Improved Permeability Estimates in
Carbonate Reservoirs Using Electrofacies
Characterization: A Case Study of the
North Robertson Unit, West Texas
Trond Mathisen,* SPE, Sang Heon Lee,** SPE, and Akhil Datta-Gupta, SPE, Texas A&M U.
Summary
We propose a simple, cost-effective approach to obtaining perme-
ability estimates in heterogeneous carbonate reservoirs using com-
monly available well logs. Our approach follows a two-step pro-
cedure. First, we classify the well-log data into electrofacies types.
This classification does not require any artificial subdivision of the
data population but follows naturally based on the unique charac-
teristics of well-log measurements, reflecting minerals and litho-
facies within the logged interval. A combination of principal com-
ponent analysis (PCA), model-based cluster analysis (MCA), and
discriminant analysis is used to identify and characterize electro-
facies types. Second, we apply nonparametric regression techniques
to predict permeability using well logs within each electrofacies.
Our proposed method has been successfully applied to the
North Robertson Unit (NRU) in Gaines county, west Texas. Pre-
vious attempts to derive permeability correlations at the NRU have
included rock-type identification using thin-section and pore-
geometry analysis that can sometimes be expensive and time-
consuming. The proposed approach resulted in improved perme-
ability estimates, leading to an enhanced reservoir characteriza-
tion, and can potentially benefit both daily operations and reservoir
simulation efforts. The successful field application demonstrates
that the electrofacies classification used in conjunction with sound
geologic interpretation can significantly improve reservoir descrip-
tions in complex carbonate reservoirs.
Introduction
Permeability estimates are a critical aspect of a reservoir descrip-
tion. In sandstone reservoirs, a linear relationship normally exists
between porosity and the logarithm of permeability. Thus, perme-
ability predictions in sandstones can be achieved with acceptable
accuracy using porosity from well logs. In carbonates, however,
petrophysical variations rooted in diagenesis, grain-size variation,
and cementation can significantly alter the direct relationship be-
tween porosity and permeability.1 Statistical regression has been
proposed as a more versatile solution to the problem of perme-
ability estimation. Conventional statistical regression generally has
been done parametrically with multiple linear or nonlinear models.2–4
Several limitations inhibit multiple regression techniques,
many arising from the inexact nature of the relationship between
petrophysical variables. Conventional parametric regression re-
quires a priori assumptions regarding functional relationships be-
tween the independent and dependent variables. In complex car-
bonate reservoirs, such underlying physical relationships are not
known in advance, making traditional multiple regression tech-
niques inadequate and often leading to biased estimates.5,6
* Now with ExxonMobil Development Co.
** Now with ChevronTexaco.
Copyright © 2003 Society of Petroleum Engineers
This paper (SPE 84920) was revised for publication from paper SPE 70034, first presented
at the 2001 SPE Permian Basin Oil and Gas Recovery Conference, Midland, Texas, 15–16
May. Original manuscript received for review 2 July 2001. Revised manuscript received 20
February 2003. Paper peer approved 7 April 2003.
176
A variety of approaches have been proposed to partition well-
log responses into distinct classes to improve permeability predic-
tions. The simplest approach uses flow zones or reservoir layer-
ing.4,6 Other approaches have used lithofacies information identi-
fied from cores, as well as the concept of hydraulic flow units
(HFUs).7–11 However, in carbonate reservoirs, such classification
is complicated by the extreme petrophysical variations rooted in
diagenesis and complex pore geometry, even within a single zone
or class. A major difficulty in this regard has been discrimination
of classes from well logs in uncored wells.12,13
This paper discusses the application of a simple and cost-
effective approach to estimating permeability in a complex car-
bonate reservoir based on a two-step approach. The first step in-
volves classification of well-log data into electrofacies types. By
electrofacies, we imply a “rock type” or sediment exhibiting a
similar set of well-log responses.14 This classification does not
require any artificial subdivision of the data but is based on the
unique characteristics of well-log measurements, reflecting min-
erals and lithofacies within the interval. A combination of PCA,
MCA, and discriminant analysis is used to describe and identify
the electrofacies types. The electrofacies classification can be im-
proved further by incorporating geologic interpretation. Then, a
nonparametric regression technique, the Alternating Conditional
Expectation (ACE) algorithm, is used to develop permeability cor-
relations using well logs within each electrofacies. Such tech-
niques are completely data-driven and do not require a priori
assumptions regarding functional forms for correlating permeabil-
ity and well logs. In a previous paper,12 we have discussed the
application of this method to the Salt Creek Field Unit in the
Permian Basin, west Texas. In this paper, we focus on the North
Robertson Unit (NRU) in Gaines County, west Texas. The pro-
ducing horizons in the NRU are the Glorieta and Clearfork for-
mations, often referred to as the Upper and Lower Clearfork. The
hydrocarbon-bearing interval extends from the top of the Glorieta
to the base of the Lower Clearfork, between the correlative depths
of approximately 5,874 and 7,440 ft.
Previous attempts8,9 to derive permeability correlations at the
NRU have had mixed results. This includes rock-type identifica-
tion using thin sections and pore-geometry analysis, a procedure
that can be expensive and time-consuming. We present a simple,
cost-effective, and computationally efficient approach for perme-
ability estimation using only commonly available well logs. For
the examples studied here, our approach appears to result in better
permeability predictions, leading to an enhanced reservoir charac-
terization based on flow, permeability (rather than storage), or
porosity. This can have potential benefits both in daily operations
and in reservoir simulation efforts. Our successful field applica-
tions also demonstrate the power and versatility of electrofacies
characterization in improving reservoir descriptions in complex
carbonate reservoirs.
Methodology
In this section, we briefly discuss the underlying principles behind
electrofacies characterization and the development of permeability
correlations. For further details, the readers are referred to our
previous papers.5–7,12 The electrofacies classification is based on
June 2003 SPE Reservoir Evaluation & Engineering
attempts to identify clusters of well-log responses with similar
characteristics. This is a three-step procedure discussed next.
Principal Component Analysis (PCA). PCA is used to summa-
rize multiple well-log data effectively and to reduce the dimen-
sionality of the data without a significant loss of information.12–18
First, the well logs are standardized by subtracting the mean from
each reading and dividing by the standard deviation.12,13
Suppose we have a data set of log responses, Xn×p, in which the
element aij is the reading of the jth log taken at a certain depth i (i=
1,…, n, and j=1,…, p). The covariance of the matrix X can be
defined as
1 —T—
⌺= 共X − n−111TX兲T 共X − n−111TX兲 = 共XTX − nx x兲,
共n − 1兲
. . . . . . . . . . . . . . . . . . . . . . . . . . . (1)
where x៮ ⳱ 1TX/n is the row vector of means of the log variables.
Because ⌺ is a non-negative definite symmetric matrix, the co-
variance matrix can be factored into a combination of diagonal and
orthogonal matrices with the aid of spectral decomposition.
⌺ = QT⌳Q, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (2)
where ⌳ is a diagonal matrix of eigenvalues of ⌺ and Q is an
orthogonal matrix with column vectors that consist of eigenvectors
associated with the eigenvalues ␭1, . . . , ␭n, respectively. Because
a row vector of X can be represented as a point in p-dimensional
space, the so-called log space, X forms a cloud in the space.
Theoretically, the eigenvectors are the coordinates of the principal
component axes of the cloud; thus, Q is used in transforming the
original log data into principal components. Principal components
constitute an alternative form of displaying the data, thereby al-
lowing better knowledge of its structure without changing the in-
formation. The total variance in a data set can be defined as the
sum of the variances associated with each principal component.
Typically, the first two or three principal components explain most
of the variation in the original variables and reveal the structure in
the data. This can reduce the dimensionality of the problem and the
complexity in the cluster and discriminant analysis, as dis-
cussed next.12,15,17
Cluster Analysis. The aim of cluster analysis is to classify a data
set into groups that are internally homogeneous and externally
isolated on the basis of a measure of similarity or dissimilarity
between the groups. In this study, model-based clustering, a hier-
archical agglomerative clustering technique, is used.19,20 This ap-
proach can give much better performance than traditional proce-
dures such as single-link (nearest neighbor) and k-mean cluster-
ing, 21 which often fail to identify groups that are either
overlapping or of varying sizes and shapes. Another advantage of
the model-based approach is that there is an associated criterion for
assessing the model. This provides a means of selecting not only
the parameterization of the model but also the number of clusters,
without the subjective judgments necessary in other conventional
cluster analysis techniques.12,17–21
Banfield and Raftery19 developed a model-based framework
for clustering by parameterizing the covariance matrix in terms of
its eigenvalue decomposition. It was subsequently extended by
Fraley and Raftery20 to simultaneously determine the number of
clusters and cluster membership using an Expectation Maximiza-
tion (EM) algorithm. This model has had considerable success in
a number of practical applications, including character recognition,
tissue segmentation, minefield and seismic fault detection, identi-
fication of textile flaws from images, and classification of astro-
nomical data. We used MCA in our approach.
The identified clusters can be viewed as distinct electrofacies
groups that reflect the hydrologic, lithologic, and diagenetic char-
acteristics. If we have additional information such as core obser-
vations or geological insight, the identified electrofacies groups
could be calibrated to ensure their interpretable geological meaning.
Discriminant Analysis. Discriminant analysis is a multivariate
method for assigning an individual observation vector to two or
June 2003 SPE Reservoir Evaluation & Engineering
more predefined groups on the basis of measurements.12,21 The
discriminant analysis requires prior classification of the data into
relatively homogeneous subgroups with characteristics that can be
described by the statistical distributions of the grouping variables
associated with each subgroup. Typically, the classification is per-
formed by defining the distinct groups based on the unique char-
acteristics of well-log measurements or by applying known exter-
nal geologic criteria such as core-derived lithofacies information.
However, because in many situations a training data set with ab-
solutely known classifications is not easily obtained, a method like
MCA is commonly used. In our analysis, group-specific probabil-
ity density functions were determined by the distinct electrofacies
groups defined in the MCA using a training data set.
Permeability Correlation
Once the electrofacies are identified, the next step is to develop
correlations between permeability and well-log responses for each
electrofacies. We use nonparametric regression techniques that
do not require a priori assumptions regarding functional forms
to model the data. Specifically, we have used the ACE algo-
rithm originally proposed by Breiman and Friedman22 and subse-
quently used by various authors for data correlation in petrole-
um engineering.5–7,12,23,24
The ACE algorithm provides a method for estimating optimal
transformations for multiple regression that result in a maximum
correlation between a dependent (response) variable and multiple
independent (predictor) variables.22 Such optimal transformations
can be derived by minimizing the variance of a linear relationship
between the transformed response variable y (e.g., permeability)
and the sum of transformed predictor variables x1,…, xp (e.g., well
logs). Our approach proceeds as follows:
1. Develop optimal nonparametric transforms for permeability
and well-log variables based on cored wells: ␪(y), ␾1(x1),…,␾p(xp).
Such transformations are generated in a flexible data-defined man-
ner without any assumptions regarding the underlying functional
forms.
2. For an uncored well, given a set of well-log responses
{x 1i ,…,x pi }, first estimate the corresponding transforms
{␾1*(x1i),…,␾p*(xpi)} from Step 1.
3. Estimate the optimal transform for permeability using the
following relationship.22
p
␪* 共yi兲 = 兺␾
l=1
l
*
共xli兲. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (3)
4. Finally, predict permeability through a back transformation:
y ipre = ␪*−1 冋兺 册p
l=1
␾l* 共xli兲 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (4)
The complete procedure is carried out using a PC Windows-based
public-domain software GRACE (GRaphical Alternating Condi-
tional Expectations),25 that was developed at Texas A&M U. and
typically requires a few seconds of computation time.
The nonparametric regression techniques are very versatile and
can easily handle variables of mixed type. For example, we can
easily incorporate categorical variables such as rock types and
lithofacies into the correlation, and we can handle missing data
values without additional complications.7 This approach provides
us with a rapid and powerful alternative to traditional multiple
regression techniques for building correlations for a variety of appli-
cations, particularly in the presence of several predictor variables.
Field Application
We have applied the proposed technique to the NRU, a heteroge-
neous carbonate reservoir in west Texas. Conventional multiple
regression did not provide satisfactory permeability estimates in
this field. The NRU is located in the northern part of the Central
Basin Platform of the Permian Basin (Figs. 1 and 2). There are
two main producing horizons, the Glorieta and Clearfork forma-
tions (often referred to as the Upper and Lower Clearfork). The
productive, or hydrocarbon-bearing, interval extends from the top
177
Fig. 1—General location of the North Robertson Unit, west Texas.
of the Glorieta to the base of the Lower Clearfork, between the
correlative depths of approximately 5,874 and 7,440 ft.
Reservoir sediments accumulated in the Clearfork area of the
Central Basin Platform were deposited in shallow, agitated waters,
providing a favorable environment for carbonate generation. Early
winnowing in this high-energy environment provided for primary
intergranular porosity in some carbonates. However, fluctuations
in relative sea level resulted in cyclic deposition typical of many
Permian stratigraphic units in west Texas. Also, leaching and dis-
solution of skeletal grains, as well as selective dolomitization,
modified porosity distribution significantly. In addition, local frac-
tures and other late diagenetic processes affected the reservoir
quality of the deposits, and the end result is a very heterogeneous
and complex reservoir.8,9
The NRU has been producing since the early 1950s on a 40-
acre spacing with a total of more than 140 producing wells, which
represent the primary well development. Secondary recovery pro-
jects, such as waterflooding and infill drilling to a 20-acre well
spacing, was implemented starting in early 1987. At the time of the
unitization, the original oil in place was estimated to be approxi-
mately 260 million STB, and estimated ultimate recovery was only
13.5%, including 6% from the secondary recovery projects. Sub-
sequent developments increased the field unit oil production from
670 STB/D to approximately 3,200 STB/D, with a gas/oil ratio
(GOR) decrease from 1,550 scf/STB to 1,400 scf/STB and water-
production rates from 500 BW/D to more than 11,000 BW/D. The
total field unit water-injection rates were roughly 20,000 BWI/D.
Data Acquisition and Processing. The NRU is like many other
carbonate reservoirs; it exhibits extreme heterogeneity, making
permeability predictions very difficult. Conventional statistical
analysis of the permeability/porosity/facies relationship from the
cored-well data did not reveal any reliable correlations that could
be used to predict permeability in uncored wells. As a result,
surveillance and reservoir management efforts had relied on a
reservoir characterization built from a storage model based on
porosity, rather than a flow model based on permeability. Scattered
high-permeability streaks have the potential to induce prematurely
high GOR, production declines, and reduced waterflood use, all of
which may lead to poor economic performance. As the waterflood
or secondary recovery projects mature, the need for detailed and
accurate reservoir descriptions has increased.
Permeability predictions at the NRU are complicated by the
fact that different depositional environments have similar ranges of
permeability and porosity. The data presented in this analysis have
been gathered from six cored wells; of these, two wells (1509 and
3533) were left out to verify our correlations using blind tests.
Although intuitive relationships existed between individual well
logs and permeability, quantitative relationships were complex and
varied from zone to zone as well as from region to region. Before
performing any type of regression analysis on well-log responses
or core data, all data must be carefully depth shifted to make sure
that all values are properly adjusted to the same depth. The depth
178
Fig. 2—Location of cored wells in the North Robertson Field
Unit, Gaines County, Texas.
correlation between logs and core data is performed by comparing
density-derived porosity values (bulk density, RHOB) with core
porosity data and adjusting the reported core-data depth to match
the well-log depth.
Finally, with the consideration of the data quality and fieldwide
availability, a suite of well logs is selected for the analysis. In this
field, we have various well logs in different wells: caliper, spon-
taneous potential (SP), gamma ray (GR), two different resistivity
logs (lateral log deep, or LLD, and lateral log shallow, or LLS),
acoustic transit-time or sonic (DT), neutron (neutron log-derived
porosity, NPHI), RHOB, and photoelectric (PEF) logs. Among the
nine well logs, only six logs (GR, LLS, LLD, DT, NPHI, and
RHOB) were chosen because of their fieldwide availability.
We made no attempts to group the data according to preset
stratigraphical zones; in general, we tried to honor the original data
set as much as possible and rejected only those points that ap-
peared to be obvious outliers. In particular, the NPHI and DT well
logs are indirect measures of porosity and thus are indicative of
permeability, even though they are poorly correlated. Similarly,
GR logs can identify shale content (particularly in sandstones) and
thus a measure of permeability, especially in low-permeability
zones. The LLS and LLD logs measure the resistivity in the im-
mediate vicinity of the well and deeper into the formation, respec-
tively. In the presence of high-permeability zones, fluid invasion
can be indicated by the ratio of LLS and LLD.
Electrofacies Characterization and Identification. PCA is used
to summarize the data effectively and to reduce the dimensionality
without any significant loss of information. This analysis is applied
to obtain the principal components PCj (j⳱1,…,6) from the well-
log data after normalization. Fig. 3 shows the scree plot, a bar plot
of the variance of the principal components labeled by
j
兺␭ Ⲑ trace共⌺兲, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (5)
i=1
i
which often provides a convenient visual method of identifying the
important principal components. Only three principal components
explain approximately 90% variation of the whole data set. We can
also explore the relationship between reservoir properties and the
three major principal components using the scatterplot in Fig. 4.
The first principal component (PC1) appears to indicate porosity of
the formation, while the second principal component (PC2) shows
a stronger correlation with gamma ray readings. The eigenvectors
of the covariance matrix ⌺ provide coefficients of the principal
components’ transformation (Table 1). For example, PC1 and PC2
are given by the following relationships:
June 2003 SPE Reservoir Evaluation & Engineering
Fig. 3—A scree plot, or a bar plot representing the variances
described by the PCs.
PC1 = −0.16GR + 0.46log共LLD兲 + 0.46log共LLS兲
− 0.46DT + 0.41RHOB − 0.42NPHI
PC2 = 0.93GR + 0.13log共LLD兲 + 0.20log共LLS兲
− 0.21DT + 0.06RHOB − 0.16NPHI. . . . . . . . . . . . . . . . (6)
A scatterplot of PC1 vs. PC2 is shown in Fig. 5.
Next, we apply MCA to define distinct groups based on the
unique characteristics of the well-log measurements. Each cluster
can be treated as an electrofacies that reflects the hydrologic, litho-
logic, and diagenetic characteristics. The clustering algorithm in-
dicates nine distinct groups based on the well-log measurements
(Fig. 6). A qualitative interpretation of the clusters is aided by the
relationship indicated in Fig. 4. For example, we can say that the
first electrofacies group (EF1) represents porous media with high
gamma ray reading, and the ninth electrofacies group (EF9) rep-
resents tight media with low gamma ray reading.
Data Correlation, Model Validation, and Comparison. Once
we partition the well-log responses into electrofacies groups, the
next step is to apply statistical regression techniques to model the
correlation between permeability and well-log responses within
each electrofacies group. In this study, the nonparametric regres-
sion technique ACE is used to develop such correlations.
To validate our approach, we predicted the permeability in
wells 1509 and 3533, the two cored wells that were left out for
blind tests. First, we defined the electrofacies from the well-log
responses in 1509 and 3533 using a discriminant analysis. On the
basis of the nine clusters defined in the discriminant analysis, an
allocation function was determined. Through the allocation func-
tion, we can define the group membership of the log responses in
new or uncored wells. Fig. 7 shows the electrofacies profile as
obtained from discriminant analysis in well 3533, while the cor-
June 2003 SPE Reservoir Evaluation & Engineering
Fig. 4—A scatterplot of different well logs and PCs.
responding plot for well 1509 is presented in Fig. 8. The “true
electrofacies” in these plots corresponded to the classification
when all the cored wells were included in the electrofacies char-
acterization. Once the electrofacies are identified, permeability es-
timates are obtained using the appropriate correlation for these
electrofacies. The results are shown in Figs. 9 and 10. Overall, the
predicted permeabilities based on well-log data in both wells seem
to agree reasonably well with the core data. The predicted vs. the
measured permeabilities for the two wells are shown in Figs. 11
and 12, and they both indicate a fairly consistent trend.
We compared our results with a previous study at the NRU that
involved a detailed thin section and pore-geometry analysis to
obtain the different rock types.8,9 A total of eight different rock
types were identified, of which four were selected to have reservoir
potential, and linear regression equations relating porosity and the
logarithm of permeability were developed for each of the indepen-
dent rock types. The rock-log model was developed based on the
petrophysical and lithological studies of several thousand feet of
whole core from eight wells in the NRU. This study incorporated
detailed sedimentologic observations of the whole core, thin-
section analysis, x-ray diffraction, special core analysis, and
pore-geometry investigations using scanning electron microscope
image analysis.8,9
The independent rock types were identified using various litho-
logical characteristics together with volume ratios of pore types.26
Briefly, the discrimination of rock types 1 through 4 in comparison
to rock types 5 through 9 was done by plotting sonic travel time vs.
neutron porosity at the well. The discrimination for independent
rock types 1 through 4 was determined by plotting the resistivity of
the flushed zone (Rxo) vs. the resistivity of the uninvaded zone (Rt)
and applying various cutoffs. After finalizing the identification of
rock types 1 through 4, several different procedures were imple-
mented to separate each of the independent rock types 5, 6, 7, and
8. These procedures are based on the difference in apparent matrix
density, apparent photoelectric capture cross section, gamma ray
response, and lithology. The permeability predictions at well 1509
based on this rock-type classification are shown in Fig. 13. For
comparison purposes, we have also superimposed the predictions
based on the electrofacies classification. The proposed approach
appears to significantly outperform the rock-type classification for
this case.
It is worth mentioning here that the primary difficulty with the
rock-type approach was that of assigning appropriate rock types at
the uncored wells based on the well-log response. Our previous
experience has also shown that the discrimination error associated
with such models can lead to poor predictions, although the cor-
relations within the rock types may be quite satisfactory. The
success of the electrofacies approach can be attributed mainly to its
improved pattern recognition and better discriminatory power.12
Computer programs (PC-Windows) for electrofacies characteriza-
179
 Fig. 5—A scatterplot of PC1 and PC2.

tion (EFACIES) and data correlation using the ACE model
(GRACE) can be obtained at the web address found in Ref. 25.
Conclusions
The principal underlying idea in this paper is that by preclassifying
the well-log responses into several distinct clusters or electrofacies
and finding the optimum permeability correlation model for each
cluster, improved permeability estimates may be obtained, particu-
larly in heterogeneous carbonate reservoirs. The following conclu-
sions can be drawn based on the results presented here.
1. We demonstrate a field application of the electrofacies charac-
terization and nonparametric regression for permeability esti-
mates using well logs in a complex carbonate reservoir, the
North Robertson Unit, west Texas. The proposed approach is
cost effective and computationally efficient and also appears to
result in improved permeability estimates compared to a previ-
ous study based on rock-type identification.
2. The success of our proposed approach can be attributed to the
robust discriminatory power of the electrofacies characterization
and the reliability of the nonparametric regression model. The
primary difficulty with the rock-type approach appears to be in
assigning appropriate rock types at the uncored wells based on
well-log responses. This can impact the permeability estimates
even though the permeability correlations within each rock type
may be quite satisfactory.
3. Nonparametric regression techniques (e.g., the ACE algorithm)
offer a powerful, versatile, and fully automated tool for building
permeability correlations in heterogeneous carbonate reservoirs
without subjective assumptions regarding the functional rela-
tionship between permeability and well logs.

Fig. 6—Electrofacies characterization through cluster analysis.

180 June 2003 SPE Reservoir Evaluation & Engineering

 Fig. 7—Predicted vs. true electrofacies in Well 3533. Fig. 8—Predicted vs. true electrofacies in Well 1509.

Fig. 9—Measured vs. predicted permeability based on the elec- Fig. 10—Measured vs. predicted permeability based on the elec-
trofacies classification and the ACE algorithm (Well 3533). trofacies classification and ACE (Well 1509).

June 2003 SPE Reservoir Evaluation & Engineering 181

Fig. 11—Measured vs. predicted permeability based on the electrofacies classification and the ACE algorithm (Well 3533—
reservoir interval 7,050 to 7,300 ft).

Nomenclature ⌺ ⳱ covariance matrix

k ⳱ permeability, md ␪ ⳱ parameters of the multivariate distribution model
Q ⳱ orthogonal matrix ␪(•) ⳱ transformation of dependent variable
Rt ⳱ resistivity of the uninvaded zone ␾(•) ⳱ eigenvalue
Rxo ⳱ resistivity of the flushed zone ␭ ⳱ diagonal matrix of eigenvalues of ⌺
trace ⳱ summation of diagonal elements in the matrix ⌳ ⳱ mean vector
T ⳱ transpose of a vector or matrix ␾l(•) ⳱ transformation of independent variable l
xi ⳱ ith data point for the jth well log 1 ⳱ column vector containing all elements with 1
x1,…,xp ⳱ independent or predictor random variables
x៮ ⳱ row vector of means of the log variables
X ⳱ a data set of log responses Subscripts and Superscripts
y ⳱ dependent or response random variables i, j ⳱ indices for data observation
␧ ⳱ Gaussian random noise l ⳱ index for independent or predictor variables

Fig. 12—Measured vs. predicted permeability based on the electrofacies classification and the ACE algorithm (Well 1509—
reservoir interval 6,150 to 6,575 ft).

182 June 2003 SPE Reservoir Evaluation & Engineering

 Fig. 13—Comparison of measured vs. predicted permeability in Well 1509 using two different approaches.
References
1. Chilingarian, G.V., Mazzullo, S.J., and Rieke, H.H.: Carbonate Res-
ervoir Characterization: a geologic-engineering analysis, Elsevier,
New York City (1992).
2. Allen, J.R.: “Prediction of Permeability from Logs by Multiple Regres-
sion,” 6th Annual European Logging Symposium Transactions: Soci-
ety of Professional Well Log Analysts (1979).
3. Jensen, J.L. and Lake, L.W.: “Optimization of Regression-Based Po-
rosity-Permeability Predictions,” paper presented at the 1985 CWLS
Symposium, Calgary, 29 September–2 October.
4. Wendt, W.A., Sakurai, S., and Nelson, P.H.: “Permeability Prediction
From Well Logs Using Multiple Regression,” Reservoir Characteriza-
tion, L.W. Lake and H.B. Carroll Jr. (eds.), Academic Press Inc., Or-
lando, Florida (1986) 181.
5. Xue, G. et al.: “Optimal Transformations for Multiple Regression:
Application to Permeability Estimation From Well Logs,” SPEFE
(June 1997) 85.
6. Barman, I. et al.: “Permeability Predictions in Carbonate Reservoirs
Using Optimal Nonparametric Transformations: An Application at the
Salt Creek Field, Kent County, Texas,” paper SPE 39667 presented at
the 1998 SPE/DOE Symposium on Improved Oil Recovery, Tulsa,
19–22 April.
7. Datta-Gupta, A., Xue, G., and Lee, S.H.: “Nonparametric Transforma-
tions for Data Correlation and Integration: From Theory to Practice,”
Reservoir Characterization—Recent Advances, R. Schatzinger and J.
Jordan (eds.), AAPG Memoir (1999) 71, 381–396.
8. Davies, D.K. and Vessell, R.K.: “Flow Unit Characterization of a Shal-
low Shelf Carbonate Reservoir: North Robertson Unit, West Texas,”
paper SPE 35433 presented at the 1996 SPE/DOE Symposium on
Improved Oil Recovery, Tulsa, 21–24 April.
9. Doublet, L.E. et al.: “An Integrated Geologic and Engineering Reser-
voir Characterization of the North Robertson (Clearfork) Unit: A Case
Study, Part 1,” paper SPE 29594 presented at the 1995 SPE Joint
Rocky Mountain Regional/Low-Permeability Reservoirs Symposium,
Denver, 20–22 March.
10. Amaefule, J.O. et al.: “Enhanced Reservoir Description: Using Core
and Log Data to Identify Hydraulic (Flow) Units and Predict Perme-
ability in Uncored Intervals/Wells,” paper SPE 26436 presented at the
1993 SPE Annual Technical Conference and Exhibition, Houston, 3–6
October.
June 2003 SPE Reservoir Evaluation & Engineering
11. Moline, R.G. and Bahr, J.M.: “Estimating Spatial Distributions of Het-
erogeneous Subsurface Characteristics by Regionalized Classification
of Electro-facies,” Mathematical Geology (1995) 27, No. 1, 3.
12. Lee, S.H. and Datta-Gupta, A.: “Electrofacies Characterization and
Permeability Predictions in Carbonate Reservoirs: Role of Multivariate
Analysis and Nonparametric Regression,” paper SPE 56658 presented
at the 1999 SPE Annual Technical Conference and Exhibition, Hous-
ton, 3–6 October.
13. Bucheb, J.A. and Evans, H.B.: “Some Applications of Method Used in
Electrofacies Identification,” The Log Analyst (January–February
1994) 35, 14.
14. Serra, O. and Abbott, H.T.: “The Contribution of Logging Data to
Sedimentology and Stratigraphy,” SPEJ (February 1982) 117.
15. Wolff, M. and Pelissier-Combescure, J.: “FACIOLOG—Automatic
Electrofacies Determination,” paper presented at the 1982 SPWLA
Annual Logging Symposium, Corpus Christi, Texas, 6–9 July.
16. Doveton, J.H. and Prensky, S.E.: “Geological Applications of Wireline
Logs—A Synopsis of Developments and Trends,” The Log Analyst
(May–June 1992) 33, 286.
17. Lim, Jong-Se, Kang, J.M., and Kim, J.: “Multivariate Statistical Analy-
sis for Automatic Electrofacies Determination From Well Log Mea-
surements,” paper SPE 38028 presented at the 1997 SPE Asia Pacific
Oil and Gas Conference, Kuala Lumpur, 14–16 April.
18. S-Plus 2000 Guide to Statistics, Volume 2, Data Analysis Products
Division, MathSoft, Seattle, Washington (2000).
19. Banfield, J.D. and Raftery, A.E.: “Model-based Gaussian and Non-
Gaussian Clustering.” Biometrics (1993) 49, No. 3, 803.
20. Fraley, C. and Raftery, A. E.: “Mclust: Software for Model-Based
Cluster and Discriminant Analysis,” Technical Report No. 342, Dept.
of Statistics, U. of Washington, Seattle, Washington (May 1998).
21. Ripley, B.D.: Modern Applied Statistics with S-Plus, Springer-Verlag,
New York City (1994) 301.
22. Breiman, L. and Friedman, J.H.: “Estimating Optimal Transformations
for Multiple Regression and Correlation,” J. of the American Statistical
Assn. (September 1985) 80, No. 391, 580.
23. McCain, W.D. Jr. et al.: “Correlation of Bubblepoint Pressures for
Reservoir Oils—A Comparative Study,” paper SPE 51086 presented at
the 1998 SPE Eastern Regional Meeting, Pittsburgh, Pennsylvania,
9–11 November.
24. Nomura, M. and Sato, K.: “Continuity Assessment Through Flow-Test
and Production Data in a Volcanic Formation,” SPE 56679 paper pre-
183
 sented at the 1999 SPE Annual Technical Conference and Exhibition, assistant at Texas A&M U. developing improved techniques for
Houston, 3–6 October. reservoir characterization, focusing primarily on field demon-
25. GRACE Programs and User Manual, Texas A&M U., College Station, stration. He holds BS and MS degrees in petroleum engineering
Texas (1997); available at http://people.tamu. edu/∼adgupta. from Texas A&M U. Sang Heon Lee is currently a reservoir en-
26. Mathisen, Trond: “Reservoir Characterization Using Nonparametric gineer with ChevronTexaco in Houston. Previously, he was a
Regression Techniques,” MS thesis, Texas A&M U., College Station, postdoctoral research associate in the Dept. of Petroleum En-
gineering at Texas A&M U. His research interests include reser-
Texas (2000).
voir engineering, reservoir simulation, and characterization. He
holds BS and MS degrees in mineral and petroleum engineer-
ing from Seoul Natl. U. in Korea and a PhD degree in petroleum
SI Metric Conversion Factors engineering from Texas A&M U. Akhil Datta-Gupta is an asso-
ciate professor and holder of the Rob. L. Adams Professorship
acre × 4.046 873 E + 03 ⳱ m2
in Petroleum Engineering at Texas A&M U. Previously, he
bbl × 1.589 873 E – 01 ⳱ m3 worked with BP Exploration/Research and the Lawrence
ft × 3.048* E – 01 ⳱ m Berkeley Natl. Laboratory. His research interests include reser-
ft3 × 2.831 685 E – 02 ⳱ m3 voir characterization, simulation, and environmental remedia-
tion. Datta-Gupta holds a PhD degree in petroleum engineer-
*Conversion factor is exact. ing from the U. of Texas at Austin. He is a member of the SPE
Editorial Review Committee and the Engineering Professional-
ism Committee. He is a recipient of the AIME Rossitter W. Ray-
Trond Mathisen currently works as a drilling engineer with mond Award (1992) and the SPE Cedric K. Ferguson Certificate
ExxonMobil Corp. in Houston, primarily with HP/HT applications (2000) and has been an SPE Distinguished Lecturer (1999–2000)
in south Texas. He previously worked as a graduate research and a Distinguished Author (2000).

184 June 2003 SPE Reservoir Evaluation & Engineering