Professional Documents
Culture Documents
1. Introduction
Density-driven flow in porous media is relevant in a number of frequently occur-
ring problems like saltwater intrusion, upconing phenomena and propagation of dense
plumes. Several model problems have been set up to investigate the behavior analytically
as well as numerically. The Elder problem and its reformulations have been designed to
investigate the transient behavior of instable situations where convection is set up only
by density differences. Since the early days of this problem two different stable steady
state solutions have been found numerically for the prescribed initial conditions. Even
though existence and uniqueness can be shown only under special conditions (Boussi-
nesq approximation, no-flow boundary conditions, see [3,9]) there are no reasonable
doubts that these results hold for the equations considered here. The appearance of two
different asymptotic solutions is due to numerical errors and the highly sensible char-
acter of this model problem. Recently, grid convergence studies have been carried out
confirming that the downwelling (“one-finger”) steady state solution corresponds to the
initial conditions [6]. For this simulation the Boussinesq approximation has been used.
Similar investigations using the full set of equations have not been carried out yet.
170 K. Johannsen / On the validity of the Boussinesq approximation
The Elder problem [7,8] is an example of density driven flow where the convection
is purely induced by density differences. The original thermally driven flow problem set
up by Elder has been reformulated by Voss and Souza [20], among others. Here the den-
sity differences are due to presence of heavy brine. Furthermore the geometry has been
resized keeping the Rayleigh number and therefore the physical behavior unchanged.
The governing equations for the coupled density-driven flow in porous media are
derived from mass-conservation principles. Assuming incompressibility, it can be de-
scribed by a system of two partial differential equations (see [2,14,18])
∂(nρ)
+ ∇ · (ρv) = 0, (1a)
∂t
∂(nρω)
+ ∇ · (ρωv − ρDm ∇ω) = 0, (1b)
∂t
K. Johannsen / On the validity of the Boussinesq approximation 171
Table 1
Parameters of the Elder problem.
Symbol Value Unit Symbol Value Unit
n 0.1 – ωmax 20% –
Dm 3.565 · 10−6 m2 s−1 ρmin 1.0 · 10−3 kg m−3
µ 10−3 kg m−1 s−1 ρmax 1.2 · 10−3 kg m−3
K 4.845 · 10−13 m2 |g| 9.81 m s−2
where ω denotes the salt mass fraction, ρ the density, n the porosity, Dm the molecular
diffusion coefficient and v the mass-averaged velocity given by Darcy’s law
K
v=− (∇p − ρg). (2)
µ(ω)
K is the permeability, µ denotes the viscosity and g the gravity vector. Dispersive effects
are neglected (the diffusion in equations (1) equals the molecular diffusion coefficient)
and permeability and viscosity are constant. The density depends on the salt mass frac-
tion only and is given by the ideal mixing relation (see [13,14])
1 ω 1 ω 1
:= 1 − + , (3)
ρ ωmax ρmin ωmax ρmax
where ωmax is the maximum salt mass fraction and ρmin and ρmax are the minimum and
maximum density, respectively. The parameters are given in table 1.
The geometry and boundary conditions are given in figure 1 together with a coarse
grid. The boundary conditions indicated for ω and p are Dirichlet boundary conditions.
On the other parts of the boundary no-flux conditions are specified
nv = 0, (4a)
n(vω − Dm ∇ω) = 0. (4b)
Equations (1) form a differential algebraic system, therefore initial conditions have to be
specified only for the salt mass fraction
ω(x) = 0 for t = 0. (5)
172 K. Johannsen / On the validity of the Boussinesq approximation
Steady state solutions are obtained solving the corresponding steady state equations
∇ · (ρv) = 0, (6a)
∇ · (ρωv − ρDm ∇ω) = 0, (6b)
∇ · v = 0, (7a)
∂(nω)
+ ∇ · (ωv − Dm ∇ω) = 0. (7b)
∂t
Correspondingly the steady state equations read
∇ · v = 0, (8a)
∇ · (ωv − Dm ∇ω) = 0. (8b)
[K(ρmax − ρmin)|g|h]
R := , (9)
[Dm µn]
where h = 150 m is the characteristic length of the problem. For the Elder problem we
have R ≈ 400.
The density driven flow equations are solved using the program package d 3 f . It
uses a finite volume scheme for discretization in space. Central differencing is applied
leading to a nonmonotonous scheme which is second-order accurate. In case of the
transient problem (1), (7), the backward Euler method is used for time integration. Fur-
thermore a consistent velocity formulation is used [11]. The nonlinear discrete system
of equations is solved using Newton’s method. The linear subproblems are inverted ap-
plying an appropriate multigrid scheme accelerated by a Krylov subspace method. The
calculations were carried out on a MIMD parallel computer using up to 32 processors.
For details see [10,16].
We use square grids to perform the numerical simulations. They are derived from
the coarse grid containing 4 squares by uniform refinement. In figure 1 grid level 3 is
displayed having 297 grid points. We carried out the calculations on grid level 5 to 8
corresponding to n5 = 4.257, n6 = 16.705, n7 = 66.177 and n8 = 263.425 grid points,
respectively.
K. Johannsen / On the validity of the Boussinesq approximation 173
3. Dynamical systems
In this section we consider the time evolution process of the density-driven flow
from a dynamical systems point of view. We are interested in its dependence on the
Rayleigh number R. Therefore the value of ρmax from table 1 is considered to be a
function of R, which is implicitly defined by (9). We briefly recall some basic facts
from the theory of dynamical systems and focus on path continuation methods. A so-
called pseudo-arclength continuation method is described, which is used in section 4
to calculate some solution branches connecting the different steady state solutions. For
details, see [1,17].
Definition 1. A dynamical system is a pair (ϕ t , X), where X is the state space and
ϕ t : X → X a continuous time evolution operator satisfying
ϕ t x = x, ∀t ∈ R+
0. (11)
ϕ t y → x, t → ∞ ∀y ∈ U, (12)
Definition 3. The bifurcation diagram of the steady state solutions of a dynamical sys-
tem (ϕαt , X, I ) depending on a real parameter α ∈ I ⊂ R is defined to be
S := (x, α) ∈ X × I | ϕαt x = x, ∀t ∈ R+ (13)
Due to the lack of analytical results the bifurcation diagrams Sf and Sb have to
be approximated numerically. To this end we use a path continuation method. Below,
we give a short introduction and describe the algorithm implemented in d 3 f , which has
been used to obtain the results presented in section 4.
Let f denote either fb or ff and let B be a solution branch of either Sb or Sf . Let
(x, r) ∈ B and ·, · be an appropriate scalar product. Furthermore let A denote the
Jacobian of f with respect to x only (r fixed). In generic situations one of the following
cases apply to construct a new solution (x̃, r̃) ∈ B:
C1. A is regular. Then there exists a δr = 0 such that f (x̃, r + δr) = 0 has a unique
solution x̃ with (x̃, r + δr) ∈ B.
C2. A is singular. Then there is a s with As = 0 and a δ > 0 such that the extended
system f (x̃, r̃) = 0 ∧ x̃ − x, s = δ has a unique solution in B.
In case C2 the system undergoes a Fold-bifurcation. Here, two steady state solu-
tions appear varying the Rayleigh number. On this base the pseudo-arclength continua-
tion method with secant prediction is realized, which additionally provides a prediction
for the nonlinear iteration and a step size control.
K. Johannsen / On the validity of the Boussinesq approximation 175
Algorithm 1.
1. Initialization
Choose appropriate δr = 0 and (x0 , r0 ) with f (x0 , r0 ) = 0.
Set i = 0, k = 0, η = 1.
2. Newton step
Set ri+1 = ri + δr and solve for xi+1 :
f (xi+1 , ri+1 ) = 0
If i = 0, use xi as initial guess,
else use 2xi − xi−1 as initial guess.
If a solution can be found, set i = i + 1, repeat step 2.
3. Extended Newton step
If i = 0, exit with error.
If η < 1e–5, exit with error.
Set s = xi − xi−1 . Solve for (xi+1 , ri+1 ):
f (xi+1 , ri+1 ) = 0, xi+1 − xi , s = ηs, s.
Use (2xi − xi−1 , r + ηδr) as initial guess.
If no solution can be found, set η = η/2, repeat step 3.
If η < 1, set η = 2η, k = 0,
else, set k = k + 1.
If (ri+1 − ri )(ri − ri−1 ) < 0, set δr = −δr, k = 0.
Set i = i + 1.
If k < 5, repeat step 3,
else, set k = 0, repeat step 2.
The adaption of the parameter η turns out to be necessary in practice. The nonlin-
ear system to be solved in step 2 is carried out as described in section 2.2. The extended
nonlinear system in step 3 is solved using Newton’s method. The arising linear subprob-
lems are solved using a Schur-complement method which again involves a multigrid
method accelerated by a Krylov-subspace method.
Remark 1. Fold bifurcation points are local extrema of the Rayleigh number compo-
nent of the solution branches. Algorithm 1 calculates discrete subsets of the branches.
The corresponding discrete local extrema serve as an approximation to fold bifurcation
points. To improve the approximation the initial parameter step size δr on a fixed spatial
grid has to be decreased.
Remark 2. If the problem has internal symmetries, case C2 may describe fold- as well as
pitchfork bifurcations. In case of a pitchfork bifurcation solutions branches with broken
symmetry appear and the local structure of the solutions is more complex. For a detailed
discussion, see [4,19].
176 K. Johannsen / On the validity of the Boussinesq approximation
4. Results
In this section we present numerical results for the Elder problem. We compare
steady state solutions obtained using the full set of equations (1) on the one hand and the
Boussinesq approximation (7) on the other hand. For Rayleigh number R = 400, we
show that to every solution of equations (1) corresponds a solution of equations (7) and
vice versa. Then we investigate the bifurcation diagram in both cases. It turns out that
the bifurcation diagrams are topologically not equivalent. All solutions presented in this
section could be represented on every of the 4 grids.
The results for the full set of equations have been reported already in [15]. Eleven
steady state solutions have been shown to exist. Three of them are stable. They have
been documented first by Frolkovic and Schepper [12]. Furthermore eight unstable
steady state solutions have been documented. They have not been mentioned in liter-
ature before. Additionally we present the corresponding solutions for the Boussinesq
approximation.
The stability of a steady state s has been investigated by transient simulation. As
initial conditions a small, random perturbation of s has been used.
Figures 2 and 3 show the stable and unstable solutions of the Elder problem, respec-
tively. They display the contour lines of the salt mass fraction with the values ω = 0.1,
0.2, . . . , 0.9ωmax . Stable solutions are denoted by s, unstable solutions by u. Solutions
obtained using the Boussinesq approximation take a subscript b, solutions of the full set
Figure 3. Elder problem: unstable, steady state solutions. All solutions correspond to R = 400. Left
column: Boussinesq approximation, right column: full set of equations. The isolines correspond to the
values 0.1, 0.2, . . . , 0.9ωmax .
of equations a subscript f. The second subscript is used for enumeration purpose. The
superscript denotes the grid level. Figure 2 shows additionally the unique steady states
8
(sb,0 8
and sf,0 ) of the linear Elder problem (R = 0). The four pairs correspond clearly.
We express this by
8
sb,i ≈ sf,i
8
, i = 0, 1, 2, 3. (17)
Figure 3 shows the eight unstable steady states for both cases. Again, the solutions
correspond clearly, with a possible uncertainty in the correspondence of the pair (u6b , u7b )
with (u6f , u7f ). We express this by
u8b,i ≈ u8f,i , i = 1, 2, . . . , 8. (18)
The same holds true for grid level l = 5, 6, 7.
178 K. Johannsen / On the validity of the Boussinesq approximation
In this section we investigate the bifurcation diagrams Sb (14) and Sf (15). The
solution branches of Sb and Sf have been approximated using algorithm 1. To visualize
the bifurcation diagrams we use the projection
Pl : Xl × IR → R2 , Pl (xl , r) := L(xl ), r (19)
with
L(xl ) = L(ωl , pl ) = ωl (x, y)f1 (x)f2 (y) dx dy,
/
f1 (x) = c0 + c1 h(x, 300) + c2 h(x, 230) + c3 h(x, 190),
3 y − 150
f2 (y) = 1 − (150 − y) exp ,
4 10
−(x − x0 )2 −(x + x0 − 600)2
h(x, x0 ) = exp + exp ,
500 500
c0 = 1.68 · 10−6 , c1 = −3.64 · 10−4 , c2 = −9.57 · 10−4 , c3 = 1.19 · 10−3 ,
where the integral is evaluated using a fourth order quadrature rule. Instead of a solution
branch Bl we display its image B l under Pl . The projections Pl have been designed to
achieve a resonable separation of the eleven solutions of the original Elder problem.
The numerical approximation of a solution branch of Sb or Sf is a discrete set
of points (xl , r) calculated on one of the grids introduced in section 2.2. The initial
parameter step size δr for algorithm 1 was δr = −1 or −2, depending on the solution
branch. The initial Rayleigh number was r0 = 400.
Both for the Boussinesq approximation and the full set of equation we found six
solution branches. They are denoted by Bb,i l l
and Bf,i , where i = 1, . . . , 6 is an enumer-
ation index. Again, l denotes the grid level. It has been found numerically that
l
Bb,i l
, Bf,i are symmetric ∀i = 1, . . . , 6, l = 5, . . . , 8. (20)
They form discrete approximations of the bifurcation diagrams Sb , Sf
6
6
Sbl := l
Bb,i , Sfl := l
Bf,i . (21)
i=1 i=1
Sbl and Sfl are displayed in figures 4 and 5, respectively. A linear interpolation of the dis-
crete points is used. The horizontal axis corresponds to the Rayleigh number, the vertical
to L(xl ). In the figures all branches corresponding to the different grids are displayed.
∗ ∗
This is indicated by an asterisk for the level index, e.g., Bb,1 , sb,0 in figure 4. Therefore
every (type of) branch is four-fold, giving an idea of the grid convergence we obtained
∗ ∗
on the successively refined grids. E.g., the branches Bb,1 , Bf,1 show fast convergence,
the branches corresponding to the different grid levels nearly coincide. The branches
K. Johannsen / On the validity of the Boussinesq approximation 179
Figure 4. Solution branches of the Elder problem using Boussinesq approximation. The solutions corre-
spond to the ones in figures 2 and 3, left column.
Figure 5. Solution branches of the Elder problem using the full set of equations. The solutions correspond
to the ones in figrures 2 and 3, right column.
∗ ∗
Bb,3 , Bf,3 show relatively slow convergence. However, the topology of the branches does
not change, when the spatial grid is refined. We claim that this topology corresponds to
the topology of the continuous solutions. Note, that intersection of two branches does
not mean two solutions coincide, since L(xl ) = L(yl ) does not imply xl = yl .
180 K. Johannsen / On the validity of the Boussinesq approximation
Remark 3. The Elder problem has one symmetry of type (16). In this case pitchfork
l l
bifurcations may appear. They have been observed on the solutions branches Bb,2 , Bb,3 ,
l l
Bf,2 , and Bf,3 . Here, solution branches with broken symmetry intersect with the symmet-
ric solution branches. Hence steady state solutions with broken symmetry exist. Without
exception these solutions are unstable. A detailed discussion is beyond the scope of this
paper.
5 5
Remark 4. The unique solutions sb,0 and sf,0 can be found directly by solving equa-
5 5 5 5 5 5
tions (1) and (7), respectively. Solutions sb,1 , sb,2 , sb,3 and sf,1 , sf,2 , sf,3 can be found
5 5 5 5
by transient simulation. sb,1 , sb,2 and sf,1 , sf,2 frequently occur in the simulation of the
5 5
Elder problem. Solutions sb,3 and sf,3 can be found choosing appropriate initial condi-
tions. The unstable solutions ub,1 , ub,3 and u5f,1 , u5f,6 can be found using algorithm 1.
5 5
The remaining unstable solutions can be found using grids too coarse to represent the
bifurcation diagram correctly (l < 5) leading to a changement of its topology. Interpo-
lation between different grids together and path following can be exploited. Finally, the
solutions on grid level l = 6, 7, 8 can be found by interpolation.
Table 2
Rayleigh numbers of fold bifurcation points of the Elder problem on grid level 8, rounded to integers.
Bifurcation points creating stable solutions are set in boldface.
Branch Boussinesq approximation Full eq.
1 – –
2 67 66
3 151, 234, 231, 235, 231, (291, 291, )399, 396 151, 230, 229, 244, 240
4 387 300
5 142 139
6 299 298
K. Johannsen / On the validity of the Boussinesq approximation 181
points. In table 2, two of them are set in braces, because it remains unclear if they dis-
l l
appear using finer spatial discretization. Nevertheless, Bb,3 and Bf,3 show a qualitatively
l l l
different behavior. Furthermore, Bb,3 connects to solution ub,3 , whereas Bf,3 connects to
l l l
solution uf,6 . This change in topology can be observed for Bb,4 and Bf,4 , too. This leads
us to the main result of this paper, which we state in:
Theorem 1. For the bifurcation diagrams Sb and Sf of the Elder problem there is no
topological equivalence respecting the correspondence (17) and (18).
Sketch of proof. Let l ∈ [5, 8]. From figure 4 we see that branch Bb,3 l l
connects sb,3
l l l
with ub,3 . Assume Sb and Sf are topological equivalent. Then there is a homeomorphism
l ∗
T : Xl ×IR → Xl ×IR with T (Sbl ) = Sfl . Therefore Bf,3 := T (Bb,3l
) is a solutions branch
l l ∗ l
of Sf . Assume T respects the correspondences (17) and (18). Then Bf,3 connects sf,3
l ∗
with ulf,3 . From figure 5 we see that Bf,3 l l
connects sf,3 with ulf,6 . Therefore Bf,3 l
∩ Bf,3
l ∗
= ∅. Since Bf,3l
= Bf,3 , on both branches there is a pitchfork bifurcation. Due to (20)
l l ∗
the branches Bf,3 , Bf,3 are symmetric. This is a contradiction to remark 2. Hence the
assumed homeomorphism does not exist.
We assume, this argumentation carries over to the continuous case.
5. Discussion
We have investigated the Elder problem from the dynamical systems point of view.
Applying a path continuation method we could determine eleven steady state solutions
for the full system of equations as well as for the Boussinesq approximation. A corre-
spondence between the solutions of both approaches has been established. Parts of the
corresponding bifurcation diagrams have been investigated. We showed that with respect
to the stable steady states the Boussinesq approximation and the full set of equations be-
have similarly. Including the unstable solutions in our considerations, we observed a
significant difference between the bifurcation diagrams. It turned out that they are topo-
logically not equivalent. Existence and uniqueness of the transient solution has been
shown up to now only for the Boussinesq approximation. The non-equivalence of the
bifurcation diagrams indicates that essentially different mathematical techniques have
to be applied to show existence and uniqueness for the full set of equations. Whether
the results have any consequences for the transient solution process of the original Elder
problem remains unclear. Further investigations are necessary.
References
[1] E.L. Allgower and K. Georg, Numerical Continuation Methods: An Introduction (Springer, New
York/Berlin, 1990).
[2] J. Bear and Y. Bachmat, Introduction to Modeling of Transport Phenomena in Porous Media (Kluwer
Academic, Dordrecht, 1991).
182 K. Johannsen / On the validity of the Boussinesq approximation
[3] P. Clement, C.J. Van Duijn and L. Shuanhu, On a nonlinear elliptic–parabolic partial differential
equation system in a two-dimensional groundwater flow problem, SIAM J. Math. Anal. 23(4) (1992)
836–851.
[4] M. Dellnitz M. and B. Werner, Computational methods for bifurcation problems with symmetries –
with special attention to steady state and Hopf bifurcation points, J. Comput. Appl. Math. 26 (1989)
97–123.
[5] H.J.G. Diersch, FEFLOW finite element subsurface flow and transport simulation system – User’s
manual/Reference manual/White paper, Release 4.9 WASY, Berlin (2001).
[6] H.J.G. Diersch and O. Kolditz, Variable-density and transport in porous media: Approaches and chal-
lenges, Adv. Water Resour. 25(8–12) (2002) 899–944.
[7] J.W. Elder, Steady free convection in a porous medium heated from below, J. Fluid Mech. 27 (1967)
29–48.
[8] J.W. Elder, Transient convection in a porous medium, J. Fluid Mech. 27 (1967) 609–623.
[9] P. Fabrie and M. Langlais, Mathematical analysis of miscible displacement in porous medium, SIAM
J. Math. Anal. 23(6) (1992) 1375–1392.
[10] E. Fein, ed., d3f – Ein Programmpaket zur Modellierung von Dichtestroemungen (GRS, Braun-
schweig, 1998).
[11] P. Frolkovic, Consistent velocity approximation for density driven flow and transport, in: Advanced
Computational Methods in Engineering, Part 2, Maastrich, eds. R. Van Keer et al. (Shaker Publishing,
1998) pp. 603–611.
[12] P. Frolkovic and H. De Schepper, Numerical modeling of convection dominated transport with density
driven flow in porous media, Adv. Water Resour. 24(1) (2001) 63–72.
[13] A.W. Herbert, C.P. Jackson and D.A. Lever, Coupled groundwater flow and solute transport with fluid
density strongly dependent on concentration, Water Resour. Res. 24 (1988) 1781–1795.
[14] E. Holzbecher, Modeling Density-Driven Flow in Porous Media (Springer, Berlin, 1998).
[15] K. Johannsen, The Elder problem – Bifurcations and steady state solutions, in: eds. S.M. Has-
sanizadeh et al., Comput. Methods Water Resour. 47(1) (2002) 485–492.
[16] K. Johannsen, W. Kinzelbach, S.E. Oswald and G. Wittum, The saltpool benchmark problem – nu-
merical simulation of saltwater upconing in a porous medium, Adv. Water Resour. 25 (2002) 335–348.
[17] Y.A. Kuznetsov, Elements of Applied Bifurcation Theory (Springer, Berlin, 1995).
[18] A. Leijnse, Three-dimensional modeling of coupled flow and transport in porous media, Ph.D. thesis,
University of Notre Dame, Indiana, USA (1992).
[19] Z. Mei, Bifurcation Problems and Numerical Analysis. An Introduction (Xi’an Jiaotong University,
1993).
[20] C.I. Voss and W.R. Souza, Variable density flow and solute transport simulation of regional aquifers
containing a narrow freshwater–saltwater transition zoner, Water Resour. Res. 26 (1987) 2097–2106.