account for fluid flow directly; they are exemplified by the Dykstra-Parsons and
measures include dispersivity and chanelling factors which directly relate to the
efficiency of a miscible displacement. We elucidate the distinction between
these two phenomena and review the means by which they are quanitified.
INTRODUCTION
both theoretical and field studies examining the impact of heterogeneity and
describing techniques to assess and mitigate its effects. Usually, the theoretical
SPE 2 01 56
2
DEFINITION OF HETEROGENEITY
locally
SPE 2 015 6
3
account for all variations in reservoir properties which impact the process
under test.
process used during the flow experiment closely parallels the process which is
expected to be applied to the reservoir, the results are most directly applicable
experiments.
are frequently used because of the relatively low cost to obtain estimates and
~Var(k)
C v = E(k)
random variable ln(k) is normally distributed with mean Jl and variance cr2.
Cv = ~ exp(cr2)-1. ( 1)
in the definition
V _ log(ko.so) - log(ko.a4)
- log (ko.so) ·
SPE 2 015 6
8
This indeed could be the case if the permeability distribution is log-normal and,
for some reason, we wanted to· have a measure of the variability in the log
which is what fluid flow responds to--the quantity V is probably misleading for
two reasons. First, as Eq. (1) shows, the coefficient of variation strictly
depends upon the difference log(ko.so) - log(ko.a4) i.e., the ratio ko.so/ko.a4 with no
explicit dependence upon the median. V can change without Vop changing and
vice versa. Second, the form of Eq. (2) where, in the numerator, percentiles are
notion of "robust" statistics. From this point of view, averages and standard
percentiles because the latter are less subject to influence by erroneous data.
on to correlate values of Vop with oil recovery from core-scale water floods.
Since their study, several reports11,12 have appeared examining and discussing
the merits of the correlation. Other studies11-16 have used the coefficient in
low sensitivity of models to variations in Vop when Vop ~ 0.5 while, for the large
reflects, in part, the nature of Vop which is to force all situations into a finite
discerned.
SPE 2 01 56
9
subtle but significant changes to the procedure for Vop estimation while still
calling the result "the Dykstra-Parsons coefficient" or referencing Dykstra and
Parson's work. For instance, two authors17,18 do not mention the relative
weighting of the central and extreme portions of the plot to determine the "best-
fit" line. This omission increases the number of possible "equivalent" reservoirs.
Another, more significant change is that several descriptions12,15,19,2o do not
mention the "best-fit" line at all. The quantities ko.so and k0 . 84 are estimated
directly from the data. This redefinition eliminates the subjectivity of the line
where kA and kH are the arithmetic and harmonic averages of the permeability
data. Eq. (3) follows from two features of the log-normal distribution:
E (k)=exp(J.1+0.5cr2) and E(k-1 )=exp( -J.L+0.5cr2). 3
data is being used in the estimation procedure. Jensen and Lake9 show that
and
SPE 20156
10
sv = 1.49•a•exp(-a)tfr1,
A
where mv is the bias and sv is the standard error of V DP· The bias is usually
quite small but, nonetheless, causes the estimates, on average, to
underestimate the true Vop. The situation is more significant in the case of the
standard error. For example, if Vop=0.60, then to have a two-in-three chance of
A
0.55::::;Vop::::;0.65 requires that 120 samples be used in the estimation.
A
When the permeability is not log-normally distributed, V DP values
obtained from estimates of ko.so and ko.s4 not longer bear a relationship to Vop
values obtained using either the "best-fit" line or assuming the permeabilities
are log-normally distributed .. Jensen and Currie21 establish some relationships
assuming a family of forms for the permeability distribution.
Lambert1 o also investigated the effects of including porosity, <f>, in the
A A
computation of Vop. She compared Vop obtained using permeability only with
A
V oP based on the ratio kl<f>. The ratio kl<f> reflects the interstitial velocity of a
V ko.1s - ko.so
c- ko.so · (4)
SPE 2 015 6
11
V c. ko.so - ko.s4
ko.s4
Vop and Vc are related by Vop = Vc/(1 +VcY. While Q$;Vop$;1, Q$;Vc$;oo so that
limit. Claridge23 states that V c "gives a somewhat more linear relationship with
actual permeability variation." He could also have considered using a instead
exp(•), they are equal to the first order. From the same type of analysis used by
Jensen and Lake,9 the bias and standard error properties of 'fJ c, when estimated
A A A
using the percentile estimates ko.so and ko.s4, are identical to those for Vop.
Jensen and Lake9 proposed a pair of heterogeneity measures, U and p,
where
U _ ko.1s- ko.s4
- ko.s4
in a different way. Rather than propose a new measure, they suggest a more
efficient method of estimating V0 p. They use the asymmetry factor p along with
" which have 50°/o less
a maximum likelihood approach to obtain estimates Vop
statistical error than the traditional method. This means that only one-quarter of
the data required for a Vop estimate based on percentile estimates is needed to
is the Lorenz coefficient. The Lorenz curve (Fig. 2) is a plot of cumulative flow
(5a)
and
(5b)
permeability so that k(1) is the layer with thickness h(1) and the largest
permeability while k(n) is the layer with thickness h(n) and the smallest
layer ordering, the Lorenz curve monotonically increases from m=1 to m=n with
in Fig. 2) between the Lorenz curve ABC and the diagonal AC. If the medium is
homogeneous, all the permeability values are identical and the Lorenz curve is
A
the straight line AC. Hence, Lc=O. Increasing levels of heterogeneity are
indicated by movement of the Lorenz curve, ABC, away from the diagonal AC
A A
with Lc increasing but always less than unity. Typical values for Lc for
00
and
00
x is a parameter which varies from 0 to oo along the Lorenz curve. For the
Lorenz curve to exist, E(k)<oo, which may not always be the case.9
00 k1
Lc = 1 - E~k) J p(k1) J k2p(k2)dk2dk1
0 0
and O~Lc~1. While the Lorenz curve is uniquely defined (to within a scale
factor) for any given distribution,2s Lc is not uniquely associated with any
SPE 2 015 6
14
As was the case for the Dykstra-Parsons plot, there is an inherent issue
of probability associated with the Lorenz plot. Comparing Eqs. (5) and (6), we
represented by each of the n intervals, the height may not accurately reflect the
the Stiles27 and the Suder and Calhoun2B models. Schauer,29 in a field study of
required for the first production increase. Lake et al.3o used Lc when they
i=m
L<t>(i).h(i)
i=1
Cm=-.- - -
l=n
L<t>(i).h(i)
i=1
SPE 2 01 56
15
If porosity is constant, the Lorenz curve remains unaltered. The data must be
ordered according to the ratio kl<j>. The inclusion of porosity variations will not
A
substantially increase Lc.
(7)
is equivalent to using the trapezoidal rule for the area. Note that Eq. (7)
requires no ordering of the data and assumes all k's have equal probability.
An analytic assessment of the sampling properties of Lc, such as that
done for \top when permeability is log-normally distributed, is not simple; the
results can only be expressed in terms of integrals which must be numerically
evaluated. Jensen and Lake9 show the results of a Monte Carlo analysis,
assuming permeability to be log-normally distributed, for bias and standard
error. Lc is substantially negatively biased (greater than 5°/o) for small sample
sizes (n<40) and heterogeneous distributions (Lc>0.6). Thus, uncompensated
heterogeneity assessments using Lc will understate the heterogeneity in the
reservoir. It is, however, more precise than the Dykstra-Parsons estimator.
SPE 2 015 6
16
1. Establish a field-wide correlation from the logs and other available data.
2. At each well and for each unit or layer, estimate the net-to-gross ratio, nij,
where i is the well number (i=1, 2, ... , q wells) and j is the layer or unit
s.2J --q-1.
-
1 t
1=1
(n··IJ -N·)2
J
4. H F i• the heterogeneity factor for the jth unit or layer, is given by the
S·
HFJ·-~
-N·· J
basis if possible, the HF's for all layers and/or units within the deposit.
coefficient of variation.
Polasek and Hutchinson also listed the HF values for nine reservoirs and
their qualitative recovery behavior. While the correlation was generally good,
SPE 2 0 l!) 6
18
the authors point out that process parameters (e.g., mobility ratio) also may
have influenced the reservoir performances listed. They also report, for five
watered-out reservoirs, the mobility ratio, HF, and bypass factor. Again, the
correlation between HF and bypass factor was moderately good.
reflects prefect correlation of a reservoir property between two wells while <;=0
wells, the variability of the property is about the same at the two wells, whereas
A.;eO represents a change in the variability. t is the ratio of gross pay between
two wells. Pirson makes some general remarks about the relationship of these
three measures to reservoir performance, but does not give any specific
examples.
directly from fluid flow. These measures are more complicated than static
measures because they also depend on the properties of the fluids comprising
which obviates dealing with the troublesome topics of wettability and local
surface characterization.
simultaneously released into a flowing stream at the inlet end of the medium
course, no two particles can occupy the same location and isoconcentration
scale variance is a~ The difference between these two (aE-a~), the channeling
variance, is a measure of the tendency of the fluid flow to channel through the
medium. Subtracting variances like this implies that channeling and mixing
occur independently. Whether such independence, in fact, exists is unknown;
however, indirect evidence regarding the growth rate of unstable miscible
displacements 35 suggest that there is some degree of dependency. This
difference, which is always non-negative, approaches zero as the
isoconcentration lines become vertical and parallel. On the other hand, a~
approaches zero in the limit of all the particles remaining on the same
isoconcentration line during the flow.
The physical quantity describing the mixing is the dispersion coefficient
which, above some relatively low velocity, is proportional to the local fluid
velocity. The constant of proportionality is the dispersivity, the basic dynamic
measure of heterogeneity which describes mixing. A displacement dominated
by this type of mixing exhibits good lateral sweep efficiency and a variance
which grows asymptotically in proportion to (travel distance)112 or t1/2.
SPE 2 015 6
21
Autocorrelation
Autocorrelation is a statistical measure of the likelihood of two quantities
spatially separated by a distance h having the same value. There are a large
number of empirical functions describing autocorrelation. The impact of
autocorrelation functions on dynamic heterogeneity measures is discussed at
some length in Gelhar and Axness36 and Arya37. Correlation models range
from simple exponentials having only a single scale, through models having
multiple scales38 up to quite complex representations which have a continuum
of correlation scales39 Unfortunately, much of this complexity is actually
observed in the distribution of permeability in realistic media. 4 0
Autocorrelation is important because it determines whether the fluid flow
is mixing or channeling dominated. Flow in media where autocorrelation is
nonexistent or sm·all (compared to the medium length) tends to exhibit mixing 14.
Such media are called random or uncorrelated. The absence of channeling in
these cases is evident even at large heterogeneity levels and modest degrees
SPE 2 015 6
22
of instability 41 ,42 except when there is a large amount of mixing. When the
autocorrelation length is even a modest fraction of the medium length, the fluid
flow will tend to channel. Media with very long autocorrelation lengths are
feature which has been impossible until the advent of imaging. Thus,
Dispersivity
this paper. The most recent contribution by Dagan 45 shows how the tensorial
(3)
In this equation and all other versions of such statistical transport properties,
the property is related to the statistics of the velocity field, not to those of the
and velocity statistics does not exist, being influenced by flow geometry,
boundary conditions and even the type of calculation method employed.45.46
Channeling Factors
Reservoir engineers, on the other hand, have tended to represent a~
through channeling factors. Such factors have been intimately associated with
the propagation of instabilities47-53 and most workers have not attempted to
separately account for heterogeneity. One of the few which does, however, is
the heterogeneity factor HK in the work of Koval. 47 HK is defined as the pore
volumes of fluid required to bring the effluent concentratio~ of a stable, miscible
displacement up to 98°/o of the injected concentration. This is about the most
direct measure of fluid flow performance that a heterogeneity measure can have
because 1/HK now represents the pore volumes of fluid injected before
breakthrough. HK ranges between 1.5 and 5 in laboratory cores 47 but can be
much larger for field-scale displacements. For uniformly layered media HK is
empirically related to V DP by 13
(8)
See also reference 23. Such a relationship has not been established for other
types of heterogeneity. HK, as expressed through Eq. (8), does an excellent job
of "straightening out" the nonlinearity in V 0 p described previously. For
example, changing V 0 p from 0.2 to 0.3 changes HK (and therefore the
breakthrough and sweepout times) relatively little; however, changing V 0 p from
0.7 to 0.8, the normal range from routine core data, changes HK substantially.
SPE 2 015 6
24
Scaling
Unfortunately, the dynamic measures do not scale well. Scaling is the
process of calibrating a measure against laboratory data and then using the so-
calibrated model in predictions. The reasons behind the lack of scaling are
many-fold: sensitivity to fluid properties, scale-dependence of media
correlation, and lack of a sound theoretical base for these measures.
We can, however, tie all of these factors together through the notion of
flow regimes, illustrated schematically in Fig. 4. The large-scale variance of a
displacement in a realistic medium grows neither strictly linear with distance nor
with the square root of distance. Instead, there may be portions which are of
one type of the other separated by transitions regions. The change in the rate
of growth may be caused by a number of factors: progression to larger scales of
heterogeneity, stabilization of instabilities or the initiation of instabilities from
large-scale perturbations. The actual at in Fig. 4 grows initially with
(distance) 112 and then changes to linear growth.
Scaling entails matching at at some small scale (laboratory in Fig. 4)
and then using this match to prediction of at a large scale (field). Using a
mixing model, where the variance grows as (distance)1/2 throughout, yields a
prediction of aE which is too small. To match aE at the field scales requires an
increase in the apparent dispersivity and, indeed, experience has shown that a
does appear to increase with length of measurement scale.1 4 On the other
hand, scaling with a channeling model will overpredict the field-scale variance.
Experience has also shown that such models must be attenuated in order to
match field results. 49
whatever manner is desired or (b) to used both the channeling and mixing
models simultaneously. The latter solution is by far the better solution since
SUMMARY
features of such flows and may provide a basis for classification. Past
experience on the last point has not been encouraging; however, such
ACKNOWLEDGEMENTS
Technology Program; and the Center for Enhanced Oil and Gas Recovery
Engineering.
NOMENCLATURE
l
Cm = Cumulative storage capacity
Cv = coefficient of variation
Hm = Cumulative thickness
k = permeability
K = sum of permeabilities
n = number of data
nij = net-to-g ross ratio for ith well and jth layer
q = number of wells
X = dummy variable
SEE 2 015 6
27
a. = dispersivity
<t> = porosity
Super- or sub-scripts
A = arithmetic average
H = harmonic average
(i) = ith element in an ordered set
0.16 = 16th percentile
0.50 = 50th percentile (median)
0.84 = 84th percentile
{\
= estimated quantity
APPENDIX A
concentration, G, and the trapezoidal rule for estimating Lc. We assume that
each datum, ki, has probability~ and is ordered such that k 1 ~k 2 ~····~kn. From
n
where ko = 0 and K = L',ki. The following steps then yield the desired result.
i=1
1 n
2nKI, [
j=1
n
i=j
:»i -
.
Jkj];
and, since the last two of the three right-side terms are each zero. we obtain
REFERENCES
Larry W. Lake and H. B. Carroll (eds), Academic Press, Inc., New York
(1986), 141-175.
41. Waggoner, John R., Ph.D. Dissertation, The University of Texas at Austin,
in progress 1989.
42. Araktinga, U. G. and Orr, F. M. Jr., "Viscous Fingering in Heterogeneous
Porous Media," SPE 18095, presented at the 63rd Annual Technical
Conference and Exhibition in Houston, Texas, October 2-5, 1988.
43. Dagan, G., "Stochastic Modeling of Groundwater by Unconditional and
Conditional Probabilities, 2. The Solute Transport," Water Resour. Res.,
18, 835-848, 1982
44. Gelhar, L. W. and Axness, C. L., "Stochastic Analysis of Macrodispersion
in a Stratified Aquifer," Water Resour. Res., 15, 1387-1397.
45. Dagan, Gideon, "Time-Dependent Macrodispersion of Solute Transport
in Anisotropic Heterogeneous Aquifers, Water Resour. Res., 24, no. 9, pp.
1491-1500.
46. Yang, An-Ping, Ph.D. Dissertation, The University of Texas at Austin, in
progress 1989.
47. Koval, E. J., "A Method for Predicting the Performance of Unstable
Miscible Displacement in Heterogeneous Media," Soc. Pet. Eng. J., 3
(June 1963}, 145-154.
48. Dougherty, E. L., "Mathematical Model of an Unstable Miscible
Displacement in Heterogeneous Media," Soc. Pet. Eng. J., 3 (June
1963}, 155-163.
49. Todd, M. R. and Longstaff,W. J., "The Development, Testing, and
Application of a Numerical Simulator for Predicting Miscible Flood
Performance," J. Pet. Tech., (July 1972}, 874-882.
50. Fayers, F. J., "An Approximate Model with Physically Interpretable
Parameters for Representing Viscous Fingering," SPE Reservoir
Engineering (May 1988), 551-558.
SPE 2 015 6
34
"Best-fit" line
Probability Scale
1.0
0.0
0.0 1.0
Mixing:
Average cr8
C(x,y,t)
lsoconcentrat1on
Mixing+
Channeling
C(x,t)
Channeling
(-t')
\
Q)
()
c:
<U
·.:::::
<U
>Q) '-Mixing
<U
()
(-t 1/2)
CJ)
I
Q)
0>
L.
<U
_J
Lab Field
Scale Scale
Distance
1.0
I Area B I
0.0
0.0 H1 1.0