. HS78]. The instantaneous scenario Ti gives at the free surface the initial seabed deformation with a slightly lower amplitude (the factor that we obtained was typically about 0.4: Physical parameters used in the numerical computations times.5 0. 1/m Angular frequency. ◦ Normal angle.32 Parameter Young modulus. The water has a highpass ﬁlter eﬀect on the initial solid boundary deformation. We also give plots of the velocity components on the free surface a few seconds (physical) after the instantaneous deformation (Figure 1.13 it is clear that the velocity ﬁeld is really negligible in the beginning of wave formation.14. From Figure 1.
0. ﬁve minutes and eight minutes and a half after the initial seabed deformation.1.6. 5 in dimensionless time.12: Freesurface elevation at t = 0. .21 Jan 2008 Figure 1.1 Waves generated by a moving bottom 33 tel00194763. version 2 . In physical time it corresponds to one second.01. one minute. 3.
that is one second after the initial seabed deformation. .21 Jan 2008 Figure 1. version 2 .13: Components u.34 Tsunami generation tel00194763. v and w of the velocity ﬁeld computed along the free surface at t = 0.01.
one minute. In physical time it corresponds to one second.14: Bottom pressure at t = 0. 3. 0.6. ﬁve minutes and eight minutes and a half after the initial seabed deformation.1.21 Jan 2008 Figure 1.1 Waves generated by a moving bottom 35 tel00194763. .01. 5 in dimensionless time. version 2 .
etc. it contributed to the tsunami and may explain the surprising tsunami heights. The present work focuses on tsunami generation by earthquakes. the 2004 Indian Ocean tsunami shed some doubts about this belief. [DDK06] used linear theory to show that indeed diﬀerences may exist between the standard passive generation and the active generation that takes into account the dynamics of seaﬂoor displacement. According to Synolakis and Bernard [SB06]. and a model for the formation of surface gravity waves resulting from the deformation of the seaﬂoor. falls of asteroids. For the resulting water motion. The transient wave generation due to the coupling between the seaﬂoor motion and the free surface has been considered by a few authors only. In a recent note. Even though this slip was relatively slow. volcanic eruptions. In this study. For example. 2 As pointed out by Geist et al. one can mention submarine mass failures. In particular. Kajiura analyzed the applicability of the passive approach using Green’s functions. Two steps in modeling are necessary for an accurate description of tsunami generation: a model for the earthquake fed by the various seismic parameters. it is essential to understand the ﬁrst minutes of a tsunami. In the absence of sophisticated source models. this approach is sometimes called the piston model of tsunami generation. Tuck and Hwang [Tuc74] solved the linear longwave equation in the presence of a moving bottom and a uniformly sloping beach. [GTS06]. slides. Hammack [Ham73] generated waves experimentally by raising or lowering a box at one end of a channel. its propagation and ﬁnally the resulting inundation and impact on structures. version 2 . There are diﬀerent natural phenomena that can lead to a tsunami. one often uses analytical solutions based on dislocation theory in an elastic halfspace for the seaﬂoor displacement [Oka85].1 This approach leads to a wellposed initial value problem with zero velocity. The focus of the present work is on the generation process. An open question for tsunami forecasting modelers is the validity of neglecting the initial velocity. We refer to the recent review on tsunami science [SB06] for a complete bibliography on the topic. 1 tel00194763.2 Comparison between threedimensional linear and nonlinear tsunami generation models Tsunami wave modeling is a challenging task. we will call this approach the passive generation approach.2 BenMenahem and Rosenman [BMR72] calculated the twodimensional radiation pattern from a moving source (linear theory). In the tsunami literature. One of the reasons is that it is commonly assumed that the source details are not important. the standard practice is to transfer the inferred seaﬂoor displacement to the free surface of the ocean. The measurements from land based stations that use the Global Positioning System to track ground movements revealed that the fault continued to slip after it stopped releasing seismic energy.36 Tsunami generation 1. Dutykh et al.21 Jan 2008 . Houston and Garcia [HG74] were the ﬁrst to use more geophysically realIn the pioneering paper [Kaj63].
1.2 Comparison of tsunami generation models
37
istic initial conditions. For obvious reasons, the quantitative diﬀerences in the distribution of seaﬂoor displacement due to underwater earthquakes compared with more conventional earthquakes are still poorly known. Villeneuve and Savage [VS93] derived model equations which combine the linear eﬀect of frequency dispersion and the nonlinear eﬀect of amplitude dispersion, and included the eﬀects of a moving bed. Todorovska and Trifunac [TT01] considered the generation of tsunamis by a slowly spreading uplift of the seaﬂoor. In this work, we mostly follow the standard passive generation approach. Several tsunami generation models and numerical methods suited for these models are presented and compared. The focus of our work is on modelling the ﬂuid motion. It is assumed that the seabed deformation satisﬁes all the necessary hypotheses required to apply Okada’s solution. The main objective is to conﬁrm or inﬁrm the lack of importance of nonlinear eﬀects and/or frequency dispersion in tsunami generation. This result may have implications in terms of computational cost. The goal is to optimize the ratio between the complexity of the model and the accuracy of the results. Government agencies need to compute accurately tsunami propagation in real time in order to know where to evacuate people. Therefore any saving in computational time is crucial (see for example the code MOST used by the National Oceanic and Atmospheric Administration in the US [TS98] or the code TUNAMI developed by the Disaster Control Research Center in Japan). Liu and Liggett [LL83] already performed comparisons between linear and nonlinear water waves but their study was restricted to simple bottom deformations, namely the generation of transient waves by an upthrust of a rectangular block, and the nonlinear computations were restricted to twodimensional ﬂows. Bona et al. [BPS81] assessed how well a model equation with weak nonlinearity and dispersion describes the propagation of surface water waves generated at one end of a long channel. In their experiments, they found that the inclusion of a dissipative term was more important than the inclusion of nonlinearity, although the inclusion of nonlinearity was undoubtedly beneﬁcial in describing the observations. The importance of dispersive eﬀects in tsunami propagation is not directly addressed in the present study. Indeed these eﬀects cannot be measured without taking into account the duration (or distance) of tsunami propagation [Tuc79].
tel00194763, version 2  21 Jan 2008
1.2.1
Physical problem description
In the whole chapter, the vertical coordinate is denoted by z, while the two horizontal coordinates are denoted by x and y, respectively. The sea bottom deformation following an underwater earthquake is a complex phenomenon. This is why, for theoretical or experimental studies, researchers have often used simpliﬁed bottom motions such as the vertical motion of a box. In order to determine the deformations of the sea bottom due to an earthquake, we use the analytical solution obtained for a dislocation in an elastic
38 parameter Dip angle δ Fault depth df , km Fault length L, km Fault width W , km Magnitude of Burger’s vector D, m Young’s modulus E, GPa Poisson ratio ν
Tsunami generation value 13◦ 3 6 4 1 9.5 0.23
tel00194763, version 2  21 Jan 2008
Table 1.5: Typical parameter set for the source used to model the seaﬂoor deformation due to an earthquake in the present study. The dip angle, Young’s modulus and Poisson ratio correspond roughly to those of the 2004 Sumatra event. The fault depth, length and width, as well as the magnitude of Burger’s vector, have been reduced for computation purposes.
halfspace [Oka85]. This solution, which at present time is used by the majority of tsunami wave modelers to produce an initial condition for tsunami propagation simulations, provides an explicit expression of the bottom surface deformation that depends on a dozen of source parameters such as the dip angle δ, fault depth df , fault dimensions (length and width), Burger’s vector D, Young’s modulus, Poisson ratio, etc. Some of these parameters are shown in Figure 1.3. More details can be found in [DD07c] for example. A value of 90◦ for the dip angle corresponds to a vertical fault. Varying the fault slip D does not change the coseismic deformation pattern, only its magnitude. The values of the parameters used in the present study are given in Table 1.5. A typical dipslip solution is shown in Figure 1.15 (the angle φ is equal to 0, while the rake angle θ is equal to π/2). Let z = ζ(x, y, t) denote the deformation of the sea bottom. Hammack and Segur [HS74] suggested that there are two main kinds of behaviour for the generated waves depending on whether the net volume V of the initial bottom surface deformation V =
R2
ζ(x, y, 0) dxdy
is positive or not.3 A positive V is achieved for example for a “reverse fault”, i.e. when the dip angle δ satisﬁes 0 ≤ δ ≤ π/2 or −π ≤ δ ≤ −π/2, as shown in Figure 1.16. A negative V is achieved for a “normal fault”, i.e. when the dip angle δ satisﬁes π/2 ≤ δ ≤ π or −π/2 ≤ δ ≤ 0. The conclusions of [HS74] are based on the Korteweg–de Vries (KdV) equation and were
However it should be noted that the analysis of [HS74] is restricted to onedimensional unidirectional waves. We assume here that their conclusions can be extended to twodimensional bidirectional waves.
3
1.2 Comparison of tsunami generation models
39
tel00194763, version 2  21 Jan 2008
Figure 1.15: Typical seaﬂoor deformation due to dipslip faulting. The parameters are those of Table 1.5. The distances along the horizontal axes x and y are expressed in kilometers.
8
6
Normal fault
4
Normal fault
2
0
V
−2
−4
−6
Reverse fault
Reverse fault
−8
−150
−100
−50
0 δ
50
100
150
Figure 1.16: Initial net volume V (in km3 ) of the seaﬂoor displacement as a function of the dip angle δ (in ◦ ). All the other parameters, which are given in Table 1.5, are kept constant.
40
Tsunami generation
tel00194763, version 2  21 Jan 2008
in part conﬁrmed by their experiments. If V is positive, waves of stable form (solitons) evolve and are followed by a dispersive train of oscillatory waves, regardless of the exact structure of ζ(x, y, 0). If V is negative, and if the initial data is nonpositive everywhere, no solitons evolve. But, if V is negative and there is a region of elevation in the initial data (which corresponds to a typical Okada solution for a normal fault), solitons can evolve and we have checked this last result using the FNPF equations (see Figure 1.17). In this study, we focus on the case where V is positive with a dip angle δ equal to 13◦ , according to the seismic data of the 26 December 2004 SumatraAndaman event (see for example [LKA+ 05]). However, the sea bottom deformation often has an N −shape, with subsidence on one side of the fault and uplift on the other side as shown in Figure 1.15. In that case, one may expect the positive V behaviour on one side and the negative V behaviour on the other side. Recall that the experiments of Hammack and Segur [HS74] were performed in the presence of a vertical wall next to the moving bottom and their analysis was based on the unidirectional KdV wave equation. We now consider the ﬂuid domain. A sketch is shown in Figure 1.9. The ﬂuid domain Ω is bounded above by the free surface and below by the rigid ocean ﬂoor. It is unbounded in the horizontal x− and y− directions. So, one can write Ω = R2 × [−h(x, y) + ζ(x, y, t), η(x, y, t)]. Before the earthquake the ﬂuid is assumed to be at rest, thus the free surface and the solid boundary are deﬁned by z = 0 and z = −h(x, y), respectively. For simplicity h(x, y) is assumed to be a constant. Of course, in real situations, this is never the case but for our purpose the bottom bathymetry is not important. Starting at time t = 0, the solid boundary moves in a prescribed manner which is given by z = −h + ζ(x, y, t), t ≥ 0.
The deformation of the sea bottom is assumed to have all the necessary properties needed to compute its Fourier transform in x, y and its Laplace transform in t. The resulting deformation of the free surface z = η(x, y, t) is to be found as part of the solution. It is also assumed that the ﬂuid is incompressible and the ﬂow irrotational. The latter implies the existence of a velocity potential φ(x, y, z, t) which completely describes the ﬂow. By deﬁnition of φ the ﬂuid velocity vector can be expressed as q = ∇φ. Thus, the continuity equation becomes ∇ · q = ∆φ = 0, (x, y, z) ∈ Ω. (1.35)
The potential φ(x, y, z, t) must satisfy the following kinematic boundary conditions on the
1.2 Comparison of tsunami generation models
41
tel00194763, version 2  21 Jan 2008
Figure 1.17: Wave proﬁles at diﬀerent times for the case of a normal fault (δ = 167◦ ). The seaﬂoor deformation occurs instantaneously at t = 0. The water depth h(x, y) is assumed to be constant.
∂t 2 z = η(x. ˜ ˜ t = c0 t/λ. y. Tsunami generation z = −h + ζ(x. y. Diﬀerent choices lead to diﬀerent models.38) (1. In dimensionless form. version 2 . (1. t). Further assuming the ﬂow to be inviscid and neglecting surface tension eﬀects. 2 (1. ∂z ∂t ∂x ∂x ∂y ∂y ∂φ ∂ζ ∂φ ∂ζ ∂φ ∂ζ = + + . z = εη(x. z = εη(x.36) tel00194763. η= . However the choice of the reference lengths and speeds is subtle.39) (1. µ = d/λ. one can write the dynamic condition to be satisﬁed on the free surface as ∂φ 1 + ∇φ2 + gη = 0. t). y. ∂φ ∂x 2 +µ 2 + ∂φ ∂z where two dimensionless numbers have been introduced: ε = a/d. (1.37) ∂φ ∂η = µ2 + εµ2 ∂z ∂t ∂φ ∂ζ = µ2 + εµ2 ∂z ∂t 1 µ + ε µ2 ∂t 2 2 ∂φ ∂φ ∂η ∂φ ∂η + ∂x ∂x ∂y ∂y .42 free surface and the solid boundary.41) . y. t). ˜ z = z/d. y. t). the equations become ∂ 2φ + µ2 ∂z 2 ∂2φ ∂2φ + ∂x2 ∂y 2 = 0. t). (1. . where λ is the horizontal scale of the motion and d a typical water depth.21 Jan 2008 where g is the acceleration due to gravity. The atmospheric pressure has been chosen as reference pressure. The speed c0 is √ the long wave speed based on the depth d (c0 = gd). and after dropping the tildes.40) ∂φ ∂ζ ∂φ ∂ζ + ∂x ∂x ∂y ∂y ∂φ ∂y 2 h z = − + εζ(x. φ= ˜ a a agλ where a is a characteristic wave amplitude. z) ∈ Ω. y. y. t). respectively: ∂η ∂φ ∂η ∂φ ∂η ∂φ = + + . ∂z ∂t ∂x ∂x ∂y ∂y z = η(x. (x. d + µ2 η = 0. ˜ y = y/λ. ζ= . Let the new dependent variables be c0 η ˜ ζ ˜ φ. Let the new independent variables be x = x/λ. The equations are more transparent when written in dimensionless variables.
(x.2 Linear theory First we perform the linearization of the above equations and boundary conditions. 0]. It is equivalent to taking the limit of (1. we now switch back to the physical variables.5 × 10−2 . version 2 . Combining equations (1.37)–(1. Indeed the satellite altimetry observations of the 2004 Boxing Day tsunami waves obtained by two satellites that passed over the Indian Ocean a couple of hours after the rupture process occurred gave an amplitude a of roughly 60 cm in the open ocean. (1. For the sake of convenience.45) (1.40) as ε → 0. ∂t The bottom motion appears in equation (1. The typical wavelength estimated from the width of the segments that experienced slip is between 160 and 240 km [LKA+ 05]. z = 0. Collecting terms of the lowest order in ε yields the linear approximation.43) (1. z = −h. ∂z ∂t ∂φ + gη = 0. The linearized problem can also be obtained by expanding the unknown functions as power series of the small parameter ε.46) (1.1. is diﬃcult to solve because of the nonlinearities in the boundary conditions and the unknown computational domain.5 × 10−4 < ε < 6 × 10−4 . The details of wave motion are completely neglected during the time . (1.47) Most studies of tsunami generation assume that the initial freesurface deformation is equal to the vertical displacement of the ocean bottom and take a zero velocity ﬁeld as initial condition. both numbers ε and µ are small. 1. 2 ∂t ∂z z = 0. The water depth ranges from 4 km towards the west of the rupture to 1 km towards the east.21 Jan 2008 The waterwave problem.44) and yields the single freesurface condition ∂φ ∂2φ +g = 0.2 Comparison of tsunami generation models 43 For the propagation of tsunamis. The linearized problem in dimensional variables reads ∆φ = 0.46) ∂η ∂φ = . Therefore average values for ε and µ in the open ocean are ε ≈ 2 × 10−4 and µ ≈ 2 × 10−2 . y.44) (1.2.42) tel00194763. A more precise range for these two dimensionless numbers is 1. (1.45). z) ∈ R2 × [−h. ∂z ∂t ∂φ ∂ζ = . z = 0. 4 × 10−3 < µ < 2. either in the form of an initial value problem (IVP) or in the form of a boundary value problem (BVP).
50) 1. t) = ζ0 (k. t.44 Tsunami generation that the source operates.45) becomes ∂φ = 0. there are some speciﬁc cases where the time scale and/or the horizontal extent of the bottom deformation may become an important factor. y.3 Active generation This case was extensively studied in Section 1. cosh(mh) (1.21 Jan 2008 1. tel00194763. y. y). [DDK06] showed that even in the case of an instantaneous seaﬂoor deformation.1. ℓ)ei(kx+ℓy) sin ωt dk. version 2 . This was emphasized for example by Todorovska and Trifunac [TT01] and Todorovska et al. Initial velocities are assumed to be zero. ℓ)ei(kx+ℓy) ω sin ωt dk. y. [THT02]. cosh(mh) kζ0 (k. So. we do not repeat the computations here and give only ﬁnal expressions (on the free surface and for t > 0) which correspond to instantaneous bottom motion: 1 (2π)2 R2 η(x. y) is the seaﬂoor deformation. cosh(mh) ω (1. ℓ)ei(kx+ℓy) cos ωt dkdℓ.49) w(x.4 Passive generation In this case equation (1.48) u(x. Recently Dutykh et al. ∂z and the initial condition for η now reads η(x.51) . Our claim is that it is important to make a distinction between two mechanisms of generation: an active mechanism in which the bottom moves according to a given time law and a passive mechanism in which the seaﬂoor deformation is simply translated to the free surface. y. 0) = ig 4π 2 R2 (1. t. 0) = − 1 4π 2 R2 ζ0 (k. z = −h.2. However they did not use realistic source models.2. who considered the generation of tsunamis by a slowly spreading uplift of the seaﬂoor in order to explain some observations related to past tsunamis. While tsunami modelers often justify this assumption by the fact that the earthquake rupture occurs very rapidly. there may be diﬀerences between these two generation processes. where ζ0 (x. 0) = ζ0 (x. (1.
ℓ. From equation (1. t. 0. z. ℓ. ℓ) cos ωt. ω From (1.43). z. t) tanh(mh).21 Jan 2008 φ(k. (1. ℓ. (1.55) √ where m = k 2 + ℓ2 . ℓ) sin ωt cosh(mz) + tanh(mh) sinh(mz) .53): B(k. y). t) = − ζ0 (k.47) become d2 φ − (k 2 + ℓ2 )φ = 0. y.52) (1. ℓ. ℓ. t) = 0.58) ζ0 (k. version 2 .53) (1. t)z=βh .51) and (1. t) and B(k.57) (1. t) = A(k. ℓ.52) is again given by (1.58) gives the Fourier transform of the velocity potential φ(x. (1. t) = 1 (2π)2 R2 (1.2 Comparison of tsunami generation models 45 Again we apply the Fourier transform in the horizontal coordinates (x. z. Taking the Fourier transform of u(x. t) = − ζ0 (k. t) + g (k. Inverting the Fourier transform provides the general integral solution η(x. y. −h. The relationship between the functions A(k. dz ∂2φ ∂φ (k. ℓ. ℓ. ℓ.56) (1. t) = 0.46).1. t) = ζ0 (k.54) tel00194763. ℓ) cos ωt ei(kx+ℓy) dkdℓ. the freesurface elevation becomes η(k. dz 2 dφ (k. Equations (1. t) = A(k.59) Let us now evaluate the velocity ﬁeld in the ﬂuid domain. t) cosh(mz) + B(k. z. t). ℓ. 2 ∂t ∂z A general solution to Laplace’s equation (1. t) can be easily found from the boundary condition (1. ℓ) sin ωt.54) and the initial conditions one ﬁnds g A(k. ℓ. t) sinh(mz).55) yields g φ(k. ℓ. Equation (1. y. β) = ∇h φ(x. y. ℓ. The Laplace transform is not applied since there is no substantial dynamics in the problem. ω Substituting the expressions for the functions A and B in (1. 0.
ℓ. β) = − g 4π 2 R2 m sin ωt ζ0 (k. (1.46 yields Tsunami generation u(k. 0) = − 1 4π 2 R2 ω sin ωt ζ0 (k. ω Along the free surface β = 0. ℓ) sin ωt i(kx+ℓy) e dk. Using the dispersion relation. ω z=βh Inverting this transform yields w(x. cosh(mh) The bottom pressure deviation from the hydrostatic pressure is then given by pb (x. β) = ig 4π 2 R2 kζ0 (k. the horizontal velocity vector becomes u(x. It is easy to obtain the Fourier transform w(k. one has ∂φ ∂t z=−h =− g (2π)2 R2 ζ0 (k. z=−h . ℓ. ℓ)ei(kx+ℓy) dk. t. y. ℓ.61) All the formulas obtained in this section are valid only if the integrals converge. ℓ) sinh(βmh) + tanh(mh) cosh(βmh) . βh. At z = −h. ℓ) sin ωt cosh(βmh) + tanh(mh) sinh(βmh) ei(kx+ℓy) dk. t)k g = i ζ0 (k. t. β) at a given vertical level z = βh. one can write the vertical component of the velocity along the free surface (β = 0) as w(x. ℓ) sin ωt cosh(βmh) + tanh(mh) sinh(βmh) k. t. β) = −iφ(k. y. 0) = ig 4π 2 R2 kζ0 (k. version 2 . ℓ) sinh(βmh)+ ω tanh(mh) cosh(βmh) ei(kx+ℓy) dk for −1 ≤ β ≤ 0. y. t. β) by diﬀerentiating (1.60) tel00194763. y. y. Again.21 Jan 2008 Next we determine the vertical component of the velocity w(x. t. ω Inverting the Fourier transform gives the general formula for the horizontal velocities u(x. t. ℓ)ei(kx+ℓy) cos ωt dk. t) = − ρ ∂φ ∂t − ρgζ. t. ω (1. y. t. one can compute the bottom pressure. β) = ∂φ ∂z = −mg sin ωt ζ0 (k. ℓ.58): w(k.
High frequencies are attenuated in the moving bottom solution. who studied landslide tsunamis. the eﬀects of bottom friction become important and must be included in the codes. which explicitly take into account this oscillatory behaviour.21 Jan 2008 1. one obtains the linear shallow water wave equation.4. so that the seaﬂoor is given by z = −1 + εζ. For simplicity. At the lowest order of approximation.50) with equations (1. The NSW equations are the most commonly used equations for tsunami propagation (see in particular the code MOST developed by the National Oceanic and Atmospheric Administration in the US [TS98] or the code TUNAMI developed by the Disaster Control Research Center in Japan).9. They are also used for generation and runup/inundation. Our analysis will focus on the NSW equations.2.2. which lowpass ﬁlters the source spectrum. we assume below that h is constant. Numerical results will be shown in Section 1. version 2 . The only diﬀerence between pb and ρgη is the presence of an additional cosh(mh) term in the denominator of pb . In the discussion section. The main diﬀerences between passive and active generation processes are that (i) the wave amplitudes and velocities obtained with the instantly moving bottom are lower than those generated by initial translation of the bottom motion and that (ii) the water column plays the role of a lowpass ﬁlter (compare equations (1.2. All integrals were computed with Filon type numerical integration formulas [Fil28].2 Comparison of tsunami generation models 47 Again. In the next level. pb reduces to − ρφt z=−h . For wave runup.59)– (1. away from the deformed seabed. . tel00194763. So the ﬁlter favors long waves. Since there are many diﬀerent ways to go to this level of approximation. 1. also commented on the 1/ cosh(mh) term. The next level of approximation provides the nondispersive nonlinear shallow water equations (NSW). Therefore one can take h as reference depth.1.1 Numerical method for the linear problem All the expressions derived from linear theory are explicit but they must be computed numerically.5 Nonlinear shallow water equations Synolakis and Bernard [SB06] introduced a clear distinction between the various shallowwater models.61)). It is not a trivial task because of the oscillatory behaviour of the integrand functions. Ward [War01]. there are a lot of diﬀerent types of Boussinesq equations. we will come back to the diﬀerences between passive generation and active generation. dispersive terms are added and the resulting equations constitute the Boussinesq equations.48)–(1.
62) µ2 φ0t − µ4 (utx + vty ) + µ2 η + εµ2 (u2 + v 2 ) = 0. (1. one formally expands the potential φ in powers of µ2 : φ = φ0 + µ2 φ1 + µ4 φ2 + · · · .38). partial derivatives are denoted by subscripts. the water is considered to be shallow. When µ2 is a small parameter. if one keeps terms of order εµ2 and µ4 in the dynamic boundary condition (1. Powers of ε will appear when expanding in Taylor series the freesurface conditions around z = 0. t) = Z (∆ux + ∆vy ) + ε ε ∇ζ2 − Z 3 ∆ζ (ux + vy ) 24 2 6 3 ε z − Z 3 ∇ζ · ∇(ux + vy ) − ∆ζt + ε∆(uζx + vζy ) + 3 6 z(−1 + εζ) ε∇ζ · ∇ ζt + ε(uζx + vζy ) − ε2 ∇ζ2 (ux + vy )− 1 (−1 + εζ) ∆ζt + ε∆(uζx + vζy ) .40) and in the kinematic boundary condition (1. t).48 Tsunami generation 1. Thus the vertical velocity component is zero and the horizontal velocity components are independent of the vertical coordinate z at lowest order. The lowestorder term in Laplace’s equation is φ0zz = 0.63) 6 tel00194763.2. y. Solving Laplace’s equation and taking into account the bottom kinematic condition yield the following expressions for φ1 and φ2 : 1 φ1 (x. t) = − Z 2 (ux + vy ) + z [ζt + ε(uζx + vζy )] . For the shallow water theory.6 Mathematical model In this subsection. z. The next step consists in retaining terms of requested order in the freesurface boundary conditions. 2 2 µ2 [ηt + ε(uηx + vηy ) + 1 + ε(η − ζ) (ux + vy ) − ζt − ε(uζx + vζy )] = 1 4 µ (∆ux + ∆vy ). t) and v := φ0y (x. version 2 . This expansion is substituted into the governing equation and the boundary conditions. y. For example. y. y. The boundary conditions imply that φ0 = φ0 (x. y. Let us introduce the notation u := φ0x (x. t). 2 1 4 z2 1 φ2 (x.21 Jan 2008 . 2 where Z = 1 + z − εζ. z. one obtains 1 1 (1.
also known as the Boussinesq equations.70) (1.65) and (1.66) (1.74) (1. ηt + [u(1 + ε(η − ζ))]x + [v(1 + ε(η − ζ))]y = ζt . t) and using the fact that (u. The NSW equations with dispersion (1. t) equal to 1 − εζ(x.66). y. one can rewrite the system of NSW equations as ε ut + (u2 + v 2 )x + ηx = 0.65) Equations (1. 6 (1.73) (1.71) (1.68) (1. ε t The system of equations (1. 2 The kinematic condition (1.75) Kulikov et al. 2 (1. v) is the horizontal gradient of φ0 .64)–(1.72) 1 ηt + [u(h∗ + εη)]x + [v(h∗ + εη)]y = − h∗ . [KML05] have argued that the satellite altimetry observations of the Indian Ocean tsunami show some dispersive eﬀects.64). vt + ε(uvx + vvy ) + ηy = 0. The socalled fundamental NSW equations which contain no dispersive eﬀects are obtained by neglecting the terms of order µ2 : tel00194763.67) (1.21 Jan 2008 ut + ε(uux + vuy ) + ηx = 0. 2 2 1 2 1 ∗ ∗ ηt + [u(h + εη)]x + [v(h + εη)]y − µ (∆ux + ∆vy ) = − h∗ .69) Going back to a bathymetry h∗ (x.66) contain in fact various shallowwater models. can be written in the following form: 1 ε ut + (u2 + v 2 )x + ηx − µ2 ∆ut = 0.(1.62) ﬁrst with respect to x and then with respect to y gives a set of two equations: 1 ut + ε(uux + vvx ) + ηx − µ2 (utxx + vtxy ) = 0. y. However the steepness is so small that the .64) (1. 2 1 vt + ε(uuy + vvy ) + ηy − µ2 (utxy + vtyy ) = 0. 2 2 1 ε 2 2 vt + (u + v )y + ηy − µ2 ∆vt = 0.1. 2 ε vt + (u2 + v 2 )y + ηy = 0. 6 ε t (1.2 Comparison of tsunami generation models 49 Diﬀerentiating (1. Note that this model does not include any bottom friction terms.63) becomes 1 (η − ζ)t + [u(1 + ε(η − ζ))]x + [v(1 + ε(η − ζ))]y = µ2 (∆ux + ∆vy ).72) has been used for example by Titov and Synolakis [TS98] for the numerical computation of tidal wave runup. (1.70)–(1. version 2 .
An accurate code was developed by Grilli et al.35) is transformed into the boundary integral equation α(xl )φ(xl ) = Γ ∂φ ∂G (x)G(x. which are not encountered in the context of tsunami hydrodynamics in deep water. one rewrites the nonlinear kinematic . [GKC96]. that is ∂G/∂n = ∇G · n. However they can be encountered in experiments such as those of Hammack [Ham73] because the parameter µ is much bigger. The boundary Γ is divided into various parts with diﬀerent boundary conditions. xl ) ∂n ∂n dΓ. dispersive eﬀects are necessary only when examining steep gravity waves. The notation ∂G/∂n means the normal derivative. yl .2. zl ) are position vectors for points on the boundary. [GGD01]. The vectors x = (x. The detailed description of this scheme is given in Appendix A. On the free surface. this algorithm reduces the computing complexity from O(N 2 ) to nearly O(N ) up to logarithms. 1. As pointed out in [KS06]. Laplace’s equation (1. 1.50 Tsunami generation origin of these eﬀects is questionable. and α(xl ) = θl /(4π) is a geometric coeﬃcient. a ﬁnitevolume approach is used.76) where G is the threedimensional free space Green’s function. [GHM98] compared Boussinesq and shallowwater models and came to the conclusion that the eﬀects of frequency dispersion are minor. This numerical method satisﬁes the conservative properties at the discrete level. z) and xl = (xl . based on the fast multipole algorithm [FD06].8 Numerical method for the full equations The fully nonlinear potential ﬂow (FNPF) equations (1. By using Green’s second identity. y. version 2 . (1. The eﬃciency of the code was recently greatly improved by introducing a more eﬃcient spatial solver. with θl the exterior solid angle made by the boundary at point xl . which was introduced by Ghidaglia et al.2. where N is the number of nodes on the boundary. It uses a highorder threedimensional boundary element method combined with mixed Eulerian–Lagrangian time updating. For example LeVeque [LeV98] used a highorder ﬁnite volume scheme to solve a system of NSW equations. By replacing every matrix–vector product of the iterative solver and avoiding the building of the inﬂuence matrix. based on secondorder explicit Taylor expansions with adaptive time steps.7 Numerical method tel00194763.37)–(1. Guesmia et al.40) are solved by using a numerical model based on the Boundary Element Method (BEM). xl ) − φ(x) (x.21 Jan 2008 In order to solve the NSW equations. Here the ﬂux scheme we use is the characteristic ﬂux scheme. with n the unit outward normal vector.
They are done only once at each time step. (ii) NSW equations and (iii) fully nonlinear equations. version 2 . Solving the linear system is another time consuming task.77) Dt 1 Dφ = −gz + ∇φ · ∇φ. The matrices are built with the numerical computation of the integrals on the reference element. For the comparisons shown below. The integrals on the boundary are converted into a sum on the elements. The introduction of the fast multipole algorithm reduces considerably the complexity of the problem. (1. Grilli et al.2 Comparison of tsunami generation models and dynamic boundary conditions in a mixed EulerianLagrangian form. tel00194763. The matrix is no longer built. each one being calculated on the reference element. since the same matrix is used for both systems. The boundary is discretized into N collocation nodes and M highorder elements are used to interpolate between these nodes. 51 DR = ∇φ. Within each element. The integral equations are solved by BEM. They included the bottom motion due to the landslide. Even with the GMRES algorithm with preconditioning. ∂ 2 φ/∂t∂n)) are full and non symmetric.2. The material derivative is deﬁned as D ∂ = + q · ∇. Assembling the matrix as well as performing the integrations accurately are time consuming tasks.78) Dt 2 with R the position vector of a freesurface ﬂuid particle. secondorder explicit Taylor series expansions are used to ﬁnd the new position and the potential on the free surface at time t + ∆t. the computational complexity is O(N 2 ). (1. the boundary geometry and the ﬁeld variables are discretized using polynomial shape functions. This time stepping scheme presents the advantage of being explicit. ∂φ/∂n) and one for the pair (∂φ/∂t. Far away nodes are placed in groups. The hierarchical structure involved in the algorithm gives automatically the distance criteria for adaptive integrations.1. The linear systems resulting from the two boundary integral equations (one for the pair (φ. . we only used the passive approach: we did not include the dynamics of the bottom motion. (1.21 Jan 2008 1.79) Dt ∂t For time integration. so less time is spent in numerical integrations and memory requirements are reduced. which is the same as the complexity of the assembling phase. and the use of spatial derivatives along the free surface provides a better stability of the computed solution. [GVW02] used the earlier version of the code to study tsunami generation by underwater landslides.9 Comparisons and discussion The passive generation approach is followed for the numerical comparisons between the three models: (i) linear equations.
In the NSW equations. with a leading wave of depression to the left and a leading wave of elevation to the right.18 shows proﬁles of the freesurface elevation along the main direction of propagation (y−axis) of transient waves generated by a permanent seaﬂoor deformation corresponding to the parameters given in Table 1. dispersive terms µ d (1. The small dimensionless numbers ε and µ2 introduced in (1. as ε → 0 and µ → 0.2. terms of order O(ε2 ) and O(µ2 ) have been neglected. Therefore it is in some sense a worst case scenario.15.75) is that the Ursell number is O(1). Another important aspect of models is the time scale of their validity.5. All the ﬁgures shown below are twodimensional plots for convenience but we recall that all computations for the three models are threedimensional. version 2 . with a corresponding Ursell number equal to 5. Therefore one expects these terms to make an orderone relative contribution on a time scale of order min(ε−2 . the length refers to the local wave shape independent of the exciting conditions.52 Tsunami generation As shown in Section 1. The small dimensionless numbers are roughly ε = 5 × 10−4 and µ = 10−2 . When tsunamis are generated along subduction zones. Figure 1. Hence. has been translated to the free surface. the symbol O(·) is used informally in the way that is common in the construction and formal analysis of model equations for physical phenomena. One can see that the front system splits in two and propagates in both directions. which has been plotted in Figure 1. The Stokes number is deﬁned speciﬁcally for the excitation of a closed basin while the Ursell number is used in a more general context to describe the evolution of a long wave system. We are concerned with the limits ε → 0 and µ → 0. The water depth is 100 m. S = O(1) means that. one moves quickly inland while the second heads toward the open ocean. 4 tel00194763.2 (one can have a look at [DDK06]). We emphasize here that the Ursell number does not convey any information by itself about the separate negligibility of nonlinear and frequency dispersion eﬀects. respectively. The three models are One ﬁnds sometimes in the literature a subtle diﬀerence between the Stokes and Ursell numbers. This deformation.41) represent the magnitude of the nonlinear terms and dispersive terms in the governing equations. in qualitative agreement with the satellite and tide gauge measurements of the 2004 Sumatra event. this generation process gives the largest transientwave amplitudes for a given permanent deformation of the seaﬂoor. Here. Therefore only the characteristic length is diﬀerent. they usually split in two. In the Ursell number. the relative importance of the nonlinear and the dispersive eﬀects is given by the parameter S= nonlinear terms ε aλ2 = 2 = 3 .4 An important assumption in the derivation of the Boussinesq system (1. S takes values that are neither very large nor very small. Thus. µ−2 ).80) which is called the Stokes (or Ursell) number [Urs53].21 Jan 2008 . Both involve a wave amplitude multiplied by the square of a wavelength divided by the cube of a water depth.73)–(1. For the Stokes number the length is the usual wavelength λ related to the √ frequency ω by λ ≈ 2π gd/ω.
18: Comparisons of the freesurface elevation at x = 0 resulting from the integration of the linear equations (· · · ). µ2 = 10−4 and consequently S = 5. t = 191 s). NSW equations (−−) and nonlinear equations (−) at diﬀerent times of the propagation of transient waves generated by an earthquake (t = 0 s. The water depth is h = 100 m. .5. t = 95 s. t = 143 s. The parameters for the earthquake are those given in Table 1.1.21 Jan 2008 Figure 1. One has the following estimates: ε = 5 × 10−4 .2 Comparison of tsunami generation models 53 tel00194763. version 2 .
except for the water depth now equal to 500 m.54 Tsunami generation tel00194763. One can study the evolution of the surface elevation during the generation time at each gauge.20).5. µ2 = 2. The water depth is h = 500 m.04. at least for these particular choices of ε and µ.5 × 10−3 and consequently S = 0. Figure 1.19: Comparisons of the freesurface elevation at x = 0 resulting from the integration of the linear equations (· · · ). version 2 . NSW equations (−−) and nonlinear equations (−) at diﬀerent times of the propagation of transient waves generated by an earthquake (t = 52 s.18. t = 104 s. The linear and nonlinear proﬁles cannot be distinguished within graphical accuracy. The small dimensionless numbers are roughly ε = 10−4 and µ = 5 × 10−2 . almost undistinguishable at all times: the waves propagate with the same speed and the same proﬁle. with a corresponding Ursell number equal to 0. The parameters for the earthquake are those given in Table 1.21 . Figure 1. Only the NSW proﬁle is slightly diﬀerent. One has the following estimates: ε = 10−4 .21 Jan 2008 Figure 1.04. Let us now decrease the Ursell number by increasing the water depth. t = 157 s). Let us introduce several sensors (tide gauges) at selected locations which are representative of the initial deformation of the free surface (see Figure 1. Nonlinear eﬀects and dispersive eﬀects are clearly negligible during the ﬁrst moments of transient waves generated by a moving bottom.19 illustrates the evolution of transient water waves computed with the three models for the same parameters as those of Figure 1.
The lower oval area represents the initial subsidence while the upper oval area represents the initial uplift.20: Top view of the initial free surface deformation showing the location of six selected gauges.5. with the following coordinates (in km): (1) 0.−3 . (5) −2.5 .2 Comparison of tsunami generation models 55 tel00194763. (3) 0. (2) 0. . (6) 1.1. version 2 .21 Jan 2008 Figure 1. (4) 10.10.3 .0 .
56 0. m 0. linear and fully nonlinear models are essentially undistinguishable at all times: the waves propagate with the same speed and the same proﬁle. The physical parameters are those of Figure 1.05 0 0. version 2 .05 −0.05 0 −0. The time t is expressed in seconds.02 −0.1 0 20 40 60 80 100 120 140 160 t.22 and 1.05 t.23 illustrate the evolution of transient water waves computed with the three models for the same parameters as those of Figure 1. m z.04 Tide gauge 3 0.005. but dispersion eﬀects are still small.02 0. m 0 z.04 0 20 40 60 80 100 120 140 160 t.1 Tide gauge 5 Tide gauge 6 0. m 20 40 60 80 100 120 140 160 0 −0.05 0 20 40 60 80 100 120 140 160 t. s 0. with a corresponding Ursell number equal to 0. Let us decrease the Ursell number even further by increasing the water depth.05 0. Since the fully nonlinear results cannot be distinguished from the linear ones.06 0 20 40 60 80 100 120 140 160 z. Figures 1. The small dimensionless numbers are roughly ε = 5 × 10−5 and µ = 0. Again there is a slight diﬀerence between the linear and the NSW models.06 Tide gauge 1 0.04 −0. s t. On the . −− nonlinear shallow water model.21: Transient waves generated by an underwater earthquake. s Figure 1. s 0. at least in this context.20: −.05 −0. except for the water depth now equal to 1 km.02 t. m 20 40 60 80 100 120 140 160 0 0 −0.15 0. They are plotted on the same graph for comparison purposes. s Tide gauge 4 z.1 Tsunami generation Tide gauge 2 z. they are not shown. linear model .02 −0. m 0 −0.18.1. On one hand. Nonlinear eﬀects are clearly negligible during the ﬁrst moments of transient waves generated by a moving bottom.05 0 −0. shows freesurface elevations corresponding to the linear and nonlinear shallow water models.19. s tel00194763.21 Jan 2008 linear NSWE z.04 0. Comparisons of the freesurface elevation as a function of time at the selected gauges shown in Figure 1.
In general it will remain small. And so is the dimensionless number µ which determines the importance of the dispersive eﬀects. We refer to the review [Kir03] on Boussinesq models and their applications for a complete description of modern Boussinesq theory. µ−2 ) and also the neglected eﬀects to make an orderone relative contribution on a time scale of order min(ε−2 . Moreover. Figure 1. This oscillatory behaviour has been observed for the water waves computed with the linear and fully nonlinear models and is due to the presence of frequency dispersion. one should replace the NSW equations with Boussinesq models which combine the two fundamentals eﬀects of nonlinearity and dispersion. The NSW equations.1. version 2 . Dispersive eﬀects come into the picture essentially because the waves are shorter compared to the water depth. [BCL05]. which predicts wave heights. lead to diﬀerent speed and amplitudes. which contain no dispersive eﬀects. dispersive eﬀects do not play a role for long enough waves. So. On the contrary. phase speeds and particle kinematics more accurately than the standard weakly nonlinear approximation ﬁrst derived by Peregrine [Per67] and improved by Nwogu’s modiﬁed Boussinesq model [Nwo93].24 shows the transient waves at the gauges selected in Figure 1. [WKGS95] provided comparisons for twodimensional waves resulting from the integration of a Boussinesq model and the twodimensional version of the FNPF model described above. These results show that one cannot neglect the dispersive eﬀects any longer. From a physical point of view.2 Comparison of tsunami generation models 57 other hand. the nonlinear shallow water model leads to a higher speed and the diﬀerence is obvious for the points away from the generation zone. Wei et al. the numerical solution obtained with the NSW model gives slightly diﬀerent results.21 Jan 2008 . the results shown in this study are in agreement with these estimates. one can expect the solutions to the long wave models to be good approximations of the solutions to the full waterwave equations on a time scale of the order min(ε−1 . ε−1 µ−2 ).20. we emphasize that the wavelength of the tsunami waves is directly related to the mechanism of generation and to the dimensions of the source event. tel00194763. except of course when the tsunami waves approached the coast. the oscillatory behaviour just behind the two front waves is no longer present. As shown in the previous examples. In fact they used a fully nonlinear variant of the Boussinesq model. Considering the 2004 Boxing Day tsunami. Waves computed with this model do not propagate with the same speed and have diﬀerent amplitudes compared to those obtained with the linear and fully nonlinear models. it is clear that dispersive and nonlinear eﬀects did not have suﬃcient time to develop during the ﬁrst hours due to the extreme smallness of ε and µ2 . Even though we have not computed precisely the constant in front of these estimates. Adapting the discussion by Bona et al. One can see that the elevations obtained with the linear and fully nonlinear models are very close within graphical accuracy. µ−4 .
22: Comparisons of the freesurface elevation at x = 0 resulting from the integration of the linear equations (− · −).005. version 2 .5. The water depth is 1 km. One has the following estimates: ε = 5 × 10−5 . The parameters for the earthquake are those given in Table 1. . NSW equations (−−) and FNPF equations (−) at diﬀerent times of the propagation of transient waves generated by an earthquake (t = 50 s. µ2 = 10−2 and consequently S = 0. t = 100 s).21 Jan 2008 Figure 1.58 Tsunami generation tel00194763.
t = 200 s).22 for later times (t = 150 s. .1. version 2 .21 Jan 2008 Figure 1.2 Comparison of tsunami generation models 59 tel00194763.23: Same as Figure 1.
s t. m 0.05 −0. Comparisons of the freesurface elevation as a function of time at the selected gauges shown in Figure 1. s 0. s Figure 1.04 −0.05 Tide gauge 5 0. .21 Jan 2008 0.02 −0.02 −0.23.15 0. m 0 z.05 z.06 0 20 40 60 80 100 120 140 160 z.05 0 −0. −− nonlinear shallow water model.02 Tide gauge 1 0. s 0. m 0 −0.1 Tide gauge 2 z. s Tide gauge 6 0. The time t is expressed in seconds.04 0.05 0 −0.60 Tsunami generation 0.20: −. s t.06 0.1 t.24: Transient waves generated by an underwater earthquake.1 0 20 40 60 80 100 120 140 160 t. m z.22 and 1.04 Tide gauge 3 Tide gauge 4 0.05 0 20 40 60 80 100 120 140 160 t. linear model . The FNPF results cannot be distinguished from the linear results.05 z.04 0 20 40 60 80 100 120 140 160 t. m 20 40 60 80 100 120 140 160 0 −0.02 tel00194763. m 20 40 60 80 100 120 140 160 0 0 −0. The physical parameters are those of Figures 1. version 2 .05 0 −0.
All times must be multiplied √ by a factor 1000.06 z. Results are presented for waves computed with the linear model.04 0. Both scenarios give roughly the same wave proﬁles. the water depth is h = 500 m. m 0 −0.04 0.2 Comparison of tsunami generation models 61 0. and the water depth is h = 500 m.03 Tide gauge 4 0.1 0. m 0 −0.02 −0.25: Transient waves generated by an underwater earthquake.1.26.06 0.08 0.06 z.01 −0.03 Tide gauge 3 tel00194763.02 −0.04 −0.04 0. The computations are based on linear wave theory. As shown in Figure 1. The time t is expressed in seconds. the two generation models diﬀer.05 0 −0.02 0. Let us now consider a slightly diﬀerent set of parameters: the only diﬀerence is the water depth which is now h = 1 km.08 0. The solid line represents the solution with an instantaneous bottom deformation while the dashed line represents the passive wave generation scenario.25 shows the waves measured at several artiﬁcial gauges.2 Tide gauge at (2. In particular. The physical parameters are those of Figure 1. a mistake was introduced in the time scale.06 0.01 0 −0.3) Tide gauge 1 0. Comparisons of the freesurface elevation as a function of time at selected gauges for active and passive generation processes. m Active generation Passive generation 0. m 0.04 −0. 5 . Figure 1. The passive mechanism gives higher wave amplitudes.1 0 20 40 60 80 100 time (s) 120 140 160 180 200 0 20 40 60 80 100 time (s) 120 140 160 180 200 0.02 z. Recall that the classical approach consists in translating the sea bed deformation to the free surface and letting it propagate.25 0.05 −0. We show the major diﬀerences between the classical passive approach and the active approach of wave generation by a moving bottom.19. version 2 .5.15 0. The parameters are those of Table 1. which extends the results given in [DDK06]5 .02 −0.21 Jan 2008 0 20 40 60 80 100 time (s) 120 140 160 180 200 0 20 40 60 80 100 time (s) 120 140 160 180 200 Figure 1. Let us quantify this diﬀerence by considering the relative diﬀerence between the two In Figures 1 and 2 of [DDK06]. Let us conclude this section with a discussion on the generation methods.02 z.
59) of Section 1.04 0.01 −0.01 0.08 0. First of all.1 −0.04 −0.28). y. ηactive ∞ Intuitively this quantity represents the deviation of the passive solution from the active one with a moving bottom in units of the maximum amplitude of ηactive (x.05 0 0 −0. The diﬀerences can be easily explained by looking at the analytical formulas (1. The second feature is more subtle. t). . Results are presented on Figures (1.06 0 20 40 60 80 100 time (s) 120 140 160 180 200 0 tel00194763. in the framework of the NSW equations. These diﬀerences.48) and (1.06 −0.06 Tide gauge 3 0.62 Tsunami generation 0.26: Same as Figure 1.06 0. m z.1 0. mechanisms deﬁned by r(x.03 0. except for the water depth. which can be crucial for accurate tsunami modelling.02 −0. m Active generation Passive generation 0.02 −0. version 2 . they will be attenuated in the moving bottom solution because of the presence of the hyperbolic cosine cosh(mh) in the denominator which grows exponentially with m.27) and (1.03 0 20 40 60 80 100 time (s) 120 140 160 180 200 Figure 1. The numerical experiments show that this diﬀerence is about 6% in the ﬁrst case and 20% in the second case. which is equal to 1 km. In other words. the wave amplitudes obtained with the instantly moving bottom are lower than those generated by the passive approach (this statement follows from the inequality cosh mh ≥ 1). m 0 −0.2 Tide gauge at (2.02 z. t) .02 −0.04 Tide gauge 1 0. Tuc74]. are twofold. there is no diﬀerence between the passive and the active approach for an instantaneous seabed deformation [Tuc79.05 −0. m z.02 z.02 Tide gauge 4 0.08 0. Incidently.15 0 20 40 60 80 100 time (s) 120 140 160 180 200 0 20 40 60 80 100 time (s) 120 140 160 180 200 0.2.15 0. y. y. 3) 0. t) = ηactive (x. if the initial deformation contains high frequencies.25. y. The water column has an eﬀect of a lowpass ﬁlter.2.25 0. t) − ηpassive (x.04 −0.21 Jan 2008 −0.
1 /η 0 η 0 50 100 time (s) 150 200 0.1.2 Comparison of tsunami generation models 63 0.02 0.04 0.04 0.05 ηactive − ηpassive/ηactive 0.05 Tide gauge 3  Tide gauge 4 0.2 Tide gauge at (2.05 0 0 50 100 time (s) 150 200 0. .01 0 0 50 100 time (s) 150 200 active 0.07 Tide gauge 1 ηactive − ηpassive/ηactive 0.3) 0.3) 0.25.03 0.25 0.15 0.3  active active 0.2 Tide gauge 3  active 0.02 0.06 0.01 0 tel00194763.2 Tide gauge 4 active passive active 0.02 0.03 0.04 0.1 η active 0.25  Tide gauge 1 0. The time t is expressed in seconds.01 0 active 0.15 passive /η 0. version 2 .05 0 0.06 0.05 Tide gauge at (2.03 0.08 0.27: Relative diﬀerence between the two solutions shown in Figure 1.2 0.05 0.1 −η active −η η 0. 0.01 0  η /η active −η passive 0 50 100 time (s) 150 200 0.26.05 0 0 50 100 time (s) 150 200 Figure 1.04 0.1 0.05 0.28: Relative diﬀerence between the two solutions shown in Figure 1.03 0.02 0.06 0.15  0.15 /η passive −η −η η 0 50 100 time (s) 150 200 passive active /η 0.21 Jan 2008 0 50 100 time (s) 150 200 η active −η passive /η 0 50 100 time (s) 150 200 Figure 1.35 0.06 0.
in the framework of linear water wave equations. height and location as fast as possible must deﬁnitely be taken into account (in other words. There are two main conclusions that are of great importance for modelling the ﬁrst instants of a tsunami and for providing an eﬃcient initial condition to propagation models. version 2 . a very good agreement is observed from the superposition of plots of wave proﬁles computed with the linear and fully nonlinear models. This is why the process of tsunami generation must be modelled as accurately as possible. Even though the constraint of being able to predict tsunami arrival time. the linearized solution tends to the instantaneous wave generation scenario. 1. the results will depend on the characteristic time of the seabed deformation.2. obviously depends on the construction of the initial condition.10 Conclusions tel00194763.64 Tsunami generation If we prescribe a more realistic bottom motion as in [DD07c] for example. So. Todorovska and Trifunac [TT01] have shown that ampliﬁcation of one order of magnitude may occur when the sea ﬂoor uplift spreads with velocity similar to the long wave tsunami velocity. Comparisons between the NSW equations and the FNPF equations for modeling tsunami runup are left for future work.21 Jan 2008 Comparisons between linear and nonlinear models for tsunami generation by an underwater earthquake have been presented. than the waves considered in the present work. However. compared to the water depth. from generation to inundation. a tradeoﬀ must be found between the precision and the speed of computation of . 1.3. the nonlinear shallow water model was not suﬃcient to model some of the waves generated by a moving bottom because of the presence of frequency dispersion. so that the NSW model is also suﬃcient to describe tsunami generation by a moving bottom. Secondly.1 The accuracy of the computation of the whole life of a tsunami. To begin with. However classical tsunami waves are much longer. the error made on the initial condition cannot be corrected by the numerical method used to propagate the tsunami. during slow events. Another aspect which deserves attention is the consideration of Earth rotation and the derivation of Boussinesq models in spherical coordinates. Moreover. one cannot exceed the passive generation amplitude with an active process.3 Tsunami generation by dynamic displacement of sea bed due to dipslip faulting Introduction 1. When the characteristic time of the bottom motion decreases.
Several studies have already been performed to understand wave formation due to diﬀerent prescribed bottom motions. When simplifying hypotheses such as homogeneity or isotropy are removed. Hammack [Ham73] generated waves experimentally by raising or lowering a box at one end of a channel.3 Tsunami generation by dynamic displacement of sea bed 65 the initial condition). We will refer to this method as passive approach. one neglects the rupture velocity and the rise time. The studies of Hammack [Ham73] and Dutykh & Dias [DD07c] take into account the rise time. tel00194763. DDK06]. and considered various laws for the rise or the fall of the box. After the pioneer work of Kajiura [Kaj70] it has become a common practice in the tsunami community to translate the static sea bed deformation generated by an underwater earthquake onto the free surface and let it propagate. analytical solutions are no longer available and the governing equations must be solved numerically. In particular. Threedimensional analytical expressions derived from Volterra’s formula applied to the general study of dislocations [MS71. When one uses as initial condition a static seismic source together with the translation of the sea bed deformation onto the free surface. In their review paper.21 Jan 2008 . version 2 . Similar analytical expressions for twodimensional problems were also derived by Freund & Barnett [FB76]. Moreover the consequences for the resulting tsunami waves were not pointed out. their computation is easy and inexpensive. or by introducing some type of rupture velocity. Obviously. However neither in [Mas03] nor in [MBS05] were the dynamical aspects and the coupling with hydrodynamics considered. either by introducing some type of rise time. [MBS05] also used the FEM to compare numerical results with analytical solutions. Anisotropy and heterogeneity turned out to be the most important factors in this type of modelling. Dutykh & Dias [DD07c] generated waves theoretically by multiplying the static deformations caused by slip along a fault by various time laws: instantaneous. The validity of this technique was already discussed in [OTM01. Megna et al. who used the theory of analytic functions of a complex variable.1. who used several dislocation models based on the ﬁniteelement method (FEM) to estimate the importance of diﬀerent physical hypotheses. Todorovska & Trifunac [TT01] studied the generation of waves by a slowly spreading uplift of the bottom. simply by imposing the socalled smooth closure condition on the slip: the slip is zero at the ends. A nice feature of the solution of Freund & Barnett is that nonuniform slip distributions can be easily considered. Oka85] are used to construct the static initial deformation. the popularity of these analytic solutions can be explained by their relatively simple explicit form. For example. Static deformations caused by slip along a fault have been extensively simulated by Masterlark [Mas03]. Thus. slip distributions which remove the singular behavior of the internal stresses at the ends of the slip zone can be dealt with. we believe that so far the scientiﬁc community has not payed enough attention to the crucial subject of tsunami generation.
[OTM01] also considered rise time and rupture velocity but unfortunately Let b(x. They found that the eﬀects of rupture velocity are much smaller than those of rise time when the rise time is assumed to be long (over 10 min). y. where T is the rise time and V the rupture velocity. and the vertical coordinate z are denoted by x = (x. The two horizontal coordinates x and y. depth d and dip angle δ. In fact he considered both rise time and rupture velocity.81) implies that at t = 0 a fracture front is established instantaneously over a width W of the y−axis at depth d. t) denote the fault displacement function and b0 (x) the ﬁnal displacement. t) was considered by Haskell: 0 t − ζ/V < 0 b(x. Ohmachi et al.29: Geometry of the source model. Haskell [Has69] was one of the ﬁrst to take into account the rupture velocity. The coordinate ζ is a coordinate along the fault. More recently. Equation (1. Let us consider the source shown in Figure 1. linear. z). The fault has width W . length L.21 Jan 2008 W δ −L 2 −d L 2 Figure 1. t) = (1.29. Okumura & Kawata [OK07] used Haskell’s approach to investigate the eﬀects of rise time and rupture velocity on tsunami generation. The following form for b(x. The front propagates unilaterally at constant velocity V over a length L cos δ of the x−axis. version 2 . z x O y tel00194763. They considered two cases of sea bottom motion: (i) with only rise time and (ii) with both rise time and rupture velocity. trigonometric.81) (b /T )(t − ζ/V ) 0 < t − ζ/V < T 0 b0 t − ζ/V > T . At any ﬁxed point on the fault plane the relative displacement increases at constant velocity from 0 at t = ζ/V to a constant ﬁnal value b0 at t = T + ζ/V .66 Tsunami generation exponential.
The numerical method for the solid motion is validated in Section 1. uy and uz .5. In the earth domain. DD07c]. For isotropic media it means that the Lam´ coeﬃcients e . Unfortunately Madariaga does not provide any plots of the deformation of the free surface and does not consider the coupling with the water layer. He considers a dipslip dislocation propagating in a halfspace. In Section 1.4. 1. Section 1.1. The displacements are denoted by ux . Another interesting attempt to understand dynamical eﬀects is that of Madariaga [Mad03].3. The present study can be considered as an attempt to understand the coupling between seismic faulting and hydrodynamics by integrating numerically the timedependent elasticity equations as well as the timedependent ﬂuid equations. The horizontal coordinates are denoted by x and y. In the ﬂuid domain.3 Tsunami generation by dynamic displacement of sea bed 67 the dynamics is not clearly explained in their paper. version 2 .3. tel00194763. We choose the KelvinVoigt viscosity model [Qui97] which consists in using complex elastic coeﬃcients (with negative imaginary parts in order to dissipate wave energy).21 Jan 2008 The work is organised as follows. We reveal numerically that the dynamical aspects of tsunami generation can for example lead to a leading depression wave when one expects a leading elevation wave.1 Dynamic fault model We assume that the fault is inside a geological viscous medium.3. z = 0 denotes the sea bed at rest (assumed to be ﬂat). Apparently they did not solve the elastodynamic equations with the secondorder time derivative (see next section). z = 0 denotes the sea surface at rest.2. while the vertical coordinate is denoted by z.3 contains details on the numerical methods used to solve the governing equations. the modeling is performed for threedimensional problems.2 Mathematical models Even though the numerical results shown in this study are for twodimensional conﬁgurations.2 we brieﬂy describe the mathematical models.3. provides a comparison between the traditional approach to tsunami generation (in which the static sea bed deformation is translated onto the free surface) and the more realistic approach of dynamic generation (in which the wavetrain is generated by the motion of the bottom).3. He solves the elastodynamic equations by using the double Laplace transform. 1. The main feature of the present study is the use of a more realistic earthquake source model. Earth’s crust is assumed to be a viscoelastic material of density ρ. The authors have already adressed the problem of tsunami generation in [DDK06. The last Section 1. We use diﬀerent origins along the vertical axis for the solid and ﬂuid motions. both for solid and ﬂuid motions. which are used in the present study.3. The solution is elegant but it is relatively complex.
µi The factors QP and QS measure the viscosity of the geological medium. uy . which e is responsible for small wave damping. Let cP and cS be the classical velocities for the propagation of P and S waves in a medium of density ρ: λr + 2µr µr . we come to the following linear viscoelastodynamic problem6 : ∂ 2u ∇ · λ (∇ · u)I + µ (∇u + ∇ u) = ρ 2 . ∂t ∗ ∗ t 6 (1.21 Jan 2008 c∗ = cP P 1+ i . QP c∗ = cS S 1+ i . λi + 2µi QS = − µr . e tel00194763. µr > 0 and λi . cS = . Thus. But we keep it for the sake of completeness. µi ≥ 0. . µ∗ = µr − iµi . version 2 . Let σ represent the stress tensor. t) = (ux . uz ) satisﬁes the classical elastodynamic equations from continuum mechanics: ∇·σ =ρ law through the strain tensor ε = 1 2 ∂ 2u . Mathematically it means that 1/QP ≪ 1 and 1/QS ≪ 1. QS where the coeﬃcients QP and QS are deﬁned as follows: QP = − λr + 2µr . y. cP = ρ ρ Complex Lam´ coeﬃcients yield complex velocities for wave propagation. z.84) We use the preﬁx “visco” due to the presence of the imaginary part in the Lam´ coeﬃcients. elasticity is suﬃcient and the assumption of a KelvinVoigt viscous material is unnecessary. The displacement ﬁeld u(x.68 have a nonpositive imaginary part: λ∗ = λr − iλi .82) It is common in seismology to assume that the stress tensor σ is determined by Hooke’s ∇u + ∇t u . Tsunami generation where λr .83) σ = λ∗ (∇ · u)I + 2µ∗ ε. The classical elasticity equations are obtained by choosing λi ≡ 0 and µi ≡ 0. ∂t2 (1. In this study we restrict our attention to the weakly viscous case. Notice that on the timescales relevant to our problem. Therefore (1.
However we will assume that they are constant in the numerical applications.e. The NSWE in dimensional form read ∂η ∂h + ∇ · ((h + η)v) = − . C (1. The validity of the NSWE model and the question of the importance of dispersive eﬀects have already been addressed by the authors in the previous Section (one can have a look at our publication [KDD07]). ∂t 2 The eﬀect of the moving bottom appears in the source term −∂h/∂t in the ﬁrst equation. The displacement vector is increased by the amount of the Burgers vector b along any contour C enclosing the dislocation surface (or line).3 Tsunami generation by dynamic displacement of sea bed 69 Recall that the mechanical characteristics ρ. it is natural to choose the nonlinear shallow water equations (NSWE) as hydrodynamic model. Since the NSWE are essentially obtained from depthaveraging the Euler equations. y. g the acceleration due to gravity and z = −h(x. . especially in codes for operational use [TS98.3. The unknowns η and v are functions of time and of the horizontal coordinates x and y. the dependence on the vertical coordinate z disappears from the equations. y. t) the bathymetry.21 Jan 2008 We let a dislocation run at speed V along a fault inclined at an angle δ with respect to the horizontal. i. Rupture starts at position x = 0 and z = −d (it is supposed to be inﬁnitely long in the transverse y−direction) and propagates with constant rupture speed V for a ﬁnite time L/V in the direction δ stopping at a distance L. A dislocation is considered as a surface (in threedimensional problems) or a line (in twodimensional problems) in a continuous medium where the displacement ﬁeld is discontinuous. du = b. Let η denote the freesurface elevation with respect to the still water level z = 0.2 Fluid layer model Since the main purpose is to model tsunami generation processes and since tsunamis are long waves. The fault is modeled as a dislocation inside a viscoelastic material. These equations are widely used in tsunami modelling. This type of model is widely used for the interpretation of seismic motion. 1. The rise time is assumed to be 0. On the fault located in the interval 0 < ζ < L slip is assumed to be constant.2. λ∗ and µ∗ can possibly depend on the spatial coordinates (x. vy ) the horizontal velocity vector. v = (vx .1. z). ∂t ∂t ∂v 1 + ∇v2 + g∇η = 0. version 2 .85) tel00194763. SB06]. Let ζ be a coordinate along the dislocation line.
the displacement ﬁeld is discontinuous and satisﬁes the following relation: u+ (x. The propagation of Burger’s vector along the fault is given by b(x. t) = b0 H(t − ζ/V ).87) where V is the rupture velocity. respectively. t) = b(x. H the Heaviside unit step function and ζ a coordinate along the dislocation line. Across the fault. We apply the P2 ﬁniteelement discretization procedure. while for the hydrodynamic part we take advantage of the hyperbolic structure of the governing equations and use a solver based on the ﬁnitevolume (FV) scheme (see for example [BQ06. we assume that the sea bed is a free surface. we take a computational domain which is suﬃciently large. (1.70 Tsunami generation The coupling between the earth and ﬂuid models is made through the function h(x. For the numerical computations.21 Jan 2008 where the signs ± denote the upper and lower boundary of the dislocation surface. that is σ · n = 0 at z = 0. the freely available code FreeFem++ [HPHO] is used.3 Numerical methods In the present study we made two natural choices.3. Dirichlet type boundary conditions u = 0 are applied. The other boundaries are assumed to be ﬁxed or. . element sizes are considerably bigger than in the fault vicinity. t). Next we discuss the implementation of the dislocation surface. y.3. t) which describes the moving sea bottom bathymetry. (1. 1.84) in variational form. This approach is not computationally expensive since we use adaptive mesh algorithms [HPHO] and in the regions far away from the fault. The timederivative operator is discretized through a classical secondorder ﬁnitediﬀerence scheme.1 Discretization of the viscoelastodynamic equations In order to apply the FEM one ﬁrst rewrites the governing equation (1. in other words. In order to avoid the reﬂection of seismic waves along the boundaries during the simulation time. 1. version 2 . Let us say a few words about the boundary conditions and the treatment of the dislocation in the program.86) tel00194763. In such problems implicit schemes become advantageous since the velocity of propagation of seismic waves is of the order of 3 to 4 km/s.3. KDD07]). The solid mechanics equations of the model are solved using the FEM with fully implicit time integration. The authors are aware of the reﬂective properties of this type of boundary conditions. t) − u− (x. The method we use is fully implicit and has the advantage of being free of any CourantFriedrichsLewytype condition. Concerning the boundary conditions.
This property together with the use of an approximate Riemann solver makes all the computations very eﬃcient. version 2 . There are other schemes that have been proposed for the NSWE (see [BQ06] for example). In any case this situation does not occur in nature. And ﬁnally. GKC01] and described in details in Appendix A.3. Mathematically it means that the vector b is tangent to the dislocation surface at each point.1.3. In this section we consider an application of the characteristic ﬂuxes scheme to NSWE. the characteristic ﬂux is more versatile than Roe’s scheme in the sense that it does not rely on an algebraic property of the ﬂux. This approach is a particular scheme of FV methods. The ﬂux of the NSWE system is given by (h + η)vx (h + η)vy 1 F(V) = 2 v2 + gη (1. In the present study the physical variables coincide with the conservative ones. . Then it is not based on the solution to the Riemann problem and consequently we do not impose a onedimensional wave structure which is no longer relevant in two dimensions. First of all. In this case one can explicitly ﬁnd eigenvalues. as well as left and right eigenvectors. v) is the vector of conservative variables7 . the two sides of the fault cannot detach physically. 1 2 0 v + gη 2 (1.21 Jan 2008 where V = (η. It is a multidimensional scheme and we do not need to split operators in order to treat separately each dimension. u. 1 v2 + gη n 2 It follows that the Jacobian matrix is given by An := ∂ F(V) · n = ∂V v · n (h + η)n gn v⊗n . In this study we adopt the characteristic ﬂux (CF) approach proposed in [GKC96.88) 0 . Let v = (vx .3 Tsunami generation by dynamic displacement of sea bed 71 Remark 1 Due to the presence of huge hydrostatic pressures in the crust. vy ). The ﬂux F(V) projected on the normal n becomes (h + η)(v · n) F·n= .3 = v · n ± cs . the CF scheme is easy to implement.2 Finitevolume scheme tel00194763. 1.89) It has three distinct eigenvalues λ1. We chose this method for the following reasons. 7 λ2 = 0.
Use is made of the wellknown result described for example by Freund & Barnett [FB76] or Okada [Oka85]. .91) x d(d cos δ − x sin δ) x 1 sin δ arctan − . It can be expressed as the diﬀerence between two contributions. The matrix of left eigenvectors can also be computed analytically through R−1 .6. In order to simplify the expressions. The sketch of the domain is given in Figure 1.72 where cs = g(h + η) is the velocity of long gravity waves. In fact the most appropriate expression is that given by equation (24) in [Mad03].92) are given in Figure 1. The fault has inﬁnite width (W → ∞). Its length is L. The values of the other parameters are given in Table 1.3. and the numerical scheme is thus completely deﬁned. We checked that it is in full agreement with the limit of Okada’s solution as the width becomes inﬁnite. that from a ﬁrst dislocation located at the beginning of the fault and that from a second dislocation located at the end of the fault. s gny −nx c2 + vx (v · n) gny s Tsunami generation (1. we only consider the twodimensional case (in other words. version 2 .4 Validation of the numerical method In this section we consider an analytic solution to the line dislocation problem in the static case. (1. Good qualitative . since it plays the most important role in tsunami formation. δ − Uz d x − L cos δ .δ dL . The results of the comparison with solution (1. This problem was solved by FEM after neglecting the dynamic terms.92) For the validation of our numerical method we chose the most violent fault which corresponds to the dip angle δ = π/2. The right eigenvectors constitute the columns of the matrix R: −cs (c2 /g)v · t cs s R = gnx ny c2 − vy (v · n) gnx . One has uz = b0  Uz where Uz x . In the present work we only give the vertical displacement component uz along the free surface.90) where t is the tangent vector.29. tel00194763.δ = d π d x2 + d2 (1.30. Let dL = d − L sin δ.21 Jan 2008 1. Using these expressions it is straightforward to compute the sign matrix deﬁned as sign(An ) := R sign(Λ)R−1 . the fault is inﬁnite in the y−direction). its depth d and its dip angle δ.
In their conclusion. Then the bottom conﬁguration is assumed to remain frozen during the rest of the simulation. [MBS05]. The passive generation approach was introduced in [Kaj70]. version 2 . compared the twodimensional FEM results with the analytical solution in the case of a normal fault.21 Jan 2008 Figure 1.3 Tsunami generation by dynamic displacement of sea bed 73 tel00194763. The static sea bed deformation obtained with the analytical solution is depicted in Figure 1.5 Results of the simulation In this section. we consider the Heavisidetype approach (1. We proceed by computing the ﬁrst eight or ﬁfteen seconds of the earthquake dynamics. 1. the active approach uses the bottom motion for wave generation.3. and quantitative agreement can be seen. One could use instead a dislocation for which Burger’s vector b0 is also spacedependent. we use the set of physical parameters given in Table 1. It consists in translating the static sea bed deformation onto the free surface and letting it propagate under gravity. Concerning the dynamical aspects of rupture propagation. Megna et al. On the other hand.31.87) where the dislocation propagates along the fault with rupture velocity V .31 is the value of the dip angle.6. Note that the only diﬀerence between Figures 1. But the main goal of the present study .30: Comparison between analytical and numerical solutions for a static twodimensional fault with a dip angle equal to π/2. they state that it is for the vertical component of the surface displacement that the discrepancies are the largest.30 and 1. who also considered static displacement due to uniform slip across a normal fault.1. In order to illustrate the numerical computations we chose several test cases of active/passive tsunami generation.
λi Damping coeﬃcient. L. m/s2 Value 9.8 Table 1. b0 . d. The water depth as well as the spatial extent in the main direction of propagation were chosen so that dispersive eﬀects can be neglected. ◦ Burger’s vector length. . version 2 . ν Damping coeﬃcient. g.6: Typical physical parameters used in the numerical computations.27 500 200 4 2 13 10 400 9. m Acceleration due to gravity. km Fault length. km Dip angle. δ.6.21 Jan 2008 Figure 1. GPa Poisson ratio.31: Static deformation due to the dislocation corresponding to the parameters given in Table 1. Parameter Young modulus. m Water depth (uniform). µi Fault depth. h0 . E.74 Tsunami generation tel00194763.5 0.
31 is translated onto the free surface as initial condition. Of course the waveform is diﬀerent. while the depression wave is slightly shifted.34. Thus. The second comparison focuses on the inﬂuence of the rupture velocity at two separate times (Figures 1. In particular. we can conclude from this ﬁrst comparison that if one is only interested in tsunami travel time or even in rough inundation zone estimation.38 shows the eﬀect of a longer earthquake. Figure 1. The most interesting comparison is the third one.36 and 1. The reader may have the impression that the passive solution does not evolve. We show below the diﬀerences between the passive and the dynamic approaches. a fast event with V = 2. The freesurface elevations are computed until the wave enters the purely propagation stage.32. in which the static solution shown in Figure 1.3 Tsunami generation by dynamic displacement of sea bed 75 is to make an attempt to include the dynamic displacement of the sea bed.33 and 1. The free surface at the beginning of the tsunami generation process is shown on Figure 1. It can be explained by the larger extent of the dynamic solution. In the ﬁrst comparison we use a strong coupling between the dynamic displacement of the sea bed and the ﬂuid layer equations and compare it with the passive approach. Recall that our somewhat artiﬁcial deﬁnition of earthquake duration is the time at which we stop the bottom motion. we consider three cases: the limiting case where the rupture velocity V is inﬁnite. This corresponds to Figure 1. One can see that the active approach gives at the beginning an amplitude which is almost twice larger but the amplitudes become comparable a few seconds later.37). which focuses on the duration of the earthquake.34. The rupture velocity V is assumed to be inﬁnite. This question has already been addressed by the authors [DDK06] in the framework of the linearized potential ﬂow equations and of a simpliﬁed model for bottom deformation. it evolves with the fast time scale. It could be an tel00194763. the passive approach can be used. the distinction between leading elevation wave and leading depression wave is not as clear when using the dynamical analysis. In the dynamical approach. The diﬀerences between the fast and the relatively slow rupture velocities are small.21 Jan 2008 . One sees that the shapes of the wave train obtained with the dynamic analysis look more complicated than that obtained by the passive analysis. After that time. The fast time scale is provided by the earthquake (P − and S−waves) and the slow one by water gravity waves. the sea bottom remains frozen. Since the active generation solution is directly coupled to the bottom dynamics.331.5 km/s and a slower event with V = 1 km/s.35.1. the explanation lies in the presence of two diﬀerent time scales in this problem. One can see as well that the location of the elevation wave is the same. Moreover the earthquake dynamics is computed during the ﬁrst eight seconds. version 2 . Further steps of this process are given in Figures 1. One notices that the resulting wave amplitude and velocity are almost the same. In fact. It is interesting to compare Figures 1.
76 Tsunami generation tel00194763. Figure 1.21 Jan 2008 Figure 1.32: Water free surface at the beginning of the earthquake (t = 2s) according to two approaches of tsunami generation: passive versus active (with inﬁnite rupture velocity).33: Same as Figure 1. version 2 .32 for times t = 4s and t = 6s. .
version 2 .32 for times t = 150s and t = 250s. The wave is leaving the generation zone (left plot) and starting to propagate (right plot).21 Jan 2008 Figure 1.35: Same as Figure 1. tel00194763.1.34: Same as Figure 1. Figure 1. .36: Same as Figure 1.35 (left plot) for two rupture velocities: V = 1 km/s (left) and V = 2.32 for times t = 7s and t = 10s.3 Tsunami generation by dynamic displacement of sea bed 77 Figure 1.5 km/s (right).
The computational results speak by themselves. Two methods for tsunami generation have been compared: static versus dynamic. explanation for the discrepancies between modeled and recorded time series of water levels at various locations along the California coast for the 1960 Chilean tsunami [BUST06]. These signiﬁcant diﬀerences only occur during .21 Jan 2008 Figure 1.3. 1.38: Water free surface at t = 150s for a longer earthquake. The principal novelty of the present study is the coupling of the resulting displacement ﬁeld with the hydrodynamic model. version 2 .37: Same as Figure 1. One can say that the dynamic approach leads to higher water levels in the nearfault area. The right plot compares the case of a slow rupture (V = 1 km/s) with the passive case. The left plot compares the inﬁnite rupture velocity solution with the passive approach.36 at time t = 250s.6 Conclusions An approach to model the dynamical character of sea bed deformations during an underwater earthquake was presented. The governing elastodynamic equations were solved by a ﬁniteelement method.78 Tsunami generation Figure 1. The rupture motion stops after 15s. tel00194763.
Consequently the distinction between leading elevation wave and leading depression wave may not be as clear as anticipated. tel00194763. version 2 .21 Jan 2008 . However it was also observed that the shape of the wave train can be altered by dynamical eﬀects.3 Tsunami generation by dynamic displacement of sea bed 79 the ﬁrst instants of the surface deformation and level oﬀ later on.1. Not surprisingly more accurate tsunami computations require ﬁner initial conditions such as those obtained by the active generation methodology used in the present study. Of course the present method is computationally more expensive but there is an overall gain in accuracy. In future work we intend to extend this modeling to three space dimensions since it is evident that the twodimensional computations presented here have little interest beyond academics.
version 2 .tel00194763.21 Jan 2008 80 Tsunami generation .
. . . . . . .3. . . . .1 2.8 Improvement of the linear dispersion relations . . . . 110 2.4 Alternative version of the Boussinesq equations . . . . . . .3. . . . . . . . . . 2. . . . . . . . . . . . . . . . . . . . . .1 2. . . . . . . . . . . . 100 2. 105 Bottom friction . . 108 2. . . . . . . . Blaise Pascal (1623 – 1662) Contents 2. . . . . 2. .3 Linearized potential ﬂow equations . . . . . . . . . . . . . . . . . . . . . . 113 Comparison between the dissipative models . . . . .4. . . .2 Introduction . . 104 Regularization of Boussinesq equations . . . . . . . . . . . . version 2 . . . .6 2. . . . . . . . . . . . . . . . . . . . . . . . . . . 2. . . 115 81 . . . . . . . . . 107 Spectral Fourier method . . . . . . . 113 2. . . . . . . .1 2. . . . . . . . . . .9. . . . .1 2. . . . . . . . . Dissipative Boussinesq equations . . 82 85 89 93 93 95 99 99 Derivation of the Boussinesq equations . . . .7 2. .3 Asymptotic expansion . . . . . . . .Chapter 2 Dissipative Boussinesq equations tel00194763. .21 Jan 2008 We are usually convinced more easily by reasons we have found ourselves than by those which have occurred to others. . . . . . . . . . . . . . .3. . . . .1 Derivation of the equations . . . . . . .5 2. . . . . . . . . Discussion . . . . . . . . .9. . .2. . . . . . . . . .9 Numerical results . . . . .2 Construction of the initial condition . . Analysis of the linear dispersion relations . . . . . . . . . . . 2. . . . . . . . .2 2. . . . . .8. .1 Validation of the numerical method . . . . . . . . . . . . . . . . . .
.21 Jan 2008 . . . pp. Bou72] included dispersive eﬀects for the ﬁrst time in the SaintVenant equations [dSV71]. 1871. . 34–38 and pp. Boussinesq equations contain more physics than the SaintVenant equations but at the same time they are more complicated from the mathematical and numerical points of views. There have been a lot of further developments of tel00194763. 755–759). . pp. . . there are several contributions of Boussinesq. In the Volumes 72 and 73 of the “Comptes Rendus Hebdomadaires des S´ances de l’Acad´mie des Sciences”. . On November 13. approximate models such as the Boussinesq equations must be used. which will be published in 1872 in the Journal de Math´matiques Pures et Appliqu´es (17. These equations can also be used to model tidal oscillations.1 Introduction Boussinesq equations are widely used in coastal and ocean engineering. . . . It is in 1871 that Valentin Joseph Boussinesq received the Poncelet prize from the Academy of Sciences for his work. avec application aux crues des rivi`res et ` l’introduction des mar´es dans leur lit” e a e (73. . e e 55–108). 147–154 and pp. . . . . pp. these types of wave motion are perfectly described by the NavierStokes equations. . The years 1871 and 1872 were particularly important in the development of the Boussinesq equations.10 Conclusions . which will be extended later in the note entitled “Th´orie g´n´rale des mouvements qui sont propag´s e e e e dans un canal rectangulaire horizontal” (73. On June 19. version 2 . Boussinesq presents the now famous note on the solitary wave entitled “Th´orie de l’intumescence liquide appel´e onde solitaire ou de e e translation. Of course. se propageant dans un canal rectangulaire” (72. One should mention that Boussinesq’s derivation was restricted to 1+1 dimensions (x and t) and a horizontal bottom. but currently it is impossible to solve fully threedimensional (3D) models in any signiﬁcant domain. . 237–240). SaintVenant presents a couple of notes of Boussinesq entitled “Sur le mouvement permanent vari´ de l’eau dans les tuyaux e de conduite et dans les canaux d´couverts” (73. . . SaintVenant e himself publishes a couple of notes entitled “Th´orie du mouvement non permanent des e eaux. . en communiquant au liquide contenu dans ce canal des vitesses sensiblement pareilles de la surface au fond”. One example among others is tsunami wave modelling. 119 2.82 Dissipative Boussinesq equations 2. . . which cover respectively e e the sixmonth periods January–June 1871 and July–December 1871. . . 101–105). pp. 1871. These equations possess a hyperbolic structure (the same as in the nonlinear shallowwater equations) combined with highorder derivatives to model wave dispersion. Thus. 256–260). . Articles [Bou71a. pp. All these notes deal with shallowwater theory. . . . Boussinesq submits a paper entitled “Th´orie des ondes et des remous e qui se propagent le long d’un canal rectangulaire horizontal.
while taking into account nonhydrostatic eﬀects. WKGS95. e pp. The goal of this type of modelling is to reduce 3D problems to twodimensional (2D) ones. It should be e pointed out that the famous treatise on hydrodynamics by Lamb has six editions. ν the kinematic viscosity of the ﬂuid and k the wavenumber of the decaying wave. The paragraphs on wave damping are not present in the ﬁrst edition (1879) while they are present in the third edition (1906). This diversity is due to diﬀerent possibilities in the choice of the velocity variable. There are a lot of forms of the Boussinesq equations. This is the principal physical diﬀerence with the nonlinear shallowwater (NSW) equations. Formally. 8588). (12) in [Bou95]) and Lamb both showed that dα = −2νk 2 α(t).1) leads to the classical law for viscous decay of waves of amplitude α. It is wellknown that the eﬀect of viscosity on free oscillatory waves on deep water was studied by [Lam32].21 Jan 2008 . en s’y r´duisant ` une houle e e e a simple.1 Introduction 83 these equations like in [Per67. (ii) “Lois de l’extinction de la houle en haute mer” (121. so it is possible that Boussinesq and Lamb published similar results at the same time. The resulting model performance is highly sensitive to linear dispersion properties.2. This is done by assuming a polynomial (usually linear) vertical distribution of the ﬂow ﬁeld. 1520). Boussinesq (see Eq. Let α denote the wave amplitude. pp. There is another technique used by [BCS02]. toute agitation confuse mais p´riodique des ﬂots” (121. Nwo93. The Boussinesq equations are intended to describe the irrotational motion of an incompressible homogeneous inviscid ﬂuid in the long wave limit. It provides an elegant way to improve the properties of the linear dispersion relation and it gives a quite general mathematical framework to study these systems. A good review is given by [Kir03]. Boussinesq wrote three related papers in 1895 in the “Comptes Rendus Hebdomadaires des S´ances de l’Acad´mie des Sciences”: (i) “Sur l’extinction graduelle de la e e houle de mer aux grandes distances de son lieu de production : formation des ´quations du e probl`me” (120. namely tel00194763. pp. 13811386). Let us outline the physical assumptions. The main purpose of this work is to include dissipative eﬀects in the Boussinesq equations. one can transform higherorder terms by invoking lowerorder asymptotic relations. version 2 . The authors did not have access to the second edition (1895). Indeed Lamb derived the decay rate of the linear wave amplitude in two diﬀerent ways: in paragraph 348 of the sixth edition by a dissipation calculation (this is also what [Bou95] did) and in paragraph 349 by a direct calculation based on the linearized NavierStokes equations. MS98]. (2. The right choice of the velocity variable can signiﬁcantly improve the propagation of moderately long waves. What is less known is that Boussinesq himself studied this eﬀect as well. (iii) “Sur la mani`re dont se r´gularise au loin.1) dt Equation (2. In most cases one chooses the velocity at an arbitrary water level or the depthaveraged velocity vector.
We will investigate numerically whether corresponding solutions remain close or not. Switching to dimensionless variables by introducing a characteristic velocity U . .].... 1 ∂u + u · ∇u = −∇p + ∆u. a nondissipative model means that there is no energy loss. which is not pertinent from a physical point of view. + u · ∇u = − ∂t∗ ρ ρ where u∗ (x. In other words. z ∗ . So. the NS equations become ∇ · u = 0. Let us consider the incompressible NavierStokes (NS) equations for a Newtonian ﬂuid: ∇ · u∗ = 0. F∗ the body force vector. First of all. ∂u∗ F∗ ∇p∗ ∗ ∗ ∗ + ν∆u + . In the present study. version 2 . y. (13) in [Bou95] after a few calculations). ∂t Re 1 The presence or absence of body forces is not important for discussing viscous eﬀects.84 Dissipative Boussinesq equations tel00194763.] it was found that the inclusion of a dissipative term was much more important than the inclusion of the nonlinear term. One may ask why dissipation is needed in Boussinesq equations. y ∗ . This physical eﬀect is “translated” in the language of partial diﬀerential equations by dissipative terms (e.21 Jan 2008 α ∼ exp(−2νk 2 t) (see Eq. So one expects the solutions to be diﬀerent. ρ the constant ﬂuid density and ν the kinematic viscosity. t) = (u∗ . In the “R´sum´” e e section one can read [. we use two diﬀerent models for dissipation and derive corresponding systems of longwave equations. the Laplacian in the NavierStokes equations). v ∗ . a characteristic length L and a characteristic pressure ρU 2 . The complexity of the mathematical equations due to the inclusion of this term is negligible compared to the beneﬁt of a better physical description. it is natural to have analogous terms in the long wave limit. w∗ )(x∗ . There are several methods to derive the Boussinesq equations but the resulting equations are not the same.. real world liquids are viscous. z.g. neglecting body forces1 in this discussion. Let us mention an earlier numerical and experimental study by [BPS81]. p∗ the pressure. t∗ ) is the ﬂuid velocity vector. since any ﬂow is accompanied by energy dissipation. They pointed out the importance of dissipative eﬀects for accurate long wave modelling. although the inclusion of the nonlinear term was undoubtedly beneﬁcial in describing the observations [.
we begin with the full waterwave problem. In our approach a suitable choice of the eddy viscosity. LH92. Zel91. It can signiﬁcantly simplify computational and modelling aspects. Let Ωt be the ﬂuid domain in R3 which is occupied by an inviscid and incompressible ﬂuid. which is a function of both space and time. which we use here as characteristic length scale.2 Derivation of the Boussinesq equations 85 where Re is the wellknown dimensionless parameter known as the Reynolds number and deﬁned as UL Finertia . The kinematic viscosity ν depends on the temperature but its order of magnitude for water is 10−6 m2 /s. HH70] who already included in Boussinesq models adhoc dissipative terms into momentum conservation equations in order to model wave breaking. DDZ07. In this section we choose the domain Ωt to be unbounded in the horizontal directions in order . RFF91] who included dissipation due to viscosity in potential ﬂow solutions and there are also authors [KCKD00. SVBM02. can model wave breaking at least as well as in the articles cited above. = Re = Fviscous ν From a physical point of view the Reynolds number is a measure of the relative importance of inertial forces compared to viscous eﬀects. This simple estimate clearly shows that the ﬂow is turbulent (as many other ﬂows in nature). There are several authors [Tuc74.2 Derivation of the Boussinesq equations In order to derive the Boussinesq equations. So the inclusion of dissipation can be viewed as the simplest way to take into account the turbulence. t′ ) and above by the free surface z ′ = η ′ (x′ . z ′ ) is used. y ′ . SKP96. is about 1 m. Considering that as the tsunami approaches the coast both the particle velocity and the wave amplitude increase. y ′ . For typical tsunami propagation applications the characteristic particle velocity U is about 5 cm/s and the characteristic wave amplitude. version 2 . y ′ . one can write that the corresponding Reynolds number is of the order of 105 or 106 . Modelling this eﬀect is not the primary goal of the present work.21 Jan 2008 2. since the ﬂow is no longer irrotational after wave breaking. Nevertheless scientists and engineers continue to use these equations even to model the runup on the beach. The subscript t underlines the fact that the domain varies with time and is not known a priori. A Cartesian coordinate system (x′ . tel00194763. The domain Ωt is bounded below by the seabed z ′ = −h′ (x′ .2. It is a common practice in ﬂuid dynamics (addition of an “eddy viscosity” into the governing equations for Large Eddy Simulations) to ignore the smallscale vortices when one is only interested in largescale motion. t′ ). Strictly speaking the Boussinesq equations can no longer be valid at this stage. with the x′ − and y ′ −axis along the still water level and the z ′ − axis pointing vertically upwards.
21 Jan 2008 0 z’ a L y’ O −0.5 −2 4 2 0 −2 −4 −2 0 2 4 y’ x’ Figure 2.y’.t’) −1 x’ h0 −1.86 Dissipative Boussinesq equations 0.5 z’ −h’(x’.y’. version 2 .5 η(x’.t’) tel00194763.1: Sketch of the ﬂuid domain .
′ φ′z′ = ηt′ + ∇′ φ′ · ∇′ η ′ . (2. 1 2 φ′t′ + ∇′ φ′  + gη ′ = 0. [JTPS96]): ′ Model I: Dφ′ := δ1 φ′ . [RFF91. For simplicity. z′ = η′. Thus. ∂x′ ∂y ′ ∂z ′ T . the question of the wellposedness of the BVP for the Boussinesq equations is essentially open.4) [DDZ07] pointed out that a viscous correction should also be added to the kinematic boundary condition if one takes into account the vortical component of the velocity. φ′t′ + ∇′ φ′  + gη ′ + Dφ′ = 0. wall. we complete the freesurface dynamic boundary condition (2. The reason is twofold. (2. η ′ ]. generating and/or absorbing boundary. ZV97. version 2 . This correction will be added in future work. First of all.one can choose a constant dissipation model (referred hereafter as Model I) which is often used (e. 2 φ′z′ + h′t′ + ∇′ φ′ · ∇′ h′ = 0. Then we write down the following system of equations for potential ﬂow theory in the presence of a free surface: tel00194763. z ′ ) ∈ Ωt = R2 × [−h′ .g. y ′ .2 Derivation of the Boussinesq equations 87 to avoid the discussion on lateral boundary conditions. It has been assumed implicitly that the free surface is a graph and that the pressure is constant on the free surface (no forcing). z ′ = −h′ . η ′ + h′ > 0 (there is no dry zone). If the ﬂow is assumed to be irrotational one can introduce the velocity potential φ′ deﬁned by u′ = ∇′ φ′ .2) by adding a dissipative term to account for the viscous eﬀects2 : 1 2 ′ z′ = η′. A typical sketch of the domain Ωt is given in Figure 2. 2 ′ In this work we investigate two models for the dissipative term Dφ′ .e. runup on a beach) depends on the application under consideration and secondly. the choice of the boundary value problem (BVP) (e. . Moreover we assume that the total water depth remains positive.1.3) There is a more physically realistic dissipation model which is obtained upon balancing of normal stress on the free surface (e. . this system of equations does not contain any dissipation. z′ = η′. where u′ denotes the velocity ﬁeld.2. i. (2. (x′ .g. Primes stand for dimensional variables.g. DDZ07]): Model II: 2 ′ Dφ′ := δ2 φ′z′ z′ . As written.2) where g denotes the acceleration due to gravity (surface tension eﬀects are usually neglected for longwave applications). ∇′ := ∂ ∂ ∂ .21 Jan 2008 ∆′ φ′ = 0.
2 2 z φz + φz = µ2 ηt + εµ2 ∇φ · ∇η. y. ε := . ∂x ∂y T .5) (2. (2. (x.88 Dissipative Boussinesq equations The derivation of Boussinesq equations is more transparent when one works with scaled variables.7) tel00194763. h0 ℓ In these equations and hereafter the symbol ∇ denotes the horizontal gradient: ∇ := ∂ ∂ . µ2 ht + εµ2 ∇φ · ∇h = 0. x= . h0 a0 ga0 ℓ where h0 . After this change of variables. R1 Model II: Dφ = 1 φzz . The dissipative term Dφ is given by the chosen model (2. ℓ ℓ h0 ℓ √ gh0 ′ η′ h′ φ. From this dimensional analysis. respectively. z) ∈ Ωt .4): Model I: Dφ = 1 φ.3) or (2.9) µ := . one can conclude that the dimension of the coeﬃcient δ1 is [s−1 ] and that of δ2 is [m2 s−1 ]. the set of equations becomes µ2 (φxx + φyy ) + φzz = 0. . t= t. φ= h= . z= .8) ε where ε and µ are the classical nonlinearity and frequency dispersion parameters deﬁned by h0 a0 (2. R2 := 1 δ2 ga ℓ √ 0 gh0 . z = εη.6) (2. Thus.21 Jan 2008 1 1 µ2 φt + εµ2 ∇φ2 + εφ2 + µ2 η + εDφ = 0. wavelength and wave amplitude. η= . it is natural to call the ﬁrst coeﬃcient viscous frequency (since it has the dimensions of a frequency) and the second one kinematic viscosity (by analogy with the NS equations). R2 where the following dimensionless numbers have been introduced: R1 := 1 δ1 h2 0 ga0 ℓ √ gh0 . z = εη (2. ℓ and a0 denote a characteristic water depth. Let us introduce the following independent and dependent nondimensional variables: √ gh0 ′ y′ z′ x′ y= . version 2 . z = −h.
that is 8 × 106 . 2. = 0.2 Derivation of the Boussinesq equations parameter Acceleration due to gravity g. The ratio 1 between inertial forces and viscous forces is 2 εµ2 ∇φ2 /εDφ . we will use the notation νi := 1/Ri . As an example. . km Water depth h0 . This will allow us to unify the physical origin of the numbers Ri with the eddyviscosity approach.1 Asymptotic expansion Consider a formal asymptotic expansion of the velocity potential φ in powers of the small parameter µ2 : φ = φ0 + µ2 φ1 + µ4 φ2 + . since we know how to relate the value of δ2 to the kinematic viscosity of water. In other words. one needs to increase the importance of viscous terms in the governing equations in order to account for turbulent dissipation.5) and the boundary conditions. the main eﬀect of turbulence is energy dissipation. . . m/s2 Amplitude a0 .2. . m Wave length ℓ.11) Then substitute this expansion into the continuity equation (2. Thus.10) In numerical computations the authors took the coeﬃcient µ = 0.21 Jan 2008 It is interesting to estimate R2 . They modeled long wave propagation by using a modiﬁed dissipative Korteweg–de Vries equation: 3 1 ηt + ηx + ηηx − µηxx − ηxxt = 0. . we will “forget” about the origin of these coeﬃcients. (2. It means that. for the sake of convenience.1. After substitution.6 × 10−3 . This value gave good agreement with laboratory data.) + φ0zz + µ2 φ1zz + µ4 φ2zz + . the Laplace equation becomes µ2 (∇2 φ0 + µ2 ∇2 φ1 + µ4 ∇2 φ2 + . It clearly shows that the ﬂow is turbulent and eddyviscosity type approaches should be used.1: Typical values of characteristic parameters in tsunami applications tel00194763. m2 /s value 10 1 100 4 10−6 89 Table 2. 2 6 (2. Its order of magnitude is µ2 R2 . Typical parameters which are used in tsunami wave modelling are given in Table 2. we refer one more time to the work by [BPS81].014. . at zerothorder approximation. km Kinematic viscosity δ. . . because their values can be given by other physical considerations. version 2 . From now on. For these parameters R2 = 5 × 109 and µ2 = 1.2.
90 Dissipative Boussinesq equations Collecting the same order terms yields the following equations in the domain Ωt : µ0 : φ0zz = 0. y. v)T .17) µ2 : φ1z + 1 ht + ∇φ0 · ∇h = 0.19) ε .21 Jan 2008 φ0 = φ0 (x. φ1 . φ2 . Performing the same computation for the bottom boundary condition yields the following relations at z = −h: µ0 : µ4 : φ0z = 0. (2.13) (2. ε and integrating one more time with respect to z gives the expression for φ1 : 1 φ1 = − (z + h)2 ∇ · u − z 2 1 ht + u · ∇h . y.13) one can express φ1 in terms of the derivatives of φ0 : φ1zz = −∇ · u.15) one immediately concludes that tel00194763.12) (2. y.16): 1 φ1z = −(z + h)∇ · u − ht − u · ∇h. u = (u. (2. t). φ2z + ∇φ1 · ∇h = 0. ε From equation (2. etc. y. Using (2. t) := ∇φ0 . t) can be found by using condition (2. The unknown function C1 (x.14): 1 φ2zz = (z + h)2 ∇2 (∇ · u) + (h + z)∇2 h + ∇h2 ∇ · u+ 2 1 2 + 2(h + z)∇h · ∇(∇ · u) + z ∇ ht + ∇2 (u · ∇h) . Let us deﬁne the horizontal velocity vector u(x. ε (2. For this purpose we use equation (2. (2. The expansion of Laplace equation in powers of µ2 gives recurrence relations between φ0 .14) µ2 : φ1zz + ∇2 φ0 = 0. µ4 : φ2zz + ∇2 φ1 = 0.12) and the boundary condition (2. Integrating once with respect to z yields φ1z = −z∇ · u + C1 (x. t).18) Now we will determine φ2 . version 2 .16) (2.15) (2.
it is possible to obtain highorder Boussinesq equations. y. The amplitude of the bottom motion has to be of the same order of magnitude as the resulting waves. y) is the static seabed and ζ(x. y.17) yields the following expression for φ2 : φ2 = 1 1 1 (h + z)4 ∇2 (∇ · u) + (z + h)3 ∇2 h + z 2 ∇h2 ∇ · u 24 6 2 1 z3 1 2 + (z + h)3 ∇h · ∇(∇ · u) + ∇ ht + ∇2 (u · ∇h) 3 6 ε 1 1 h − zh ∇2 ht + u · ∇h + ∇h · ∇ ht + u · ∇h − ∇h2 ∇ · u . t) is the dynamic component due to a seismic event or a landslide (see for example [DD07c] for a practical algorithm constructing ζ(x. t) := h0 (x. Of course.6): φ0z + µ2 φ1z + µ4 φ2z = µ2 ηt + εµ2 ∇φ0 · ∇η + O(ε2 + εµ4 + µ6 ). (2. No substantial diﬀerence was revealed. Recently. In the present study we restrict our attention to dispersion terms up to order O(µ2 ). First of all. We will also assume that the UrsellStokes number [Urs53] is O(1): S := ε = O(1).2 Derivation of the Boussinesq equations 91 Integrating twice with respect to z and using the bottom boundary condition (2. t). y. since ε2 = S 2 µ4 = O(µ4 ). µ2 This assumption implies that terms of order O(ε2 ) and O(εµ2 ) must be neglected. y. [Ben06] performed a comparative study between fullynonlinear equations [WKGS95] and Boussinesq equations with optimized dispersion relation [Nwo93]. (2. t) has the following special form in dimensionless variables: h(x. y) − εζ(x. version 2 . We would like to emphasize that this term is O(1). since in problems of wave generation by a moving bottom the bathymetry h(x. εµ2 = Sµ4 = O(µ4 ).22) . Now. we substitute the asymptotic expansion (2. Thus (1/ε)ht = −ζt = O(1).20) 2 ε ε tel00194763. we are ready to derive dissipative Boussinesq equations in their simplest form. For highorder asymptotic expansions we refer to [WKGS95. (2. z = εη.11) into the kinematic freesurface boundary condition (2. We decided not to take this research direction. t) in the absence of a dynamic source model). MS98].21 Jan 2008 Remark 2 In these equations one ﬁnds the term (1/ε)ht due to the moving bathymetry.2. y. since we assume the ﬂuid to be inviscid and incompressible.21) where h0 (x.
µ2 2 h∇ 2 · ut = 0.20) one can evaluate φ1z and φ2z on the free surface: 1 φ1z z=εη = −(h + εη)∇ · u − ht − u · ∇h.92 Dissipative Boussinesq equations The ﬁrst term on the left hand side is equal to zero because of Eq. respectively. ε φ2z z=εη = h3 2 ∇ (∇ · u) + h2 ∇h · ∇(∇ · u) + h 6 h 2 ∇ h + ∇h2 ∇ · u 2 h2 1 2 1 − ∇ ht − h ∇ht · ∇h + O(ε). The last step consists in diﬀerentiating the above equations with respect to the horizontal coordinates in order to obtain equations for the evolution of the velocity.15). so we have the obvious relation ∂u ∂v = . We also perform some minor transformations using the fact that the vector u is a gradient by deﬁnition.18) and (2. 2 ε ε Substituting these expressions into (2. After performing all these operations one can write down the following equations: Model I: Model II: ε ε φ0t + 2 u2 + η + ν1 µ2 φ0 − ν1 ε 2 h∇ 2 ε φ0t + 2 u2 + η − ν2 ε∇ · u − ·u− µ2 2 h2 ∇ · ut = 0.21 Jan 2008 ηt + ∇ · (h + εη)u = − 1 + µ2 2 2 h ∇ + µ2 h∇h · ∇ 2 h3 1 ht + µ2 ∇2 (∇ · u) ε 6 h 2 + µ2 h h∇h · ∇(∇ · u) + ∇ h + ∇h2 ∇ · u .7). φ1t and φ1zz along the free surface z = εη and then substitute the expressions into the asymptotic form of (2. 2 where. This derivation will depend on the selected dissipation model.7): 1 µ2 φ0t + µ4 φ1t + εµ2 ∇φ0 2 + µ2 η + εν2 µ2 φ1zz = O(ε2 + εµ4 + µ6 ). (2.22) and retaining only terms of order O(ε+µ2 ) yields the freesurface elevation equation: tel00194763. ∂y ∂x The resulting Boussinesq equations for the ﬁrst and second dissipation models. For both models one has to evaluate φ1 . Using expressions (2. 2 The equation for the evolution of the velocity ﬁeld is derived similarly from the dynamic boundary condition (2. dissipative terms are given according to the second model. as an example. are given below: ηt + ∇ · (h + εη)u = − 1 + µ2 2 2 h ∇ + µ2 h∇h · ∇ 2 h3 1 ht + µ2 ∇2 (∇ · u) ε 6 . version 2 .
Substituting this form of solutions into equations (2. φz = ηt .3. version 2 .25) 2. z = −h. 2 2 1 1 ut + 2 ε∇u2 + ∇η = εν2 ∇2 u + 2 µ2 ∇(h2 ∇ · ut ).1 tel00194763. ϕ0 Straightforward computations give the solution to this problem: ϕ(z) = −i η0 ϕ0 ek(2h+z) + e−kz e2kh − 1 ω .26) (2. z) ∈ R × [−h. η(x. t) = η0 ei(kx−ωt) .30) where ϕ0 and η0 are constants. The next step consists in choosing a special form of solutions: φ(x. so h = const. (2. 2 93 (2. As above the term Dφ depends on the selected dissipation model and is equal to δ1 φ or δ2 φzz .27) and (2. (2. z. φz = 0.26). 0].29) yields the following boundary value problem for an ordinary diﬀerential equation: ϕ′′ (z) − k 2 ϕ(z) = 0. z ∈ [−h.3 Analysis of the linear dispersion relations + µ2 h h∇h · ∇(∇ · u) + h 2 ∇ h + ∇h2 ∇ · u . (2.24) (2.3 Analysis of the linear dispersion relations For simplicity.27) (2. 0]. z = 0. η0 ϕ′ (0) = (−iω). we will consider in this section only 2D problems. φt + gη + Dφ = 0.2.23) Model II: 1 Model I: ut + 1 ε∇u2 + ∇η + ν1 Su = 2 εν1 ∇(h2 ∇ · u) + 1 µ2 ∇(h2 ∇ · ut ). The generalization to 3D problems is straightforward and does not change the analysis.21 Jan 2008 Linearization of the full potential ﬂow equations with dissipation First we write down the linearization of the full potential ﬂow equations in dimensional form. z = 0. 2.29) Remark 3 In this section the water layer is assumed to be of uniform depth. ϕ′ (−h) = 0. (x. t) = ϕ0 ei(kx−ωt) ϕ(z). (2. after dropping the primes: ∆φ = 0. k .28) (2.
35) − kh 2 2 It can be shown that in order to keep the phase velocity unchanged by the addition of dissipation.4) one obtains the following relation: ω=± gk tanh(kh) − (2. ∀k in order to avoid the exponential growth of certain wavelengths. The expressions for the phase velocity are obtained from the corresponding dispersion relations (2. The problem is that ω and k cannot be arbitrary. Dφ = δ1 φ and the dispersion relation is given implicitly by ω 2 + iδ1 ω − gk tanh(kh) = 0.32) If δ1. . 2k (2.33) ω = ± gk tanh(kh). similar dissipative terms must be included in both the kinematic and the dynamic boundary conditions [DDZ07]. k The phase velocity is directly connected to the speed of wave propagation and is extremely important for accurate tsunami modelling since tsunami arrival time obviously depends on the propagation speed.2 ≡ 0 one easily recognizes the dispersion relation of the classical waterwave problem: (2.94 Dissipative Boussinesq equations The dispersion relation can be thought as a necessary condition for solutions of the form (2. 4 2 For the second dissipation model (2.31) tel00194763. For our analysis it is more interesting to look at the phase speed which is deﬁned as ω(k) cp (k) := . since ei(kx−ω(k)t) = eIm ω(k)t · ei(kx−Re ω(k)t) . One can easily solve this quadratic equation for ω as a function of k: ω = ± gk tanh(kh) − δ2 k 2 2 2 − iδ2 2 k . after substituting this solution into (2.34) tanh(kh) δ2 k 2 iδ2 − k. which is called the dispersion relation. 2 (2. Remark 4 It is important to have the property Im ω(k) ≤ 0.21 Jan 2008 ω 2 + iδ2 ωk 2 − gk tanh(kh) = 0. version 2 .32): c(1) (k) = ± gh p c(2) (k) = ± p gh δ1 tanh(kh) − kh 2k 2 − iδ1 .28). When the dissipative term is chosen according to model I (2. or in explicit form by 2 iδ1 δ1 − .3).31) and (2.30) to exist. (2. We obtain the required relation ω = ω(k).
2 Dissipative Boussinesq equations The analysis of the dispersion relation is even more straightforward for Boussinesq equations. The corresponding phase velocities are given by (1) cpb = gh 1 + 1 (kh)2 6 1 1 + 2 (kh)2 − − δ1 2k 2 − 2 iδ1 .3.3 Analysis of the linear dispersion relations 95 2. 2 + (kh)2 (2. 2 2 This system admits nontrivial solutions if its determinant is equal to zero. In the case of the ﬁrst model.21 Jan 2008 η = η0 ei(kx−ωt) . 6 1 ut + gηx + δ1 u = 2 δ1 h2 uxx + 1 h2 uxxt . one obtains the following homogeneous system of linear equations: 1 (−iω)η0 + ikh 1 + (kh)2 u0 = 0. u = u0 ei(kx−ωt) .37) .36) (2) cpb = gh 1 + 1 (kh)2 6 1 + 1 (kh)2 2 δ2 k 2 + (kh)2 − (2. we switch to dimensional variables. It gives the required dispersion relation: ω 2 + iωδ1 − ghk 2 1 1 + 6 (kh)2 1 + 1 (kh)2 2 = 0. 6 iω δ1 gikη0 + −iω + δ1 + (kh)2 − (kh)2 u0 = 0. 2k iδ2 k .2. 2 1 ut + gηx = δ2 uxx + 2 h2 uxxt . version 2 . As usual we begin with the (1 + 1)D linearized equations: ηt + hux = Model I: Model II: h3 uxxx . In order to be coherent with the previous subsection. where η0 and u0 are constants. Now we substitute a special ansatz in these equations: tel00194763. A similar relation is found for the second model: iωδ2 ω + k 2 − ghk 2 1 2 1 + 2 (kh) 2 1 + 1 (kh)2 6 1 1 + 2 (kh)2 = 0.
version 2 .8 Re cp/(gh)1/2 Full.6 0.2 with a zoom on long waves.2 0 0 0.1 0 0 2 4 6 kh 8 10 12 14 Figure 2.1 0.8 0.5 0.21 Jan 2008 0.2 Full. Same as Figure 2.4 0.4 0. Zoom 1 0. Phase speeds.6 Figure 2.5 0. Real part 1 0. no dissipation Full equations with dissipation−1 Dissipative Boussinesq−1 0.6 0.3 kh 0. no dissipation Full equations with dissipation−1 Dissipative Boussinesq−1 tel00194763.3: Dissipation model I.7 Re cp/(gh)1/2 0. .2: Dissipation model I.2 0.96 Dissipative Boussinesq equations Phase speeds. Real part.9 0. Real part of the phase velocity.4 0.3 0.
12 tel00194763.9 0.1 kh 0. version 2 . Imaginary part of the frequency. no dissipation Full equations with dissipation−2 Dissipative Boussinesq−2 Figure 2. Phase speeds.15 0.08 −0. Real part 1 0. . Branch − Full equations with dissipation−1.21 Jan 2008 −0.4 0.5: Dissipation model II.3 Analysis of the linear dispersion relations 97 Dispersion relation.14 0 0. Imaginary part 0 Full equations with dissipation−1.5 0.05 0.2 Figure 2.2. Real part of the phase velocity. Branch + Dissipative Boussinesq−1 −0.06 −0.4: Dissipation model I.1 −0.2 0.3 0.02 −0.6 0.04 Im ω(k) −0.1 0 0 2 4 6 kh 8 10 12 14 Full.8 0.7 Re cp/(gh)1/2 0.
Dispersion relations. Same as Figure 2.21 Jan 2008 −18 −20 0 2 4 6 kh 8 10 12 14 Figure 2.1 −0.6 with a zoom on long waves.4 Im ω(k) −0.5 1 1.7: Dissipation model II. version 2 . Imaginary part 0 −2 −4 −6 −8 Im ω(k) −10 −12 −14 −16 Full equations with dissipation−2. Imaginary part 0 −0. Imaginary part of the frequency.8 −0.7 −0. .5 3 3.98 Dissipative Boussinesq equations Dispersion relations.2 −0.5 −0. Branch + Dissipative Boussinesq−2 tel00194763. Branch − Full equations with dissipation−2.9 −1 0 0.6: Dissipation model II. Branch + Dissipative Boussinesq−2 Figure 2.5 2 kh 2.3 −0.5 4 Full equations with dissipation−2.6 −0.
as is often the case in physics. the ﬁrst dissipation model prefers very long waves. which provides slightly diﬀerent governing equations. (2. one can see that dissipation is very selective. The real and imaginary parts of the phase velocities (2. On the other hand we would like to point out that the second model admits a critical wavenumber kc such that the phase velocity (2. Per72]. this property disappears for physically realistic values of the parameters g.1).37)) depends on the roots of the following polynomial equation: (kh)4 + 8 − 2 3δ2 gh3 tel00194763. h and δ2 (see Table 2.3 that very long linear waves are not advected in the ﬁrst dissipation model. 2. but only dissipated. Moreover one can see from the expressions (2.4 Alternative version of the Boussinesq equations In this section we give an alternative derivation of Boussinesq equations. Clearly. Ben74.2 – 2.14. Together with the dissipative models we also plotted for comparison the wellknown phase velocity corresponding to the full conservative (linearized) waterwave problem: tanh(kh) cp (k) = gh . In this example the parameters are given by δ1 = 0.21 Jan 2008 (kh)2 + 12 = 0.3 Discussion Let us now provide a discussion on the dispersion relations. the hyperbolic structure is the .35) becomes purely imaginary with negative imaginary part. it can be clearly seen in Figure 2.34). since the real part of their phase velocity is identically equal to zero. Furthermore. δ2 = 0.14. We use another classical method for deriving Boussinesqtype equations [Whi99.34) – (2.37) for the full and long wave linearized equations are shown graphically on Figures 2.4 Alternative version of the Boussinesq equations 99 2. 2 3δ2 gh3 This equation does not have real roots since ≪ 1. That is why we suggest to make use of the second model in applications involving very long waves such as tsunamis. version 2 . When one switches to the Boussinesq approximation. kh First of all.7. (2) Let us clarify this situation. The qualitative behaviour of the phase velocity cpb (see equation (2.2.3. From a physical point of view it means that the waves shorter than kc are not advected. while the second model dissipates essentially short waves. Namely.36) that the phase velocity has a 1/k singular behaviour in the vicinity of kh = 0 (in the long wave limit).
2. the derivation process can be performed in a similar fashion for model I (2. y. t) = n=0 z n φn (x. y. in our opinion. t) = n=0 (−1)n µ2n z 2n 2n z 2n+1 ∇ φ0 + ∇2n φ1 .6). z. ∂z )φ: 1 u0 = ∇φ0 . but the dispersive terms diﬀer.38): ∞ φ(x. w0 = 2 φ1 . µ2 It is straightforward to ﬁnd the relations between u0 . . (2. y. dissipative eﬀects which are modelled this time according to model II (2. z. µ . it is just convenient to use this notation in asymptotic expansions but in practice. this model is more appropriate for long wave applications. 1. y.4. t). z. take into account a moving bathymetry and. This leads to the classical recurrence relation φn+2 (x. In other words. When we substitute the expansion (2. φ1 if one remembers that ∂ (u.1 Derivation of the equations The starting point is the same: equations (2. y. (n + 1)(n + 2) n = 0. w0 and φ0 . we have an inﬁnite polynomial in z. t). y. t). (2.4) because. t).38) We would like to emphasize that this expansion is only formal and no convergence result is needed. (2n)! (2n + 1)! The following notation is introduced: u0 := u(x. It is not necessary to justify the convergence of the sum with three or four terms. 0. version 2 . Using this relation one can eliminate all but two unknown functions in (2. 2. This time the procedure begins with representing the velocity potential φ(x. .5). The derivation follows closely the paper by [MS98]. y.3). 0. Anyhow.5). Requiring that φ formally satisﬁes Laplace equation implies that the coeﬃcients of each power of z vanish (since the righthand side is identically zero). t) as a formal expansion in powers of z rather than of µ2 : tel00194763. t) = − µ2 ∇2 φn (x.7) and (2. . seldom more than four terms are used. w) = (∇.8). In numerical simulations we suggest to use this system of equations.21 Jan 2008 ∞ φ(x.38) into Laplace equation (2. The main diﬀerences are that we neglect the terms of order O(µ4 ). of course. w0 := 1 w(x.100 Dissipative Boussinesq equations same. y. (2.
(2n + 1)! (2n)! . t) + . .2. . In this asymptotic framework the above formulas become much simpler: z µ2 z 2 ∇2 φ0 + ∇2 φ1 + O(µ4 ). In order to establish the relation between w0 and u0 one uses the bottom kinematic boundary condition (2.8). (2. (2. (2. y. In the present work we neglect the terms of order O(µ4 ) and higher.21 Jan 2008 µ2 z 2 ∇(∇ · u0 ) + O(µ4 ).42) + ε w0 − µ2 ∇2 (∇ · u0 ) − 6 2 In order to obtain the expression of w0 in terms of u0 one introduces one more expansion: w0 (x. t) + µ2 w0 (x.39). This (0) (1) leads to the following explicit expressions for w0 and w0 : 1 (0) w0 = − ht − ∇ · (hu0 ). . version 2 .4 Alternative version of the Boussinesq equations 101 Using the deﬁnition of the velocity potential φ one can express the velocity ﬁeld in terms of u0 . y.40). t) = w0 (x. . (0) (1) (2. w0 : ∞ u= n=0 (−1)n ∞ z 2n+1 z 2n 2n µ ∇ ∇2n−2 (∇ · u0 ) + µ2n+2 ∇ ∇2n w0 (2n)! (2n + 1)! z 2n+1 z 2n 2n+2 2n µ2n+2 ∇2n (∇ · u0 ) + µ ∇ w0 .42). which has the following form after substituting the asymptotic expansions (2.43) We insert this expansion into the asymptotic bottom boundary condition (2.41) w = µ2 w0 − zµ2 ∇ · u0 + O(µ4 ). ε h h2 ∇h · ∇(∇ · u0 ) + ∇2 (∇ · u0 ) = 2 3 1 h 1 2 − h ∇h · ∇ht + ∇h · ∇ ∇ · (hu0 ) + ∇ ht + ∇2 (∇ · (hu0 )) ε 2 ε (1) w0 . u = u0 + zµ ∇w0 − 2 2 (2. w= n=0 (−1)n − These formulas are exact but not practical. y.40) (2.41) in it: ht + ε∇h · u0 − hµ2 ∇w0 − µ2 h2 ∇(∇ · u0 ) 2 µ2 h2 2 h3 ∇ w0 + O(µ4 ) = 0.39) φ = φ0 + zφ1 − 2 3 tel00194763. (2.
46) In this work we apply a trick due to [Nwo93]. In fact.44) in terms of u0 .47) Remark 5 Behind this change of variables there is one subtlety which is generally hushed up in the literature.21 Jan 2008 w=− (2. one can improve the linear dispersion relation and this is important for wave modelling. (2. ε 2 µ2 ht − µ2 ∇ · (hu0 ) + z∇ · u0 + O(µ4 ).5).40)(2. It seems that this point is not clearly mentioned in the literature on this topic. (2. w) = O when u and w are expressed in terms of the variable uα . First we evaluate (2.102 Dissipative Boussinesq equations Substituting these expansions into (2.46) or. (2.45). Technically this change of variables is done as follows. But on the other hand.41) become u = u0 − z µ2 z2 ∇ht − µ2 z∇ ∇ · (hu0 ) + ∇(∇ · u0 ) + O(µ4 ).45) at z = zα .7). . the wave motion is assumed to be irrotational since we use the potential ﬂow formulation (2. Namely. one loses the property that the ﬂow is irrotational. ε (2. By construction rot(u. ε 2 (2.6). which gives the connection between u0 and uα : uα = u0 − zα z2 µ2 ∇ht − µ2 zα ∇ ∇ · (hu0 ) + α ∇(∇ · u0 ) + O(µ4 ).45) tel00194763. ε 2 Using the standard techniques of inversion one can rewrite the last expression as an asymptotic formula for u0 in terms of uα : u0 = uα + zα z2 µ2 ht + µ2 zα ∇ ∇ · (huα ) + α ∇(∇ · uα ) + O(µ4 ). version 2 . in other words. (2.43) and performing some simpliﬁcations yields the required relation between u0 and w0 : 1 w0 = − ht − ∇ · (hu0 ) ε − µ2 ∇ · h2 h2 h3 ∇ht + ∇ ∇ · (hu0 ) − ∇(∇ · u0 ) + O(µ4 ). w) = O when u and w are computed according to (2.8) of the waterwave problem. (2. When one turns to the velocity variable uα deﬁned at an arbitrary level.44) 2ε 2 6 Now one can eliminate the vertical velocity w0 since one has its expression (2. we introduce a new velocity variable uα deﬁned at an arbitrary water level zα = −αh. That is to say. Equations (2. a direct computation shows that rot(u. The purpose of this remark is simply to inform the reader about the price to be paid while improving the dispersion relation properties. in terms of the variable u0 .
We prefer to introduce this function in order to eliminate the division by ε in the source terms since this division can give the impression that stiﬀ source terms are present in our problem. Technically it is done by using the relation (2.2.7). The ﬁrst method consists in integrating the continuity equation (2. the equation for the horizontal velocity ﬁeld is derived from the dynamic freesurface boundary condition (2. which is not the case. 3 . When the bathymetry is static. In this section we also assume that the StokesUrsell number S is of order O(1). It consists in using directly the kinematic freesurface boundary condition (2.2 and we do not insist on this point: ε u0t + ∇ u0 2 + ∇η − εδ∇2 u0 = O. The second way is more straightforward. 2 Switching to the variable uα yields the following governing equation: ε z2 uαt + ∇uα 2 + ∇η + µ2 zα ∇ ∇ · (huα ) + α ∇(∇ · uα ) = 2 2 t = εδ∆uα + µ2 (zα ∇ζt )t . There are two diﬀerent methods to obtain the freesurface elevation equation.6) and perform several simpliﬁcations. It is done exactly as in section 2.21).5) over the depth and then use the kinematic freesurface and bottom boundary conditions.49) We already discussed this point on page 91.47) between u0 and uα . we switch to the variable uα . (2. version 2 . t) is deﬁned according to (2.4 Alternative version of the Boussinesq equations 103 Let us now derive the Boussinesq equations.48) 2 2 As above.21 Jan 2008 µ2 ∇ · (h2 ∇ζt ). y.6): ηt + ε∇φ · ∇η − 1 φz = 0. In order to be able to optimize the dispersion relation properties.39) into (2. (2. Neglecting all terms of order O(ε2 + εµ2 + µ4 ) yields the following equation3 : ηt + ∇ · (h + εη)u0 + h3 µ2 ∇ · h2 ∇ ∇ · (hu0 ) − ∇(∇ · u0 ) = 2 3 = ζt + tel00194763. 2 Recall that ζ(x. ζ ≡ 0. The result is given below: ηt + ∇ · (h + εη)uα h h 2 h2 + µ ∇ · h zα + ∇ ∇ · (huα ) + zα − ∇(∇ · uα ) = 2 2 3 h = ζt + µ2 ∇ · h zα + ∇ζt . µ2 Then one can substitute (2.
104 Dissipative Boussinesq equations In several numerical methods it can be advantageous to rewrite the system (2.49) in vector form: Ut + µ2 L(U)t + ∇ · F(U) + µ2 ∇ · P(U) = S(x. S := ζt + µ2 ∇ · h zα + h 2 µ2 (zα ∇ζt )t 0 0 0 D := 0 1 0 . at least linearly.21 Jan 2008 · uα ) h2 3 . y.5 Improvement of the linear dispersion relations As said above. Usually Boussinesqtype models with good dispersion characteristics are linearly wellposed as well. . But these authors started with a desired dispersion relation and artiﬁcially added extra terms to the momentum equation in order to produce the desired characteristics.48). MMS91]. We refer to [BCS02] as a general reference on this topic. ∇ · F := + . (2. the idea of using one free parameter α ∈ [0. 1] to optimize the linear dispersion relation properties appears to have been proposed ﬁrst by [Nwo93]. ∇ ∇ · (huα ) + O h 2 2 zα − ∇(∇ · uα ) . ε 2 0 uα  + η 2 L := P := h zα + 0 zα ∇ ∇ · (huα ) + h 2 2 zα ∇(∇ 2 tel00194763. Remark 6 When one plays with the dispersion relation it is important to remember that the resulting problem must be wellposed. See for example [Mur89. ∂x ∂y vα (h + εη)vα (h + εη)uα ε F := 2 uα 2 + η . 0 0 1 ∇ζt . 2. We prefer to follow the ideas of [Nwo93]. G := 0 . t) + εδ∇ · (D∇U)). The idea of manipulating the dispersion relation was wellknown before 1993. where η ∂F ∂G U := uα . version 2 .
48).50) to be exact up to order O ((kh)4 ) one immediately obtains an equation for αopt : √ αopt (2 − αopt ) 5 2 = ⇒ αopt = 1 − ≈ 0. Indeed we want to concentrate our attention on the propagation properties which are more important. gh kh 3 15 If one insists on the dispersion relation (2. We are going to use lower order asymptotic relations in order to play with dispersive terms.49).50) tel00194763. In the present work we choose α by comparing the coeﬃcients in the Taylor expansions of the phase velocity in the vicinity of kh = 0. from (2. For simplicity we assume the bottom to be static h = h(x. . 6 15 5 We suggest using this value of α in numerical computations. The result is 1 1 − α − α + 3 (kh)2 c2 (k) α(2 − α) 1 b 2 = (kh)4 + O (kh)6 .52) ut = −∇η + O(ε + µ2 ).33) in the least square sense. It was already extensively exploited in [BCS02] and other publications.33): tanh(kh) 1 2 c2 (k) (2.55. y) and we drop out viscous terms. (2. The inﬂuence of this modiﬁcation on dispersion relation properties will be discussed below.6 Regularization of Boussinesq equations 105 In order to look for an optimal value of α we will drop dissipative terms.49).2. The choice for the parameter α depends on the optimization criterion.48)(2. One can also use Pad´ approximants [Wit84] since rational functions e have better approximation properties than polynomials. This operation will allow us to obtain very gentle system of equations from the numerical point of view.21 Jan 2008 On the other hand one can write down the phase velocity of the full linearized equations (2. version 2 . We brieﬂy describe the procedure. which corresponds to the longwave limit. 2. Boussinesq equations (2. Namely. nonviscous. (2.6 Regularization of Boussinesq equations In this section we are going to modify further just obtained Boussinesq equations (2. First of all one has to obtain the phase velocity of the linearized. = 1 − (kh)2 + α 2 gh 3 6 1 − α 2 − 1 (kh) 2 (2. Another possibility is to match the dispersion relation of the full linearized equations (2.48). (2. The main idea behind is very simple.49) one can deduce two following relations: ηt = −∇ · (hu) + O(ε + µ2 ).51) = = 1 − (kh)2 + (kh)4 + O (kh)6 .
0800 (let us recall that α = 0 corresponds to the free surface and α = 1 to the bottom level). In practice.53).48): µ2 ∇ · h zα + h ∇ ∇ · (huα ) 2 = −µ2 ∇ · h zα + h ∇η 2 + O(εµ2 + µ4 ). With its help we can modify the ﬁrst dispersive term in the equation (2. version 2 .55) with optimal values of α. (2. The main purpose of this little modiﬁcation is to decrease the order of dispersive operator which can be stiﬀ from the numerical point of view: (1 − ∂xx )ut + ux + uux = 0 → ut + (1 − ∂xx )−1 (ux + uux ) = 0.21 Jan 2008 ε z2 uαt + ∇uα 2 + ∇η + µ2 zα ∇ ∇ · (huα ) + α ∇(∇ · uα ) = 0.106 Dissipative Boussinesq equations In the present work we will only use the relation (2. the regularized system of equations reads: ηt + ∇ · (h + εη)uα − µ2 ∇ · h zα + h ∇η 2 + µ2 ∇ · h 2 h2 z − ∇(∇ · uα ) = 0.51) and (2.50).54) is the following: c2 (k) r := gh 1+ 1+ 1 2 1 1 2 3 − α2 1+α 1− − α (kh)2 α 2 = (kh)2 (2.52).53) 2 α 3 t tel00194763. the experiments show that the regularized equations (such as BBM) completed by dissipative terms seem to be excellent in description of long water waves [BPS81].54) 2 2 t Just obtained system of equations has the same relation with classical Boussinesq equations [Per67] as KdV equation ut + ux + uux + uxxx = 0 is related to BBM model [BBM72] ut + ux + uux − uxxt = 0. 2 12 2 30 This equation has only one root which has physical sense and lead to linearly wellposed problem αopt := 0. (2. In this way we get the following equation for α: 1 3 13 2 1 1 α − α + α− = 0. it can drastically change the CFL condition from ∆t = O(∆x2 ) for KdV (see [Tre00] for example) to ∆t = O(∆x) for BBM equation (or even unconditionally stable schemes if one considers an equivalent integral representation of solutions [BPS81.55) 1 13 1 1 1 = 1 − (kh)2 − α3 − α2 + α − (kh)4 + O (kh)6 3 2 12 2 6 Here again we can choose the parameter α so that the dispersion relation (2. We have to say that this operation does not change the formal order of approximation.55) be exact up to terms of order O (kh)4 . We suggest to use this value in numerical simulations. Moreover. The dispersion relation of just obtained equations (2.8 we plot dispersion relations (2. (2. . (2. t Thus. In Figure 2. BPS85]).
2 0. version 2 . • Ch´zy law: e Sf = −Cf g where Cf is the Ch´zy coeﬃcient.8 0. one switches back to dimensional variables. Most frequently. We give here these models in dimensional form.5 0. So it is highly recommended to introduce these source terms in numerical models. e u u .6 0.3 0. The following terms have to be added to the source terms of Boussinesq equations when one wants to include bottom friction modelling.7 Bottom friction In this section.2. h+η . There is no unique bottom friction law. Both laws have similar structures.21 Jan 2008 0 0 1 2 3 4 kh 5 6 7 8 Figure 2.1 Euler equations Boussinesq−Nwogu model Regularized equations 107 cp(k)/(gh)1/2 tel00194763.7 Bottom friction Phase velocities with optimal value of α 1 0. Nwogu version of Boussinesq equations and its regularized version 2. Ch´zy and DarcyWeisbach e laws are used.4 0. since they have the form of source terms that do not involve diﬀerential operators.9 0.8: Dispersion relations of complete water wave problem. In the Boussinesq and nonlinear shallow water equations there is also a possibility to include some kind of empirical terms to model these physical eﬀects.7 0. It is a common practice in hydraulics engineering to take into account the eﬀect of bottom friction or bottom rugosity. From the mathematical and especially numerical viewpoints these terms do not add any complexity.
Then we derive corresponding long wave equations which take the following form in the absence of dispersive terms: ηt + ∇ · ((h + η)u) = 2ν∆η + ν π 0 t ∇·u √ dτ. Remark 7 In chapter 4 we develop an alternative approach to bottom friction modelling. spatial derivatives are computed with the following algorithm: . 8(h + η) where λ is the socalled resistance value. Sf = −k 2 g 4 .8 Spectral Fourier method In this study we adopted a wellknown and widely used spectral Fourier method. • ManningStrickler law: u u tel00194763.21 Jan 2008 (h + η) 3 where k is the Manning roughness coeﬃcient. The resulting viscouspotential formulation contains local and nonlocal dissipative terms. First.108 • DarcyWeisbach law: Dissipative Boussinesq equations λu u . Typically ks can vary from 1mm for concrete to 300mm for bottom with dense vegetation. which depends on the composition of the bottom. Here ks denotes the friction parameter. 2.84(h + η) . there are very powerful tools for the fast and accurate computation of discrete Fourier transforms (DFT). the diﬀerentiation operation in Fourier transform space is extremely simple due to the following property of Fourier transforms: f ′ = ikf . This parameter is determined according to the simpliﬁed form of the ColebrookWhite relation: Sf = − 1 √ = −2. Secondly. Our analysis is based on formal asymptotic expansion of NavierStokes equations in ﬂuid and in the bottom boundary layer. version 2 . The eﬀectiveness of this method is explained by two main reasons. 2 The principal diﬀerence with just presented engineer’s approach consists in at least formal justiﬁcation of the governing equations. So.03 log λ ks 14. t−τ 1 ut + ∇u2 + g∇η = 2ν∆u. The main idea consists in discretizing the spatial derivatives using Fourier transforms.
21 Jan 2008 where k is the wavenumber. Let us apply the Fourier transform to both sides of equations (2. Let us now consider the discretization of the dissipative Boussinesq equations. 1 ∆t . This approach. We chose equations (2. (2.57) constitute a system of ordinary diﬀerential equations to be integrated numerically. 1 k4 + O(∆t6 ). yn ). Equations (2.23). this type of method is appropriate.56) 2 1 1 ut + εiku2 + ikη + εν2 k2 u − µ2 ikh2 ∇ · ut = 0.25): 1 µ2 µ2 µ2 η t = −ik · (h + εη)u − ht − h2 ∇2 ht − h∇h · ∇ht + h3 ∇2 ∇ · u ε 2ε ε 6 µ2 2 2 + h ∇ h∇ · u + µ2 h ∇h2 ∇ · u + µ2 h2 ∇h · ∇(∇ · u). The second drawback is that we can only handle simple geometries. (2. N . . yn + k3 ) . (2. k1 = ∆tf (tn .25). In the present study we use the classical explicit fourthorder RungeKutta method.56) and (2. For the purpose of academic research. (2. yn + k1 k2 = ∆tf tn + 2 2 ∆t 1 k3 = ∆tf tn + .2. ky ) denotes the Fourier transform parameters. (2. which is extremely eﬃcient.25) in order to avoid cumbersome expressions and make the description as clear as possible. 6 Remark 8 A lot of researchers who integrated numerically the KdV equation noticed that the stability criterion has the form λ ∆t = 2 . The other systems are discretized in the same way. namely. We show in detail how the discretization is performed on equations (2. 1 1 1 yn+1 = yn + k1 + k2 + k3 + 6 3 3 3: f ′ ← iﬀt (v) . yn + k2 2 2 k4 = ∆tf (tn + ∆t. The ﬁrst drawback consists in imposing periodic boundary conditions since we use DFT. version 2 .8 Spectral Fourier method 1: f ← ﬀt (f ) 109 2: v ← ikf tel00194763.57) 2 2 where k = (kx . has the drawbacks of almost all spectral methods.23).23). Cartesian products of 1D intervals.
but it cannot ﬁx the dependence on 1/N 2 . We do not have this diﬃculty because we use regularized dispersive terms. one can easily handle nonlinearities.21 Jan 2008 (h + εη)u = ﬀt (h + ε Re iﬀt (η)) · Re iﬀt (u) . 2 We look for solitarywave solutions travelling to the left in the form η(x. . x0 are constants. The regularization eﬀect can be seen from equation (2.8.57). Unfortunately. In order to increase the time integration step ∆t they solved exactly the linear part of the partial diﬀerential equation since the linear term is the one involving high frequencies and constraining the stability. Let us brieﬂy explain how we treat the nonlinear terms.110 Dissipative Boussinesq equations where λ is the CourantFriedrichsLewy (CFL) number and N the number of points of discretization. With this simpliﬁcation several solitary wave solutions can be obtained. We follow closely the work of [Che98]. But for validation purposes.10).58) (2. The same idea was exploited by [BPS81]. Since c is constant (there is no acceleration). 6 (2. In other words we choose a new frame of reference which moves with the same celerity as the solitary wave.1 Validation of the numerical method One way to validate a numerical scheme is to compare the numerical results with analytical solutions.59) µ2 ut + ηx + εuux − uxxt = 0. 2. t) = Bη(ξ). the observer moving with the wave will see a steady picture. t) = η(ξ) = η(x0 + x + ct). This method. one can neglect the viscous term. For example the term (h + εη)u is computed as follows: tel00194763. allows an increase of the CFL number up to a factor ten. where we introduced the new variable ξ = x0 + x + ct and B. From the physical point of view this change of variables is nothing else than Galilean transformation. the authors did not succeed in deriving analytical solutions to the (1 + 1)D dissipative Boussinesq equations over a ﬂat bottom. In (1 + 1)D in the presence of a ﬂat bottom. who used the modiﬁed KdV equation (2. Since the time integration scheme is explicit. The other nonlinear terms are computed in the same way. c. the Boussinesq system without dissipation becomes ηt + ux + ε(uη)x − µ2 uxxx = 0. which is usually called the integrating factor method. u(x. version 2 .
21 Jan 2008 µ2 ε cu + η + u2 − c u′′ = 0. Compatibility conditions are obtained by comparing the coeﬃcients of corresponding terms in equations (2. The unique solution of those equations is B2 = Choosing B > 0 so that c > 0 leads to 6 B=√ . 15 1 c= √ .61): 1 2 1 B − Bc = 1. 6 These relations can be thought as a system of linear equations with respect to B 2 and Bc. t) = lim u(x. 15 12 .59) gives cη ′ + u′ + ε(uη)′ − cu′ + η ′ + εuu′ − c µ2 ′′′ u = 0.61) 2 2 In order to have nontrivial solutions both equations must be compatible. t) = 0. 6 µ2 ′′′ u = 0. since a priori the solution behaviour at inﬁnity is known: the solitary wave is exponentially small at large distances from the crest. 2 One can decrease the order of derivatives by integrating once: cη + u + εuη − µ2 ′′ u = 0. Mathematically it can be expressed as ξ→±∞ lim η(x.60)(2.58)(2.8 Spectral Fourier method 111 In the following primes denote derivation with respect to ξ. 6 (2. ξ→±∞ Now we use the relation u(ξ) = Bη(ξ) to eliminate the variable u from the system: (c + B)η − B (1 + cB)η − cB µ2 ′′ η = −εBη 2 . (2. version 2 . 2 2 The solution is integrable on R and there are no integration constants.60) µ2 ′′ ε η = − B 2η2. 6 . Substituting this special form into the governing equations (2. 6 tel00194763. 2 2 1 2 B − Bc = 0. 5 c= B .2.
Moreover. 4ε 2µ √ √ 7 7 15 2 u(x.63) Note that this exact solitary wave solution is not physical. β be real constants.62) yields the following solution: √ 7 7 2 η(x. the error ǫN is bounded below by the maximum between the error due to the time integration algorithm and ﬂoating point arithmetic precision. This result shows an excellent performance of this spectral method with an exponential convergence rate. In any case. .62) tel00194763. ˜ 1≤i≤N where {xi }1≤i≤N are the discretization points. 10ε 2µ (2. The exponential convergence rate to the exact solution is one of the features of spectral methods. one diﬀerentiates once equation (2. version 2 . Che98]): Lemma 1 Let α. It explains the success of these methods in several domains such as direct numerical simulation (DNS) of turbulence. The solution to this equation is wellknown (see for example [New77.61): 7η ′ − µ2 η ′′′ = −12εηη ′ .112 Dissipative Boussinesq equations These constants determine the amplitude and the propagation speed of the solitary wave. the equation αη ′ (ξ) − βη ′′′ (ξ) = η(ξ)η ′ (ξ) has a solitarywave solution if αβ > 0.9 shows the graph of ǫN as a function of N . Figure 2. T ) . At the end of the computations one computes the L∞ norm of the diﬀerence between the analytical solution (2. the solitarywave solution is η(ξ) = 3α sech 2 where ξ0 is an arbitrary constant. We choose a solitary wave as initial condition and let it propagate during a certain time T with the spectral method. Indeed the velocity is negative whereas one expects it to be positive for a depression wave propagating to the left. In general. the goal here is to validate the numerical computations by comparing with an exact solution. The methodology is simple. t) = − sech (x + ct + x0 ) . T ): ˜ ǫN := max η(xi .63) and the numerical one η (x.21 Jan 2008 Applying this lemma to equation (2. t) = − sech (x + ct + x0 ) . One of the main drawbacks of spectral methods consists in the diﬃculties in handling complex geometries and various types of boundary conditions. In order to ﬁnd the shape of the wave. 1 2 α (ξ + ξ0 ) β (2. T ) − η (xi .
64) h + (h − hl ) 1 + sin ∆x (x − x0 − 2 ∆x) . 2.2.2. r 2. Its classical construction is as follows. version 2 . The parameters used in this computation are given in Table 2. 1 π 1 h(x) = (2.9 Numerical results Error between exact and numerical solutions 113 10 0 10 −2 10 −4 εn 10 −6 10 −8 10 −10 tel00194763. x0 < x < x0 + ∆x.9: Error on the numerical computation of a solitary wave solution.24) and (2.9 Numerical results This test case is interesting from a practical point of view since it clearly illustrates the phenomena of long wave reﬂection by bottom topography. y) is given by hl . x ≤ x0 . Here T = 1.21 Jan 2008 10 −12 100 150 200 250 300 350 400 N − degrees of freedom 450 500 550 Figure 2. A typical function h(x.1 Construction of the initial condition We propagate on the free surface a socalled approximate soliton. We begin with the nondissipative Boussinesq equations on a ﬂat . All values are given in nondimensional form. they have been performed with the 2D version of the code.25). The bathymetry z = −h(x. In this section we perform comparisons between the two dissipation models (2. y) is chosen to be a regularized step function which is translated in the y−direction. l 2 r h. x ≥ x0 + ∆x. Even though the computations we show deal with a 1D wave propagating in the negative x−direction.9.
6 Since the water depth is h = 1 + εη = 1 + O(ε). 2 2 and look for u in the following form: ηt + (1 + εη)u − x bottom: (2. Replacing time derivatives by spatial ones in (2.005 µ ν 1 . 6 µ2 ε ut + ηx + (u2 )x − uxxt = 0.69) yields 1 Px = ηηx . 3 By integration (using the fact that solitary waves tend to zero at inﬁnity).68)(2. P.68) (2. one obtains 1 P = η2.3 Dissipative Boussinesq equations ε 0.5 ∆x 0.114 parameter hl value 0. Consequently one has the following relations: ∂f ∂f = + O(ε + µ2 ).5 hr x0 1. ∂t ∂x f ∈ {η.06 0. Q}. 2 1 Qx = − ηxxx . 4 1 Q = − ηxx 3 (2. 6 (2. One substitutes this asymptotic expansion into the governing equations and retains only the terms of order O(ε + µ2 ): ηt − ηx + εPx + µ2 Qx − 2εηηx + −ηt + ηx + εPt + µ2 Qt + εηηx + µ2 ηxxx = O(ε2 + εµ2 + µ4 ).14 Table 2.69) 1 µ2 : Qx + Qt + 1 ηxxx + 2 ηxxt = 0. (2.65) tel00194763. the approximate solitary wave should travel to the left with a celerity c = 1+O(ε) and depend on the variable x+ct = x+t+O(ε).67) µ2 ηxxt = O(ε2 + εµ2 + µ4 ). (2. 2 Add these two equations and set the coeﬃcients of ε and µ2 equal to 0: ε: Px + Pt − ηηx = 0.70) .2: Typical values of the parameters used in the numerical computations µ2 uxxx = 0.21 Jan 2008 u = −η + εP + µ2 Q + O(ε2 + εµ2 + µ4 ). version 2 .66) It is precisely at this step that one makes an approximation.0 −0. ν2 0.
24) and (2. It can be explained by the presence of the term ν1 Su which is bigger in magnitude than εν2 ∇2 u.72) tel00194763. t0 ) (divided by 10 for clarity’s sake) during and just after reﬂection by the step is given on Figure 2.2 Comparison between the dissipative models The snapshot of the function η(x.25). This approximate soliton is used in the numerical computations.13 due to the interaction with bathymetry. the wave amplitude grows. . 115 (2. 4 3 Substituting this expression for u into (2. First of all. version 2 . In particular we see an important drawback of the dissipation model I: just after the wave crest the free surface has some kind of residual deformation which is clearly nonphysical.10. The velocity u is obtained from (2. We could hardly predict this eﬀect directly from the equations without numerical experiments. the wave amplitude before the interaction is equal exactly to 8 (without including dissipation) and over the step it becomes roughly 9.4.12 one can see that diﬀerences between the two solitons continue to grow. Then we compare the two sets of equations (2. there is almost no diﬀerence between the conservative case and model II. . (2. To do so we look at the section of the free surface at y = 0 along the propagation direction.11 shows that even at the beginning of the computations the two models give slightly diﬀerent results. The amplitude of the pulse obtained with model I is smaller. . 2. We would like to point out several soliton transformations in Figure 2. Within graphical accuracy. (2. Figure 2.23). In Figure 2. since the depth decreases.2. Quantitatively speaking. .71) (2. On the other hand the soliton becomes less symmetric which is also expected. Because of periodic numerical boundary conditions we also observe the residue of the freesurface deformation coming through the left boundary.71) by simple substitution. Our numerical experiments show that the amplitude of this residue depends almost linearly on the parameter ν1 .66) connecting η and u becomes µ2 ε u = −η + η 2 − ηxx + .9 Numerical results and the relation (2.9.23).21 Jan 2008 where c > 1. y. 2 6 which admits solitary wave solutions of the form η = η(x + ct): η(x.67) yields a classical KdV equation for η: 3 µ2 ηt − 1 + εη ηx − ηxxx = 0. t) = 2(c − 1) sech 2 ε 1 2µ 6(c − 1)(x + ct) .
0) 5 4 3 2 1 0 −1 −3 −2 −1 0 x 1 2 3 −0.21 Jan 2008 (a) t = 2.1 t = 0.1s (b) t = 2.11: Freesurface snapshot before the interaction with the step: (left) the curves corresponding to the three models are almost superimposed.02 0 −0.10: Interaction between a leftrunning solitary wave and a step.06 1 2 6 η(x.02 −0.4 0. version 2 .400 10 9 8 7 No dissipation Model 1 Model 2 0. (right) diﬀerence between model II and model I. Difference between model 2 and 1 at t = 0.116 Dissipative Boussinesq equations tel00194763.5s Figure 2.08 −4 −3 −2 −1 0 x 1 2 3 4 Figure 2.04 η −η 0.08 0.06 0.04 −0. .
0) 5 4 3 2 1 No dissipation Model 1 Model 2 tel00194763. version 2 .13: Beginning of the solitary wave deformation under the change in bathymetry .0) 5 4 3 2 1 0 −1 −3 −2 −1 0 x 1 2 3 No dissipation Model 1 Model 2 Figure 2.100 10 9 8 7 6 η(x.2.100 10 9 8 7 6 η(x.9 Numerical results 117 t = 1.12: Free surface just before the interaction with the step t = 2.21 Jan 2008 0 −1 −3 −2 −1 0 x 1 2 3 Figure 2.
14: Initiation of the reﬂected wave separation t = 2.0) 5 4 3 2 No dissipation Model 1 Model 2 tel00194763.800 10 9 8 7 6 η(x.118 Dissipative Boussinesq equations t = 2.250 10 9 8 7 6 η(x.15: Separation of the reﬂected wave . version 2 .0) 5 4 3 2 1 0 −1 −3 −2 −1 0 x 1 2 3 No dissipation Model 1 Model 2 Figure 2.21 Jan 2008 1 0 −1 −3 −2 −1 0 x 1 2 3 Figure 2.
in which the decay is proportional to the second derivative of the velocity. We refer one more time to the theoretical as well as experimental work of [BPS81]. Moreover our dissipative Boussinesq equations are in the same relationship with the classical Boussinesq equations [Per67] as Euler and NavierStokes equations.16 show the process of wave reﬂection from the step at the bottom. This is a second argument towards the physical pertinency of the results obtained with model . 2.21 Jan 2008 0 −1 −3 −2 −1 0 x 1 2 3 Figure 2. Their work shows excellent agreement between experiments and numerical solutions to dissipative KdV equation. appears to be better. there is an indirect evidence. At this stage we cannot show comparisons with laboratory experiments in order to demonstrate the performance of model II.200 10 9 8 7 6 η(x.0) 5 4 3 2 1 No dissipation Model 1 Model 2 119 tel00194763. they added to the Korteweg–de Vries equation an adhoc dissipative term in the form of the Laplacian (but in 1D). version 2 . Nevertheless. In order to model wave trains.10 Conclusions Comparisons have been made between two dissipation models.2.14. The reﬂected wave clearly moves in the opposite direction.15 and 2. Model II.10 Conclusions t = 3.16: Two separate waves moving in opposite directions Figures 2. This term coincides with the results of our derivation if we model dissipation in the equations according to the second model. 2.
120 II.21 Jan 2008 . Dissipative Boussinesq equations tel00194763. version 2 .
. . . . . . . . . .1 3.1 3. . . . 138 Diﬀusive ﬂuxes computation . .5 Sign matrix computation . . . . . . . 137 TVD and MUSCL schemes . . . . [. version 2 . . . .2. . . . and if one has really a sound insight. . . . . because the discrepancy may well be due to minor features that are not properly taken into account and that will get cleared up with further development of the theory. . . . . . . . . . . . .1 3. . . . . . If there is not complete agreement between the results of one’s work and experiment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Paul Adrien Maurice Dirac (1902 – 1984) Contents 3. 129 Finite volume scheme on unstructured meshes . . . . . 144 Solution interpolation to mesh nodes . . . . . 126 Equation of state . .4 3. . . . . . . . . . . . namely that it is more important to have beauty in one’s equations that to have them ﬁt experiment. . . . . . . . . one should not allow oneself to be too discouraged. . . 127 Formal limit in barotropic case . . . 146 121 . . . . . . . . . .3 3. . . . . . . one is on a sure line of progress. . . . . . .4.] It seems that if one is working from the point of view of getting beauty in one’s equations. . . . .2 3. . . .21 Jan 2008 I think that there is a moral to this story. . . . . . . . 122 Mathematical model . . . . . . .2. .4. . . 125 3. . . . . . . . . . 131 3. .. . 136 Second order scheme . . . . . . . . . . . . .. . . . . . . . . . . . . . . .4. .4. . . . . . . .4 Sound speed in the mixture .4. . .2 Introduction .3 3. . . . . .2 3. . . . . . . . .Chapter 3 Two phase approach to free surface compressible ﬂows tel00194763. . . . .
.5 Two phase ﬂows Time stepping methods . 157 Water drop test case . . . our contribution consists in suggesting to use this model for the numerical simulation of freesurface aerated ﬂows in various containers. . . . . Finally. . Namely. . . . . . . . . . . . .1 3. . . . . . Twophase ﬂows in nature include many interesting examples such as clouds. . . .5. . . pressure and temperature for both phases). In the present study we do not take into account this eﬀect since in our applications it is unimportant. . .7 3. .5. . . . . . . . . . . . . . version 2 . . we are going to use the homogeneous model (with equal velocity. . . . . . . . . . . . . . . . . fountains and gas/oil slicks. . Several features make twophase ﬂow an interesting and challenging branch of ﬂuid mechanics. . . .122 3. . . foam. Another important application of twophase ﬂows is pump cavitation.21 Jan 2008 Introduction Twophase or even multiphase ﬂows are very common in nature. In the present chapter we are going to present a mathematical model which is issued from twophase ﬂows ﬂuid mechanics. . . . .4. . which is signiﬁcantly diﬀerent from the singlephase case. . 151 Numerical results . The sound speed changes dramatically for materials undergoing phase change. . . .4. .6 3. . . . . . 162 3. . Then. . . Similar diﬀerences are typical of water liquid/water vapor densities and can make the computation very stiﬀ. . . . . They occur in a system containing gas and liquid1 with a meniscus separating the two phases. . . . .5. . . . . . the phase changes are not instantaneous. . . . . . Basically. . . . . waves on the sea. . . . . rain. nuclear reactors use water to remove heat from the reactor core using twophase ﬂow. . . . . . bubbles.3 Convergence test . and can be orders of magnitude diﬀerent (see Figure 3. . 153 3. . . . . Probably. . . the most important applications of twophase ﬂow are in largescale power systems. and the liquid vapor system will not necessarily be in phase equilibrium. . . . . . . . . groundwater ﬂow. .2 3. . in the case of air and water at standard temperature and pressure. . . . . . Even more critically. . . . . . 148 Boundary conditions treatment . . . First of all. . . . . This introduces strong compressible eﬀects into the problem. 154 Falling water column . . . . . . the density of the two phases diﬀers by a factor of about 1000. . .1 tel00194763. . . . . . . surface tension complicates considerably all dynamical problems.6 Conclusions .2). The most important applications are issued from petroleum 1 They can occur even in a system containing two diﬀerent liquids as well. . . The design of power stations requires a detailed understanding of twophase ﬂow heattransfer and pressure drop behaviour. . . . 157 3.
They are produced when the front of the breaker is almost parallel to the wall at the instant of impact. many engineers disregard wave impact loading in the belief that the duration of the impulse is too short for the structure to respond.1 shows an image of a broken wave. This image is taken from [BOPB07]. Figure 3. We refer to [Per03] as general excellent review of waterwave impact problem. If the wave overturns as it hits the wall. The authors report that even for series of supposedly regular waves. version 2 .3. Accurate predictions of wave loading are crucial to the design of coastal structures. Interesting experimental results are obtained in [BOPB07]. When a wave breaks directly onto a vertical or inclined wall.1: Example of homogeneous water/air mixture just after the wave breaking. industry and simulation of wave impacts onto a wall or breakwater. impact pressures can be extremely large in comparison to the pressures exerted by nonbreaking waves. The compressibility of the trapped or entrained air will aﬀect the dynamics and is often thought to reduce the maximum pressure due to a cushioning eﬀect. an air pocket will . However. Presently. there is great variability in the violence of the impacts. the loading due to breaking waves is diﬃcult to predict and is poorly understood in general. It gives an example of homogeneous water/air mixture and shows the physical situation which we aim to model. Despite all this.1 Introduction 123 tel00194763. The greatest impact pressures are highly localized in space and time.21 Jan 2008 Figure 3. it can trap an air pocket.
. sometimes accompanied by clear evidence of entrapment. the impulse associated with the impact may well be higher. It is plotted as a function of gas void fraction α.2 0. also tend to distribute the impact pressures more widely so that the overall force on the wall may not be reduced. It suggests that the greater persistence of air in seawater waves may not help to protect structures from impact damage as is often assumed. gas void fraction 0. On the other hand. Wave impacts are classiﬁed as lowaeration when the measurements indicate that the water adjacent to the wall contains relatively little air (typically a voids ratio < 5%). version 2 . [BOPB07] showed that lowaeration impacts have temporally and spatially localised pressure spike of short rise time and duration (≈ 80 to 200 ms) while in highaeration situation the rise time. Even when the pressures during a highaeration impact are lower. When this is combined with the longer duration.21 Jan 2008 0 0 0.8 1 Figure 3.2: Sound speed cs in the water/air mixture at normal conditions. m/s s 500 tel00194763. fall time and duration are longer (≈ 100 to 400 ms) and the pressure spike is less localised. an impact is called a highaeration impact if it contains a higher level of entrained air. Bullock et al.124 Sound velocity in the water/air mixture 1500 Two phase ﬂows 1000 c .4 0.6 α. the fact that the impact is generally less spatially localised than a lowaeration impact reduces the chance of the resultant force being lower.
we need to provide the socalled equation of state (EOS) p = p± (ρ± . p where ρ := α+ ρ+ + α− ρ− (the total density). For example. u. u.3). It is possible to rewrite these equations as a system of balance laws ∂w + ∇ · F(w) = S(w).2 the dependence of sound velocity in the mixture of water and air as a function of gas volume fraction at normal conditions. for example). We use superscripts ± to denote any quantity which is related to liquid and gas respectively. E. e). if the problem under consideration does not require explicitly the compressibility (like in acoustics. for example).1) . this eﬀect can usually be neglected since it complicates considerably the study.2) (3. version 2 . e.3. p and E and only four governing equations (3. + ∇ · ρHu = ρg · u. α+ and α− denote the volume fraction of liquid and gas and obviously satisfy the condition α+ + α− = 1. Then. The reader can see that the sound velocity drops down very quickly (to about 50 m ) when we add a small quantity of air and the Mach number is not s so small anymore.1) (3. we depict on Figure 3. ∂t More precisely.2. If the liquid is pure (water. The above system contains ﬁve unknowns α± ρ± . p. the importance of its compressibility is questionable2 .3) (ρu)t + ∇ · ρu ⊗ u + pI = ρg. we have the following classical quantities: ρ± . the velocity ﬁeld vector.2 Mathematical model tel00194763. we need to introduce the notation which will be used throughout this chapter. The construction of the EOS will be discussed below (see Section 3. momentum and energy lead to the four following equations: (α± ρ± )t + ∇ · (α± ρ± u) = 0. In this section we present the equations which govern the motion of two phase mixtures in a computational domain Ω. E = e + 1 u2 (the total energy).1). g which denote the density of each phase. Recall that according to [BOPB07] lowaerated impacts contain about 5% of air.2).21 Jan 2008 We would like to begin this section by explaining our choice of a compressible model for the liquid phase. In the present work we aim to model the mixtures of gas and liquid which typically occur in oil containers and broken waves (see Figure 3. ρE t (3. H := E + ρ (the speciﬁc enthalpy). In order to close the system.2 Mathematical model 125 3. 2 .(3. the internal and total energy and the acceleration due to gravity correspondingly. thus dropping down signiﬁcantly the sound speed. Conservation of mass (one equation for each phase). the pressure. First of all. In order to show the importance of compressibility eﬀects. This system can be seen as the single energy and inﬁnite drag 2 limit of the more conventional six equations model [Ish75].
α− ρ− . ρu.3) is hyperbolic. One can conclude that the system (3.1) – (3.126 where the conservative variables in the 2D case are deﬁned as follows: w = (wi )5 := (α+ ρ+ . ρHun ) = w3 n 1 + w4 n 2 w3 n 1 + w4 n 2 w3 n 1 + w4 n 2 . i=1 Two phase ﬂows (3.2. Its expression in the physical variables is un c − −un c+ c + n1 c + n2 0 −un c− un c + c − n1 c − n2 0 ∂p ∂p ∂p ∂p ∂p un2 + ∂w4 n1 n1 An = −uun + ∂w1 n1 −uun + ∂w2 n1 un + un1 + ∂w3 n1 ∂w5 ∂p ∂p ∂p ∂p ∂p −vun + ∂w n2 −vun + ∂w n2 vn1 + ∂w3 n2 un + vn2 + ∂w4 n2 n ∂w5 2 1 2 ∂p ∂p ∂p ∂p ∂p un ∂w1 − H un ∂w2 − H un ∂w3 + Hn1 un ∂w4 + Hn2 un 1 + ∂w5 This matrix has three distinct eigenvalues: λ 1 = un − c s . The ·n)(w) jacobian matrix An (w) := ∂(F∂w can be easily computed. λ2. version 2 . ρuun + pn1 .4) The ﬂux projection on the normal direction n = (n1 . . where cs is deﬁned in the next section. In the present setting they are deﬁned by s formulas (3. ρvun + pn2 .3.21 Jan 2008 where un := u · n = un1 + vn2 is the velocity projection on the normal direction n. λ 5 = un + c s . The computation of the eigenvectors is trickier but can still be performed analytically. n2 ) can be expressed in physical and conservative variables F · n = (α+ ρ+ un . ρv. ρ(α+ ρ− (c− )2 + α− ρ+ (c+ )2 ) s s where c± are the sound velocities in each phase.4 = un . w2 . α− ρ− un . w3 + pn1 .5) tel00194763. ρE). This hyperbolicity represents the major advantage of this model. We do not give here the ﬁnal expressions since they are cumbersome. 3. w1 w1 + w2 w1 + w2 w1 + w2 w3 n 1 + w4 n 2 w3 n 1 + w4 n 2 + pn2 .1 Sound speed in the mixture In [Ghi08] it is shown by direct computation of the eigenvalues of the Jacobian An (w) that the square of the sound speed in the mixture is given by c2 := s ρ+ ρ− (c+ )2 (c− )2 s s .8). (w5 + p) w4 w1 + w2 w1 + w2 (3.
π0 are constants. pure water is well described v when we take N = 7 and π0 = 2. the constants c± can be calculated after simple algebraic manipulav tions of equations (3. we make the additional assumption that the two phases are in thermodynamic equilibrium: p+ = p− .3. e+ = c+ T + + v π0 . v (3. N .6) while the heavy ﬂuid is modeled by Tate’s law3 [GZI+ 79. HA71] p+ + π0 = (N − 1)ρ+ e+ . .21 Jan 2008 For example. The sound velocities in each phase are given by the following formulas: (c− )2 = s γp− . T + = T −. To do this. ρ In the literature Tate’s law is sometimes called the stiﬀened gas law. (3. For example.7) and matching with experimental values at normal conditions: c− ≡ v c+ ≡ v p0 .8) N p0 + π0 . we ﬁrst introduce the following notation for the mass fractions of each phase in the mixture c± := 3 α ± ρ± .1 × 109 P a. (N − 1)N ρ+ T0 0 In order to construct an equation of state for the mixture. Now. (3. ρ+ (3. (γ − 1)ρ− T0 0 tel00194763.2.7) where the quantities γ.2 Equation of state In the present work we assume that the light ﬂuid is described by an ideal gas type law p− = (γ − 1)ρ− e− .2 Mathematical model 127 3. version 2 .6). let us construct a system of nonlinear equations which will help us obtain the pressure in the mixture. N ρ+ (3. ρ− (c+ )2 = s N p+ + π0 . c± . e− = c− T − . for an air/water mixture under normal conditions we have the values given in Table 3.1. Remark 9 In practice.9) Below values of the common pressure and common temperature will be denoted by p and T respectively.
After substituting the expressions (3. it does not correspond to any physical situation. We would like to reduce the number of unknowns. Nevertheless. version 2 . From (3.72 kg·K J 646. .9) we have an additional relation p ≡ p+ (ρ+ .1: Values of the parameters for an air/water mixture under normal conditions. π0 e+ e− − − = . (3. we choose this value in order to accelerate all dynamic processes in our test cases.4 7 2.21 Jan 2008 Table 3. T + (ρ+ .10) (3.7) for T ± . we express the internal energies e± of each phase as a function of the other thermodynamic variables and constants.0 kg·K 100 m/s2 Two phase ﬂows tel00194763. D c − N c + ρ+ v v 1 π0 c+ . cv v Let us analyse the obtained results. We would like to comment on the rather high value of the acceleration due to gravity.11) c where D := c− + c+ . e− ) .1 × 109 P a J 166. + cv cv N c + ρ+ v This linear system can be easily solved to give us the expressions for e± e+ = e− = + − 1 e π0 c− + . e+ ) = p− (ρ− .6).29 kg/m3 300 K 1. e− Dc+ N ρ+ v (3. Let us write the two following relations: c+ e+ + c− e− = e. e− ). To do this. e+ ) = T − (ρ− .128 parameter p0 ρ+ 0 ρ− 0 T0 γ N π0 c+ v c− v g value 105 P a 103 kg/m3 1. Obviously. our system becomes c+ e+ + c− e− = e.
The superscript + is used to . one can express ρ− in terms of ρ+ and the conservative variables w1. It is done by solving the following system of nonlinear equations: (N − 1)ρ+ e ρ− π0 c− π0 c+ − (γ − 1) + e − = Dπ0 . ρ+ ρ− Finally. α + ρ+ = w1 .3 Formal limit in barotropic case 129 It means that the pressure p is completely determined if we know (ρ+ . e− ).3 + Formal limit in barotropic case The propagation of waves at the interface between a heavy compressible ﬂuid of density ρ and a light compressible ﬂuid of density ρ− is investigated. ρ± ). version 2 .3. Once ρ+ is determined.10) and. + c − N c + ρ+ cv N ρ+ v v w1 w2 + = 1.21 Jan 2008 (N − 1)ρ+ ρ− π0 c− π0 c+ e − (γ − 1) + e − = Dπ0 . (3. 3.11).2 as ρ+ w2 ρ− = + . using the second equation. it is suﬃcient to determine these four quantities (α± . the pressure p = (N − 1)ρ+ e+ − π0 . e+ ) or (ρ− . The last two equations can be easily eliminated to reduce the system to two equations tel00194763. cv N cv This equation can be solved analytically (we do not give here the solution’s expression since it is cumbersome) or by a Newtontype numerical method. + c − N c + ρ+ cv N ρ+ v v α+ + α− = 1. ρ± ) by expressions (3. ﬁnally. α − ρ− = w2 . Since internal energies e± are expressed in terms of (α± . ρ − w1 This expression is then substituted in the pressure equality condition to give the following single equation with respect to ρ+ : (N − 1) (N − 1)π0 c− + e + + ρ (ρ − w1 ) + (ρ − w1 ) c− N c+ v v e π0 c+ − (γ − 1) + ρ+ w2 + (γ − 1) + w2 = Dπ0 (ρ+ − w1 ).10). we can ﬁnd e+ from (3.
(3. δ±γ ∇ · u = 0. ρ := α+ ρ+ + α− ρ− and the equation of state reads: p ≡ p+ (ρ+ ) = p− (ρ− ) .3. with the relations α+ + α− = 1.3: Sketch of the ﬂow. The governing equations are (α± ρ± )t + ∇ · (α± ρ± u) = 0 .15) .13) where u = (u.14) The speeds of sound are given by (c± )2 = s Let γ± = Let also δ±γ = γ ± − γ ∓ . ρ± the densities of each ﬂuid. A typical geometry is shown in Figure 3. w) is the common velocity. (ρu)t + ∇ · (ρu ⊗ u) + ∇p = ρg .12) (3. z z = η(x. α± are the volume fractions of each ﬂuid. p γ = α+ γ − + α− γ + . y tel00194763. ∂ρ± ρ± (c± )2 s . γ (3. v.21 Jan 2008 Figure 3. (3.130 Two phase ﬂows denote the heavy ﬂuid while the superscript − is used to denote the light ﬂuid. t) −: light ﬂuid 0 +: heavy ﬂuid x. Then one can show that ± αt + (u · ∇)α± + α+ α− ∂p . version 2 .
version 2 . In electromagnetism.16) (3. Discontinuous Galerkin methods have proven to be more eﬃcient [CLS04].4 Finite volume scheme on unstructured meshes γ +γ − ∇ · u = 0. ∂t 4 (3. we can show that our model automatically degenerates into a discontinuous twoﬂuid system where two pure phases are separated by an interface. When compared to other discretization methods such as ﬁnite elements or ﬁnite diﬀerences.4 Finite volume scheme on unstructured meshes Finite volume methods are a class of discretization schemes that have proven highly successful in solving numerically of a wide class of conservation law systems. 131 (3. momentum and total energy4 . Equations (3. This simple computation shows an interesting property of our model.12) and (3. the primary interests of ﬁnite volume methods are robustness. Hence. . t)) . Physically this substitution means that we consider two pure phases separated by an interface.3. with this type of discretization.18) This statement is true in the absence of source terms and appropriate boundary conditions.1) – (3. ut + (u · ∇)u + ρ This system of equations is nothing else than the system of a discontinuous twoﬂuid system with an interface at z = η(x.21 Jan 2008 ρt + (u · ∇)ρ + ρ∇ · u = 0 .3) we rewrite the governing equations in the following form ∂w + ∇ · F(w) = ∇ · (D∇w) + S(w). for example. t). More precisely let pt + (u · ∇)p + p α− := H(z − η(x. These systems often come from compressible ﬂuid dynamics. we conserve “exactly” the mass. γ Consider the case where α− is either 0 or 1. 3. v) · ∇x η = w . and the intrinsic local conservation properties.17) where H is the Heaviside step function. applicability on very general unstructured meshes.15) for α− gives ηt + (u. For the important special case of barotropic equations. ∇p = g. In order to solve numerically the system of balance laws (3. Substituting the expression (3. It follows that α+ α− = 0 .13) become tel00194763.17) into the equation (3.
132
Two phase ﬂows
where w(x, t) : Rd × R+ → Rm is the vector of conservative variables (in the present study d = 2 or 3 and m = 5), F(w) is the advective ﬂux function, D the diﬀusion matrix and S(w) the source term. We must say that the governing equations (3.1) – (3.3) considered in this chapter do not possess viscous terms. However, we prefer to include in the discussion their discretization since it can be useful for NavierStokes or viscous SaintVenant and Boussinesq equations [DD07a]. The system (3.18) should be provided with initial condition w(x, 0) = w0 (x) and appropriate boundary conditions. nKL (3.19)
tel00194763, version 2  21 Jan 2008
L O K ∂K
Figure 3.4: An example of control volume K with barycenter O. The normal pointing from K to L is denoted by nKL .
The computational domain Ω ⊂ Rd is triangulated into a set of non overlapping control volumes that completely cover the domain. Let T denote a tesselation of the domain Ω with control volume K such that ¯ ¯ ∪K∈T K = Ω, ¯ K := K ∪ ∂K.
For two distinct control volumes K and L in T , the intersection is either an edge (2D) or face (3D) with oriented normal nKL or else a set of measure at most d − 2 (in 2D it is just a vertex, in 3D it can also be a segment, for example). We need to introduce the following notation for the neighbourhood of K N (K) := {L ∈ T : area(K ∩ L) = 0} ,
3.4 Finite volume scheme on unstructured meshes
133
a set of all control volumes L which share a face (or an edge in 2D) with the given volume K. In this chapter, we denote by vol(·) and area(·) the d and d − 1 dimensional measures5 respectively. The choice of control volume tesselation is ﬂexible in the ﬁnite volume method. In the present study we retained a socalled cellcentered approach (see Figure 3.5), which means that degrees of freedom are associated to cell barycenters. There exists an alternative vertexcentered method [BJ89, BO04] (see Figure 3.6) which necessitates the construction of dual mesh even for ﬁrstorder schemes. In the cellcentered ﬁnite volume scheme, the triangles themselves serve as control volumes with solution unknowns attributed to triangles barycenters. In the vertexcentered ﬁnite volume scheme, control volumes are formed as a geometric dual to the triangle complex and solution unknowns stored on vertex basis. Remark 10 Except for the construction of dual mesh in vertexcentered approach, these two methods are almost equivalent in the interior of the computational domain Ω. However, the boundary conditions treatment is diﬀerent and it is harder (or less natural in author’s opinion) when data is stored at vertices. This is one more reason why we retained the cell centers to store the solution’s information.
tel00194763, version 2  21 Jan 2008
Storage location Control volume
Figure 3.5: Illustration for cellcentered ﬁnite volume method
The ﬁrst steps in Finite Volume (FV) methods are classical. We start by integrating equation (3.18) on the control volume K (see Figure 3.4 for illustration) and we apply GaussOstrogradsky theorem for advective and diﬀusive ﬂuxes. Then, in each control
In other words, in 3D the notation area(·) and vol(·) are very natural and mean area and volume respectively, while in 2D they refer to the area and the length.
5
134
Two phase ﬂows Storage location Control volume
Figure 3.6: Illustration for vertexcentered ﬁnite volume method
tel00194763, version 2  21 Jan 2008
volume, an integral conservation law statement is imposed: d dt w dΩ +
K ∂K
F(w) · nKL dσ =
∂K
(D∇w) · nKL dσ +
K
S(w) dΩ .
(3.20)
Physically an integral conservation law asserts that the rate of change of the total amount of a substance (for example: mass, momentum, total energy, etc) with density w in a ﬁxed control volume K is balanced by the ﬂux F, diﬀusion D of the substance through the boundary ∂K and the production of this quantity S inside the control volume. The next step consists in introducing the socalled control volume cell average for each K∈T 1 wK (t) := w(x, t) dΩ . vol(K) K After the averaging step, the ﬁnite volume method can be interpreted as producing a system of evolution equations for cell averages, since d dt w(x, t) dΩ = vol(K)
K
dwK . dt
Godunov was the ﬁrst [God59] who pursued and applied these ideas to the discretization of the gas dynamics equations. However, the averaging process implies piecewise constant solution representation in each control volume with value equal to the cell average. The use of such representation renders the numerical solution multivalued at control volume interfaces. Thereby the calculation of the ﬂuxes ∂K (F(w)·nKL ) dσ at these interfaces is ambiguous. Next fundamental aspect of ﬁnite volume methods is the idea of substituting the true ﬂux at interfaces by a
3.4 Finite volume scheme on unstructured meshes numerical ﬂux function F(w) · n
∂K∩∂L
135
←− Φ(wK , wL ; nKL ) : Rm × Rm → Rm ,
a Lipschitz continuous function of the two interface states wK and wL . The heart of the matter in ﬁnite volume method consists in the choice of the numerical ﬂux function Φ. In general this function is calculated as an exact or even better approximate local solution of the Riemann problem posed at these interfaces. In the present study we decided to choose the numerical ﬂux function according to VFFC scheme extensively described in Appendix A. The numerical ﬂux is assumed to satisfy the properties: Conservation. This property ensures that ﬂuxes from adjacent control volumes sharing an interface exactly cancel when summed. This is achieved if the numerical ﬂux function satisﬁes the identity Φ(wK , wL ; nKL ) = −Φ(wL , wK ; nLK ). Consistency. The consistency is obtained when the numerical ﬂux with identical state arguments (in other words it means that the solution is continuous through an interface) reduces to the true ﬂux of the same state, i.e. Φ(w, w; n) = (F · n)(w). After introducing the cell averages wK and numerical ﬂuxes into (3.20), the integral conservation law statement becomes dwK + dt area(L ∩ K) Φ(wK , wL ; nKL ) = vol(K) = 1 vol(K) (D∇w) · nKL dσ + 1 vol(K) S(w) dΩ .
tel00194763, version 2  21 Jan 2008
L∈N (K)
L∈N (K)
K∩L
K
The discretization of diﬀusive ﬂuxes will be discussed in Section 3.4.4. We denote by SK 1 the approximation of the following quantity vol(K) K S(w) dΩ. In practice, the source term discretization is discussed in Section A.1.3. Thus, the following system of ordinary diﬀerential equations (ODE) is called a semidiscrete ﬁnite volume method: dwK + dt area(L ∩ K) Φ(wK , wL ; nKL ) − (D∇w) · nKL = SK , vol(K) ∀K ∈ T . (3.21)
L∈N (K)
136
Two phase ﬂows
The initial condition for this system is given by projecting (3.19) onto the space of piecewise constant functions 1 w0 (x) dΩ . wK (0) = vol K K This system of ODE should also be discretized. There is a variety of explicit and implicit n time integration methods. Let wK denote a numerical approximation of the cell average solution in the control volume K at time tn = n∆t. The simplest time integration method is the forward Euler scheme n+1 n dwK ∼ wK − wK . = dt ∆t When applied to (3.21) it produces the fullydiscrete ﬁnite volume scheme. The time integration approach used in this study is detailed in Section 3.4.6.
tel00194763, version 2  21 Jan 2008
3.4.1
Sign matrix computation
In the context of the VFFC scheme (see Appendix A for more details), we need to compute the socalled sign matrix which is deﬁned in the following way Un := sign(An ) = R sign(Λ)L, where R, L are matrices composed of right and left eigenvectors correspondingly, and Λ = diag(λ1 , . . . , λm ) is the diagonal matrix of eigenvalues of the Jacobian. This deﬁnition gives the ﬁrst “direct” method of sign matrix computation. Since the advection operator is relatively simple, after a few tricks, we can succeed in computing analytically the matrices R and L. For more complicated twophase models it is almost impossible to perform this computation in closed analytical form. In this case, one has to apply numerical techniques for eigensystem computation. It turns out to be costly and quite approximative. In the present work we use physical information about the model in numerical computations. There is another way which is less expensive. The main idea is to construct a kind of interpolation polynomial which takes the following values P (un ± cs ) = sign(un ± cs ), P (un ) = sign(un ).
These three conditions allow us to construct a second degree interpolation polynomial. Obviously, when P (λ) is evaluated with λ = An we obtain the sign matrix Un as a result. The construction of Lagrange interpolation polynomial P (λ) is simple and we give only its
4. we can always make the mesh spacing extremely small but it cannot be a solution since it makes the scheme ineﬃcient. we understand that in fact. whenever possible we suggest to use the computation of the Jacobian eigenstructure. it seems that we can expect to be ﬁrst order accurate at most. They proposed a kind of posttreatment procedure currently known as solution reconstruction or MUSCL6 scheme. It can lead to instabilities and diminish overall code robustness.4. Godunov has shown [God59] that all linear schemes that preserve solution monotonicity are at most ﬁrst order accurate. On the other hand.E. we have only one degree of freedom per data storage location. Kolgan [Kol72. Kol75] and van Leer [vL79].3. This rather negative result suggests that a higher order accurate scheme has to be essentially nonlinear in order to attain simultaneously a monotone resolution of discontinuities and high order accuracy in continuous regions. the authors often cite the paper by van Leer [vL79].1 Historical remark In general. 3. Of course.2. version 2 . Thus. However. From the theoretical point of view the situation is even worse since 1 an O(h 2 ) L1 norm error bound for the monotone and Eﬂux schemes [Osh84] is known to be sharp [Pet91]. although an O(h) solution error is routinely observed in numerical experiments. A signiﬁcant breakthrough in the generalization of ﬁnite volume methods to higher order accuracy is due to N. when we read numerical articles which use the MUSCL scheme. In the above papers the authors used linear reconstruction (it will be retained in this study as well) but we must say that this method was already extended to quadratic approximations in each cell [BF90]. It is commonly believed in the scientiﬁc commu6 MUSCL stands for monotone upstreamcentered scheme for conservation laws. it shows a rather bad numerical behaviour.21 Jan 2008 Second order scheme If we analyze the above scheme.4 Finite volume scheme on unstructured meshes expression P (λ) = 137 sign(un + cs ) − 2 sign(un ) + sign(un − cs ) (λ − un )+ 2c2 s sign(un ) − sign(un − cs ) (λ − un + cs ) + sign(un − cs ). . when we approach pure phase states. Our experience shows that the interpolation method is quicker and gives correct results in most test cases. Hence. In the numerical community ﬁrst order schemes are generally considered to be too inaccurate for most quantitative calculations.2 tel00194763. 3. cs In our research code we have implemented both methods.
version 2 .E.22) This motivated Harten [Har83] to introduce the notion of discrete total variation of numerical solution uh := {uj } T V (uh ) := j uj+1 − uj  . In the presence of shock waves. Kolgan [Kol72. the Central Aerodynamical National Laboratory near Moscow. The following theorem was proved in [Har83]: .138 Two phase ﬂows nity that B. one can establish the following inequality +∞ +∞ −∞ du(x.” tel00194763. that a scheme closely resembling MUSCL (including limiting) was developed in this laboratory by V.22) T V (un+1 ) ≤ T V (un ).4. formerly of TsAGI. the works of N. For compactly supported or periodic solutions. and the discrete counterpart to (3. We would like to underline the fact that the ﬁrst publication of Kolgan came out seven years before van Leer’s paper. Kol75] remained unknown for a long time. which was probably observed for the ﬁrst time by Peter Lax [Lax73]: The total increasing and decreasing variations of a diﬀerentiable solution between any pair of characteristics are conserved. Vladimir Sabelnikov. Kolgan (1972). Van Leer seems to be aware of this situation since in his recent review paper [vL06] one can ﬁnd “A historical injustice” section: “It has been pointed out to me by Dr.21 Jan 2008 Thus. 3. P. t2 ) ≤ −∞ du(x. van Leer was ﬁrst to propose the gradient reconstruction and slope limiting ideas. t2 ≥ t1 . his work apparently received little notice outside TsAGI. it is desirable to correctly attribute this result in future publications.3 TVD and MUSCL schemes There is a property of scalar nonlinear conservation laws. information is lost and the total variation decreases. Kolgan died young. Unfortunately. (3. because of political reasons. h h If this property is fulﬁlled we say that a ﬁnite volume scheme is total variation diminishing (TVD). t1 ) .
3.4 Finite volume scheme on unstructured meshes Theorem 1 (i): Monotone schemes are TVD
139
(ii): TVD schemes are monotonicity preserving, i.e. the number of solution extrema is preserved in time. Remark 11 From the mathematical point of view it would be more correct to say “the total variation nonincreasing (TVNI) scheme” but the “wrong” term TVD is generally accepted in the scientiﬁc literature. In one space dimension the construction of TVD schemes is not a problem anymore. Let us recall that in this study we are rather interested in two space dimensions (or even three in future work). In these cases the situation is considerably more complicated. Even if we consider the simplest case of structured cartesian meshes and apply a 1D TVD scheme on a dimensionbydimension basis, a result of Goodman and Leveque shows [GV85] that TVD schemes in two or more space dimensions are only ﬁrst order accurate. Motivated by this negative result, weaker conditions yielding solution monotonicity preservation should be developed. In this chapter we will describe the construction and practical implementation of a secondorder nonlinear scheme on unstructured (possibly highly distorted) meshes. The main idea is to ﬁnd our solution as a piecewise aﬃne function on each cell. This kind of linear reconstruction operators on simplicial control volumes often exploit the fact that the cell average is also a pointwise value of any valid (conservative) linear reconstruction evaluated at the gravity center of a simplex. This reduces the reconstruction problem to that of gradient estimation given cell averaged data. In this case, we express the reconstruction in the form ¯ wK (x) = wK + (∇w)K · (x − x0 ), K ∈ T , (3.23) ¯ where wK is the cell averaged value given by ﬁnite volume method, (∇w)K is the solution gradient estimation (to be determined) on the cell K, x ∈ K and the point x0 is chosen to be the gravity center for the simplex K. It is very important to note that with this type of representation (3.23) we remain absolutely conservative, i.e. 1 vol(K) ¯ wK (x) dΩ ≡ wK
tel00194763, version 2  21 Jan 2008
K
due to the choice of the point x0 . This point is crucial for ﬁnite volumes because of intrinsic conservative properties of this method. In next sections we describe brieﬂy two common techniques: GreenGauss integration and least squares methods for solution gradient estimation on each cell. There are other
140
Two phase ﬂows
available techniques. We can mention here an implicit gradient reconstruction method proposed in [MG96] and reused later in [AMS04], for example. We decided not to implement this approach in our research code since this procedure is computationally expensive7 . 3.4.3.1 GreenGauss gradient reconstruction
∂K
n
O
N2 e N1
tel00194763, version 2  21 Jan 2008
K
Figure 3.7: Illustration for GreenGauss gradient reconstruction. Control volume K with barycenter O and exterior normal n.
This gradient reconstruction technique can be easily implemented on simplicial meshes. It is based on two very simple ideas: the mean value approximation and GreenGaussOstrogradsky formula. Consider a control volume K with barycenter O. The exterior normal to an edge e ∈ ∂K is denoted by ne . This conﬁguration is depicted on Figure 3.7. In order to estimate the solution gradient on K (or in other words, to estimate its value at gravity center O) we make the following mean value approximation (∇w)K = (∇w)O ∼ = 1 vol(K) ∇w dΩ,
K
and if we apply GreenGaussOstrogradsky formula (∇w)K ∼ =
7
1 vol(K)
∂K
w ⊗ n dσ =
1 vol(K) e∈∂K
e
w ⊗ ne dσ ∼ =
e∈∂K
area(e) we/2 ⊗ ne , vol(K)
In order to reconstruct the solution gradient we have to solve a linear system of equations. Recall that the gradient is estimated at each time step on each control volume. This factor slows down considerably explicit time discretizations.
3.4 Finite volume scheme on unstructured meshes
141
where we/2 denote the solution value at the face (or edge in 2D) centroid. The face value needed to compute the reconstruction gradient can be obtained from a weighted average of the values at the vertices on the face [HC89]. In 2D it simply becomes we/2 = wN1 + wN2 . 2
This approximation yields the following formula for gradient estimation: (∇w)K ∼ = area(e) (wN1 + wN2 ) ⊗ ne . vol(K) 2
e∈∂K
tel00194763, version 2  21 Jan 2008
The gradient calculation is exact whenever the numerical solution varies linearly over the support of the reconstruction. This procedure requires the knowledge of the solution values at the mesh nodes {Ni }. Recall that a cell centered ﬁnite volume scheme provides us with data located at cell gravity centers. Thus, an interpolation technique is needed. We have to say that the quality of GreenGauss gradient reconstruction greatly depends on the chosen interpolation method. The method retained in this study is explained in Section 3.4.5. 3.4.3.2 Leastsquares gradient reconstruction method
O2
T2 O K T3 O3 O1 T1
Figure 3.8: Illustration for leastsquares gradient reconstruction. We depict a triangle control volume with three adjacent neighbors.
In this section we consider a triangle8 control volume K with three adjacent neighbors T1 , T2 and T3 . Their barycenters are denoted by O(x0 ), O1 (x1 ), O2 (x2 ) and O3 (x3 )
8
Generalization to other simplicial control volumes is straightforward.
142
Two phase ﬂows
respectively. In the following we denote by wi the solution value at gravity centers Oi : wi := w(xi ), w0 := w(x0 ).
Our purpose here is to estimate ∇w = (∂x w, ∂y w) on the cell K. Using Taylor formula, we can write down the three following relations: w1 − w0 = (∇w)K · (x1 − x0 ) + O(h2 ), (3.24) (3.25) (3.26)
w2 − w0 = (∇w)K · (x2 − x0 ) + O(h2 ),
w3 − w0 = (∇w)K · (x3 − x0 ) + O(h2 ).
tel00194763, version 2  21 Jan 2008
where ∆xi = xi − x0 , ∆yi = yi − y0 . For further developments it is convenient to rewrite our constraints in abstract form [L1 , L2 ] · (∇w)K = f . (3.27)
If we drop higher order terms O(h2 ), these relations can be viewed as a linear system of three equations for two unknowns9 (∂x w, ∂y w). This situation is due to the fact that the number of edges incident to a simplex mesh in Rd is greater or equal (in this case see Remark 12) to d thereby producing linear constraint equations (3.24) – (3.26) which will be solved analytically here in a least squares sense. First of all, each constraint (3.24) – (3.26) is multiplied by a weight ωi ∈ (0, 1) which will be chosen below to account for distorted meshes. In matrix form our nonsquare system becomes ω1 (w1 − w0 ) ω1 ∆x1 ω1 ∆y1 ω2 ∆x2 ω2 ∆y2 (∇w)K = ω2 (w2 − w0 ) , ω3 (w3 − w0 ) ω3 ∆x3 ω3 ∆y3
We use a normal equation technique in order to solve symbolically this abstract form in a least squares sense. Multiplying on the left both sides of (3.27) by [L1 L2 ]t yields G(∇w)K = b, G = (lij )1≤i,j≤2 = (L1 · L1 ) (L1 · L2 ) (L2 · L1 ) (L2 · L2 ) (3.28)
(L1 · f ) . The socalled normal (L2 · f ) equation (3.28) is easily solved by Cramer’s rule to give the following result where G is the Gram matrix of vectors L1 , L2 and b = (∇w)K =
9
1 2 l11 l22 − l12
l22 (L1 · f ) − l12 (L2 · f ) . l11 (L2 · f ) − l12 (L1 · f )
This simple estimation is done for scalar case only w = (w). For more general vector problems the numbers of equations and unknowns have to be changed depending on the dimension of vector w.
3.4 Finite volume scheme on unstructured meshes
143
The form of this solution suggests that the least squares linear reconstruction can be eﬃciently computed without the need for storing a nonsquare matrix. Now we have to discuss the choice of weight coeﬃcients {ωi }3 . The basic idea is i=1 to attribute bigger weights to cells barycenters closer to the node N under consideration. One of the possible choices consists in taking a harmonic mean of respective distances ri = xi − xN . This purely metric argument takes the following mathematical form: ωi = xi − xN −k , 3 −k j=1 xj − xN 
where k in practice is taken to be one or two (in our numerical code we choose k = 1). Remark 12 When we have a triangle sharing an edge with boundary ∂Ω (see Figure 3.13 for illustration), the gradient reconstruction procedure becomes even simpler, since the number of constraints is equal to d and linear system (3.24) – (3.26) becomes completely determined: w1 − w0 = (∇w)K · (x1 − x0 ) + O(h2 ), or in componentwise form it reads x1 − x0 y1 − y0 x2 − x0 y2 − y0 (∇w)K = w1 − w0 w2 − w0 . w2 − w0 = (∇w)K · (x2 − x0 ) + O(h2 ),
tel00194763, version 2  21 Jan 2008
The unique solution to this linear system is given again by Cramer’s rule (y2 − y0 )(w1 − w0 ) − (y1 − y0 )(w2 − w0 ) (x1 − x0 )(w2 − w0 ) − (x2 − x0 )(w1 − w0 ) (x1 − x0 )(y2 − y0 ) − (x2 − x0 )(y1 − y0 )
(∇w)K = 3.4.3.3 Slope limiter
.
The idea of incorporating limiter functions to obtain nonoscillatory resolution of discontinuities and steep gradients dates back to Boris and Book [BB73]. When the limiter is identically equal to 1, we have the unlimited form of the linear interpolation. In the 1D case one can easily ﬁnd in the literature about 15 diﬀerent limiter functions such as CHARM, minmod, superbee, van Albada and many others. On unstructured meshes the situation is quite diﬀerent. In the present study we decided to retain the BarthJespersen limiter proposed in [BJ89]. Here we do not discuss its construction and properties but just give the ﬁnal formula. We need to introduce the following notation
min wK := min wL , L∈N (K) max wK := max wL . L∈N (K)
it can be shown that it yields ﬁnite volume schemes possessing a global extremum diminishing property. where it is assumed that αK ∈ [0. 3. Also this limiter produces the least amount of slope reduction which can be advantageous for accuracy. K∈T . 1]. . It was proposed for nonconforming cartesian meshes in [Coi94] and rigorous justiﬁcation was given later in [CVV99]. In this section we brieﬂy describe a method for diﬀusive ﬂuxes discretization.23) is given by the following modiﬁed reconstruction operator ¯ wK (x) = wK + αK (∇w)K · (x − x0 ).9: Diamond cell constructed around face f . version 2 . min ¯ ¯ wK −wK BJ min αK := min if wK (xf ) < wK . ¯ ∀f ∈∂K wK (xf )−wK 1 otherwise. Obviously. Note that in practical implementation little modiﬁcations are required to prevent near zero division for almost constant solution data. Barth and Jespersen [BJ89] propose the following choice of αK : wK −wK max ¯ max wK (xf )−wK if wK (xf ) > wK .4 Diﬀusive ﬂuxes computation O1 K N2 f N1 O2 L Figure 3. Although this limiter function does not fulﬁll all the requirements of ﬁnite volume maximum principle on unstructured meshes [BO04]. tel00194763.21 Jan 2008 where xf denotes the face f centroid. the choice αK = 0 corresponds to the ﬁrst order scheme while αK = 1 is the unlimited form.4.144 Two phase ﬂows The limited version of (3.
while nKL is determined by mesh geometry. In order to express it in terms of cell averages {wL }L∈T we apply once again GreenGauss technique.4 Finite volume scheme on unstructured meshes 145 When we look at the integral conservation law (3. These volumes share a common face (or an edge in 2D) f with nodes N1 . . Actually it is rather a complicated topic. The quantity (D∇w) is called the diﬀusion ﬂux and should be properly discretized in this section. Consider two simplicial control volumes K and L with barycenters O1 and O2 respectively. O1 N2 .1): (∇wf /2 ) ∼ = 1 vol(f ⋄ ) ∇w dΩ = 1 vol(f ⋄ ) ℓ∈∂f ⋄ w ⊗ nℓ dσ ∼ = f⋄ f⋄ ℓ∈∂f ⋄ area(ℓ) area(ℓ) (wP1 + wP2 ) (∇wf /2 ) ⊗ nℓ ∼ ⊗ nℓ . We have not discussed the implementation of boundary conditions for diﬀusive ﬂuxes. The diﬀusion matrix D is given by physical properties. Remark 13 There is a little question omitted in this section. .9. Once the diamond cell is deﬁned. Consider a control volume L having a face ℓ on the boundary ∂Ω. N2 . N2 . . For the sake of clarity we construct the diamond cell in the 2D case. We have to start by constructing a dual mesh T ∗ . we can apply to its boundary the GreenGauss method (see Section 3. the diﬀusion term has the following form: 1 vol(K) f ∈∂K (D∇w) · nf dσ. For the time being 10 1 In 1D it simply degenerates to a very well known middle point rule −1 g(x) dx ∼ 2g(0).3. O2 N1 } and P1.3.2 ∈ {N1 .21 Jan 2008 Thus. O1 . in order to compute the diﬀusive ﬂux. vol(K) f ∈∂K tel00194763. version 2 . We assume that in our physical problems viscosity plays a secondary role. This situation is schematically depicted on Figure 3. In the present study we choose f ⋄ as the closed polygon formed by nodes N1 O1 N2 O2 N1 (see Figure 3. we have just to estimate the solution gradient ∇w but this time on the face centroid denoted by f /2.20) for control volume K. O2 }.9). The last integral can be further simpliﬁed by applying a second order Gauss quadrature10 : area(f ) D (∇wf /2 ) · nf . It is composed of diamond cells created around each interior face (diﬀusive ﬂux computation through a boundary face is another topic which is brieﬂy discussed in Remark 13). = . We denote by f ⋄ the diamond cell associated to the face f .4. N2 O2 . f where the integration is performed along all the faces of the cell K. = vol(f ⋄ ) vol(f ⋄ ) 2 ⋄ ℓ∈∂f where ℓ ∈ {N1 O1 .
10 for illustration). using the Taylor formula or.5 Solution interpolation to mesh nodes We have seen above that several gradient reconstruction procedures (in particular gradient estimation on the faces) require the knowledge of the solution at mesh nodes (or vertices).4. In other words it means that the gradient on the boundary is replaced by its value on the cell L. Here d(N ) is the degree of vertex N in the sense of graph theory. 3. equivalently.21 Jan 2008 Oi+1 N Oi Oi−1 Figure 3. (3. tel00194763. This information is not directly given by the ﬁnite volume method since we retained the cellcentered approach. One interesting technique was proposed in [HC89] and further improved in . Let us consider a node N (xn .10: Triangles with their barycenters Oi sharing the same vertex N . One of the possible ways to overcome this contradiction is the averaging. After discretization this condition takes the following form: (D∇wℓ ) · n := (D∇wL ) · n.23) we can estimate the solution value at the node N ¯ wN = wKi + (∇w)Ki · (xN − xi ). The MUSCL procedure provides us solution gradient on each cell. yi ) having this node as a vertex (see Figure 3.146 Two phase ﬂows we impose a Neumann type condition on diﬀusive ﬂuxes. yn ) of the tesselation T and a control volume Ki with barycenter Oi (xi .29) The problem is that we will have d(N ) diﬀerent values of the solution in the same point depending on the control volume under consideration. representation (3. Thus. version 2 .
which was to minimize the cost function given by (3.3. The diﬀerence lies in the factor of 2 ri ≡ ON − OOi 2 which was introduced in [KMC03].30) is equivalent to minimizing the function L deﬁned by 1 L= 2 which leads to ∆ωi = d(N ) d(N ) d(N ) ri ∆ωi i=1 2 −λ i=1 ωi (xi − xn ) − µ i=1 ωi (yi − yn ) → min λ(xi − xn ) + µ(yi − yn ) . we slightly modiﬁed this method.31) with the constraints (3. (3. the original optimization problem.30). let us look for the vertex value wN as a weighted sum of the values wNi computed by formula (3.4 Finite volume scheme on unstructured meshes 147 [KMC03]. The weighting factors {ωi }d(N ) are made to satisfy the condition of zero pseudoLaplacian i=1 d(N ) d(N ) L(xn ) ≡ i=1 ωi (xi − xn ). the weighting factor ωi is written as ωi = 1 + ∆ωi . The weights {ωi } are determined by solving an optimization problem in which the costfunction to be minimized is deﬁned as 1 2 d(N ) ri ∆ωi i=1 2 → min (3. The algorithm implemented in our code is brieﬂy described here.30) tel00194763. Employing the method of Lagrange multipliers. It should be noted that the cost function is slightly diﬀerent from the original formulation.29) from each surrounding cell ¯ wN = d(N ) i=1 ωi wNi d(N ) i=1 ωi .31) with two constraints given by (3. 2 ri . In our turn. They are imposed so that the method be exact for aﬃne data over the stencil. This modiﬁcation eﬀectively allows larger values of weight ∆ωi for those cells closer to the node in question. L(yn ) ≡ i=1 ωi (yi − yn ) . As in the original formulation by Holmes and Connell [HC89].21 Jan 2008 These conditions have a very simple interpretation. ¯ First of all. version 2 .
3. 2 ri Iyy = i=1 (yi − yn )2 . we would like to give another idea (based on a purely metrics argument) of how to construct the weights {ωi }d(N ) .148 The two Lagrangian multipliers.10): xN − xi −k .4. 2 Ixx Iyy − Ixy d(N ) µ= rx Ixy − ry Ixx . are obtained from λ= where rx = i=1 d(N ) Two phase ﬂows ry Ixy − rx Iyy . d(N ) ry = i=1 (yi − yn ). 2 Ixx Iyy − Ixy d(N ) (xi − xn ). ωi := d(N ) −k j=1 xN − xi  where k in practice is taken 1 or 2.4.21 Jan 2008 Remark 14 The above algorithm is not computationally expensive since the weights {ωi }d(N ) i=1 depend only on the tesselation T geometry. 1]. in order to calculate ωi one can simply take the harmonic mean of distances between the node N under the question and respective cell barycenter Oi (see Figure 3. This approach is considerably simpler than optimization i=1 problem solving and it was already used in the least squares gradient reconstruction method (see Section 3. λ and µ. 2 ri to the range [0. Remark 15 Even if we suggest to use the above method (since it gives slightly better results). In fact. d(N ) Ixx = i=1 (xi − xn )2 .2). In the present work we decided to retain the socalled Strong StabilityPreserving (SSP) .32) In order to obtain a fully discrete scheme. It means that they can be computed and stored before the main loop in time and reused during the computations later. we have to discretize the time evolution operator. version 2 .6 Time stepping methods In previous sections we considered the spatial discretization procedure with a ﬁnitevolume scheme.3. We have to say that this choice does not guarantee exact interpolation of globally linear functions. The last step consists in renormalizing the weights {ωi }d(N ) i=1 tel00194763. This approach is called a method of lines: FV ut + ∂x f (u) = S(u) =⇒ ut = L(u) (3. It is a common practice in solving timedependent PDEs to ﬁrst discretize the spatial variables. 2 ri Ixy = i=1 (xi − xn )(yi − yn ) .
m. .k L(u(k) ) . . A general mstage RungeKutta method for (3.34) (3. 11 . GST01.k αi. Usually the relevant norm is the total variation11 norm: TV(un ) := j un − u n j j−1 and TVD discretizations have the property TV(un+1 ) ≤ TV(un ).35) with βi. .33) – (3. Historically these methods were initially called Total Variation Diminishing (TVD) time discretizations. un+1  ≤ un . Thus a stronger measure of stability is usually required: Deﬁnition 1 A sequence {un } is said to be strongly stable in a given norm · provided that un+1  ≤ un  for all n ≥ 0.k ≥ 0.21 Jan 2008 (3.35) u(i) = un+1 = u k=0 (m) In [SO88] the following result is proved Theorem 2 If the forward Euler method is strongly stable under the CFL restriction ∆t ≤ ∆tF E un + ∆tL(un ) ≤ un  . c = min i.k .k Here we give a few examples of SSP schemes which are commonly used in applications: We have to say that the notion of total variation is used essentially for onedimensional discrete solutions. αi. SR02]. i = 1. Remark 16 Special approaches are needed for hyperbolic PDEs since they contain discontinuous solutions and the usual linear stability analysis is inadequate.32) can be written in the form u(0) = un .32).k u(k) + ∆tβi. Then. The main idea behind SSP methods is to assume that the ﬁrst order forward Euler method is strongly stable (see the deﬁnition below) under a certain norm for our method of lines ODE (3. (3.33) αi.k ≥ 0 is SSP.4 Finite volume scheme on unstructured meshes 149 time discretization methods described in [Shu88. version 2 . then the RungeKutta method (3. . we try to ﬁnd a higher order scheme.3. . βi. provided the following CFL restriction is fulﬁlled: ∆t ≤ c∆tF E . i−1 tel00194763.
2) & RK2 SSPRK(3. version 2 .11: Linear absolute stability regions for diﬀerent time stepping methods Figure 3.4) RK4 3 2 1 0 −1 −2 −3 tel00194763.21 Jan 2008 −6 −5 −4 −3 −2 −1 0 1 2 Figure 3.3) & RK3 SSPRK(3.150 Two phase ﬂows Absolute stability regions SSPRK(2.12: Nonlinear absolute stability regions .
version 2 . u(n+1) = 2 2 2 • Optimal third order threestage SSPRK(3. u(n+1) = 3 3 3 151 tel00194763.2 and we refer to [GP05] for a general discussion. . 3 3 6 1 u(n+1) = u(3) + ∆tL(u(3) ). That is why we decided to look also at nonlinear absolute stability [CP92]. 2 1 u(2) = u(1) + ∆tL(u(1) ). Note that the linear absolute stability region for corresponding RK and SSPRK schemes is the same.4) due to its accuracy and wide stability region. We tested these diﬀerent schemes in our numerical code and we decided to adopt SSPRK(3. It is a consequence of linearization since these schemes are clearly diﬀerent for nonlinear problems.12 the linear and nonlinear absolute stability regions of these schemes.4) scheme with CFL = 2: 1 u(1) = u(n) + ∆tL(u(n) ).3) scheme with CFL = 1: u(1) = u(n) + ∆tL(u(n) ).7 Boundary conditions treatment In this section we show how to implement the wall boundary condition in the ﬁnite volumes framework. 12 Optimality in the sense of CFL condition. 4 4 4 1 (n) 2 (2) 2 u + u + ∆tL(u(2) ). 2 2 (n) 1 (2) 1 u(3) = u + u + ∆tL(u(n) ). and compared them with the classical RungeKutta methods respectively. 1 (n) 1 (1) 1 u + u + ∆tL(u(1) ). In our opinion this scheme represents a very good tradeoﬀ between precision and robustness.2) scheme with CFL = 1: u(1) = u(n) + ∆tL(u(n) ).11 and 3. The ﬂavor of boundary conditions treatment for hyperbolic systems is given in Section A.3.21 Jan 2008 • Third order fourstage SSPRK(3. 3 (n) 1 (1) 1 u(2) = u + u + ∆tL(u(1) ). 3.4. 2 We depicted on Figures 3.4 Finite volume scheme on unstructured meshes • Optimal12 second order twostage SSPRK(2.
36) tel00194763. pb n1 . ρvun + pn2 .13 for illustration).21 Jan 2008 ∂Ω K Figure 3. Consider the case of a rigid wall boundary u(x.1) – (3.3). and the hyperbolic system (3. 0. ρuun + pn1 . and if we take into account physical condition (3. pb := px∈∂Ω un ≡ u · n. The last identity means that we need only to know the pressure value on the wall in order to reconstruct the advective ﬂux (F · n)x∈∂Ω . version 2 . In this situation.152 Two phase ﬂows From the numerical point of view. pb n2 . ρHun ).4. The advective ﬂux13 through an oriented boundary face with exterior normal n is given by (3.4 . t) · n = 0. α− ρ− un . 0). we need to employ the boundary conditions when we have a control volume sharing a face (or an edge in 2D) with the computational domain boundary (see Figure 3. the numerical method must be provided by an additional information (such as the value of normal advective or diﬀusive ﬂuxes on ∂K).5) F · n = (α+ ρ+ un . Physical conditions completed if necessary by numerical ones allow us to reconstruct the normal ﬂux F · n through the boundary face. x ∈ ∂Ω.36) it becomes (F · n)x∈∂Ω = (0. 13 A few words about numerical boundary conditions for diﬀusive ﬂuxes are given in Section 3.13: Control volume sharing a face with boundary ∂Ω. (3.
The reader can believe us that our numerical code passed successfully these classical tests (see Figure 3. version 2 .21 Jan 2008 Figure 3. we decided not to put here classical test cases such as the shock tube of Sod [Sod78] etc.14: Shock tube of Sod. the boundary pressure pb measures the force “exerted” by the wall in order to annihilate completely the normal velocity un . Remark 17 The last formula has a very interesting physical interpretation.14). Actually.3. where the righthand side is evaluated in the boundary cell center. In order to reduce the document size.5 Numerical results 153 tel00194763.5 Numerical results In this section we present a few test cases which show our research code perfomance. . 3. . After some straightforward computations (for more details see [GP05]) one establishes the following formula for the boundary pressure pb = p + ρun cs . This plot represents the density.
We remark that in our implementation of the secondorder scheme the leastsquares reconstruction seems to give slightly more accurate results than the GreenGauss procedure. On this ﬁgure we represented one more curve (the highest one) which corresponds to GreenGauss gradient reconstruction on ﬂuxes (it seems to be very natural in the context of VFFC scheme).154 Two phase ﬂows 3. it has almost compact support in order to reduce the inﬂuence of boundary conditions.5. this kind of data is extremely computer dependent but the qualitative behaviour is the same on all systems. Here again one can notice that the leastsquares method is slightly faster than the GreenGauss procedure. On this plot. So. the blue curve corresponds to the ﬁrst order upwind scheme while the other two (red and black) correspond to the MUSCL scheme with leastsquares (see Section 3. The other two curves have almost the same slope equal to 1. we constructed a sequence of reﬁned meshes.3. The slope of these curves represents an approximation to the theoretical convergence rate. Our numerical tests show that this method is quite expensive from the computational point of view and we decided not to retain it.2) and GreenGauss (see Section 3.4. Moreover.1 Convergence test We begin the presentation of numerical tests by the simplest one – convergence test. To do it.21 Jan 2008 with smooth14 initial conditions. It is obvious that this equation will just translate the initial form in the direction u0 . On the other hand. version 2 . On Figure 3.15 we depict the error of numerical method in L∞ norm as a function of the mesh characteristic size. Obviously.90 indicating a second order convergence rate for the MUSCL scheme. One can see that the blue curve slope is equal approximatively to 0. we have an analytical solution which can be used to quantify the numerical method error. to measure the convergence rate. We would like to show the accuracy of the MUSCL scheme implementation.3. On Figure 3.4. Then we have two almost superimposed (black and red) curves referring to the secondorder gradient reconstruction on variables. ∂t u0 ∈ R2 tel00194763.1) gradient reconstruction procedures respectively.97 which means ﬁrst order convergence. We intentionally choose a smooth initial condition since the discontinuities can decrease the overall accuracy of the scheme. we solve numerically the following scalar linear advection equation ∂v + u0 · ∇v = 0.16 one can see that the “fastest” curve is the blue one (ﬁrst order upwind scheme). The next ﬁgure represents the measured CPU time in seconds again as a function of the mesh size. 14 .
version 2 . Conv.15: Numerical method error in L∞ norm.9012 10 −4 10 −2 10 10 h − average edge length −1 0 10 1 Figure 3. rate = 1. Conv. .21 Jan 2008 10 −1 L∞ error 10 −2 10 −3 Upwind.97731 LeastSq. rate = 0. rate = 1. Conv.3.9051 GGO on Variables.5 Numerical results 155 Convergence rate in L norm 10 0 ∞ tel00194763.
version 2 .156 Two phase ﬂows 10 3 CPU time in seconds Upwind 1st order Green−Gauss Least Sqs GG on fluxes 10 2 tel00194763.16: CPU time for diﬀerent ﬁnite volume schemes.21 Jan 2008 10 1 10 T 0 10 −1 10 −2 10 −3 10 −2 10 −1 10 h 0 10 1 Figure 3. .
24 the maximal pressure on the right wall as a function of time: t −→ max p(x.5 Numerical results 157 3.y)∈1×[0. The mesh used in this computation contained about 108000 control volumes (in this case it was triangles). t). we depicted on Figure 3.21 Jan 2008 We performed another computation when gas was modelled as α+ = 0.1 g 0. The values of the other parameters are given .2 Falling water column The geometry and initial condition for this test case are depicted on Figure 3. 1 0. version 2 .25.05.9 α+ = 0. On the other hand.05 0 0. Geometry and initial condition. The results of this simulation are presented on Figures 3.3.9 α− = 0.17: Falling water column test case.1] tel00194763. (x.26.17. y.1 α− = 0.3 Water drop test case The geometry and initial condition for this test case are depicted on Figure 3. Initially the velocity ﬁeld is taken to be zero. The pressure is recorded as well and this result is ploted on Figure 3.5. α− = 0. The values of the other parameters are given in Table 3.9 α+ = 0. One can see that the peak value is higher and the impact is more localised in time.23.95 mixture.1.3 0. 3.7 1 Figure 3.5.65 0. Initially the velocity ﬁeld is taken to be zero.18 – 3.
158
Two phase ﬂows
tel00194763, version 2  21 Jan 2008
(a) t = 0.005
(b) t = 0.06
Figure 3.18: Falling water column test case. Initial condition and the beginning of column dropping process.
(a) t = 0.1
(b) t = 0.125
Figure 3.19: Falling water column test case. Splash creation due to the interaction with step.
3.5 Numerical results
159
tel00194763, version 2  21 Jan 2008
(a) t = 0.15
(b) t = 0.175
Figure 3.20: Falling water column test case. Water strikes the wall  I.
(a) t = 0.2
(b) t = 0.225
Figure 3.21: Falling water column test case. Water strikes the wall  II.
160
Two phase ﬂows
tel00194763, version 2  21 Jan 2008
(a) t = 0.3
(b) t = 0.4
Figure 3.22: Falling water column test case. Splash is climbing the wall.
(a) t = 0.5
(b) t = 0.675
Figure 3.23: Falling water column test case. Turbulent mixing process.
3.5 Numerical results
Maximal pressure on the right wall 1.8 1.7 1.6 1.5 pmax/p0 1.4 1.3 1.2 1.1
161
tel00194763, version 2  21 Jan 2008
1 0.9
0
0.1
0.2
0.3 t, time
0.4
0.5
0.6
0.7
Figure 3.24: Maximal pressure on the right wall. Heavy gas case.
in Table 3.1. The mesh used in this computation contained about 92000 control volumes (in this case it was triangles). The results of this simulation with our code are presented on Figures 3.27 – 3.33. On Figure 3.34 we plot the maximal pressure on the bottom as a function of time: t −→
(x,y)∈[0,1]×0
max
p(x, y, t).
From this ﬁgure, it is clear to see that the pressure exerted on the bottom reaches 2.5p0 due to the drop impact at t ≈ 0.16. Remark 18 Beginning with Figure 3.32 one can see some asymmetry in the solution. It is not expected since the initial condition, computational domain and forcing term are fully symmetric with respect to the line x = 0.5. This discrepancy is explained by the use of unstructured meshes in the computation. The arbitrariness of the orientation of the triangles introduces small perturbations which are suﬃcient to break the symmetry at the discrete level. At the present time, symmetry preserving schemes on unstructured meshes represent a big challenge in numerical analysis. The author is not aware of any work in this direction.
162
Maximal pressure on the right wall 2.2
Two phase ﬂows
2
1.8
pmax/p0
1.6
1.4
1.2
tel00194763, version 2  21 Jan 2008
1
0.8
0
0.1
0.2
0.3 t, time
0.4
0.5
0.6
0.7
Figure 3.25: Maximal pressure on the right wall as a function of time. Light gas.
3.6
Conclusions
In this chapter we presented a simple mathematical model for simulating water waves impacts. The validation of this approach is the subject of future publications. Namely, we are going to perform qualitative and quantitative comparison with the more general six equations model [Ish75]. On the other hand, we presented a numerical approach for discretizing the governing equations. It is a second order ﬁnite volume scheme on unstructured meshes. This method was implemented in our research code. By construction, our code has excellent mass, momentum and total energy conservation properties. Diﬀerent numerical tests from Section 3.5 validate the numerical method. In the future we would like to carry out a parametric study with our solver. The inﬂuence of aeration, gas properties and other factors on the impact pressures is very important for industrial applications.
3.7 α+ = 0.005 (b) t = 0.15 g 0. . version 2 .21 Jan 2008 α+ = 0.1 α− = 0.27: Water drop test case.1 tel00194763. Initial conﬁguration and the beginning of the fall.9 0 0. (a) t = 0.6 Conclusions 163 1 R = 0.9 α− = 0.26: Geometry and initial condition for water drop test case.5 1 Figure 3.075 Figure 3.
21 Jan 2008 (a) t = 0.125 Figure 3. (a) t = 0. Drop/bottom compressible interaction.29: Water drop test case. version 2 .135 (b) t = 0.1 (b) t = 0. .15 Figure 3.28: Water drop test case.164 Two phase ﬂows tel00194763. Drop approaching container bottom.
version 2 .30: Water drop test case. (a) t = 0. .225 (b) t = 0.275 Figure 3. Side jets crossing. Vertical jets formation.2 Figure 3.21 Jan 2008 (a) t = 0.3.6 Conclusions 165 tel00194763.31: Water drop test case.175 (b) t = 0.
166 Two phase ﬂows tel00194763.45 Figure 3. Side jets interﬂow at the center.4 (b) t = 0.33: Water drop test case. (a) t = 0. version 2 .32: Water drop test case.35 Figure 3. .325 (b) t = 0. Central jet reﬂection from the bottom.21 Jan 2008 (a) t = 0.
2 t.5 tel00194763.15 0.25 0.05 0.21 Jan 2008 2 /p p 0 max 1. time 0. .35 0.6 Conclusions 167 Maximal pressure on the bottom 2.34: Water drop test case. version 2 .3 0.4 Figure 3.5 1 0 0. Maximal bottom pressure as a function of time.3.1 0.
version 2 .tel00194763.21 Jan 2008 168 Two phase ﬂows .
. 180 4. . 182 4. . . . . . . . . . . . . . .5 4. . . . . . . . . . . . . . . . . 187 4. . . . . 169 Anatomy of dissipation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viscous eﬀects cannot be neglected under certain circumstances. . .6. . . . . . . . . . . . . . . .21 Jan 2008 Albert Einstein (1879 – 1955) Contents 4. . . . . . . . . . . . . . . . . version 2 . . . . . . . . . . . . . .1 4. Indeed the question of dissipation in potential ﬂows of ﬂuid with a free surface is a very important one.1 Introduction Even though the irrotational theory of freesurface ﬂows can predict successfully many observed wave phenomena. . . . . . . .4. . . As stated by [LH92]. . . . . . . . tel00194763. . . . . . .6 Attenuation of linear progressive waves . . .1 4. 187 Discussion . . . . 172 Derivation . . but not simpler. . . . . . . . . 189 .1 4. . . . . . . . . . . . . . . . . 174 4. . . . . . 179 Dispersion relation .6. . . . . . . . . . . . .3. .2 4. . . . . . . .1 Discussion . . . . . . . . . 182 Numerical results . . . . .3 Introduction .Chapter 4 Weakly damped free surface ﬂows Everything should be made as simple as possible. . 192 4. . . . . . .2 Approximate solitary wave solution . . . . . . . . .7 Conclusion 4. . . . . . . . . . . . . . it would be convenient to have equations and boundary 169 . . . . . . . . . . . . . .4 Dissipative KdV equation . . . . . . . . . . . . . . . . . . . . . . . . . .
For example if we take a tsunami wave and estimate its Reynolds number.170 Viscous potential ﬂows conditions of comparable simplicity as for undamped freesurface ﬂows.. What is less known is that Boussinesq studied this eﬀect as well [Bou95].] Another author [Wu81] points out this defect of the classical water waves theory: [... For example.21 Jan 2008 where α denotes the wave amplitude. The eﬀects of viscosity on gravity waves have been addressed since the end of the nineteenth century in the context of the linearized Navier–Stokes (NS) equations.. the amplitude disagrees somewhat.1) tel00194763.. namely 2 α(t) = α0 e−2νk t .] However. The research community agrees at least on one point: the molecular viscosity is unimportant. This discrepancy can be ascribed to the neglect of the viscous eﬀects in the theory [.] it was found that the inclusion of a dissipative term was much more important than the inclusion of the nonlinear term. It is wellknown that Lamb [Lam32] studied this question in the case of oscillatory waves on deep water. although the inclusion of the nonlinear term was undoubtedly beneﬁcial in describing the observations [. dt (4. Here we are going to quote a few of them.2) The importance of viscous eﬀects for water waves has been realized by numerous experimental studies for at least thirty years. On the other hand.. and we suppose that this might be due to the viscous dissipation [.. in laboratory experiments the Reynolds number is much more moderate and sometimes we can neglect this ..] the peak amplitudes observed in the experiments are slightly smaller than those predicted by the theory.. ν the kinematic viscosity of the ﬂuid and k the wavenumber of the decaying wave.] And ﬁnally in the classical article [BPS81] one ﬁnds the following important conclusion: [. In this particular case they both showed that dα = −2νk 2 α. version 2 . Now let us discuss more debatable statements. in [ZG71] one can ﬁnd [. we ﬁnd Re ≈ 106 ..] Water wave energy can be dissipated by diﬀerent physical mechanisms. the ﬂow is clearly turbulent and in practice it can be modelled by various eddy viscosity models. The peculiarity here lies in the fact that the viscous term in the Navier–Stokes (NS) equations is identically equal to zero for a velocity deriving from a potential. So. (4. This equation leads to the classical law for viscous decay...
The diﬀerences will be explained below. while in the present study we remove these two hypotheses and all the computations are done in 3D. the most important (and the most challenging) mechanism of energy dissipation is wave breaking. We have to say that the boundary layer mechanism is especially important for long waves since they “feel” the bottom. The present chapter is a direct continuation of the recent study [DD07b. Nowadays the practitioners can only be happy to model roughly this process by adding adhoc diﬀusive terms when the wave becomes steep enough. We brieﬂy discuss the free surface boundary layer and explain why we do not take it into account in this study. The inclusion of this term is natural since it represents the correction to potential ﬂow due to the presence of a boundary layer. Unfortunately we made this discovery after publishing our ﬁrst paper [DD07b] on the subject and. As a result we obtain a qualitatively diﬀerent formulation which contains a nonlocal term in the bottom kinematic condition. In that work the authors considered twodimensional (2D) periodic waves in inﬁnite depth. Then one can think about boundary layers. We present here a novel approach. √ Moreover. CG07]. Even though we try to read the literature on a regular basis. In this work we keep the features of undamped freesurface ﬂows while adding dissipative eﬀects. we sometimes rederive what others did before. Other researchers have obtained nonlocal corrections but they diﬀer from ours [KM75]. The classical theory of viscous potential ﬂows [JW04] is based on pressure and boundary conditions corrections due to the presence of viscous stresses. DDZ07]. Liu. Consequently. this term is predominant since its magnitude scales with O( ν). while other terms in the freesurface boundary conditions are of order O(ν). This process is extremely diﬃcult from the mathematical but also the physical and numerical points of view since we have to deal with multivalued functions.1 Introduction 171 eﬀect. tel00194763. There have also been several attempts to introduce dissipative eﬀects into long wave modelling [Mei94. DD07a. topological changes in the ﬂow and complex turbulent mixing processes.4. Currently. it is impossible to be aware of all papers. This point is important since the vorticity structure is more complicated in 3D. DKZ04. we could not cite the original work of P. In this chapter we will deal essentially with the bottom boundary layer. Recently we discovered that the nonlocal term in exactly the same form was already derived in [LO04]. ZKPD05]. potential ﬂows with adhoc dissipative terms are used for example in direct numerical simulations of weak turbulence of gravity waves [DKZ03. version 2 . Finally.21 Jan 2008 . In other words we consider a general wavetrain on the free surface of a ﬂuid layer of ﬁnite depth. It happened to us and we ensure that it was done completely independently. consequently.
∂t ρ ρ where τ is the viscous stress tensor τij = 2µεij . z Sf O x Rf Ri Rb tel00194763. In order to make some estimates we introduce the notation which will be used in √ this section: µ is the dynamic viscosity. bottom boundary layers. We assume that the ﬂow is governed by the incompressible NavierStokes equations: ∇ · u = 0.2 Anatomy of dissipation In this section we are going to discuss the contribution of diﬀerent ﬂow regions into water wave energy dissipation. On this ﬁgure Sf and Sb stand for free surface and bottom respectively. t0 is the characteristic time. version 2 .21 Jan 2008 Sb Figure 4. Rf and Rb denote the interior region.172 Viscous potential ﬂows 4. δ = O( µ) is the boundary layer thickness characterization. Then.1. a0 is the characteristic wave amplitude and ℓ is the wavelength or basin length (in laboratory experiments. We conventionally [Mei94] divide the ﬂow into three regions illustrated on Figure 4. We multiply the second equation by u and integrate over the domain Ω to get the . εij = 1 ∂ui ∂uj + 2 ∂xj ∂xi .1: Conventional partition of the ﬂow region into interior region and free surface. Ri . for example). 1 1 ∂u + u · ∇u + ∇p = g + ∇ · τ . free surface and bottom boundary layers.
The lefthand side is the total rate of energy change in Ω. The second term is the ﬂux of energy across the boundary. the ﬁrst integral represents the rate of work by surface stresses acting on the boundary. 1 A free surface boundary layer also exists. We estimate the order of magnitude of the rate of dissipation in various regions of the ﬂuid. Another extreme case is when the free surface is heavily contaminated. The energy dissipation rate is then O TRi ∼ a 1 µ µ t0 ℓ 2 · ℓ3 = µ a t0 2 ℓ ∼ O(µ) . 3 From the physical point of view it is weaker. We start by the interior region Ri . The second integral is the rate of work done by the gravity force throughout the volume.21 Jan 2008 In this identity each term has a precise physical meaning. since only the zero shear stress condition on the free surface is required. Inside the bottom boundary layer the normal gradient of the solenoidal part of u dominates the strain rate. by oil slicks. The stress in the free surface boundary layer can then be as great as in . Because of the small volume O(δℓ2 ) the rate of dissipation in the free surface boundary layer is only O TRf ∼ a 1 µ µ t0 ℓ 2 · δℓ2 = µδ a t0 2 ∼ O(µ 2 ) . and the third integral T is the rate of viscous dissipation. so that O TRb ∼ 1 a µ µ t0 δ 2 · δℓ2 = µ aℓ δ t0 2 ∼ O(µ 2 ) . T tel00194763. Outside the boundary layers. Consider ﬁrst the classical case of a clean surface.4.2 Anatomy of dissipation following energy balance equation: 1 2 Ω 173 ∂ 1 ρu2 dΩ + ∂t 2 ∂Ω ρu2 u · n dσ = = ∂Ω −pI + τ n · u dσ + Ω ρg · u dΩ − 1 2µ Ω τ : τ dΩ . Its importance depends on the free surface conditions. On the righthand side. We will focus our attention on the last one T . version 2 . The stress is mainly controlled by the potential velocity ﬁeld which is of the same order as in the main body of the ﬂuid. for example. it is reasonable to 0 expect that the rate of strain is dominated by the irrotational part whose scale is a0 and t the length scale is the wavelength ℓ.
We represent v = (u. ρ the ﬂuid density and g the acceleration due to gravity. ψ2 . .3 tel00194763. ψ = (ψ1 . ∂t Next we discuss the boundary conditions. t) . v. φt + p − p0 + gz = 0.174 Viscous potential ﬂows the boundary layer near a solid wall. After linearization it becomes simply ηt = w. ψ3 ). version 2 . . (4. p the pressure. The previous scalings suggest the following diagram which represents the hierarchy of dissipative terms: O µ 2 ֒→ O(µ) ֒→ O µ 2 ֒→ . (4. ∂t ρ ∇ · v = 0. all 3 dissipative terms of order O(µ 2 ) and higher will be neglected.21 Jan 2008 Derivation Consider the linearized 3D incompressible NS equations describing freesurface ﬂows in a ﬂuid layer of uniform depth h: ∂v 1 = − ∇p + ν∆v + g.3). . ρ ∂ψ = ν∆ψ. one notices that the equations are veriﬁed provided that the functions φ and ψ satisfy the following equations: ∆φ = 0.4) into (4. Consequently.4) After substitution of the decomposition (4. We assume that the velocity ﬁeld satisﬁes the conventional noslip condition at the bottom vz=−h = 0. Dynamic condition states that the forces must be equal on both sides of the free surface: [σ · n] ≡ −(p − p0 )n + τ · n = 0 at z = η(x. while at the free surface we have the usual kinematic condition. Rb Ri Rf 1 3 It is clear that the largest energy dissipation takes place inside the wall boundary layer.3) with v the velocity vector. 4. w) in the form of the Helmholtz–Leray decomposition: v = ∇φ + ∇ × ψ. In the present study we do not treat such extreme situations and the surface contamination is assumed to be absent. We take into account only the two ﬁrst phenomena from this diagram. which can be stated as ηt + v · ∇η = w.
4) into these two identities yields 2 ∂ 2 ψ2 ∂ 2 ψ1 ∂ 2 ψ3 ∂ 2 ψ2 ∂2φ + − + − = 0. ∂y∂z ∂x∂y ∂y 2 ∂z 2 ∂x∂z z = 0. We assume that all the functions involved in the present computation satisfy the necessary regularity requirements and have suﬃcient decay at inﬁnity so that the integral transforms can be applied. (4. n= 1 1+ ∂η ( ∂x )2 ∂η − ∂x 1 . ˆ s where m2 := k2 + ν and ϕ+ . at z = 0.5) 2 z = 0. determined by the initial and appropriate boundary conditions. Let us use ﬁrst those related to the tangential stresses (the third one will be used later). ψ in the transform space. version 2 . s)e−mz . ϕ− . LF ˆ f (x. There are three dynamic conditions on the free surface. Using Fourier–Laplace transforms.6) . 1).21 Jan 2008 we can determine the structure of the unknown functions φ. C3 ) are unknown functions of the ˆ0 ˆ0 transform parameters (k. The explicit expressions (in two dimensions for simplicity) of τ and n are τ = ρν 2 ∂u ∂x ∂u + ∂w ∂z ∂x ∂u ∂z + ∂w ∂x 2 ∂w ∂z . n is the normal to the free surface and τ the viscous part of the stress tensor σ. s)e−kz . where µ = ρν: σxz = µ ∂w ∂u + = 0. t) −→ f (k. which we denote by LF ≡ L ◦ F. k = (kx . The solution for φ is obtained from the transformed continuity equation LF ˆ ˆ ∆φ = 0 −→ φzz − k2 φ = 0 and ψ from the corresponding transformed equation LF ˆ ˆ ˆ ψt = ν∆ψ −→ sψ = ν ψzz − k2 ψ ˆ ˆ φ = ϕ+ (k. ∂x ∂z σyz = µ ∂w ∂v + ∂y ∂z = 0. ψ0 . s). ky ) tel00194763. s). ˆ0 0 ˆ ˆ ψi = ψi0 (k. s)ekz + ϕ− (k. 0. C := (C1 . Substituting decomposition (4. [f ] denotes the jump of a function f across the free surface. s) emz + Ci (k. C2 .4. (4. After linearization the normal vector n equals to (0.3 Derivation 175 where σ is the stress tensor. ∂x∂z ∂x2 ∂x∂y ∂y∂z ∂z 2 ∂2φ ∂ 2 ψ2 ∂ 2 ψ1 ∂ 2 ψ1 ∂ 2 ψ3 + − + − = 0.
9) (4. (4.8) can be rewritten as k (ϕ+ − ϕ− ) ˆ0 ˆ0 = η. since the derivation dealt with the inﬁnite depth case.176 Viscous potential ﬂows The next step consists in taking the Fourier–Laplace transform to these relations. ∂t ∂z ∂x ∂y In transform space it becomes ˆ ˆ sˆ = k (ϕ+ − ϕ− ) + iky ψ10 (1 + C1 ) − ikx ψ20 (1 + C2 ). ˆ (4. we have to modify the dynamic freesurface condition as well. We do not give here the explicit expressions since this operation is straightforward. we consider a semiinﬁnite ﬂuid layer as it is usually done in boundary layer theory.7) Let us turn to the freesurface kinematic condition ∂φ ∂ψ2 ∂ψ1 ∂η =w≡ + − .7) and (4. In this derivation we use the pure Leray decomposition . In order to include this modiﬁcation. ˆ F (4. The second step in our derivation consists in introducing a boundary layer correction at the bottom. so Bernoulli’s equation becomes 3 1 − ∂ 2 ψ1 ∂y∂z φt + gη + 2νφzz + O(ν 2 ) = 0. version 2 . This is done using the balance of normal stresses at the free surface: σzz = 0 at z = 0 ⇒ p − p0 = 2ρν Using (4.10) Using (4. ∂z ∂z 2 ∂x∂z ∂y∂z = O(ν 2 ).12) Since we only consider weak dissipation (ν ∼ 10−6 − 10−3 m2 /s). ˆ0 ˆ0 m2 + k2 (4.5) + (−iky )(4. we neglect terms of order o(ν).10) one can replace the rotational part in the kinematic freesurface condition: ηt = φz + L−1 −2ν k2 η = φz + 2ν∆η. The ﬂuid occupies the domain z > −h. η ˆ0 ˆ0 (4. this was not done in the previous study [DDZ07]. ˆ ν(m2 + k2 ) ˆ ˆ iky ψ10 (1 + C1 ) − ikx ψ20 (1 + C2 ) = −2ν k2 η . The combination (−ikx )(4. Obviously.21 Jan 2008 Equations (4.8) z = 0.10) one can show that ∂ 2 ψ2 ∂x∂z ∂ 2 ψ2 ∂ 2 ψ1 ∂2φ ∂w ≡ 2ρν + − .6) gives the important relation ˆ ˆ iky ψ10 (1 + C1 ) − ikx ψ20 (1 + C2 ) = − 2 k3 (ϕ+ − ϕ− ).11) In order to account for the presence of viscous stresses. tel00194763.
Analytical solut