HAL Id: hal-02510722
https://hal.inria.fr/hal-02510722v2
Preprint submitted on 23 Sep 2021
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
The Navier-Stokes system with temperature and salinity for free surface flows. Numerical scheme and validation
Léa Boittin, François Bouchut, Marie-Odile Bristeau, Anne Mangeney, Jacques Sainte-Marie, Fabien Souillé
To cite this version:
Léa Boittin, François Bouchut, Marie-Odile Bristeau, Anne Mangeney, Jacques Sainte-Marie, et al..
The Navier-Stokes system with temperature and salinity for free surface flows. Numerical scheme and validation. 2021. �hal-02510722v2�
The Navier-Stokes system with temperature and salinity for free surface flows
- Numerical scheme and validation
L. Boittin
1,4, F. Bouchut
2, M.-O. Bristeau
1, A. Mangeney
1,3, J.
Sainte-Marie
1, and F. Souillé
1,41
Inria Paris, 2 rue Simone Iff, CS 42112, 75589 Paris Cedex 12 and Sorbonne Université, Univ. Paris Diderot, SPC, CNRS, Laboratoire
Jacques-Louis Lions, LJLL, F-75005 Paris
2
Laboratoire d’Analyse et de Mathématiques Appliquées (UMR 8050), CNRS, Univ. Gustave Eiffel, UPEC, F-77454, Marne-la-Vallée, France
3
Univ. Paris Diderot, Sorbonne Paris Cité, Institut de Physique du Globe de Paris, Seismology Group, 1 rue Jussieu, Paris F-75005, France
4
Risk Management Solutions, Peninsular House, 30 Monument Street, London, EC3R 8NB, UK
5
EDF R&D LNHE - Laboratoire National d’Hydraulique et Environnement, 6 quai Watier, F-78400 Chatou
September 23, 2021
Abstract
In this paper, we propose a numerical scheme for the layer-averaged Euler with variable density and the Navier-Stokes-Fourier systems presented in part I (Boittin et al., 2020). These systems model hydrostatic free surface flows with density variations. We show that the finite volume scheme presented is well balanced with regards to the steady state of the lake at rest and preserves the positivity of the water height. A maximum principle on the density is also proved as well as a discrete entropy inequality in the case of the Euler system with variable density.
Some numerical validations are finally shown with comparisons to 3D analytical solutions and experiments.
Keywords: Navier-Stokes equations, free surface flows, variable density flows, layer- averaged formulation, finite volume scheme
Contents
1 Introduction 2
2 The layer-averaged models 4
2.1 The layer-averaged Navier-Stokes-Fourier model . . . 6
2.2 The layer-averaged Euler system . . . 10
3 Numerical scheme for the layer-averaged Euler system 10 3.1 Strategy for the time discretization . . . 11
3.2 Semi-discrete (in time) scheme . . . 11
3.3 Eigenvalues for the advection and pressure part . . . 13
3.4 Finite volume formalism for the Euler part . . . 15
3.4.1 The horizontal fluxes and the pressure terms . . . 17
3.4.2 The hydrostatic reconstruction technique . . . 18
3.4.3 The vertical exchange terms . . . 19
3.4.4 The variable update . . . 20
3.4.5 Properties of the numerical scheme . . . 21
3.5 Kinetic fluxes . . . 24
3.6 Discrete entropy inequality . . . 25
4 Numerical scheme for the layer-averaged Navier-Stokes-Fourier sys- tem 36 4.1 Semi-discrete (in time) scheme . . . 36
4.2 Spatial discretization of the diffusion terms . . . 36
5 Numerical validation 38 5.1 Parabolic bowl . . . 38
5.2 Lock exchange . . . 42
5.3 Comparison with the Boussinesq assumption . . . 43
5.3.1 Temperature diffusion . . . 44
5.3.2 Thermal equilibrium . . . 47
6 Conclusion 48
1 Introduction
In this paper we present a numerical scheme for the 3D incompressible Navier-Stokes- Fourier system with variable density and free surface, as well as numerical test cases.
This model describes variable density flows with free surface, the density variations coming from differences in temperature and/or salinity. The model is presented in the companion paper (Boittin et al., 2020), in which a layer-averaged formulation is also given. The layer-averaged formulation suppresses the need for moving meshes (Decoene
and Gerbeau, 2009), (Donea et al., 2004). It allows to perform 3D simulations with a 2D fixed mesh.
Variable density flows are frequently studied by oceanographers. Different systems of coordinates exist, among which terrain-following coordinates and isopycnal coordi- nates. For a discussion of the advantages and disadvantages of the various coordinates frequently used in ocean models, the reader is referred to (Griffies et al., 2000) and (Song and Hou, 2006). The layer-averaged model presented here is not a terrain-following co- ordinate model. Though the layer thicknesses are defined as fractions of the total water height, it is not aσ-coordinate system. The model also differs from isopycnal coordinate models because in the layer-averaged formulation, the layers exchange mass between themselves, which means that the internal layer boundaries are actually not physical.
For the Euler part of the Navier-Stokes-Fourier system with variable density, a finite- volume formalism is adopted. The hydrostatic reconstruction technique is used (Audusse et al., 2004). Therefore, the topography is accurately represented and the scheme is well- balanced. Yet the discretization of the nonconservative pressure terms demands special care. For the viscosity terms, we use finite elements as in (Allgeyer et al., 2019).
In (Audusse et al., 2011), a similar model was studied and simulated. The scheme presented in (Audusse et al., 2011) was a 2D (x−z) scheme and relied on a kinetic interpretation. In the present work we present a fully 3D scheme which is more flex- ible in a certain sense, because the kinetic flux is only one of the possible choices for the numerical flux. With any flux consistent with the semi-discrete in time Euler sys- tem, the resulting scheme is well-balanced and preserves the nonnegativity of the water depth. A maximum principle on the density is satisfied. In order to prove an in-cell entropy inequality, we adopt a kinetic flux, already used in the context of the Shallow Water equations in (Perthame and Simeoni, 2001), (Audusse et al., 2011). The entropy inequality is satisfied for a constant topography and includes third-order rest terms.
Moreover, the unknowns in (Audusse et al., 2011) were not the same, which resulted in a complicated numerical scheme - nonlinear systems were solved at each step and a Newton fixed-point method was used. The present scheme is simpler and does not involve nonlinear systems. Finally, the present scheme is more stable than the scheme in (Audusse et al., 2011), the CFL condition of which could actually degenerate and give a time step equal to zero. With the proposed scheme, the computational cost of the simulation of a non-Boussinesq flow is not greater than that of the simulation of a Boussinesq flow.
The proposed numerical scheme is validated on three test cases simulated with the Freshkiss3d code Freshkiss3d (2020). Each of these test cases allows to validate an aspect of the numerical scheme: wet/dry interface treatment, buoyancy terms, diffusion effects, second order extensions (space and time). . . A real application of hydrodynamics in a river with chlorides entries will be presented in a forthcoming paper (the research report is available (Souillé et al., 2017), in French).
The first test is a convergence test towards an analytical solution (Bristeau et al., 2020) for the Euler system with variable density. In the second test, a lock exchange simulation is performed and the results are compared with experimental data available
from the literature (Adduce et al., 2012). Finally, in two diffusion cases, the differ- ences between the Navier-Stokes-Fourier with variable density and Boussinesq models are evidenced.
The paper is organized as follows. In section 2, the layer-averaged Navier-Stokes- Fourier system and the Euler with variable density introduced in (Boittin et al., 2020) are recalled. A numerical scheme for the layer-averaged Euler model with variable density is presented in Section 3, its properties are studied. An extension of this scheme for the layer-averaged Navier-Stokes-Fourier model with variable density is presented in section 4. The numerical test cases are presented in section 5.
2 The layer-averaged models
hsec:reminder_modelsiFrom now on, we just call the Euler system, the Euler system with variable density. We briefly recall the features of the multilayer models presented in (Boittin et al., 2020) and studied here. The multilayer Navier-Stokes-Fourier model is a layer-averaged version of the incompressible, hydrostatic Navier-Stokes-Fourier system
∇ ·U=−ρ0(T)
ρ2cp ∇ ·(λ∇T) +µ|∇x,yu|2+µ
∂u
∂z
2!
, (1) eq:NSF_1_p2
∂ρ
∂t +∇ ·(ρU) = 0, (2)?eq:NSF_2_p2?
∂(ρu)
∂t +∇x,y·(ρu⊗u) + ∂(ρuw)
∂z +∇x,y Z η
z
ρgdz1 =µ∆x,yu+µ ∂
∂z ∂u
∂z
,(3) eq:NSF_3_p2
whereU(t, x, y, z) = (u, v, w)T is the velocity,u= (u, v)T is the horizontal velocity vector and ρ is the density. The notation ∇ denotes ∇ =
∂
∂x,(∂y∂,(∂z∂T
, ∇x,y corresponds to the projection of ∇ on the horizontal plane i.e. ∇x,y =
∂
∂x,∂y∂ T
and the quantity
|∇x,yu|2 means |∇x,yu|2 = (∇x,yu) : (∇x,yu)T.
The temperature T is linked to the density ρ via the state equation T =T(ρ). The viscosity coefficient is denoted by µ, the heat conductivity by λ and the specific heat capacity at constant pressure bycp.
The energy balance for model (1)-(3) is
∂
∂t
ρ|u|2 2 +ρe
+∇x,y·
u
ρ|u|2 2 +
Z η z
ρgdz1+ρe
−µ∇|u|2 2
=∇.(λ∇T), with e is the internal energy of the fluid governed by
∂(ρe)
∂t +∇x,y·(ρeu) =Z η z
ρgdz1
∇x,y·u+µ|∇x,yu|2+µ
∂u
∂z
2
+∇.(λ∇T).
We consider a free surface flow, therefore we assume
zb(x, y)≤z ≤η(t, x, y) :=h(t, x, y) +zb(x, y),
Figure 1: Flow domain with water height h(t, x, y), free surface η(t, x, y) and bottom zb(x, y).
hfig:notations3d_rhoi
with zb(x, y)the bottom elevation and h(t, x, y)the water depth, see Fig. 1.
Let nb and ns be the unit outward normals at the bottom and at the free surface respectively defined by
nb = 1
p1 +|∇x,yzb|2
∇x,yzb
−1
, and ns= 1
p1 +|∇x,yη|2
−∇x,yη 1
. On the bottom we prescribe an impermeability condition
U.nb = 0, (4) eq:bottom_p2
and a friction condition given e.g. by a Navier law µ
q
1 +|∇x,yzb|2 ∂u
∂nb =−κu, (5)?eq:fric_p2?
with κ a Navier coefficient.
On the free surface, the kinematic boundary condition
∂η
∂t +u(t, x, y, η)· ∇x,yη−w(t, x, y, η) = 0, (6) eq:free_surf_p2
is satisfied, along with the no stress condition µ∂u˜
∂ns
= 0, (7)?eq:bound3ns_p2?
with u˜ = (u,0)T.
On solid walls, we prescribe a slip condition
U·n = 0, (8)?eq:slip?
coupled with an homogeneous Neumann boundary condition µ∂u
∂n = 0, n being the outward normal to the considered wall.
Boundary conditions for the temperature also have to be considered, we can choose either Neumann or Dirichlet conditions namely at the bottom
λ∂T
∂nb =φT|zb, (9) BC:neumann_bottom_p2
or
Tb =Tb0 (10) BC:dirichlet_bottom_p2
and at the free surface
λ∂T
∂ns =φT|η (11) BC:neumann_surface_p2
or
Ts=Ts0 (12) BC:dirichlet_surface_p2
where φT|zb, φT|η are two given temperature fluxes and Tb0, Ts0 are two given tempera- tures. SinceT =T(ρ), the boundary conditions forρnaturally ensue from the boundary conditions forT.
The system is completed with some initial conditions
h(0, x, y) =h0(x, y), ρ(0, x, y) =ρ0(x, y), U(0, x, y, z) =U0(x, y, z).
The system (1)-(3) was derived from the compressible Navier-Stokes system in (Boittin et al., 2020). More specifically, the derivation consisted in performing the incompressible limit. This model respects the second principle of thermodynamics (non-decreasing entropy).
2.1 The layer-averaged Navier-Stokes-Fourier model
?hsec:nsf_multilayeri?We consider a discretization of the fluid domain by layers (see Fig. 2) where the layer α contains the points of coordinates (x, y, z) with z ∈ Lα(t, x, y) = (zα−1/2, zα+1/2) and {zα+1/2}α=1,...,N is defined by
zα+1/2(t, x, y) = zb(x, y) +Pα
j=1hj(t, x, y), α ∈[0, . . . , N],
hα(t, x, y) = zα+1/2(t, x, y)−zα−1/2(t, x, y) =lαh(t, x, y), (13) eq:layer and PN
α=1lα = 1.
Using the notations (13), let us consider the spacePN,t0,h of piecewise constant functions defined by
PN,t0,h =
1z∈Lα(t,x,y)(z), α∈ {1, . . . , N} , (14)?eq:P0_space?
Figure 2: Notations for the layerwise discretization.
hfig:free_p2i
where 1z∈Lα(t,x,y)(z) is the characteristic function of the layer Lα(t, x, y). Using this formalism, the projection of ρ, u, v and w on PN,t0,h is a piecewise constant function defined by
XN(t, x, y, z,{zα}) =
N
X
α=1
1z∈Lα(t,x,y)(z)Xα(t, x, y), (15)?eq:ulayer?
for X ∈(ρ, u, v, w). When the quantities {ρα(t, x, y)}α=1,...,N are known, if the function T =T(ρ)is known, it is possible to recover the temperature using the formula
TN(t, x, z,{zα}) =
N
X
α=1
1z∈Lα(t,x,y)(z)T(ρα(t, x, y)).
The layer-averaged Navier-Stokes-Fourier system introduced in (Boittin et al., 2020) reads
∂h
∂t +
N
X
α=1
∇x,y·(hαuα) =−
N
X
α=1
ρ0(Tα)
ρ2αcp (ST ,α− Sµ,α), (16) eq:massesvml_ns_p2
∂ραhα
∂t +∇x,y·(ραhαuα) =ρα+1/2Gα+1/2−ρα−1/2Gα−1/2, α = 1, ..., N, (17)?eq:massesvml1_ns_p2?
∂ραhαuα
∂t +∇x,y·(ραhαuα⊗uα) +∇x,y hαpα
=pα+1/2∇x,yzα+1/2−pα−1/2∇x,yzα−1/2 +uα+1/2ρα+1/2Gα+1/2−uα−1/2ρα−1/2Gα−1/2+∇x,y·(µhα∇x,yuα)
+Γα+1/2(uα+1−uα)−Γα−1/2(uα−uα−1)−καuα, α= 1, ..., N, (18) eq:mvtsvml_ns_p2
with
Gα+1/2 =−
N
X
j=1
Xα
p=1
lp−1j≤α
∇x,y·(hjuj) +
α
X
j=1
ρ0(Tj)
ρ2jcp (ST ,j−Sµ,j), (19)?eq:Qalphabis_T_p2?
κα =
(κ if α= 1 0 if α 6= 1 ,
ST,α =
λ∇x,y·(hα∇x,yTα) + 2λα+1/2Tα+1−Tα
hα+1+hα −2λα−1/2Tα−Tα−1
hα+hα−1
, (20) S_T_alpha λα+1/2 =λ for α= 1, . . . , N −1,
Tα =T(ρα).
Forα= 0,2λα+1/2Thα+1−Tα
α+1+hα =φT|zb if the Neumann boundary condition (9) is chosen, or h0 =h1,T0 =Tb0if the Dirichlet boundary condition (10) is chosen. Likewise, forα=N, 2λα+1/2Thα+1−Tα
α+1+hα = φT|η with the boundary condition (11), or hN+1 = hN, TN+1 = Ts0 with the boundary condition (12). The dissipation term due to the viscous effects is
Sµ,α =−hαµ|∇x,yuα|2−Γα+1/2|uα+1−uα|2
2 −Γα−1/2|uα−uα−1|2
2 −κα|uα|2, (21) S_mu_alpha with
Γα+1/2 = 2µα+1/2 hα+1+hα
, (22) Gamma_alpha_plus
µα+1/2 =
0 if α= 0
µ if α= 1, . . . , N −1 0 if α=N.
(23)?mu_alpha_plus?
The term |∇x,yuα|2 actually denotes
|∇x,yuα|2 = (∇x,yuα) : (∇x,yuα)T
=
∂uα
∂x 2
+ ∂uα
∂y 2
+ ∂vα
∂x 2
+ ∂vα
∂y 2
. The velocitiesuα+1/2 and the densities ρα+1/2 at the interfaces are defined by
vα+1/2 =
vα if Gα+1/2 ≤0
vα+1 if Gα+1/2 >0 (24)?eq:upwind_uT_p2?
forv =u, ρ.
The pressure termspα,pα+1/2 are given by
pα=g ραhα 2 +
N
X
j=α+1
ρjhj
!
and pα+1/2 =g
N
X
j=α+1
ρjhj. (25)?eq:palpha1_p2?
The pressure is hydrostatic. The terms Gα+1/2 represent the mass exchanges between the layers. Notice that some of the viscous and diffusion terms have been simplified, see (Boittin et al., 2020). Since the right-hand side of the total height conservation equation (16) is nonzero, we expect to observe dilatation and contraction due to the temperature diffusion and to the viscosity.
We recall here the following result, obtained in (Boittin et al., 2020).
?hcontinuous_energy_balance_nsi?Proposition 2.1 The system (16)-(18) completed with the equation
∂
∂t(ραhαeα) +∇x,y·(ραhαuαeα) =ρα+1/2eα+1/2Gα+1/2−ρα−1/2eα−1/2Gα−1/2
+pαρ0(Tα)
ρ2αcp (ST ,α− Sµ,α) +ST ,α− Sµ,α admits, for smooth solutions, the energy balance
∂
∂tEα+∇x,y·(uα(Eα+hαpα−µhα∇x,yuα)) +Γα+1/2|uα+1|2− |uα|2
2 −Γα−1/2
|uα|2− |uα−1|2 2
= ρα+1/2u2α+1/2
2 +gρα+1/2zα+1/2
!
Gα+1/2+pα+1/2 Gα+1/2−∂zα+1/2
∂t
− ρα−1/2
u2α−1/2
2 +gρα−1/2zα−1/2
!
Gα−1/2−pα−1/2 Gα−1/2− ∂zα−1/2
∂t
−1
2 ρα+1/2(uα+1/2−uα)2+ghα(ρα+1/2−ρα)
Gα+1/2 +1
2 ρα−1/2(uα−1/2 −uα)2−ghα(ρα−1/2−ρα)
Gα−1/2+ST ,α, (26) eq:energy_mcl_ns_p2
with
Eα =ραhα|uα|2
2 + ραghαzα
2 +eα. (27)?eq:energ_al_NS_p2?
Note that in (26), we use the notation uα∇x,yuα =
uα∂u∂xα +vα∂v∂xα uα∂u∂yα +vα∂v∂yα
=∇x,y|u2α| 2 . The sum of Eqs. (26) for α= 1, . . . , N gives
∂
∂t
N
X
α=1
Eα+
N
X
α=1
∇x,y·uα(Eα+hαpα)
= −
N
X
α=1
ρα+1/2|uα+1−uα|2
2 |Gα+1/2|+
N
X
α=1
ST ,α
−g 2
N
X
α=1
hα(ρα+1/2−ρα) +hα+1(ρα+1/2−ρα+1)
Gα+1/2. The sum of ST ,α over the layers gives
N
X
α=1
ST ,α =λ
N
X
α=1
∇x,y·(hα∇x,yTα)− ∇T|s·ns+∇T|b·nb.
As explained in (Boittin et al., 2020), the terms on the last line of the right-hand side are third-order terms.
2.2 The layer-averaged Euler system
What we refer to hereafter as the layer-averaged Euler system is the system (16)-(18) without viscosity and without diffusion terms, i.e.
∂h
∂t +
N
X
α=1
∇x,y·(hαuα) = 0, (28) eq:massesvml_p2
∂ραhα
∂t +∇x,y·(ραhαuα) =ρα+1/2Gα+1/2−ρα−1/2Gα−1/2, α = 1, ..., N, (29)?eq:massesvml1_p2?
∂ραhαuα
∂t +∇x,y·(ραhαuα⊗uα) +∇x,y hαpα
=pα+1/2∇x,yzα+1/2−pα−1/2∇x,yzα−1/2
+uα+1/2ρα+1/2Gα+1/2−uα−1/2ρα−1/2Gα−1/2, α = 1, ..., N. (30) eq:mvtsvml_p2
The quantity Gα+1/2 (resp. Gα−1/2) corresponds to mass exchange accross the interface zα+1/2 (resp. zα−1/2) and Gα+1/2 is defined here by
Gα+1/2 =
α
X
j=1
∂hj
∂t +∇x,y·(hjuj)
=−
N
X
j=1
Xα
p=1
lp−1j≤α
∇x,y·(hjuj), (31) eq:Qalphabis_p2
forα = 1, . . . , N. Notice that the mass conservation for the layer α writes
∂hα
∂t +∇x,y·(hαuα) = Gα+1/2−Gα−1/2, (32) eq:exchange_terms
and the sum (total or partial) of the previous equations with G1/2 = GN+1/2 = 0 (see (Boittin et al., 2020, Prop. 3.1)) gives Eqs. (28) and (31).
The energy balance verified by the system (28)-(30) is given in (Boittin et al., 2020).
It is very similar to the balance (26), obviously without the viscosity and diffusion terms. In the balance for the Euler system, the internal energy eα does not intervene.
It is actually equal to 0 because there is no volume variation.
3 Numerical scheme for the layer-averaged Euler sys- tem
hsec:euler_numi
In this section, a numerical scheme for the layer-averaged Euler system is designed and analyzed. It extends the work done by some of the authors in (Allgeyer et al., 2019;
Audusse et al., 2011). Before specifying the scheme for the Euler system, a common strategy for the time discretization of the Navier-Stokes-Fourier and Euler systems is presented. The discretization of the diffusion terms does not present any additional difficulty, however including these terms considerably lengthens the equations. This is why it seems preferable to explain the numerical scheme for the Euler system first.
The advantages of the numerical scheme are the following
• it gives a 3D approximation of the Navier-Stokes-Fourier system, while only 2D situations were considered in (Audusse et al., 2011)
• it can be implemented with any flux that is consistent with the homogeneous Saint- Venant system; the kinetic flux is used only for the discrete entropy property stated for a constant topography
• the scheme is endowed with strong stability properties (well-balanced, positivity of the water depth)
• convergence curves towards a 3D non-stationary analytical solution with wet-dry interfaces are obtained, see paragraph 5.1.
3.1 Strategy for the time discretization
The system (16)-(18) has the form
∂U
∂t +∇x,y·F(U) = Sp(U, zb) +Se(U, ∂tU, ∂xU) +Sv,f(U), (33) eq:glo_p2 where the vector of unknowns is
U= (h, ρ1h1, . . . , ρNhN, qx,1, . . . , qx,N, qy,1, . . . , qy,N)T ,
with qx,α = ραhαuα, qy,α = ραhαvα. We denote by F(U) = (Fx(U), Fy(U))T the fluxes of the conservative part and by
Sp(U, zb) =
0, . . . , p3/2∂z3/2
∂x −p1/2∂z1/2
∂x , . . . , p3/2∂z3/2
∂y −p1/2∂z1/2
∂y , . . . T
, the non-conservative part of the pressure terms. The source terms areSe(U, ∂tU, ∂xU) and Sv,f(U), representing respectively the mass and momentum exchanges and the vis- cous and friction effects. Notice that, as a consequence of the layer-averaged discretiza- tion, the system (33) is made of only 2d(x, y) partial differential equations with source terms. Hence, the spacial approximation of the considered PDEs is performed on a 2d planar mesh.
We consider discrete times tn with tn+1 = tn+ ∆tn. For the time discretisation of the layer-averaged Navier-Stokes-Fourier system (33) we adopt the following scheme
Un+1 =U−∆tn(∇x,y·F(U)− Sp(U, zb)) + ∆tnSen+1+ ∆tnSv,fn+l, (34) eq:glo_dis where the integer l = 0,1/2,1 will be precised below. In (34) and wherever there is no
ambiguity the superscript n has been omitted.
3.2 Semi-discrete (in time) scheme
From now on and until the end of this section, the system considered is the Euler system.
Similarly to (Allgeyer et al., 2019), the semi-discrete in time scheme (34) withSv,fn+l = 0 writes
hn+1/2α = hα−∆tn∇x,y(hαuα) (35) eq:nscont_layer_d
(ραhα)n+1/2 = ραhα−∆tn∇x,y·(ραhαuα) (36) eq:nsml22_dbis
(ραhαuα)n+1/2 = ραhαuα−∆tn
∇x,y·(ραhαuα⊗uα) +∇x,y hαpα
−pα∇x,yhα+ραghα∇x,yzα
, (37) eq:nsml22_d
hn+1 = hn+1/2 =
N
X
α=1
hn+1/2α =h−∆tn
N
X
α=1
∇x,y·(hαuα), (38) eq:nsml11_d
(ραhα)n+1 = (ραhα)n+1/2+ ∆tn
ρn+1α+1/2Gα+1/2−ρn+1α−1/2Gα−1/2
, (39) eq:nsml33bis_d
(ραhαuα)n+1 = (ραhαuα)n+1/2+ ∆tn
un+1α+1/2ρn+1α+1/2Gα+1/2−un+1α−1/2ρn+1α−1/2Gα−1/2
(40), eq:nsmlbis_d
with
Gα+1/2 = −
N
X
j=1
Xα
p=1
lp−1j≤α
∇x,y·(hjuj). (41) eq:nsf_6d
Since the relations
pα±1/2 =pα∓ ραghα
2 , zα = zα+1/2+zα−1/2 2
hold, in (37) we often use the identity
pα+1/2∇x,yzα+1/2−pα−1/2∇x,yzα−1/2 =pα∇x,yhα−ραghα∇x,yzα. (42) eq:S_P_alpha
Notice that hn+1α =lαhn+1 and hn+1α 6=hn+1/2α , see Eq. (32).
The first three equations (35)-(37) consist in an explicit time scheme where the horizontal fluxes and the pressure terms are taken into account whereas in Eqs. (39)-(40) an implicit treatment of the exchange terms between layers is proposed. The implicit part of the scheme requires to solve a linear problem, see paragraph 3.4.3 and Allgeyer et al. (2019).
Following (34), Eqs. (35)-(38) also writes
Un+1/2 =U−∆tn(∇x,y·F(U)− Sp(U, zb)), (43) eq:glo_dis1
these computations being detailed in paragraphs 3.4.1 and 3.4.2. Equations. (38)-(40) can be reformulated under the form
Un+1 =Un+1/2 + ∆tnSen+1, (44)?eq:glo_dis2?
this implicit step being precised in paragraph 3.4.3.
3.3 Eigenvalues for the advection and pressure part
In order to propose a finite volume discretisation for Eqs. (35)-(38), an estimation of the eigenvalues is necessary and given below.
Without the exchange terms, the system (28)-(30) writes
∂h
∂t +
N
X
α=1
∇x,y·(hαuα) = 0, (45) eq:massesvml_se
∂(ραhα)
∂t +∇x,y·(ραhαuα) = 0, (46)?eq:massesvml1_se?
∂(ραhαuα)
∂t +∇x,y·(ραhαuα⊗uα) +pα∇x,yhα−ραghα∇x,yzα = 0, (47) eq:mvtsvml_se
with
pα∇x,yhα−ραghα∇x,yzα=ghα
N
X
j=α+1
∇x,y(ρjhj) + 1
2∇x,y(ραhα)
!
+ ραghα
2 ∇x,yhα+ραghα
α−1
X
j=1
∇x,yhj+∇x,yzb
! . This system is the continuous version of system (35)-(38).
We rewrite the system (45)-(47) under the form
∂h
∂t +
N
X
j=1
(ljh∇x,y·uj) +
N
X
j=1
(ljuj)· ∇x,yh= 0, (48) eq:massesvml_se1
∂ρα
∂t +uα.∇x,yρα = 0, (49)?eq:massesvml1_se1?
∂uα
∂t + (uα· ∇x,y)uα+ 1
ραhα(pα∇x,yhα−ραghα∇x,yzα) = 0, (50) eq:mvtsvml_se1
and the quasilinear form of the system (48)-(50) writes
∂U˜
∂t +A( ˜U)∇x,yU˜ =sb( ˜U), (51) eq:quasi_lin
with U˜ = (h,u1, . . . ,uN, ρ1, . . . , ρN)T, and
A( ˜U) =
A1( ˜U) A2( ˜U) A3( ˜U) A4( ˜U)
,
A1( ˜U) =
PN
j=1ljuj l1h . . . lNh
˜
p1 u1 0 . . . 0
˜
p2 0 u2 0 . . . 0 ... 0 ... uj ... 0
... 0 ... 0 ... 0
˜
pN 0 0 . . . 0 uN
,
A2( ˜U) =
0 . . . 0
gh21
2 0 ... 0
gh2h1 gh222 0 0 ... 0 gh
2 j
2 0
ghNh1 0 0 gh22N
,
A3( ˜U) = 0N×N+1, A4( ˜U) =diag(uj),
˜ pj = g
ρj ρjlj+ρj
j−1
X
i=1
li+
j−1
X
i=1
ρili
! .
For the sake of simplicity, the expression of the matrixA( ˜U)given below corresponds to the 1D case i.e. forvi = 0,i= 1, . . . , N. Notice thatA2( ˜U)and A3( ˜U)T are rectangular matrices with N + 1 rows and N columns.
The following proposition holds, consisting in a version of the Cauchy’s interlace theorem (Hwang, 2004) in the case of non symmetric matrix.
Proposition 3.1 The system (51) is strictly hyperbolic for h > 0 and the eigenvalues of A( ˜U) belong to the interval (λmin, λmax) with
λmin = min
j {uj, vj} − max{ρj} min{ρj}
pgh, λmax = max{uj, vj}+max{ρj}
min{ρj} pgh.
hprop:hyperi
Proof of prop. 3.1 Since A( ˜U) is a block-matrix, its eigenvalues consist in the eigen- values ofA1( ˜U) completed with the set{ui}Ni=1. Writing the characteristic polynomial of A1( ˜U) under the form (development e.g. along the first row)
PA1 = ΠNi=1(λ−ui) λ−
N
X
j=1
ljuj−
N
X
i=1
lih˜pi λ−ui
! ,
the eigenvalues of A1( ˜U) satisfy λ−
N
X
j=1
ljuj =QA1(λ),
Figure 3: The two functions λ 7→ QA1(λ) and λ 7→ λ−PN
j=1ljuj, each intersection of the two curves is an eigenvalue of A1( ˜U).
hfig:pol_caraci
withQA1(λ) = PN i=1
lihp˜i
λ−ui. ForN = 3, the functions λ7→QA1(λ) andλ7→λ−PN j=1ljuj are depicted over Fig. 3 and it is easy to see that the four eigenvalues λi exists with the interlacing
λ1 < u1 ≤λ2 ≤. . .≤u3 < λ4. Moreover we have
QA1(λmax) =
N
X
i=1
lih˜pi λmax−ui
≤ min{ρj} max{ρj}
N
X
i=1
lih˜pi
√gh ≤ max{ρj} min{ρj}
pgh≤λmax−
N
X
j=1
ljuj, and likewise
QA1(λmin) =
N
X
i=1
lihp˜i λmin−ui
≥ −min{ρj} max{ρj}
N
X
i=1
lihp˜i
√gh ≥ −max{ρj} min{ρj}
pgh≥λmin−
N
X
j=1
ljuj,
therefore the eigenvalues {λi}Ni=1 of A1( ˜U) satisfy
λmin ≤λi ≤λmax, i= 1, . . . N, proving the result.
3.4 Finite volume formalism for the Euler part
hsubsec:fviIn this paragraph, we propose a space discretization for the model (35)-(40) completed with (41). We first recall the general formalism of finite volumes on unstructured meshes.
LetΩdenote the computational domain with boundaryΓ, which we assume is polygonal.
LetTh be a triangulation of Ω for which the vertices are denoted byPi with Si the set of interior nodes and Gi the set of boundary nodes. The dual cells Ci are obtained by joining the centers of mass of the triangles surrounding each vertex Pi. We use the following notations (see Fig. 4):
• Ki, set of subscripts of nodesPj surrounding Pi,
• |Ci|, area of Ci,
• Γij, boundary edge between the cells Ci and Cj,
• Lij, length of Γij,
• nij, unit normal toΓij, outward to Ci (nji =−nij).
IfPi is a node belonging to the boundary Γ, we join the centers of mass of the triangles adjacent to the boundary to the middle of the edge belonging toΓ (see Fig. 4) and we denote
• Γi, the two edges of Ci belonging to Γ,
• Li, length ofΓi (for sake of simplicity we assume in the following thatLi = 0 if Pi
does not belong to Γ),
• ni, the unit outward normal defined by averaging the two adjacent normals.
(a) (b)
Figure 4: (a) Dual cell Ci and (b) Boundary cellCi.
hfig:meshi
We define the piecewise constant functionsUn(x, y)on cellsCi corresponding to time tn as
Un(x, y) =Uni, for (x, y)∈Ci,
with Uni = (hni, ρn1,ihn1,i, . . . , ρnN,ihnN,i, qx,1,in , . . . , qx,N,in , qy,1,in , . . . , qny,N,i)T i.e.
Uni ≈ 1
|Ci| Z
Ci
U(tn, x, y)dxdy.
For the topography, we choose a piecewise constant approximation under the form zi ≈ 1
|Ci| Z
Ci
zb(x, y)dxdy.
We will also use the notation
Unα,i ≈ 1
|Ci| Z
Ci
Uα(tn, x, y)dxdy, with Uα defined by
Uα = (hα, ραhα, ραhαuα, ραhαvα)T. (52) eq:u_alpha 3.4.1 The horizontal fluxes and the pressure terms
hsubsubsec:horizontal_discreteiA finite volume scheme for solving the system (35)-(38) is a formula of the form Un+1/2i =Ui− X
j∈Ki
σi,jFi,j−σiFe,i+X
j∈Ki
σi,jSp(Ui,Uj, zb,i, zb,j), (53) eq:upU0 where using the notations of (34)
X
j∈Ki
Li,jFi,j ≈ Z
Ci
∇x,y·F(U)dxdy, (54) eq:flux_dis1
with
σi,j = ∆tnLi,j
|Ci| , σi = ∆tnLi
|Ci| . Here we consider first-order explicit schemes where
Fi,j =
Fi,jh Fi,jρ1h1
...
Fi,jρNhN Fi,ju1
...
Fi,juN
. (55) eq:flux_def
and
Fi,jh =
N
X
α=1
Fi,jhα, (56) eq:flux
and for the boundary nodes
Fi,e=
Fi,eh Fi,eρ1h1
...
Fi,eρNhN Fi,eu1
...
Fi,euN
. (57) eq:fluxbis
The fluxesFi,jhα,Fi,jραhα,Fi,juα appearing in expressions (55),(56), (57) are numerical fluxes such that
Fi,jmα =Fmα(Uα,i,Uα,j,ni,j), with m=h, ρh,u, α= 1, . . . , N.
Relation (53) tells how to compute the values Un+1/2i knowing Ui and discretized valueszb,i of the topography. Following (54), the term Fi,j in (53) denotes an interpola- tion of the normal component of the flux F(U).ni,j along the edge Ci,j. The functions F(Ui,Uj,ni,j)∈R2N+1 are the numerical fluxes, see (Bouchut, 2004).
Until now, the expression for the numerical fluxes is not detailed since any numerical fluxes (Rusanov, HLL,. . . ) can be used (Bouchut, 2004). In paragraph 3.5 we define F(Ui,ni,j)using kinetic fluxes and we prove a discrete entropy inequality for the system.
The computation of the value Ui,e, which denotes a value outside Ci (see Fig. 4-(b)), defined such that the boundary conditions are satisfied, and the definition of the bound- ary flux F(Ui,Ue,i,ni) are described in (Allgeyer et al., 2019). Notice that we assume a flat topography on the boundaries i.e. zb,i =zb,i,e.
For the discretization of the pressure source term Sp(U, zb), we adopt a strategy defined below.
3.4.2 The hydrostatic reconstruction technique
hsubsubsec:HRiThe hydrostatic reconstruction scheme (HR scheme for short) for the Saint-Venant sys- tem has been introduced in (Audusse et al., 2004) in the 1d case and described in 2d for unstructured meshes in (Audusse and Bristeau, 2005). The HR in the context of the kinetic description for the Saint-Venant system has been studied in (Audusse et al., 2016).
In order to take into account the topography variations and to preserve relevant equilibria, the HR leads to a modified version of (53) under the form
Uin+1/2 =Ui −X
j∈Ki
σi,jFi,j∗ −σiFi,e+ X
j∈Ki
σi,jSp,i,j∗ , (58) eq:upU0_HR where
Fi,j∗ = F(Ui,j∗ , Uj,i∗ ,ni,j), (59)?eq:flux_HR?
Sp,i,j∗ = Sp(Ui, Ui,j∗ , zb,i, zb,j,ni,j)
=
0
˜
p∗1,i,j(h1,i,j−h1,i)ni,j −gρ1,i˜h∗1,i,j(z1,i,j−z1,i)ni,j ...
˜
p∗α,i,j(hα,i,j−hα,i)ni,j −gρα,i˜h∗α,i,j(zα,i,j−zα,i)ni,j ...
(60) eq:Sp_HR
with
zb,i,j∗ = max(zb,i, zb,j), h∗i,j = max(hi+zb,i−zb,i,j∗ ,0),
Ui,j∗ = (h∗i,j, ρ1,il1h∗i,j, . . . , ρN,ilNh∗i,j, ρ1,il1h∗i,ju1,i, . . . , ρN,ilNh∗i,juN,i, . . .)T, zα,i,j∗ =zb,i,j∗ +
lα
2 +Pα−1 j=1 lj
h∗i,j, zα,i,j =zα,j,i= z
∗
α,i,j+zα,j,i∗
2 ,
hα,i,j =hα,j,i= h
∗
α,i,j+h∗α,j,i
2 ,
˜h∗α,i,j = hα,i+h
∗ α,i,j
2 ,
˜
p∗α,i,j = pα,i+p
∗ α,i,j
2 ,
(61) eq:state_HR
and
zα =zb+ lα 2 +
α−1
X
j=1
lj
!
h, pα = ραghα
2 +
N
X
j=α+1
ρjghj. Throughout this work, the∗ refers to the HR technique.
Remark 3.2 Since the quantity S(U, zb) appearing in (33) contains non conservative terms, its integration over the cell Ci is not straightforward and we have used the result proposed by Bouchut (Bouchut, 2004, Proposition 5.3) to obtain the expression (60). 3.4.3 The vertical exchange terms
hsubsubsec:exchangesiWe give the fully discrete expression of the step for the vertical exchanges, described by equations (39)-(40). The step for the vertical exchanges consists in
Uin+1 =Uin+1/2+ ∆tnGin+1, (62) eq:upU1_HR
with
Gin+1 =
0 ρn+13/2,iG3/2,i ρn+15/2,iG5/2,i−ρn+13/2,iG3/2,i
...
ρn+1N−1/2,iGN−1/2,i−ρn+1N−3/2,iGN−3/2,i
−ρn+1N−1/2,iGN−1/2,i
un+13/2,iρn+13/2,iG3/2,i
un+15/2,iρn+15/2,iG5/2,i−un+13/2,iρn+13/2,iG3/2,i ...
un+1N−1/2,iρn+1N−1/2,iGN−1/2,i−un+1N−3/2,iρn+1N−3/2,iGN−3/2,i
−un+1N−1/2,iρn+1N−1/2,iGN−1/2,i
.
The mass conservation for the layerαis governed by Eq. (31) and its discretization gives hn+1α,i =hn+1/2α,i + ∆t(Gα+1/2,i−Gα−1/2,i).