Kol Ditz 1998
Kol Ditz 1998
Kol Ditz 1998
21-46, 1998
0 1997 Elsevier Science Ltd
All rights reserved. Printed in Great Britain
PII: SO309-1708(96)00034-6 0309-1708/98/$17.00+0.00
ELSEVIER
This work examines variable density flow and corresponding solute transport in
groundwater systems. Fluid dynamics of salty solutions with significant density
variations are of increasing interest in many problems of subsurface hydrology.
The mathematical model comprises a set of non-linear, coupled, partial
differential equations to be solved for pressure/hydraulic head and mass
fraction/concentration of the solute component. The governing equations and
underlying assumptions are developed and discussed. The equation of solute mass
conservation is formulated in terms of mass fraction and mass concentration.
Di:fferent levels of the approximation of density variations in the mass balance
equations are used for convection problems (e.g. the Boussinesq approximation
and its extension, fully density approximation). The impact of these simplifica-
tions is studied by use of numerical modelling.
Numerical models for nonlinear problems, such as density-driven convection,
must be carefully verified in a particular series of tests. Standard benchmarks for
proving variable density flow models are the Henry, Elder, and salt dome
(HYDROCOIN level 1 case 5) problems. We studied these benchmarks using two
finite element simulators - ROCKFLOW, which was developed at the Institute
of Fluid Mechanics and Computer Applications in Civil Engineering and
FE:FLOW, which was developed at the Institute for Water Resources Planning
and Systems Research Ltd. Although both simulators are based on the Galerkin
finite element method, they differ in many approximation details such as temporal
discretization (Crank-Nicolson vs predictor-corrector schemes), spatial dis-
cretization (triangular and quadrilateral elements), finite element basis functions
(linear, bilinear, biquadratic), iteration schemes (Newton, Picard) and solvers
(direct, iterative). The numerical analysis illustrates discretization effects and
defects arising from the different levels of the density of approximation. We
contribute new results for the salt dome problem, for which inconsistent findings
exist in literature. Applications of the verified numerical models to more complex
problems, such as thermohaline and three-dimensional convection systems, will
be presented in the second part of this paper. 0 1997 Elsevier Science Ltd. All
rights reserved
Il”IIl
D spatial numerical dispersion resulting from Qld source terms of the solute component in
spatial discretization [m2 s-t] terms of mass fraction [kg me3 s-‘1
num
D temporal numerical dispersion resulting from Ras Rayleigh number of solute HI
temporal discretization [m2 s-l] Raeff effective Rayleigh number of solute [l]
6 = rDmd + c+$ + (aL - cq) 4 = Dap s contour of A [m*-’ 1
tensor of hydro d”ynamic dispersion specific storativity of a porous medium
[m2 s-l] with respect to pressure changes [Pa-‘]
Dm coefficient of molecular diffusion specific storativity of a porous medium
[m2 s-‘1 with respect to hydraulic head changes
density of the extensive quantity E [m-7
rate of internal production of an t + 8,At time collocation point bl
extensive quantity E At, time-step increment at time level n [s]
fl,h known functions of position and time V characteristic value of the macroscopic
g vector of gravity acceleration [m se21 velocity [m s-l]
h=p/gp”-z hydraulic head related to pw V macroscopic velocity [m s-l]
height of the aquifer ;:; V” microscopic velocity of a continuum
nodal source term of fluid mass element [m SC’]
first nodal source term of solute mass yE velocity of particles belonging to an E-
second nodal source term of solute continuum [m s-‘1
mass VS microscopic average velocity of the
i microscopic diffusive flux relative to the solute component [m s-t]
advective one Fg m -2 s-‘I VW macroscopic fluid velocity with respect
1
‘*
microscopic diffusive mass flux relative to a fixed coordinate system [m s-l]
to the mass-weighted velocity v*, v’ microscopic mass- and volume-
[kg m -2 s-‘I weighted velocities of the fluid [m s-‘1
.I
J microscopic diffusive mass flux relative AZ vertical element size [ml.
to the volume-weighted velocity
[kg m -2 s-7 Greek symbols
J* macroscopic dispersive mass flux in (Y upwind parameter
terms of mass fraction [kg mV2 SC’] (yL, aT longitudinal and transverse disper-
J’ macroscopic dispersive mass flux in terms sivities [ml
of mass concentration [kg m -2 s-7 coefficients of compressibility of the
tensor of permeability of a porous porous medium due to fluid pressure or
medium b21 hydraulic head variations, the compres-
tensor of hydraulic conductivity of a sibility coefficient of the porous medium
porous medium [m s-‘1 is related to the rock compressibility by
Al characteristic element length [ml the void ratio op = n/( 1 - n)/&
L,(P)
characteristic length of density varia- Pa-‘I, WI
tions [ml coefficient of expansivity resulting from
n porosity [m3 mp3] the change of the mass concentration of
I’ outward unit vector PI solute at constant pressure [m3 kg-‘]
N finite element basic and shape function coefficient of compressibility of the fluid
HI resulting from the change of the
P dynamic pressure Pal hydraulic head at constant mass frac-
AP, nodal pressure difference Pal tion of the solute Cd
Pe Peclet number 111 coefficient of compressibility of the fluid
pg grid Peclet number HI resulting from the change of the fluid
q specific flux vector (specific discharge or pressure at constant mass fraction of
mean Darcy velocity of the bulk fluid the solute Pa-‘1
flow) [m s-‘1 coefficient of rock compressibility[Pa-‘1
QC source terms of the solute component in coefficient of expansivity resulting from
terms of mass concentration the change of the mass fraction of
[kg m -3 s-‘I solute at constant pressure [kgW3kg31
PQP source term of the fluid mass Y diffusion impedance factor due to tor-
[kg m -3 s-‘I tuosity VI
PjQp' nodal source term of the fluid mass at r boundary
the time collocation point [kg rnp3 SC’] 8 Kronecker-delta (unity tensor) PI
Coupled groundwater jlow and transport 29
weighting factor of the time collocation Groundwater density may vary slightly as a result of
point VI either pressure or small temperature variations. However,
dynamic viscosity of the fluid (bulk the density can alter considerably in geothermal areas or
fluid viscosity) [Pa s-l] concentrated brine pools. In most of such situations
microscopic mass density [kg mp31 groundwater dynamics are strongly influenced by
ew, es microscopic mass densities of water and destabilizing density differences producing convective
solute components [kg mm31 currents. Density gradients can introduce gravitational
P macroscopic mass density of the fluid instabilities that give rise to recirculating groundwater
(bulk fluid density) [kg mm31 systems. Concentrated salt solutions occur, e.g. near salt
P”9PS macroscopic mass densities of water domes41 in deep crystalline rocks, and in bedded-salt
and concentrated salt solution, respec- formations. Commonly, saturated brine densities are
tively (mass of component per unit of higher than 1200 kg rnp3.
volume of the fluid) [kg mp31 Usually, the coupling between flow and transport in
W mass fraction of concentrated salt solu- groundwater systems is reduced to advective and
tion (solute) [kg3 K31 velocity-dependent hydrodynamic dispersion mechan-
Aw characteristic difference of mass frac- isms. This approximation holds for low concentration
tion of solute [kg3 K31. situations if all material properties such as density and
viscosity are nearly constant and the overall fluid
Special symbols
motion is independent of the motion of individual
A difference species. Groundwater-brine systems are more strongly
V Nabla operator coupled by the buoyancy term of the momentum
V2 Laplace operator. equation (i.e. generalized Darcy’s law). Furthermore,
Subscripts osmotic effects may become important under certain
i,.i, k 1 nodal values circumstances.17 For high concentration situations,
ff,P spatial coordinates Hassanizadeh16 developed a more general form of
0 reference values. Darcy’s and Fick’s laws, which contribute to these
cross-coupling effects. Recently, Voge149 reported on
artefacts arising from inconsistent formulations of the
2 INTRODUCTION balance equations with respect to velocities of fluid mass
or fluid volume.
2.1 Background The reasons for hydrodynamic dispersion in porous
media are twofold. First, kinematic effects such as
A number of environmentally important problems streamline dividing result from the structure of a porous
require the analysis of brine or seawater dynamics in medium, whereas dynamic effects result from fluid
subsurface systems, e.g. disposal of hazardous (toxic and velocity fluctuations. Interaction between dispersion
radioactive) waste in crystalline or salt rock formations, and convection phenomena is quite complex. On the
infiltration of leachates from landfills and industrial one hand, buoyancy effects produce additional fluid
waste disposals, and saltwater intrusion in exploited motion and, therefore, support macroscopic hydro-
coastal aquifers or aquifers overlying salt formations. In dynamic dispersion, which depends on the microscopic
the past, numerous coides have been developed to velocity field. On the other hand, dispersion causes
simulate groundwater systems including density effects additional mixing and, therefore, reduces the potential
(Table 1). for convective currents, i.e. density differences become
smoothed. Recently, Hassanizadeh and Leijnse18 pre- et al.” introduced a more general drag term for flow
sented a nonlinear theory of high concentration- through a porous medium containing a dispersive
gradient dispersion in porous media, which is supported contribution to the viscous force due to mass fraction
by their theoretical and experimental studies of hydro- gradients. They showed that this additional term can be
dynamically stable flow regimes in vertical columns. neglected, if the salt concentration does not change
appreciably over the dispersion length. Hassanizadeh
2.2 Prior work concerning variable density flow and and Leijnse17 considered extensions of Darcy’s and
transport Fick’s laws in order to account for coupling effects. By
means of numerical experiments, the potential signifi-
In the laboratory, density effects have been frequently cance, e.g. of osmotic processes is illustrated. Further-
studied in columns and cavities. Wooding” performed more, they discussed the issue of compatibility between
experiments in saturated soil columns with aqueous balance equations and corresponding boundary condi-
solutions denser than water. Elders,” experimental tions for variable density problems. Senger and Fogg,46
studies concerning thermal free convection in a Hele- and Evans and Raffensperger,12 formulated variable
Shaw cell became famous and are frequently employed for density flow in terms of stream functions and equivalent
the verification of numerical codes. The Hele-Shaw freshwater heads. They found that stream functions
analogy is based on the similarity of flow between should be defined in terms of mass flux. Otherwise, the
two plane walls to this in a porous medium having an Boussinesq approximation is involved implicitly. Schin-
equivalent permeability equal to (gap of the ce11)2/12. cariol et a1.43 studied mixed convective flows observed
The Hele-Shaw analogy is constrained to situations during their flow container experiments.42 They con-
where advection and rapid diffusion of vorticity are cluded that the classic Rayleigh number criterion alone
negligible. Schincariol and Schwartz42 and Oostrom et is inappropriate to evaluate the stability of mixed
a1.36 conducted experiments in flow containers to study convective flow systems. It was found that initial
mixed convective motions.They observed complex con- perturbations can grow or decay in dependence on the
centration distributions that develop as lobes of dense wavelength of the perturbing function (e.g. harmonic
fluid moving downward from the plume, while less dense oscillations of the brine boundary condition). Recently,
fluid migrates upward into the plume. Field evidence of Oldenburg and Pruess35 studied the Henry problem as
convective phenomena is complicated. There exist some well as the Elder problem to verify the finite difference
investigations on plume behaviour under landfills (e.g. simulator TOUGH2. Additionally, they investigated the
van der Molen and van Ommen48). salt dome problem in detail and determined new results
The modelling of variable density flow has been which will be discussed and compared in Section 5.3.
addressed by numerous authors. In the past, a number In this paper we present two numerical approaches
of codes were developed to simulate density-dependent for the simulation of variable density flow and
groundwater systems. Table 1 summarizes some of these transport in groundwater systems: ROCKFLOW-
simulators. Internal natural convections in closed cells DM2, developed at the Institute of Fluid Mechanics
were treated already by Horton and Rogers22 and and Computer Applications in Civil Engineering; and
Lapwood who examined the linear stability providing FEFLOW (Release 4.9, developed at the Institute for
a pertubation analysis. They found that instabilities Water Resources Planning and Systems Research Ltd.
appear if a certain Rayleigh number of 4n2 is exceeded. Three benchmark tests are studied: the Henry problem;
The Rayleigh number can be physically interpreted as the Elder problem; and the salt dome problem
the ratio of the rate of thermal energy release to the rate (HYDROCOIN level 1 case 5). The paper will be
of viscous dissipation of energy (Nield and Bejan34). continued in a second part, in which thermohaline and
First numerical models were introduced, e.g. by Desai three-dimensional variable density flow and transport
and Contractor,’ Elder,” Frind,13 Pinder and Cooper3’ will be considered.
and Segol et aI.& Diersch6 presented a finite element
model based on primitive variable formulations. He
used the Elder problem for verifying the numerical 3 BASIC EQUATIONS
model. Voss and Souza” and Herbert et al.20 discussed
the problem of consistent approximation of terms In this section we will summarize the governing
involved in fluid velocity calculations. Herbert et al.20 equations of variable density groundwater flow and
preferred mixed finite element interpolations and com- solute transport, which are described in detail, e.g. by
puted continuous velocity distributions by a global Bear and Bachmat, Hassanizadeh and Gray” and
smoothing technique for velocity components. They also Nield and Bejan.
discussed the question of how the viscous force between
the matrix skeleton and a multicomponent fluid has to General balance equation
be modified if there are significantly different densities The general form of the microscopic differential balance
of the fluid mixture components. Moreover, Herbert equation of an extensive thermodynamic property (e.g.
Coupled groundwater flow and transport 31
mass, linear momentum, energy) is given by above microscopic relationships to a macroscopic level.
To this end, averaging rules are employed, which
g + V.(evE) = uE. correspond to a representative elementary volume
(REV).2>‘5 The following discussion is restricted to the
Equation (1) is a classical balance law of continuum macroscopic level. Therefore, no special symbols will be
mechanics expressing the conservation principle of any used to indicate the quantities are macroscopic ones.
thermodynamic property. Considering a continuum
consisting of a number OFparticles: the particle velocity Assumptions
must be replaced by a representative velocity of the A number of assumptions, common for groundwater
continuum, which will be a certain average velocity of a flow in porous media, will be introduced:
cloud of particles (e.g mass- or volume-weighted
velocities). Now, the tota. flux of an extensive quantity (1) The inertia and convective acceleration terms can
E may be divided into two parts: an advective flux be neglected (Stokes flow).
corresponding to an ave:rage velocity; and a diffusive (2) The viscous drag of the fluid on the skeleton
flux relative to this average velocity surface is assumed to be proportional to the velocity,
evE = eve + e(vE - v”) = eve + j. and stress is simply a pressure term.
(2)
(3) Macroscopic dispersive mass fluxes in porous
Substituting eqn (2) into eqn (1) the balance law takes media may be written in analogy to Fick’s law (Table 3).
the general form (4) Moreover, it is assumed that the advective flux by
the bulk fluid flow is much larger than the sum of
$ +V.(U’) + 0-j = dE. diffusive and dispersive ones, i.e. effects of solute
dispersion to the mass flux of the whole fluid mixture
are negligible.
D@usive fluxes
Diffusive mass fluxes can be expressed by Fick’s law. The IPVlz+ WPI (5)
use of a certain formulation of Fick’s law in terms of mass
Equation (5) can be derived from a dimensional analysis
fraction (Table 2(a)) or mass concentration (Table 2(b))
of the balance equation of the fluid mass if the following
implies that diffusive mas’s fluxes are defined relative to
two conditions are introduced:2 (i) the characteristic
mass- or volume-weighted velocities. Subscript s is used to
length of density variations is much larger than the
denote the solute component of the fluid. Quantities
maximum value of dispersivity L(p) > aL; and (ii) the
without subscripts are related to the fluid.
diffusive mass flux is much smaller than the advective
The above equations are written for a binary fluid
one, which is expressed by large values of the Peclet
consisting of two components, water and concentrated
number Pe = (L@‘v)/D, > 1. The assumption (5)
salt solution (solute). The mass- and volume-weighted
implies that differences between volume-weighted and
velocities are related by the following expression:2
mass-weighted velocities vanish.
Linearized forms
p = pW[l + Pn(h - ho) + PCC
or
P dn dn=;dh= (1 -n)
n$+V.(npv)+- -= PQ,, (6b)
l-n dt
where vw is the fluid velocity with respect to a fixed
(12)
coordinate system, while v is the fluid velocity relative to Employing the equations of state for the bulk fluid
the solid; d/dt denotes the material derivative when density (Table 4) and the porosity, eqn (11) or eqn (12),
moving with the solid phase. the mass balance equation of the fluid phase can be
Depending on which formulation for dispersive fluxes written as
is chosen (Table 3), a corresponding equation of solute
mass conservation can be derived, written either in terms ~Sfl g + V.(pnv) = pQp - pnP, $
of mass fractions
S: = n& + cxP (13)
% + v(nvpw) - V(npbVw) = Q, (7) in terms of pressure and mass fraction and
or in terms of mass concentrations
dnC Sk g + V.(nv) = QP - n& g - nv&VC)
dt + V+rvC) - V@VC) = Qc. (8)
The momentum balance equation for variable-density so=nph + ah (14)
fluid flow in a porous medium leads to the generalized in terms of hydraulic head and mass concentration,
form of Darcy’s law where Sopand Sk represent the specific storativity of the
^
porous medium with respect to pressure or hydraulic
q=nv=- k.(Vp-/& = -k.
CL (Vh_!?_gE
I P g>
head changes, respectively. Both eqns (13) and (14) are
equivalent. Note, eqn (14) is fully developed by using
(9) the equation of state for the bulk fluid density, Table
where h = p/gp” - z is the hydraulic head related to the 4(b). The formulations of the mass balance equations
mass density of water p”. of the fluid phase given in eqns (13) and (14) are
The dimensional analysis of the momentum balance implemented in the simulators ROCKFLOW and
equation for fluid flow provides characteristic numbers FEFLOW, respectively.
that characterize the flow regime, e.g. the Rayleigh
number of solute
3.2 Boussinesq approximation of the mass balance of the
KHP, Aw KH& AC
Ras= nD = nD (10) fluid
m m
where AC and Aw represents characteristic differences Commonly, the Boussinesq approximation is introduced
of mass concentration and mass fraction. which neglects density variations within the mass
balance equation of the fluid phases. Density variations
Equations of state
are included by the buoyancy term of the Darcy equa-
The functional dependencies of the density on the variables
tion only. Accordingly, the equation of mass conserva-
of state are derived by using a total differential representa-
tion of the fluid phase eqn (6) takes the simplified forms
tion. For isothermal situations, the equations of state of the
bulk fluid density and their linearized forms are given in
Sig + V.(nv) = QP (15)
Table 4. An equation of state for the bulk fluid viscosity of
saline groundwater is e.g. presented by Lever and Jackson.32
in terms of pressure and
It is assumed that changes of the porosity result from
the compressibility of the skeleton of the porous
medium due to fluid pressure variations.
Sk E + V.(nv) = Qp (16)
in terms of hydraulic head. For the flow of a non-
dn=$dp=(l-n) deformable medium this simplification substitutes the
conservation of mass by the conservation of fluid
(11) volume. Analyzing eqn (6), it can be easily shown that
Coupled groundwater $00~ and transport 33
this approach holds only if the streamlines follow the balance equation of solute mass are used either in terms
density-isolines, i.e. for barocline flow. of mass fraction or in terms of mass concentration. These
The Boussinesq approximation is inappropriate, if simplifications result in different approximation levels of
significant density changes arise from solute concen- density variations in the balance equations for fluid and
tration variations. For this case, the contours of solute masses as summarized in Table 5. The impact of
equipotential pressures are not orthogonal to the these different approximation levels will be considered in
velocity vectors. Evans and Raffensperger12 studied the the present study, especially in the discussion of the Elder
limitation of the Boussinesq approximation for a problem of transient free convection.
problem which is similar to the Elder problem. They
found significant differences in the salinity distributions
up to 9% by comparing results of both formulations of 4 FINITE ELEMENT MODELS
the mass balance of the fluid phase by eqns (6) and (15).
Two finite element simulators, ROCKFLOW38’52 and
3.3 Different formulations of the balance equation of FEFLOW,* were developed to model variable density
solute mass flow and transport in porous media. Both simulators
differ in several points concerning the formulations of
By use of the continuity eqn (6), the divergent form of the basic equations as well as the numerical methods.
the equation of solute mas:s conservation, eqn (7), can be While ROCKFLOW is based on balance equations in
transformed to a convective formulation terms of pressure and mass fraction, eqns (9), (13) and (17),
FEFLOW use the adequate formulations in terms of
n$ + np”*Vw - V+z@Vw) + wpQp = Q,. (17) hydraulicheadandmassconcentration,eqns(9),(14)and(8).
To develop the finite element model, all variables are
Developing the equation of solute mass conservation in discretized as piece by piece continuous functions using
terms of mass concentration, eqn (8) we obtain the either linear basis functions on triangular elements, or
following relation bilinear and biquadratic basis functions on quadrilateral
elements in two-dimensional problems. To keep numer-
rr$ + nV.VC - V$nij.VC) + CQP = Qc + C ical dispersion as low as possible, no upwinding is used.
In general, time derivatives are evaluated by finite
dh #DC
(1 - nbh -dt + n/k- dt + m(p,vc)
.
(18) difference schemes. Both simulators provide some
specialities concerning the temporal approximation.
In particular, ROCKFLOW employs the explicit
If density and porosity variations are neglected for the
Euler-Taylor-Galerkin (ETG) method for the discre-
balance equation of the solute mass conservation, eqn
tization of the advection-diffusion equation.’ The basic
(18) can be written in a simplified form
idea is to expand the mass concentration in Taylor series
n !$ + nv*VC - V.( nD.VC) + CQ, = Qc. with respect to time. Time derivatives of up to third
(19)
order can be expressed by space derivatives obtained by
the differentiation of the pure advection equation, which
Levels of the density approximation normally is the main cause of severe numerical diffusion.
As mentioned above, the Boussinesq approximation was Diffusion, dispersion and source terms are added
employed by some authors in order to simplify the afterwards. The resulting differential equation, first
mathematical model of variable density flow, i.e. to derive order in time and second order in space, is handled
semi-analytical solutions” or to introduce stream with Galerkin’s method in conjunction with the Green-
functions.” Moreover, different formulations of the Gauss theorem.
34 0. Kolditz et al.
In combination with Euler time-stepping, the ETG change the physical conditions actually simulated by the
scheme leads to a truncation error of fourth order in time. numerical model. Especially for the investigations of
The element matrices become symmetric and diagonally convection problems, therefore, appropriate schemes
dominant. Highly efficient methods, e.g. PCG (precon- should be preferred to keep the numerical dispersion as
ditioned conjugate gradients) exist for solving this type of low as possible.
equation. The flow equations are set up with pressure as
single degree of freedom per node. The coefficients of the 5 BENCHMARKEXAMPLES
global matrix then have exactly the same pattern of
storage, no index tables have to be recalculated. Crank- The usual verification procedure by use of analytical
Nicholson time-stepping avoids fourth order error terms solutions’4 is complicated due to the nonlinear nature of
with the disadvantage of non-symmetric matrices where variable density flow and transport. In general, there are
standard PCG cannot be applied.38 no exact solutions for this type of problem for
For stability of this higher order formulation, time-step comparison, excepting a semi-analytical solution by
size limitation according to the Courant criterion is Henry” for a steady-state case. Therefore, a cross-
generally required.27 The iterative coupling between the verification of results obtained by different codes is
discretized flow and transport equations is realized by a required to test numerical models. A particular series of
fixpoint (Picard) iteration scheme, i.e. the actual densities benchmarks will be carefully analyzed by use of the
of a new iteration step are calculated from the field simulators ROCKFLOW and FEFLOW to show that
variables (pressure and salt concentration) of the previous the numerical models accurately represent the physics
iteration step. Details of the approximate model imple- implied by the governing differential equations and,
mented in ROCKFLOW are given in the Appendix. finally, to build confidence in the numerical modeling of
For the detailed description of the general purpose convection phenomena.
simulator FEFLOW, we refer to Diersch’ and to the We present results for three benchmark problems:
second part of this paper. (5.1) the classic seawater intrusion problem of Henry,”
termed the Henry problem; (5.2) the free convection
Truncation errors and resulting efective Rayleigh numbers problem by Elder” (fingering problem), called the Elder
The spatial and temporal discretization can introduce problem; and (5.3) the salt dome problem, which is the
spurious dispersion effects where the amount of test case 5 of level 1, that was proposed by the
(physical) hydrodynamic dispersion is enlarged by the international HYDROCOIN47 project for the verifica-
numerical dispersion tion of groundwater models. In addition to density
Jeff = Dphys+ ~num. variations, both simulators ROCKFLOW and
(20)
FEFLOW can also account for viscosity variations.
The numerical dispersion results from the spatial and The considered standard benchmarks are stated for
temporal discretization. To estimate the actual disper- variable density, but constant viscosity flow. Therefore,
sion effective in the numerical approach, the truncation to make our results comparable with prior works, we
errors are determined keep the viscosity constant through our simulations.
Il”lll Since viscosity variations can gain importance in brine
D spatia,M + + O(A1*), systems, the standard benchmarks for saltwater intru-
sion should be extended to variable-viscosity flow in
Il”IIl
D temporalM (0, - ;)v*Al, + O(At;) (21) forthcoming cross-verification tests, accordingly.
where AZ is the characteristic element length, At,, is the 5.1 The Henry problem
time-step increment at time level n, (I: is the upwind
parameter, and 8, is the weighting factor of the time Definition of the problem
collocation point. The Henry problem describes the advance of a saltwater
Now, the numerical solutions can be characterized by front in a confined aquifer which was initially saturated
an effective Rayleigh number with uncontaminated fresh water. Henry” developed a
semi-analytical solution technique for this problem.
Ra:’ =
Ras
(22) Based on the Boussinesq approximation he derived
1 +TPg+(Bt -4)CrPg’
analytical expressions for the stream function and the
The upwind parameter is zero for the Galerkin finite salt concentration in the form of Fourier series. The
element method and unity for the streamline upwind resulting algebraic equations for determining the coeffi-
method. The weighting factor of the time collocation cients of the Fourier series must be solved by numerical
point is 0.5 for the Crank-Nicolson as well as second techniques.
order Adams-Bashforth/Trapezoid-Rule schemes and is Using quite different approximation methods, a
unity for fully implicit schemes. It is obvious that number of authors obtained similar results.415~‘3128~35~37,44,50
especially upwind and fully implicit techniques can However, the ‘mystery’ of Henry’s solution is, that no
Coupled groundwater flow and transport 35
(4 ~ 300m
h=O
4 /
7E
-1
x
c=o
600m
(a)
Fig. 4. Elder problem. Effect of spatial discretization on the computed salinity evolution at 4, 10, 15 and 20 years simulation time;
positions of the 20 and 60% isochlors: (a) conrse mesh (1170 grid points, 1100 finite elements), similar to the discretization used by
Elder” and Voss and Son~a;~’ (b) fine mesh (4539 grid points, 4400 bilinear finite elements); (c) very fine mesh (10 108 grid points,
9900 finite elements).
However, we observe strong discretization effects for A mesh analysis is provided by use of the grid Peclet
these coarse meshes. .4t first, we investigated the number, which is defined as the ratio of (grid) advective
influence of different temporal discretization schemes, to (grid) dispersive fluxes. The vertical grid Peclet
where time marching schemes of different order of number can be written by
accuracy are applied: (i) a fully implicit, and Crank’-
Nicholson schemes in c’ombination with the Newton-
IK(+&AC)l AZ
Raphson method; and (ii) one-step predictor-corrector
Pg, =~Azz
methods of first- and second-order accuracy embedded nD, n&
in a full Newton method. The discrepancies in the results
KAh KH&AC
indicate that a convergent approximate solution has not = -- -AZ
n& nD, H
yet been reached.25
Therefore, we consecutively refine the meshes until kgpwAh
~- Ra * (23)
grid convergence is achieved (Fig. 4). To this end, a = pnD, SH
spatial discretization consisting of at least 4400 bilinear
finite elements (Fig. 4(b), 88 elements in horizontal and where AC is the maximum difference of mass concen-
50 elements in vertical d:irection) will be required. Note, tration, Ah is the maximum difference of hydraulic
the number of convection cells is doubled for the fine head, and AZ is the characteristic vertical element
mesh calculations and, :furthermore, the resulting con- length.
vection patterns are quite different. There is a central Note, a relative large diffusivity is used instead of
upward flow rather th,an a central downward flow, hydrodynamic dispersion for the Elder problem. If the
resulting from the coarse mesh calculations. Finally, our vertical velocity is estimated by the buoyancy term only,
high-resolution result obtained with both simulators a maximum grid Peclet number can be determined by
(Figs 4(c) and 6) compare very well with those given by use of the Rayleigh number, which is equal to 400 for
Oldenburg and Pruess.35 the Elder problem. The maximum difference of the
38 0. Kolditz et al.
m
(4 (W
Fig. 5. Elder problem. Computed salinity distributions (20% and 60% isochlors) at 2,4, 10 and 20 years simulated time) from left to
right) - different levels of the density approximation: (a) level 1; (b) level 2; (c) level 3 (FEFLOW simulation).
calculated hydraulic head is equal to 15 m. Therefore, mass of solute are neglected. In this case, the set of mass
the following estimation of the vertical grid Peclet balance equations is given by eqns (16) and (19).
number is given for the Elder problem Level 2. The extended Boussinesq approximation
accounts for density variations in the balance equation
Pg; M 200$. of the fluid mass, according to eqn (14).
Level 3. The full density approximation accounts for
Now, the corresponding grid Peclet numbers of the density variations in all balance equations, according to
several meshes above can be characterized by the eqns (13) and (17) in terms of mass fractions and
vertical discretization AZ/H and are equal to Pgz = 8 according to eqns (14) and (8) in terms of mass
for the coarse mesh (Fig. 4(a)), Pg, = 4 for the fine mesh concentrations.
(Fig. 4(b)), and Pg_ = 2.5 for the very fine mesh (Fig.
4(c)). In fact, these grid Peclet numbers are still larger In order to evaluate the impact of the Boussinesq
than the theoretical stability criterion of linear advec- approximation, we compare Fig. 5(a) with (b) showing
tion-diffusion problems (e.g. Pg 5 2 for bilinear finite the salinity evolutions corresponding to the level 1 and
elements23). The grid convergence test (Fig. 4) may be level 2 approximations. Differences are apparent at early
the best proof of stability of the solution of this non- stages, when large concentration contrasts are involved.
linear problem, where no exact discretization criteria Up to 10 years, considerable discrepancies in the solute
exist. This study reveals, that a grid convergent distributions are found. Concerning the long-term
numerical solution can be obtained for grid Peclet behaviour, where density differences are already
numbers smaller than 5 for the Elder problem. smoothed, these deviations decrease again. Finally, the
concentration distributions become nearly identical at
Levels of the density approximation 20 years.
According to Table 5, we distinguish three different Regarding the density variations in the equation des-
levels of the density approximation: cribing conservation of the mass of solute, the results of
the level 2 simulation and the full density approximation
Level 1. Usually, the Boussinesq approximation of level 3 are analyzed. If comparing the computed
implies that density effects are considered only in the salinity evolutions in Fig. 5(b) and (c), we find the
buoyancy term of the momentum equation. Density isohalines propagate more rapidly into the depth of the
variations in the balance equations of mass of fluid and cavity for the full density approximation, i.e. while
Coupled groundwater flow and transport 39
(4 h=20.456 m
Salt Dome
Fig. 7. Definition of the salt dome problem (HYDROCOIN level 1 case 5) - domain and boundary conditions: (a) hydraulic head;
(b) streamline; (c) velocity and salinity distributions.
Velocities are computed continuously by applying a stepping procedure for the numerical simulation of
separate finite element scheme for the velocity function. hydrothermal convections in thermoelastic fractures.
Furthermore, Herbert et ~1.~’employed a steady-state In contrast to Herbert et al.,” we calculate the fully
strategy to obtain the long-term solution. They used transient evolution of the convection system. Steady-
parameter stepping for the diffusivities and dispersiv- state solutions were reached after about 200 years total
ities. However, such parameter increments must be small simulation time. We used two different techniques to
enough to guarantee satisfactor convergence by the find the steady-state solution. First, the ROCKFLOW
Newton-Raphson iteration scheme. Kolditz and simulations were performed with comparatively small
Diersch” applied successfully a similar parameter time-steps (governed by the Courant criterion in the
Coupled groundwater $0~ and transport 41
Table 8. Simulation parameters for the salt dome problem (HYDROCOIN level 1 case 5)
explicit approach). The comparison of the 200 with the molecular diffusivity (Fig. 8(b)). We computed the
400 years results, where no more changes of the salinity models for both values of molecular diffusivity (Figs
distributions were detected, indicated that steady-state 7(c) and 8(c)) and found nearly identical salinity
was reached. Secondly, the FEFLOW simulations were distributions. Oldenburg and Pruess35 calculated a
conducted with an automatic time-step control based on quite different salinity distribution. They called their
a predictor-corrector method (second order Adams- results a ‘fully swept-forward’ pattern (Fig. 8(b)). There
Bashforth/trapezoid scheme). After about 100 years, the is a substantial difference in the number of convection
time-steps grew very rapidly, which means that only cells. This swept-forward pattern covers only one
very small changes in pressure and salinity distributions instead of two recirculating cells near the aquifer
occur. We stopped the simulation at time of about 200 bottom (Fig. 7(b)). Oldenburg and Pruess35 argued
years. Oldenburg and Pruess3’ achieved steady-state that for large dispersive fluxes (D > 2 x 10e7 m2 s-i) the
solutions after about 100 years (starting from a fresh amount of brine rising from the bottom source is
water pool as initial condition). sufficient to cause a clockwise recirculation against the
overall counterclockwise flow imposed by the pressure
difference along the aquifer top. In contrast, the
Results and discussion
buoyancy forces are insufficient to overwhelm the
In past, the salt dome problem was analyzed for
pressure-driven flow for small dispersive fluxes (i.e.
different values of molecular diffusivity and mechanical
D < lop7 m2 s-l). As a consequence, the brine is swept
dispersivity (Table 9). Kolditz25 studied different
forward from left to right with the overall flow.
diffusive/dispersive models of the salt dome problem
Kolditz25 obtained similar swept-forward regimes for
and compared the results with prior findings presented
the case of lower Rayleigh numbers, i.e. if buoyancy
by Herbert et aL2’ and Oldenburg and Pruess.35 In
effects are less significant.
general, our results agree with those of Herbert et aL2’
Different levels of the approximation of the density
rather than with those by Oldenburg and Pruess.35
variations are employed for the analysis of the salt dome
Recently, Rivera and Jones4 confirmed the prior
problem, i.e. level 2 (Fig. 8(c)) and level 3 (Fig. 8(d)).
NAMMU results by Herbert et aL2’ using the new
The good agreement of the simulation results indicates
6.2 version of this simulator.
that the different levels of density approximations have
The most interesting c,ases for that inconsistent results
no influence on the steady-state salinity distribution for
exist will be discussed here (Table 9, emphasized values).
this case.
Figure 8 shows the computed steady-state salinity
distributions for the case of Q.,, = 1.39 x 10-8m2s-1
and ‘Ye= 20 m, d!r = 2 m. Note, the results by Herbert 6 CONCLUSIONS
et aL2’ correspond to values of D, = 1.39 x lo-* m2 s-i
(Fig. 8(a)), whereas Oldenburg and Pruess35 used a zero This work examines the modelling of variable density
Table 9. Values of molecular dilkivlty and dlspersivities for tbe salt dome problem
Parameters + Molecular Longitudinal Transverse
1 Model diffusivity D, dispersivity oL dispersivity or
Diffusive 5 x 10e6 m2 s-’ 0 0
5 x lo-’ m2 s-’ 0 0
5 x low8 m2 SK’ 0 0
Diffusrve-dispersive 5 x lo-’ m2 s-l 20m 2m
1.39 x lo-’ m* s-l 20m 2m
Dispemive 0 20m 2m
42 0. Kolditz et al.
(a)
1
-100
.\--.---
..-..2----
.\------
...~-N----
..w--w---d
-200
,\..~5~------d
\ \.~w----
. . ..I--___
-300
0 300 600 900
Fig. 8. Results for the salt dome problem: steady-state salinity contours corn uted by (a) Herbert et al.;2o (b) Oldenburg and
Pruess? (c) FEFLOW simulation; (d) ROCK FpLOW simulation.
groundwater flow and solute transport. The correspond- and the Galerkin method in combination with a one-
ing balance equations are formulated in terms of mass step Newton predictor-corrector scheme).
fractions and in terms of mass concentrations, which are Numerical simulations have been conducted in order
incorporated in the finite element codes ROCKFLOW to study the impact of different levels of the approx-
and FEFLOW. Different approximation and iteration imation of density variations in the mass balance
schemes are implemented in the codes for solving the equations, such as the Boussinesq approximation, its
nonlinear equations (e.g. the Euler-Taylor-Galerkin extension, and fully density approximations. Differ-
method in combination with a Picard iteration scheme ences occur in the convection pattern during the
Coupled groundwater JEow and transport 43
strongly dependent on concentration. Water Resour. Res., 40. Rivera, A. & Johns, R., Reverification of the HYDRO-
24 (1988) 1781-95. COIN test case 5 with NAMMU6.2 and comparison with
21. Holzbecher, E., Numerische Modellierung von Dich- TOUGH2, Presented on 1st GAMM Seminar on Modelling
testrijmungen im poriisen Medium. In Mitteilungen Nr. and Computation in Environmental Sciences, Okt. 12-13,
117, Institut fur Wasserbau und Wasserwirtschaft, Tech- Stuttgart, Germany, 1995.
nische Universitat, Berlin, 1991. 41. Schelkes, K. & Vogel, P., Paleohydrological information
22. Horton, C. W. & Rogers, F. T., Convection currents in a as an important tool for groundwater modeling of
porous medium. J. Appl. Phys., 16 (1945) 367-9. Gorleben site, Paleohydrological Methodr and their Appli-
23. Huyakorn, P. S. & Pinder, G. F., Computational Methods cations to Radioactive Waste Disposal, Proc. OECDINEA
in Subsurface Flow. Academic Press, New York, 1983. Workshop, Paris, France, 1992.
24. Kolditz, 0. & Diersch, H.-J., Quasi-steady-state strategy 42. Schincariol, R. A. & Schwartz, F. W., An experimental
for numerical simulation of geothermal circulation in hot investigation of variable density flow and mixing in
dry rock fractures. Znt. .Z. Non-Linear Mech. 28 (1993) homogeneous and heterogeneous media. Water Resour.
467-81. Res. 26 (1990) 2317-29.
25. Kolditz, O., Benchmarks and examples for numerical 43. Schincariol, R. A., Schwartz, F. W. 8~ Mendoza, C. A., On
groundwater simulations. In Diersch H.-J., FEFLOW the generation of instabilities in variable density flow.
Manual, WASY - Gesellschaft fur wasserwirtschaftliche Water Resour. Res. 30 (1994) 913-27.
44. Segol, G., Pinder, G. F. & Gray, W. G., A Galerkin finite
Planung und Systemforschung, Berlin, 1994.
element technique for calculating the transient position of
26. Kolditz, O., Ratke, R., Zielke, W. & Diersch, H.-J.,
the saltwater front. Water Resour. Res. 11(1975)343-7.
Coupled physical modelling for the analysis of ground-
45. Segol, G., Classic Groundwater Simulations - Proving and
water systems. In Notes on Numerical Fluid Mechanics,
Zmproving Numerical Models. PTR Prentice Hall, Engle-
Vol. 51. Vieweg, Braunschweig-Wiesbaden, 1995.
wood Cliffs, NJ, 1994.
27. Krohn, K.-P., Simulation von Transportvorgangen im
46. Senger, R. K. & Fogg, G. E., Stream functions and
kliiftigen Gestein mit der Methode der Finiten Elemente, equivalent freshwater heads for modeling regional flow of
Bericht Nr. 29/1991, Institut fur Stromungsmechanik, variable-density groundwater: 1. Review of theory and
Universitat Hannover, Dissertationsschrift, 1991. verification. Water Resour. Res., 26 (1990) 2089-96.
28. Kriihn, K.-P. & Zielke, W., FE-Simulation von Trans- 47. Swedish Nuclear Power Inspectorate, HYDROCOIN -
portvorglngen im khiftigen Gestein. Deutsche Gewiisser- an international project for studying groundwater hydro-
kundliche Mitteilungen, 35 (1991) 82-8. logy modeling strategies, Level 1 Final Report, Stockholm,
29. Lapwood, E. R., Convection of a fluid in a porous 1986.
medium. Proc. Cambridge Phil. Sot., 44 (1948) 508-21. 48. van der Molen, W. H. & van Ommen, H. C., Transport
30. Leijnse, T. & Hassanizadeh, S. M., Verification of the of solutes in soils and aquifers. J. Hydrol., 100 (1988)
METROPOL code for density dependent flow in porous 433-51.
media, Report No. 728528002, Rijksinstituut voor Volks- 49. Vogel, P., Zur Theorie binlrer Fluidgemische in poriisen
gezondheid en Milieuhygiene RIVM, Bilthoven, 1989. Medien, Technischer Bericht, Archiv-Nr. 113882, Bunde-
31. Leijnse, T., Three-dimensional modeling of coupled flow sanstalt fur Geowissenschaften und Rohstoffe, Hannover,
and transport in porous media, Ph.D. Thesis, University of 1995.
Notre Dame, Indiana, 1992. 50. Voss, C. I. & Souza, W. R., Variable density flow and
32. Lever, D. A. & Jackson, C. P., On the equation for the solute transport simulation of regional aquifers containing
flow of concentrated salt solution through a porous a narrow freshwater-saltwater transition zone. Water
medium, US Department of Energy, Report No. DOE/ Resour. Res., 23 (1987) 1851-66.
RWj85.100, 1985. 51. Wooding, R. A., Free convection of fluid in a vertical tube
33. Mendoza, C. A., VapourT users guide (version 2.11). filled with porous material. J. Fluid Mech., 13 (1962) 126-
Waterloo Center for Groundwater Research, University of 44.
Waterloo, Canada, 1990. 52. Zielke, W., Helmig, R., Krohn, K.-P., Shao, H. &
34. Nield, D. A. & Bejan, A., Convection in Porous Media. Wollrath, J., Discrete modeling transport processes in
Springer, Berlin, 1992. fractured porous rock, Proc. Znt. Congress on Rock
35. Oldenburg, C. M. & Pruess, K., Dispersive transport Mechanics. Aachen, 1991.
dynamics in a strongly coupled groundwater-brine flow
system. Water Resour. Res., 31 (1995) 289-302.
36. Oostrom, M., Hayworth, J. S., Dane, J. H. & Guven, O.,
Behavior of dense aqueuos phase lechate plumes in
APPENDIX
homogeneous porous media. Water Resour. Res., 28
(1992) 2123-34.
37. Pinder, G. F. & Cooper, H. H., A numerical technique for ROCKFLOW - finite element formulations
calculating the transient position of the saltwater front.
Water Resour. Res., 6 (1970) 875-82. Substituting Darcy’s law, eqn (9), into the balance
38. Ratke, R., Zur Liisung der Strbmungs- und Transport- equation for the overall fluid mass (12), a combined flow
gleichung bei veranderlicher Dichte, Technischer Bericht,
equation can be derived
Institut fur Striimungsmechanik und Elektronisches
Rechnen im Bauwesen, Universitlt Hannover, 1995.
39. Reeves, M., Ward, D. S., Johns, N. D. & Cranwell, R. M.,
Theory and implementation of SWIFT II, the Sandia
‘waste-isolation’ flow and transport model for fractured
media, Report no. SAND83-1159, Sandia National
(Al)
Laboratory, Albuquerque, 1986. The weighted residual method applied to this flow
Coupled groundwater flow and transport 45
equation yields the following weak formulation mass conservation, which contains additional terms due
to the higher order approximation in time.
1
a%
w apax,ax,
PW
x (PF’
+ e&k) dA
>
1 J
second order derivatives we obtain
X NjzdA AWk = (PQpw)p) NiNkdA
P A
NiNjNkdA -- 81 ?pj
to the right hand side to maintain an overall mass function is introduced after the reduction of second
balance at the nodes. order derivatives by the Green-Gauss formula. The
To avoid non-symmetrical terms within the discre- setup condition for the global system is the same as for
tized transport equation, the interpolation of the density the flow equation.