Percolation Theory

PR King
1
, SV Buldyrev
2
, NV Dokholyan
2,4
, S Havlin
5
, E Lopez
2
, G Paul
2
, HE Stanley
2


1 Department of Earth Science & Engineering, Imperial College, SW7 2BP, London, UK
2 - Department of Engineering, Cambridge University, Cambridge, CB2 1PZ, UK
3 - Center for Polymer Studies, Boston University, Boston, MA 02215, USA
4 Department of Chemistry & Chemical Biology, Harvard, Cambridge MA02138, USA
5 - Minerva Center & Department of Physics, Bar-Ilan University, Ramat Gan, Israel


Draft 27/5/02

Introduction:
Oil reservoirs are very complex having heterogeneities on all length scales from
microns to tens of kilometres. These heterogeneities affect all aspects of the flow and
have to be modelled to make reliable predictions of future performance. However, we
have very few direct measurements of the flow properties. Core plugs directly
measure the permeability but they represent a volume of roughly 10
-13
of a typical
reservoir. Well logs and well tests measure large volumes (10
-4
and 10
-7
respectively)
but the results have to be interpreted to infer flow properties. The flow itself takes
place on the scale of the pores which are typically around 10
-21
of the volume of the
reservoir. So there is a great deal of uncertainty about the spatial distribution of the
heterogeneities which influence the flow.

The conventional approach to this is to build detailed reservoir models (note that the
largest of these has around 10
7
grid blocks so they fall very short of the actual level
heterogeneity that we know about), upscale or coarse grain them to around 10
4
or
10
5
grid blocks and then run flow simulations. These models need to be taken from a
whole range of possible models with a suitable probability attached to each to
determine the uncertainty in performance. The problem with this approach is that it is
computationally very expensive. Therefore, there is a great incentive to produce much
simpler models which can predict the uncertainty in performance. These models must
be based on the dominant physics that control the displacement process.

It has long been understood that flow in heterogeneous porous media is largely
controlled by continuity of permeability contrasts flow barriers e.g. shales or high
permeability streaks or faults. Although there are other influences these are the
predominant features affecting flow. With this in mind we look to ways of modelling
reservoir flow which concentrate on the connectivity of permeability contrasts. The
basic mathematical model of connectivity is called percolation theory. Whilst there is
a very extensive literature on percolation theory in both the mathematical and physics
literature it is mostly not very accessible to the general geoscientist. The aim of this
article is to attempt to redress this balance.

Percolation theory
Percolation theory is a general mathematical theory of connectivity and transport in
geometrically complex systems. The remarkable thing is that many results can often
be encapsulated in a small number of simple algebraic relationships. First let us
describe what percolation theory is. There are many different variants which turn out
to be identical in almost all important aspects so we shall describe the simplest
version.

Take a square grid and occupy sites on this grid with a probability p. For small values
of this probability we see mostly isolated occupied sites with occasional pairs of
neighbouring sites that are both occupied. If neighbouring sites are both occupied we
call it a cluster. As the occupancy probability increases we get more isolated clusters,
some clusters grow and some merge. So the clusters on the whole get larger (see
Figs). Then at a particular value of the occupancy probability one cluster dominates
and becomes (infinitely) large. Above this the other clusters become absorbed into
this one until at p=1 every site is occupied.

Series of Figures. ->








Note that a very peculiar thing happens at one particular value of the occupancy
probability. Suddenly one cluster becomes infinitely large (for these purposes we are
discussing infinitely big lattices, we shall discuss what happens on finite size lattices
later). This is called the spanning cluster as it spans the entire lattice. This sudden
onset of a spanning cluster occurs at a particular value of the occupancy probability
known as the percolation threshold (p
c
) and is the fundamental characteristic of
percolation theory. The exact value of the threshold depends on which kind of grid is
used and strongly on the dimensionality of the grid. A table of values is given below.

Lattice Site Bond
Hexagonal 0.692 0.6527
Square 0.592746 0.5
Triangular 0.5 0.34729
Diamond 0.43 0.388
Cubic 0.3116 0.2488
BCC 0.246 0.1803
FCC 0.198 0.119

Note that we have described the connectivity of sites on the lattice, so this is known as
site percolation. Instead we could have occupied the edges of the sites (or the bonds),
this problem is known as bond percolation. Again this affects the percolation
threshold, but not the other fundamental properties, we shall return to this point later.

Sudden changes like this are common in other branches of physics. For example a
magnet, when heated, looses its magnetisation at a particular temperature (the Curie
Temperature). In general these are known as phase transitions (or critical
phenomena) and the percolation threshold is just another of these. It turns out that
many of the properties close to this transition can be described in very simple terms.







Not all occupied sites are in the infinite (or spanning) cluster. If we look at the
probability that an occupied site is in the infinite cluster (P(p)) then clearly this must
be zero (since there is no spanning cluster) below the percolation threshold. Above the
threshold it can be described in very simple analytical terms

( )
c c
p p p p p P >

~ ) (

1


P



0
0 p
c
1
p

This is known as a power law or scaling law and the exponent is known as a critical
exponent. This has the remarkable feature that it is entirely independent of the kind of
lattice being studied or whether it is bond or site percolation, it only depends on the
dimensionality of space (i.e. 2D or 3D). This is known as universality and is an
important aspect of percolation theory (and indeed critical phenomena in general).
Broadly speaking it means that the large scale behaviour of these systems can be
described by (relatively) simple mathematical relationships which are entirely
independent of the small scale construction. Clearly this is a very powerful concept as
it enables us to study and understand the behaviour of a very wide range of systems
without needing to worry too much over much of the detail. One key factor which is
not universal however is the percolation threshold. But the scaling laws and critical
exponents are. A table of values for this exponent is given below.

2D 3D
5/36~0.139 0.41

Note that in two dimensions it is often possible to determine exact values for the
exponents, whereas in three dimensions there are only approximate results or
numerical estimates. There are many other critical exponents that can be defined
which describe the properties of the percolating system at or near the threshold. There
are too many to describe in an introductory article like this and the literature should be
consulted for a more complete description (ref to Stauffer & Aharony). Here we shall
describe only the most useful for the application described above. Consider first the
size of the clusters. First we have to be a bit careful by what we mean by cluster size.
For these purposes we start with the two point correlation function g(r). This is the
probability that if one point is in a cluster then another point a distance r away is in
the same cluster. Then this typically has an exponential decay given by a correlation
length ()


/
~ ) (
r
e r g



Clearly at a low value of p these are small (typically clusters of size one or two), this
increases until the threshold when the spanning cluster dominates and is infinite in
size, so the cluster size diverges. What about above the threshold? Well we have to
remove the infinite cluster from our calculation otherwise it will always dominate and
only consider the other clusters. As the clusters get absorbed into the spanning cluster

0
0
P
the typical size of those left goes back down again. So we have a cluster size which
increases, diverges at the threshold and then decreases again. This can be described in
a mathematical form as

c
p p ~

Here is another critical exponent. As with the connectivity exponent () it is
universal (independent of details of the lattice) but does depend on the dimension of
the system. A table of values is given below.

2D 3D
4/3~1.333 0.88

There is a huge literature on percolation theory which defines and calculates a large
number of critical exponents. The intention (even were it possible) is not to review all
of this material but to focus on the issues pertinent to the questions asked in the
introduction. Before we do this there are two issues that must be covered. Everything
so far has been defined for an infinite lattice. What happens if a) the lattice is finite
and b) there is no lattice at all.

Finite size scaling
The problem of how to deal with finite size lattices is known as finite size scaling. It is
a useful introduction to the style of theoretical argument that is often used in
percolation theory. We shall look at the connectivity, P.

Consider a lattice of size L. For the sake of argument we shall assume it is square so
the number of cells is L
d
in d dimensions (physicists like to consider the general case
wherever possible, although many of our illustrations will be in 2D where it easier to
visualise things). The main thing to notice is that things are less clear cut. Consider
the following configurations on a 5x5 grid (notice that this is very small and one
doesnt really expect these kind of scaling arguments to apply but these give useful
insights into the real problems).

a) b)



































p=0.6 P=0.67 p=0.6 P=0.47 p=0.6 P=0.93

The thing to notice is that you can get connectivity (here defined as connectivity from
left to right) at very much less than the percolation threshold (p
c
~0.592... for a square
lattice in 2D infinite system) (0.2 for Figure a) but not get it at a much higher
occupancy (0.8 for Figure b) for a different realisation. Also if we look at some
realisations at exactly the same occupancy we can find very different connected
fractions (fraction of occupied sites belonging to the connected cluster). This is a
consequence of the fact that a finite size system can only sample from the entire
distribution of possible configurations and so there is a sample size uncertainty. If we
plotted the connected fraction as a function of the occupancy sampled over a large
number of configurations we would get a scatter of points. As the size of the system
got larger the scatter would reduce until we return to the plot for the infinite system
shown earlier.
Average curve
P




p

Instead of considering the whole scatter we can look at the average connectivity (that
is averaged over all realisations at the same occupancy fraction). Now we get the
following curve. Notice that there is no longer a sharp transition, it gets smeared
out. As we increase the size of the system the smearing gets less.




L
increasing


This phenomenon is again familiar from other thermodynamic phase transitions where
small size systems have smeared transitions. We can describe this smearing in
simple mathematical terms. First we look at the length scales in the problem. There
are only two the system size and the correlation length. Clearly if the system is much
larger than the correlation length (which we recall represents the typical size of the
clusters) then the clusters dont really notice the finite boundaries and the system must
behave like the infinite system. On the other hand when the cluster size sees the
boundaries a new behaviour must be introduced. So the important parameter must be
the dimensionless ratio of these two lengths, /L (in fact for later convenience we use
an equivalent dimensionless parameter z=(/L)
1/
=(p-p
c
)L
1/
). Then at a given value of
the system size the connected fraction must be a function of the form

( ) [ ]
/ 1
) , ( L p p f L p P
c L
=

where f
L
is some function. Now consider how the behaviour of the system changes
under an arbitrary change in size. Let L L . Under this change of scale we expect
the essential percolation behaviour to be unaltered, that is ) , ( ) ( ) , ( L p P c L p P .
The only function that has this property is the power so we can write

( ) [ ]
/ 1
) , ( L p p F L L p P
c
A
=

where F is a universal function and A is a power to be determined. To do this we
consider the asymptotic behaviour of P. As L we must recover the infinite
system critical law. Hence

z z F ] [ for large L. To cancel the L dependence we

must have A=-/. So the finite size scaling law for the connectivity can be written.

( ) [ ]
/ 1 /
) , ( L p p F L L p P
c
=



If we plot L
/
P against (p-p
c
)L
1/
all the curves found previously should lie on top of
each other to form a single universal curve. We can then use this to great effect
because if we want to know the finite size connectivity for any system that we havent
performed simulations of we can unscramble the result we require from this
universal curve.


L
/
P








(p-p
c
)L
1/





Continuum percolation
This is very straightforward because of the universality principle. There is no reason
why we need to use a grid at all. We can just geometrical objects randomly and
independently (formally this is called a Poisson process) in a continuum space.
Connectivity is defined as the overlap of the objects. In place of the occupancy
probability p we have the volume fraction of objects (or the probability that a point
chosen at random lies within one of the objects). For the sake of clarity we give this
the same letter, p. We get the same threshold phenomenon of a single cluster growing
and dominating the system. The percolation threshold depends only on the shape of
the objects, but for circles it is 0.678 and for squares it is 0.668 (similarly in 3D for
spheres it is 0.28 and for cubes 0.276) so the difference is not very large and
numerical experiments indicate that for reasonable convex (i.e. not very spiky) objects
the threshold is around the same value. This is known as continuum percolation.

Then the principle of universality applies (and has been extensively demonstrated by
numerical experiments) that the same scaling laws (e.g. finite size scaling) with the
same critical exponents (with the same numerical values). This is a remarkable result
and a very powerful one that we can now use.

Application to reservoir modelling

Imagine a typical reservoir model constructed with an object based technique. That is
geometrical objects (representing geological entities, e.g. shales, fractures, sand
bodies etc.) are placed randomly in space. Then the connectivity can be estimated
directly by percolation theory. Take a concrete example of sand bodies in an
otherwise impermeable (or low permeability) background. The net to gross ratio is the
volume fraction of the good sand and is, therefore, identical to the occupancy
probability p. Suppose we have a reservoir of size L and a pair of wells separated by a
Euclidean distance r. We can ask question about the probability that the two wells are
connected, or in percolation language, in the same cluster. This is just the two point
correlation function defined previously. Suppose we want to know what fraction of
the sand in contact with the wells is connected to both wells. This is just the
connectivity function P defined earlier. We can use finite size scaling to estimate this
fraction. Also we can use related scaling laws to estimate the uncertainty. Note that
these are algebraic laws with no spare parameters. The percolation threshold is
defined by the shape of the objects, but it is largely unimportant whether we model
the sand units as rectangles or ellipsoids or other shapes (provided they are not too
exotic). The scaling laws and exponents are determined from lattice models (and this
has been done very extensively in the literature) and can be straightforwardly applied.
As the expressions are simple algebraic they are very rapid to calculate. Compare this
with the building of a typical 3D reservoir model which can be computationally very
expensive. Here we show results for a North Sea fluvial system of intermediate net to
gross. A conventional cross sectional model was built to determine the connected sand
fraction. The sand bodies were very long which is why a cross sectional rather than a
full 3D model was built. The sand bodies had typical dimensions of 300mx2m and the
reservoir interval was 100m thick. Well spacing was taken as 1km. Sensitivities to
the net to gross ration were considered as this was uncertain. These results were
compared with the predictions from percolation theory. It can be seen that the two are
in good agreement. In particular the percolation approach is able to estimate the
uncertainty which is not possible for the single realisation reservoir models. The
percolation calculations were done in a fraction of a second in a spreadsheet.

Net to Gross ratio
Connected sand fraction
(conventional model)
Connected sand fraction
(percolation prediction)
25% 6% 2.7 4 %
35% 10% 10 8 %

Percolation can do more than predict static connectivity. There are scaling laws for
the effective permeability. It may be noted that percolation clusters have dead ends
which cannot be swept. Only oil in the so called backbone can be swept. This is a
fractal object of known dimensions and so obeys a scaling law and can be estimated.
We shall describe in more detail how breakthrough time can be estimated.

Estimation of breakthrough time
For simplicity we consider a pair of wells (one injector, one producer) separated by a
distance r. we shall also only consider transport of a passive tracer (or a unit mobility
ration miscible flood). We can then determine the probability distribution (over
different realisations of the sandbody locations). This is the conditional probability
that the breakthrough time is t
br
given that the reservoir size (measured in
dimensionless units of sandbody length) is L and the net to gross is p, i.e. P(t
br
|r,L,p).
In previous studies (Dokholyan et al., 1998; Lee et al., 1999) we have shown that this
distribution obeys the following scaling

P(t
br
|r, L, p) ~

1
1 2 3
r
t
r
f
t
r
f
t
L
f
t
d
b r
d
g
b r
d
b r
d
b r
d
t t
t
t t t
[
\
|

)
j
[
\
|

)
j
[
\
|

)
j
[
\
|

)
j
~

(1)


f
1
(x) = exp(-a x
~
)
f
2
(x) = exp(-b x
q
)
f
3
(x) = exp(-c x

)

Currently the best estimates of the various coefficients and powers (as found from
detailed numerical experiments on lattices and theory, see Andrade et al. 2000) in this
are:
d
t
= 1.330.05 ; g
t
= 1.900.03 ; a = 1.1 ; b = 5.0 ; c = 1.6(p<p
c
) 2.6(p>p
c
)
= 3.0 ; = 3.0 ; = 1.0 and = |p - p
c
|
-
= 4/3 ; p
c
= 0.6680.003 (for
continuum percolation).

In this paper we will not discuss the background to this scaling relationship but
concentrate on how well it succeeds in predicting the breakthrough time for a realistic
permeability field. However, it is sufficient to mention that this is typical of the kind
of scaling result that percolation is able to provide. However, it is worth spending
some time describing the motivation behind the form of the various functions. The
first expression (f
1
) is an extension to the expression developed by others (see Havlin
& Ben-Avrahim, 1987 for a detailed discussion) for the shortest path length in a
percolating cluster between two points. The breakthrough time is strongly correlated
with the shortest path length (or chemical path).

To this there are some corrections for real systems. In a finite size system very large
excursions of the streamlines are not permitted because of the boundaries so there is a
maximum length permitted (and also a maximum to the minimum transit time). This
cut-off is given by the expression f
2
. Away from the percolation threshold the clusters
of connected bodies have a typical size (given by the percolation correlation length,
) which also truncates the excursion of the streamlines. This leads to the cut-off
given by the expression f
3
. The multiplication together of these three expressions is an
assumption that has been tested by Dokholyan et al. (1998). Also a more detailed
derivation of this form is given there and the references therein. Here we shall
concentrate on using this scaling form to make predictions about the distribution of
breakthrough times for a realistic data set.

Rather than considering the theoretical aspects we shall look at an application to real
field data. We took as an example a deep water turbidite reservoir. The field is
approximately 10km long by 1.5km wide by 150m thick. The turbidite channels,
which make up most of the net pay (permeable sand) in the reservoir, are typically
8km long by 200m wide by 15m thick. These channels have their long axes aligned
with that of the reservoir. The net to gross ratio (percolation occupancy probability, p)
is 50%. The typical well spacing was around 1.5km either aligned or perpendicular to
the long axis of the field. In order to account for the anisotropy in the shape of the
sand bodies and the field we first make all length units dimensionless by scaling with
the dimension of the sand body in the appropriate direction (so the field dimensions
are then L
x
, L
y
and L
z
in the appropriate directions). Then scaling law for the
breakthrough time can be applied with the minimum of these three values (L=min(L
x
,
L
y
, L
z
) ). The validity of using just the minimum length has been previously tested
(Andrade et al., 2000).

The real field is rather more complex than this, and a more realistic reservoir
description was made and put into a conventional flow simulator. We could then enter
these dimensions into the scaling formula. It should be noted that first the
dimensionless units were converted into real field units to compare with the
conventional simulation results. Using these data we find breakthrough times of
around one year. The full probability distribution of breakthrough times from the
scaling law is given by the solid curve in the figure below.
0.0
0.2
0.4
0.6
0.8
0 1 2 3 4
Breakthrough Time (years)
P
r
o
b
a
b
i
l
i
t
y




In addition conventional numerical simulations were carried out for the field. We
could then collect the statistics for breakthrough times for the various well pairs to
compare with this theoretical prediction. Not all pairs exhibited breakthrough in the
timescale over which the simulations were run and there were only three injectors so
there were only 9 samples. The histogram of breakthrough times is also shown on the
figure. Clearly with such a small sample these results cannot be taken as conclusive
however, certainly they are indicative that the percolation prediction from the simple
model is consistent with the results of the numerical simulation of the more complex
reservoir model. The agreement with the predictions is certainly good enough for
engineering purposes. We would hope that if the simulation had been run for longer
and more well pairs had broken through that better statistics could have been
collected. The main point being that the scaling predictions took a fraction of a second
of cpu time (and could be carried out on a simple spreadsheet) compared with the
hours required for the conventional simulation approach. This makes this a practical
tool to be used for making engineering and management decisions.

Conclusions
In this article we have introduced many of the concepts of percolation theory and
described the simplest of the scaling laws used to describe the behaviour. We have
also demonstrated that it can be a practical tool for answering many of the questions
that arise when considering geometrically complex systems where connectivity is the
primary issue, such as hydrocarbon recovery from oil reservoirs.

There are many issues which we have not considered. suppose the bodies are not
placed independently of each other (so called correlated percolation) such as may be
found for stacking patterns of particular sand units, or the distribution of faults. Does
this alter the percolation properties? Well essentially no, the same scaling formalism
applies. If the spatial correlation between the bodies is of a finite range then this
doesnt alter any of the previous discussion. If the spatial correlation has no finite
range (such as a power law correlation function) then some of the critical exponents
may change their numerical value.

If the permeability cannot be simply split into good and bad then again the approach
has to be modified slightly but if the permeability distribution is very broad we can
apply a cutoff to the permeability such that we are sitting just at the percolation
threshold. This cutoff value then dominates the flow.

These issues, and other, extensions to the simple predictions described here (such as
post breakthrough behaviour) are the subjects of our current research into applying
percolation theory to oil recovery problems.