Professional Documents
Culture Documents
net/publication/226045321
CITATIONS READS
6 210
4 authors, including:
Jean-Claude Latché
Institut de Radioprotection et de Sûreté Nucléaire (IRSN)
77 PUBLICATIONS 1,576 CITATIONS
SEE PROFILE
All content following this page was uploaded by Raphaèle Herbin on 16 May 2014.
Abstract
We propose a discretization for the MAC scheme of the viscous dissipation term
τ (u) : ∇u (where τ (u) stands for the shear stress tensor associated to the velocity
hal-00576268, version 1 - 13 Aug 2011
field u), which is suitable for the approximation of this term in a conservation
equation for a scalar variable. This discretization enjoys the property that the
integral over the computational domain Ω of the (discrete) dissipation term is equal
to what is obtained when taking the inner product of the (discrete) momentum
balance equation by u and integrating over Ω. As a consequence, it may be used
as an ingredient to obtain an unconditionally stable scheme for the compressible
Navier-Stokes equations. It is also shown, in some model cases, to ensure the strong
convergence in L1 of the dissipation term.
1 Introduction
Let us consider the compressible Navier-Stokes equations, which may be written as:
∂t ρ + div(ρu) = 0, (1a)
∂t (ρu) + div(ρu ⊗ u) + ∇p − div(τ (u)) = 0, (1b)
∂t (ρe) + div(ρeu) + pdivu + div(q) = τ (u) : ∇u, (1c)
ρ = ℘(p, e), (1d)
where t stands for the time, ρ, u, p and e are the density, velocity, pressure and internal energy
in the flow, τ (u) stands for the shear stress tensor, q for the energy diffusion flux, and the
function ℘ is the equation of state. This system of equations is posed over Ω × (0, T ), where Ω
is a domain of Rd , d ≤ 3. It must be supplemented by a closure relation for τ (u) and for q,
assumed to be:
2µ
τ (u) = µ(∇u + ∇t u) − divu I, q = −λ∇e, (2)
3
Viscous dissipation term discretization with the MAC scheme
(ii) to identify the integral of the dissipation term at the right-hand side of the discrete
counterpart of (1c) with the (discrete) L2 inner product between the velocity and the
hal-00576268, version 1 - 13 Aug 2011
where e(x) stands for the first vector of the canonical basis of R2 . We denote by σi,j x the right
x x
face of Ki− 1 ,j , i.e. σi,j = {xi } × (yj− 1 , yj+ 1 ). Splitting the boundary integral in (4), the part
2 2 2
dif
of T̄i− x , also referred to as the viscous flux through σ x , reads:
associated to σi,j
1
,j i,j
2
hZ 2
i Z Z
(x) x
− τ (u) n dγ · e = −2 µ ∂x u dγ + µ (∂x ux + ∂y uy ) dγ,
x
σi,j x
σi,j 3 x
σi,j
2
Viscous dissipation term discretization with the MAC scheme
hxi− 1
2
hxi
yj+ 3
2
uxi− 1 ,j+1
2
hyj+ 1 x
: Ki− 1
,j
yj+ 1 2 2
2
x
yj uxi− 3 ,j uxi− 1 ,j uxi+ 1 ,j hyj : σi,j
2 2 2
yj− 1 x
: σi− 1
2 ,j+ 1
2 2
uxi− 1 ,j−1
yj− 3 2
2x xi− 1 xi+ 1
i− 3 2
xi−1 2
xi 2
and the usual finite difference technique yields the following approximation for this term:
4 2
Z Z
x
− µ ∂x u dγ + µ ∂y uy dγ
3 σi,j
x 3 σi,j
x
4 hyj 2 hyj
≈ − µi,j x (ui+ 1 ,j − ui− 1 ,j ) + µi,j y (uyi,j+ 1 − uyi,j− 1 ), (5)
x x
3 hi 2 2 3 hj 2 2
x . Similarly, let σ x
where µi,j is an approximation of the viscosity at the face σi,j =
i− 1 ,j+ 1 2 2
(xi−1 , xi ) × {yj+ 1 } be the top edge of the cell. Then:
2
hZ i Z
− τ (u) n dγ · e(x) = − µ (∂y ux + ∂x uy ) dγ
σx 1 1 σx 1 1
i− 2 ,j+ 2 i− 2 ,j+ 2
h hxi− 1 hxi− 1 i
≈ −µi− 1 ,j+ 1 2
(uxi− 1 ,j+1 − uxi− 1 ,j ) + 2
(uyi,j+ 1 − uyi−1,j+ 1 ) ,
2 2 hyj+ 1 2 2 hxi− 1 2 2
2 2
x
where µi− 1 ,j+ 1 stands for an approximation of the viscosity at the edge σi− .
1
2 2 ,j+ 1 2 2
h 4 hy y
2 hj i
dis j x x y y
Ti,j (u, v) = µi,j − (u − u ) + (u − u ) (v xi− 1 ,j − v xi+ 1 ,j )
3 hxi i+ 21 ,j i− 12 ,j 3 hyj i,j+ 21 i,j− 21 2 2
y y
h 4 uxi+ 1 ,j − uxi− 1 ,j x x
2 ui,j+ 21 − ui,j− 21 i v i+ 12 ,j − v i− 12 ,j
= µi,j hyj hxi 2 2
− . (6)
3 hxi 3 hyj hxi
3
Viscous dissipation term discretization with the MAC scheme
x
Similarly, the term associated to σi− appears multiplied by v xi− 1 ,j and by −v xi− 1 ,j+1 , and
1
2
,j+ 21 2 2
we get:
dis
Ti− 1
,j+ 1
(u, v) = µi− 1 ,j+ 1 hxi− 1 hyj+ 1
2 2 2 2 2 2
– For the other derivatives, we introduce another mesh which is vertex-centred, and we
xy
denote by K xy the generic cell of this new mesh, with Ki+ 1
,j+ 1
= (xi , xi+1 ) × (yj , yj+1 ).
hal-00576268, version 1 - 13 Aug 2011
2 2
xy
Then, ∀x ∈ Ki+ 1
,j+ 1
:
2 2
and: Z
dis
Ti− 1
,j+ 1
(u, v) = µi− 1 ,j+ 1 (∂yM ux + ∂xM uy ) ∂yM v x dx.
2 2 2 2
K xy 1 1
i− 2 ,j+ 2
Let us now perform the same operations for the y-component of the velocity. Doing so, we are
y
lead to introduce an approximation of the viscosity at the edge σi− 1
,j+ 1
= {xi− 1 } × (yj , yj+1 )
2 2 2
x
(see Figure 2). Let us suppose that we take the same approximation as on σi− . Then,
1
,j+ 1 2 2
the same argument yields that multiplying each discrete equation for ux and for uy by the
corresponding degree of freedom of a velocity field v, we obtain a dissipation term which reads:
Z
T dis (u, v) = τ M (u) : ∇M v dx, (10)
Ω
where ∇M is the discrete gradient defined by (8)-(9) and τ M the discrete tensor:
2µ ∂xM ux µxy (∂yM ux + ∂xM uy )
τ M (u) = − 2 µ (∂xM ux + ∂yM uy ) I, (11)
µxy (∂yM ux + ∂xM uy ) 2µ ∂yM uy 3
where µ is the viscosity defined on the primal mesh by µ(x) = µi,j , ∀x ∈ Ki,j and µxy is the
xy
viscosity defined on the vertex-centred mesh, by µ(x) = µi+ 1 ,j+ 1 , ∀x ∈ Ki+ 1
,j+ 1
.
2 2 2 2
4
Viscous dissipation term discretization with the MAC scheme
y x y
: Ki,j− 1 : σi− 1
,j+ 1
: σi− 1
,j+ 1
2 2 2 2 2
uyi,j+ 3
yj+ 3 2
2
uyi,j− 1
yj− 1 2
2
xi− 3 xi− 1 xi+ 1 xi+ 3
2 2 2 2
Now the form (10) suggests a natural to discretize the viscous dissipation term in the internal
energy balance in order for the consistency property (ii) to hold. Indeed, if we simply set on
each primal cell Ki,j :
1
Z
(τ (u) : ∇u)i,j = τ M (u) : ∇M u dx, (12)
|Ki,j | Ki,j
holds. Furthermore, we get from Definition (11) that τ M (u)(x) is a symmetrical tensor, for any
i, j and x ∈ Ki,j , and therefore an elementary algebraic argument yields:
1
Z
(τ (u) : ∇u)i,j = τ M (u) : ∇M u dx
|Ki,j | Ki,j
1
Z
τ M (u) : ∇M u + (∇M u)t dx ≥ 0.
=
2 |Ki,j | Ki,j
Remark 1 (Approximation of the viscosity) Note that, for the symmetry of τ M (u) to hold, the
x y
choice of the same viscosity at the edges σi− 1
,j+ 1
and σi− 1
,j+ 1
is crucial even though other
2 2 2 2
choices may appear natural. Assuming for instance the viscosity to be a function of an additional
variable defined on the primal mesh, the following construction seems reasonable:
1. define a constant value for µ on each primal cell,
2. associate a value of µ to the primal edges, by taking the average between the value at the
adjacent cells,
3. x
finally, split the integral of the shear stress over σi− in two parts, one for the part
1
,j+ 1 2 2
included in the (top) boundary of Ki−1,j and the second one in the boundary of Ki,j .
5
Viscous dissipation term discretization with the MAC scheme
zk+ 1
2
zk− 1
2
yj+ 1
2
xy
: Ki+ 1
,j+ 1 ,k
2 2
xi+ 1
2
xy
Figure 3: The xy-staggered cell Ki+ 1
,j+ 1 ,k
, used in the definition of ∂yM ux , ∂xM uy , and
2 2
τ M (u)x,y = τ M (u)y,x .
hal-00576268, version 1 - 13 Aug 2011
x y
Then the viscosities on σi− 1
,j+ 1
and σi− 1
,j+ 1
coincide only for uniform meshes, and, in the
2 2 2 2
general case, the symmetry of τ M (u) is lost.
– The partial derivatives of the velocity components are then defined as piecewise constant
functions, the value of which is obtained by natural finite differences:
– Then, define four families of values for the viscosity field, µ, µxy , µxz and µyz , associated
to the primal and the three edge-centred meshes respectively.
– The shear stress tensor is obtained by the extension of (11) to d = 3.
– And, finally, the dissipation term is given by (12).
6
Viscous dissipation term discretization with the MAC scheme
with u and f two scalar functions, f ∈ L2 (Ω). Let us suppose that this problem is discretized by
the usual finite volume technique, with the uniform MAC mesh associated to the x-component
of the velocity. We define a discrete function as a piecewise constant function, vanishing on the
left and right sides of the domain (so on the left and right stripes of staggered (half-)meshes
adjacent to these boundaries), and we define the discrete H1 -norm of a discrete function v by:
Z
||v||1 = (∂xM v)2 + (∂yM v)2 dx.
2
Ω
Let (M(n) )n∈N be a sequence of such meshes, with a step hn tending to zero, and let (u(n) )n∈N be
hal-00576268, version 1 - 13 Aug 2011
the corresponding sequence of discrete solutions. Then, with the variational technique employed
in the preceding section (i.e. multiplying each discrete equation by the corresponding unknown
and summing), we get, with the usual discretization of the right-hand side:
Z Z
(n) 2 M (n) 2 M (n) 2
||u ||1 = (∂x u ) + (∂y u ) dx = f u(n) dx. (14)
Ω Ω
Since the discrete H1 -norm controls the L2 -norm (i.e. a discrete Poincaré inequality holds [2]),
this yields a uniform bound for the sequence (u(n) )n∈N in discrete H1 -norm. Hence the sequence
(u(n) )n∈N converges in L2 (Ω) to a function ū ∈ H10 (Ω), possibly up to the extraction of a subse-
quence [2], and he discrete derivatives (∂xM u(n) )n∈N and (∂yM u(n) )n∈N weakly converge in L2 (Ω)
to ∂x ū and ∂y ū respectively. This allows to pass to the limit in the scheme, and we obtain
that ū satisfies the continuous equation (13) (so, since the solution to (13) is unique, the whole
sequence converges). Thus, taking ū as a test function in the variational form of (13):
Z Z
2 2
(∂x ū) + (∂y ū) dx = f ū dx.
Ω Ω
Since the discrete gradient weakly converges and its norm converges to the norm of the limit,
the discrete gradient strongly converges in L2 (Ω)2 to the gradient of the solution. Let us now
imagine that Equation (13) is coupled to a balance equation for another variable, the right-hand
side of which is |∇u|2 ; this situation occurs in several physical situations, as the modelling of
7
Viscous dissipation term discretization with the MAC scheme
Joule effect [1], or RANS turbulence models [10, 3]. The discretization (12) of the dissipation
term in the cell K, which reads here:
1
Z
|∇u(n) |2 K = (∂ M u(n) )2 + (∂yM u(n) )2 dx,
|K| K x
thus yields a convergent right-hand side, in the sense that, for any regular function ϕ ∈ C∞
c (Ω),
we have:
XZ Z
|∇u(n) |2 K ϕ dx = |∇u|2 ϕ dx.
lim
n→∞ K Ω
K
(A declination of) this argument has been used to prove the convergence of numerical schemes
in [1, 10, 3].
References
[1] A. Bradji, R. Herbin: Discretization of the coupled heat and electrical diffusion problems
by the finite element and the finite volume methods. IMA Journal of Numerical Analysis,
28, 469–495 (2008).
hal-00576268, version 1 - 13 Aug 2011
[4] F.H. Harlow, J.E. Welsh: Numerical calculation of time-dependent viscous incompressible
flow of fluid with free surface. Physics of Fluids, 8, 2182–2189 (1965).
[5] F.H. Harlow, A.A. Amsden: A numerical fluid dynamics calculation method for all flow
speeds. Journal of Computational Physics, 8, 197–213 (1971).
[7] R. Herbin, J.-C. Latché: A kinetic energy control in the MAC discretization of compressible
Navier-Stokes equations. International Journal of Finite Volumes 2 (2010).
[8] R. Herbin, W. Kheriji, J.-C. Latché: An unconditionally stable Finite Element-Finite Vol-
ume pressure correction scheme for compressible Navier-Stokes equations. In preparation
(2011).
[9] ISIS: a CFD computer code for the simulation of reactive turbulent flows,
https://gforge.irsn.fr/gf/project/isis.
[10] A. Larcher, J.-C. Latché: Convergence analysis of a finite element-finite volume scheme for
a RANS turbulence model. Submitted (2011).