Statistical Scale-Up of Reservoir Properties: Concepts and Applications

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

Journal of Petroleum Science and Engineering 44 (2004) 27 39

www.elsevier.com/locate/petrol

Statistical scale-up of reservoir properties:


concepts and applications
Larry W. Lake *, Sanjay Srinivasan
Department of Petroleum and Geosystems Engineering, University of Texas at Austin, 1 University Station Stop CO300, Austin, TX 78712, USA

Abstract
All petrophysical quantities are used at a scale different from the one on which they were measured. This necessitates an
adjustment of the measured values before they are used to develop a reservoir model, a step referred to as scale-up. Scale-up is
complicated by the properties being heterogeneously distributed in space and self- or autocorrelated. The autocorrelation means
that the heterogeneity itself must be scaled up so that the adjusted measurements correctly reflect the property at the coarser
scale.
This paper attempts to understand the uncertainty in assigning scaled up values to finite regions of space. The region may be
a formation interval or a cell thickness for numerical simulation. We use the notion of the variance of the mean of a random
variable to understand the scale-up process. The behavior of the variance of the mean is used to investigate the definition of a
representative elementary volume (REV), and of the behavior of lateral and vertical permeabilities with scale and the resultant
impact on uncertainty distributions for reservoir properties.
The paper demonstrates that the notion of the variance of a mean can be used to understand and refine the concept of
representative elementary volume. The change in horizontal and vertical permeability with scale can also be explained using the
variance of the mean using reasonable autocorrelation functions. The impact of scaling up on the autocorrelation structure of the
simulated field is demonstrated in the paper. The results point to the importance of employing rigorous procedures to scale
heterogeneity in order to derive robust estimates of uncertainty.
D 2004 Elsevier B.V. All rights reserved.
Keywords: Scale-up; Variance of the mean; Representative elementary volume; Kriging; Block Kriging; Permeability scaling; Scaling
heterogeneity

1. Introduction
Recognizing and reconciling differences in measurement scales associated with data from different
sources is an important aspect of reservoir modeling.
Typically, the measurement scale (fine scale) is

* Corresponding author. Fax: +1-512-471-3161.


E-mail address: larry_lake@mail.utexas.edu (L.W. Lake).
0920-4105/$ - see front matter D 2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.petrol.2004.02.003

smaller than the application scale (coarse scale). Since


properties are heterogeneously distributed in space
and self- or autocorrelated, the heterogeneity itself
must be scaled up so that the adjusted measurements
correctly reflect the property at the coarser scale. That
this is so is often overlooked, with scale-up traditionally being done on the properties only, not on their
heterogeneity. As demonstrated below by analytical
calculation and simulation, the intrinsic heterogeneity
plays a role in the scale-up. Furthermore, traditionally

28

L.W. Lake, S. Srinivasan / Journal of Petroleum Science and Engineering 44 (2004) 2739

scaled up quantities and fine-scale measurements can


easily have a different spatial distribution from each
other.
Let Z be a scalar, spatially continuous, Gaussian,
random variable distributed in one-dimensional field.
The variance of the average of Z over a distance L is
given by:
Z nL Z gn

2r2

VarZ 2
qgdgdn
L
n0
g0

where r2 is the variance of the entire field (the


population variance), and q is the spatial autocorrelation function. The variance of the mean should
decrease as L increases; as L approaches zero, the
variance of the mean approaches the population
variance since the averaging volume is now a point.
The variance of Z at a point within L is expected to
increase as L increases (properties within large volumes are more heterogeneous than within small volumes). This is given through Kriges relationship
(Journel and Hujbregts, 1978):
r2o=D

r2o=L

r2L=D

2
where ro/D
is the variance of Z at a point (o) in a large
2
volume D, ro/L
is the variance of Z at a point in a
2
small volume L ( < D), and rL/D
is the variance of the
average of Z over L within D.
2
2
In the notation of Eq. (1) Var(Z) = rL/D
and ro/D
=
2
r . These equations have identified the population
with the volume D. We will call L the measurement
support below. The variance of a point within L is
then given by:

r2  2r
VarZ r2  VarZ
L2
Z nL Z

n0

gn

Considering two types of autocorrelation functions:


Multiscale (K) exponential:

k K
X

fk eg=kk

k1

That gives rise to the following analytical expression


for Var{Z}


k K
2r2 X
kk
L=kk

VarZ
fk kk 1  1  e

5
L k0
L
This model allows specification of the scales of the
variability of Z (through the kk) and the fraction that
each contributes to the total variance (through the fk).
These scales can, in turn, be related to other observations on the basis of laboratory, bed and interval
scales, or even to P
such vagaries as micro, macro and
mega. Since 1 kK
k1 fk , Eq. (4) is a 2K-parameter
model ((K  1)fk, r2, and Kkk).
n The stable autocovariance model:
q eg=k

a is a roughness parameter. Many geophysical data


seem to be fit by this model wherein a is usually
around 0.2 (Neuman, 1994). When Eq. (6) is used, the
range parameter k is usually set to the volume size D.
The limits of Eq. (6) are instructive; a = 0 is a pure
nugget model; a = 1 is a (single scale) exponential;
and a = 2 is the Gaussian model (Jennings, 1999). The
variance of the mean cannot be integrated analytically
from Eq. (6), but it can be expanded as a series and
integrated term-by-term to give (Mahadevan et al.,
2003):
 ma
ml
X
1k Lk
2

VarZ 2r
7
m!ma 1ma 2
m0

qgdgdn

g0

3
A difficulty with Eq. (3) is deciphering what constitutes a point and a volume in data. Nevertheless, all the integrals in the expression can be
evaluated, analytically for simple autocovariance
functions, and numerically for nearly all other of
interest.

The series converges rapidly for small a.


Expressions for the variance of the average can be
developed for combinations of the three-scale and the
stable models. The following section discusses the
application of Eqs. (5) and (7) to explain commonly
observed scaling phenomena. Robust models of uncertainty are possible only after complete understanding of the scale-up characteristics of the reservoir
attributes has been gained.

L.W. Lake, S. Srinivasan / Journal of Petroleum Science and Engineering 44 (2004) 2739

29

2. Representative elementary volume (REV)


The most fundamental postulate in practical reservoir engineering is the existence of the representative
elementary volume abbreviated as REV. Fig. 1 illustrates the traditional explanation (Bear, 1972).
The plot shows the average of porosity on the y-axis
against the averaging volume V (x-axis) (any other
petrophysical quantity could be on the y-axis). Beginning at large V, the plot shows that the average of the
property changes as V decreases, becoming stable
(constant) at intermediate V, and then beginning to
fluctuate again as V approaches zero. The REV is the
volume of V at which the small-scale fluctuations begin
when approached from large V. Bear refers to the largescale fluctuation as inhomogeneity (heterogeneity).
Fig. 1 is the result of a thought experiment because
an actual measurement cannot be doneat least not
over enough scales to confirm the behavior shown in
the plot. V is meant to be a three-dimensional volume.
In this work, L is a one-dimensional proxy for V.
Another way to look at Fig. 1 is in terms of the
variance of the mean of the porosity. Were this to be
plotted on the same scale as in Fig. 1, this variance
would be large at the large scale, zero (or small) at the
intermediate scale and become large again for V less
than the REV. Interestingly, this behavior cannot
happen for the autocorrelation models discussed
above, or, for that matter, any two point autocorrelation function that decreases with distance up to a
range and stabilizes to the stationary variance thereafter. The reproduction of the variability at the large

Fig. 2. Variance of the mean of Z for Eq. (5) (normalized by the 2r2
term) as a function of L for a K = 3 model where f1 = 0.9, f2 = 0,
f3 = 0.1, k1 = 0.01, k2 = 1, k3 = 100. The solid straight line has slope
of  1, which indicates spatially independent Z.

scale is not possible for such a stabilized semivariogram. Fig. 2 illustrates.


In Fig. 2, the small-scale variability is 90% of the
total variance, and the large-scale variability contributes the remaining 10%. The intermediate scale
variability has been omitted. The two remaining
scales are clearly visible on the x-axis as the starting
and ending locations of the knee. The slope of the
plot for L>k3 = 100 approaches  1 as it should for
independent Z. As is evident from Fig. 2, the
variance of the mean continually decreases as length
increases and the rate of the decline ultimately
approaches a slope of  1 at length scales greater
than the largest range of correlation of Z, in this case
k3 = 100.
We can use the information in Fig. 2 to generate a
plot comparable to Fig. 1 by letting Z be porosity. Z is
now a locally binary (ones and zeros) variable signifying pore or grain spaces at points within a volume L.
The average porosity of the volume is simply the
expected value of the binary indicator and the population variance of porosity is:
r2 Var/ E/1  E/

Fig. 1. The representative elementary volume (REV) concept


applied to porosity. Adapted from Bear (1972).

If the random variable Z is assumed to be Gaussian


with mean of E(/) and variance Z obtained using Eq.

30

L.W. Lake, S. Srinivasan / Journal of Petroleum Science and Engineering 44 (2004) 2739

(8), then realizations of average porosity Z can be


generated by Monte Carlo sampling. Fig. 3 shows one
such realization.
Fig. 3 resembles Fig. 2 in that the curve becomes
especially erratic as the averaging distance L
approaches zero. Furthermore, the average porosity
becomes less variable as the volume increases and
approaches the population average for large L. There,
however, is one significant difference between Figs. 1
and 3. There exists no region of stable porosity at
intermediate distance even though the parameters
were chosen (the contribution for the middle scale is
zero and the contribution from the extreme scales are
factor of ten different) to emphasize such stability.
These results are consistent with general observation. The variability of the average porosity in a set of
1-in.-diameter core plugs is not zero (albeit the
variability is probably less than that shown in Fig.
3), and this variability is certainly smaller than the
variability of porosity on the scale of say 100 Am.
What is surprising, however, is that a substantial
fraction of the observed small-scale variability comes
from large-scale heterogeneity. The large-scale heterogeneity, which starts at L = 100 in Figs. 2 and 3,
persists to all smaller scales. In other words, the only
way to obtain a stable average is for the averaging
volume to exceed the largest scale of the heterogeneity. This is an unsatisfactory outcome because it says
that the REV, as defined by Fig. 1, is the largest scale

of heterogeneity in the field. Since it is known that the


largest scale of heterogeneity can be very large
approaching the field or even basin scale, this definition suggests that any cell-by-cell computation or any
measurement is below the REV scale and, hence,
there is added variance (uncertainty) of the modeled
porosity values.

3. Uncertainty assessment accounting for scale-up


The preceding section discussed the systematic
change in the variance of a petrophysical property
with volume. Here, we discuss how scale-up affects
assessment of uncertainty and spatial autocorrelation.
The variance is a measure of the uncertainty in
assigning a property to a reservoir block of certain
volume support. The uncertainty in assigning reservoir properties to two blocks is represented through
the block autocovariance between two points a lag
distance h apart. The scale-up of values measured at a
small scale causes systematic changes in the block
autocovariance structure, in a manner similar to the
changes observed above for the variance of the mean
(Frykman and Deutsch, 2002). This means that the
scale-up will affect the assessment of uncertainty in a
property, and this is demonstrated by the simulations
in this section. These simulations yield the uncertainty
of the petrophysical properties caused by the disparity

Fig. 3. One realization of mean porosity for E(/) = 0.3. Same semivariogram parameters as in Fig. 2.

L.W. Lake, S. Srinivasan / Journal of Petroleum Science and Engineering 44 (2004) 2739

in volume support (the value of L above) of the


respective data.
Data integration is often a primary quest of reservoir modeling. Several techniques for assimilating
data from different sources have been proposed (Varela et al., 2002; Datta-Gupta et al., 1995). Data
integration that specifically accounts for the volume
scaling relationships has been less well studied, despite the considerable research effort expended on up
scaling fine-scale geological models to coarse reservoir flow simulation grids. Data integration, in the
case of scale-up for flow simulation, consists of using
petrophysical properties that are measured at small
scale (often on a core) support and then analyzing
these data for autocorrelation measures. The measures
inferred from the point support data are used for
simulating spatial variations of the phenomena at
block support. Therefore, the scale-up of both variance and autocovariance is an issue.
The semivariogram of block support values c(V,VV)
can be expressed in terms of the semivariogram of
point support data c(m,mV):

; V V
cV

V X
VV
X
1
cm; mV
AV NV VA m1 mV1

Eq. (1) expressed the variance of the mean ZV over


the volume V as a function of the fine-scale semivariogram cm(h). Eq. (9) extends that notion of variance of the mean to the semivariance between the
means ZV and ZV V of two blocks; specific comparisons between Eqs. (1) and (9) are:
(a) Eq. (9) is written in terms of the semivariance
rather than the autocorrelation function. Eq. (1)
can be converted to this form using Eq. (2).
(b) The origin and destination scales are notated
differently. In Eq. (1), the origin scale is a point
(o); in Eq. (9), it is a finite volume V V. The
destination scale in Eq. (1) is L; in Eq. (9), it is V.
(c) Eq. (9) is for multi-dimensions whereas Eq. (1)
is in one dimension. The sums in Eq. (9), which
are approximations to integrals, are threefold
integrations.
The discrete sums in Eq. (9) occur because the
continuous reservoir domain is commonly assumed to

31

be pieced into a discrete set of cells for reservoir


modeling. The cells represent the local or point
support of the property. An important remark pertaining to Eq. (9) is that at a volume support V greater
than the REV, the block-to-block autocovariance
c (V,V V) simply reverts to the prior block variance since
the means of the blocks are then independent of each
other. For any scale less than the REV, the c(V,V V)
value will be finite and different from the variance of
the mean in Eq. (1).
Integration of measured values at a small scale into
the geological model requires that the cross-autocovariance (or semivariance) between the small-scale
values and the cell values also be known. These can
be expressed as:
V; m
cV

VV
1 X
cm; mV
AV VA m V1

10

Reiterating, if the volume support V is greater than the


REV, the cross-autocovariance between a block and
any point-measured value (left-hand side of Eq. (10))
falling outside the block will identify the prior block
variance.
These scaled up variance and auto/cross-variances
(or semivariograms) can be used to model the spatial
variability of a petrophysical property. The sequential
approach to stochastic simulation consists of constructing conditional cumulative distributions at each
simulation location by Kriging. The mean of the
conditional distribution (assumed Gaussian) is given
by:
X
X
Z*
ka Z a
tb Z b
11
a

Za, a = 1,. . .,n are the conditioning block data and Zb,
b = 1,. . .,nVare the measured small-scale data. If only
small-scale data values are available for modeling the
reservoir, the number of conditioning block data n will
be zero. The n + nVweights ka and tb are obtained by
solving a co-Kriging system:
n
X

a ; VaV
kaV cV

aV1
n
X
aV1

a ; mb
kaVcV

nV
X
b1
nV
X

a ; Vo
aV ; mb cV
tb cV
b ; Vo
tb cmbV ; mb cm

bV1

12

32

L.W. Lake, S. Srinivasan / Journal of Petroleum Science and Engineering 44 (2004) 2739

Fig. 4. Semivariogram model (left) and location of point support data in the simulation grid (right) used for demonstrating stochastic simulation
using the scale-up relations.

The simulation node is represented by Vo . The


corresponding Kriging variance is
r2K

X
a

o ; Va 
ka cV

o ; mb
tb cV

13

Assuming the local conditional distribution at Vo is


Gaussian with mean and variance given by Eqs. (11)
and (13), a simulated value is obtained by randomly
sampling from the Gaussian distribution with mean
from Eq. (11) and variance from Eq. (13). This
simulated value is added to the conditioning data set
for the next node. The simulation is completed by
visiting all the nodes in the simulation domain following a random path (Deutsch and Journel, 1998). The

number of conditioning block data, even if zero at the


start of the simulation process will increase as the
simulation proceeds.
The semivariogram model and the conditioning
data used in the example are in Fig. 4. The autocorrelation is represented by an anisotropic semivariogram that has its largest correlation length at 45j to
the x-axis.
Taking the semivariogram model (left in Fig. 4) and
the conditioning data to be representative of point scale
variability as shown on the right in Fig. 4, the objective
is to simulate the spatial variations in porosity on a
scaled up grid. The scaled up grid consists of 10  10
cells each of dimension 50  50 ft. Fig. 5 shows one
realization of the scaled up model obtained following

Fig. 5. The figure on the left is a realization of the simulated field using the volume scaling relations. A fine-scale realization (center) was also
linearly upscaled to give the realization shown on the right.

L.W. Lake, S. Srinivasan / Journal of Petroleum Science and Engineering 44 (2004) 2739

the procedure outlined above. For comparison, a suite


of fine-scale models, at a resolution of 50  50 grid
blocks, each of dimension 10  10 ft, were simulated
and subsequently upscaled to the 10  10 grid using a
linear (arithmetic) average of the property Z within
each coarse grid. Fig. 5 also shows one realization of
the scaled up model obtained in this fashion. The finescale simulation assumes that the conditioning data in
Fig. 4 is in fact at the volume support of the fine-scale
grid and hence no scale-up was performed.
Comparing the scaled up simulation results (left in
Fig. 5) with the upscaled model results (right in Fig. 5),
we see that the upscaled model retains the variability
observed in the fine-scale model. This is evident from
the strongly contrasting shades in the pixel map. The
semivariogram corresponding to the upscaled map
(Fig. 5, right) probably exhibits a smaller range of
autocorrelation than that corresponding to the scaled
up simulation map (Fig. 5, left). The autocorrelation
structure (i.e., the heterogeneity itself) is scaled up in
order to generate the scaled up realization. In contrast,
the upscaled realization (Fig. 5, right) is obtained by
simply averaging the fine-scale values on a block
support, i.e., the heterogeneity is not scaled up in the
up-scaling procedure.
To demonstrate the effect of using the correct
scale-up relations in the simulation, the average of
all blocks in the northeast quadrant of the pixel map
was computed for both the scaled up and upscaled

33

models. If the quantity Z being simulated is porosity,


this quadrant average serves as a surrogate for the
pore volume in a particular region of the reservoir.
Fig. 6 plots the variability in quadrant average observed over 50 realizations of the two models.
The uncertainty of the scale-up realizations is the
smallest. This is expected since direct simulation on
block support uses the scaled up semivariogram with
a longer autocorrelation range (corresponding to
linear averaging), which implies reduction in variance of simulated values. In contrast, any single
realization obtained by up scaling a fine-scale model
exhibits more variability, i.e., the semivariogram has
a shorter range that causes a wider uncertainty
distribution. However, ergodicity requires the spatial
average of the semivariogram (as computed in the
scale-up simulations) to be equivalent to the average
semivariogram characteristics computed on a large
number of upscaled realizations. This implies that
the two uncertainty distributions depicted in Fig. 5
would converge to the same mean given enough
realizations. However, depending on the range of the
starting point-support semivariogram, the number of
upscaled realizations required for convergence could
be large. This is an important result since it implies
that robust estimates of uncertainty using the fewest
representations of the reservoir model require implementing a proper scale-up procedure in stochastic
simulations.

Fig. 6. The variation in volume averages over a portion of the reservoir obtained using a suite of scale-up simulations is plotted on the left. The
variation computed over a suite of upscaled realizations is on the right.

34

L.W. Lake, S. Srinivasan / Journal of Petroleum Science and Engineering 44 (2004) 2739

Fig. 7. A typical acoustic log indicating porosity variations adjacent to a vertical well. The measurement interval is 0.5 ft.

The simulation approach outlined above is similar


to the missing scale simulation approach proposed
by Tran (1995). In that approach, the rank-order of the
point support data are used to simulate the spatial
variability at the block support. That simulation approach also uses scaled up semivariogram for the
spatial simulation of block support values.

4. A well log example


Another interesting application of the scale-up
procedure using the notion of the variance of the
mean is for analyzing the vertical variability typically
observed in well logs. Well logs record changes in
petrophysical properties in the vicinity of a well.
Response variables such as acoustic travel time exhibit different characteristics adjacent to producing
intervals than in non-producing intervals. Fig. 7
shows a typical travel time log recorded over a

1800-ft interval. The log is characterized by relatively


wide depth intervals of high or low acoustic travel
time and also isolated peaks and lows. In a typical
reservoir characterization exercise, logs from adjacent
wells are analyzed and a reservoir model is constructed by correlating the log traces at the wells.
Several methods for correlation analysis have been
proposed (Jennings, 1999). Frequently in these techniques, the window size is varied until the maximum
correlation between the traces at adjacent wells
occurs.
The fluctuations observed in the well logs represent point scale vertical variations in petrophysical
properties. These point scale variations must be scaled
up to a REV scale. At the REV scale, adjacent vertical
blocks have attribute values that are independent of
each other. The conjecture here is that autocorrelation
analysis between adjacent wells at this REV scale
would yield a much better idea of the compartmentalization of the reservoir.

Fig. 8. Vertical semivariogram and model fit computed using the well log data in Fig. 7. The semivariogram fit is shown on a log log scale on
the right.

L.W. Lake, S. Srinivasan / Journal of Petroleum Science and Engineering 44 (2004) 2739

35

Fig. 9. Variance of the block mean (left) and its derivative (right) plotted as a function of block size. The variation in slope of the variance plot is
plotted as a function of the block size on the right. The REV is defined as the onset of the portion with slope =  1.

The vertical semivariogram computed using the


acoustic log information and the model fit are shown
in Fig. 8. A two-structure exponential semivariogram
model with f1 = 0.6 and k1 = 6 ft, f2 = 0.4 and k2 = 20 ft
(Eq. 4) was selected.
Using these semivariogram model parameters, a
scale-up plot similar to Fig. 2 can be plotted. The
systematic change in the variance of the mean computed using vertical blocks of increasingly larger size
is shown in Fig. 9 (left). As discussed earlier, at small
scale (L), the variance of the block averages is large
2
and about equal to the population variance ro/D
. The
variance decreases as the vertical block size is increased and ultimately the data take on a slope of  1
indicating that blocks are at or above the REV scale.
We plot the slope of the variance of the mean plot on
the right of Fig. 9 to better illustrate the approach to
 1. Based on Fig. 9, the REV for the well in Fig. 7 is
about 80 ft. The procedure thus yields the optimum
size of the averaging window (80 ft) based on the

vertical variogram model. The averaging window size


will always be greater than the largest autocorrelation
length but how much more will depend on the point
scale semivariogram structure. If exponential, the
scale will not be much more than the small scale
support. If the point scale structure is linear (nonstationary), the scale will be larger.
Fig. 10 compares the scaled up log against the
original log response. As expected, the high frequency
variations are smoothed by the linear averaging. These
scaled up blocks now become the basis for the
subsequent log-correlation analysis between wells.
Well log data was available for seven other wells in
the reservoir. Semivariogram analysis revealed that
wells 2, 3 and 5 exhibit similar point-support semivariogram characteristics. Wells 6 and 7 were alike,
but different from wells 2, 3 and 5. The semivariogram model parameters are summarized in Table 1.
The volume variance relationships for these wells
are in Fig. 11. As expected, the REV (given by the

Fig. 10. Acoustic well log (continuous line) scaled up using a REV block size of 80 ft.

36

L.W. Lake, S. Srinivasan / Journal of Petroleum Science and Engineering 44 (2004) 2739

Table 1
A summary of the semivariogram parameters for all the wells

Wells 2,3 and 5


Well 4
Wells 6 and 7

f1

k1, ft

f2

k2, ft

0.6
0.6
0.8

6
15
20

0.4
0.4
0.2

30
55
40

In all cases, a two-scale autocorrelation model was used.

block size after which the slopes of the straight lines


in Fig. 11 attain a slope of  1) is larger for all other
wells since the semivariogram range in general is
larger for the other wells. The REVs range from 80
ft for well 1 to 110 ft for well 4.
The scaled up well logs can now be analyzed for
horizontal correlation. For illustration, the seven
wells were assumed to lie along a cross section
through the reservoir at irregular spacing. The horizontal variograms were computed assuming large lag
tolerances in order to have enough number of pairs
informing each lag. Fig. 12 shows the horizontal
semivariogram calculated on the basis of the scaled
up logs. The sub-REV variations are represented by
the nugget effect. The Gaussian shape (parabolic) of
the semivariogram at short lags indicates strong
lateral autocorrelation (continuity). The horizontal
range is approximately 1250 ft. The horizontal semivariogram computed using the original log data (at
0.5 ft measurement interval) has a much larger
nugget effect and a shorter range. The high nugget
effect is directly a result of the spacing of the wells
being large compared to the range of the fine-scale
semivariogram model. Reservoir modeling directly
using the well log information is therefore likely to
yield wider uncertainty distributions.

Fig. 11. Plot showing REV for all the wells calculated on the basis
of well log data and the semivariogram models in Table 1.

Fig. 12. Standardized horizontal semivariograms computed on the


basis of scaled up (REV) data. For comparison, the semivariogram
computed using the original point scale measurements is also shown.

It is common practice to aggregate well log information into property variations within layers. The
procedure described above is a viable scheme to
accomplish the task of defining layers in the reservoir
accounting for the pattern of variability exhibited by
the logs. Also note that the issue of how much small
scale variability should be retained in the scaled up
models is important. Perhaps one approach would be
to establish the scaling laws (linear averaging or nonlinear/ power averaging) rigorously through calibration before performing the scale-up operations described above.

5. Average permeability
The variance of a mean can also be used to say
something about the tendency of averages with scale
for properties that have non-Gaussian distributions
such as permeability. This involves investigating the
increase in variance of average permeability with

Fig. 13. Effect of scale on lateral permeability and its distribution.


Adapted from Kiraly (1975).

L.W. Lake, S. Srinivasan / Journal of Petroleum Science and Engineering 44 (2004) 2739

scale in the presence of inherent non-linearity in the


scale-up relationships.
Fig. 13 shows how lateral (horizontal or bed
parallel) permeability increases with scale. This observation is among the most common observations in
history matching performance data: well performance
based on core or log-derived permeability must be
increased to match field data. This behavior is at odds
with current upscaling practice wherein the mean
permeability either decreases or remains constant with
increasing scale. The effect is also one of the most
expected types of behavior because there are formations from which fluids are readily produced, but
corresponding cores from which will pass very little
fluid.
The same scale effect occurs with vertical (bed
normal) permeability except now the vertical permeability decreases with scale as illustrated in Fig. 14.
The vertical permeability decreases by two factors of
ten as the scale increases by 10. Figs. 13 and 14 are
not directly comparable. Fig. 14 shows the ratio of
vertical to horizontal permeability, wherein, an increase as in Fig. 13, is present in the denominator and
may partially account for the decrease in the ratio. The
x-axis in Fig. 14 is volume, not length as in Fig. 13.
The decrease, nevertheless, is undeniable and is consistent with the common need to reduce vertical
permeability from the core to cell scale to effect
history matches in numerical flow simulation.
Both the effects in Figs. 13 and 14 are usually
explained using geology, the vertical permeability
effect, in particular, by the presence of partially
permeable or impermeable shales that are discontinu-

37

ous. There is, however, a statistical explanation for


both effects as well.
Recall that Z in Eq. (1) is a Gaussian scalar random
variable. Hence, k = eZ is a log-normally distributed
random variable. Further, we assume the arithmetic
average (expectation) of k denoted by kH, is a surrogate for the horizontal permeability and the harmonic
average of k, kV is a surrogate for the vertical
permeability. Identifying the arithmetic average with
the horizontal permeability and the harmonic average
with the vertical permeability means, of course, that
the calculated properties apply to a uniformly layered
permeability medium. The scale L in the subsequent
paragraphs is perpendicular to these layers.
The non-centered moments of order j for a lognormal distribution are given as (Aitchison and
Brown, 1976):
kj V e

jl 12 j2 r2o=L

14

In the above equation, l = E(Z) = E(ln k) = ln kG where


kG is the geometric mean of the log-normally distrib2
uted k. ro/D
is related to Var(Z) from Eq. (2). Hence,
as the scale L increases, the variance of the mean
Var(Z) decreases (Eq. (5) and Fig. 2), and the variance
of a point within the scale increases, resulting in
changes in the non-centered moments of a log-normal
distribution.
Eq. (14) for j = 1 yields, the arithmetic average or
horizontal permeability:
kH kG e

r2 VarZ
2

15

The vertical permeability is equivalent to a harmonic


2
average and since Z Harmonic Z Geometric
=Z Arithmetic, Eq.
(14) yields the vertical permeability
kV kG e

r2 VarZ
2

16

The vertical to horizontal permeability ratio becomes:


kV
2

er VarZ
kH

Fig. 14. Variation of vertical to horizontal permeability ratio with


scale. Adapted from Corbett et al. (1996).

17

Fig. 15 plots Eq. (15) corresponding to the stable (one


scale) autocorrelation model of Eq. (7). The parameters, shown in the caption, have been adjusted to yield
approximately the same variations as in Fig. 13.
The figure also shows the F 1 standard deviation
of the estimate since this is available from the variance

38

L.W. Lake, S. Srinivasan / Journal of Petroleum Science and Engineering 44 (2004) 2739

Fig. 15. Calculated increase of horizontal permeability with scale


calculated with Eq. (15) and the one-scale stable autocorrelation
model. kG = 100, r2 = 20, a = 0.2, k = 10 000.

of the mean. The parameters used to generate Fig. 15


are reasonable. The range parameter k = 10 000 is
greater than the maximum dimension shown with
a = 0.2 being consistent with other literature (Jennings, 1999).
The decrease in vertical permeability with scale is
the same as that shown in Fig. 15 but decreasing
rather than increasing; comparing Eqs. (16) and (17).
The ratio of vertical to horizontal permeability is
similarly easy to calculate. Eqs. (15) and (17) provide
a statistically sound approach to scale-up point scale
permeability values before obtaining a history match
in flow simulation. It is quite likely that predictions of
future performance made on the basis of permeability
models that are correctly scaled up will yield robust
estimates of uncertainty.

6. Conclusions
The variance of the mean and how it depends
on averaging scale leads to refinement (if not
abandonment) of the representative elementary volume concept. It also can explain, at least qualitatively, the change in horizontal and vertical
permeability with scale using reasonable autocorre-

lation functions and reasonable parameters for the


autocorrelation functions. This is a consequence of
the various means depending on variability (for
non-Gaussian distributions) and the variability, in
turn, depending on scale. Since the averages now
depend on scale, it is possible that the ideas could
be used to generate scale-up factors if the parameters of the autocorrelation function can be inferred
from data. The predicted increase in horizontal
permeability is probably greater that what would
be observed in practice as would the predicted
decrease in vertical permeability, mainly because
the ideas were based on uniformly layered reservoirs,
which are idealizations. Nevertheless, the calculations could form limits for starting points of more
refinement.
It is also apparent that direct scaling up (taking the
averages over volumes) has the possibility of significantly altering the autocorrelation structure of a
simulated field. In the example shown here, the semivariogram of a well log that appears to be dominated
by the nugget effect actually exhibit strong autocorrelation after scale-up. The well-to-well correlation
became much clearer once the semivariogram was
computed from averaged properties, the averaging
volume being determined from the properties of the
local semivariogram.
Nomenclature
E( )
Expectation operator
fk
Fraction of total variance within scale k
K
Number of discrete scales in multiscale
autocorrelation model
k
Permeability
kG
Geometric mean permeability
kH, kV Horizontal and vertical permeabilities
L
Averaging scale
Var( ) Variance operator
Z*
Kriging estimate of Z
Za, Zb Cell mean and point values of conditioning
data at locations ua and uh
a
Roughness parameter in stable semivariogram model
c
Semivariance
c (V,V V) Semivariance between averages over volumes V and VV
v,v V
Point supports within volumes V and VV
kj V
Noncentered moment of order j

L.W. Lake, S. Srinivasan / Journal of Petroleum Science and Engineering 44 (2004) 2739

k a, t b
kk
el
/
q
r2
2
ro/D
2
ro/L
2
rL/D

r2k

Kriging weights
Autocorrelation length of scale k
Geometric mean of lognormal distribution
Porosity
Autocorrelation function
Population variance
Variance of a point property within a large
volume
Variance of a point property with a volume L
Variance of the average of a property over L
within D
Simple Kriging variance

References
Aitchison, J., Brown, J.A.C., 1976. The Lognormal Distribution.
Cambridge Univ. Press.
Bear, J., 1972. Dynamics of Fluids in Porous Media. Elsevier,
Amsterdam.
Corbett, P.W.M., Good, T.R., Jensen, J.L., Lewis, J.J.M., Pickup,
G.E., Ringrose, P.S., Sorbie, K.S., 1996. Reservoir description
in the 1990s: a perspective from the flow simulation through
layercake parasequence flow units. In: Glennie, K., Hurst, A.
(Eds.), AD 1995: NW Europes Hydrocarbon Industry. Geological Society, London, pp. 167 176.
Datta-Gupta, A., Vasco, D.W., Long, J.C.S., 1995. Detailed char-

39

acterization of a fractured limestone formation using stochastic inverse approaches. SPE Formation Evaluation 10 (3),
133 140.
Deutsch, C.V., Journel, A.G., 1998. GSLIBGeostatistical Software
Library and Users Guide. Oxford Univ. Press, New York City.
Frykman, P., Deutsch, C.V., 2002. Practical application of geostatistical scaling laws for data integration. Petrophysics 43 (3),
153 171.
Jennings Jr., J.W. 1999. How much core-sample variance should a
well-log reproduce. SPE Reservoir Evaluation & Engineering 2
(5), 442 450.
Journel, A.G., Hujbregts, C., 1978. Mining Geostatistics. Academic
Press, New York City.
Kiraly, L., 1975. Rapport sur le`tat actuel des connaissances dans le
domaine des characte`res physiques des roches karstiques. In:
Burger, A., Dubertret, L. (Eds.), Hydrogeology of Karstic Terrains. International Union of Geological Sciences. Series B, vol.
3, pp. 53 67.
Mahadevan, J., Lake, L.W., Johns, R.T., 2003. Estimation of true
dispersivity in field scale permeable media. Society of Petroleum Engineers Journal, Sept., 272 279.
Neuman, S.P., 1994. Generalized scaling of permeabilities: validation and effect of support scale. Geophysical Research Letters
21 (5), 349 352 (March 1).
Tran, T.T., 1995. The missing scale and direct simulation of block
effective properties. Journal of Hydrology 183, 37 56.
Varela, O.J., Torres-Verdin, C., Lake, L.W., Assessing the value of
3D seismic data in reducing uncertainty in reservoir production
forecasts, SPE 77359, presented at the 2002 SPE Annual Technical Conference and Exhibition, San Antonio, Texas, September 29 October 2.

You might also like