SPE 95167

Permeability Estimation Using a Hybrid Genetic Programming and Fuzzy/Neural

Inference Approach
D. Xie, D. Wilkinson, and T. Yu, Chevron ETC
Copyright 2005, Society of Petroleum Engineers
This paper was prepared for presentation at the 2005 SPE Annual Technical Conference and
Exhibition held in Dallas, Texas, U.S.A., 9 12 October 2005.
This paper was selected for presentation by an SPE Program Committee following review of
information contained in a proposal submitted by the author(s). Contents of the paper, as
presented, have not been reviewed by the Society of Petroleum Engineers and are subject to
correction by the author(s). The material, as presented, does not necessarily reflect any
position of the Society of Petroleum Engineers, its officers, or members. Papers presented at
SPE meetings are subject to publication review by Editorial Committees of the Society of
Petroleum Engineers. Electronic reproduction, distribution, or storage of any part of this paper
for commercial purposes without the written consent of the Society of Petroleum Engineers is
prohibited. Permission to reproduce in print is restricted to a proposal of not more than 300
words; illustrations may not be copied. The proposal must contain conspicuous
acknowledgment of where and by whom the paper was presented. Write Librarian, SPE, P.O.
Box 833836, Richardson, TX 75083-3836, U.S.A., fax 01-972-952-9435.

Abstract
We have developed a methodology that provides
permeability estimates for all rock-types or lithologies, for
a wide range of permeability. This is a hybrid Genetic
Programming and Fuzzy/Neural Net inference system
and which utilizes lithologic and permeability facies as
indicators. This work was motivated by a need to have a
volumetric estimate of permeability for reservoir
modeling purposes. To this end, for our purposes,the
inputs to this process are limited to properties that can
be estimated from seismic data. The permeability
transform is first estimated at the well locations using
core permeability, elastic parameter logs and porosity.
The output from the process can then be used, in
conjunction with estimates of these properties from 3D
seismic data, to provide an estimate of permeability on a
volume basis. The inputs are then, the volume of shale
(Vsh) or any other log type used to determine lithology,
the sonic and density logs, the porosity log and core
permeability measurments. The transform system is
composed of three distinct modules. The first module
serves to classify lithology and separates the reservoir
interval into user-defined lithology types. The second
module, based on Genetic Programming, is designed to
predict permeability facies within lithology type. A
permeability facies is defined as as a low, medium or
high permeability set associated with each lithology type.
A Fuzzy/Neural Net inference algorithm makes up the
third module of the system, in which a TSK fuzzy logic
relationship is formed, for each permeability facies and
lithology.
The system has been applied in two oil fields, both
offshore West Africa. In comparison with current
estimation approaches, this system yields more
consistent estimated permeability. The results from

conducting cross-validation suggest this methodology is

robust in estimating permeability in complex
heterogeneous reservoirs. This system is designed to
use elastic log properties inverted from seismic data,
such as acoustic velocity and density as input so
permeability volume can be obtained.

Introduction
Knowledge of the spatial variability of permeability is of
great importance in reservoir modeling. However, core
permeability measurments are very limited and tend to
be biased. While it is commonly observed that
permeability has some correlation with wire-line data,
there is no theoretical equation to describe such a
relationship. In general, developing a generic system
that produces good permeability estimates for all types
of lithologies is a difficult task. Permeability estimation
continues to pose a significant challenge to reservoir
characterization and simulation.
In the past few years, tremendous effort has been
expended in the generation of a large variety of
approaches to estimating permeability utilizing wire-line
logs: multi linear regression, principal components
analysis and clustering, non-linear Neural Network, and
fuzzy logic among them. All those methods involve using
some form of indicator, such as lithology type, electrofacies (Mathisen, et al, 2001, Lee, et al, 2002), grainsize, litho-facies (Suryanarayana, et al, 2003), or
hydraulic flow unit (Badarinadh, et al, 2002, Fahad, et al,
2000) to improve the permeability estimation. Introducing
such indicators allows one to establish different
relationships which can then be applied to the
associated intervals or units of interest.
In complex heterogeneous reservoirs, however, these
methods happen to miss-estimate the high permeability
layers or produce low resolution permeability logs. In
addition, some of those indicators, such as grain-size
indicator, are not easy to obtain and are certainly not
available on a reservoir wide scale. The permeability of
reservoir sequences in these types of depositional
environments exhibits significant changes even within
the same lithologic indicator or electro-facies
indicator. It is not unusual that the permeability related
to reservoir lithologys range from 10 md to 1000 md. To
attempt to account for this wide variation we decided to
use permeability facies (high, low, medium) as an

SPE 95167

indicator. Moreover, the permeability facies could have

different meaning for different lithology types. For
instance, the high permeability facies of sandstone might
indicate a permeability range from 500 to 1000 md while
for shaly sand the range may be from 50 to 100 md. This
indicator scheme or partitioning is a flag process, that
is, estimated values are categorical, instead of
continuous. Categorical values are very difficult to assign
using wire-line logs with linear or non-linear regression
type techniques. Our studies show that Genetic
Programming (GP) provides more stable flag or
categorical estimation in comparison to multi-linear
regression, Neural Network, and decision tree methods
(Quinlan, 1993).
Now that we have partitioned the input dataset into
sets characterized by permeability facies and lithology
type, we wish to estimate permeability for each of these
sets independently. In this study, we utilize a combined
system of Fuzzy Inference and Neural Network to do the
permeability estimation. One advantage of using Fuzzy
Inference is that we can account for the imprecision and
uncertainty in the input values.
We will briefly discuss the Genetic Programming and
Fuzzy Inference algorithms, and then introduce our
hybrid permeability estimation process with the
applications into two oil fields, both offshore West Africa.
Genetic GP and Fuzzy Inference
GP is a symbolic regression method (Koza, 1992). GP
training cycle (Figuer 1) starts with a population of
randomly generated classifiers. Depending on their
fitness, calculated by the percentage of the training data
that are correctly classified, two classifers with higher
fitness in the tournament selection become the winner
for reproduction. If two candidates are tied in their hit
rate, the mean squared error measurement is used to
select the winner. Reproduction is done by a
combination of crossover and mutation. The two
generated offsprings replace the worst classifiers in the
population. This select-reproduce-replace cycle continue
for many generations until the termination criterion is met
(Yu, et al., 2003).
Adaptive Neural-Fuzzy Inference System (ANFIS) is
a fuzzy modeling tool that generates Takagi-SegenoKang (TSK) type fuzzy inference system which is based
on the given input and output data. A TSK fuzzy system
has the following structure (Jang, 1993).
The first component is a set of input membership
functions (MFs). Membership functions map crisp data to
linguistic values or labels. Fuzzy Inference system uses
conditional statement in if-then format. The if-part
consists of linguistic values and fuzzy operators (AND,
OR, and NOT). The then-part is a first order linear
equation. Based on the conditions or fuzzy rules, the
inference system interprets input values and assigns
output values. In TSK system, the output values are
calculated based the weight of each rule, as shown in
Figure 2. The construction of ANFIS is automatic
through a two step learning process: (1) the network
structure is built by a subtractive clustering algorithm;

and (2) the input membership function parameters are

tuned by a back propagation algorithm while the output
parameters are tuned by a least square estimation.
Hybrid GP and Fuzzy/Neural Permeability Estimation
System
Our hybrid system consists of three modules. The first
module serves to classify lithology and separates
reservoir interval into user-defined lithology types. The
second module (GP based) is designed to determine
permeability facies within each lithology type.
Permeability facies are defined as low, medium or high
permeability zones associated with lithologies. A
Fuzzy/Neural Net inference algorithm makes up the third
module of the system in which a TSK fuzzy logic
relationship is derived for each permeability range.
Given the input data, the system analyzes the lithology
types, classifies the permeability facies, and then
estimates the permeability values (Figure 3).
For the permeability facies estimation we decided to
have high, medium, and low permeability facies related
to each lithology type. In some cases, however, two
facies are enough. The number of facies may also be
dependent on the number and distribution of core
permeability samples. The GP module captures the
errors and restarts the process if the GP
misclassification reaches unacceptable levels. This
second module is of vital importance of the overall
system performance as generates the permeability
facies (categorical values) which will be pipelined to the
Fuzzy/Neural permeability estimation process.
The Fuzzy/Neural module estimates the actual
permeability values. The process is very similar to nonlinear regression methods. The main difference is that its
internal engine utilizes fuzzy rules. Each input well log
has its own rules in the form of membership functions
and weight factors trained with the input logs and core
permeability. The estimated permeability is the result of
inferring or interpreting those fuzzy rules with core
permeability.

Applications
The hybrid GP and Fuzzy/Neural system has been
applied in two oil fields, both offshore West Africa, to
estimate the permeability of reservoir sequences. In
those two deep water fields, the reservoirs are highly
heterogeneous and the permeability is highly variable
not only among different lithology types and but also
within the same lithology type. As a result, the
relationships of core permeability to the measured logs
are extremely difficult to derive.
In the first field, five wells are available with core
permeability. Each of these wells is at least half a mile
from its neighbors. Wire-line logs, including sonic and
density, have been recorded in each well. Figure 4
shows the plot of the core permeability against the
volume of shale (Vsh) indicating the changes in
permeability are significant both among and within the
lithology types. It is clear that the variations could lead

to considerable uncertainty by using simple regresion to

estimate the permeability. The critical decision that
needs to be made is how to partition Vsh into different
lithology types so that each one has similar geological
properties or easily recognizable patterns of permeability
distribution. One way to define lithology type is based on
the cutoffs of Vsh. It is possible to define these partitions
with either a sharp or fuzzy boundary. In this study, we
simply apply sharp cutoffs to determine each lithology
type. Based on the Vsh distribution and formation
evaluation (FE) analysis, five lithology types are defined:
sand (Vsh < 15%), shaly sand (Vsh 15-40%), high
impedance sand (Vsh < 15% and P-wave velocity >
7200 ft/s), sandy shale (Vsh 40-75%), and shale (Vsh >
75%). The high impedance sand, which has a P-wave
velocity greater than 7200 ft/second, only has high
permeability.
The input data, beside core permeability, are total
porosity, P-wave velocity and bulk density. During the
training process, we specify permeability facies as high,
medium, and low. Note that the shale has low and
medium permeability facies and high impedance sand
requires only one facies, i. e., high permeability facies.
The cutoffs for the three facies depend on the lithology
type and are different as shown in Table 1. As
expected, the permeability facies for the sand and shale
are easy to model for all the five wells. In contrast, the
sandy shale has the most complicated relationship to the
core permeability. As a result, its classifiers have the
lowest accuracy rate. The detailed accuracy of the GP is
given in Table 2.
The system generates permeability estimations that
are well correlated with the core permeability for all
lithology types. The correlation coefficient values (R2)
are between 0.9 and 0.95. We combined the GP
classifiers with the TSK system and then ran the overall
system on the input data. Figure 5 shows the results.
Among the five lithology types, the sandy shale has the
worst match between the estimated and the core
permeability. This is due to the misclassification of
permeability facies by the GP classifiers.
Our comparison study shows that the hybrid GPfuzzy system outperforms the current system on all the
data from the five wells (Figure 6). We also made
detailed performance comparisons with the previous
system using data from one well (well C). Figure 7
shows that the hybrid system (* Perm-FIS) produces
permeability estimations that match the core permeability
(Perm-Core) better than the two previous works. Note
that Perm-Log is the current systems while Perm-Vsh is
the system used Vsh as the only input. For instance, the
Core-Perm (+ in Figure 7) between depth 6340 and
6360 feet are with high values (>1000 md). The values
decrease between 6335 and 6340 feet and raise again
at 6330 feet. The hybrid system yields permeability
estimations that match this trend of high-low-high pattern
very well. In contrast, Perm-Vsh matches badly to the
Core-Perm trend. Another previous work, Perm-Log
gives a better match than Perm-Vsh does. However, its
estimations match poorly between the depth of 6325 and

SPE 95167

6340 feet where Core-Perm drops.

For a totally blind test, this system is applied to the
second field where one well is used. In this field, the
overall core permeability is much lower than that from
the previous field. Four lithology types: sand, shaly sand,
sandy shale, and shale, are determined. There is no
high impedance sand. The cutoffs for permeability facies
are listed in Table 3. The input logs, again, are total
porosity, sonic and bulk density. The input data set is
equally divided into three sub-datasets which are used
for training, validation, and testing. The results show a
very good match to the core permeability (Figure 8).

Conclusions
A hybrid GP-Fuzzy/Neural system has been shown to be
robust in estimating permeability from elastic parameter
input. The divide-and-conquer approach is very effective
in dealing with highly heterogeneous reservoirs. This is
demonstrated in the exceptional performance of the
hybrid system. Once the lithology type and permeability
facies are identified, the estimation of permeability
becomes more robust and easier. In comparison with
current estimation approaches, this system yields the
estimated permeability that matches core permeability
more consistently. This system is capable of using input
data which is derivable from seismic as input and
creating an output permeability volume.

References
1. Mathisen T., S. Lee, and A. Datta-Gupta, 2001,
Improved permeability estimates in carbonate
reservoirs using electrofacies characterization: a
case study of the North Robertson Unit, West Texas,
SPE 70034 paper at the 2001 SPE Annual Technical
Conference and Exhibition.
2. Lee S., A. Kharghoria, and A. Datta-Grpta, 2002,
Electrofacies characterization and permeability
predictions in complex reservoirs, SPE Reservoir
Evaluation and Engineering.
3. Suryanarayana K. and K. Al Amari, 2003,
Permeability and RRT estimation from conventional
logs in a Middle East carbonate reservoir using neural
network approach, SPE 81473 paper at the 2003 SPE
Annual Technical Conference and Exhibition.
4. Badarinadh V., K. Suryanarayana and F. Youssef,
2002, Log-derived permeability in a heterogeneous
carbonate reservoir of Middle East, Abu Dhabi, using
artificial neural network, SPE 74345 paper at the 2002
SPE Annual Technical Conference and Exhibition
5. Fahad A. and S. Holditch, 2000, Permeability
estimation using hydraulic flow units in a Central
Arabia reservoir, SPE 63254 paper at the 2000 SPE
Annual Technical Conference and Exhibition
6. Quinlan Ross, 1993, C4.5, Programming for machine
learning, Morgan Kaufmann Publishers.
7. Koza, J., 1992 Genetic Programming: On the
programming of computers by means of Natural
Selection. Cambridge, MA: MIT Press.
8. Yu T., D. Wilkinson, and D. Xie, 2003, A hybrid

SPE 95167

approach for reservoir characterization: Genetic

Programming theory and Practice, edited by R. Riolo
and B. Worzel, Kluwer Academic Publishers.
9. Jang, J-S R., 1993, ANFIS: Adaptive-Network-Based
fuzzy inference system, IEEE Transactions on
Systems, Man and Cybernetics 12 (3): 665-685.

Replace the two

losers (in the
selected four) with
the two new

Select four classifiers

and compete to pick
two final winners

a population
of classifiers

Evaluate the
new classifiers

Authors acknowledge Chevron for granting permission

for the work to be published. The discussions with
Sundar Sundararaman, Dave Larue, and Feng Jian have
been very useful.

Use crossover/mutation to
create two new classifiers

* /

Figure 1 Genetic Programming cycle

B1

A1

z1 =
p1*x+q1*y+r1

Z=
w1*z1+w2*z

A2

B2

w
y

Input MFs

Fuzzy Rules

Figure 2 TSK fuzzy inference system structure

Acknowledgements

z2 =
p2*x+q2*y+r2
Fuzzy Inference

w1+w

SPE 95167

High Perm Fuzzy

System
Med Perm Fuzzy
System
Low Perm Fuzzy
System

Sand Perm
Classifier

M
L

High Perm Fuzzy

System
Med Perm Fuzzy
System
Low Perm Fuzzy
System

Shaly Sand
Perm
Classifier

M
L

High Perm Fuzzy

System
Med Perm Fuzzy
System
Low Perm Fuzzy
System

W ell log data

Sandy Shale
Perm
Classifier

Lithology
Analyzer

M
L

Med Perm Fuzzy

System
Low Perm Fuzzy
System

Shale
Perm
Classifier

Perm eability

High Im p sand
Fuzzy system

Figure 3 GP and fuzzy hybrid system high-level tructure

10000

1000

100
KU_D143A
Well A
KU_A02
Well B
KU_C01
Well C
Well D
KU_C04
Well E
KU_B05

10

0.1

0.01

0.001
0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Figure 4 Core permeability vs. Vsh (five wells)

Lithology
type

High
permeability

Low
permeability

Sand
High imped.
sand

>1000 md

Shale sand

>300 md

300-100 md

<10 md

Sandy shale

>300 md

300-100 md

<10 md

>10md

<10 md

Shale sand

>1000 md

Medium
permeability
1000-100
md

<10 md

Table 1 Permeability facies cutoffs for different lithology

sand
High
permeability
Medium perm.
Low
permeability

shaly
sand

sand
shale

shale

93.2%
82.9%

93.8%
79.0%

67.5%
79.0%

85.0%

100.0%

88.9%

80.6%

96.6%

Table 2 Classification accuracy of GP classifier

SPE 95167

Figure 5 Core vs. estimated permeability from

hybrid system

Sand
Shaly
sand
Sandy
shale
Shale

High
permeability

Medium
permeability

Low
permeability

>100 md

100-20md

<20md

>100 md

100-1 md

<1md

>100 md
>10md

100-1 md

<1md
<10 md

Table 3 Permeability facies cutoffs for different lithology

types of the second field

Figure 6 Core vs. estimated permeability using

existing method

SPE 95167

Well C

Figure 7 Performance comparisons of

hybrid system with two previous systems
Perm-FIS: Our GP and Fuzzy/Nerual hybrid
Perm-Log: Multi-regression using five logs
Perm-Vsh: Regression using Vsh only
Perm-Core: Core permeability samples

Well from offshore West Africa

Figure 8 The blind test results of hybrid system of the second field
(one well only, core vs. estimation)