Int.

Conference on Boundary and Interior Layers

BAIL 2006
G. Lube, G. Rapin (Eds)
c University of Gottingen, Germany, 2006
Mass conservation of nite element methods
for coupled ow-transport problems
Gunar Matthies
1
and Lutz Tobiska
2
1
Fakultat f ur Mathematik, Ruhr-Universitat Bochum,
Universitatsstrae 150, 44780 Bochum, Germany
2
Institut f ur Analysis und Numerik, Otto-von-Guericke-Universitat Magdeburg,
PSF 4120, 39016 Magdeburg, Germany
1. Introduction
Advection-diusion equations arise in a number of important applications. Their robust and
accurate numerical solution is in case of advection dominated ows still a challenge. Often
it is neglected that the physical processes are governed by a velocity eld u which itself is the
solution of a hydrodynamical model like the incompressible Navier-Stokes equations. In this
paper we address the issue of mass conservation when the underlying velocity eld u in the
transport equation is replaced by an approximation u
h
.
We consider the simplest case of a coupled ow-transport problem in a bounded domain
R
d
, d = 2, 3. The system is described by the instationary, incompressible NavierStokes
equations
u
t
u + (u )u + p = f in (0, T],
div u = 0 in (0, T],
u = u
b
on (0, T],
u(0) = u
0
in ,
(1)
and the time-dependent transport equation
c
t
c +u c = g in (0, T],
(cu c) n = c
I
u n on

(0, T],
c n = 0 on
+
(0, T],
c(0) = c
0
in .
(2)
Here, u and p denote the velocity and the pressure of the uid, respectively, and are small
positive numbers, T > 0 denes the nal time. The boundary is splitted into the inow
boundary

:= x : u n < 0 and the outow boundary

+
:=

where n is the
unit outer normal. Furthermore, c is the concentration and c
I
the concentration at the inow
boundary

. We assume that the given velocity eld u

b
on the boundary is the restriction
of a divergence free function which we denote again by u
b
. The initial velocity u
0
satises the
incompressibility constraint u
0
= 0.
Dierent discretization methods for both the instationary, incompressible NavierStokes
equations and the transport equation have been developed also in the practically important
case of 1 and 1, for an overview see [12]. We study the mass conservation of the
discretized transport equation when using stabilized schemes. For simplicity of notation we
restrict ourselves to the semidiscretization in time of the problems (1) and (2). The results can
be extended to the fully discretized problems using the discontinuous Galerkin method in time.
1
2. Algorithms for the transport equation
2.1. Weak formulation and mass conservation
Let W := H
1
(), (, ) and , )

denote the L
2
-inner products on and , respectively. A weak
formulation of the transport problem (2) is given by
Find c(t) W such that for all W : (c(0) c
0
, ) = 0 and
d
dt
(c, ) +(c, ) + (u c, ) c u n, )

= (g, ) c
I
u n, )

. (3)
Setting = 1 and using the incompressibility constraint u = 0, we get from (3) the global
mass conservation property
d
dt
_

c dx +
_

c
I
u nd +
_

+
c u nd =
_

g dx. (4)
Let the domain be polyhedral. We are given a family
_
T
h
_
h>0
of shape-regular triangulations
of into simplicial elements K. The diameter of K is denoted by h
K
while h := max
_
h
K
:
K T
h
_
. Let W
h
W be a nite element space for approximating the concentration. Then,
the standard Galerkin discretization of (3) reads
Find c
h
(t) W
h
such that for all
h
W
h
: (c
h
(0) c
0
,
h
) = 0 and
d
dt
(c
h
,
h
) +(c
h
,
h
) + (u
h
c
h
,
h
) c
h
u
h
n,
h
)

= (g,
h
) c
I
u
h
n,
h
)

(5)
where the solenoidal vector eld u has been replaced by some in general discontinuous
approximation u
h
. Now setting
h
= 1 and using elementwise integration by parts, we end up
with
d
dt
_

c
h
dx +
_

c
I
u
h
nd +
_

+
c
h
u
h
nd =
_

g dx +m
h
(c
h
, u
h
) (6)
where compared to the global mass conservation on the continuous level, see (4), the additional
term
m
h
(c
h
, u
h
) :=

KT
h
(c
h
, u
h
)
K
+

EE
h
c
h
, [u
h
n
E
]
E
)
E
(7)
is present. Here, c
h
denotes the set of inner faces E, (, )
K
and , )
E
denote the L
2
-inner
products on K and E, respectively. With each E c
h
we associate an arbitrary but xed unit
normal n
E
and dene the jump of across the common face E of the two adjacent elements K
and

K by
[]
E
:= u
h
[
e
K

E
u
h
[
K

E
where n
E
is an outer normal to K. Conditions which guarantee that m
h
= 0 will be discussed
in the following Sections.
2.2. Stabilized schemes for the transport equation
In the case that 0 < 1, the standard Galerkin discretization exhibits spurious oscillations
which can be suppressed by using some sort of stabilized schemes. We consider stabilized schemes
of the following type:
Find c
h
(t) W
h
such that for all
h
W
h
: (c
h
(0) c
0
,
h
) = 0 and
d
dt
(c
h
,
h
) +(c
h
,
h
) + (u
h
c
h
,
h
) c
h
u
h
n,
h
)

(8)
+S
h
(c
h
,
h
) = (g,
h
) c
I
u
h
n,
h
)

.
2
In the Streamline Diusion (SD) method [2] we add weighted residuals of the strong form of the
dierential equation. Thus, we have
S
sd
(c
h
,
h
) :=

KT
h

K
(c
h,t
c
h
+u
h
c
h
g, u
h

h
)
K
(9)
with a user-chosen parameter
K
.
In the subgrid scale method [6, 10], the space W
h
contains a subspace of resolvable scales
W
H
W
h
which is given by a projector P
H
: W
h
W
H
. Then, the non-resolvable scales are
stabilized by adding
S
subg
(c
h
,
h
) :=

KT
h

K
((id P
H
)c
h
, (id P
H
)
h
)
K
(10)
to the standard Galerkin approach (5).
Finally, we mention the stabilization by local projection [1] which relies on a local projection
operator

P : W
h
D
h
into a proper space of discontinuous nite elements. Here, the added
stabilizing term becomes
S
locp
(c
h
,
h
) :=

KT
h

K
_
(id

P)c
h
, (id

P)
h
_
K
. (11)
Note that in the subgrid scale method the unresolvable scales are stabilized by an articial
viscosity term whereas in the local projection method the uctuations of the gradients, i.e.,
(id

P)c
h
, are controlled in the L
2
-norm.
With respect to our purpose we see that in all considered cases the stabilizing terms vanish
for
h
= 1. Thus, the global mass conservation will be guaranteed provided that the additional
term m
h
(c
h
, u
h
) in (6) vanishes.
3. Algorithms for the incompressible NavierStokes equations
3.1. Weak formulation and Galerkin approach
Let V := H
1
0
()
d
, M := L
2
(), and Q :=
_
q M : (q, 1) = 0
_
. A weak formulation of the
NavierStokes problem (1) which is obtained in a standard way reads
Find
_
u(t), p(t)
_
(u
b
+V ) Q such that for all v V : (u(0) u
0
, v) = 0 and
d
dt
(u, v) +(u, v) +
_
(u )u, v
_
( v, p) = (f, v) v V, (12)
( u, q) = 0 q Q. (13)
Our assumption that u
b
is the restriction of a divergence free function yields
( u, 1) = u n, 1)

= u
b
n, 1)

= ( u
b
, 1) = 0. (14)
This combined with (13) implies that ( u, q) = 0 for all q L
2
().
We consider inf-sup stable discretizations of the problem (12), (13). Let V
h
V , M
h
M,
and Q
h
= M
h
Q be nite element spaces such that the inf-sup condition
inf
q
h
Q
h
sup
v
h
V
h
( v
h
, q
h
)
[v
h
[
1
|q
h
|
0
(15)
is satised with a positive constant which is independent of the mesh size parameter h. Using
the discrete spaces V
h
and M
h
, the standard Galerkin approach of (12), (13) reads
3
Find
_
u
h
(t), p
h
(t)
_
(u
b,h
+V
h
) M
h
such that for all v
h
V
h
: (u
h
(0) u
0
, v
h
) = 0 and
d
dt
(u
h
, v
h
) +(u
h
, v
h
) +
_
(u
h
)u
h
, v
h
_
( v
h
, p
h
) = (f, v
h
) v
h
V
h
, (16)
( u
h
, q
h
) = 0 q
h
Q
h
, (17)
where u
b,h
is a suitable approximation of u
b
which satises the condition u
b,h
n, 1)

= 0. As
a consequence
( u
h
, 1) =

KT
h
( u
h
, 1)
K
= u
b,h
n, 1)

EE
h
[u
h
]
E
n
E
, 1)
E
=

EE
h
[u
h
]
E
n
E
, 1)
E
.
Thus, if the normal components of u
h
are continuous over the faces we get the discrete analogon
of (14), i.e., ( u
h
, q
h
) = 0 for all q
h
M.
While discretizing the NavierStokes problem by inf-sup stable nite elements, one has to
make the fundamental decision of choosing either a continuous or discontinuous pressure approx-
imation. Due to (17), the incompressibility constraint u = 0 from (1) is fullled only in an
approximate sense. If discontinuous pressure approximations are used, the mass conservation
is satised more locally since functions with support within one element can be used as test
functions.
3.2. Stabilized schemes for the Navier-Stokes equations
In recent years, a huge number of schemes have been developed to stabilize both the eect of
dominating advection and the instabilities caused by using nite element pair V
h
and Q
h
which
do not satisfy (15), see e.g. [1, 11, 14]. In particular, the use of equal order interpolation of
velocity and pressure with the streamline diusion method or the stabilization by local projection
seems to be quite popular. However, a common feature of these types of stabilization methods
is an additional stabilizing term in the mass balance (13) of the Navier-Stokes equation which
produces an additional error for the mass conservation of the transport equation (6). For
avoiding this additional discretization error, one should try to separate the stabilization of the
two instability phenomena: dominating advection and use of unstable nite element pairs for
approximating velocity and pressure. Such a separation technique has been considered e.g. in
[3, 7, 8].
In the following we want to restrict ourselves to the solution of the Navier-Stokes equations
by inf-sup stable conforming nite element pairs. In this case, the computed velocity eld
u
h
(t) u
b,h
+V
h
belongs to H
1
()
d
and is discretely divergence free in the sense that
( u
h
, q
h
) = 0 q
h
M
h
.
Note that (17) implies that this relation holds rst for all q
h
Q
h
M
h
but the choice of the
approximation u
b,h
of the boundary data u
b
indeed guarantees its fulllness for all q
h
M
h
.
4. Mass conservative methods
We have seen in Section 2. that the mass conservation on the discrete level is guaranteed when
the term m
h
(c
h
, u
h
) vanishes. On the continuous level the term m
h
(c
h
, u) vanishes due to
incompressibility condition u = 0 and [u n
E
]
E
= 0 for all inner faces. In the following
subsections, we will discuss several possibilities which ensure m
h
(c
h
, u
h
) = 0.
4.1. Higher-order approximation of the ow problem
Let us assume that the transport equation will be solved by a method of order r 1, i.e., the
approximation error in space will satisfy
inf
W
h
[c
h
[
m
C h
r+1m
[c[
r+1
, c H
r+1
(), 0 m r + 1.
4
One could think for example to use continuous, piecewise polynomials of degree less than or
equal to r. In the following, we will shortly note this space by P
r
. We observe that m
h
(c
h
, u
h
)
vanishes if c
h
M
h
, in particular if W
h
M
h
. Thus, in the considered case we could choose
the space P
r
to be a subspace of pressure space M
h
. This means that the pressure will be
approximated in the L
2
-norm of order r + 1. In order to ensure the inf-sup condition the
velocity has to be approximated by suitable elements of order r + 1 like P
r+1
. We end up
with the following discretization for (u
h
, p
h
, c
h
) P
r+1
P
r
P
r
. Hence, the NavierStokes
problem has to be discretized by a method which is one order better than the method used for
the transport equation. Note that this technique works for both, continuous and discontinuous
pressure approximations.
A similar statement is also true for nonconforming nite element discretizations of the Navier-
Stokes equations since a nonconforming discretization for the velocities of order r +1 is assumed
to satisfy
[u
h
]
E
, r
h
)
E
= 0 r
h
P
r
(E), E c
h
.
If the discretization is completed by elements of order r for the approximation of pressure and
concentration, then we conclude from (7) that m
h
(c
h
, u
h
) = 0. Thus, again a one order higher
discretization of the ow problem results in a mass conservation of the transport problem.
However, from the practical point of view, this technique seems to be non-attractive since
the discretization of the NavierStokes problem by a higher order method is too costly.
4.2. Post-processing of the discrete velocity
Another idea for ensuring the exact mass balance on the discrete level consists in replacing
the discrete velocity solution u
h
by a dierent discrete function w
h
which is close to u
h
. This
technique was proposed in [4] for the local discontinuous Galerkin method applied to ow prob-
lems. To be precise, instead of solving the standard weak formulation of the transport equation,
see (5), the following problems will be solved
Find c
h
(t) W
h
such that for all
h
W
h
: (c
h
(0) c
0
,
h
) = 0 and
d
dt
(c
h
,
h
) +(c
h
,
h
) +(w
h
c
h
,
h
) c
h
w
h
n,
h
)

= (g,
h
) c
I
w
h
n,
h
)

(18)
Let the NavierStokes equations be discretized by the inf-sup stable nite element pair
P
bubble
r
/P
disc
r1
consisting of the velocity space of continuous, piecewise polynomials of degree r
enriched with certain bubble function and the pressure space of discontinuous, piecewise poly-
nomials of degree r 1, [9]. We will construct the function w
h
which acts as the velocity eld
in the transport equation by a post-processing. To this end, we dene on each element K T
h
the vector-valued local interpolation operator P
K
: H
1
(K)
d
P
r
(K)
d
by

(P
K
v) n
K
,
_
E
= v n
K
, )
E
E K, P
r
(E), (19)
(P
K
v, ) = (v, ) P
r1
(K), (20)
(P
K
v, ) = (v, )
r
(K), (21)
where

r
(K) :=
_
L
2
(K)
d
: (DF
T
K
) F
K

r
_
with

r
:=
_

P
r
(

K)
d
:

= 0 in

K,

n
b
K
= 0 on

K
_
.
In the above formulas we have used the reference transformation F
K
:

K K which is a
bijective mapping from the reference cell

K onto the original cell K. Further,

= F
K
. The
denition of P
K
by (19)-(21) is the customization of the interpolation operator P which was
5
introduced in [4] in the frame of local discontinuous Galerkin methods. The local interpolation
operators P
K
can be put together to a global interpolation operator P
h
in the following way
(P
h
v)[
K
:= P
K
(v
h
[
K
).
Note that, in general, the function P
h
v will not belong to H
1
()
d
.
Next we will show that the post-processed solution P
h
u
h
is piecewise divergence free. We
start with the incompressibility constraint (17) and use the conditions (19) and (20) of the
denition of P
K
to obtain
0 = ( u
h
, q
h
)
K
= (u
h
, q
h
)
K
+ u
h
n
K
, q
h
)
K
= (P
K
u
h
, q
h
)
K
+ (P
K
u
h
) n
K
, q
h
)
K
= ( P
K
u
h
, q
h
)
K
.
Here, we have used q
h
P
r1
(K)
d
and q
h
[
E
P
r
(E) for all faces E K.
Furthermore, we notice that the function which is piecewise dened by
[
K
:= P
K
u
h
belongs to Q
h
. Since
K
P
r1
(K) holds true, we have to show only that has zero integral
mean over . Indeed, we get
_

dx =

KT
h
_
K
P
K
u
h
dx =

KT
h
_
K
(P
K
u
h
) n
K
d =

KT
h
u
h
n
K
d =
_

u
h
nd = 0,
where we have used the condition (19).
Hence, can be used as a pressure test function in (17). Using this, we obtain
0 = ( P
h
u
h
, ) =

KT
h
( P
h
u
h
, P
h
u
h
)
K
which gives P
h
u
h
[
K
0, i.e., the post-processed velocity solution is piecewise divergence free.
The modied convection eld in (18) is chosen to be w
h
:= P
u
u
h
. This ensures that the rst
term of m
h
(c
h
, w
h
) vanishes. Moreover, the normal component of w
h
has no jumps across inner
faces due to condition (19) in the denition of P
K
. Indeed, we have for all P
r
(E) that

[P
h
u
h
]
E
n
E
,
_
E
= P
e
K
u
h
n
E
, )
E
P
K
u
h
n
E
, )
E
= u
h
n
E
, )
E
u
h
n
E
, )
E
= 0
where K and

K are the two elements which are adjacent to E. Since [P
h
u
h
]
E
P
r
(E) holds
true, we conclude [P
h
u
h
]
E
= 0. Hence, also the second term of m
h
(c
h
, w
h
) vanishes.
Compared to the method of Section 4.1., we solve both the transport and the Navier-Stokes
equation with a method of order r. Nevertheless, the mass conservation on the discrete level is
guaranteed.
4.3. Scott-Vogelius elements
Now, we will concentrate on nite element discretizations which guarantee that the discrete
velocity solution u
h
is piecewise divergence free. To this end, let us consider discretizations with
P
r
/P
disc
r1
-elements. It is well-known that in the two-dimensional case for r 4, this nite element
pair is inf-sup stable when special meshes which exhibit so called singular vertices, are excluded,
see [13]. Furthermore, it has been shown recently that there are macroelement meshes in two
and three dimensions on which this pair of nite elements is inf-sup stable [15] provided that the
polynomial degree r is greater than or equal to the space dimension d. Although a restriction on
the mesh is needed this pair is attractive since a discretely divergence free function is piecewise
divergence free. Indeed, due to the participating discrete spaces, the divergence of each discrete
6
velocity eld belongs to the pressure space. Hence, the weak compressibility condition (17)
yields
0 = ( u
h
, u
h
),
i.e., the discrete velocity solution u
h
is divergence free in the L
2
-sense. Of course, this results in
m
h
(c
h
, u
h
) = 0 since the continuous velocity approximation u
h
has no jumps.
References
[1] M. Braack and E. Burman, Local projection stabilization for the Oseen problem and its
interpretation as a variational multiscale method, SIAM J. Numer. Anal., 43, 25442566,
(2006).
[2] A. N. Brooks and T. J. R. Hughes, Streamline upwind/Petrov-Galerkin formulations for
convection dominated ows with particular emphasis on the incompressible Navier-Stokes
equations, Comput. Methods Appl. Mech. Engrg., 32, 199259, (1982).
[3] E. Burman and P. Hansbo, A stabilized non-conforming nite element method for incom-
pressible ow, Comput. Methods Appl. Mech. Engrg., 195, 28812899, (2006).
[4] B. Cockburn, G. Kanschat, and D. Schotzau, A locally conservative LDG method for the
incompressible Navier-Stokes equations, Math. Comp., 74, 10671095, (2005).
[5] C. Dawson, S. Sun, and M. F. Wheeler, Compatible algorithms for coupled ow and trans-
port, Comput. Methods Appl. Mech. Engrg., 193, 25652580, (2004).
[6] A. Ern and J.-L. Guermond, Theory and practice of nite elements, volume 159 of Applied
Mathematical Sciences, Springer-Verlag, New York, (2004).
[7] L. P. Franca, V. John, G. Matthies, and L. Tobiska, An infsup stable and residualfree
bubble element for the Oseen equations, Preprint 27/2006, Fakultat f ur Mathematik, Otto-
von-Guericke-Universitat Magdeburg, (2006).
[8] T. Gelhard, G. Lube, M. A. Olshanskii, and J.-H. Starcke, Stabilized nite element schemes
with LBBstable elements for incompressible ows, J. Comput. Appl. Math., 177, 243267,
(2005).
[9] V. Girault, P.A. Raviart, Finite Element Methods for NavierStokes Equations, Springer
Verlag, Berlin, 1986
[10] J.-L. Guermond, Stabilization of Galerkin approximations of transport equations by subgrid
modeling, M2AN Math. Model. Numer. Anal., 33, 12931316, (1999).
[11] P. Hansbo and A. Szepessy, A velocity-pressure streamline diusion method for the in-
compressible Navier-Stokes equations, Comput. Methods Appl. Mech. Engrg., 84, 175192,
(1990).
[12] H.-G. Roos, M. Stynes, and L. Tobiska, Numerical methods for singularly perturbed dier-
ential equations, Springer-Verlag Berlin, (1996).
[13] L. R. Scott and M. Vogelius, Norm estimates for a maximal right inverse of the divergence
operator in spaces of piecewise polynomials, RAIRO Model. Math. Anal. Numer., 19, 111
143 (1985).
[14] L. Tobiska and R. Verf urth, Analysis of a streamline dision nite element method for the
Stokes and Navier-Stokes equations, SIAM J. Numer. Anal., 33, 107127 (1996).
7
[15] Shangyou Zhang, A new family of stable mixed nite elements for the 3d Navier-Stokes
equations, Math. Comput., 74, 543554, 2004
8