VOLUME 34, NUMBER 5 NOVEMBER 1986

Fluctuations, viscous fingering, and diffusion-limited aggregation

Derek Y. C. Chan and Barry D. Hughes

Department of Mathematics, University of Melbourne, Parkville, Victoria, Australia 3052

I.incoln Paterson
Division of Geomecharucs, Commonwealth Scientific and Industrial Research Organization,
P. O. Box 54, Mount waverley, Victoria, Australia 3149
(Received 25 July 1986)
A physical interpretation of the diffusion-limited-aggregation (DLA) algorithm for simulating
viscous fingering during fluid displacement in a porous medium is given in terms of the material
properties of the porous medium. In one dimension, the flow patterns generated by a DLA simula-
tion correspond to a simple exponential distribution of "Quid capacity. "
Also, for any given distri-
bution of fluid capacity, it is possible to estabhsh a corresponding stochastic algorithm for the mov-
ing Quid interface. The discussion of this paper places the DLA simulation of viscous fingering on
a physical basis and, more generally, establishes the mathematical equivalence of two models of sto-
chastic interface evolution.

Paterson' has shown that the diffusion-limited- and P/r, respectively.

P/1 vi=(klan, The magnitudes of the volume
v„=(klan,
aggregation (DLA) process introduced by Witten and fluxes vt and v„on the left and right of the origin are
Sander can be used to model fingering patterns and given by Darcy s law':
sweep efficiencies of two-fluid displacement in a porous
medium. The original DLA simulation is only applicable )(P/1), )(P/r),
to the case of an inviscid driving fluid displacing a where k is the permeability and p is the viscosity of the
Newtonian fluid (i.e., at the limit of infinite mobility ra- displaced fluid. The instantaneous velocities of the inter-
tio) and when interfacial and wetting effects are negligi- faces are given by
ble. While there have betm ro:ent attempts to modify the
DLA algorithm to include some of these effects, the pre- dl/dt = vt/q, dr—
/dt = —v„/y, (2)
cise connection between DLA and the fluid flow problem
where y is the porosity. In an ideal porous medium where
has yet to be established.
there are no fluctuations in material properties, k and y
When an inviscid fluid is displacing another fluid of fi-
are constants and the motion of the interfaces can be
nite viscosity, the fluid interface is unstable to any pertur-
found by solving Eqs. (I) and (2) for l(t) and r(t). One
bations, however small in magnitude. 3 Even in a macros- —
finds that 1 (t) — 1 (0)=r (t) r (0)= (2kPlpy)t. — In a
copically uniform and statistically homogeneous porous
statistically homogeneous rather than ideal porous medium
material there are fluctuations in the hydrodynamic con-
the interface will not move in this deterministic manner,
ductivities and void structure over the pore scale, and over
as there is a stochastic element in the displacement pro-
larger scales, which can initiate finger growth; these fluc-
cess. We introduce this stochastic element by hypothesis
tuations do not appear in the macroscopic equation
A:
governing the flow (Darcy's law). In contrast, in the
DLA algorithm, the stochastic nature of the "random
walkers" provides the source of fluctuations in the growth In the next time step ht, the interface will advance one
of fingers. The aim of this paper is to establish a hnk be- step du& with a statistical weight that is proportional to
the magnitude of the velocity.
tween the probabilistic nature of DLA and the fluctua-
tions of microstructure in a porous material. Such a con-
This hypothesis has been proposed by several authors as
nection will hopefully allow one to devise simple proba-
a growth law or simulation algorithm to produce random
bilistic models to study more complex problems of fluid
structures. %'e now examine its consequences and its rela-
flow in porous media.
tionship with the statistical properties of the porous medi-
Consider the displacement of a Newtonian fluid of fi-
QHl.
nite viscosity by an inviscid fluid. We limit the discussion
Consider the relative evolution of the left and right in-
to a displacement process in one dimension, with the
terfaces for the one-dimensional displacement described
viscous fluid withdrawn from the origin and the interfaces
above. According to hypothesis A, the probability of the
between the fluids located at points x = — l(t) &0 and
x =r(t) ~0. A constant pressure drop P is maintained r interface moving in the next time step is
between each fluid interface and the origin. Since the
Probt r interfaces moves) =(dr/dt)/[(dr/dt)+(dl/dt)]
pressure satisfies Laplace's equation, the pressure gra-
dients on the left and right of the origin have magnitudes

34 4079 Oc1986 The American Physical Society

4080 CHAN, HUGHES, AND PATERSON 34

with a similar equation for the probability of advance of The function I(M) is exhibited in Fig. 1. We may inter-
the I interface. Applying Eqs. (1) and (2) to each inter- pret 1 I— (M) as the one-dimensional analogue of the
face, we find that areal sweep efficiency in displacement studies. In one di-
mension I(M)~0 as M~ ao (although the mean length
Prob f r interface moves I =I /(r + I}, of the region not swept out diverges}. In contrast, in two
(4) or three dimensions the analogue of I(M) converges to 1
Prob[1 interface moves) =r/(r +I) . as the size of the region initially filled with the fluid to be
These results lead to a probabilistic algorithm for the evo- displaced grows.
lution of the two interfaces: select a random number in While hypothesis A (and DLA, to which it is
[0, 1] and then move one or the other interface according equivalent} may be used as a basis for qualitatively correct
to Eqs. (4). In a lattice simulation, the interface chosen is simulation of unstable displacements, it uses no informa-
advaiiced by oiie grid spaclllg. tion about the porous material other than its bulk proper-
The probabilities given in Eqs. (4) are closely related to ties, and so cannot distinguish between microstructurally
the DLA algorithm. If we release an unbiased random different porous media. We contend that microstructure
walker at the origin, the probability of the walker reach- is the essence of fingering, and that a correct understand-
ing the r interface (I interface) before reaching the I inter- ing of the role of microstructure in fingering might lead
face (r interface) is given precisely by the first (second) of to the possibility of defining model media for which DLA
Eqs. (4). In other words, for a one-dimensional lattice, gives a quantitatively correct description of fingering. We
the DLA algorithm is mathematically identical to our hy- suggest that a key microstructural parameter which con-
pothesis A. That hypothesis A and DLA are equivalent trols fingering is the fluid capacity We. define the fluid
for a lattice of arbitrary dimensionality is well known. capacity, a dimensionIess quantity, to be the void space
The advantage of hypothesis A is that we have a method per specified length A, in one dimension, the void space per
for converting a flow equation such as Eq. (1) or its gen- specified area A, in two dimensions, and the void space
eralization (to include finite mobility, interfacial, or non- per specified volume }(,i in a three dimensions. If )(, is on
Newtonian effects) to a simple stochastic algorithm. 6 The the Darcy scale, then in three dimensions fluid capacity
difficulty in two and three dimensions is to derive the becomes identical to porosity ip, and we shall use the sym-
analogue of Eq. (1) needed to relate the interfacial velocity bol q& here to denote fluid capacity. In three dimensions,
to the position of the interface. fiuid capacity, like porosity, is constrained to be less than
For our one-dimensional model we can formally express unity; no such constraint exists in one or two dimensions.
the probabilistic advance of the two interfaces as follows. If )(, is on the pore scale, the fluctuations in the fluid capa-
Let P„(l,r) be the probability that the two interfaces be at city correspond to the pore size distribution. In a DI,A
position I and r at the nth time step. From Eqs. (4) the simulation, we identify A, as the lattice size.
recurrence equation for P„ is To illustrate the role of the fiuid capacity, consider a
model porous medium, which consists of narroio tubes of
P„(l,r) =[I/(I+r +1)]P„ i(l, r +1) comparable lengths and diameters connecting chambers
with volumes very much larger than the volumes of the
+[r/(I+r +1)]P„ i(l + l, r) narrow tubes. The narrow tubes give rise to the permea-
with the initial and boundary conditions Po(l, r) = 5& L 5, x, bility of the porous medium, while the chambers give rise
to the fluid capacity. To apply this model to a real porous
P„(L,R)=0, n ~0, P„(l,r)=0 if I ~l or r yR . medium, one should not necessarily identify the narrow
It is useful to introduce the generating function
+" "(I+r)h(l, r)
P(l, r, f)= g PP„(l,r)=(~
n&0

and in terms of the g-independent function h (l, r) defined

by Eq. (6) the difference equation to be solved becomes

(I +r)h (I, r) =rh (I + l, r)+Ih (l, r + 1),

0(1(L and 0&r &R, (7)
10
subject to the boundary conditions: h (L,R) =(L +R)
and h (l, r) =0 for I ~ L or r ~R. The numerical solution
of Eq. (7) is straightforward. 4
The quantity of interest in this one-dimensional model 10 10 M

is the displacement inefficiency, which we define to be the

FICi. 1. Displacement inefficiency I(M) [Eq. (S)), obtained
mean position of one interface when the other has reached
the origin. %e consider only the symmetric case where
by numerical I
solution of Eq. (7). It appears that (M) ~ M
as M~00, so the mean length not swept is proportional to
L =R =M for which the displacement inefficiency is M' . This may be compared with the easily derived result that
the length not swept is proportional to M'~ when the probabili-
I (M) =M '
Q IP(1,0, 1) g P(1,0, 1) ties of motion of either interface are assumed to be equal. The
inset shows the model system simulated in Table I.
34 FLUCTUATIONS, VISCOUS FINGERING, AND DISCI'USION-. . . 40S1

tubes as individual pore throats and the chambers as indi- taneous velocity of the appropriate interface. We consider
vidual pore bodies. The physical nature of the tubes and two prescriptions for the probability that the bond labeled
chambers is determined by the length scale A, . We shall i is the bond chosen for growth:
take the volumes of the chambers to be randomly and in-
dependently distributed. For a two-fluid displacement in
our model porous medium, consider flow within two of Prob{bond i chosenj =u; g uj,
the narrow tubes which are adjacent to unfilled chambers,
and comparable distances from the pressure source. The
interface cannot advance into a new tube until the Prob{bond i chosenj =Prob{ u;/p; =maxi {ui/+J j j (12)
chamber into which fluid is currently pouring is filled.
The flow rates within the tubes, and so the rates of filling
of the chambers, may be of comparable magnitude, but all The first of these equations is DLA or hypothesis A and
other things being equal, the smallest chamber will fill f
the second is hypothesis B. Let denote the probability
density function for the independent random variables q&;.
soonest, so the motion of the interface is heavily governed
by the micrastructure. The event "u;/q&; =maxj {uj./qi j" is exactly the same as
The preceding model is a discrete version of a porous
"
the event "yj &(ui/u;)p; for all j, so that Eq. (12) be-
medium in which the fluid capacity y is a random vari- comes
able. Can we find a stochastic law for interface growth
for such a medium which coincides exactly with the Prob{bond i chosen j
microstructure-insensitive hypothesis A (or DLA)? We
propose, for the one-dimensional model, hypothesis 8: pl fl +J +J
j
(V /V )P-

Prob{1 interface movesj =Prob{ dl/dt

~ ( & (
dr/dt ~ j .

For the special case in which the random variables y;

With this hypothesis only one interface moves at each have the exponential density f
(p) a exp( —
= ay), where a
is an arbitrary constant, the integrals in Eq. (13) are easily
time step„even if dl/dt and dr/dt are of comparable
magnitude. However, if we consider what happens over a evaluated and the probabihty found to be u;/g. ui.
few time steps, we find that both interfaces will advance. Hence the two growth laws (11) and (12) are exactly the
We make the plausible assumption that the effective per- same, irrespective af the precise values of the determinis-
meabilities experienced by the left and right interfaces are tic quantities u; and the value of N, provided that the ran-
nearly equal, so that all stochastic properties of the flow dom variables y; are exponentially distributed. This re-
are attributed to fluctuations in the jiuid capacity of the sult may be significant far the modeling of a variety of
porous medium. Equation (2) remains valid with p now processes in which there is both a deterministic field and a
denoting the fluid capacity, and combining this with Eqs. random transport coefficient governing interface motion,
(1) and (9) we have but for the particular problem of fiuid displacements in
porous media leads to the following conclusion: Hy-
pothesis A (or equiualently the DLA algorithm) will yield
Prob{1 interface moves) =Prob{(lyt) '&(ry„) 'j . displacement results that correspond to a porous medium
(10) for which hypothesis 8holds, prouided that there is an ex-
ponential distribution ofj7uid capacities.
One may establish an exact mathematical equivalence We conclude by reporting the results of some simula-
between hypothesis A (which is the same as DLA) and the tions which we have performed for the model system of
consequence (10) of hypothesis 8, in the particular case in identical narrow tubes of neghgible volume connecting
which the fiuid capacity y has an exponential probability chambers of randomly distributed volume shown in Fig.
density function. Indeed the equivalence is a very general 1. These are classical continuum simulations, in which
consequence of exponential fluctuations in microstructure, both interfaces are allowed to move simultaneously. The
and holds independent of the dimension of the system and distribution of the chamber volumes is taken to be the
the physical interpretation placed upon the random quan- fiuid capacity distribution f(tp). An inviscid fluid dis-
tity y. We consider any connected set of sites on a d places the viscous fluid from both sides of the origin at
dimensional lattice, with a single new site about to be add- constant driving pressure. In this model system, motion
ed to the cluster in a random fashion. We label the possi- of the inviscid fluid from one tube to a neighboring tube
ble bonds along which the growth may occur (i.e., all is forbidden until all viscous fluid is displaced from the
bonds adjacent to the cluster) with a coordinate intervening chamber. The inefficiency I
is found as the
i =1,2, . . . , N. %ith each of these bonds is associated a ratio of the number of empty tubes on one side of the ori-
deterministic positive number u; and a positive random gin when the fluid an the other side reaches the origin to
variable y;. The random variables y; are assumed in- the number of empty tubes initially on one side of the ori-
dependent. For the one-dimensional fluid displacement gin. Results of simulations are given in Table I for a
problem, N =2, with i =1 indexing the l interface and selection of fluid capacity probability densities f(y); in
i =2 the r interface; u; denotes the volume flux calculated each case there are 100 narrow tubes on each side of the
on the basis of Darcy's law (1); and u;/y; is the instan- origin. For the exponential fluid capacity distribution, we
 34

TABLE I. Displacement ineffjciency I

(M) for a one- and so can be used to compare the models (and thereby
dimensional displacement with I. =R =M =100. Each result is hypothesis 8) against the continuum simulation. For the
an average over 200 simulations. The random fluid capacity y exponential fluid capacity distribution, we find on taking
is generated from a random number z, uniformly distributed on =
M 100 the exact value 33.0% for the displacement inef-
(0, 1). ficiency, which is satisfyingly close to the value of 32.4%
Displacement
found by the simulation.
Footnote Fluid capacity inefficiency (%) While DI A and hypothesis A are equivalent, they coin-
cide with hypothesis 8 only for the particular case of an
a y=—lnz 32.4 exponential distribution of fluid capacity. For any porous
b p=z 24. 5 medium in which fingering is governed by fluctuations in
e +=z 30.6 the fluid capacity, a DLA simulation will correctly
d +=z 34. 1 describe the fingering if and only if there is a length scale
e y=(1 —5+25z)2, 5=1 30.6 k for which the distribution of fluid capacity is well ap-
e y=(1 —5+25z)2, 5=0.33 19.8
proximated by an exponential. It should be remembered
e y=(1 —5+25z)2, 5=0. 1 11.2
that the probability density function for fiuid capacity
e y = (1 —5+25z)2, 5=0 1.0
need not be scale invariant, so that a density appropriate
'On the basis of hypothesis 8, this corresponds to DLA. at one length scale A, may not be appropriate at a different
~
q uniformly distributed. length scale. We have obtained by simulation of the
Cylinders with radii uniformly distributed on (0, 1). model system shown in Fig. 1 the displacement efficiency
Spheres with radii uniformly distributed on (0, 1). for various forms of the fluid capacity probability density
—
Cylinders with radii uniformly distributed on (1 5, 1+5), cf. function f (p). The results in Table I are consistent with
Chen and Wilkinson (Ref. 7). the simulations of Chen and Wilkinson. The discussion
given in this paper gives the DLA simulation of viscous
flngering a physical basis in terms of microscopic fluctua-
tions in the porous medium. Provided that one can iden-
have proved above that DLA, hypothesis A, and hy- tify the physical origin of relevant fluctuations, the argu-
pothesis 8 are mathematically identical. Numerical solu- ments advanced here in the context of viscous fingering
tion of the difference equation (7) yields the exact dis- can be used to quantify connections between DLA and
placement inefficiency for these three equivalent models, other physical phenomena.

~L. Paterson, Phys. Rev. Lett. 52, 1621 (1984}. 5This is just the classical solution of the "gambler's ruin" prob-
&T. A. Mitten and L. M. Sander, Phys. Rev. Lett. 47, 1400 lem (see, e.g. „%'. Feller, An Introduction to Probability Theory
(1981); Phys. Rev. 8 27, 5686 (1983); R. Ball, M. Nauenberg, and Its Applications, 3rd ed. (%'iley, New York, 1968), Vol. 1.
and T. A. VA'tten, Phys. Rev. A 29, 2017 (1984). The solution of this one-dimensional problem may be thought
3In two dimensions, the flow equations for a porous medium of as a simple example of an elegant proaxiure described by
coincide with those governing flow in a Hele-Shaw cell [Na- L. A. Turkevich and H. Scher, Phys. Rev. Lett. 55, 1026
ture (London) 58, 34 (1898)], i.e., between two closely spaced (1985), for calculating growth pattern probabihties in DLA.
parallel plates. The hydrodynamic instabihty when a less 6For the displacement of a power-law fluid by an inviscid fluid,
viscous Quid displaces a more viscous one is common to both Eq. (1) is replaced by v, =8/r~, where P is the power-law in-
systems; for the Hele-Shaw cell it is called the Saffman- dex, and Eq. (4) by Prob[r interface moves) =(dr/dt)/
Taylor instability. There are many papers in the literature on [(dr/dt)+(dl/dt)]=1~/(r~+l&) For the c.ase of finite mo-
this subject; a recent one is D. Bensimon, Phys. Rev. A 33, bility ratios y: —m&/m2 (m =k/p, subscript 1 denotes the
1302 (1986). Despite the mathematical equivalence of their driving fluid), Eq. {1)becomes v, = m ~P/[R +(y —1)r]. Here
governing equations, the Hele-Shaw cell is a poor model for the driving fluid occupies the intervals — I. &x &1(t) and
porous media because of its complete lack of the microstruc- r(t) &x &R, while the displaced fluid is in —1(t) &x &r(t);
ture wherein most of the salient features of porous medium the points x =L and x =R are held at a pressure I' and the
hydrodynamics reside. Papers on QLA and the Saffman- origin is taken at zero pressure. Equation {4) be-
Taylor problem include L. P. Kadanoff, J. Stat. Phys. 39, 267 comes Prob[r interface movesj =[I. +(y —1)l]/[I. +R
(1985); C. Tang, Phys. Rev. A 31, 1977 (1985); S. K. Sarker, +(y —1)(r +l)). From these examples, it is easy to see how
ibid. 32, 3114 (1985};and J. Nittmann, G. Daccord, and H. E. to generate the appropriate stochastic algorithm for a given
Stanley, Nature (London) 314, 141 (1985). physical displacement process. The interpretation of each sto-
4J. D. Sherwood and J. Nittmann, J. Phys. (Paris} 47, 15 (1986}; chastic algorithm in terms of fluctuations in the fluid capacity
A. J. DeGregoria, Phys. Fluids 28, 2933 (1985); M. J. King is found in prinriple from Eq. (13).
and H. Scher, SPE Report No. 14366 (unpublished). ~J.-D. Chen and D. %'ilkinson, Phys. Rev. Lett. 55, 1892 (1985}.