You are on page 1of 16

Adaptive simulations of two-phase ow by discontinuous

Galerkin methods
W. Klieber, B. Rivie`re
*
Department of Mathematics, University of Pittsburgh, 301 Thackeray, Pittsburgh, PA 15260, USA
Received 4 October 2005; received in revised form 12 April 2006; accepted 3 May 2006
Abstract
In this paper we present and compare primal discontinuous Galerkin formulations of the two-phase ow equations. The wetting
phase pressure and saturation equations are decoupled and solved sequentially. Proposed adaptivity in space and time techniques yield
accurate and ecient solutions. Slope limiters valid on non-conforming meshes are also presented. Numerical examples of homogeneous
and heterogeneous media are considered.
2006 Elsevier B.V. All rights reserved.
Keywords: Error indicators; Discontinuous Galerkin; Adaptive time stepping; Slope limiters; Five-spot; NIPG; SIPG; IIPG; OBB
1. Introduction
Accurate simulations of multiphase processes are essen-
tial in problems related to the environment and the energy.
There is a need for discretization methods that perform
well on very general unstructured grids. Standard methods
such as the nite dierence methods, nite volumes and
expanded mixed nite element fail to capture the ow phe-
nomena in the case of highly heterogeneous media with
full permeability tensors. Recently, discontinuous Galerkin
(DG) methods have been applied to a variety of ow and
transport problem [21,22,2,25,24] and due to their exibil-
ity, they have been shown to be competitive to standard
methods. Furthermore, DG methods allow for unstruc-
tured meshes and full tensor coecients. Even though the
discontinuous nite element methods are more expensive
than the nite dierence methods, oil engineers are willing
to pay the price for accuracy and thus avoid costly mis-
takes [20].
In this work, the pressure-saturation formulation (also
known as the sequential formulation) of the two-phase ow
problem is discretized using several discontinuous Galerkin
methods. Description of the sequential model and other
formulations for two-phase ow can be found in [14,6].
The unknowns are the wetting phase pressure and satura-
tion and the equations are solved successively. One imme-
diate advantage is the fact that the diculty arising from
the non-linearity is removed by time-lagging the co-
ecients. The pressure equation is solved by the Oden
BaumannBabuska (OBB) method [18] whereas the
saturation equation is solved by either the OBB, the non-
symmetric interior penalty Galerkin method (NIPG) [23],
the symmetric interior penalty Galerkin method (SIPG)
[27,1] or the incomplete interior penalty Galerkin method
(IIPG) [9,26]. One can note that all four methods OBB,
NIPG, SIPG and IIPG are very similar to each other,
and can be described by the same variational formulation
with a bilinear form involving constant parameters. For
instance, OBB and NIPG only dier by the addition of a
penalty term; SIPG and NIPG only dier by a sign.
The objective of this work is to investigate adaptive
simulations in time and space on unstructured meshes.
We formulate error indicators for the spatial renement
and derenement techniques. We also present an algorithm
that allows the time step to vary during the simulation. One
0045-7825/$ - see front matter 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.cma.2006.05.007
*
Corresponding author. Tel.: +1 412 6248315; fax: +1 412 6248397.
E-mail address: riviere@math.pitt.edu (B. Rivie`re).
www.elsevier.com/locate/cma
Comput. Methods Appl. Mech. Engrg. 196 (2006) 404419
of the main diculties is the development of slope limiters
that would handle meshes with hanging nodes. We propose
a limiting technique based on the one introduced by
Durlofsky et al. [10] for conforming meshes. By the use
of adaptivity, we signicantly reduce the computational
cost while keeping the accuracy. To our knowledge, there
is little work in the literature on applications of DG meth-
ods to two-phase ow. In [19,4], simulations were per-
formed on uniformly rened meshes and with a constant
time step. In [3], DG is applied to a total pressure-satura-
tion formulation. In [17], DG and mixed nite elements
are coupled. More recently, in [13], a compressible air
water two-phase ow problem is numerically solved on
uniform meshes using the NIPG/OBB/SIPG method and
a local discontinuous Galerkin (LDG) [8] discretization
for the saturation equation. In this case, a Kircho trans-
formation is required to obtain a diusive ux from the
previous time step. The saturation equation is solved
explicitly in time, which is computationally appealing;
however this reduced cost is compensated by the introduc-
tion of an additional unknown, intrinsic to the LDG for-
mulation. Finally, in [11], fully coupled DG formulations
are considered and in this case slope limiters are not needed
even for high order of approximation. However, the solu-
tion of the fully coupled DG formulations require the con-
struction of a Jacobian matrix at each time step for the
NewtonRaphson method.
The plan of the paper is as follows. In the next section,
we present the equations describing the two-phase ow
problem. Section 3 contains the discrete schemes and nota-
tion. The adaptive strategy in space and time, as well as
the slope limiting technique, are described in Section 4.
Numerical examples are given in Section 5. Some conclu-
sions follow.
2. Model problem
The mathematical formulation of two-phase ow in
a porous medium X in R
2
consists of a coupled system
of non-linear partial dierential equations. The phases con-
sidered here are a wetting phase (such as water) and a
non-wetting phase (such as oil). For each phase, the conser-
vation of mass and a generalized Darcys law are obtained.
Under the assumption of incompressibility, a pressure-
saturation formulation is derived, for which the primary
variables are the pressure and the saturation of the wetting
phase denoted by p
w
and s
w
:
r k
t
Krp
w
r k
o
Krp
c
; 1
o/s
w

ot
r
k
o
k
w
k
t
Krp
c
_ _
r
k
w
k
t
u
t
_ _
: 2
The coecients in Eqs. (1) and (2) are dened below:
K is the permeability tensor and is spatially dependent;
for heterogeneous media, K is discontinuous.
The coecient / denotes the porosity of the medium.
k
t
= k
o
+ k
w
is the total mobility, that is, the sum of the
mobility of the non-wetting phase and the mobility of
the wetting phase. Mobilities are functions that depend
on the uid viscosities l
w
and l
o
and on the eective
wetting phase saturation s
e
. The eective saturation
depends on the residual wetting phase and non-wetting
phase saturations s
rw
and s
ro
as follows:
s
e

s
w
s
rw
1:0 s
rw
s
ro
:
The mobilities are then given by the BrooksCorey
model [5]:
k
w
s
w

1
l
w
s
4
e
; k
o
s
w

1
l
o
1 s
e

2
1 s
2
e
:
The dierence of the pressures of the two phases
p
c
= p
n
p
w
is the capillary pressure. From the
BrooksCorey model, it depends on the eective satura-
tion and a constant entry pressure p
d
:
p
c
s
w

p
d

s
e
p :
From this equation, we see that p
0
c
s
w
< 0 and we will
write: rp
c
jp
0
c
jrs
w
.
u
t
= u
o
+ u
w
is the total velocity, that is the sum of the
two phases velocities. Each phase velocity is given as
u
d
Kk
d
rp
d
; d o; w:
Let n denote the outward normal to oX. We associate to
Eqs. (1) and (2) several boundary conditions, by rst
decomposing the boundary of the porous medium oX
into disjoint parts:
oX C
p1
[ C
p2
C
s1
[ C
s2
; C
p1
\ C
p2
C
s1
\ C
s2
;:
The boundary conditions for (1) are of Dirichlet and
Neumann type:
p
w
p
dir
; on C
p1
; 3
Kk
t
rp
w
n 0; on C
p2
: 4
The boundary conditions for (2) are of Robin and
Neumann type:
s
w
u
t
K
k
o
k
w
k
t
p
0
c
rs
w
_ _
n s
in
u
t
n; on C
s1
; 5
K
k
o
k
w
k
t
p
0
c
rs
w
_ _
n 0; on C
s2
: 6
3. Scheme
In this section, we rst establish some notation for the
temporal and spatial discretization and we present our
numerical scheme. Let 0 = t
0
< t
1
< < t
N
= T be a sub-
division of the time interval (0, T). For any function v that
depends on time and space, we introduce the notation
v
i
= v(t
i
, ) for i = 0, . . . , N. We also dene the time step
Dt
i
= t
i+1
t
i
.
W. Klieber, B. Rivie` re / Comput. Methods Appl. Mech. Engrg. 196 (2006) 404419 405
The domain X is subdivided into triangular elements
that form a mesh. Because of the renements and derene-
ments, the mesh changes at every time step. Let us denote
by E
i
h
fEg
E
the mesh at time t
i+1
. Let h
i
be the maximum
diameter of the elements. Let C
h
i be the union of the open
sets that coincide with interior edges of elements of E
i
h
. Let
e denote a segment of C
h
i shared by two triangles E
k
and E
l
of E
i
h
(k > l); we associate with e, once and for all, a unit
normal vector n
e
directed from E
k
to E
l
and we dene for-
mally the jump and average of a function w on e by
w wj
E
k j
e
wj
E
l j
e
; fwg
1
2
wj
E
k j
e

1
2
wj
E
l j
e
:
If e is adjacent to oX, then the jump and the average of w
on e coincide with the trace of w on e and the normal vector
n
e
coincides with the outward normal n. The quantity jej
denotes the length of e.
For each integer r, we dene a nite element subspace of
discontinuous piecewise polynomials:
D
r
E
i
h
fv : vj
E
2 P
r
E 8E 2 E
i
h
g;
where P
r
(E) is a discrete space containing the set of poly-
nomials of total degree less than or equal to r on E. We will
approximate the wetting phase pressure and saturation by
discontinuous polynomials of order r
p
and r
s
, respectively.
We now derive the variational formulation for the two-
phase ow problem, by considering the pressure Eq. (1)
and the saturation Eq. (2) separately.
3.1. The pressure equation
We rewrite (1) by dening v Kk
o
rp
c
Kk
o
jp
0
c
jrs
w
:
r Kk
t
rp
w
r v: 7
Multiplying (7) by a test function v 2 D
rp
, and using
Greens formula on one element E yields:
b
1
b
b
E
2
2
3
E
1
E
3
E
4
E
0
Fig. 1. Slope limiting on non-conforming meshes.
Fig. 2. Renement of a triangular element.
Fig. 3. Five-well example: coarse mesh at initial time and adaptive meshes obtained at 15 and 45 days.
406 W. Klieber, B. Rivie` re / Comput. Methods Appl. Mech. Engrg. 196 (2006) 404419
_
E
Kk
t
rp
w
rv
_
oE
Kk
t
rp
w
n
E
v

_
E
v rv
_
oE
v n
E
v;
where n
E
is the outward normal to E. Summing over all the
elements in E
i
h
and using the fact that p
w
and v are smooth
enough, namely [p
w
] = 0, [Kk
t
$p
w
n
e
] = 0 and [v n
e
] = 0,
we have

E2E
i
h
_
E
Kk
t
rp
w
rv

e2C
i
h
[oX
_
e
fKk
t
rp
w
n
e
gv

e2C
i
h
_
e
fKk
t
rv n
e
gp
w

E2E
i
h
_
E
v rv

e2C
h
[oX
_
e
v n
e
v:
Making use of the boundary conditions (3) and (4), we
obtain

E2E
i
h
_
E
Kk
t
rp
w
rv

e2C
i
h
[C
p1
_
e
fKk
t
rp
w
n
e
gv

e2C
i
h
[C
p1
_
e
fKk
t
rv n
e
gp
w

E2E
i
h
_
E
v rv

e2C
i
h
[oX
_
e
v n
e
v

e2C
p1
_
e
Kk
t
rv np
dir
: 8
3.2. The saturation equation
Similarly, we dene the auxiliary vector f
k
w
kt
u
t
. Then,
(2) can be rewritten as
o/s
w

ot
r K
k
o
k
w
k
t
jp
0
c
j

rs
w
_ _
r f: 9
As for the pressure equation, we multiply by a test function
z 2 D
rs
over one element in E
i
h
, sum over all elements, and
Fig. 4. Five-well example: three-dimensional pressure contours at 15, 30, 45 and 52.5 days.
W. Klieber, B. Rivie` re / Comput. Methods Appl. Mech. Engrg. 196 (2006) 404419 407
use the regularity of s
w
and f. We nally obtain after some
algebraic manipulation:
_
X
o/s
w

ot
z

E2E
i
h
_
E
K
k
o
k
w
k
t
jp
0
c
jrs
w
rz

e2C
i
h
[oX
_
e
K
k
o
k
w
k
t
j p
0
c
jrs
w
n
e
_ _
z

e2C
i
h
_
e
K
k
o
k
w
k
t
jp
0
c
jrz n
e
_ _
s
w

E2E
i
h
_
E
f rz

e2C
i
h
[oX
_
e
f n
e
z:
Making use of the boundary conditions (5), (6) and the
continuity of pressure, we have:
_
X
o/s
w

ot
z

E2E
i
h
_
E
K
k
o
k
w
k
t
jp
0
c
jrs
w
rz

e2C
s1
_
e
s
w
u
t
n
e
z

e2C
i
h
_
e
K
k
o
k
w
k
t
jp
0
c
jrs
w
n
e
_ _
z

e2C
i
h
_
e
K
k
o
k
w
k
t
jp
0
c
jrz n
e
_ _
s
w

e2C
i
h
r
jej
_
e
s
w
z

E2E
i
h
_
E
f rz

e2C
i
h
[oX
_
e
f n
e
z

e2C
s1
_
e
s
in
u
t
n
e
z

e2C
i
h
_
e
k
w
k
t
fKk
t
rz n
e
gp
w

C
p1
_
e
Kk
w
rz n
e
p
w
p
dir
: 10
The equation above is parametrized by the coecients
2 {1, 0, 1} and r P0. For a positive penalty value r,
the choice = 1 yields the SIPG method, the choice
= 0 yields the IIPG method and the choice = 1 the
NIPG method. If r = 0 and = 1, we obtain the OBB
method.
3.3. The discrete scheme
We discretize the time derivative by nite dierence,
which yields the backward Euler scheme. The initial
approximations P
0
w
, S
0
w
are simply obtained by a L
2
projec-
tion of the initial data p
w
(t = 0) and s
w
(t = 0). Based on (8)
and (10), we formulate the following numerical method:
given P
i
w
; S
i
w
2 D
rp
D
rs
, nd P
i1
w
; S
i1
w
2 D
rp
D
rs
such that for all v; z 2 D
r
p
D
r
s
:
Fig. 5. Five-well example: three-dimensional saturation contours at 15, 30, 45 and 52.5 days.
408 W. Klieber, B. Rivie` re / Comput. Methods Appl. Mech. Engrg. 196 (2006) 404419

E2E
i
h
_
E
Kk
t
S
i
w
rP
i1
w
rv

e2C
i
h
[C
p1
_
e
fKk
t
S
i
w
rP
i1
w
n
e
gv

e2C
i
h
[C
p1
_
e
fKk
t
S
i
w
rv n
e
gP
i1
w

E2E
i
h
_
E
v
i
h
rv

e2C
i
h
[oX
_
e
v
i"
h
n
e
v

e2C
p1
_
e
Kk
t
S
i
w
rv np
dir
11
and
_
X
/
Dt
i
S
i1
w
z

E2E
i
h
_
E
K
k
o
S
i
w
k
w
S
i
w

k
t
S
i
w

jp
0
c
S
i
w
jrS
i1
w
rz

e2C
s1
_
e
S
i1
w
U
i
t
n
e
z

e2C
i
h
_
e
K
k
o
S
i
w
k
w
S
i
w

k
t
S
i
w

jp
0
c
S
i
w
jrS
i1
w
n
e
_ _
z

e2C
i
h
_
e
K
k
o
S
i
w
k
w
S
i
w

k
t
S
i
w

jp
0
c
S
i
w
jrz n
e
_ _
S
i1
w

e2C
i
h
r
jej
_
e
S
i
w
z

_
X
/
Dt
i
S
i
w
z

E2E
i
h
_
E
f
i
h
rz

e2C
i
h
[C
p1
_
e
f
i"
h
n
e
z

e2C
s1
_
e
s
in
U
i
t
n
e
Fig. 7. Two-dimensional pressure contours at 7.5, 15, 22.5 and 30 days.
INJECTION
PRODUCTION
Fig. 6. Domain and coarse mesh for quarter-ve spot.
W. Klieber, B. Rivie` re / Comput. Methods Appl. Mech. Engrg. 196 (2006) 404419 409

e2C
i
h
_
e
k
w
S
i
w

k
t
S
i
w

fKk
t
S
i
w
rz n
e
gP
i
w

e2C
p1
_
e
Kk
w
S
i
w
rz n
e
P
i
w
p
dir
; 12
where U
i
t
, f
i
h
and v
i
h
are the approximates of u
i
t
; f
i
and v
i
.
U
i
t
Kk
w
S
i
w
rP
i
w
Kk
o
S
i
w
p
0
c
S
i
w
rS
i
w
rP
i
w
;
v
i
h
Kk
o
S
i
w
jp
0
c
S
i
w
jrS
i
w
;
f
i
h

k
w
S
i
w

k
t
S
i
w

fU
i
t
g:
Because of the discontinuous approximations, there are
two values for the functions v
i
and f
i
on an interior edge.
These quantities are then replaced by the upwind numerical
uxes v
i"
h
and f
i"
h
. Upwinding is done with respect to the
normal component of the average of the total velocity U
t
8e oE
k
\ oE
l
; k > l; 8w; w
"

wj
E
k if fU
i
t
g n
e
P0;
wj
E
l if fU
i
t
g n
e
< 0:
_
From the derivations in Sections 3.1 and 3.2, we obtain the
consistency of the scheme (11) and (12).
Lemma 1. If (p
w
, s
w
) is a solution of (1), (2), then (p
w
, s
w
) is
also a solution of (11) and (12).
3.4. Local mass balance
Let us x an element E and a test function v 2 D
rp
that
vanishes outside of E. For simplicity, we assume that E is
an interior element in X. The pressure equation (11)
becomes:
_
E
Kk
t
S
i
w
rP
i1
w
rv
_
oE
fKk
t
S
i
w
rP
i1
w
n
E
gv

_
oE
1
2
Kk
t
S
i
w
rv n
E
P
i1
w

_
E
v
i
h
rv
_
oE
v
i"
h
n
E
v:
If in addition, we let v to be equal to one over E, we obtain
the local mass property satised by the approximations:

_
oE
fKk
t
S
i
w
rP
i1
w
n
E
gv
_
oE
v
i"
h
n
E
v 0:
3.5. Slope limiting
Approximations of high order yield overshoot and
undershoot in the neighborhood of the front of the injected
phase. Slope limiters are the appropriate tools for decreas-
ing the local oscillations [7,15]. To our knowledge there is
Fig. 8. Two-dimensional saturation contours at 7.5, 15, 22.5 and 30 days.
410 W. Klieber, B. Rivie` re / Comput. Methods Appl. Mech. Engrg. 196 (2006) 404419
no analysis available for slope limiters in 2D and 3D, even
on a conforming mesh. We are also not aware of limiters
that would handle non-conforming meshes. In this section,
we propose a limiting technique that can handle meshes
with hanging nodes. This procedure is successfully tested
for our two-phase ow problem. We apply the limiting
technique to the approximations P
i1
w
and S
i1
w
after each
time step t
i+1
.
In what follows, we say that one element E is active if it
belongs to the mesh E
i
h
, i.e., if it is used in the computation
of (11) and (12). The element can become inactive if it is
rened and thus its children are created and activated.
The limiting process consists of two steps.
First, we loop through all the active elements starting
from the oldest generation to the youngest (in general this
would mean that the order is in decreasing size). For exam-
ple, Fig. 1 shows an example of ve elements of dierent
generation: if G denotes the generation of the elements E
0
and E
1
, then elements E
3
and E
4
are of younger generation
G + 1 and element E
2
is of older generation G 1. Thus, it
is assumed that the limiting process has been already
applied to E
2
.
(1) Neighbor averages: We rst compute the average sat-
uration for the element to be limited and all neighboring
elements as follows. Let S
0
denote the average saturation
over E
0
and let S
j
denote a function associated to each side
j 2 {1, 2, 3} of E
0
. For E
0
and the neighbors of the same
generation, we have the usual averaging operator:
S
0
AE
0
; S
1
AE
1
; where AE
1
jEj
_
E
S
i1
w
:
To compute S
2
corresponding to the side 2 of E
0
and the
element E
2
that is of older generation, we rst locate the
barycenter b
2
of an imaginary child

C
2
of the same gener-
ation of E
0
(see dashed lines in Fig. 1). We then set
S
2
S
i1
w
j
E
2
b
2
:
We note that S
2
A

C
2
. The smaller elements E
3
and E
4
belong to a parent

E (see dotted lines in Fig. 1). If we de-
note by F
~
E
1
; . . . ; F
~
E
4
the children of

E, we can write
S
3

1
4

4
l1
BF
~
E
l
;
where the function B is dened recursively as (using the
notation F
E
l
for the lth child of E):
BE
1
jEj
_
E
S
i
w
; if E active;
1
4

4
l1
BF
E
l
; otherwise:
_

_
Fig. 9. Two-dimensional pressure contours at 7.5, 15, 22.5 and 30 days obtained on uniformly rened meshes.
W. Klieber, B. Rivie` re / Comput. Methods Appl. Mech. Engrg. 196 (2006) 404419 411
If the edge j is a boundary edge, then S
j
is dened accord-
ing to the boundary conditions.
S
j
s
in
on C
s1
; S
j
S
0
on C
s2
:
(2) Test: We then compute the saturation S
i1
w
j
E
0
m
j

evaluated at the midpoint m


j
of each edge j and we check
that this value is between S
j
and S
0
. We stop here if the test
is successful, otherwise we continue to step 3.
(3) Construction of three linears: Based on the technique
by Durlofsky et al. [10], we construct three linears using the
points b
j
and the averages S
j
. For instance, if we write the
linears as L
j
x; y a
j
0
a
j
1
x a
j
2
y, for j 2 {1, 2, 3}, they
are uniquely determined by
L
j
b
0
S
0
and L
j
b
l
S
l
; for l 6 j:
We then rank the linears by decreasing

a
j
1

2
a
j
2

2
_
and
check that for the values of the linears evaluated at the
midpoint m
l
, L
j
m
l
, is between S
l
and S
0
for 1 Pl P3.
If none of the constructed linears satisfy the test then the
slope is reduced to 0.
Second, we loop through all elements and check that
their slopes are not too large in the euclidean norm. If it
is larger than a cut-o value (set up by user), we scale it
by the ratio cut-o/norm.
4. Adaptivity strategy
In this section, we dene the error indicators and present
the adaptivity in space and time techniques. They are based
on the a posteriori error estimates obtained for a linear
convectiondiusion time-dependent problem [12], that
has some similarity with the saturation equation. However,
there is no rigorous mathematical proof for our coupled
system of equations and the error estimators of [12] are
used here as error indicators in the adaptivity algorithm.
4.1. Error indicators
We dene the following quantities:
R
vol

/
Dt
i
S
i1
w
S
i
w
r K
k
n
S
i
w
k
w
S
i
w

k
t
S
i
w

jp
0
c
S
i
w
jrS
i
w
_ _
r
k
w
S
i
w

k
t
S
i
w

U
i
t
_ _
;
R
e1
S
i1
w
;
R
e2
K
k
w
S
i
w
k
n
S
i
w

k
t
S
i
w

jp
0
c
S
i
w
jrS
i1
w
n
_ _
fU
i
t
g nS
i1
w
;
Fig. 10. Two-dimensional saturation contours at 7.5, 15, 22.5 and 30 days obtained on uniformly rened meshes.
412 W. Klieber, B. Rivie` re / Comput. Methods Appl. Mech. Engrg. 196 (2006) 404419
R
es1
s
in
U
i
t
nS
i1
w
U
i
t
nK
k
n
S
i
w
k
w
S
i
w

k
t
S
i
w

jp
0
c
S
i
w
jrS
i1
w
n;
R
es2
K
k
n
S
i
w
k
w
S
i
w

k
t
S
i
w

jp
0
c
S
i
w
jrS
i1
w
n:
Then the error indicator g
E
computed on each element E is
g
E
h
4
E
kR
vol
k
2
0;E

e2oEnoX
h
3
e
kR
e2
k
2
0;e
h
e
1kR
e1
k
2
0;e

e2oE\C
s1
h
3
e
kR
es1
k
2
0;e

e2oE\C
s2
h
3
e
kR
es2
k
2
0;e
_
1=2
;
X
S
w
0 100 200 300 400
0
0.2
0.4
0.6
0.8
1
X
P
w
0 100 200 300 400
2.4E+06
2.6E+06
2.8E+06
3E+06
3.2E+06
3.4E+06
Fig. 11. Saturation (left) and pressure (right) fronts along the diagonal line x = y at 7.5, 15, 22.5 and 30 days. The solid line corresponds to a uniform
mesh renement (h3) and the dashed line to an adaptively rened mesh.
X
S
w
0 100 200 300 400
0
0.2
0.4
0.6
0.8
1
X
P
w
0 100 200 300 400
2.4E+06
2.6E+06
2.8E+06
3E+06
3.2E+06
3.4E+06
Fig. 12. Saturation (left) and pressure (right) fronts along the diagonal line x = y at 15 and 30 days. Solid lines correspond to methods NIPG, SIPG, IIPG
with r = 10
5
and OBB. Dashed lines correspond to methods NIPG, SIPG and IIPG with r = 1.
Table 1
Number of degrees of freedom for adaptive and non-adaptive simulations
in the case (, r) = (1, 0)
t (days) DOFS press DOFS sat
AMR UNI AMR UNI
7.5 2538 25,344 1269 12,672
15 2412 25,344 1206 12,672
22.5 2142 25,344 1071 12,672
30 2466 25,344 1233 12,672
Table 2
Total number of degrees of freedom for adaptive simulations for all
methods
t (days) OBB NIPG NIPG SIPG SIPG IIPG IIPG
r = 0 r = 1 r = 10
5
r = 1 r = 10
5
r = 1 r = 10
5
7.5 3807 3348 3159 3294 3294 3348 4023
15 3618 3105 2970 3132 2997 3105 3591
22.5 3213 3348 3213 3483 3267 3375 3375
30 3699 2538 3780 2484 3834 2997 3915
W. Klieber, B. Rivie` re / Comput. Methods Appl. Mech. Engrg. 196 (2006) 404419 413
where h
E
is the diameter of the element E and h
e

maxh
E
k
; h
E
l
if the edge e is shared by elements E
k
and
E
l
. Note that the notation oEnoX means that the edges
are interior edges only.
4.2. Adaptive mesh renement technique
Let us assume that the solution P
i
w
and S
i
w
have been
obtained at the ith time step. We compute the error indica-
tor g
E
for each active element E. Then, we rst rene the
appropriate elements and second apply the derenement
technique.
Renement: We rene each element whose error indica-
tor is greater than a threshold value g
R
. Note that this
threshold value can be a percentage of the maximum of
the error indicators. Fig. 2 shows how one element (also
called parent) is rened into four smaller elements (also
called children).
Derenement: We consider a parent element for dere-
nement if (A) all of its children are active, (B) the error
indicator of each of its children is less than a threshold
value g
D
, and (C) the element was not rened during the
current time step. For each parent element meeting these
requirements, an L
2
projection is performed to retrieve
the degrees of freedom of the parent. Before actually doing
the derenement, we check that the parent error indicator
is less than g
R
. If it is, we then derene. If it is not, we
do not derene.
4.3. Adaptive time stepping technique
For time strategy, we allow the time step to vary during
the simulation. We uniformly divide the simulation interval
(0, T) into whole steps of length Dt
i
. At the start of each
whole step, we try to compute the saturation for time
t
i
+ Dt
i
, where t
i
is the current time. If the resulting satura-
tion function is satisfactory, then we record it, calculate the
new pressure function, and proceed to the next whole step.
On the other hand, if the resulting saturation function is
unsatisfactory, then we discard it and subdivide the whole
step into two half steps. We then compute the saturation
for the time at which the rst half step ends. If the result
is acceptable, we proceed to the second half step, and if
its result is also acceptable, then we continue on to the next
whole step. If one of the half steps does not yield satisfac-
tory results, then we divide it into quarter steps, proceeding
in the same manner as before, with the exception that we
accept the results of the quarter steps regardless of how
Fig. 13. Two-dimensional pressure contours at 7.5, 15, 22.5 and 30 days on an inhomogeneous medium.
414 W. Klieber, B. Rivie` re / Comput. Methods Appl. Mech. Engrg. 196 (2006) 404419
satisfactory they are. For the numerical simulations in this
paper, a resulting saturation function was deemed unsatis-
factory if the average saturation in any element exceeded
the physically permissible range by more than 0.01; other-
wise, it was considered satisfactory.
The main purpose of this time stepping technique is to
speed up computation without losing accuracy, thus to
increase the eciency of the method.
5. Numerical examples
In the following simulations, we assume that the uid
and medium properties are
l
o
0:002 kg=ms; l
w
0:0005 kg=ms; / 0:2;
s
w
t 0 0:2; p
w
t 0 3:45 10
6
Pa;
s
in
0:95; s
rw
0:15; s
ro
0;
p
d
5 10
3
Pa:
The orders of approximation are discontinuous piecewise
linears for the saturation and discontinuous piecewise qua-
dratics for the pressure. For the adaptive renements and
derenements, we chose g
R
= 1 10
1
and g
D
= (1/3)g
R
.
The well-known ve-spot problem on homogeneous and
heterogeneous media is rst considered, then simulations
with highly varying permeability are presented.
5.1. Five-spot on homogeneous medium
The permeability tensor is K = 10
11
dm
2
, where d is the
Kronecker delta tensor. Fig. 3 shows the coarse mesh and
Fig. 14. Two-dimensional saturation contours at 7.5, 15, 22.5 and 30 days on an inhomogeneous medium.
Fig. 15. Permeability eld and coarse mesh: permeability is 10
11
in white
regions and 10
16
elsewhere.
W. Klieber, B. Rivie` re / Comput. Methods Appl. Mech. Engrg. 196 (2006) 404419 415
the domain X embedded into the square (300, 300)
2
. Four
production wells are located at each corner of the domain;
the well bore corresponds to part of C
p1
, where we assume
that p
dir
= 2.41 10
6
Pa. An injection well is located in the
interior of the domain; the well bore corresponds to the
boundary C
s1
and the remainder of C
p1
and the pressure
is set p
dir
= 3.45 10
6
Pa. The ow of the phases is thus
driven by the gradient of pressure from the injection well
to the production wells.
The simulation is run for 52.5 days with a time step
varying between 0.001875 days and 0.0075 days. The
parameters in (12) are r = 0 and = 1. Three dimensional
views of contours of wetting phase pressure and saturation
at selected times are shown in Figs. 4 and 5. In order to bet-
ter analyze this example and because of the symmetry of
the problem, we re-run the simulations on one quarter of
the domain; this yields the quarter-ve spot problem shown
in Fig. 6. The injection well is at the left bottom corner
whereas the production well is at the right top corner.
The domain is now embedded into (0, 300)
2
.
The contours of wetting phase pressure and saturation
at selected times are shown in Figs. 7 and 8. The locally
rened and derened meshes are also given on these g-
ures. One can conclude that the proposed error indicators
capture well the location of the front. As expected, the
mesh is more rened in the neighborhood of the saturation
front. It also appears that the mesh stays rened at the
neighborhood of the injection well bore.
We compare the adaptive results with those obtained on
the coarse mesh rened uniformly three times. The pressure
and saturation contours are given in Figs. 9 and 10. Here,
the time step varies between 0.015 days and 0.00375 days.
The contours are similar to the adaptive ones. For better
comparison, we show the pressure and saturation proles
along the diagonal {(x, y): x = y} (see Fig. 11). Using adap-
tive renement and derenement decreases signicantly the
cost of the computation, as shown in Table 1. The columns
for AMR correspond to adaptively rened meshes, and the
columns for UNI correspond to uniformly rened meshes.
We now compare the saturation and pressure proles
obtained by varying the parameters r and . In particular,
we consider the cases (, r) 2 {(1, 1), (1, 10
5
)}, which yield
the NIPG method with small and large penalty values;
the cases (, r) 2 {(1, 1), (1, 10
5
)}, which yield the SIPG
method with small and large penalty values; the cases
(, r) 2 {(0, 1), (0, 10
5
)}, which yield the IIPG method with
small and large penalty values and the OBB method used
above. If the penalty value is small enough, i.e., r = 10
5
,
Fig. 16. Two-dimensional saturation contours at 17.5, 35, 52.5 and 70 days: ( = 1, r = 0).
416 W. Klieber, B. Rivie` re / Comput. Methods Appl. Mech. Engrg. 196 (2006) 404419
all methods produce identical solutions for both pressure
and saturation. Fig. 12 shows the proles obtained at 15
and 30 days: solid lines correspond to small penalty values
or the OBB method whereas dashed lines correspond to
large penalty values. As expected, if the penalty value
increases, numerical diusion added by the jump term
appears in the solutions. The saturation fronts are slightly
smeared. It is also interesting to note that for a xed pen-
alty value, all three penalty methods produce the same
solutions.
Finally, we give in Table 2 the total number of degrees of
freedom for pressure and saturation equations and we show
that the numerical cost is comparable for all methods.
5.2. Five spot on heterogeneous medium
This simulation is identical to the one above except for
the permeability tensor. Here, K is discontinuous and is
equal to 10
15
dm
2
in a small subdomain. In the rest of
domain, K = 10
11
dm
2
. We present the contours of the
pressure and saturation at dierent times in Figs. 13 and
14 in the case (, r) = (1, 0). Clearly, the region of low per-
meability is not invaded by the injected wetting phase. This
shows that the scheme has very little numerical diusion. It
is also interesting to note that the proposed method allows
for an arbitrary number of hanging nodes, without any
special care.
5.3. Highly varying permeability eld
We consider a square domain (0, 400)
2
with varying
permeability as shown in Fig. 15. The permeability is
10
11
Im
2
except in several small regions where it is 10
5
times smaller (see [16]). The simulation is run for 70 days.
The time step varies between 2.1875 10
3
days and
8.75 10
3
days. The vertical boundaries correspond to
C
p1
where the same pressure p
dir
as in the previous exam-
ples is imposed. The left vertical boundary corresponds
to C
s1
. We rst consider the OBB method ( = 1, r = 0).
Saturation contours on adaptively rened meshes are
shown in Fig. 16. The degrees of freedom are 8232,
13,281, 17,070 and 19,131 for the respective times 17.5,
35, 52.5 and 70 days. The gures show clearly that there
is very little numerical diusion. For comparison, we show
the contours obtained on a uniform mesh, which corre-
sponds to 38,400 degrees of freedom (see Fig. 17). The
coarse mesh has been rened twice, and this produces a
computational time of 24 h on a single processor. For
another level of renement, the simulation would run for
1 week.
Fig. 17. Two-dimensional saturation contours at 17.5, 35, 52.5 and 70 days on uniform meshes: ( = 1, r = 0).
W. Klieber, B. Rivie` re / Comput. Methods Appl. Mech. Engrg. 196 (2006) 404419 417
We repeat the experiment for the NIPG method with
three choices of penalty r 2 {10
6
, 10
5
, 1}. The saturation
contours at 35 days are shown in Fig. 18. Because of the
highly varying permeability eld, the method is more sensi-
tive to the choice of the penalty. The value r = 10
6
yields
a comparable solution to the OBB method; there is very lit-
tle numerical diusion. However, for r = 10
5
, the wetting
phase oods the region of lower permeability and for
r = 1, the method is too diusive to capture the barrier
zones. Similar conclusions can be made with the other
two methods. We show the saturation contours for
r = 10
6
for IIPG and SIPG in Fig. 19.
Fig. 18. NIPG Two-dimensional saturation contours at 35 days: r = 10
6
, r = 10
5
and r = 1.
Fig. 19. Two-dimensional saturation contours at 35 days for r = 10
6
: IIPG (left) and SIPG (right).
418 W. Klieber, B. Rivie` re / Comput. Methods Appl. Mech. Engrg. 196 (2006) 404419
6. Conclusions
This paper present adaptivity techniques in space and
time. We show that the adaptive simulations are more e-
cient than the simulations obtained on uniform meshes and
with constant time step. A slope limiting technique for
meshes with several hanging nodes per face is dened.
Numerical experiments show robustness of the proposed
DG schemes on heterogeneous media. Comparisons
between SIPG, NIPG, IIPG and OBB methods are per-
formed: if the penalty value is small enough, the resulting
numerical solutions are very similar. Increasing the penalty
value introduces numerical diusion in the approxima-
tions, in particular if the permeability eld highly varies
in space. Several future extensions are currently under
investigation: for instance, we plan to extend our computa-
tional results to three-dimensional problems using unstruc-
tured tetrahedral meshes. We are also investigating the
eects of p-adaptivity on the accuracy and the computa-
tional eciency of the scheme.
Acknowledgement
The rst author is partially supported by a CRDF grant
from the University of Pittsburgh and by a Brackenridge
fellowship. The second author is supported by a National
Science Foundation grant DMS-0506039.
References
[1] D.N. Arnold, An interior penalty nite element method with
discontinuous elements, SIAM J. Numer. Anal. 19 (1982) 742760.
[2] P. Bastian, High order discontinuous Galerkin methods for ow and
transport in porous media, Challenges in Scientic Computing CISC
2002, vol. 35, 2003.
[3] P. Bastian, Discontinuous Galerkin methods for two-phase ow in
porous media, Also technical report, 2004-28, IWR (SFB 359),
Heidelberg University, submitted for publication.
[4] B. Rivie`re, Numerical study of a discontinuous Galerkin method for
incompressible two-phase ow, in: Proceedings of ECCOMAS, 2004.
[5] R.H. Brooks, A.T. Corey, Hydraulic properties of porous media,
Hydrol. Pap. 3 (1964).
[6] G. Chavent, J. Jare, Mathematical Models and Finite Elements for
Reservoir Simulation, North-Holland, 1986.
[7] B. Cockburn, C.-W. Shu, TVB RungeKutta local projection
discontinuous Galerkin nite element method for conservative laws
II: general framework, Math. Comput. 52 (1989) 411435.
[8] B. Cockburn, C.-W. Shu, The local discontinuous Galerkin method
for time-dependent convectiondiusion systems, SIAM J. Numer.
Anal. 35 (1998) 24402463.
[9] C. Dawson, S. Sun, M.F. Wheeler, Compatible algorithms for
coupled ow and transport, Comput. Meth. Appl. Mech. Engrg. 193
(2004) 25652580.
[10] L.J. Durlofsky, B. Engquist, S. Osher, Triangle based adaptive
stencils for the solution of hyperbolic conservation laws, J. Comput.
Phys. 98 (1992) 6473.
[11] Y. Epshteyn, B. Rivie`re, Fully implicit discontinuous Galerkin
schemes for multiphase ow, Appl. Numer. Math., in press.
[12] A. Ern, J. Proft, A posteriori discontinuous Galerkin error estimates
for transient convectiondiusion equations, Appl. Math. Lett. 18 (7)
(2005) 833841.
[13] O. Eslinger, Discontinuous Galerkin nite element methods applied
to two-phase, air-water ow problems, Ph.D. thesis, The University
of Texas at Austin, 2005.
[14] R. Helmig, Multiphase Flow and Transport Processes in the
Subsurface, Springer, 1997.
[15] H. Hoteit, Ph. Ackerer, R. Mose, J. Erhel, B. Philippe, New two-
dimensional slope limiters for discontinuous Galerkin methods on
arbitrary meshes, Int. J. Numer. Meth. Engrg. 61 (2004) 25662593.
[16] L.J. Durlofsky, Accuracy of mixed and control volume nite element
approximations to Darcy velocity and related quantities, Water
Resour. Res. 30 (4) (1994) 965973.
[17] D. Nayagum, G. Schafer, R. Mose, Modelling two-phase incom-
pressible ow in porous media using mixed hybrid and discontinuous
nite elements, Comput. Geosci. 8 (1) (2004) 4973.
[18] J.T. Oden, I. Babuska, C.E. Baumann, A discontinuous hp nite
element method for diusion problems, J. Comput. Phys. 146 (1998)
491519.
[19] B. Rivie`re, The dgimpes model in ipars: Discontinuous Galerkin for
two-phase ow integrated in a reservoir simulator framework,
Technical Report 02-29, Texas Institute for Computational and
Applied Mathematics, 2002.
[20] B. Rivie`re, E. Jenkins, In pursuit of better models and simulations, oil
industry looks to the math sciences, SIAM News (January/February)
(2002).
[21] B. Rivie`re, M.F. Wheeler, Discontinuous Galerkin methods for ow
and transport problems in porous media, Commun. Numer. Meth.
Engrg. 18 (2002) 6368.
[22] B. Rivie`re, M.F. Wheeler, Non conforming methods for transport
with nonlinear reaction, in: Proceedings of an AMS-IMS-SIAM Joint
Summer Research Conference on Fluid Flow and Transport in
Porous Media: Mathematical and Numerical Treatment, 2002.
[23] B. Rivie`re, M.F. Wheeler, V. Girault, Improved energy estimates for
interior penalty, constrained and discontinuous Galerkin methods for
elliptic problems. Part I, Comput. Geosci. 3 (1999) 337360.
[24] S. Sun, B. Rivie`re, M.F. Wheeler, A combined mixed nite element
and discontinuous Galerkin method for miscible displacement prob-
lems porous media, in: Proceedings of International Symposium on
Computational and Applied PDEs, 2002, pp. 321348.
[25] S. Sun, M.F. Wheeler, Symmetric and non-symmetric discontinuous
Galerkin methods for reactive transport in porous media, SIAM J.
Numer. Anal. 43 (1) (2005) 195219.
[26] S. Sun, M.F. Wheeler, Symmetric and nonsymmetric discontinuous
Galerkin methods for reactive transport in porous media, SIAM J.
Numer. Anal. 43 (1) (2005) 195219.
[27] M.F. Wheeler, An elliptic collocation-nite element method with
interior penalties, SIAM J. Numer. Anal. 15 (1) (1978) 152161.
W. Klieber, B. Rivie` re / Comput. Methods Appl. Mech. Engrg. 196 (2006) 404419 419