J. Fluid Mech. (2006), vol. 548, pp. 87–111.

c 2006 Cambridge University Press 87
doi:10.1017/S0022112005007494 Printed in the United Kingdom

Onset of convection in a gravitationally unstable

diffusive boundary layer in porous media
By A. R I A Z, M. H E S S E, H. A. T C H E L E P I† AND F. M. O R R j r
Department of Petroleum Engineering, Stanford University, Stanford, CA 94305, USA

(Received 30 March 2005 and in revised form 22 July 2005)

We present a linear stability analysis of density-driven miscible flow in porous media

in the context of carbon dioxide sequestration in saline aquifers. Carbon dioxide
dissolution into the underlying brine leads to a local density increase that results
in a gravitational instability. The physical phenomenon is analogous to the thermal
convective instability in a semi-infinite domain, owing to a step change in temperature
at the boundary. The critical time for the onset of convection in such problems has
not been determined accurately by previous studies. We present a solution, based
on the dominant mode of the self-similar diffusion operator, which can accurately
predict the critical time and the associated unstable wavenumber. This approach is
used to explain the instability mechanisms of the critical time and the long-wave
cutoff in a semi-infinite domain. The dominant mode solution, however, is valid only
for a small parameter range. We extend the analysis by employing the quasi-steady-
state approximation (QSSA) which provides accurate solutions in the self-similar
coordinate system. For large times, both the maximum growth rate and the most
dangerous mode decay as t 1/4 . The long-wave and the short-wave cutoff modes scale
as t 1/5 and t 4/5 , respectively. The instability problem is also analysed in the nonlinear
regime by high-accuracy direct numerical simulations. The nonlinear simulations at
short times show good agreement with the linear stability predictions. At later times,
macroscopic fingers display intense nonlinear interactions that significantly influence
both the front propagation speed and the overall mixing rate. A dimensional analysis
for typical aquifers shows that for a permeability variation of 1 −3000 mD, the critical
time can vary from 2000 yrs to about 10 days while the critical wavelength can be
between 200 m and 0.3 m.

1. Introduction
1.1. Carbon dioxide sequestration
Carbon dioxide (CO2 ) sequestration in deep geological formations has been suggested
as a way of reducing greenhouse gas emissions. When CO2 is injected into an aquifer
at a depth greater than 800 m, it forms an immiscible CO2 -rich vapour phase, and a
small amount of CO2 dissolves in the brine (van der Meer 1992). In the temperature
and pressure range encountered in geological CO2 storage, the density of the CO2 -rich
vapour phase is less than the density of the brine, but the density of the brine will
increase with increasing CO2 concentration (Ennis-King & Paterson 2003). Several
mechanisms for trapping of CO2 underground should be provided by any suitable
injection site. The buoyant CO2 -rich vapour is prevented from rising back to the

† Author to whom correspondence should be addressed.

88 A. Riaz, M. Hesse, H. A. Tchelepi and F. M. Orr Jr

CO2 gas phase

z=0 g
CO2 dissolved in brine
z

Brine

Figure 1. Sketch of the CO2 sequestration process in a simple geometry. A CO2 gas phase
accumulates along the impermeable top boundary. It slowly dissolves into the underlying
brine, forming a heavier boundary layer. The resulting gravitational instability leads to the
convective transport of CO2 -saturated brine plumes.

surface primarily by an impermeable cap rock referred to as a hydrodynamic trap

(Bachu, Gunther & Perkins 1994). Another mechanism is dissolution of CO2 into
the brine. The increased density of the CO2 -rich brine prevents the CO2 from rising.
This effect is often referred to as solution trapping (Lindeberg & Wessel-Berg 1997;
Ennis-King & Paterson 2003). We will discuss the effect of density-driven convection
on the effectiveness of this mechanism. Two additional trapping mechanisms, which
will not be considered here, are the residual trapping in which CO2 is immobilized
by capillary snap-off of bubbles of CO2 (Kumar et al. 2004) and mineral trapping
in which geochemical reactions convert CO2 into solid minerals (Hitchon 1996). A
hydrodynamic trap is a prerequisite for any storage site, because it prevents the rise
of the CO2 -rich vapour during the time required for other trapping mechanisms to
immobilize it.
During the injection period, gravity forces will cause the injected CO2 to accumulate
at the top of the reservoir. We model this complex flow process by assuming that
the CO2 forms a layer that is separated from the brine below by a relatively sharp
horizontal interface. Across the interface, CO2 will dissolve into brine to form a
diffusive boundary layer that grows with time (Lindeberg & Wessel-Berg 1997). The
CO2 -rich brine in this boundary layer is heavier than the underlying brine, and
will eventually become unstable, so that fingers of dense CO2 -rich brine propagate
downward and transport the aqueous CO2 away from the interface (see figure 1).
This density-driven convection increases the rate of mass transport from the CO2 -
rich vapour into the brine, and therefore can reduce the time required for the total
dissolution of a CO2 -rich vapour plume by several orders of magnitude (Ennis-King &
Paterson 2003).
Numerical simulations by Lindeberg & Bergmo (2003) of the full two-phase problem
have shown that the interface between the CO2 -rich vapour and brine remains
relatively sharp and is not deflected by the density fingering in the underlying brine,
as is shown schematically in figure 1. Both Lindeberg & Bergmo (2003) and Ennis-
King & Paterson (2003) show that a local equilibrium between the CO2 -rich vapour
and the brine can be assumed at the interface. We assume therefore that the CO2 -rich
 Density-driven miscible flow in porous media 89
vapour layer acts as a horizontal upper boundary, with a constant CO2 concentration.
These authors propose that the gravitational instability can be studied in the context
of a finite domain, bounded at the top by a constant concentration boundary, using
the Boussinesq approximation. This problem is analogous to thermal convection in
a porous medium with insulated boundaries, that is rapidly heated from below, or
cooled from the top, at a fixed temperature. Elder (1967, 1968) was among the first to
investigate this problem in detail (Diersch & Kolditz 2002). His results suggest that
the diffusive layer at the boundary becomes unstable only after an initial period of
decaying perturbations. After this critical time, plumes or fingers of hot fluid rise into
the porous medium.

1.2. Previous work on the stability of diffusive boundary layers

The stability of a concentration base profile to small perturbations depends on the
nature of the base state. For steady base states, linear stability theory leads to an
algebraic eigenvalue problem (Drazin & Reid 1981; Nield & Bejan 1999). The diffusive
boundary layer, however, is a time-dependent base state. A common approach to such
stability problems is the use of frozen time coefficients, the so-called quasi-steady-
state-approximation (QSSA), (Lick 1964; Robinson 1976). The QSSA is valid, if the
growth rate of perturbations is large compared to the growth rate of the diffusive
boundary layer. Initially, when the boundary layer grows rapidly, the QSSA is not
justified, but it becomes valid for large times, when the base state changes relatively
slowly (cf. Tan & Homsy 1987; Riaz & Meiburg 2003a, b). Gresho & Sani (1971)
have shown that the QSSA is not valid for the onset of the instability of a thermal
boundary layer in a viscous incompressible fluid.
In order to avoid using the QSSA for small times, several authors have solved
the initial-value problem (IVP) numerically for a large number of random initial
conditions. However, the solution of the initial-value problem is sensitive to the
particular initial conditions, and therefore not unique (Foster 1965). Foster (1965,
1968), Gresho & Sani (1971) and Jhavery & Homsy (1982) solved the IVP for a
thermal boundary layer in a viscous incompressible fluid. Caltagirone (1980) and
Kaviany (1984) investigated the IVP for a thermal boundary layer in a porous
medium due to a step change in temperature. Ennis-King & Paterson (2003) extended
the analysis of Caltagirone (1980) to anisotropic porous media, and applied it to the
CO2 -sequestration problem. All of these investigations give a wide range of critical
times, depending on the particular method of measurement of the growth rate.
The stability of time-dependent base states can be investigated by energy methods
(Homsy 1973; Caltagirone 1980). Energy methods give a lower bound for the onset
of the instability, but give no information about the growth rate and wavenumber of
the most dangerous disturbance.
For unbounded domains, Pritchard (2004) and Ben, Demekhin & Chang (2002)
argue that a fundamental limitation on the accuracy of the solution at short
times is the assumption of decoupled normal modes in the streamwise direction.
Since the disturbances are confined within a narrow diffusive zone, the global
Fourier eigenfunctions of the diffusion operator do not provide an optimal basis
for streamwise perturbations (Pego & Weinstein 1994). Ben et al. (2002) note that
the Hermite polynomial-based discrete eigenspectrum of the self-similar diffusion
operator is the natural basis for streamwise perturbations in an unbounded domain.
This eigenspectrum is the solution of the stability problem for zero wavenumber. The
first mode is neutral. It has zero growth rate, whereas the rest of the modes decay with
time. Therefore, the first mode becomes dominant after a relatively short time. The
90 A. Riaz, M. Hesse, H. A. Tchelepi and F. M. Orr Jr
perturbation dynamics can then be projected onto this dominant mode, to increase
the growth rate from zero to positive values, for small wavenumbers.
A similar approach is adopted here for the stability analysis in the semi-
infinite domain. We use the self-similar transform of the diffusion operator in
the inhomogeneous direction with localized eigenfunctions. The zero-wavenumber
solution shows that the first mode of the self-similar diffusion operator decays with
time, while the rest of the spectrum decays more rapidly. We use this dominant first
mode to capture the perturbation dynamics for larger wavenumbers that at later times
shift the growth rate from negative to positive values. We show that in contrast to the
long-wave instability in an unbounded domain, a critical time, critical wavenumber,
and a long-wavelength cutoff characterize the stability problem in the semi-infinite
domain.
It is important to note that the dominant mode solution becomes inaccurate for
large times and large wavenumbers. For these cases we use the QSSA in self-similar
coordinates. We show by comparing with the initial-value problem that the QSSA
in this context gives reasonable results. Compared to poor accuracy in the original
coordinates, the success of both the QSSA and the IVP in the self-similar coordinate
system is due to localized basis functions in the streamwise direction.
In order to analyse the long-term evolution of the unstable modes predicted by
the linear stability theory, we carry out high-accuracy direct numerical simulations
(DNS) using a vorticity-based formulation (Tan & Homsy 1988). Voss & Sousa
(1987) have also carried out numerical simulations of density-driven flows. We use
the methodology developed by Ruith & Meiburg (2000) using Fourier–Galerkin
decomposition in the lateral direction and compact finite differences in the streamwise
direction to solve for the velocity field. For time integration we use a fourth-order
Runge–Kutta method. The nonlinear simulations are validated by comparing with
the linear stability analysis. Excellent agreement is observed at early times for growth
rates associated with individual wavenumbers. The long-time nonlinear behaviour is
characterized by complex fingering interactions.

2. Linear stability analysis

2.1. Governing equations
The equations describing the Boussinesq-flow in a horizontal porous layer, where
gravity points in the positive z-direction, are:
K
u=− (∇P − ρg ẑ), (2.1)
µ
∂C
φ = −u · ∇C + φD∇2 C, (2.2)
∂t
∇ · u = 0, (2.3)
ρ = ρ0 + ρC. (2.4)
C is the concentration of the heavier fluid normalized to unity, u = (u, w) is the
Darcy velocity. K is the permeability, D is the diffusion coefficient, φ is the porosity,
and µ is the viscosity. The unit vector in the direction of the acceleration due to gravity
is ẑ. Density ρ is specified as a linear function of concentration and ρ0 is the density
of the lighter fluid. The initial conditions are u(x, z, t = 0) = 0 and C(x, z, t = 0) = 0,
and the boundary conditions are given by
w(x, z = 0, t) = 0, w(x, z = H, t) = 0,
 Density-driven miscible flow in porous media 91


∂C


C(x, z = 0, t) = 1, = 0.
∂z
x,z=H,t
We choose the domain depth H as the length scale, figure 1, and the buoyancy
velocity U as the velocity scale. We define the following scaling relations,
Kρg
U= , (2.5)
µ
ρ ∗ = ρ1 − ρ0 = ρ, (2.6)
µU H
P∗ = = ρgH, (2.7)
K
φH φµH
t∗ = = . (2.8)
U Kρg
Based upon these scaling relationships the non-dimensional equations can be
expressed as,
∇ · u = 0, (2.9)
u = −(∇P  − C ẑ), (2.10)
∂C 1 2
= −u · ∇C + ∇ C. (2.11)
∂t Ra
A modified pressure P  = ∇(P − ρ0 gẑ/ρgH ) appears in the above equations. The
Rayleigh number Ra is the only non-dimensional parameter in the problem defined
as,
UH KρgH
Ra = = . (2.12)
φD φDµ
The boundary conditions are,
w(x, z = 0, t) = 0, w(x, z = 1, t) = 0,


∂C


C(x, z = 0, t) = 1, = 0.
∂z
x,z=1,t

We eliminate pressure by taking the curl of (2.10) and substitute the transverse
velocity from the continuity equation. The streamwise velocity, w, and concentration
are decomposed into the base state and perturbation components,
(w, c)(x, z, t) = (wo , Co )(z) + (ŵ, ĉ)(z, t)eikx . (2.13)
The base velocity wo is zero. Although, strictly speaking, a streamwise velocity does
not exist since wo = 0, we use this nomenclature to describe the perturbation velocity
component in the direction of acceleration due to gravity. The base concentration, Co ,
is the solution of (2.11) with u = 0 and ∂/∂x = ∂ 2 /∂x 2 = 0. The perturbation variables
are decomposed into eigenfunctions, which depend on time and the streamwise
coordinate, and normal modes in the transverse x-direction with wavenumber k. The
linearized perturbation equations can then be expressed as,
 2 
∂
− k ŵ = −k 2 ĉ,
2
(2.14)
∂z2
 2 
∂ ĉ 1 ∂ ∂Co
− − k 2
ĉ = − ŵ, (2.15)
∂t Ra ∂z2 ∂z
92 A. Riaz, M. Hesse, H. A. Tchelepi and F. M. Orr Jr
with boundary conditions,

ĉ(z = 0, t) = ŵ(z = 0, t) = 0, (2.16)

∂ ĉ


= ŵ(z = 1, t) = 0. (2.17)
∂z
z=1,t

Equation (2.11) admits a streamwise one-dimensional base-state solution, Co , which

is,

4 1
∞
Co (z, t) = 1 − sin ((n − 1/2)πz) exp(−(n − 1/2)2 π2 t/Ra).
π n=1 2n − 1

The length δ(t) over which Co is non-zero √ is the so-called ‘penetration depth’ of
the diffusive boundary layer. For δ ∝ 4t/Ra  1 the domain can be considered
semi-infinite in the positive z-direction, and the base state is given by,
  
Ra
Co (z, t) = 1 − erf z (z ∈ (0, ∞)), (2.18)
4t

with the boundary conditions,

Co (z = 0, t) = 1, (2.19)
Co (z → ∞, t) → 0. (2.20)

The problem has now been redefined as the instability of a diffusive boundary layer
in a semi-infinite domain. The parameter range over which the √ results are valid, for
the original layer geometry in a finite domain, is given by δ ∝ 4t/Ra  1. Note that
the semi-infinite domain does not impose any external length scale on the problem.
One of the internal length scales in the problem is the time-dependent penetration
depth δ(t). A Rayleigh number based on δ(t) is itself time dependent, such that a
critical Rayleigh number, Rac = Ra(δ(tc )), is merely a function of the critical time,
tc , at which the boundary layer becomes unstable. On the other hand, the Rayleigh
number can also be scaled out of the equations by specifying the length scale as a
ratio of diffusion to buoyancy velocity, H = D/U . In this case, the critical time, tc ,
is the only criterion for the onset of instability. For our analysis, we choose to work
with the latter, but retain the Rayleigh number, with an arbitrary length scale H ,
for convenience in comparing the linear stability analysis with the nonlinear results
discussed in § 3. In the discussion of the dimensional results (§ 4), we show that the
imposed length scale cancels out, so that the choice of length scale is arbitrary.
The perturbation equations, (2.14)–(2.17), can be solved in a straightforward manner
by using the quasi-steady-state approximation (QSSA). Earlier investigations have
shown that the QSSA gives reasonable results for long times. A fundamental problem
with such an approach is that the concentration eigenfunctions are localized in the
boundary layer while the eigenfunctions of the operator ∂ 2 /∂z2 , z ∈ (0, ∞) are global
modes. Hence, they do not provide an appropriate basis for streamwise perturbations
(cf. Chang, Demekhin & Kalaidin 1998). Therefore, in the context of the initial-value
problem (IVP), a finite time is required by the global eigenfunctions to represent
accurately the localized structure of streamwise perturbations. Since the current
problem can have a small critical time, the QSSA as such is ill-suited to the task of
resolving the early-time behaviour.
 Density-driven miscible flow in porous media 93
2.2. Localized eigenmodes of the diffusion operator
In a semi-infinite domain, we transform the perturbation equations such that the
eigenfunctions associated with the streamwise diffusion operator are localized around
the base-concentration front. The objective is to achieve considerable improvement in
accuracy at small times, even with the QSSA. Following
√ a coordinate transformation
to the similarity variable of the base state ξ = z Ra/4t, the base state and the
perturbation equations can be expressed as,
Co (ξ ) = 1 − erf(ξ ), (2.21)
 
Ra ∂ 2
− k ŵ = −k 2 ĉ,
2
(2.22)
4t ∂ξ 2
  
∂ ĉ 1 1 ∂ 2 ξ ∂ k2t Ra
− 2
+ − ĉ = exp(−ξ 2 )ŵ, (2.23)
∂t t 4 ∂ξ 2 ∂ξ Ra πt
with boundary conditions,
ĉ(ξ = 0, t) = ŵ(ξ = 0, t) = 0, (2.24)
ĉ(ξ = ∞, t) = ŵ(ξ = ∞, t) = 0. (2.25)
Note that the self-similarity applies only to the base concentration. The amplitude and
the spatial structure of perturbations are time dependent. The streamwise operator
of the perturbation concentration in the transformed coordinate, ξ , is
1 ∂2 ξ ∂
L= 2
+ (ξ ∈ (0, ∞)). (2.26)
4 ∂ξ 2 ∂ξ
We expand the perturbation concentration as,

∞
ĉ(ξ, t) = An (t)φn (ξ ), (2.27)
n=1

with
Lφn = λn φn (ξ ) = λn e−ξ Hn (ξ )
2
(n = 1, 2, 3, . . .). (2.28)
The eigenfunctions φn of L are the Hermite polynomials Hn (ξ ) in a semi-
infinite domain with weight function exp(−ξ 2 ) (Robinson 1976). The associated
eigenvalues are λn = − n for n = 1, 2, . . . , and the dominant mode of the perturbation
concentration is φ1 = ξ exp(−ξ 2 ). These eigenfunctions, being localized around the
base state, provide an optimal basis for streamwise perturbations in the semi-infinite
domain (Pego & Weinstein 1994).
We note from (2.22) and (2.23) that for k = 0 and using (2.27), the perturbation
amplitude can be expressed as
dAn
= λn An . (2.29)
dt
All modes decay as t −n for k = 0, i.e. the flow is stable in the long-wave limit. The
perturbation eigenfunction related to the largest eigenvalue, φ1 = ξ exp(−ξ 2 ), decays
as 1/t.
We solve an initial-value problem with (2.21)–(2.23) to show that the preferred
mode of the streamwise perturbations for k = 0 is φ1 . Figure 2(a) shows that the
concentration eigenfunction resolves into the dominant mode of L, i.e. ξ exp(−ξ 2 ),
in a very short time, (t = O(1/Ra)), for any given set of white-noise initial conditions.
We show in § 2.4 that the time required for convergence to the first eigenfunction is
94 A. Riaz, M. Hesse, H. A. Tchelepi and F. M. Orr Jr
(a) (b)
1.0
t= 10–5 t = 0.1
10–4 0.5
1.0 10–3 1.0
0.5

ĉ 0.5 ĉ 0

0 –0.5

–1.0
0 1 2 3 4 5 0 0.2 0.4 0.6 0.8 1.0
ξ z

Figure 2. Disturbance concentration profile for Ra = 500 and k = 0 at different times, obtained
from, (a) the IVP in the (ξ, t)-coordinate system with (2.21)–(2.23) and (b) the IVP in the
(z, t)-coordinate system with (2.14)–(2.18). Random initial conditions are used for both cases.
Plot (a) shows that the concentration profile resolves into the correct solution within a very
short time, starting from any given set of white-noise initial conditions. For the (z, t)-coordinate
system on the other hand, plot (b), convergence is much slower and is strongly dependent on
initial conditions.

several orders of magnitude smaller than the critical time for the onset of instability.
The IVP in (ξ, t)-coordinates can therefore yield accurate results for short times.
In order to highlight this phenomenon, figure 2(b) plots the perturbation
eigenfunction in (z, t)-coordinates, obtained by solving the IVP with (2.14)–(2.18).
Although the white-noise initial conditions converge to ĉ ∼ z exp(−z2 Ra/4t), the time
required for convergence is several orders of magnitude larger, compared to that in
the self-similar coordinates, as shown in figure 2(a). Hence, the small-time dynamics
in the (z, t)-coordinate system are obscured during the time it takes for the random
initial perturbations to resolve into the dominant mode. This explains why earlier
investigations using the IVP in the (z, t)-coordinate system did not produce accurate
results. For the (ξ, t)-coordinate system on the other hand, the localized eigenfunctions
of L converge rapidly to the exact solution, thereby giving an accurate disturbance
growth rate at small times.

2.3. Dominant mode solution

Robinson (1976) used a one-term approximation to the solutions in his QSSA analysis
in (z, t)-coordinates, and found that the error is small when compared to expansions
using many terms. In the (ξ, t)-coordinates, such a one-term approximation is expected
to be even better, because we expand in the eigenfunctions of the streamwise operator.
We use the leading-order approximation for ĉ from (2.27), substitute it into the IVP
(equation (2.23)) and integrate across the domain to obtain,

dA1 A1 A1 k 2 Ra
=− − + exp(−ξ 2 )ŵ , (2.30)
dt t Ra πt
 Density-driven miscible flow in porous media 95
where,
 ∞
exp(−ξ 2 )ŵ dξ
exp(−ξ 2 )ŵ = 0 ∞ . (2.31)
ξ exp(−ξ 2 ) dξ
0
Again, using only the leading-order estimate of ĉ to solve for ŵ, we obtain
 2 
∂ 4tk 2 4tk 2
− ŵ = − A1 ξ exp(−ξ 2 ). (2.32)
∂ξ 2 Ra Ra
This can be solved analytically for ŵ
  ξ 
t
ŵ = − A1 k exp(2k t/Ra) ξ exp(−2kx t/Ra − x 2 ) dx − B1
Ra 0
 ξ 
− exp(−2k t/Ra) ξ exp(2kx t/Ra − x ) dx − B2
2
, (2.33)
0

where
 ξ
ξ exp(−2kx t/Ra − x 2 )dx = − 12 exp(−2kξ t/Ra − ξ 2 )
0
   
πt t
− 12 k exp(k 2 ) erf ξ + k , (2.34)
Ra Ra
and
 ξ
ξ exp(2kx t/Ra − x 2 ) dx = − 12 exp(2kξ t/Ra − ξ 2 )
0
   
πt t
+ 12 k exp(k 2 ) erf ξ − k . (2.35)
Ra Ra
The constants B1 and B2 are obtained by satisfying the boundary conditions ŵ = 0
at ξ = 0, ∞. Using this value of ŵ(ξ, t), we can then solve numerically the integral
in (2.31) to obtain the amplitude equation for the dominant mode with growth rate σ ,
dA1
= σ (t; k) A1 . (2.36)
dt
Equation (2.36) shows that the perturbations grow exponentially. The growth rate
σ (t; k) is obtained without the QSSA and is expected to yield accurate results.
It is interesting to compare the above development for the semi-infinite case with
that of the infinite case analysed by Ben et al. (2002). The eigenfunctions of the
self-similar operator in the latter case are the full range of Hermite polynomials.
The associated eigenvalues are λn = − n/2 for n = 0, 1, 2, . . . , resulting in a neutral
mode for the zeroth eigenvalue. In a bounded domain on the other hand, the self-
similar operator has eigenfunctions based upon only those Hermite polynomials
which satisfy the boundary condition ĉ(ξ = 0, t) = 0. The associated eigenvalues are
−1, −2, . . .; therefore, a neutral mode is not present.
The presence of a zero eigenvalue in the unbounded case implies a long-
wave instability (σ = 0) for k = 0 such that the flow is always unstable for small
wavenumbers. For the semi-infinite domain on the other hand, the dominant mode
decays as t −1 , hence σ = − 1/t for k = 0, i.e. the flow is always stable for small
96 A. Riaz, M. Hesse, H. A. Tchelepi and F. M. Orr Jr

2 t=1

t = tc = 0.34
0

–2
σ t = 0.05
–4

–6

IVP
–8 Dominant mode

0 20 40 60 80
k

Figure 3. Growth rate vs. wavenumber curves for Ra = 500 computed by the dominant-mode
method and the IVP. The flow is stable for small times. The growth rate increases with time
to become positive at the critical time tc and the critical wavenumber kc . Flow instability
increases with time beyond tc with both short-wave and long-wave cutoffs. Comparison with
the IVP shows exact agreement for small times and small wavenumbers.

wavenumbers. A long-wave cutoff therefore exists along with a critical time at which
the flow becomes unstable.
The existence of a critical time as a function of the Rayleigh number has been
noted previously. However, its exact value as well as the fundamental mechanism
have not been given explicitly. Consider the growth rate,

1 k2 k
σ (t; k) = − − + √ F (t; k), (2.37)
t Ra π
where F (t; k) is computed numerically from (2.31). Since F (t; k) > 0, the stabilizing
effects come from the first two negative terms on the right-hand side. The first term,
−1/t, which is due to the non-zero eigenvalue of the dominant mode of L, ensures
that σ < 0 for small times. In physical terms, the flow can become unstable only
when the perturbations grow at a rate faster than the decay rate of the first mode
of L.
It is important to note that selecting only the first mode to capture the perturbation
dynamics cannot be accurate for the entire range of length and time scales of interest,
in view of the absence of a neutral mode of L. We show in § 2.4 that the dominant
mode
√ solution gives exact results in comparison with the IVP only for small values of
k t/Ra. Therefore, we√use the QSSA for (2.21)–(2.25) to solve for the growth rates
for larger values of k t/Ra. We also show that when the QSSA is used with the
governing equations in the self-similar coordinates, accurate results are obtained.
2.4. Results
The growth rate vs. wavenumber curves given by both the dominant-mode method
and the numerical solution of the IVP are shown in figure 3 for different times. High-
accuracy numerical simulations of the IVP were carried out for each wavenumber
k on a fine computational grid, by using standard methods for one-dimensional
problems. The evolution of the maximum value of either the concentration or the
 Density-driven miscible flow in porous media 97
2.5

2.0
k = 8π

1.5

Growth rate
1.0 k = 4π

0.5

IVP
0 QSSA
Dominant mode
–0.5
0 1 2 3 4
t

Figure 4. Comparison of the dominant-mode method and the QSSA with the initial-value
problem for two perturbation wavenumbers at Ra = 500. The dominant-mode solution gives
exact results for small times, but becomes inaccurate for later times particularly for large
wavenumbers. The QSSA on the other hand is reasonably accurate for all times.

velocity eigenfunction forms the basis of the growth rate plotted at a particular time
for each wavenumber. Completely stable behaviour is indicated by σ < 0 for all
wavenumbers at early time. A critical time, tc , is also shown when the growth rate
just becomes positive at a critical wavenumber kc . At larger times, the stability curve
displays a maximum growth rate at a corresponding most dangerous wavenumber,
along with both a long-wave and a short-wave cutoff. Comparison of the dominant-
mode solution √ with the IVP shows exact agreement for all wavenumbers at small
times when k t/Ra is small. For longer times, the growth rate begins to deviate from
the IVP result at larger wavenumbers. However, the critical time and the long-wave
cutoff are computed exactly by the dominant-mode solution.
√Since the dominant-mode solution does not give accurate results for large values of
k t/Ra, we use the QSSA in the self-similar coordinates to compute the growth rates.
Figure 4 compares the growth rate as a function of time obtained from the IVP to that
computed from the dominant-mode method and the QSSA, for two wavenumbers.
The dominant-mode method again gives exact results for small times, but deviates
from the IVP solution for large times, particularly for the larger wavenumber. The
QSSA, on the other hand, gives reasonably accurate results for all times. This clearly
shows that the QSSA in the self-similar coordinates can be employed to obtain reliable
results even for short times.
Figures 5(a) and5(b), respectively, show the critical time and the critical wavenumber
as a function of the Rayleigh number computed by the dominant-mode method. The
critical time varies as Ra −1 while the critical wavenumber scales linearly with Ra.
Similar scalings have been obtained by Caltagirone (1980), but our analysis eliminates
the effect of the initial condition so that we obtain the following relationships for the
critical time tc ≈ 146/Ra
√ and the critical wavenumber kc ≈ 0.07 Ra. These relationships
apply only when 4t/Ra  1. They are not applicable when either Ra is small or t
large enough to violate this condition.
The maximum growth rate as a function of time and the corresponding most
dangerous mode for various Ra, computed by the QSSA, are shown respectively in
98 A. Riaz, M. Hesse, H. A. Tchelepi and F. M. Orr Jr
(a) (b)
101 103

Critical wavenumber, kc
102
100
Critical time, tc

101
10–1

100

10–2
101 102 103 104 101 102 103 104
Ra Ra

Figure 5. (a) Critical time vs. Ra. (b) Critical wavenumber vs. Ra, given by the dominant-mode
solution. The critical time varies as the 1/Ra while the critical wavenumber scales linearly
with Ra.

(a) (b)
101 200
150

100

50
σmax

kmax

100

Ra = 500
1000
2000

10–1 20
10–1 100 101 10–1 100 101
t t

Figure 6. (a) Maximum growth rate σmax as a function of time, and (b) the most dangerous
wave-number kmax as a function of time, for different Rayleigh numbers. All the results are
obtained using the QSSA.

figures 6(a) and 6(b). The maximum growth rate, σmax , increases rapidly at early time
beyond tc , reaches a maximum value and the late-time evolution shows ≈ t 1/4 decay.
The most dangerous wavenumber kmax also displays ≈ t 1/4 scaling.
The evolution of the longwave, kl , and the short-wave cutoff, ks , is plotted in
figure 7. The former decays as ≈ t 1/5 , whereas the latter decays much faster as ≈ t 4/5
at long times.

3. Direct numerical simulations

3.1. Numerical method and validation of the numerical solution
Nonlinear fingering dynamics govern the long-term flow behaviour. We solve
the nonlinear problem with a high-accuracy vorticity-based method proposed by
 Density-driven miscible flow in porous media 99
(a) (b)
103 250
200
Ra = 500
1000 150
Longwave cutoff, kL

Shortwave cutoff, kS
2000
102
100

101 50

100 20
10–2 10–1 100 101 10–1 100 101
t t

Figure 7. (a) The long-wave cutoff wavenumber kL as a function of time for different
Rayleigh numbers, computed by the dominant-mode method. kL behaves similarly for all Ra
and varies approximately linearly with time. (b) The short-wave cutoff kS computed by the
QSSA.

Ruith & Meiburg (2000). This method has been employed successfully to obtain
highly accurate results for various miscible-flow problems in porous media (cf. Camhi,
Ruith & Meiburg 2000; Riaz & Meiburg 2003a, b). High-resolution finite-difference
simulations have been reported by Otero et al. (2004), for high-Rayleigh-number
convection in a porous layer of finite depth. They compare the numerical results with
analytical heat-flow estimates. We will compare our numerical simulations with the
results from the linear stability analysis.
The governing equations used for direct numerical simulations are (2.10) and (2.11).
We use a vorticity formulation for (2.10) to eliminate pressure. By taking the curl of
(2.10) we obtain,
∂C
ω=− = −∇2 ψ, (3.1)
∂x
where ω is the vorticity and ψ is the streamfunction which is related to velocity as
∂ψ
w= , (3.2)
∂x
∂ψ
u=− . (3.3)
∂z
Periodicity is assumed in the transverse x-direction, for the vorticity and
streamfunction, while symmetry conditions are used for concentration. The boundary
conditions are,
∂C
C(z = 0, x, t) = 1, (z = 1, x, t) = 0, (3.4)
∂z
∂C ∂C
(z, x = 0, t) = 0, (z, x = A, t) = 0, (3.5)
∂x ∂x
w(z = 0, x, t) = 0, w(z = 1, x, t) = 0, (3.6)
u(z, x = 0, t) = 0, u(z, x = A, t) = 0. (3.7)
The aspect ratio A = L/H , where L is the lateral extent of the computational domain
and H is the depth (figure 1). The streamfunction ψ = 0 on all boundaries while ω = 0
100 A. Riaz, M. Hesse, H. A. Tchelepi and F. M. Orr Jr
at x = 0 and x = A. Boundary conditions for vorticity at z = 0, 1 can be obtained from
(3.1).
We solve the Poisson equation, (3.1), by expanding ω and ψ in Fourier modes
in the x-direction. We then solve the resulting ODE for the decoupled z-direction
eigenfunctions with sixth-order compact finite differences. For details see Riaz &
Meiburg (2003). The velocities are then computed from (3.2) and (3.3) where the
derivatives of ψ are evaluated with sixth-order compact finite differences. Time
integration of (2.11) is carried out using a standard fourth-order Runge–Kutta method
where all the spatial derivatives are again evaluated with sixth-order compact finite
differences. The resulting numerical scheme resolves accurately all relevant length
and time scales. The initial condition for the concentration is given by (2.18) with a
starting time of t = 0.2. The initial condition for velocity is w = u = 0.
The explicit nature of time integration imposes a strict limit on the time steps.
Although stable time steps are given by the CFL condition, we use even smaller
time steps, of the order of 10−5 , to ensure accuracy. Spatial resolution of the
computational grid ranges from 512 × 512 grid points for small-Rayleigh-number
cases to 2048 × 2048 grid points for larger-Rayleigh-number cases. These fine spatial
and temporal resolutions produce converged results. Appropriate grid spacing for
different parameter combinations is obtained by consideration of the cutoff mode
provided by the linear stability analysis. The grid spacing is chosen such that it is
smaller than the cutoff wavelength. Additionally, the divergence of the velocity field
is checked throughout to ensure exact mass conservation for a given grid spacing.
Validation of the numerical simulations was performed by comparing the growth
rates with those obtained from the linear stability analysis. The numerical simulations
are perturbed with pure sinusoidal modes in the transverse direction, superimposed
on the initial concentration profile. The growth rate is measured using the norm of
vorticity, defined as
 1 A
ω(t) = ω(z, x, t) dx dz. (3.8)
0 0
The growth rate for the numerical simulations is then defined as
 
1 ω(t)
σDNS = ln . (3.9)
t ω(t − t)
Figure 8 shows that the growth rates for the two cases as a function of time for
two wavenumbers, although not exactly equal, are in good agreement. The reason for
this small discrepancy can be related to the two-dimensional nature of perturbations
in the numerical simulations which show a slightly higher rate of growth than the
one-dimensional initial-value problem.
3.2. The flow structure for Ra = 4000
We begin our discussion of the nonlinear dynamics by analysing the concentration
contours for Ra = 4000. An aspect ratio, A = 1 will be used throughout unless
noted otherwise. In order to observe the nonlinear behaviour, we perturb the initial
concentration with white noise. The initial disturbances are localized within the
diffusive zone to avoid unphysical conditions of C > 1 or C < 0. The wavelength
selection mechanism is largely independent of the amplitude and the particular values
of the random initial perturbations.
Figure 9 presents the concentration contours at different times. This simulation was
carried out with 1024 × 1024 grid points. At an early time of t = 1, figure 9(a) shows
 Density-driven miscible flow in porous media 101
2.5

k = 8π
2.0

1.5

Growth rate
1.0 k = 4π

0.5

Initial-value problem
0
DNS

–0.5
0 1 2 3 4
t

Figure 8. Comparison between the nonlinear direct numerical simulations and the
initial-value problem in terms of the growth rates of two different perturbation wavenumbers.
The two sets of results are in close agreement.

a multitude of competing fingers. The number of fingers is about 24, which is close
to 22, predicted by the linear stability analysis. In § 3.3, we will precisely quantify the
wavenumber for nonlinear simulations and compare with the linear stability results.
Figure 9(b) shows the concentration contours at a later time of t = 1.8. The number
of fingers is much smaller than that at t = 1. This increase in the finger wavelength
is primarily due to vigorous nonlinear interactions such as merging and shielding.
Fingering instability in this case is reminiscent of viscous and gravitational instability
of displacement type flows in porous media (Zimmerman & Homsy 1992; Tchelepi &
Orr 1994; Manickam & Homsy 1995). The driving force for instability, i.e. the density
gradient, is weakened progressively as the fingers move away from the top boundary.
In addition to the diffusive spreading, nonlinear finger interactions tend to further
smooth out the concentration gradients of a large number of competing fingers.
Consequently, at a later time of t = 2.3, shown in figure 9(c), some of the smaller
fingers disappear owing to diffusive smearing while others merge to form large-scale
structures. These larger fingers then travel more or less independently with relatively
less interaction with neighbouring fingers. This trend continues for later times, as
shown in figure 9(d) for t = 3.8. Note that at later times, large-scale fingers are
connected to the thin diffusive boundary layer at discrete locations, which serve as
feeding sites of high-density fluid for the fingers.
Fingering dynamics are revealed by analysis of the vorticity profile that generates
the rotational flow responsible for driving the fingers. Figure 10 plots the vorticity field
contours for Ra = 4000 at t = 2.3. The corresponding concentration field is shown in
figure 9(c). Superimposed on the vorticity field in figure 10 are the streamlines showing
the negative (dashed lines) and positive (solid lines) circulation paths. The vorticity
field displays a dipole structure of negative (light) and positive (dark) vorticity pairs.
Very high vorticity pairs are concentrated at the root of the fingers which act as
feeding sites for the finger portions away from the top boundary. The corresponding
streamlines show how the fluid is drawn from the sides to flow laterally along the
top boundary layer and down through the high-vorticity regions at the finger roots.
102 A. Riaz, M. Hesse, H. A. Tchelepi and F. M. Orr Jr
(a) 0

z 0.1

0.2
0 0.2 0.4 0.6 0.8 1.0

(b) 0

z 0.1

0.2
0 0.2 0.4 0.6 0.8 1.0

(c) 0

0.1
z
0.2

0.3
0 0.2 0.4 0.6 0.8 1.0

(d) 0

0.1

0.2
z
0.3

0.4

0.5
0 0.2 0.4 0.6 0.8 1.0
x

Figure 9. Concentration contours at different times for Ra = 4000. (a) t = 1, (b) t = 1.8,
(c) t = 2.3 and (d) t = 3.8 A large number of fingers, consistent with the linear stability
analysis, develop initially. Nonlinear interactions rapidly reduce the number of fingers and
give rise to large-scale structures at later times. The fingers at later times are connected to the
top boundary at discrete locations. These connections act as feeding sites of the high-density
fluid to the convecting fingers below.

Although the highest vorticity magnitude occurs at the finger roots, the tips of the
fingers also display a moderate accumulation of vorticity.

3.3. Dominant wavenumber of nonlinear flows

As noted in § 3.2, the early-time wavelength developed by the nonlinear simulations
is in close agreement with that predicted by the linear stability analysis. In order to
compare the time evolution of the preferred mode of the nonlinear flow, we define a
 Density-driven miscible flow in porous media 103
0

0.1
z
0.2

0.3
0 0.2 0.4 0.6 0.8 1.0
x

Figure 10. Vorticity contours in the background with overlapping streamlines for Ra = 4000,
t = 2.3. Corresponding concentration contours are shown in figure 9(c). The vorticity field has
a dipole structure that drives the high-density fluid through the fingers. The streamlines show
how the fluid travels laterally and then descends through the isolated feeding sites for the
fingers. The direction of fluid circulation is shown by the arrows.

dominant mode as,

 K
kE(k) dk
n̂ = 0 K
, (3.10)
E(k)dk
0
where n̂ is the dominant wavenumber, k is the Fourier mode and E(k) is the energy
spectrum associated with the Fourier transform of the vorticity field. We compute
E(k) as,
 x  1 
E(k; t) = ω(z, x, t)dz e−i kx dx (3.11)
0 0
The dominant mode n̂ is an integral measure of the transverse perturbation
spectrum. We have observed the spectrum to be highly localized around the high-
energy modes. The magnitude of n̂ is a reasonable approximation of the high-energy
modes in the spectrum. We have also observed that the energy spectrum is only weakly
dependent on the random disturbances introduced as initial conditions. Figure 11
shows the evolution of n̂ for various Rayleigh numbers. The most dangerous mode
kmax from the linear stability analysis is also shown for comparison. We observe a
good agreement of n̂ with kmax for short times. The onset of the nonlinear regime leads
to the deviation of the dominant mode away from the most dangerous mode. This
deviation from kmax is stronger and occurs earlier for higher values of Ra, indicating
a stronger influence of nonlinearity for larger Rayleigh numbers.
3.4. Influence of the Rayleigh number
Figure 11 shows that the nonlinear behaviour is strongly dependent on the Rayleigh
number. In order to estimate the influence of Ra on the flow dynamics, we
compare the relatively early-time behaviour at two Rayleigh numbers. Figure 12
plots the concentration contours for Ra = 1000 at t = 2.2 and Ra = 8000 at t = 1.6.
These simulations employ 512 × 512 and 2048 × 2048 grid points, respectively. The
concentration front moves much faster for the later case. The Ra = 1000 case develops
a few large fingers and also shows a substantial amount of diffusive smearing in
other regions of the concentration front. The large-scale fingers appear to move
independently, without interacting with neighbouring fingers. The Ra = 8000 case,
on the other hand, develops a vigorous instability that results in complex fingering
104 A. Riaz, M. Hesse, H. A. Tchelepi and F. M. Orr Jr
40
30 DNS (n̂)
linear stability (kmax)
20

Wave number
10

Ra = 500
1000
2000
2 4000

0.4 2 4 6 8 10
t

Figure 11. The dominant mode of the nonlinear simulations as a function of time for different
values of the Rayleigh number. Also shown are the most dangerous modes given by the linear
stability analysis. Good agreement is observed between the two for early times. At later times,
the onset of nonlinear behaviour leads to a significant deviation from linear results.

(a) 0

z 0.1

0.2
0 0.2 0.4 0.6 0.8 1.0

(b) 0

z 0.1

0.2
0 0.2 0.4 0.6 0.8 1.0
x

Figure 12. Concentration contours for (a) Ra = 1000, t = 2.2 and (b) Ra = 8000, t = 1.6. The
larger-Rayleigh-number case is marked by intense nonlinear interaction between competing
fingers. The small-Ra case shows larger fingers with a strong diffusive spreading.

structures. Compared to the Ra = 4000 case shown in figure 9(b) at t = 0.18, the
Ra = 8000 case already shows well-developed discrete feeding sites. Note that new
feeding sites develop as the old ones are abandoned. In other regions, multiple fingers
can attach to a single feeding site, the nonlinear dynamics then selects one over the
others as the preferential flow path.
The complexity of the fingering structures suggest that the long-term fate of
the nonlinear competition cannot be predicted from the concentration profiles at
earlier times, especially for large Rayleigh numbers. In order to analyse the fingering
dynamics at long times, we plot concentration profiles in figure 13 for various Rayleigh
numbers at times when the concentration front reaches the bottom boundary. We
designate this time as tb . The small Rayleigh case, Ra = 1000 at time tb = 8.9, shows
 Density-driven miscible flow in porous media 105
(a) 0 (b) 0

0.2 0.2

0.4 0.4
z
0.6 0.6

0.8 0.8

1.0 1.0
0 0.2 0.4 0.6 0.8 1.0 0 0.2 0.4 0.6 0.8 1.0

(c) 0 (d) 0

0.2 0.2

0.4 0.4
z
0.6 0.6

0.8 0.8

1.0 1.0
0 0.2 0.4 0.6 0.8 1.0 0 0.2 0.4 0.6 0.8 1.0
x x

Figure 13. Concentration contours at time tb , when the concentration front reaches the
bottom. (a) Ra = 1000, tb = 8.9. (b) Ra = 2000, tb = 7. (c) Ra = 4000, tb = 8.0. (d) Ra = 8000,
tb = 8.5. The late-time dynamics is governed by large-scale fingers. Large-Rayleigh-number
cases continue to display vigorous fingering interactions.

that two large fingers survive to reach the bottom boundary. Comparison with
figure 12(a) shows that the fingering configuration is completely different from that
at an earlier time of t = 2.2. The Ra = 2000 case at tb = 7 displays slightly narrower
fingers, with one isolated finger in the middle making it to the bottom boundary,
while two other fingers, attached to a single feeding site, are still in competition. For
larger-Rayleigh-number cases, Ra = 4000, tb = 8 and Ra = 8000, tb = 8.5, more fingers
reach the bottom boundary. Many of these fingers can be observed to undergo strong
interactions while others are in the process of fading out. Figures 13(c) and 13(d)
clearly show that the fading of the fingers is due to the shift in the feeding sites as well
as to the lateral pinch-off mechanism, where one diagonally moving finger cuts the
fluid supply of the neighbouring finger. A consequence of these fingering interactions
is an increase in tb for larger values of the Rayleigh number, as compared to the
small Ra cases. Note that tb = 8.9 for Ra = 1000, it decreases to 7 for Ra = 2000, but
then again increases to 8 for Ra = 4000 and then to 8.5 for Ra = 8000.
The position of the most advanced section of the concentration front, the tip
position, as a function of time is shown in figure 14 for various Rayleigh numbers.
106 A. Riaz, M. Hesse, H. A. Tchelepi and F. M. Orr Jr

100

Tip position
10–1

Ra = 500
2000
8000

10–2
10–1 100 101
t

Figure 14. The position of the most advanced portion of the front as a function of time for
different values of the Rayleigh number. The tip moves faster for larger Rayleigh numbers for
early times. Fingering interactions significantly influence the rate of front propagation at later
times, particularly for large Ra cases.

For the small Ra = 500 case, the front initially propagates as t 1/2 and then switches to
a linear growth for larger times. Higher-Rayleigh-number cases display the diffusive
t 1/2 behaviour for relatively shorter times. The Ra = 2000 case displays a faster
than linear growth of the tip position at later times. The tip travels quickly for the
Ra = 8000 case but for a very small initial time. It is clear that by t = 5, the tip is
moving faster for the Ra = 2000 case as compared the Ra = 8000 case. Consequently,
the time for a finger to reach the bottom boundary is shorter for the Ra = 2000
case. As shown by the concentration contours in figure 13, this behaviour is due to
a more intense fingering interaction for higher-Rayleigh-number cases, such that the
interacting fingers do not allow any one finger to clearly breakaway ahead of the
front. For the smaller Ra cases, on the other hand, isolated fingers travel faster owing
to the absence of interference from the neighbouring fingers.

4. Discussion
We have analysed the stability of the diffusive boundary layer in a semi-infinite
domain. This analysis is applicable when the penetration depth of the diffusive
boundary layer, δ, is small relative to the domain thickness, H . The penetration depth
at the onset of instability is given in dimensional form as δc ≈ 24µD/(Kρg). We
must have δc  H for the assumption of the semi-infinite domain to be valid. The
dimensional critical time, which is then independent of the length scale, is given by
φµ2 D
tc = 146 . (4.1)
(Kρg)2
Similarly, we define the dimensional critical wavelength λc as

2πµD
λc =
. (4.2)
0.07Kρg
To validate both the linear analysis and the numerical simulations, the predictions
of tc and λc should be compared with experimental observations. Unfortunately,
 Density-driven miscible flow in porous media 107

103 tc

102 λc

tc (yrs), λc (m), δc (m)

101
δc

100

10–1

10–2

101 102 103 104

k (mD)

Figure 15. Variation with permeability of the critical time, tc , the critical wavelength, λc
and the penetration depth at the critical time, δc , for ρ = 5 kg m−3 , φ = 0.3, µ = 0.5 cP and
D = 10−9 m2 s−1 . For this range of permeability variation, tc varies between 2000 yrs and 10 days
while λc goes from about 100 m to less than 1 m. The penetration depth δc gives information
regarding the applicability of our analysis with respect to the layer thickness H , such that
δc  H .

Elder (1968) states only the Rayleigh number of his experiments and does not give
all the data necessary to calculate tc and λc . Green & Foster (1975) have reported
experiments of a salt solution diffusing into the top of a Hele-Shaw cell. The Rayleigh
number in their experiment is Ra = KρgH /(µD) = 90 500. They do not report the
onset time, but give the wavelength of the first observed fingers as λ = 1.8 mm and
note that this may be an overestimate. For the same parameters, our linear stability
analysis predicts a critical time of tc = 4 s and a critical wavelength of λc = 0.6 mm. The
critical wavelength predicted by linear theory is generally smaller than the wavelength
first observed, because the fingers are initially not detectable in the experiments. By
the time they become observable, they have coarsened to a larger wavelength. Given
this experimental limitation on the detection of the critical parameters, high-resolution
numerical studies as reported above are valuable because the growth of perturbations
can be detected before they become visible.
The critical time and the critical wavelength can vary by orders of magnitude
depending on the properties of the geological formation. While viscosity, diffusion
and density difference have more or less similar values for typical aquifers, permeability
can vary over a large range of values. It therefore introduces the largest variation in
the critical time, tc , and the critical wavelength, λc . Figure 15 shows that δc decreases
from 55 m to about 0.07 m when the permeability increases from 1 mD to 3 D. Hence,
our results for the semi-infinite domain will apply to high-permeability layers with a
thickness of few tens of metres, but for low-permeability formations the thickness has
to be several hundred metres for our results to be applicable. The critical wavelength
decreases from λc ≈ 200 m to λc ≈ 0.3 m, while the critical time for the onset of the
instability decreases from tc ≈ 2000 yrs to tc < 10 days.
The critical wavelength, λc , is an indication of the length scales that have to be
resolved by numerical simulations to capture the convective transport and the resulting
solution trapping. In high-permeability reservoirs the fingers are much smaller than
108 A. Riaz, M. Hesse, H. A. Tchelepi and F. M. Orr Jr

1.0
e≈
102

op
Sl
Depth (m)
.5
101 ≈0
pe
Slo

100
101 102 103 104
Time (yrs)

Figure 16. The advance of the fastest finger tip is shown until the bottom of the simulation
domain is reached. A diffusive, t 0.5 , behaviour is observed at early times, while the fingers
advance proportional to t at later times.

the size of grid blocks typically used in reservoir simulations. A numerical study of
the high-permeability Sleipner injection site by Lindeberg & Bergmo (2003) shows
that a grid block size of 3 cm×4 cm is necessary to capture the initial fingering
with a standard reservoir simulator. For field-scale simulation, the grid size has to be
increased to 100 m×100 m. Lindeberg & Bergmo (2003) show that this delays the time
of onset to 100 years, i.e. two orders of magnitude higher than on the fine grid. Such
increases in the onset time may or may not be acceptable, depending on the time scale
of storage. More importantly, the failure to resolve the length scales of the critical
disturbances at small times may lead to significant errors in the nonlinear regime
even if the dominant length scales at long times are coarse enough to be resolved on
the chosen grid. Diersch & Kolditz (2002) have investigated the convergence of the
Elder problem. They show that even at late times, when most of the structures are
large enough to be resolved on a coarse grid, the solution is sensitive to the grid size.
Following the onset of instability, the strength of convective mixing can be
determined qualitatively from the speed of the most advanced finger tip, shown
in figure 16 for several values of the permeability. All other properties are held fixed
at the same values as in figure 15. For all permeability values, the position of the
finger tips is initially proportional to t 0.5 . Eventually, however, they accelerate, and the
penetration is approximately proportional to t. Note that the late-time rate of advance
is similar for all values of permeability. Larger permeability values simply lead to an
earlier switch from t 0.5 to the linear regime. In high-permeability formations, solution
trapping is therefore strongly enhanced by convection, and may be the dominant
mechanism reducing the amount of the mobile CO2 phase. To assess the safety and
effectiveness of CO2 storage in high-permeability formations, it is therefore necessary
to simulate this convective process accurately.
Heterogeneity is important for the miscible viscous fingering instability (Tchelepi &
Orr 1994; Prassad & Simmons 2003; Riaz & Meiburg 2004b). Some attempts have
been made to model the effect of heterogeneity on the mechanism of this instability
(Gounot & Caltagirone 1989), but this remains one of the directions of future research.
The fine stratigraphic layering of sedimentary rocks causes strong anisotropy of the
 Density-driven miscible flow in porous media 109
permeability. In general, the vertical permeability may be several orders of magnitude
smaller than in the horizontal direction. Ennis-King & Paterson (2003) have shown
that anisotropy has an equally strong effect on the critical time and wavelength as the
magnitude of the permeability. The analysis developed here can easily be extended to
include the effects of anisotropy.
Our analysis does not include the full dispersion tensor which can account for
the influence of velocity-induced dispersion (Yortsos & Zeybek 1988; Tchelepi et al.
1993). Work by Riaz & Meiburg (2004a) suggests that for homogeneous porous
media, velocity-induced dispersion is generally equivalent to a slight increase in the
level of molecular diffusion. However, dispersion can be expected to play a more
important role in heterogeneity dominated flows.

5. Conclusions
Accurate numerical simulation of density fingering over long times at the field scale
is one of the main challenges in predicting the movement of CO2 underground. Our
theoretical and numerical results are valuable for understanding the density-driven
convection during CO2 storage in saline aquifers. Our analysis is directly applicable
to the thermal instability problem and resolves the issue of the critical time and the
critical wavenumber. We also highlight the physical mechanisms of instability that
give rise to both the long wavenumber cutoff and the critical time for the onset, in a
semi-infinite domain.
The time for the onset of instability can be accurately determined only by the linear
stability analysis presented here. Experiments tend to over-predict this time because
the disturbances grow to observable fingers only after a finite period following the
critical time. For typical aquifers with moderate permeability, the onset time can be
as large as hundreds of years. The prediction of the critical wavenumber is crucial
in choosing the grid resolution of the numerical simulation. Although the late-time
large-scale fingers can be resolved with fewer grid blocks, the errors incurred in
ignoring the small-scale dynamics at early times, can have a significant influence on
the late time behaviour.
The simplifications necessary to treat the problem theoretically may introduce
significant errors when such results are applied to real aquifers. The important
assumptions are the homogeneity and isotropy of the porous medium, as well
as the assumption of the single-phase flow and the absence of velocity-induced
dispersion. The assumption of single-phase flow will break down if capillary forces
are significant. Additionally, physical mechanisms related to dissolution, precipitation
and geochemical reactions, which are not accounted for in our analysis, can be
expected to play a role. Some of these processes may be incorporated in the linear
stability analysis, while others can be investigated only with high-resolution nonlinear
simulations. The results for the basic gravitational fingering instability presented
here, form the basis for investigating the effect and importance of these additional
processes.

The authors would like to acknowledge helpful discussions with Dr David Pritchard
and Dr Jonathan Ennis-King. This research was supported by the Global Climate and
Energy Project and the SUPRI-B, reservoir simulation affiliates program, at Stanford
University.
110 A. Riaz, M. Hesse, H. A. Tchelepi and F. M. Orr Jr
REFERENCES
Bachu, S., Gunther, W. D. & Perkins, E. H. 1994 Aquifer disposal of CO2 : hydrodynamic and
mineral trapping. Energy Conserv. Mgmt 35, 269–279.
Ben, Y., Demekhin, E. A. & Chang, H.-C. 2002 A spectral theory for small-amplitude miscible
fingering. Phys. Fluids 14, 999.
Caltagirone, J.-P. 1980 Stability of a saturated porous layer subject to a sudden rise in surface
temperature: comparison between linear and energy methods. Q. J. Mech. Appl. Maths 33,
47–58.
Camhi, E., Ruith, M. & Meiburg, E. 2000 Miscible rectilinear displacements with gravity override.
Part 2. Heterogeneous porous media. J. Fluid Mech. 420, 259.
Chang, H.-C., Demekhin, E. A. & Kalaidin, E. 1998 Generation and suppression of dadiation by
solitary pulses. SIAM J. App. Maths 58, 1246.
Diersch, H. J. G. & Kolditz, O. 2002 Variable-density flow and transport in porous media:
approaches and challenges. Adv. Water Res. 25, 899–944.
Drazin, P. G. & Reid, W. H. 1981 Hydrodynamic Stability. Cambridge University Press.
Elder, J. W. 1967 Transient convection in a porous medium. J. Fluid Mech. 27, 609–623.
Elder, J. W. 1968 The unstable thermal interface. J. Fluid Mech. 32, 69–96.
Ennis-King, J. & Paterson, L. 2003 Role of convective mixing in the long-term storage of carbon
dioxide in deep saline formations. SPE 84344 pp. 1–12.
Foster, T. 1965 Stability of a homogeneous fluid cooled uniformly from above. Phys. Fluids 8,
1249–1257.
Foster, T. 1968 Effect of boundary conditions on the onset of convection. Phys. Fluids 11, 1257–
1262.
Gounot, J. & Caltagirone, J.-P. 1989 Stabilité et convection naturelle au sein d’une couche poreuse
non homogène. Intl J. Heat Mass Transfer 32, 1131–1140.
Green, L. L. & Foster, T. 1975 Secondary convection in a Hele Shaw cell. J. Fluid Mech. 71,
675–687.
Gresho, P. M. & Sani, R. L. 1971 The stability of a fluid layer subjected to a step change in
temperature: Transient vs. frozen time analysis. Intl J. Heat Mass Transfer 14, 207.
Hitchon, B. 1996 Aquifer disposal of carbon dioxide: hydrodynamic and mineral trapping-proof
of concept. Tech. Rep. Alberta Research Council.
Homsy, G. M. 1973 Global stability of time-dependent flows: impulsively heated or cooled fluid
layers. J. Fluid Mech. 60, 129–139.
Jhavery, B. S. & Homsy, G. M. 1982 The onset of convection in fluid layer heated rapidly in a
time-dependent manner. J. Fluid Mech. 114, 251.
Kaviany, M. 1984 Onset of thermal convection in a saturated porous medium: experiment and
analysis. Intl J. Heat Mass Transfer 27, 2101–2110.
Kumar, A., Noh, M., Pope, G. A., Sepehrnoori, K., Bryant, S. & Lake, L. 2004 Reservoir
simulation of CO2 storage in deep saline aquifers. SPE 89343 .
Lick, W. 1964 The instability of a fluid layer with time-dependent heating. J. Fluid Mech. 21,
565–576.
Lindeberg, E. & Bergmo, P. 2003 The long-term fate of CO2 injected into an aquifer. In Greenhouse
Gas Control Technologies (ed. J. Gale & Y. Kaya), vol. 1, pp. 489–495. Elsevier.
Lindeberg, E. & Wessel-Berg, D. 1997 Vertical convection in an aquiver column under a gas cap
of CO2 . Energy Conserv. Mgmt 38, SS229–SS234.
Manickam, O. & Homsy, G. M. 1995 Fingering instability in vertical miscible displacement flows
in porous media. J. Fluid Mech. 288, 75.
van der Meer, L. G. H. 1992 Investigations regarding the storage of carbon dioxide in aquifers in
the Netherlands. Energy Conserv. Mgmt 33, 611–618.
Nield, D. A. & Bejan, A. 1999 Convection in Porous Media, 2nd edn. Springer.
Otero, J., Dontcheva, L. A., Johnston, H., Worthing, R. A., Kurganov, A., Petrova, G. &
Doering, C. R. 2004 High-Rayleigh-number convection in a fluid-saturated porous layer.
J. Fluid Mech. 500, 263–281.
Pego, R. L. & Weinstein, M. I. 1994 Asymptotic stability of solitary waves. Commun. Math. Phys.
104, 305.
 Density-driven miscible flow in porous media 111
Prassad, A. & Simmons, C. T. 2003 Unstable density driven flow in heteroheneous porous media:
a stochastic study of the Elder [1967b] ‘short heater’ problem. Water Resources Res. 39, 1–21.
Pritchard, D. 2004 The instability of thermal and fluid fronts during radial injection in porous
media. J. Fluid Mech. 508, 133.
Riaz, A. & Meiburg, E. 2003a Radial source flows in porous media: linear stability analysis of
axial and helical perturbations in miscible displacements. Phys. Fluids 15, 938.
Riaz, A. & Meiburg, E. 2003b Three-dimensional miscible displacement simulations in
homogeneous porous media with gravity override. J. Fluid Mech. 494, 95.
Riaz, A. & Meiburg, E. 2004a Linear stability of radial displacements in porous media: influence
of velocity-induced dispersion and concentration-dependent diffusion. Phys. Fluids 16, 3592.
Riaz, A. & Meiburg, E. 2004b Vorticity interaction mechanisms in variable-viscosity heterogeneous
miscible displacements with and without density contrast. J. Fluid Mech. 517, 1.
Robinson, J. L. 1976 Theoretical analysis of convective instability of a growing horizontal thermal
boundary layer. Phys. Fluids 19, 778–791.
Ruith, M. & Meiburg, E. 2000 Miscible rectilinear displacements with gravity override. Part 1.
Homogeneous porous medium. J. Fluid Mech. 420, 225.
Tan, C. T. & Homsy, G. M. 1987 Stability of miscible displacements in porous media: radial source
flow. Phys. Fluids 30, 1239.
Tan, C. T. & Homsy, G. M. 1988 Simulation of nonlinear viscous fingering in miscible displacement.
Phys. Fluids 31, 1330.
Tchelepi, H. A. & Orr, F. M. Jr 1994 Interaction of viscous fingering, permeability inhomogeneity
and gravity segregation in three dimensions. SPE Res. Engng p. 266.
Tchelepi, H. A., Orr, F. M. Jr, Salin, D. & Woumeni, R. 1993 Dispersion, permeability
heterogeneity and viscous fingering: acoustic experimental observations and particle tracking
simulations. Phys. Fluids 5, 1558.
Voss, C. I. & Sousa, W. R. 1987 Variable density flow and solute transport simulation of regional
aquifers containing a narrow freshwater–saltwater transition zone. Water Resources Res 28,
1851.
Yortsos, Y. C. & Zeybek, M. 1988 Dispersion driven instability in miscible displacement in porous
media. Phys. Fluids 31, 3511.
Zimmerman, W. B. & Homsy, G. M. 1992 Viscous fingering in miscible displacements: unification
of effects of viscosity contrast, anisotropic dispersion and velocity dependence of dispersion
on nonlinear finger propagation. Phys. Fluids 4, 2348.