Transp Porous Med (2012) 95:25–54

DOI 10.1007/s11242-012-0031-z

Numerical Simulation of Natural Convection

in Heterogeneous Porous media for CO2
Geological Storage

Panneerselvam Ranganathan · Rouhollah Farajzadeh ·

Hans Bruining · Pacelli L. J. Zitha

Received: 16 February 2010 / Accepted: 5 June 2012 / Published online: 10 August 2012
© The Author(s) 2012. This article is published with open access at Springerlink.com

Abstract We report a modeling and numerical simulation study of density-driven natural

convection during geological CO2 storage in heterogeneous formations. We consider an
aquifer or depleted oilfield overlain by gaseous CO2 , where the water density increases due
to CO2 dissolution. The heterogeneity of the aquifer is represented by spatial variations of
the permeability, generated using Sequential Gaussian Simulation method. The convective
motion of the liquid with dissolved CO2 is investigated. Special attention is paid to instability
characteristics of the CO2 concentration profiles, variation of mixing length, and average CO2
mass flux as a function of the heterogeneity characterized by the standard deviation and the
correlation length of the log-normal permeability fields. The CO2 concentration profiles
show different flow patterns of convective mixing such as gravity fingering, channeling, and
dispersive based on the heterogeneity medium of the aquifer. The variation of mixing length
with dimensionless time shows three separate regimes such as diffusion, convection, and
second diffusion. The average CO2 mass flux at the top boundary decreases quickly at early
times then it increases, reaching a constant value at later times for various heterogeneity
parameters.

Keywords Density-driven natural convection · Porous media · Heterogeneity ·

CO2 · Solubility · Mass transfer

List of Symbols
Variables
A Aspect ratio of the domain H/L (–)
A0 Amplitude of sinusoidal perturbation
c Dimensionless concentration (–)
c Concentration (mol/m3 )

P. Ranganathan · R. Farajzadeh · H. Bruining · P. L. J. Zitha (B)

Department of Geotechnology, Delft University of Technology, 2628 CN, Delft, The Netherlands
e-mail: [email protected]

123
26 P. Ranganathan et al.

D Diffusion coefficient (m2 /s)

f (x, z) Permeability variation function
Fm Average CO2 mass flux
g Acceleration due to gravity (m/s2 )
H Height of the porous medium (m)
i and j Index of the mesh points in x and z direction, respectively
k Permeability (m2 )
k0 Mean permeability (m2 )
L Length of the porous medium (m)
P Pressure (Pa)
Ra Rayleigh number (–)
t Time (sec)
td Dimensionless time (–)
ux Dimensionless x-component velocity (–)
uz Dimensionless z-component velocity (–)
uX x component velocity (m/s)
uZ z component velocity (m/s)
w Width of domain (m)
w (t) Total dissolved CO2 concentration
x Dimensionless distance in X coordinate
Y Log-normal permeability distribution
z Dimensionless distance in Z coordinate

Greek Symbols
α Wave-number
βc The volumetric expansion coefficient
ε Criteria for convergence
ϕ Porosity (–)
λ0 Wavelength of sinusoidal perturbation (–)
λ Correlation length scale (m)
μ Viscosity of fluid (kg/m s)
θ Mean of log permeability
ψ Stream function (m3 m−1 s−1 )
ρ0 Pure liquid density (kg/m3 )
ρ Liquid density (kg/m3 )
ρ Change in density (kg/m3 )
σ Standard deviation
τx Horizontal correlation length (m)
τz Vertical correlation length (m)

Abbreviation
SGSIM Sequential Gaussian Simulation

123
Numerical Simulation of Natural Convection in Heterogeneous... 27

1 Introduction

Concern about global warming has increased the impetus to reduce the emissions of CO2 , one
of the main greenhouse gases, into the atmosphere. The injection of CO2 into aquifers and
depleted gas or oil fields is an option for achieving that goal. It can also be used for enhanc-
ing gas or oil recovery. Natural convection is one of the main mechanisms to enhance mass
transfer of CO2 into the liquid or oil phase during CO2 storage in aquifers and water-flooded
oil reservoirs (Firoozabadi and Cheng 2010). The natural convection phenomenon arises
because water with dissolved CO2 has a higher density than pure water. Over the last few
decades, density-driven natural convection process in porous media received much attention.
Many authors (Gebhart et al. 1988; Nield and Bejan 1999; Ennis-King and Paterson 2005;
Li et al. 2006; Riaz et al. 2006; Farajzadeh et al. 2007; Hassanzadeh et al. 2007; Pruess and
Zhang 2008; Rapaka et al. 2008; Hidalgo and Carrera 2009; Kneafsey and Pruess 2010; Pau
et al. 2010) investigated the density-driven natural convection in homogeneous porous media
in view of its application to CO2 sequestration. The effect of layering and anisotropy in the
permeability fields has also been investigated (Farajzadeh et al. 2008; Rapaka et al. 2009).
The convective mixing in a heterogeneous porous medium received less attention in litera-
ture compared to the homogeneous case. Nield and Bejan (1999), McKibbin and O’Sullivan
(1980), and Malkovsky and Pek (1999) developed analytical solutions for the onset condi-
tions of density-driven natural convection in heterogeneous media for simple geometry and
boundary conditions using linear stability analysis. However, this theory does not describe
the convective mixing after onset of natural convection. Welty and Gelhar (1991, 1992) per-
formed stochastic analyses of one-dimensional density-coupled CO2 transport and derived
mathematical expressions for an effective longitudinal, variable-density macrodispersivity.
Their analytical expression for the longitudinal macroscopic dispersivity is a function of the
geostatistical properties of the porous medium (log-permeability variance and correlation
length), concentration gradient, fluid properties, pressure gradient, and time. Schincariol et
al. (1997), Simmons et al. (1999, 2001), and Prasad and Simmons (2003) reported numerical
modelling studies of natural convection in heterogeneous porous media, with application to
groundwater flows. Schincariol et al. (1997) studied the effects of heterogeneity in terms of
the variance and the correlation length of the random permeability field on unstable behavior.
They reported that longer correlation lengths and higher variance of log-normal permeability
distribution promote stability, i.e., less gravity fingering. Simmons et al. (2001) conducted
numerical experiments to study variable-density groundwater flows in both periodically struc-
tured and random-permeability fields. They found that both the onset and subsequent growth
and decay of instabilities are mainly related to the structure and the variance of the permeabil-
ity field. Prasad and Simmons (2003) used the short heater paradigm (Elder 1967) to study the
effect of heterogeneity in permeability distributions on the solute transport by Monte Carlo
techniques. They found that an increase in the standard deviation of the log-permeability field
results in greater instabilities at earlier times but promotes stability at later times. Further,
they found that an increase in the horizontal correlation length of the log-permeability field
causes a reduction in the degree of instability. Nield and Simmons (2007) examined the effect
of heterogeneity on the onset of natural convection in porous media based on the criterion of
“rough-and-ready”. Bryant et al. (2008) studied the effect of permeability correlation length
and anisotropy on CO2 migration and residual CO2 trapping using single realisation of cor-
related permeability fields. Recently, Green and Ennis-King (2010) investigated the effect of
vertical permeability variations in a reservoir with randomly distributed impermeable bar-
riers on vertical migration of CO2 due to buoyancy using multiple realisations numerical
simulation. Nield et al. (2010) investigated the effect of strong heterogeneity (large variation

123
28 P. Ranganathan et al.

of permeability around its mean value) on the onset of convection using spatially correlated
log-normal permeability fields and analyzed the result in the statistical framework. Han et al.
(2010) investigated the effect of heterogeneity mainly on CO2 trapping mechanism in saline
formations. They found that the amount of trapped CO2 depends more on the geometry of the
geological structure than on the magnitude of the effective permeability. Recently, we inves-
tigated the effect of heterogeneity on density-driven natural convection of CO2 overlying a
brine layer (Farajzadeh et al. 2011). The spectral method of Shinozuka and Jan (1972) was
used to generate permeability heterogeneity fields. We identified several flow regimes (fin-
gering, channeling, and dispersive) based on the permeability heterogeneity of the medium,
i.e., Dykstra–Parsons coefficient (VDP ) and correlation length. We also found that the rate of
mass transfer CO2 into water is higher in heterogeneous media.
This study is a detailed extension of our previous study on the natural convection in
heterogeneous media. In the present case, the heterogeneity fields were generated using a
more standard Sequential Gaussian Simulation (SGSIM) method. The heterogeneity i.e., the
spatial variations of the permeability, is characterized by standard deviation which scales like
Dykstra–Parsons coefficient (VDP ) and the correlation length. The focus here is the effect of
the Rayleigh number, evolution of the mixing length, full time dependence of the average
CO2 mass flux, effect of boundary conditions and mesh size dependence and the height
and width of the domain. This is relevant for CO2 sequestration in aquifers in geologically
heterogeneous formations.
The remainder of the paper is organized as follows. Section 2 provides the physi-
cal model, the governing equations and the method used to generate the random per-
meability fields. Section 3 presents the method of solution implemented. Section 4 is
devoted to the results and discussion. Section 5 presents the main conclusions of this
study.

2 Model Formulation

2.1 Physical Model

We consider a heterogeneous porous medium with height H and length L (Fig. 1) saturated
with water (oil) when a CO2 -rich gaseous phase overlays the liquid column. Isothermal
conditions are assumed throughout. When CO2 is injected into a saline aquifer, it migrates
to the top of the formation due to gravity. Note that our model is also valid for oilfields at
the end of the water phase flooding. However, this study is confined to the mixing of CO2
during injection of CO2 into saline aquifers. We consider a subset of the above problem,
i.e., a vertical cross section of a single-phase flow aquifer domain, which is overlain by
CO2 . Recently, Lindeberg and Bergmo (2003) studied numerical simulation of full two-
phase problem and reported that the interface between the CO2 -rich gas and water remains
relatively sharp and is not deflected by the fingering in the underlying brine. Ennis-King and
Paterson (2003) pointed out that a local equilibrium between the CO2 -rich gas and the brine
can be assumed at the interface. Accordingly, we assume that interface between the CO2 -rich
gas and liquid is fixed and sharp, i.e., the capillary forces are neglected. Due to the overlying
free CO2 phase, the upper boundary of the single-phase aqueous region remains saturated
with CO2 . The flow domain is initially in gravity equilibrium (hydrostatic pressure gradient).
Furthermore, the flow velocities are small and hence hydrodynamic dispersion reduces to the
molecular diffusion (Bear 1972).

123
Numerical Simulation of Natural Convection in Heterogeneous... 29

Fig. 1 Schematic representation of present model with initial and boundary conditions

2.2 Governing Equations

We consider two-dimensional single-phase flow in a porous medium as depicted in Fig. 1,

where the permeability is a function of x and z within the computational domain. We adopt the
following notation for the partial derivatives: ∂x ≡ ∂/∂ x, ∂x22 ≡ ∂ 2 /∂ x 2 for any independent
variable x. The governing equations can be written as
(a) Continuity equation
∂t ρ + ∂ X (ρu X ) + ∂ Z (ρu Z ) = 0 (1)
(b) Darcy’s law
u X = −k0 μ−1 f (x, z)∂ X p (2)
−1
u Z = −k0 μ f (x, z) (∂ Z p − ρg) (3)
In these equations f (x, z) describes the permeability variations with respect to mean perme-
ability, k0 , or otherwise k (x, z) = k0 f (x, z).
(c) Concentration
 
ϕ∂t c + u X ∂ X c + u Z ∂ Z c = ϕ D ∂ X2 2 c + ∂ Z2 2 c (4)
The symbols are defined at the start of paper. We use the Boussinesq approximation, which
assumes a constant density everywhere in the equation of motion except in the buoyancy term.
This approximation is valid since it is assumed here to simplify the mass balance equation,
where density variation due to dissolution of CO2 is assumed to be small. The changes of
liquid density with the increasing CO2 concentration is assumed to be a liner function of the
form
 
ρ = ρ0 + βc c − c0 (5)
where ρ0 is the pure liquid density and βc is the coefficient of density increase with concen-
tration.

123
30 P. Ranganathan et al.

Equations (1–4) were solved assuming mainly no-flow boundary conditions at the bottom,
left, and right side of the porous medium. The no-flow boundary condition is chosen for the
lateral sides of simulation domain to ensure that simulation domain represents a laterally
infinity domain. This is a common practice in reservoir simulation studies.

2.3 Dimensionless Equations

We take H as the characteristic length and define the following dimensionless variables
X Z H H D c − ci
x= , z = , ux = u X , uz = u Z , td = 2 t, c =  (6)
H H ϕD ϕD H c0 − ci
and we also define a Rayleigh number, Ra by
ρgk0 H
Ra = (7)
ϕ Dμ
where k0 and k(x, z) are as defined above. Rayleigh Number is dimensionless number that
describes instabilities due to natural convection and is defined as the ratio of buoyancy flux
to the diffusive flux.
The governing equations then take the form

∂td ρ + ∂x (ρu x ) + ∂z (ρu z ) = 0 (8)

u x = − f (x, z) ∂x P (9)
u z = − f (x, z) (∂z P − Ra c) (10)
 
∂td c + u x ∂x c + u z ∂z c = ∂x22 c + ∂z22 c (11)

we introduce a stream function ψ so that

u x = −∂z ψ, u z = ∂x ψ, (12)

we also eliminate P, thus

   
∂x f −1 (x, z) ∂x ψ + ∂z f −1 (x, z) ∂z ψ = Ra ∂x c (13)
 
∂td c + ∂x ψ∂z c − ∂z ψ∂x c = ∂x22 c + ∂z22 c (14)

2.4 Boundary and Initial Conditions

The initial condition of the problem is

ψ = 0 and c = 0 at td = 0, (15)

The boundary conditions of the problem are

ψ = 0 and ∂z c = 0 at x = 0,
ψ = 0 and c = 1 at z = 0,
ψ = 0 and ∂x c = 0 at z = 1,
ψ = 0 and ∂x c = 0 at x = A (16)

123
Numerical Simulation of Natural Convection in Heterogeneous... 31

Table 1 Geostatistical
Case no. Standard Correlation Number
parameters
deviation length realisations

1 0.1 0.025 10
2 0.1 0.05 10
3 0.1 0.1 10
4 0.1 0.25 10
5 0.3 0.025 10
6 0.3 0.05 10
7 0.3 0.1 10
8 0.3 0.25 10
9 1 0.025 10
10 1 0.05 10
11 1 0.1 10
12 1 0.25 10
13 2 0.025 10
14 2 0.05 10
15 2 0.1 10
16 2 0.25 10

2.5 Random Permeability Fields

We assume that the flow domain represented in Fig. 1 is heterogeneous with a log-normal
permeability distribution [Y = ln(k)]. The permeability field is described using the following
main parameters: mean permeability, standard deviation, and correlation length in the vertical
direction τ y as well as in the horizontal direction τx . Its spatial correlation is quantified using
exponential autocorrelation function or the variograms (Simmons et al. 2001).
The random permeability fields were generated using Sequential Gaussian Simulation
(SGSIM) method developed by GSLib (Deutsch and Journel 1992). SGSIM program is most
commonly used to generate random fields of reservoir characterization and is based on multi-
gaussian approach in the spatial distribution of random field. This code generates a standard
normal field with a mean of zero and standard deviation of one. We have used the following
transformation (Nield et al. 2010) to obtain log-normal distributions which are used as input
to our simulation code.

Permnew = exp (θ + σ PermSGSIM ) (17)

In this study, 16 cases of different standard deviations and different correlation lengths are
considered. Each case corresponds to a different heterogeneity pattern. The parameters used
in each case are shown in Table 1. The ranges of standard deviation values are similar to
values quoted in the literature (LeBlanc et al. 1991; Barlebo et al. 2004).
The flow behavior obtained from a single realisation may be entirely different from the
flow pattern for another statically equivalent realisation. To draw any statically significant
conclusion, it is necessary to consider a large number of realisations of a single case. In
practise, however, this may not be viable as it requires excessive computation time. Since
SGSIM code is based on Monte-Carlo simulation for the generation of random field, it can
internally, effectively generate multiple equally probable property fields. Thus, a few field

123
32 P. Ranganathan et al.

realisations are used in simulations of density-driven flows (Schincariol et al. 1997; Bryant
et al. 2008). For each of the 16 cases, ten realisations of the selected permeability fields were
generated.

3 Method of Solution

The numerical simulations were done for a domain having a width of 1 m and a height of 1
m. Such small domain was chosen to insure that we could capture the details of the growth
of instabilities due to natural convection. This is required to describe the natural convection
behavior at the scale of a grid block, representing a small part of the aquifer. The question
whether such a fine scale simulation (grid dimension of 0.01m) is sufficiently accurate will
be addressed in Sect. 4.2. The flow domain was discredited in 101 × 101 grid cells. A fully
implicit finite volume approach (Guceri and Farouk 1985) was used to solve the governing
equations, to obtain transient values of the dependent variables in Eq. (14). For each time
step, the stream function was computed from Eq. (13) and then concentration profile was
obtained by solving Eq. (14). The numerical calculation for each time step was iterated until
the following criteria were satisfied

 K +1   K 
c K  ψ − ψK 
 i, j − ci, j   i, j i, j 
  ≤ ε and  ≤ε (18)
 c K +1   ψ K +1 
i, j max i, j

where i and j are the index of the mesh points in x and z direction, respectively, and the
superscript K stands for the iteration cycle. For our numerical computations, ε was set to
10−5 and the time step was chosen to be 10−6 to insure accuracy of the results. To observe
the fingering behavior it was necessary to disturb the interface. Therefore in the numerical
simulations, following the procedure in Farajzadeh et al. (2007), we used sinusoidal initial
perturbation are superimposed along with top boundary condition, i.e.,

 
2π x
c(x, z = 0, t = 0) = 1 + A0 sin (19)
λ0

where A0 is the amplitude and λ0 is the wavelength of sinusoidal perturbation. In this

study, A0 = 0.01 and λ0 = 1/12 is used. We studied the various ranges of wave num-
bers (α = 2π/λ0 ) between 10 and 270 to investigate the effect of the wavelength on the
numerical solution.

4 Results and discussion

Below we discuss the instabilities in terms of CO2 concentration profiles, variation of the
mixing length, and the average CO2 mass flux at the top boundary. The concentration profiles
are expressed as a dimensionless mass fraction of CO2 in the aquifer at any time. The mixing
length is defined as the distance from the top boundary to the location where the dimensionless
CO2 concentration is equal to 0.1. The mixing length expresses how the concentration of
dissolved CO2 is affected by hydrodynamics. The average CO2 mass flux at the top boundary
is calculated using

123
Numerical Simulation of Natural Convection in Heterogeneous... 33

Fig. 2 CO2 concentration profiles along with log-normal permeability fields of one realisation (σ = 0.3 and
λ = 0.1) for Rayleigh number of 10,000. (a) at td = 0.00005, (b) at td = 0.0001, (c) at td = 0.00015, (d) at
td = 0.0002, (e) at td = 0.00025, (f) at td = 0.0005, (g) at td = 0.00075

L 
∂c
Fm (t) = dx (20)
∂z z=1
0

4.1 Base Case

Figure 2 shows the CO2 concentration profiles along with log-normal permeability field at
various dimensionless times for a Rayleigh number Ra = 10, 000. We generated a single
realisation of a log-normal permeability field with a standard deviation of 0.3 and a corre-
lation length of 0.1 in both vertical and horizontal directions. Figure 2 shows that at early
times, e.g., 0.00015 (Fig. 2a), the number of fingers is equal to the initial perturbations,
11; the fingers grow independently of each other. At longer times, however, the number of
fingers is reduced as fingers interact with each other due to various non-linear interaction
mechanisms to form a large-scale structure that reaches the bottom of the flow domain. The

123
34 P. Ranganathan et al.

Fig. 3 A snapshot of velocity vectors along with concentration profiles at the dimensionless time of td =
0.0015 for log-normal permeability field (σ = 0.3 and λ = 0.1)

evolution mechanisms of the fingers include shielding, side-merging, tip splitting, and fading
phenomena observed earlier in homogeneous porous medium (Riaz et al. 2006; Farajzadeh
et al. 2007). Remarkably, it appears that the interaction of fingers mainly takes place at the
high-permeability zones, as shown by the circle in the figures. It is found that the interaction
occurs laterally between the dominant fingers and the slowly moving fingers. This mechanism
may be referred to as “side merging.” The lateral development of the tip of slowly moving
fingers that interact with the dominant fingers results in narrowing the neck of the domi-
nant fingers, which is indicated by the square in Fig. 2d. This mechanism may be referred
to as “necking.” The “side-merging” and “necking” mechanisms were discussed in detail
by Zimmerman and Homsy (1992). The other mechanisms of non-linear interaction are
tip-splitting and trailing lobe detachment. In tip splitting mechanism, the fingers that move
independently get flattened at the tip: this mechanism is not observed in this study. In trailing
lobe detachment, a trailing lobe of different concentration forms and detaches itself during
the interaction of the fingers. This mechanism is indicated by the ellipse in both Fig. 2e, f.
This trailing lobe detachment mechanism was previously reported by Rogerson and Meiburg
(1993) and Ghesmat and Azaiez (2008).
Since the velocity plays a major role in the convective mixing, a snapshot of velocity
vectors along with the CO2 concentration profiles are plotted in Fig. 3. It is evident from
the figures that the circulation pattern can be observed around the fingers tip, whereby the
downward velocity vectors (arrows) occur within the fingers and the upward velocity moves
toward the top boundary in the region between fingers. It is also observed from the same
figures that the velocity is smaller at the regions away from the fingers.
It is interesting to compare how instabilities grow in homogeneous and heterogeneous
porous media. Figure 4 shows the concentration profiles obtained for homogeneous media
with a Rayleigh number of 10,000. In the homogeneous case, the instabilities grow smoothly
due to uniform permeabilities. But in the heterogeneous case (Fig. 2), the instabilities become
much more ragged and grow faster than in homogeneous media. The growing instabilities
in heterogeneous media exhibit side bumps, i.e., small lateral bulges at the location where
the CO2 concentration begins to enter the high permeability regions. In the homogeneous
case at later times (td = 0.0005) fingers merge at the top and split halfway (Fig. 4). In the
heterogeneous case, the finger seems to remain separated (Fig. 2).

123
Numerical Simulation of Natural Convection in Heterogeneous... 35

Fig. 4 CO2 concentration profiles for Rayleigh number of 10,000 with various dimensionless times for
homogeneous porous medium. (a) at td = 0.00025, (b) at td = 0.0005, (c) at td = 0.00075, (d) at td = 0.001

4.2 Effect of Grid Size

In order to check the effect of grid size, we performed a number of simulations by varying
grid size for a particular realisation of the log-normal permeability fields with a standard
deviation of 0.3 and a correlation length of 0.1. Figure 5 shows the plot of the total amount
of dissolved CO2 versus dimensionless time for various grid sizes, ranging from 0.008 to
0.0167m. The total dissolved CO2 concentration is the integral quantity of the dissolution of
the CO2 and is given by
H L

w(t) = c(x, z, t)dxdz (21)

0 0

It is observed that the total dissolved CO2 is same for all grid sizes over time. Note that
the simulation with fine grid size (0.008–0.01m) predicts slightly lower total dissolved CO2
than predicted using larger grid sizes (0.0125–0.017m). Thus, it can be asserted that grid size
in the range of 0.008–0.0167m is adequate for resolving density-driven mixing.

123
36 P. Ranganathan et al.

Fig. 5 Effect of mesh sensitivity on the total dissolved CO2

4.3 Effect of Rayleigh Number

The effect of varying the Rayleigh number on the CO2 concentration profiles is now exam-
ined. Our simulations are based on generated permeability fields leading to Rayleigh numbers
of 1000, 2000, 5000, and 10,000. The standard deviation and the correlation length in all
fields is σ = 0.3 and λ = 0.1. For each condition ten realisations are made. Figure 6a
shows the log-normal permeability field for a single realisation. We first examine the behav-
ior of the instabilities at the early stages. Figure 6b shows CO2 concentration profiles for
a single realisation for the Rayleigh number Ra = 1000 at td = 0.001, Ra = 2000 at
td = 0.0005, Ra = 5000 at td = 0.00025, and Ra = 10, 000 at td = 0.0001. Figures
show that the concentration front moves faster as the Rayleigh number increases. For the
lower Rayleigh numbers (Ra = 1000 − 2000), there are only a few large fingers, which
move independently without interacting with each other. A significant dispersive spreading
is also visible in the concentration front. For higher Rayleigh numbers, i.e., (5,000–10,000),
the instabilities are more developed and form small-scale complex fingering structures. In
addition, the concentration front becomes steeper as the Rayleigh number increases.
Now, we consider the long time behavior of the fingers. The concentration profiles for
various Rayleigh numbers at the time that the CO2 plume reaches the bottom boundary are
plotted in Fig. 6c. For all values of Rayleigh numbers, it is found that small fingers merge
together to form two larger fingers, which reach the bottom of the flow domain. A dispersive
spreading in the concentration fronts is also observed for low Rayleigh numbers, whereas
the concentration profiles are steeper for high Rayleigh numbers. Fingers grow faster with
increasing Rayleigh number.

4.4 Effect of Standard Deviation

The effect of standard deviation on the instabilities was investigated using the following values
of standard deviations 0.1, 0.3, 1.0, and 2. Figure 7a–c shows the concentration profiles and
the log-normal permeability field for the values of the standard deviations examined, together

123
Numerical Simulation of Natural Convection in Heterogeneous... 37

Fig. 6 (a) Log-normal permeability field of one realisation, (b) CO2 concentration profiles at earlier time,
(c) at later time for various Rayleigh numbers near the time at which a finger hits the bottom

123
38 P. Ranganathan et al.

Fig. 6 continued

with different correlation length at a dimensionless time of 0.0005. At very small standard
deviations (σ → 0), i.e., for small degrees of heterogeneity, the instability behavior is similar
to the homogeneous case (Farajzadeh et al. 2007). As σ increases, there is a significant
change in the concentration profiles, as can be seen from Fig. 7. For lower standard deviation
(σ = 0.3), the estimated average logarithm of the permeability field obtained from a SGSIM
simulation varies between 0.7 and 1.4. For higher standard deviation, the estimated log-
permeability field varies from 500 to 3500. In the literature, the behavior of instabilities
defined was categorized into fingering, channeling, and dispersive behavior (Waggoner et al.
1992; Sorbie et al. 1994; Farajzadeh et al. 2011). The fingering pattern means that instabilities
grow along the preferential flow path, which is formed by permeability variations. Channeling
behavior of the instabilities show that instabilities move toward the high permeable zone
and are controlled by the permeability distribution pattern. The dispersive flow pattern of
instabilities occurs when the length scale of instabilities is smaller than the correlation length
scale of the permeability field and is characterized by the square root of time.

123
Numerical Simulation of Natural Convection in Heterogeneous... 39

Fig. 7 Log-normal permeability fields and CO2 concentration profiles of one realisation for Rayleigh number
of 10,000 at dimensionless time td = 0.0005 with different standard deviations for correlation lengths of
(a) 0.025, (b) 0.1, (c) 0.25

123
40 P. Ranganathan et al.

Fig. 7 continued

123
Numerical Simulation of Natural Convection in Heterogeneous... 41

Fig. 7 continued

123
42 P. Ranganathan et al.

Figure 7 shows that the fingering behavior of instabilities occurs for all values of the cor-
relation length (0.025–0.25) at lower value of standard deviation (σ = 0.1). The channeling
behavior of instabilities occurs for all values of the correlation length at standard deviations
below σ = 0.3; at a higher standard deviation i.e., 1 and 2 it occurs for a correlation length
of 0.25 (Fig. 7). For the correlation length range (0.025–0.1) at higher standard deviations
σ = 1 and σ = 2, the instabilities exhibit a dispersive pattern. The dispersive character of
the instabilities grows fast at earlier times and become passive at later times due to increase
in the dispersive mixing. These findings are consistent with the work of Prasad and Simmons
(2003), who observed that an increase in the standard deviation of the log-permeability field
caused an increase in instabilities at earlier times and decrease of instabilities at later times.

4.5 Effect of Correlation Length

To investigate the effect of the correlation length on the instabilities, we consider the same
correlation length in vertical and horizontal directions and set the correlation length to
λ = 0.025, 0.05, 0.1, and 0.25. Note that each realisation gives a different average per-
meability where the correlation length becomes comparable to the domain size. Figure 8a, b
shows the concentration profiles with log-normal permeability fields for a single realisation
with different correlation lengths at standard deviations of 0.3 and 1.0, respectively. The
instabilities exhibit a channeling behavior for the correlation lengths examined (0.025–0.25)
at a standard deviation of 0.3. The reason for this behavior is that for larger correlation length
(0.25), the log-normal permeability field generates high permeability zones toward which the
instabilities tend to move. Also the length scale of permeability is somewhat larger than that
of the instabilities formed in the domain. The dispersive pattern occurs at higher standard
deviations e.g., for a standard deviation of 1.0 and for a correlation length in the range of
0.025–0.1(Fig. 8a–c). This may be due to the fact that for smaller correlation length, the log-
normal permeability field leads to local fluctuations in the velocity due to a local change in
permeability leading to strong mixing. These observations are consistent with the findings by
Prasad and Simmons (2003), who observed that increasing the horizontal correlation length
reduction in the extent of fingering.

4.6 Evolution of the Mixing Length

Figure 9 shows the variation of the mixing length as a function of the dimensionless time
for all realisations of different correlation lengths with standard deviations between 0.3 and
1.0. For a low standard deviation (σ = 0.3), three regimes can be observed, namely dif-
fusion, convection and second diffusion. In the first regime, the mixing length scales like
the square root of time. This indicates that initially diffusion is dominant and convection
does not affect the flow significantly. In the second regime, the mixing length varies lin-
early with time. This shows that the flow is dominated by convection. There is a char-
acteristic time for the onset of convection, i.e., for the transition from the diffusion and
convection dominated regimes. In a third regime (second diffusion), the mixing length
varies at a rate which is lower than the second regime, but is slightly higher than that
of the first regime. This drop in mixing length growth rate is probably due to the pres-
ence of non-linear interaction mechanisms such as merging, shielding, tip splitting, and
fading.
For a higher standard deviations and for all realisations, only two regimes are observed
at lower correlation lengths i.e., between 0.025 and 0.05. A higher correlation length of 0.1

123
Numerical Simulation of Natural Convection in Heterogeneous... 43

Fig. 8 Log-normal permeability fields and CO2 concentration profiles of one realisation for Rayleigh number
of 10,000 at dimensionless time td = 0.0005 with different correlation lengths for standard deviations of
(a) 0.3, (b) 1.0

123
44 P. Ranganathan et al.

Fig. 8 continued

123
Numerical Simulation of Natural Convection in Heterogeneous... 45

Fig. 9 Mixing length variation as a function of dimensionless time for multiple realisations of different
correlation lengths with standard deviations of 0.3 and 1.0

123
46 P. Ranganathan et al.

Fig. 10 Mixing length growth with time (a) various Rayleigh numbers, (b) various standard deviations, (c)
various correlation lengths

and 0.25, only a few realisations give three regimes but most of the realisations show only
two regimes. These two regimes are similar to the second and third regimes that are obtained
earlier (Fig. 9a). The occurrence of only two regimes can be explained by the fact that for
high correlation lengths and at earlier times, the growth of the fingers is dominant, which
gives an increasing mixing length. At later times, the higher standard deviation promotes
lateral dispersive mixing, which leads to slower motion of the most advanced fingers. The
data given in the figure also shows that for low standard deviations all realisations have the
same mixing length and for high standard deviations there is a large variation of the mixing
length. However, for the high standard deviations the mixing length obtained from different
realisations show large deviation.
Figure 10a–c shows the variation of the mixing length versus dimensionless time obtained
by averaging ten realisations as a function of the Rayleigh number, standard deviation, and
correlation length of the log-normal permeability fields. Fig. 10a clearly exhibits the three
regimes discussed in Fig. 9a, for various Rayleigh numbers. It can be seen that the transition
time from diffusion dominated flow to convection dominated flow decreases as the Rayleigh
number increases. Furthermore, it is found similarly to the above graphs that the three mixing
length regimes reduce to two regimes when the standard deviation increases from 0.3 to 1.0.
For all correlation lengths, the profiles of the mixing length variation show a similar pattern,
as in Fig. 10c. Thus, the effect of correlation length on the growth of mixing length is
small.

4.7 Average CO2 Mass Flux

Figure 11 shows the average CO2 mass flux at the top boundary as a function of the dimen-
sionless time for ten realisations examined by varying the correlation length at a standard

123
Numerical Simulation of Natural Convection in Heterogeneous... 47

Fig. 11 CO2 mass flux at top boundary as a function of dimensionless time for multiple realisations of
different correlation lengths with standard deviations of 0.3 and 1.0

123
48 P. Ranganathan et al.

Fig. 12 Plot of average CO2 mass flux at the top boundary versus dimensionless time (a) various Rayleigh
numbers, (b) various standard deviations, (c) various correlation lengths

deviation of 0.3 and 1.0, respectively. For all realisations, the CO2 mass flux has a rapidly
decreasing trend at the early times and after slight fluctuations it decreases slowly toward a
constant value. This can be explained by observing that at early times, the mass transport of
CO2 from the top boundary is only controlled by diffusion. In cases where the CO2 mass
flux fluctuates, the first peak represents the onset of convection. The constant value of the
average mass flux may be explained by the fact that the domain is nearly saturated with dis-
solved CO2 after the fingers reach the bottom boundary. It is also shown from the figure that
for lower standard deviations the average flux of CO2 trend is the same for all realisations.
However, for larger standard deviations the mass flux obtained from different realisations
shows broader variations.
Figure 12a–c shows the variation of the average CO2 mass flux from the top boundary
obtained by averaging ten realisations, as a function of the Rayleigh number, standard devi-
ation and correlation length of the log-normal permeability field. First, the CO2 mass flux
decreases quickly at earlier times for all Rayleigh numbers, increases slightly and then levels
off to a plateau (Fig. 12a). It is also shown that the time at which the CO2 mass flux increases,
diminishes for higher Rayleigh numbers. Figure 12b shows the average CO2 mass flux at the
top boundary as a function of dimensionless time for different standard deviations. For the
lower standard deviation of 0.3 and 1.0, the average CO2 mass flux as a function of dimen-
sionless time decreases quickly at early times. Subsequently increases before it decreases
again. For the higher standard deviation of 1 and 2, the average CO2 mass flux decreases
with time and then reaches a stabilized value of the mass flux at later times. This shows that
the transport of mass is mainly controlled by dispersive mixing. The effect of the correlation
length on the average CO2 mass flux is shown in Fig. 12c. We observe that there is little
variation in the calculation of the average CO2 mass flux for different correlation lengths. In

123
Numerical Simulation of Natural Convection in Heterogeneous... 49

Fig. 13 Effect of various boundary conditions on the CO2 concentration profiles of one realisation (λ = 0.1
and σ = 0.3) at the dimensionless time of td = 0.001 (a) no-flow conditions for all boundaries, (b) no-flow
conditions for lateral boundaries and constant bottom boundary, (c) periodic condition for lateral boundaries
and no-flow at bottom boundary

general, it can be concluded that the effect of correlation length on the average CO2 mass
flux is small.

4.8 Effect of the Boundary Conditions

A series of simulations were carried out to test the influence of the boundary conditions
on CO2 concentration profiles. The boundary conditions examined are: (a) no-flow condi-
tions for lateral and bottom boundaries, (b) periodic conditions for lateral boundaries and
no-flow at bottom boundary, and (c) no-flow conditions for the lateral boundaries and zero
concentration at bottom boundary. The concentration profiles and average CO2 mass flux
at the top boundary obtained from these simulations are shown in Figs. 13 and 14, respec-
tively. The concentration profiles are rather similar for the different boundary conditions
tested, while there is considerable difference in concentration profiles at the lateral posi-

123
50 P. Ranganathan et al.

Fig. 14 Plot of average CO2 mass flux at the top boundary versus dimensionless time for Rayleigh number
of 10,000 with various boundary conditions

tions for periodic boundary conditions. For the no-flow boundary condition, a few gravity
fingers are shown to develop along the lateral boundaries, but the extent of these fingers
is less than the extent of the fingers in the interior of the region. For the periodic bound-
ary condition, the CO2 concentration allows lateral flow, which tends to spread the CO2
plume. The constant concentration condition (c = 0) at the bottom boundary gives an unre-
alistic pattern at the bottom of the domain. This may be due to the fact that CO2 is lost
through diffusion across the bottom boundary. Figure 14 reveals that the CO2 mass flux at
the top boundary with time has nearly the same trend for all boundary conditions. At later
times the average CO2 mass flux is slightly higher for a periodic condition compared to
other boundary conditions. Most probably the CO2 plume spreads more when lateral flow is
admitted.

4.9 Effect of Height and Width of Domain

To investigate the effect of width and height of the flow domain on the development of
instabilities particularly at longer times, we plot the CO2 concentration profiles at the time
when the fingers reach the bottom boundary for the different values of height and width of the
domain. The height and width of the domain examined in the present study are 1.0, 2.0, and
4.0. The concentration profiles of one of the realisations (correlation length of 0.1, standard
deviation of 0.3, and Rayleigh number of 10,000) for various height and width of domain are
shown in Figs. 15 and 16, respectively. It can be observed from Fig. 15 that a few large-scale
dominant fingers are formed due to merging of small-scale fingers and reach the bottom
boundary of the domain. The number of the large-scale dominant fingers decreases as the
height of the domain increases. The effect of the width of the domain on instabilities shows
that the interactions between fingers reduce, and the fingers move independently to reach the
bottom boundary. This trend is more pronounced when the width of domain increases. It is
also observed that the time when fingers reach the bottom boundary increases as the width
of the domain increases.

123
Numerical Simulation of Natural Convection in Heterogeneous... 51

Fig. 15 Effect of the height of the domain on CO2 concentration profiles with log-normal permeability fields
of one realisation (λ = 0.1 and σ = 0.3) for Rayleigh number of 10,000. (a) H = 1 at td = 0.0005, (b)
H = 2 at td = 0.00125, (c) H = 4 at td = 0.0025

5 Conclusions

Density-driven natural convection in heterogeneous porous media, where a CO2 -rich gaseous
phase overlays a liquid saturated medium, was investigated through numerical modeling. Ten
realisations of log-normal permeability fields generated by the sequential gaussian simula-
tion (SGSIM) method were used to do the numerical analyses. The convective motion was
examined in terms of concentration profiles, variation of mixing length, and the average CO2
mass flux at the top boundary as a function of the Rayleigh numbers, the standard deviation
of heterogeneity characteristic, and the correlation length of the permeability distributions.
The following main conclusions can be drawn from this study:

• For various correlation lengths at lower standard deviation, a gravity-induced fingering

pattern was observed; at moderate standard deviations of the log-normal permeability
field, the flow exhibits a channel pattern, which is controlled by the permeability field. At
higher standard deviations the flow was dispersive when the correlation length is small,
whereas a high correlation lengths, the flow was characterized by a channel pattern.
• For various Rayleigh numbers, three different regimes were deduced from the plot of the
mixing length variation with dimensionless time. In the first regime, the mass transfer

123
52 P. Ranganathan et al.

Fig. 16 Effect of the width of the domain on CO2 concentration profiles with log-normal permeability fields
of one realisation (λ = 0.1 and σ = 0.3) for Rayleigh number of 10,000. (a) w = 1 at td = 0.0006, (b) w = 2
at td = 0.00075, (c) w = 4 at td = 0.0001

is dominated by diffusion; in the second regime convection starts and mass transfer is
controlled by convective mixing and in the third regime, the rate of mixing length growth
is lower than that in the second regime, but is nearly equal to that of the first regime. For
higher standard deviations, there are only two regimes, i.e., the first diffusion dominated
regime is absent. The mixing length is only weakly dependent on the correlation length.
• For all values of the Rayleigh numbers, at first the average CO2 mass flux decreases
quickly at earlier times, increases slightly at later times and then levels off to a plateau.
The time at which the CO2 mass flux increases, diminishes for higher Rayleigh numbers.
It was found that at early times and for the lower standard deviation, the average CO2
mass flux decreases rapidly; however, for later times it increases and then decreases. For
higher standard deviations, the average CO2 mass flux decreases with time and reaches
a constant value at later times. We found that the effect of the correlation length on the
average CO2 mass flux is small.
• The effect of boundary condition shows significant changes in CO2 concentration profile
while a little change in CO2 mass flux at later times. With increasing the height of the
domain, a few large-scale dominant fingers are formed due to interactions between fingers.
With increasing the width of the domain, fingers move independently without the interact
each other.

Acknowledgments This study was partly done as part of the CATO II programme. The authors are grateful
to the Dutch Ministry of Economic Affaires and the sponsoring companies for funding. The authors are also
grateful to the anonymous refereed whose comments and suggestions have help improve manuscript.

123
Numerical Simulation of Natural Convection in Heterogeneous... 53

Open Access This article is distributed under the terms of the Creative Commons Attribution License which
permits any use, distribution, and reproduction in any medium, provided the original author(s) and the source
are credited.

References

Barlebo, H.C., Hill, M.C., Rosbjer, D.: Investigating the macrodispersion experiment (MADE) site in
Columbus, Mississippi, using a three-dimensional inverse flow and transport model. Water Resour.
Res. 40, W04211 (2004)
Bear, J.: Dynamics of Fluids in Porous Media. Dover Publications, New York (1972)
Bryant, S.L., Lakshminarasimhan, S., Pope, G.A.: Buoyancy-dominated multiphase flow and its effect on
geological sequestration of CO2 . SPE J. 13, 447–454 (2008)
Deutsch, C.V., Journel, A.G.: GSLIB Geostatistical Software Library and Users Guide. Oxford University
Press, New York (1992)
Elder, J.W.: Transient convection in a porous medium. J. Fluid Mech. 27, 609–623 (1967)
Ennis-King, J., Paterson, L.: Role of convective mixing in the long-term storage of carbon dioxide in deep
saline formations. SPE 84344, 1–12 (2005)
Ennis-King, J., Paterson, L. : Rate of dissolution due to convective mixing in the underground storage of
carbon dioxide. In: Gale, J., Kaya, Y. (eds.) Greenhouse Gas Control Technologies, vol. 1, pp. 507–510.
Elsevier, Amsterdam (2003)
Farajzadeh, R., Salimi, H., Zitha, P.L.J., Bruining, J.: Numerical simulation of density-driven natural convec-
tion in porous media with application for CO2 injection projects. Int. J. Heat Mass Transf. 50, 5054–
5064 (2007)
Farajzadeh, R., Farshbaf Zinati, F., Zitha, P.L.J., Bruining, J.: Density driven natural convection in layered
and anisotropic porous media. In: Proceeding of the 11th European Conference on Mathematics in Oil
Recovery (ECMOR X1), Norway (2008)
Farajzadeh, R., Ranganathan, P., Zitha, P.L.J., Bruining, J.: The effect of heterogeneity on the character of
density-driven natural convection of CO2 overlying a brine layer. Adv. Water Resour. 34, 327–339 (2011)
Firoozabadi, A., Cheng, P.: Prospects for subsurface CO2 sequestration. AIChE J. 56, 1398–1405 (2010)
Gebhart, B., Jaluria, Y., Mahajan, R.L., Sammakia, B.: Buoyancy-Induced Flows and Transport. Hemisphere,
Harper and Row, New York (1988)
Ghesmat, K., Azaiez, J.: Viscous fingering instability in porous media: Effect of anisotropic velocity-dependent
dispersion tensor. Transp. Porous Med. 73, 297–318 (2008)
Green, C., Ennis-King, J.: Effect of vertical heterogeneity on long-term migration of CO2 in saline forma-
tions. Transp. Porous Med. 82, 31–47 (2010)
Guceri, S., Farouk, B.: Numerical solutions in laminar and turbulent natural convection. In: Kakac, S., Aung, W.,
Viskanta, R. (eds.) Natural Convection, Fundamentals and Applications, pp. 615–655. Hemisphere, Bris-
tol, PA (1985)
Han, W.S., Lee, S., Lu, C., McPherson, B.J.: Effect of permeability on CO2 trapping mechanism and buoyancy-
driven CO2 migration in saline formations. Water Resour. Res. 46, 1–20 (2010)
Hassanzadeh, H., Pooladi-Darvish, M., Keith, D.W.: Scaling behavior of convective mixing, with application
to geological storage of CO2 storage. AIChE J. 55, 1121–1131 (2007)
Hidalgo, J.J., Carrera, J.: Effect of dispersion on the onset of convection during CO2 sequestration. J. Fluid
Mech. 640, 441–452 (2009)
Kneafsey, T.J., Pruess, K.: Laboratory flow experiments for visualizing carbon dioxide-induced, density-driven
brine convection. Transp. Porous Med. 82, 123–139 (2010)
LeBlanc, D.R., Garabedian, S.P., Hess, K.M., Gelhar, L.W., Quadri, R.D., Stollenwerk, K.G., Wood,
W.W.: Large-scale natural gradient tracer test in sand and gravel, Cape Cod, Massachusetts. 1. Experi-
mental design and observed tracer movement. Water Resour. Res. 27, 895–910 (1991)
Li, Z., Dong, M., Shirif, E.: Transient natural convection induced by gas diffusion in liquid-saturated vertical
porous columns. Ind. Eng. Chem. Res. 45, 3311–3319 (2006)
Lindeberg, E., Bergmo, P.: The long-term fate of CO2 injected into an aquifer. In: Gale, J., Kaya, Y. (eds.) Green-
house Gas Control Technologies, vol. 1, pp. 489–494. Elsevier, Amsterdam (2003)
Malkovsky, V.I., Pek, A.A.: Onset conditions of free thermal convection of a single-phase fluid in a horizontal
porous layer with depth-dependent permeability. Izvestiya Phys. Solid Earth 35, 990–994 (1999)
McKibbin, R., O’Sullivan, M.J.: Onset of convection in a layered porous medium heated from below. J. Fluid
Mech. 96, 375–393 (1980)
Nield, D.A., Bejan, A.: Convection in Porous Media. Springer, New York (1999)

123
54 P. Ranganathan et al.

Nield, D.A., Simmons, C.T.: A discussion on the effect of heterogeneity on the onset of convection in a porous
medium. Transp. Porous Med. 68, 413–421 (2007)
Nield, D.A., Kuznetsov, A.V., Simmons, C.T.: The effect of strong heterogeneity on the onset of convection
in a porous medium: 2D/3D Localization and spatially correlated random permeability fields. Transp.
Porous Med. 83, 465–477 (2010)
Pau, G.S.H., Bell, J.B., Pruess, K., Almgren, A.S., Lijewski, M.J., Zhang, K.: High resolution simulation
and characterization of density–driven flow in CO2 storage in saline aquifers. Adv. Water Resour. 33,
443–455 (2010)
Prasad, A., Simmons, C.T.: Unstable density-driven flow in heterogeneous porous media: a stochastic study
of the Elder [1967b] “short heater” problem. Water Resour. Res. 39, 1–21 (2003)
Pruess, K., Zhang, K.: Numerical Modeling Studies of the dissolution-diffusion-convection process dur-
ing CO2 storage in saline aquifers. Lawrence Berkeley Laboratory Report LBNL-1243E, Berkeley,
CA (2008)
Rapaka, S., Chen, S., Pawar, R.J., Stauffer, P.H., Zhang, D.: Nonmodal growth of perturbations in density-
driven convection in porous media. J. Fluid Mech. 609, 285–303 (2008)
Rapaka, S., Chen, S., Pawar, R.J., Stauffer, P.H., Zhang, D., Chen, S.: Onset of convection over a transient
base-state in anisotropic and layered porous media. J. Fluid Mech. 641, 227–244 (2009)
Riaz, A., Hesse, A., Tchelepi, M., Orr, F.M.: Onset of convection in a gravitationally unstable diffusive
boundary layer in porous medium. J. Fluid Mech. 548, 87–111 (2006)
Rogerson, A., Meiburg, E.: Numerical simulation of miscible displacement processes in porous media flows
under gravity. Phys. Fluids A 5, 2644–2660 (1993)
Schincariol, R.A., Schwartz, F.W., Mendoza, C.A.: Instabilities in variable density flows: stability and sensi-
tivity analyses for homogeneous and heterogeneous media. Water Resour. Res. 33, 31–41 (1997)
Shinozuka, M., Jan, C.M.: Digital simulation of random processes and its applications. J. Sound Vib. 25,
357–368 (1972)
Simmons, C.T., Narayan, K.A., Wooding, R.A.: On a test case for density-dependent groundwater flow and
solute transport models: the salt lake problem. Water Resour. Res. 35, 3607–3620 (1999)
Simmons, C.T., Fenstemaker, T.R., Sharp, J.M.: Variable-density groundwater flow and solute transport in
heterogeneous porous media: approaches, resolutions and future challenges. J. Contam. Hydrol. 52, 245–
275 (2001)
Sorbie, K.S., Farag, F., Pickup, K.S., Ringrose, P.S., Jensen, J.L.: Flow regimes in miscible displacements in
heterogeneous correlated random fields. SPE Adv. Technol. Ser. 2, 78–87 (1994)
Waggoner, J.R., Castillo, J.L., Lake, L.W.: Simulation of EOR processes in stochastically generated permeable
media. Soc. Pet. Eng. Form. Eval. 7, 173–180 (1992)
Welty, C., Gelhar, L.W.: Stochastic analysis of the effects of fluid density and viscosity variability on macrodis-
persion in heterogeneous porous media. Water Resour. Res. 27, 2061–2075 (1991)
Welty, C., Gelhar, L.W.: Simulation of large-scale transport of variable density and viscosity fluids using a
stochastic mean model. Water Resour. Res. 28, 815–827 (1992)
Zimmerman, W.B., Homsy, G.M.: Viscous fingering in miscible displacement: unification of effects of vis-
cosity contrast, anisotropic dispersion, and velocity dependence of dispersion on nonlinear finger prop-
agation. Phys. Fluids A 4, 348–2359 (1992)

123