Advances in Water Resources 24 (2001) 63±72

www.elsevier.com/locate/advwatres

Numerical modelling of convection dominated transport coupled

with density driven ¯ow in porous media
Peter Frolkovic, Hennie De Schepper *
Department of Mathematical Analysis, University of Gent, Galglaan 2, 9000 Gent, Belgium
Received 12 July 1999; received in revised form 7 February 2000; accepted 12 April 2000

Abstract
In this paper, we present a numerical model for a problem of coupled ¯ow and transport in porous media. We use a barycentre
based ®nite volume method (FVM), which, in the case of convection dominated transport, is combined with suitable upwind
methods, in order to avoid numerical instabilities. We present some relevant and new numerical results for the Elder problem, which
o€er a better understanding of mutually non-compatible results in other papers, by showing the dependence of the recirculating
patterns on the level of grid re®nement and on the numerical scheme, as well as on (numerical) perturbations. Ó 2000 Elsevier
Science Ltd. All rights reserved.

Keywords: Density driven ¯ow; Convection dominated transport; Finite volume methods; Upwind methods; Discrete maximum
principle; Elder problem

1. Introduction schemes. Hence, for more involved situations, numerical

methods which preserve known analytical or physical
When studying ¯ow and transport phenomena in properties are preferable in view of the reliability of the
porous media, one may observe that the transported computed solutions. In this paper, we design discret-
component in¯uences the ¯ow patterns, leading to a isation schemes which preserve a local mass conserva-
fully coupled problem of transport and ¯ow. Typically, tion property, and which moreover lead to physically
the varying concentration of the transported species acceptable values for the unknown concentration of the
directly a€ects the ¯uid density, which can result in ¯ow transported species.
pattern formation, due to the gravity forces. Such phe- The paper is organized as follows.
nomena are usually indicated as density driven ¯ow, In Section 2, we brie¯y recall the mathematical
[13], or buoyancy induced ¯ow [11]. model.
A motivation for the mathematical problem consid- In Section 3, we introduce so-called vertex centered
ered may be the modelling of groundwater ¯ow near salt ®nite volume methods (FVMs), which preserve a local
domes, where the ¯uid is a mixture of pure water and mass balance. They are related to ®nite element methods
brine, see [8,12,13,23] or the modelling of an evaporat- (FEMs), as the ®nite volume mesh is constructed com-
ing salt lake, where dense brine overlies a less dense plementary to a standard, well-chosen ®nite element
¯uid, [26]. However, our paper rather has an academical mesh ± triangular or quadrangular. Here, we restrict
purpose. ourselves to a 2D-approach, corresponding e.g., to a
The mathematical model of nonlinear coupled ¯ow vertical cross-section of the physical domain. These
and transport is quite involved. Analytical solutions are FVMs were presented ®rst in various applications of
known for simpli®ed situations only, see [29]. The same ¯ow and transport, see e.g. [24]. More recently, they
holds for the error analysis of well-established numerical have also been studied from the theoretical point of
view, see [2,10,15,28].
* In Section 4, we prove a discrete minimum and
Corresponding author. Tel.: +32-9-264-48-97; fax: +32-9-264-49-87;
http://cage.rug.ac.be/hds/nufarg/.
maximum property for the numerical solution of the
E-mail addresses: pfrolko@cage.rug.ac.be (P. FrolkovicÏ), hds@ca- transport equation. This allows us to predict non-
ge.rug.ac.be (H. De Schepper). physical behaviour of the numerical solution (e.g.,

0309-1708/01/$ - see front matter Ó 2000 Elsevier Science Ltd. All rights reserved.
PII: S 0 3 0 9 - 1 7 0 8 ( 0 0 ) 0 0 0 2 5 - 7
64 P. Frolkovic, H. De Schepper / Advances in Water Resources 24 (2001) 63±72

under- and overshooting, non-physical oscillations,. . .). given (injected) mass fraction of externally supplied
Such behaviour may occur either due to the approxi- brine …Q > 0†, or C  c for extraction wells …Q < 0†, and
mation of complex (anisotropic, variable) di€usion± D ˆ D…x; v† is the di€usion±dispersion matrix, see e.g.
dispersion ¯uxes, or due to the convection dominated [4]. Finally, v is the mass averaged velocity, given by
character of the transport equation. Darcy's law, see e.g. [4],
In Section 5, we consider the case, where the mini-
k
mum and maximum principle mentioned above fails, vˆÿ …rp ÿ q…c†g†; …3†
due to convection dominated transport. In this case we l…c†
propose a class of upwind methods, retaining the idea of where p ˆ p…t; x† stands for the unknown ¯uid pressure.
a discrete local mass balance. Several advanced types of Through (3), Eqs. (1) and (2) form a (nonlinear and fully
upwind methods were presented already in [24] and have coupled) system for two unknown functions c and p.
also been studied theoretically, e.g., in [2,14]. Further data, entering v, are the ¯uid viscosity l ˆ
Finally, in Section 6, we present numerical results for l…c†, the permeability matrix k ˆ k…x† and the gravity
the Elder experiment, [6]. Moreover, for this bench- vector g.
mark, we show the advantages of some advanced up- Suitable initial and boundary conditions (ICs and
wind schemes. Although the Elder problem originates BCs), see [4,8,12,20,21] must accompany (1) and (2). For
from thermodynamics, it has become a popular bench- the sake of simplicity we will only consider standard
mark in mass transfer with variable density ¯ow (in Dirichlet and (homogeneous) Neumann BCs. Formally,
particular brine transport, see [1,3,18,19,21,23,30]), due an IC needs to be supplied for c only. However, from the
to its complex dynamical behaviour. This example has numerical point of view, a reasonable initial pressure
been chosen amongst others, as the computations o€er distribution can be useful, see e.g. [8,21].
clear insight into the e€ectiveness of the numerical
schemes.
Compared to existing results in the literature, see
[2,9,14,24], the main contribution of this paper concerns 3. Finite volume methods
the following aspects. We extend the results on the dis-
crete minimum and maximum principle to the general In this section, we describe the vertex centered ®nite
case of nonlinear transient coupled ¯ow and transport, volume discretisation of the system (1)±(3). To ®x the
using FVMs based on both triangular and quadrangular ideas, let X be polygonal. We consider a triangulation T
meshes. Moreover, we consider the more complex case of X, in the sense of [5], consisting of triangular and/or
of anisotropic di€usive±dispersive transport, and we quadrangular elements T e ; e ˆ 1; . . . ; E which is used to
derive a general class of upwind methods in this context. de®ne a FE-mesh. The corresponding set of nodes, de-
Furthermore, our results for the Elder problem are new. noted by xi ; i ˆ 1; . . . ; N , consists of the vertices of all
They allow for a better judgment of some mutually non- elements T e in T. Further, we consider time points
compatible numerical results and conclusions in di€er- 0 ˆ t0 < t1 <


and we denote sm ˆ tm ÿ tmÿ1 ; m 2 N0 .
ent papers, see [6,19,21,23,30]. In general, the data functions in Eqs. (1)±(3) will
show a number of discontinuities in X. To incorporate
this in our numerical model, we assume that T is con-
structed such that the data are continuous inside each
2. Mathematical model
element.
At any time point tm , we consider the functions
The mathematical model below describes a miscible
displacement of salt water (brine) and fresh water, X
N

constituting a ``binary ¯uid mixture'', e.g. [4,12,13, c~  c~m …x† ˆ cmi Ni …x†;
iˆ1
21,23,30]. The unknown mass fraction of brine is de-
noted by c ˆ c…t; x†, where x 2 X  R2 and t P 0. The X
N
p~  p~m …x† ˆ pim Ni …x†; x 2 X; …4†
variability of the ¯uid density is expressed through a iˆ1
(nonlinear) equation of state, q ˆ q…c† P qw ; qw de-
noting the density of fresh water. The dependence of q where Ni ; i ˆ 1; . . . ; N , are the usual, globally continuous
on the ¯uid temperature or on the compressibility of the ®nite element basis functions, de®ned piecewise and
¯uid is neglected. The ¯ow and transport equations are obeying Ni …xj † ˆ dij ; j ˆ 1; . . . ; N . The function rNi is
N
well-de®ned per element. In (4), fcmi ; pim giˆ1 are unknown
/ ot q…c† ‡ r
…q…c†v† ˆ q…C†Q; …1†
discrete variables, approximating the nodal values
c…tm ; xi † and p…tm ; xi †, respectively.
/ ot …q…c†c† ‡ r
q…c†…cv ÿ Drc† ˆ q…C†CQ: …2† N
We set up an algebraic system for fcmi ; pim giˆ1 , which
Here, / ˆ /…x† is the porosity of the soil, Q ˆ Q…t; x† can be solved sequentially for m ˆ 1; 2; . . ., the values c0i
describes external sources and sinks, C either denotes a (and preferably also pi0 ) being determined by ICs. To
 P. Frolkovic, H. De Schepper / Advances in Water Resources 24 (2001) 63±72 65

this end, we ®rst introduce a local mass balance in in- tion of the integrand at a well-chosen point xeij 
tegral form, starting from (2), i.e., xeji 2 Ceij , taken to be xij for the convective part. Other
Z Z quadrature rules could also be considered, see [14] for
/q…c…t††c…t† dx ÿ /q…c…^t††c…^t† dx e.g., the Crank±Nicholson scheme or the consistent
V
Z tZ
V mass scheme.
In order to write the resulting algebraic system in a
ˆÿ q…c†n
…cv ÿ Drc† dc dt
^t oV transparent form, some further notations must be in-
Z tZ troduced. Eventual discontinuities of the space depen-
‡ q…C†CQ dx dt …5† dent coecient functions between two elements are
^t V
taken into account by the notation fie or fije for the ex-
and similarly for (1). Here, V  X and …^t; t† is an arbi- tension of a function f on T e to xi 2 oT e or to xij 2 oT e ,
trary time interval. respectively. Moreover for time dependent data we omit
Next, to derive vertex centered FVMs, we introduce a the upper index m (e.g., ci  cmi ) and we write
dual mesh, by constructing the ®nite volume Vi 6ˆ ; for cij  c~m …xij †, ceij  c~m …xeij †, rceij  r~ cm …xeij † and anal-
each vertex xi . Here, we deal with barycenter based ®nite ogously for p. Finally, by jVi e j and by jCeij j we, re-
volumes, but other types of vertex centered ®nite vol- spectively, denote the area of Vi e and the length of Ceij .
umes may be considered as well, see [2,9,14,16]. We We arrive at the system
de®ne the ®nite volume Vi as the polygonal, bounded by X XX
jVi e j/ei q…ci † ‡ sm jCeij jq…cij †neij
veij
Ci (when xi 62 oX), or by Ci [ C0i (when xi 2 oX), with e2Ki e2Ki j2Kei
Ci ˆ [e2Ki [j2Kei Ceij and C0i ˆ [j2K0i C0ij . Here, Ki is the set X ÿ ÿ

of all indices e, such that xi 2 oT e ; when e 2 Ki ; Kei is the ˆ jVi e j /ei q cimÿ1
e2Ki
set of all indices j, such that xi and xj are neighbouring

nodes on oT e , and ®nally, for xi 2 oX, K0i is the set of all ‡ sm q…Cie †Qei ; i ˆ 1; . . . ; I …6†
indices j such that xi and xj are neighbouring nodes on
and
oX. Further, Ceij is the line segment, connecting the X XX
barycenter xe of the element T e with the midpoint xij of jVi e j/ei q…ci †ci ‡ sm jCeij jneij
xi xij , while C0ij stands for the line segment xi xij , when xi e2Ki e2Ki j2Kei
  X ÿ ÿ
mÿ1
and xj 2 oX. An illustration of the notations is given by

q…cij †cij veij ÿ q…ceij †Deij rceij ˆ jVi e j /ei q cmÿ1
i ci
Fig. 1. e2Ki
We consider (5) and its analogon for the ¯ow equa-

‡ sm q…Cie †Cie Qei ; i ˆ 1; . . . ; I: …7†
tion for V ˆ Vi and for the generic time interval
…tmÿ1 ; tm †. To simplify the treatment of the BCs, we as- Here, neij is the unit normal vector on Ceij , pointing from
N
sume that fcmi ; pim giˆI‡1 are prescribed by Dirichlet BCs, Vi to Vj . If no special approximation of v is used (see
which allows us to consider only i ˆ 1; . . . ; I. Moreover, [7,17,30] for the consistent velocity approximation), we
at the remaining part of oX we impose zero ¯ux BCs, simply put
causing eventual integrals over C0i to drop out. For a kije  e 
more involved treatment of BCs, we refer to [8,9]. veij ˆ ÿ rpij ÿ q…cij †g : …8†
Furthermore, we approximate c and p by (4) and we l…cij †
apply suitable quadrature rules, deliberately chosen as Notice that equations can be assembled elementwise, as
follows: for time integrals we use an implicit Euler in FEMs, the resulting algebraic system being nonlinear,
scheme, by evaluation of the integrand at t ˆ tm ; for coupled and sparse.
integrals over Vi e  Vi \ T e , we resort to a mass lumping
scheme, by evaluation of the integrand at x ˆ xi ; ®nally,
Remark 1. Eqs. (6) and (7) can be interpreted as discrete
line integrals over Ceij will be approximated by evalua-
local mass conservation properties in the ®nite volume
Vi , during the time interval …tmÿ1 ; tm †. These mass bal-
ances also hold in the union of two adjacent ®nite vol-
umes, leading, by continuation, to discrete conservation
laws in the whole domain. When adapting the grid be-
tween subsequent time points, see [18], we should retain
these mass balances. In particular,
X
N X X
N X
q…ci † jVi e j/ei and q…ci †ci jVi e j/ei ;
iˆ1 e2Ki iˆ1 e2Ki

should be preserved, as they represent in a discrete sense

Fig. 1. Notations for ®nite volume meshes. the total mass of ¯uid, respectively, of salt in X at t ˆ tm .
66 P. Frolkovic, H. De Schepper / Advances in Water Resources 24 (2001) 63±72

We also give an equivalent form of (7), which will and below ± by extreme values of some given data in the
help us to study the discrete minimum and maximum mathematical model. Such a minimum and maximum
principle and to introduce upwind methods. Invoking principle has been proved for the analytical solution of a
(6), (7) may be rewritten in the form linear transport equation, where q and l are not de-
X ÿ
ÿ
XX pendent on c, see e.g. [27]. Moreover for an appropriate
jVi e j/ei q cmÿ1 ci ÿ cmÿ1 ‡ sm jCeij jneij
i i numerical discretisation scheme, also a discrete mini-
e2Ki e2Ki j2Kei
    mum and maximum principle holds, under some as-

q…cij † c^eij ÿ ci veij ÿ q…ceij †Deij rceij sumptions, see [2,14]. Here, we formulate similar results
X ÿ
for nonlinear, coupled ¯ow and transport. For trian-
‡ sm jVi e jq…Cie † ci ÿ Cie Qei ˆ 0; i ˆ 1; . . . ; I: …9† gular elements and isotropic di€usion without disper-
e2Ki
sion this was studied in [9]. In this paper, we allow for
Above, we formally introduced the notation the more general ®nite elements considered above, as
well as for a general di€usion±dispersion matrix.
c^eij ˆ ci ‡ b^eij …cj ÿ ci †; j 2 Kei ; e 2 Ki ; …10† Moreover, we derive upwind schemes in this general
where b^eji
ˆ b^eij
ˆ 1=2, implying that c^eij  cij . Hence, (7) context: cf. Section 5.
and (9) are equivalent. In Sections 5 and 6, this will be In the following two remarks we ®rst discuss some
referred to as the no upwind scheme. features of keik , corresponding to particular choices of the
Let k 2 Ke when xk 2 T e . Introducing the scalars quadrature nodes xeij , which appear in (11), for the ap-
proximation of the di€usive±dispersive ¯uxes.
1 X e
keik ˆ jC jq…ceij †neij
Deij rNk …xeij †; k 2 Ke n fig;
q…cik † j2Ke ij Remark 3. As for the approximation of the convective
i

…11† term, we can choose xeij ˆ xij . Here, in general, keik 6ˆ keki
and, moreover, we cannot judge the sign of keik . How-
ceij ˆ jCeij jneij
veij ; j 2 Kei ; …12† ever, for the isotropic case (i.e., D  /dI), with a grid
consisting either of equilateral triangles or of rectangles,
which characterise the di€usive±dispersive and the con- we obtain
vective part of the numerical ¯uxes, respectively, (9) can
be rewritten as j Ceik j
ÿ
X e eÿ
keik ˆ keki ˆ /e d e > 0; k 2 Kei …14†
q cmÿ1 jVi j/i ci ÿ cmÿ1 j xi ÿ xk j ik ik
i i
e2Ki
X X  
‡ sm q…cik † keik ÿ b^eik ceik …ci ÿ ck † and zero otherwise, where dike ˆ d…xeik †: see e.g. [2,14].
e2Ki k2Ke nfig
X ÿ
Remark 4. We may also choose xeij ˆ xe . Then, we get
‡ sm jVi e jq…Cie †Qei ci ÿ Cie ˆ 0; i ˆ 1; . . . ; I;
e2Ki 0 1
…13† q…c e
† X
keik ˆ @ j Ceij j neij A
De rNk …xe †; k 2 Ke n fig:
where we formally take b^eik ˆ 1=2, and ceik ˆ 0 when q…cik † j2Ke
i
k 2 Ke n Kei ; k 6ˆ i.
If T e is a rectangle or a triangle then we obtain that
Remark 2. Eqs. (6) and (13) form a system of nonlinear keik ˆ keki (cf. [15] for the case of a triangle). Moreover for
algebraic equations, to be solved by some nonlinear it- the isotropic case and when all angles of the triangular
erative solver, where in each iteration a linearisation of element T e do not exceed p=2, we also have keik P 0: see
the nonlinear algebraic system has to be provided. For [2,14]. Again, in general, no conclusion concerning the
our numerical simulations we choose a fully coupled, sign of keik is possible.
nonlinear solver of Newton type. All derivatives in the
Jacobi matrix with respect to the unknown discrete Below, the form (13) of the discrete transport equa-
variables are computed analytically and no simpli®ca- tion is the starting point for our considerations.
tions or decoupling techniques like Pickard iterations
are used: see e.g. [25].
Theorem 5. Let fci ; i ˆ 1; . . . ; N g fulfill (13). Denote
cimin ˆ miniˆ1;...;I^fci g and cimax ˆ maxiˆ1;...;I^fci g, where 1 6
4. Discrete minimum and maximum principle I^ 6 I. Furthermore, suppose that for i ˆ imin and i ˆ imax
we have
Physically, the range of values for the concentration
of the transported species should be bounded ± above keik P b^eik ceik for e 2 Ki ; k 2 Ke n fig: …15†
 P. Frolkovic, H. De Schepper / Advances in Water Resources 24 (2001) 63±72 67

Then, and to which numerical quadrature has been applied.

^ j 2 Ke ; e 2 Ki g; Yet, we can interpret upwind methods as the replace-
cimin P minfminfcj j j > I; min
ment of c~ by another approximation of c, which is more
cimÿ1
min
; minfCiemin j e 2 Kimin gg; suitable for the convective±di€usive ¯uxes.
^ j 2 Ke ; e 2 Kimax g;
cimax 6 maxfmaxfcj j j > I; We proceed as follows: see [2,24]. De®ne, for some
function u…s†; s 2 …0; 1†,
cimÿ1
max
; maxfCiemax j e 2 Kimax gg:
Jije  Jije …u…s†† ˆ ceij …u…s† ÿ ci † ÿ keij os u…s†;
Proof. Consider (13) for i ˆ imin and i ˆ imax . It suces i ˆ 1; . . . ; I; e 2 Ki ; j 2 Kei :
to notice that all coecients in this equation are non-
negative and some are strictly positive.  For u~…s†  c~…x…s††, with x…s†  xi ‡ …xj ÿ xi †s, we have
u~…s† ˆ ci ‡ …cj ÿ ci †s. The corresponding values
The theorem above allows us to study the numerical
Jije …~
u…1=2†† appear exactly in (13). Moreover, in the sit-
solution locally, for any subset of fci j i ˆ 1; . . . ; Ig, in-
uation of Remark 3 we have
cluding the case of a single equation. In particular, when  
(15) is not ful®lled, this theorem is no longer valid and j Ceij j neij
veij …c…tm ; x…s†† ÿ ci † ÿ dije rc…tm ; x…s††
non-physical behaviour of the numerical solution could
be observed. This may occur in the following two situ- ˆ Jije …c…tm ; x…s†††:
ations. First, if keik < 0 in (15), the condition fails when
We aim at replacing the function u~, approximating
ceik P 0. In other words, (15) can be violated, even in the
c…tm ; x…s††, by a more appropriate interpolation pro®le u^,
case of zero convection …ceik ˆ 0†, due to the approxi-
leading to a constant value of Jije …^
u…s†† for s 2 …0; 1†. This
mation of the di€usive±dispersive ¯uxes. However, also
function is found as the exact solution of the equation
when keik P 0, (15) can fail, viz., when ceik =keik > 1=b^eik ˆ 2.
Mathematically, this corresponds to the numerical rep- os Jije …u…s†† ˆ 0; s 2 …0; 1†; with u…0† ˆ ci ; u…1† ˆ cj :
resentation of so-called convection dominated trans- …18†
port, dealt with in Section 5.
For keij ˆ 0, i.e., the hyperbolic case, only the BC at the
in¯ow boundary is imposed, i.e., u…0† ˆ ci for ceij > 0 or
5. Upwind algorithms u…1† ˆ cj for ceij < 0. Consequently, the function u^ is
given by
The term convection dominated transport is used u^  u^…s†
when the ratio of the relative contribution of convection 8 exp…P e s†ÿ1
versus di€usion±dispersion to the solute transport is >
> ci ‡ …cj ÿ ci † exp…Pije †ÿ1 ; keij 6ˆ 0; ceij 6ˆ 0;
>
< ij
larger then one. For the discrete transport equation (13), keij ˆ ceij ˆ 0;
ˆ ci ‡ …cj ÿ ci †s; 6 0; …19†
the relation between the e€ectiveness of discrete mass >
> c; keij ˆ 0; ceij > 0;
>
: i
transport due to convection and to di€usion±dispersion cj ; keij ˆ 0; ceij < 0:
can be expressed in a convenient way by the (dimen-
sionless) grid Peclet numbers, Next, we de®ne J^ije ˆ …keij ÿ beij ceij †…ci ÿ cj †, where
81
ceik 1 e e
Pike ˆ ; when keik 6ˆ 0 for e 2 Ki ; k 2 Ke n fig: < Pije ÿ exp…Pije †ÿ1 ; kij > 0; cij 6ˆ 0;
>
keik e
bij ˆ 0; e
kij 6 0; ceij > 0; …20†
>
: 1;
…16† keij 6 0; ceij < 0:
Hence, when keik > 0, Eq. (15) can be reformulated as Note that if keij P 0, then J^ije  Jije …^
u†. Moreover, we do
Pike 6 2; e
e 2 Ki ; k 2 K n fig: …17† not need to de®ne beij when ceij ˆ 0. When keij < 0, to
avoid so-called downwind schemes, the value of beij cor-
If (17) fails, the minimum and maximum principle above
responds to the case keij ˆ 0 in (19). If keij ˆ keji , we can
does not hold, and we can expect instabilities (non-
easily show that beji ˆ 1 ÿ beij , as Pjie ˆ ÿPije . In general,
physical oscillations) in the numerical simulations. In
we define b^eij , entering (10), by
order to avoid the restriction (17), we resort to other
methods when dealing mathematically with the case of a b^eij ˆ beij for ceij > 0 and b^eij ˆ 1 ÿ b^eji for ceij < 0:
convection dominated discrete transport equation. We
…21†
present so-called upwind methods, which retain the
concept of local mass conservation. As ˆ ceij ÿceji ,
this de®nition is consistent. Together with
The idea is to modify (13), by no longer taking (20) it de®nes the so-called exponential upwind scheme.
b^eij ˆ 1=2. Consequently, we are not dealing anymore Moreover, from (10), we have that c^eij  c^eji and hence
with a direct approximation of the exact mass balance the global discrete mass conservation property is
(5), in which c and p have been replaced by c~ and p~, (4) preserved.
68 P. Frolkovic, H. De Schepper / Advances in Water Resources 24 (2001) 63±72

Using (21) in the discrete transport equation (13), (15) 6. Numerical experiments
is seen to be ful®lled for keij P 0. Hence, in this case the
discrete minimum and maximum principle is ensured, 6.1. Description of the problem
even for convection dominated transport. The only
drawback of the scheme (20) may be the expensive The Elder problem is an example of density driven
evaluation of the exponential function. To avoid this, ¯ow, motivated by experimental measurements of ther-
other upwind schemes are considered, which are based mally driven convection in porous media, see [6]. A
on a suitable modi®cation of the exponential interp- solutal analogue of the thermal convection was de®ned
olation pro®le (19), see [2,24] for details. The simplest ex- in [30] and since then the Elder problem became a widely
ample is the full upwind scheme, where (20) is replaced by used benchmark for density driven ¯ow simulators, see
[1,18,19,21,23], or see [3,13,22] for closely related prob-
beij ˆ 0 for ceij > 0 and beij ˆ 1 for ceij < 0: …22† lems. It describes an instable situation in a closed box,
where a dense ¯uid lies on top of a less dense ¯uid, see
Notice that, for keij 6 0, (22) appears in (20). Another
also the modelling of an evaporating salt lake [26]. The
popular method is the partial upwind (or ``hybrid'')
complexity of the Elder problem may be presumed from
scheme: see [14,24]. Here (20) is replaced by
the fact that many authors report qualitatively di€erent
81
> ; keij > 0; ÿ2 6 Pije 6 2; numerical results.
>
> 2
>
> 1
; keij > 0; Pije > 2; In this paper, we use the Elder problem to illustrate
>
< Pije
e the discretisation method outlined above (including
beij ˆ 1 ‡ P1e ; kij > 0; Pije < ÿ2; …23† upwind methods), and we moreover give an explanation
>
> ij
>
> 0; e e
kij 6 0; cij > 0; for seemingly incompatible results presented in other
>
>
:
1; keij 6 0; ceij < 0: papers. Furthermore, we try to provide a reference nu-
merical solution for this problem.
Eq. (23) constitutes the smallest modi®cation of (13), The domain is taken to be rectangular, corresponding
required to ful®ll (15). In fact for ÿ2 6 Pije 6 2, no up- to a vertical cross-section of the physical domain.
wind is applied in (23). Moreover, by the symmetry of the original problem, it
For a comparison of the schemes based on (20), (22) suces to consider only (the left) half of the rectangle
and (23), we refer to Fig. 2 below, where the dependence considered in [6,30].
of beij on Pije ˆ ceij (with keij  1) is shown. From Fig. 2, it Hence, the computational domain has dimensions
is clear that visible di€erences between the exponential 300 m  150 m, with left bottom corner …x; y† ˆ …0; 0†
and the partial upwind scheme only occur for jPije j < 6, and right top corner …x; y† ˆ …300; 150†.
while the full upwind scheme di€ers signi®cantly from Initially, c…0; x; y† ˆ 0 inside the domain. For the
the others for jPije j < 15. BCs, we strictly follow the model outlined in [30] and
Finally, we note that the exponential and the partial implemented by other authors also, although a more
upwind scheme depend on the discrete unknowns appropriate choice could be made, see e.g. [8,20]. For
through beik ˆ beik …Pike † and Pike ˆ Pike …ci ; ck †. Hence the the concentration the BCs are given by the Dirichlet
linearisation of a nonlinear iterative solver has to take conditions
into account the derivatives of beik w.r.t. ci and ck . In
(20), beik depends smoothly on Pike , in contrast to (23), c…t; x; 0† ˆ 0 for x 2 …0; 300†; t P 0;
where the ®rst derivative jumps for Pike ˆ 2. c…t; x; 150† ˆ 1 for x 2 …150; 300†; t P 0

and by zero ¯ux conditions elsewhere. For the pressure,

a zero ¯ux BC is prescribed everywhere, except at the
left top corner, where the pressure is supposed to be
®xed at the value p ˆ 0. Further, we need to specify the
data appearing in the mathematical model. First, the
equation of state is given by
 ÿ1
1ÿc c
q…c† ˆ ‡ ;
qw qb
where qw ˆ 1000 and qb ˆ 1200. Other data are / ˆ 0:1,
k ˆ 4:845  10ÿ13 , l ˆ 10ÿ3 , g ˆ …0; ÿ9:81†, D  /dI,
with d ˆ 3:565  10ÿ6 . All values, as well as the speci®c
form of the equation of state, were chosen as in [23].
Fig. 2. Comparison of beij …Pije † for the upwind schemes (20), (22) and The discretisation method described in this paper was
(23). implemented in the D3 F (distributed density driven
 P. Frolkovic, H. De Schepper / Advances in Water Resources 24 (2001) 63±72 69

¯ow) package, see [8], itself being based on the numer-

ical library UG (unstructured grids), [3]. In addition, we
applied the algorithm of consistent velocity approxi-
mation, see e.g. [7]). As we aim at comparing our
computations with results from other papers, where no
upwind schemes were applied, we have used the partial
upwind scheme (23), which only introduces a minimal
modi®cation with respect to the no upwind scheme. A
discussion of di€erent upwind schemes follows at the
end of this section.
The nonlinear algebraic system was solved by a fully
coupled, nonlinear iterative solver of Newton type [3],
with analytical computations of all derivatives with re-
spect to the unknown discrete variables in the corre-
sponding Jacobi matrix. After each Newton iteration the
discrete L2 norm of the defect (residual) of the system
was computed. The iterative process was stopped when
the norm of the initial defect (computed from the initial
guess) was reduced by a factor 10ÿ6 compared to the
norm of the last defect. As the convergence for the
chosen time step was very good, the iterative procedure
required in average only three iterations, after which the
norm of the initial defect was reduced in average by a
factor 10ÿ11 .

6.2. Discretisation method and numerical results

We discretise the domain using a structured, uniform

grid with bilinear square elements. Such a grid can be
characterised by its level, de®ned as follows: the grid at Fig. 3. Results for t ˆ 1; 2; 2:5; 3; 4; 5; 6; 7; 10; 12; 15 and 20 yr
the lth level, (l ˆ 0; 1; 2; . . .), consists of 22l‡1 identical at grid level 7.
square elements. The computations were executed at
grid level 7: (i.e., with 32 768 elements) and with a uni- ®nally in (10 yr, 12 yr). As a consequence, when t > 12
form time step s ˆ 0:025 yr. The corresponding nu- yr, only a half ®nger of brine, accompanied of one re-
merical results are presented in Fig. 3, where the mass circulation cell, can still be observed in Fig. 3. This
fraction isolines c ˆ 0:2; 0:4; 0:6 and 0.8 are shown, as constitutes a major di€erence with the results in
well as the velocity ®eld v at chosen points in the do- [19,21,23], where a complete ®nger of brine is observed
main, at 12 di€erent time points. Grey arrows in the for t ˆ 20 yr.
velocity plot indicate that the length of the velocity To estimate the truncation error from the time dis-
vector v was cut beyond some critical value, in order to cretisation, the ``reference'' simulations were performed
avoid long arrows in the graphics. for the time step s ˆ 0:00625 yr as well. No important
In what follows, we aim at interpreting the results visible di€erences were observed with respect to Fig. 3.
obtained and at comparing them with those presented in In [23], the computations were executed for grid level
[19,21,23]. 5 and for a non-uniform grid which closely resembles
Due to the initial density di€erence at the midpoint of our grid level 6. As the results obtained did not di€er
the upper boundary, a ¯ow pattern is developed in the signi®cantly for these two ``consecutive'' grids, the
form of a recirculation cell. Thus, due to the convective authors concluded to achieve so-called ``grid conver-
transport, a so-called ``®nger'' of brine appears, ac- gence''. This conclusion is also quoted by other authors,
companied of two recirculation cells, see Fig. 3 for t ˆ 1 see e.g. [19,26], but seems to be incorrect. To be able to
yr. Next, density di€erences in the neighbourhood of the argue this more profoundly, we have executed some
®rst ®nger cause the development of a second one, see computations on grids of the levels 5 and 6. Here, we
Fig. 3 for t ˆ 2 yr. In total, three ®ngers and a half are have obtained results which are very similar to those
developed (7 in the original domain), see Fig. 3 for presented in [1,19,21,23], as can be seen in Fig. 4.
t ˆ 2:5 and t ˆ 3 yr. However, meanwhile, mutual in- From these computations, we may also observe that,
teractions cause the merging of two ®ngers at some time for t ˆ 4 yr, the numerical solutions for the grid levels 6
t in the interval (2 yr, 3 yr), as well as in (5 yr, 6 yr), and and 7 are still very resembling, with a minor di€erence
70 P. Frolkovic, H. De Schepper / Advances in Water Resources 24 (2001) 63±72

Fig. 4. Results for t ˆ 4; 7 and 20 yr at grid level 6. Fig. 6. Di€erent stationary solutions at grid level 7.

concerning the half ®nger at the top right corner. On the presented in Fig. 6 for grid level 7 ± however, all three
coarser grid, this half ®nger disappears and the 2 re- can be observed for various grid levels.
maining ®ngers merge at a later time, see t ˆ 7 and The ®rst solution, shown in the left picture of Fig. 6
t ˆ 20 yr. For grid level 5 the half ®nger does not even was obtained by computing the Elder problem for grid
occur at all for t ˆ 4 yr. In general, the ®ner the grid, the level 7 and for large simulation times. The second one, in
more ®ngers we may observe at an early stage in the time the middle, was found by computation of the stationary
process. This leads us to the conclusion that the nu- solution for grid level 6, followed by a uniform re®ne-
merical solution at grid level 7 is closer to the exact ment of the grid and transient computations for grid
solution of the Elder problem than the one at grid level 7. Finally, the third solution, at the right of Fig. 6
level 6. was obtained by an analogous strategy, now starting
To support this conclusion we resorted to a further, from grid level 4.
local adaptation of the grid, as the computational cost Consequently, one may expect very di€erent numer-
became too high for a uniform re®nement up to level 8. ical solutions, when imposing even slightly modi®ed ICs.
In fact, the grid was uniformly re®ned up to level 6, but To illustrate this assertion we consider a very small and
the next two grid levels were created by a local re®ne- local perturbation of the standard ICs. We take e.g.,
ment procedure, using an a posteriori error estimator, c…0; 300; 150 ÿ 150=64† ˆ 0:01, i.e., for one of the ver-
described in [18]. Moreover, this grid re®nement (and tices of the square element at the upper right corner of
coarsening) was also applied during the time simula- the domain, we replace the original zero IC by the value
tions. The error estimator required to re®ne all elements 0.01. The numerical results, obtained for grid level 6 are
near the top boundary side, causing the grid to resemble presented in Fig. 7 at three di€erent times. At t ˆ 0:5 yr,
strongly the one of level 8 at this part of the domain. one can still observe a strong resemblance with the un-
The numerical results correspond very well to the results perturbed problem, but at t ˆ 4 yr, the di€erences are
at grid level 7: see Fig. 5 for an illustration. Note that the already very clear. In fact, the solution for the perturbed
unstructured grid in Fig. 5 not only contains square problem will asymptotically reach the stationary solu-
elements, but also triangles and general quadrangles: see tion, where a half ®nger of brine occurs, while for the
[3,18]. unperturbed problem, the stationary solution shows a
The observation of di€erent ``asymptotic'' behaviour complete ®nger.
of the numerical solutions suggests that non-unique Di€erent transient behaviour of the numerical solu-
stationary solutions exist for the Elder problem. In this tions can also be caused by the numerical method itself,
context, the notion of a (numerical) stationary solution see e.g. [22] for the problem of Benard convection. In
is used for the transient problem, if the numerical so- particular, when the grid is not aligned according to the
lution no longer changes while continuing with the time sides of the domain, the interpolation of the ICs can
process. For the qualitatively very similar problem of cause a perturbation which, eventually, may determine
Benard convection, the existence of several stationary the character of the numerical solution. When however,
solutions has been proved, see [11,13,22]. Here, we have the re®nement of the grid does not a€ect the character of
found three di€erent numerical stationary solutions, the perturbation (e.g., by uniform re®nement of the
which di€er in the number of ®ngers and, correspond- initial unstructured grid), one can observe grid conver-
ingly, in the number of recirculation cells. They are gence of the numerical solution, but the result may di€er

Fig. 5. The locally adaptive grid and the corresponding results for Fig. 7. Results for the perturbed problem at t ˆ 0:5, 2 and 4 yr at grid
t ˆ 4 yr. level 6.
 P. Frolkovic, H. De Schepper / Advances in Water Resources 24 (2001) 63±72 71

from the one presented in Fig. 3. Nevertheless, we may For the coarse grid level 4 the large negative oscilla-
expect the uniform and aligned grid with square ele- tions develop (c  ÿ1). They persist in the computations
ments to be ``free'' of perturbation. up to grid level 6, but they are getting smaller and they
obtain a rather local character. The computations for
the exponential upwind scheme are presented in Fig. 10.
6.3. Application of various upwind methods
The comparison of Figs. 9 and 10 suggests that the
numerical solutions for grid level 5 for both schemes
To illustrate the features of the di€erent upwind
represent well the exact solution for this special choice of
schemes, for the Elder problem, we compare the cases of
ICs, but only the upwind scheme produces the solution
the no upwind scheme, the exponential upwind scheme
with no non-physical oscillations. For the coarse grid
(20) and the full upwind scheme (22).
level 4 only the upwind method seems to produce a
First, we have executed our computations on a coarse
reasonable numerical approximation of the solution.
grid of level 4, see Fig. 8 for a comparison of the results
As in [13,19,23] one can estimate the ``global'' Peclet
at t ˆ 3:5 yr. When no upwind is applied, we observe
number (also referred to this type of problems as the
large negative values of c near the middle of the upper
``Rayleigh'' number) by
boundary. For both upwind methods no negative values
of c occur, but only the results based on the exponential …qb ÿ qw †gkH
scheme resemble the solution with no upwind, while P :ˆ ; …24†
l/d
those corresponding to full upwind are signi®cantly
di€erent. where g ˆ jgj and H is the height of the domain. For the
In general, the ®ner the grid, the less apparent the Elder problem we have P  400. This means that for
di€erences between all methods become. In fact for grid grid level 0 we can estimate the local Peclet numbers by
level 7 the di€erences between the no upwind scheme, Pike < 400, for grid level 4 we have Pike < 25 and ®nally for
the exponential scheme (20) and the partial upwind grid level 7 we have Pike < 3; 125. From this point of view
scheme (23) are negligible and not visible in Fig. 3. The the behaviour of the no upwind scheme and of the up-
negative oscillations in case of no upwind only occur at wind schemes for di€erent grid levels corresponds to our
the early stage of the time process and disappear later results on the discrete maximum principle.
on. On the other hand, the full upwind scheme produced Finally, we can con®rm a slightly better convergence
the numerical solution analogous to the Fig. 4, even for of the iterations for computations with the exponential
grid level 7. upwind scheme than with the partial upwind scheme.
As it is clear from the previous subsection, very ®ne The convergence was very good in both cases and the
grid re®nements are necessary for the Elder problem to di€erence only became more pronounced for larger time
resolve the correct development of the numerical solu- steps.
tion from the ICs. To explain clearly the advantages of As in practice, the computational memory is always
advanced upwind schemes we have computed the Elder restricted, advanced upwind methods, like the partial or
problem for di€erent ICs, represented by the results of the exponential upwind scheme, o€er several advan-
the full upwind scheme in the right picture of Fig. 8 at tages. They produce a physically acceptable numerical
t ˆ 3:5 yr. Starting from these ICs, grid convergence was solution, without restrictions on the grid Peclet num-
reached very quickly. This is con®rmed by the following bers, moreover avoiding excessive grid re®nement in the
computations. The results for the no upwind scheme are case of convection dominated transport. The full up-
presented in Fig. 9, at t ˆ 7 yr for grid levels 4, 5 and 6. wind scheme, on the other hand, causes unnecessary
large smoothing (or damping) e€ects and should rather
be replaced by more advanced schemes.
It is well-known that upwind methods can impose a
grid e€ect (arti®cial di€usion) on the numerical solution,
if the grid is not aligned to the convective velocity ®eld,
see [24]. To reduce such grid e€ects the method of
aligned ®nite volumes was implemented in D3 F: see [16].
Fig. 8. Comparison of no upwind with the exponential and the full
upwind scheme.

Fig. 10. The exponential upwind scheme for the grid levels 4, 5 and 6
Fig. 9. The no upwind scheme for the grid levels 4, 5 and 6 at t ˆ 7 yr. at t ˆ 7 yr.
72 P. Frolkovic, H. De Schepper / Advances in Water Resources 24 (2001) 63±72

As described in [8,9], upwind methods can successfully [15] Jianguo H, Shitong X. On the ®nite volume element method for
be implemented in this context as well. general self-adjoint elliptic problems. SIAM J Numer Anal
1998;35(5):1762±74.
[16] Johannsen K. Aligned 3D-®nite-volumes for convection±di€u-
sion-problems. In: Benkhaldoun F, Vilsmeier R, editors. Finite
Acknowledgements volumes for complex applications. Paris: Hermes, 1996. p. 291±
300.
One of the authors (PF) was ®nancially supported by [17] Knabner P, Frolkovic P. Consistent velocity approximations in
the VEO-project no. 011 VO 697. The authors thank ®nite element and volume discretizations of density driven ¯ow in
porous media. In: Aldama AA, et al., editors. Computational
R. Van Keer, coordinator of this project, for stimulating
methods in water resources XI. Southampton: Computational
discussions and for his critical reading of the text. Mechanics Publications, 1996. p. 93±100.
[18] Knabner P, Tapp C, Thiele K. Adaptive ®nite volume discret-
ization of density driven ¯ows in porous media. Acta Math Univ
References Com 1998;67(1):115±36.
[19] Kolditz O, et al. Coupled groundwater ¯ow and transport: 1.
[1] Ackerer Ph, Younes A, Mose R. Modelling variable density ¯ow Veri®cation of variable density ¯ow and transport models. Adv
and solute transport in porous media: 1. Numerical model and Wat Res 1998;21(1):27±46.
veri®cation. Transp Porous Media 1999;35(3):345±73. [20] Konikow LF, Sanford WE, Campbell PJ. Constant-concentration
[2] Angermann L. An introduction to ®nite volume methods for boundary conditions: lessons from the HYDROCOIN variable-
linear elliptic equations of second order. Institut f
ur Angewandte density groundwater benchmark problem. Wat Res Res
Mathematik, Universit at Erlangen, 1995. Preprint no. 164. 1997;33(10):2253±61.
[3] Bastian P, et al. UG ± a ¯exible software toolbox for solving [21] Leijnse A. Three-dimensional modeling of coupled ¯ow and
partial di€erential equations. Comput Visual Sci 1997;1(1):27±40. transport in porous media, PhD thesis. Indiana: University of
[4] Bear J, Verruijt A. Modelling groundwater ¯ow and pollution. Notre Dame, 1992.
Dordrecht: Reidel; 1987. [22] Leijnse A, Oostrom M. The onset of instabilities in the numerical
[5] Ciarlet PG. The ®nite element methods for elliptic problems. simulation of density-driven ¯ow in porous media. In: Peters A,
Amsterdam: North-Holland; 1978. et al., editors. Computational methods in water resources X.
[6] Elder JW. Transient convection in a porous medium. J Fluid Dordrecht: Kluwer Academic Publisher; 1994. p. 489±96.
Mech 1967;27(3):609±23. [23] Oldenburg CM, Pruess K. Dispersive transport dynamics in a
[7] Frolkovic P. Consistent velocity approximation for density driven strongly coupled groundwater±brine ¯ow system. Wat Res Res
¯ow and transport. In: Van Keer R, et al., editors. ACOMEN'98. 1995;31(4):289±302.
Maastricht: Shaker; 1998. p. 603±11. [24] Patankar SV. Numerical heat transfer and ¯uid ¯ow. New York:
[8] Frolkovic P. Discretization. In: Fein E, editor. d3 f ± a simulator Hemisphere; 1980.
for density-driven ¯ow. Gesellschaft fuer Anlagen-und Reaktorsi- [25] Putti M, Paniconi C. Picard and Newton linearization for the
cherheit, Braunschweig, 1999. coupled model of saltwater intrusion in aquifers. Adv Wat Res
[9] Frolkovic P. Maximum principle and local mass balance for 1995;18(3):159±70.
numerical solutions of transport equation coupled with variable [26] Simmons CT, Narayan KA, Wooding RA. On a test case for
density ¯ow. Acta Math Univ Com 1998;67(1):137±57. density-dependent groundwater ¯ow and solute transport models:
[10] Fuhrmann J, Langmach H. Stability and existence of solutions of the salt lake problem. Wat Res Res 1999;35:3607±20.
time-implicit ®nite volume schemes for viscous nonlinear conser- [27] Sperb RP. Maximum principles and their applications. New York:
vation laws. Weierstrass-Inst. fur Angew. Analysis und Stochas- Academic Press; 1981.
tik, Berlin, 1998. Preprint 437. [28] Trujillo RV. Error analysis of the ®nite volume element method
[11] Gebhart B, et al. Buoyancy-induced ¯ows and transport. New for elliptic and parabolic partial di€erential equations, PhD thesis.
York: Hemisphere; 1988. Denver: University of Colorado; 1996.
[12] Hassanizadeh SM, Leijnse T. On the modeling of brine transport [29] van Duijn CJ, Peletier LA, Schotting RJ. Brine transport in
in porous media. Wat Res Res 1988;24:321±30. porous media: self-similar solutions. Adv Wat Res 1998;
[13] Holzbecher EO. Modeling density-driven ¯ow in porous media. 22(3):285±97.
Berlin: Springer; 1998. [30] Voss CJ, Souza WR. Variable density ¯ow and solute transport
[14] Ikeda T. Maximum principle in ®nite element models for convec- simulation of regional aquifers containing a narrow freshwater±
tion±di€usion phenomena. Amsterdam: North-Holland; 1983. saltwater transition zone. Wat Res Res 1987;23:1851±66.