Stochastic modeling and simulations for solute

transport in porous media

Abstract

A stochastic model for solute transport in aquifers is studied based on the

concepts of stochastic velocity and stochastic diffusivity. By applying finite difference
techniques to the spatial variables of the stochastic governing equation, a system of stiff
stochastic ordinary differential equations is obtained. Both the semi-implicit Euler
method and the balanced implicit method are used for solving this stochastic system.
Based on the Karhunen-Loeve expansion, stochastic processes in time and space are
calculated by means of a spatial correlation matrix. Four types of spatial correlation
matrices are presented based on the hydraulic properties of physical parameters.
Simulations with two types of correlation matrices are presented.

1. Introduction

In the last few decades solute transport in porous media has attracted much
attention. This research topic is very important in both environmental protection and
resource management. However, due to the inaccuracy or insuf-ficiency of
hydrogeological and hydrochemical information, there are many uncertainties in
modeling solute transport by the deterministic advection-dispersion equation. To
overcome these difficulties, researchers usually use stochastic approaches that treat
aquifer properties and flow variables as spatial random fields characterised by a limited
number of statistical parameters rather than by well defined deterministic values. Dagan
[3] considered the concentration of a solute as a random variable and presented a
stochastic model with random coefficients. Deterministic equations are derived for the
moments of the concentration. The weakness of this type of stochastic approach is the
limitation on the magnitude of the random parameters required by the perturbation
expansion solution. Rubin [8] recently surveyed the development in this research
direction.
Another approach is to simulate the stochastic governing equation directly if the
number of stochastic sources is more than one. Hydraulic properties may be based on
stochastic coefficients, a stochastic driving force, and/or stochastic boundary and initial
conditions [5, 9]. The source of randomness is used more directly in order to deal with
larger random variations. This numerical approach not only provides statistical properties
of the concentration but also gives stochastic simulations of the solute transport. With the
progress in computing facilities and numerical methods for stochastic differential
equations, this approach is becoming more and more attractive.
In the stochastic simulation approach, stochastic ordinary differential equations
(SODEs) are obtained after applying a finite difference scheme to spatial variables of the
stochastic governing equation. One of the difficulties in designing effective numerical
methods for strong solutions of SODEs is the stability requirements when the SODEs are
stiff in both the deterministic and stochastic components. There have been some attempts
in the literature for improving stability properties of numerical methods for stiff SODEs
[2]. However, due to the complexity of the SODEs in this paper, the semi-implicit Euler
method and the balanced implicit method will be used in simulations.
The outline of this paper is as follows. In Section 2 we first discuss the stochastic
governing equation. In Section 3 we apply a semi-discretization process to the stochastic
governing equation and obtain a system of SODEs. The calculation of stochastic
processes in time and space is also discussed in this section. Numerical simulations are
presented in Section 4.

2. Stochastic model

For a three-dimensional deterministic problem, the solute transport in an

aquifer is described by the advection-dispersion equation [1]

∂ (ϕC )
+ div(ϕJ ) = 0, [1]
∂t

with v = −(K/ ϕ)∇φ and J = vC−D∇C .

Here C is the solute concentration at position (x1, x2, x3) and time t, D is a dispersion
tensor which accounts for the molecular diffusion and pore-scale dispersion, v = (vx1 , vx2
, vx3)T is the Darcy velocity vector, ∇φ is the gradient of piezometric head, ϕ is the
porosity (volumetric water content) and K is the hydraulic conductivity.
In an isotropic porous medium, if aligning an axis with the direction of the
velocity vector, the diffusivity matrix can be written as a diagonal matrix D = diag(Dx1
,Dx2 ,Dx3 ). For the 3-dimensional steady flow in homogeneous and isotropic soil, it is
assumed that K and ϕ are constant, D is a constant tensor, and ∇φ satisfies

∂ 2φ ∂ 2φ ∂ 2φ
+ + =0 [2]
∂x12 ∂x22 ∂x32

Substituting v and J into the governing equation [1] and applying the homogeneous and
isotropic condition [2], the following equation gives the 3-dimensional solute transport in
homogeneous and isotropic aquifers
∂C 3 ∂C 3 ∂ 2C
= − ∑ v xi + ∑ Dxi 2 [3]
∂t i =1 ∂xi i =1 ∂xi
Since the velocity of a solute particle is central in formulating a model of transport
process, we first consider the concept of stochastic velocity for the motion of a tracer in a
porous medium with uncertainties due to the pore structure:
K
v = − ∇φ + s1 , [4]
ϕ
T
where s1(x, t) = (s11, s12, s13) is a stochastic vector in time and space. Apart from the
velocity of a tracer, another important physical phenomenon involved in the solute
transport is dispersion which is used to describe solute movement carried by the
convection fluid. Dispersion may play a significant role in describing the solute transport
if the velocity is very small. The second stochastic concept considered here is the
stochastic dispersion
D = D + s2 , [5]
where the stochastic components s2(t, x) = diag(s21, s22, s23) has the same structure of D.
Substituting [4] and [5] into [3], the following random field gives the solute transport in
porous media with uncertainties:
∂C 3 ∂C 3 ∂ 2C
= − ∑(v xi + s1i ) + ∑( Dxi + s2i ) 2 [6]
∂t i =1 ∂xi i =1 ∂xi
Due to the limit of space, here we just consider simulations for one-dimensional solute
transport in an homogeneous and isotropic aquifer. For this one-dimensional problem, the
stochastic governing equation is
∂C ∂C ∂ 2C
= − (v + s1 ) + ( D + s2 ) 2 [7]
∂t ∂x ∂x
Note that model [7] is different from that proposed by Kulasiri [5] in which the
derivatives of stochastic processes were considered.
The initial condition for 1-dimensional solute transport is C(x, 0) = c0(x)
(0 ≤ x ≤ l) and we consider the upstream boundary condition
⎧c1 (t ),0 < t < t 0 , ∂C
C (0, t ) = ⎨ = c 2 (t ). [8]
⎩0, t > t 0 , ∂x x =l

3. Stochastic differential equations

For a given equidistant mesh on [0, l] with xn = nhx (n = 0, 1, . . . ,N + 1), forward and
central differences are used to approximate the first and second partial differentials in
space, respectively. Denoting Cn = C(xn, t) and sjn = sj(xn, t) (j = 1, 2), stochastic equation
[7] is approximated by
dC n C − Cn C − 2C n + C n −1
= −(v + s1n ) n+1 + ( D + s 2 n ) n +1 , [9]
dt hx hx2
for n = 1, . . . ,N . Denoting C = (C1, . . . ,CN)> and s j = (sj1, . . . , sjN)T, equations [9]
become
N
d C = ( Ao + A1 C )dt + ∑ [B1n dξ1n + B2 n dξ 2 n ]C + A2 [10]
n =1
where
⎛D D − vh ⎞
A0 = ⎜⎜ 2 C 0 ,0,...,0, 2
C N +1 ⎟⎟
⎝ hx hx ⎠
1
A1 = 2 tri − diag ( D,a1 , a2 )
hx
1
A2 = 2 (C0 dξ 21 ,0,...,0) T
hx
with a1 = vh − 2D and a2 = D − vh . Here dξjn = sjndt and ξjn = ξj(xn, t) . C0 and CN+1 are
determined by the boundary conditions. The sparse matrices Bjn = (b kl (jn) )N×N have non-
zero elements:

1 ⎧− 1, k = n, l = n + 1,
bkl(1n ) = ⎨ n = 1,..., N - 1;
hx ⎩1, k = l = n,

1 ⎧− 2, k = l = 1,
bkl( 21) = ⎨
hx2 ⎩1, k = 1, l = 2;

1 ⎧− 2, k = l = n,
bkl( 2 n ) = ⎨ n = 2,..., N - 1;
hx2 ⎩1, k = n, l = n − 1 or l = n + 1,

1 ⎧1, k = N , l = N − 1,
bkl( 2 N ) = ⎨
hx2 ⎩− 1, k = N , l = N ;

Now we consider two different simulation methods for solving SODEs. Forthe general
SODEs in ItÔ form
d
dy = g 0 ( y )dt + ∑ g i ( y )dWj (t ) , [11]
j =1

The semi-implicit Euler method takes the form [2]

d
yi +1 = yi + g 0 ( yi +1 )ht + ∑ ∆W j( i ) g i ( y j ). [12]
j =1

Here the ∆W j(i ) are the Wiener increments for time step i + 1 and are independent
N(0, ht). The balanced implicit method takes the form [7]

d
y i +1 = y i + g 0 ( y i )ht + ∑ ∆W j(i ) g i ( y j ) + U i ( yi − y i +1 ), [13]
j =1

with
d
U i = u 0 ( y i )ht + ∑ ∆W j( i ) u j ( y i ).
j =1

Here uj (j = 0, 1, . . . , d) represent m × m-matrix-value functions if y is a m-dimensional

vector. These matrices are chosen appropriately to get good damping properties.
Applying the semi-implicit Euler method [12] to [10] gives

[ ]
N
C i +1 = C i + ( A0 + A1 C i +1 )ht + ∑ B1n ∆ξ1(ni ) + B2 nξ∆(2i )n C i + ∆A2 ,
n =1
where C i = C (t i ) with ti = iht , and ∆A2 = (C0∆ξ21/ hx2 ,0,…,0)T
Using the Karhunen-Loeve expansion [4], the stochastic process
∆ξ(x, t) = ξ(x, t + ht) − ξ(x, t) has the truncating expansion
N
∆ξ ( x, t ) = ∑ λn ∆η n (t ) f n ( x),
i =1

where the ∆η n (t ) ~ N (0, ht ) are independent Gaussian processes. Then the stochastic
vector ∆ξ (t ) = ∆ξ ( x1 , t ),..., ∆ξ ( x N , t )) T can be expressed as ∆ξ (t) = F∆η with
F = (Fij)N×N = (fj(xi))N×N and
∆η = ( λ1∆η1 (t ) ,..., λ N ∆η N (t ), ) T
A spatial correlation matrix Q = (Qij)N×N is defined with eigenvalues λI and the
corresponding eigenvectors en = (fn(x1), . . . , fn(xN))T . There are four types of spatial
correlation matrices based on the hydraulic properties of physical parameters.

Type 1. The spatial independent matrix

Qij = qδij . [14]
Type 2. The adjacent correlation matrix
Qij = tri-diag(q, 2q, q).
Type 3. The exponential correlation matrix [4]
xi − x j
Q ji = q exp− ,
r
with a parameter r for adjusting the correlation decay rate between two points of the
process.
Type 4. The triangular correlation matrix [4]
⎡ 1⎤
Q ji = q(1 − r xi − x j ), xi − x j ∈ ⎢0, ⎥, [15]
⎣ r⎦
where r is a parameter for adjusting the correlation between two points of the process.
Here q in all four types of correlation matrices is a scalar with different values for the
stochastic components of the velocity and diffusivity, respectively. Applying the
balanced implicit method [13] to [10], gives

[ ]
N
C i +1 = C i + ( A0 + A1 C i +1 )ht + ∑ B1n ∆ξ1(ni ) + B2 nξ∆(2i )n C i + ∆A2
n =1

⎧
[ ]
⎫
N
+ ⎨ A1ht + ∑ B1n ∆ξ1(ni ) + B2 n ξ∆(2i )n ⎬(C i − C i +1 )
⎩ n =1 ⎭

The implicit nature of this method is adjusted by choosing the matrix A1 and the
B jn appropriately.
Matrix A1 is determined by the properties of A1. As the eigenvalues of A1 are all
negative, we choose A1 = −A1 . In this case the deterministic component is implicit. For
stochastic components, the non-zero elements of the matrices B jn = b kl ( )
( jn )
N ×N can be
(1n ) ( 2n)
chosen as: b nn = µ1 / hx and b nn = µ 2 / hx2 (n = 1,..., N ). We have that
N

∑B
n =1
jn ∆ξ (jni ) = λ j diag ( ∆ξ j1 ,..., ∆ξ jn ), j=1,2,…

with λ1 = µ1/hx and λ2 = µ2/ hx2 . With positive µ1 and µ2, it can be proved that the matrix
N N
I − A1 + ∑ B1n ∆ξ1(ni ) + ∑ B 2 n ∆ξ 2( in)
n =1 n =1
has an inverse and every eigenvalue of the inverse matrix is less than 1. Compared with
the semi-implicit Euler method, the balanced implicit method can be implemented with
better stability properties.

4. Numerical results

In this section we present stochastic simulations of the solute transport in an aquifer by

using the semi-implicit Euler method and the balanced implicit method. Due to the limit
of space, we just give simulations with Types 1 and 4 correlation matrices. Hydrology
K
parameters are: Darcy velocity: υ = − ∇φ = 1 cm/hour; diffusivity: D = 0.03 cm2/hour
ϕ
[10], and step sizes are hx = 0.05 and ht = 0.01. The initial condition is C(x, 0) = 0 (0 ≤ x
≤5) and the boundary conditions are c1(t) = 1 , t0 = 3 and c2(t) = 0 in [8]. The scalars q in
[14] are q1 = 0.002 and q2 = 0.00002 while those in [15] are q1 = 0.005 and q2 = 0.00005
for ∆ξ1n and ∆ξ2n, respectively. Figure 1 gives simulations obtained by the semi-implicit
Euler method with the spatial independent matrix (left) and the triangular correlation
matrix (r = 1/l = 0.2) (right) at t = 1, 4, 7. Fluctuations of the concentration in the left
figure are larger than those in the right figure. For the triangular

Figure 1: Simulations of the semi-implicit Euler method with two differentspatial

correlation matrices given by Eqs [14] (left) and [15] (right). (t =1: solid line; t = 4:
dashed line; t = 7: dash-doted line).
Correlation matrix we have also simulated the system with other r in [15]. Fluctuations in
concentration are larger if a larger r is used.
As concentrations in deterministic simulations are in the range of [0, 1],
concentrations beyond this range in stochastic simulations may be derived from
stochastic processes in the governing equation, the finite difference procedure for
discretising the governing equation, and the accuracy and stability properties of
numerical methods for solving [10]. Here we use the balanced implicit method to study
the impact of stability properties of numerical methods on simulation results.
We can choose different µ1 and µ2 in the balanced implicit method to adjust
stability properties. Numerical simulations with all four types of spatial correlation
matrices indicate that, if larger µ1 and µ2 are used, fluctuations in concentration are
smaller but computational errors are larger. If µ1 and µ2 are too small, the difference
between simulations obtained by the semi-implicit Euler method and by the balanced
implicit method is very small. For the problem in this paper it is recommended to use µ1
and µ2 in the range between 0.5 and 2. In this case the stability properties of the balanced
implicit method are improved and the computational errors are acceptable.

Figure 2: Simulations of the balanced implicit method with two different spatial
correlation matrices given by Eqs [14] (left) and [15] (right). (t = 1: solid line; t = 4:
dashed line; t = 7: dash-doted line).

Figure 2 gives simulations using the balanced implicit method with µ1=µ2=1 . For both
types of spatial correlation matrices, negative concentrationsin Figure 2 are much smaller
than those in Figure 1. Similar observation can also be found for concentration
fluctuations which are greater than 1. This indicates that good stability properties of
numerical methods have positive impact on fluctuations in stochastic simulations. Thus
numerical methods with good stability properties can reduce fluctuations in stochastic
simulations.
5. Conclusions

A stochastic model for solute transport in aquifers has been considered based on
the concepts of stochastic velocity and stochastic diffusivity. The semiimplicit Euler
method and the balanced implicit method have been used to solve the derived SODEs.
The drawback of the semi-implicit Euler method is the large non-physical values of
concentrations (C < 0 or C > 1). It has been shown that the good stability properties of the
balanced implicit method have significant impact on stochastic simulations.
Both the semi-implicit Euler method and the implicit balanced method are of
strong order half. Considering the forward difference scheme applied to the governing
equation (7), the accuracy of numerical simulations in this paper is just O( ht + hx ). In
order to improve the accuracy of stochastic simulations, future work should be based on
the development of implicit numerical methods for solving SODEs with high
convergence order and with better stability properties. In addition it would be very
important to employ other discretization schemes for the governing equation (7) in order
to improve stability properties, which is a key issue in numerical solutions of partial
differential equations. Candidate schemes include the up-wind scheme and the Dufort-
Frankel scheme [6]. All of these are the topics for future work.