Expert Systems with Applications 37 (2010) 5353–5363

Contents lists available at ScienceDirect

Expert Systems with Applications

journal homepage: www.elsevier.com/locate/eswa

Hybrid computational models for the characterization of oil and gas reservoirs
Tarek Helmy *,1, Anifowose Fatai, Kanaan Faisal
Information and Computer Science Department, College of Computer Science and Engineering, King Fahd University of Petroleum and Minerals, Mail Box. 413,
Dhahran 31261, Saudi Arabia

a r t i c l e i n f o a b s t r a c t

Keywords: The process of combining multiple computational intelligence techniques to build a single hybrid model
Hybrid computational intelligence has become increasingly popular. As reported in the literature, the performance indices of these hybrid
Petroleum reservoir characteristics models have proved to be better than the individual components when used alone. Hybrid models are
Fuzzy Logic extremely useful for reservoir characterization in petroleum engineering, which requires high-accuracy
Support Vector Machines
predictions for efficient exploration and management of oil and gas resources.
Functional Networks
In this paper, we have utilized the capabilities of data mining and computational intelligence in the
prediction of porosity and permeability, two important petroleum reservoir characteristics, based on
the hybridization of Fuzzy Logic, Support Vector Machines, and Functional Networks, using several
real-life well-logs. Two hybrid models have been built. In both, Functional Networks were used to select
the best of the predictor variables for training directly from input data by using its functional approxima-
tion capability with least square fitting algorithm. In the first model (FFS), the selected predictor variables
were passed to Type-2 Fuzzy Logic System to handle uncertainties and extract inference rules, while Sup-
port Vector Machines made the final predictions. In the second, the selected predictor variables were
passed to Support Vector Machines for training by transforming them to a higher dimensional space,
and then to Type-2 Fuzzy Logic to handle uncertainties, extract inference rules and make final predic-
tions.
The simulation results show that the hybrid models perform better than the individual techniques
when used alone for the same datasets with their higher correlation coefficients. In terms of execution
time, the hybrid models took less time for both training and testing than the Type-2 Fuzzy Logic, but
more time than Functional Networks and Support Vector Machines. This could be the price for having
a better and more robust model. The hybrid models also performed better than a combination of two
of the individual components, Type-2 Fuzzy Logic and Support Vector Machines, in terms of higher cor-
relation coefficients as well as lower execution times. This is due to the effective role of Functional Net-
works, as a best-variable selector in the hybrid models.
Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction tained in a reservoir and its ability to flow. These properties

Petroleum reservoir characterization is a process for quantita-
tively describing various reservoir properties in spatial variability
by using available field data. It plays a crucial role in modern res-
ervoir management: making sound reservoir decisions and
improving the reliability of the reservoir predictions. The ultimate
goal is a reservoir model with realistic tolerance for imprecision
and uncertainty. Porosity and permeability are the two fundamen-
tal reservoir properties which relate to the amount of fluid con-
* Corresponding author.
E-mail addresses: helmy@kfupm.edu.sa (T. Helmy), anifowo@kfupm.edu.sa (A.
Fatai), kanaan@kfupm.edu.sa (K. Faisal).
1
Tarek Helmy is on leave from College of Engineering, Department of Computers
Engineering & Automatic Control, Tanta University, Egypt.
make significant impacts on petroleum field operations and reser-
voir management (Lim, 2005).
A good number of studies have been carried out on the use of
various Computational Intelligence (CI) schemes, such as Logistic
Regression (LR), K-Nearest Neighbor (KNN), Multilayer Perceptrons
(MLP), Radial-Basis Function (RBF), Bayesian Belief Networks
(BBN), Naïve Bayes (NB), Random Forests (RF), Functional Net-
works (FunNets), Support Vector Machines (SVM), Artificial Neural
Networks (ANN), Probabilistic Networks (PN), Adaptive-Neuro
Fuzzy Systems (ANFIS) and Decision Trees (DT) (Duch, Adamczak,
& Jankowski, 1997; Giovanni & Vincenzo, 2005; Guojie, 2004; He,
2008; Hosmer & Lemeshow, 2000; Lauría and Peter Duchessi,
2006; Salah, Rahman, & Nath, 2005), to predict the characteristics
of oil and gas reservoirs such as depth, temperature, pressure, vol-
ume, drive mechanism, structure and seal, well spacing, well-bore

0957-4174/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2010.01.021
5354 T. Helmy et al. / Expert Systems with Applications 37 (2010) 5353–5363
integrity, porosity and permeability. CI covers many existing
branches of science viz. Artificial Neural Networks, Fuzzy Systems,
Evolutionary Computing and Hybrid Systems. It also embraces
techniques that use Swarm Intelligence, Fractals and Chaos Theory,
Artificial Immune Systems, Wavelets, etc. (He, 2008).
The combination of two or more Computational Intelligence
schemes as a single model is called Hybrid Computational Intelli-
gence (HCI) and is becoming increasingly popular. This increased
popularity lies in the extensive success of hybrid systems in many
real-world complex problems (Giovanni & Loia, 2005). A key prere-
quisite for the merging of technologies is the existence of a ‘‘com-
mon denominator” to build upon (Andreas, Symeonidis, & Pericles
Mitkas, 2005). In our case, part of the ‘‘common denominator” for
Fuzzy Logic (FL), Support Vector Machines (SVM) and Functional
Networks (FN) is the inference procedures they deploy and their
excellent predictive capabilities. We attempt to combine the indi-
vidual capabilities of FL, SVM and FN in an HCI scheme, to predict
two characteristics of oil and gas reservoirs, namely porosity and
permeability, with better performance indices.
Our motivations for this work include the quest for higher per-
formance accuracy in the prediction of oil and gas characteristics,
the recently increasing popularity of hybrid intelligent systems,
the reported success of these systems in many real-world complex
problems, the need to complement the weaknesses of one algo-
rithm with the advantages of the others and hence to combine
the cooperative and competitive characteristics of the individual
techniques and the existing theoretic and experimental justifica-
tions (Munakata, 2008; Nong, 2003; Peddabachigari, Abrahamb,
Grosan, & Thomas, 2007; Phillips-Wren, Ichalkaranje, & Jain,
2008; Sun et al., 2007; Triantaphyllou & Felici, 2006; Wang & Fu,
2005) that hybrids produce more accurate results than the individ-
ual techniques used separately.
The rest of this paper is organized as follows: Section 2 reviews
the literature on how Artificial Intelligence (AI) can be employed in
Petroleum Engineering. Section 3 presents a brief overview of Fuz-
zy Logic, Support Vector Machines, Functional Networks and Hy-
brid Systems. Section 4 explains the Well log data and tools,
along with the methodology used in this work. Section 5 discusses
the criteria of quality measurement used to evaluate the validity in
this work while conclusion, with a detailed plan of future work is
described in Section 6.
2. Literature survey
The application of the capabilities of Artificial Intelligence (AI)
has been widely appreciated in petroleum engineering, as well as
in other fields. Some of the areas of petroleum technology in which
AI has been used with success include seismic pattern recognition,
porosity and permeability predictions, identification of sandstone
lithofacies, drill bit diagnosis, and analysis and improvement of
gas well production (Ali, 1994; El Ouahed, Tiab, & Mazouzi,
2005; Goda, Maier, & Behrenbruch, 2007). Porosity and permeabil-
ity measurements are frequently made on plugs extracted from the
core of wells drilled for oil and gas exploration. The data are valu-
able for linking permeability to porosity, and these both serve as
standard indicators of reservoir quality in the oil and gas industry.
Porosity is the percentage of voids and open space in a rock or
sedimentary deposit. The greater the porosity of a rock, the greater
its ability to hold water and other earth materials, such as oil
(Excellence in Educational Development, 2007). Porosity is very
important in evaluating the potential volume of hydrocarbons it
may contain. Permeability is the ease with which fluid is transmit-
ted through a rock’s pore space. Although a rock may be very por-
ous, it is not necessarily very permeable. Permeability is a measure
of how interconnected the individual pore spaces are in a rock or
sediment. Permeability is a key parameter associated with the
characterization of any hydrocarbon reservoir. In fact, it is not pos-
sible to have accurate solutions to many petroleum engineering
problems without having accurate permeability value (K.S. Univer-
sity, 2007).
El Ouahed et al. (2005) produced 2-D fracture intensity and
fracture network maps in a large block of field using Artificial Neu-
ral Network and FL. The results showed that the proposed ap-
proach is a practical methodology to map the fracture network.
Ali and Chawathe (2000) used neural networks to predict perme-
ability from petrographic data while using Fuzzy Logic to screen
and rank the predictor variables with respect to the target variable.
The result demonstrated the generalizing capability of the neural
network. A similar study was done by Saemi, Ahmadi, and Varjani
(2007) when they proposed a new method for the auto-design of
neural networks based on genetic algorithm (GA).
The traditional FL, now referred to as Type-1 FL has featured in a
number of research efforts, especially in reservoir characterization.
Type-2 FL has also featured in many recently published articles in
various fields and especially in reservoir properties modeling (El
Ouahed et al., 2005). One of the earliest references to the applica-
tion of FL in the petroleum industry was by Fang and Chen (1997)
who presented a fuzzy modeling for predicting porosity and per-
meability from the compositional and textural characteristics of
sandstones. They found that fuzzy modeling is assumption-free,
tolerant of outliers, and capable of making both linguistic and nu-
meric predictions based on qualitative and quantitative data.
SVMs have been used extensively in many areas, including oil
and gas (Jian & Wenfen, 2006), with very promising results. Tabo-
ada, Matías, Ordóñez, and García (2007) used different kinds of
SVMs: SVM classification (multi-class one-against-all), ordinal
SVM and SVM regression, and they found that the SVMs are per-
fectly comparable to kriging (a statistical model) and have better
control of outliers.
Like the other AI techniques described in previous sections,
FunNets have also featured in a number of research studies. Castil-
lo (1998) and Castillo, Hadi, and Lacruz (2001a) gave a comprehen-
sive demonstration of the application of FunNets in Statistics and
Engineering.
Most hybrids found in the literature usually contain Neural Net-
works fused with one other technique due to its wide use in the
computational intelligence. Other hybrid systems include different
combinations of several techniques such as SVM, FL, Decision
Trees, Extreme Learning Machines, Genetic Algorithm, Hidden
Markov Model and Radial-basis Functions, as applied in various
fields including reservoir characterization (Chikhi, 2006; Chikhi &
Batouche, 2004).
A fuzzy linear programming Support Vector Machine (LP-SVM)
was used for multiclass classification problems by Shigeo (2004),
demonstrating the superiority of the hybrid over the conventional
SVMs. Deny, Wilkinson, Yu, and Ramon (2005) developed a hybrid
genetic programming and fuzzy/neural network inference system
to estimate the permeability of reservoirs. When compared with
contemporary estimation approaches, the hybrid yielded more
consistent and robust estimated results. Chikhi and Batouche
(2004) combined a neural method with radial-basis functions to
enhance the classification of lithofacies of certain wells. This was
followed by a similar hybridization effort by Chikhi (2006), who
combined Neural Networks with Hidden Markov Models to obtain
the lithological identification of the same wells. The results ob-
tained by the hybrid are close to those obtained by the fuzzy Adap-
tive Resonance Theory (ART) approach applied to the same
borehole with the same well-logs. A hybridization of SVMs and
the Interval Type-2 FL System (FLS) was performed by Chen, Li,
Harrison, and Zhang (2007) to better handle uncertainties existing
in real classification data.
 T. Helmy et al. / Expert Systems with Applications 37 (2010) 5353–5363
Fig. 1. The structure of a Type-2 FLS.
Fig. 2. Gaussian MF with uncertain mean.
3. Fuzzy Logic, Support Vector Machines, Functional Networks
and Hybrids
3.1. Fuzzy Logic
Type-2 FLS was introduced as an extension of the concept of
Type-1 FLS. Type-2 FLS has membership grades that are them-
selves fuzzy. For each value of a primary variable (e.g., pressure
and temperature) the membership is a function (not just a point
value). The secondary Membership Function (MF), whose domain
is in the interval [0, 1], and whose range may also be in [0, 1].
Hence, the MF of a Type-2 FLS is three dimensional, and it is the
new third dimension that provides new degrees of design freedom
for handling uncertainties. Type-2 FL does not obtain good perfor-
mance when the number of training data is small, but it can per-
form better when the number of training prototypes is large
(Fang & Chen, 1997; Karnik & Mendel, 1999; Mendel, 2003; Zaran-
di, Rezaee, Turksen, & Neshat, 2007). Figs. 1 and 2 show the struc-
ture of a Type-2 FLS and Gaussian MF with uncertain mean,
respectively.
3.2. Support Vector Machines
Support Vector Machines are a set of related supervised learn-
ing methods used for classification and regression. They belong
to a family of generalized Linear Classifiers. They can also be con-
sidered as a special case of Tikhonov Regularization. SVMs map in-
put vectors to a higher dimensional space where a maximal
separating hyperplane is constructed (Burges, 1998; Littman,
2003). This is shown in Fig. 3.
5355
Regression in Primal
Y
Training
Dataset
Fig. 3. Mapping input vectors to a higher dimensional space in SVM.
3.3. Functional Networks
Functional Networks are extensions of Neural Networks which
consist of different layers of neurons connected by links. Each com-
puting unit or neuron performs a simple calculation: a scalar typ-
ically monotone function f of a weighted sum of inputs. The
function f, associated with the neurons, is fixed and the weights
are learned from data using some well-known algorithms. A Func-
tional Network consists of: a layer of input units containing the in-
put data; a layer of output units containing the output data; one or
several layers of neurons or computing units which evaluate a set
of input values, coming from the previous layer, and which give a
set of output values to the next layer of neurons or output units.
The computing units are connected to each other, in the sense that
the output from one unit can serve as part of the input to another
neuron or to the units in the output layer. Once the input values
are given, the output is determined by the neuron type, which
can be defined by a function (Castillo, 1998; Castillo et al.,
2001a; Castillo, Gutiérrez, Hadi, Lacruz, 2001b; Chen et al., 2007).
Fig. 4 shows the structure of a Functional Network and its
simplification.
3.4. Hybrids
An approach resulting from the combination of two or more ap-
proaches is called a hybrid. It has also been defined as an approach
that combines different theoretical backgrounds and algorithms
such as data mining and soft computing methodologies. The main
idea behind hybridization is to complement the weaknesses of one
technique with the strength of other techniques. Since no single
technique is good for everything and in all situations, there is a
need to combine the individual capabilities of each technique to
obtain a more versatile and robust technique.
5356 T. Helmy et al. / Expert Systems with Applications 37 (2010) 5353–5363

form in North America (site 1) and the three for permeability from
a drilling site in the Middle East (site 2). The datasets from site 1
have six predictor variables for porosity, while the dataset from
site 2 has eight predictor variables for permeability. These are
shown in Tables 1 and 2.

5. Experimental design and model validation

5.1. Experimental design

The methodology in this work is based on the standard Compu-

tational Intelligence approach to hybridization of Artificial Intelli-
gence technique using Fuzzy Logic (FL), Support Vector Machines
(SVM) and Functional Networks (FN). The hybrid models were de-
signed to benefit immensely from the strength of the individual
techniques, and to complement the weaknesses of one technique
with the advantages of the others, and hence to combine the coop-
erative and competitive characteristics of the individual
techniques.
Fig. 4. Structure of Functional Networks and its simplification. The hybrid models were implemented by using mainly MATLAB
codes and MATLAB toolboxes for SVM, FN and FL. In order to show
the reasons for the choices of FL, SVM and FN, Table 3 presents the
Table 1 areas of strengths and weaknesses of the techniques. The choice of
Predictor variables for site 1 well log for porosity. Type-2 FLS lies in its ability to determine an exact membership
Predictors for porosity function for a fuzzy set; hence, it is useful for handling uncertain-
1 Core
ties (Karnik & Mendel, 1999; Mendel, 2003). It is also strong in the
2 Top interval process of extracting rules, to be used for inferencing, directly from
3 Grain density the input data. However, its disadvantage lies in its complexity of
4 Grain volume implementation and hence too much time spent in tuning the
5 Length
parameters used for inferencing during the process of training,
6 Diameter
compared to other techniques. Also, Type-2 FL does not obtain
good performance when the number of training data is small, but
it performs better when the number of training prototypes is large
Table 2
Predictor variables for site 2 well log for permeability. (Mendel, 2003). This weakness is intended to be complemented
with the ability of SVM to handle small a dataset. Although the
Predictors for permeability Full meaning
use of SVMs in applications has only recently begun, application
1 GR Gamma ray log developers have already reported state-of-the-art performances
2 PHIE Porosity log
in a variety of applications in pattern recognition, regression esti-
3 RHOB Density log
4 SWT Water saturation mation, and time series prediction (Jian & Wenfen, 2006). The
5 RT Deep resistivity strengths of SVM lie mainly in its relative ease of training, there
6 MSFL Microspherically focused log is no local optimal similar to neural networks, scalability to high
7 NPHI Neutron porosity log dimensional data, ability to explicitly control the tradeoff between
8 CALI Caliper log
complexity and error, and ability to handle non-traditional data
like trees as input to the system, instead of feature vectors. It is also
4. Description of data known to have the capability of using a small training dataset (Jian
& Wenfen, 2006). However, it is weak in the sense that it needs a
Well-logs for porosity and permeability from six wells were ‘‘good” kernel function (Littman, 2003). This weakness is intended
used for the validation of this work. The three well-logs for poros- to be complemented with the ability of FunNet and FL to learn di-
ity were obtained from a drilling site in the Northern Marion Plat- rectly from the input data.

Table 3
Strengths and weaknesses of FL, SVM and FN.

Technique Strength Weakness

Type-2 Ability to determine an exact membership functions for a fuzzy set Complexity of implementation
FLS Ability to handle uncertainties Too much time spent in tuning the parameters used for inferencing during the
training process
Ability to extract rules directly from input data Does not obtain good performance when the number of training data is small
SVM Ease of training It is weak in the sense that it needs a ‘‘good” kernel function
No local optima, unlike in neural networks
It scales relatively well to high dimensional data
Ability to explicitly control the tradeoff between complexity and error
It is also known to have the capability of using small training dataset
FN There is no need to include weights associated with links
Functional approximation
Ability to select the best among functions by minimizing the sum of
square errors
 T. Helmy et al. / Expert Systems with Applications 37 (2010) 5353–5363 5357

Handling Transformation to
Uncertainties from higher dimensional
Least Square
best-variable data space for regression
Fitting

Input FunNets Fuzzy Logic SVM Predicted

Data Output

Data with best

input
Rule Extraction
variables
from best-variable
data

Fig. 5. The conceptual design framework of the FFS hybrid model.

Handling
Least Square Uncertainties from
Fitting best-variable data

Input FunNets SVM Fuzzy Logic Predicted

Data Output

Data with best Transformation to Rule Extraction from

input variables higher dimensional best-variable data
space for regression

Fig. 6. The conceptual design framework of FSF hybrid model.

works, for regression. Fig. 5 shows the conceptual design frame-

Table 4 work of this model.
Division of datasets into training and testing. In the second model, Functional Networks–SVM–Fuzzy Logic
Wells Site 1 (Porosity) Site 2 (Permeability)
(FSF), Support Vector Machines are used to transform the best of
the predictor variables from Functional Networks to a higher
1 2 3 1 2 3
dimensional space suitable for training the Type-2 Fuzzy Logic that
Data size 415 285 23 355 477 387 will handle uncertainties, extract inference rules and make the fi-
Training (70%) 291 200 16 249 334 271
nal predictions. Fig. 6 shows the conceptual design framework of
Testing (30%) 124 85 7 106 143 116
this model.

5.3. Model implementation and validation

The learning process of a Functional Network consists of obtain-
ing the neural functions from a set of training data based on min-
imizing the sum of squared errors between the input and the target
output by suggesting an approximation to each of the functions
and selecting the best among them (Castillo et al., 2001b).
5.2. Conceptual framework of the hybrid models
A Functional Network was used as the base for the two hybrid
models. This is due to its functional approximation capability to se-
lect the best of the predictor variables directly from the training
data. In the first model, Functional Networks–Fuzzy Logic–SVM
(FFS), Type-2 Fuzzy Logic is used to handle uncertainties and to ex-
tract inference rules directly from the best predictor variables ob-
tained from the Functional Networks block. The output of this is
passed to Support Vector Machines that have previously been
trained with the best predictor variables from Functional Net-
The available data for each of the wells are divided into training
and test data by using a stratified sampling technique. Seventy per-
cent of the entire data goes for training, and the remaining 30%
goes for testing. To further ensure fairness and integrity of the re-
sults obtained, several iterations were made, and the average of the
runs was obtained. Table 4 shows the two well-logs with their
sizes and divisions into training and test sets.
5.4. Control experiment: removing functional from the Hybrid models
In order to appreciate the role performed by the FN block in the
hybrid models, another set of experiments was performed, using
only two of the components namely: Type-2 Fuzzy Logic and
SVM. This reduced architecture was chosen because the two hybrid
models have similar configurations and demonstrate very compet-
itive levels of performance.
5358 T. Helmy et al. / Expert Systems with Applications 37 (2010) 5353–5363

Table 5
Result of the porosity prediction site 1, well 1.

Model Correlation coefficient RMSE Execution time

Training Testing Training Testing Training Testing
SVM 0.820348 0.94502 7.54717 2.80199 15.458333 0.000000
FunNet 0.835855 0.957491 6.17685 4.15040 0.098958 0.000000
Fuzzy Logic 0.840697 0.816021 6.85104 7.52180 155.395833 55.937500
Hybrid-FFS 0.920608 0.969179 7.06432 3.72873 60.979167 9.093750
Hybrid-FSF 0.916415 0.957003 6.61915 5.67851 61.562500 8.890625

Table 6
Result of the porosity prediction site 1, well 2.

Model Correlation coefficient RMSE Execution time

Training Testing Training Testing Training Testing
SVM 0.806304 0.775851 6.75723 7.17699 7.354167 0.000000
FunNet 0.803918 0.775465 6.77203 7.18608 0.093750 0.000000
Fuzzy Logic 0.798557 0.723302 7.31978 7.69430 71.526042 25.817708
Hybrid-FFS 0.82147 0.80758 9.18052 7.09860 28.442708 4.411458
Hybrid-FSF 0.8147 0.813319 9.09662 9.45112 31.567708 4.296875

Table 7
Result of the porosity prediction site 1, well 3.

Model Correlation coefficient RMSE Execution time

Training Testing Training Testing Training Testing
SVM 0.980730 0.898397 1.64195 2.95456 0.315625 0.000000
FunNet 0.824680 0.803017 4.75388 6.26507 0.062500 0.000000
Fuzzy Logic 0.910924 0.599756 4.26323 7.04103 0.518750 0.129688
Hybrid-FFS 0.93055 0.914020 6.37047 3.34832 0.493750 0.042188
Hybrid-FSF 0.920961 0.928830 6.81454 8.32623 0.471094 0.051042

Fig. 7. Correlation coefficients comparisons for porosity training and testing.

5.5. Criteria for performance evaluation The RMSE is one of the most commonly used measures of suc-
cess for numeric prediction, computed by taking the square root of
In order to establish a valid and reasonable evaluation of this the average of the squared differences between each predicted va-
work, we have used Correlation Coefficient (CC), Root Mean- lue xn and its corresponding actual value yn. It gives the error value
Squared Error (RMSE), and Execution Time (ET) as criteria for mea- the same dimensionality as the actual and predicted values
suring the performance. The CC measures the statistical correlation
between the predicted and actual values. A value of ‘‘1” means per- qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ðx1 y1 Þ2 þðx2 y2 Þ2 þþðxn yn Þ2
fect statistical correlation and a ‘‘0” means there is no correlation at The formula is : n :
all. where n is the size of data
P
ðx  x0 Þðy  y0 Þ
The formula is : qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P P
ðx  xÞ2 ðy  y0 Þ2 The ET is simply the total time taken for a technique to run from the
beginning to the end, and it is computed as: T 2  T 1 , where T2 is the
where x and y are the actual and predicted values, while x0 and y0 are CPU time at the end of the run and T1 is the CPU time at the begin-
the mean of the actual and predicted values. ning of the run.
 T. Helmy et al. / Expert Systems with Applications 37 (2010) 5353–5363 5359

Fig. 8. Execution times comparisons for porosity training and testing.

Fig. 9. Execution time comparison for well 3 porosity training and testing.

Table 8
Result of the permeability prediction for site 2, well 1.

Model Correlation coefficient RMSE Execution time

Training Testing Testing Testing Training Testing
SVM 0.854109 0.832785 0.64724 0.68445 10.052083 0.000000
FunNet 0.852724 0.833204 0.64785 0.68113 0.140625 0.000000
FL 0.869565 0.844982 0.62187 0.66946 187.494792 68.828125
Hybrid-FFS 0.872843 0.858610 1.10615 0.66580 43.333333 6.231250
Hybrid-FSF 0.827372 0.901920 0.72800 0.64289 44.630208 6.531250

Table 9
Result of the permeability prediction for site 2, well 2.

Model Correlation coefficient RMSE Execution time

Training Testing Testing Testing Training Testing
SVM 0.864894 0.881033 0.63803 0.61685 18.661458 0.005208
FunNet 0.868102 0.877959 0.62799 0.62760 0.182292 0.000000
Fuzzy Logic 0.860039 0.866123 0.65863 0.67214 355.026042 128.739583
Hybrid-FFS 0.88142 0.895119 1.11178 0.60509 79.062500 12.463542
Hybrid-FSF 0.88144 0.903620 0.73102 0.72689 87.697917 14.140625

6. Experimental results
We implemented and validated the individual techniques as
well as the two hybrid models in the prediction of porosity
and permeability by using the training and testing data
described in the previous section. Several iterations were made,
and the average of results was taken. Tables 5–7 present the
result of the prediction of porosity for the three wells in
site 1.
These are summarized graphically in Figs. 7–9.
Tables 8–10 present the result of the prediction of Permeability
for the three wells in site 2.
These are also summarized graphically in the plots shown in
Figs. 10 and 11 for both training and testing.
In order to appreciate the role performed by the FN block in the
FFS and FSF hybrid systems, another set of experiments was per-
formed by using only two of the components namely: Type-2 Fuz-
zy Logic and SVM.
The results of the simulations for the three Porosity and three
Permeability wells are shown in Tables 11 and 12, respectively.
5360 T. Helmy et al. / Expert Systems with Applications 37 (2010) 5353–5363

Table 10
Result of the permeability prediction for site 2, well 3

Model Correlation coefficient RMSE Execution time

Training Testing Testing Testing Training Testing
SVM 0.800820 0.767553 0.73625 0.76481 12.802083 0.010417
FunNet 0.787706 0.730623 0.74712 0.76140 0.177083 0.000000
Fuzzy Logic 0.753439 0.741249 0.83188 0.86444 241.822917 89.036458
Hybrid-FFS 0.768380 0.789960 1.10064 0.73154 51.885417 7.791667
Hybrid-FSF 0.769102 0.801785 0.78078 0.78106 59.447917 11.098958

Fig. 10. Correlation coefficients comparisons for permeability training and testing.

Fig. 11. Execution time comparisons for porosity training and testing.

Table 11
Results of the simulations for the three porosity wells.

Wells CC RMSE Execution time (s)

Site 1, well 1 0.811098 0.789864 7.33084 8.64583 297.864583 56.411458
Site 1, well 2 0.781196 0.700457 7.18849 8.20909 137.381250 26.159375
Site 1, well 3 0.815213 0.708821 4.61251 5.32429 1.380208 0.125000

Comparisons of the performance of the Type-2-SVM above with
the FFS and FSF Hybrids are shown in Figs. 12–15.
6.1. Discussion of results
In the prediction of porosity and permeability, it is clear from
the results that the hybrid models performed better than, or com-
petitively equal to, the three individual techniques used separately,
in terms of their correlation coefficient. A special characteristic was
observed in the result of site 1 well 3 for Porosity, where SVM dem-
onstrated its ability to withstand a shortage of training data but FL
demonstrated otherwise. Still, the hybrid models performed better
than the individual components due to the cooperative spirit that
was been built into them.
In terms of execution time, the results showed that Functional
Networks are the fastest in terms of both training and testing, fol-
lowed by SVM. Type-2 Fuzzy Logic took the most time for both
training and testing, due to its complexity as described in Sec-
tion 3.1 and Table 3. The hybrid models proved to be faster than
the Fuzzy Logic component, but not for SVM and FN. This is the
price for obtaining better models in terms reliability and
robustness.
When compared to the Type-2-SVM Hybrid, the FFS and FSF Hy-
brids also proved to be better in terms of both correlation coeffi-
cient and execution time. The better performance in correlation
coefficient can be attributed to the role of Functional Networks
in the FFS and FSF hybrids. The Functional Networks block serves
as a best-variable selector, which extracts from the input variables
 T. Helmy et al. / Expert Systems with Applications 37 (2010) 5353–5363 5361

Table 12
Results of the simulations for the three permeability wells.

Wells CC RMSE Execution time (s)

Site 2, well 1 0.814240 0.831428 0.61296 0.61174 358.395833 69.317708
Site 2, well 2 0.823580 0.817343 0.64555 0.68116 644.093750 125.03125
Site 2, well 3 0.728086 0.715730 0.90577 0.86452 564.942708 109.015625

Fig. 12. Correlation coefficient comparisons for porosity training and testing.

Fig. 13. Execution time comparisons for porosity training and testing.

Fig. 14. Correlation coefficient comparisons for permeability training and testing.

only those variables that are most relevant to the prediction The better performance of the FFS and FSF hybrids in terms of
system. execution times is also due to the above reason. In the process of
5362 T. Helmy et al. / Expert Systems with Applications 37 (2010) 5353–5363
Fig. 15. Execution time comparisons for permeability training and testing.
best-variable selection by the Functional Networks block, the
dimensionality of the input variable that goes to the next block is
reduced. This works in favor of the Type-2 FLS that does not work
well with an input data of very high dimensionality.
7. Conclusion
Two conceptual design frameworks for the hybridization of Fuz-
zy Logic, Support Vector Machines and Functional Networks have
been implemented and presented. They were tested by using two
well-logs containing six porosity and permeability datasets. The
results showed that the hybrids performed better than the individ-
ual techniques used separately for all the datasets in the prediction
of porosity and permeability. With any novel data, the hybrid will
be able to select the best model to use, extract inference rules, and
handle uncertainties that might be present, directly from the data,
and perform the required prediction. The hybrid models are also
seen to be very reliable, robust and effective, given the reported
and observed good performance of the individual components.
Despite the good performance of the hybrid, we could not fully
validate its performance due to lack of data that contain uncertain
measurements. The data we used had no such uncertainty. Conse-
quently, the design, implementation and validation of the models
were based on well-logs and not on expert knowledge. Also, the
Type-2 FLS has some limitations, and our framework was built
on these limitations.
Motivated by the success of this work, part of our future work
will be directed towards modifying this hybrid model in order to
solve classification and pattern recognition problems in the oil
and gas industry such as Lithofacie and History Matching. More
experiments will be done on these hybrids if data is available with
uncertain numerical measurements, and the frameworks will be
validated by using expert data instead of well-logs.
Acknowledgments
We would like to thank both KACST and King Fahd University of
Petroleum and Minerals for providing the computing facilities and
support. Special thanks go to Mr. David Birkett for his help in proof
reading of the paper.
References
Ali, J. K. (1994). Neural networks: A new tool for the petroleum industry? In
European petroleum computer conference, Aberdeen, UK, March 1994.
Ali, M., & Chawathe, A. (2000). Using artificial intelligence to predict permeability
from petrographic data. Computers & Geosciences, 26, 915–925.
Andreas, L., Symeonidis, L., & Pericles Mitkas, A. (2005). Agent intelligence through
data mining, Multiagent systems, artificial societies, and simulated organizations
series, Vol. 14. International Book Series.
Burges, C. J. C. (1998). A tutorial on support vector machines for pattern recognition.
Data Mining and Knowledge Discovery, 2, 121–167.
Castillo, E. (1998). Functional networks. Neural Processing Letters, 7, 151–159.
Castillo, E., Gutiérrez, J. M., Hadi, A., & Lacruz, B. (2001). Some applications
of functional networks in statistics and engineering. Technometrics 43,
10–24.
Castillo, E., Hadi, A., & Lacruz, B. (2001). Optimal transformations in multiple linear
regression using functional networks. In International work-conference on
artificial and natural neural networks. Lecture Notes in Computer Science (Vol.
2084, pp. 316–324).
Chen, X., Li, Y., Harrison, R., & Zhang, Y. Q. (2007). Type-2 fuzzy logic based classifier
fusion for support vector machines. Applied Soft Computing Journal. doi:10.1016/
j.asoc.2007.02.019.
Chikhi, S. (2006). A fuzzy ART versus hybrid NN-HMM methods for lithology
identification in the Triassic Province. IEEE Transactions, 7803-9521-2/06.
Chikhi, S., & Batouche, M. (2004). Probabilistic neural method combined with
radial-bias functions applied to reservoir characterization in the Algerian
Triassic Province. Journal of Geophysics and Engineering, 1, 134–142.
Deny, X., Wilkinson, D., Yu, T., & Ramon, S. (2005). California, permeability
estimation using a hybrid genetic programming and fuzzy/neural inference
approach. In 2005 Society of petroleum engineers annual technical conference and
exhibition held in Dallas, TX, USA, 9–12 October, 2005.
Duch, W., Adamczak, R., & Jankowski, N. (1997). Initialization and optimization of
multilayered perceptrons. In: Third conference on neural networks and their
applications, Kule, Poland, 1997.
El Ouahed, A. K., Tiab, D., & Mazouzi, A. (2005). Application of artificial intelligence
to characterize naturally fractured zones in Hassi Messaoud Oil Field, Algeria.
Journal of Petroleum Science and Engineering, 49, 122–141.
S. Excellence in Educational Development, Science Lab Project (2007). <http://
www.seed.slb.com/en/scictr/lab/porosity/index.htm>.
Fang, J. H., & Chen, H. C. (1997). Fuzzy modeling and the prediction of porosity and
permeability from the compositional and textural attributes of sandstone.
Journal of Petroleum Geology, 20(2), 185–204.
Giovanni, A., & Vincenzo, L. (2005). Using FML and Fuzzy technology in adaptive
ambient intelligence environments. International Journal of Computational
Intelligence Research, 1(2), 171–182., ISSN: 0973-1873. <http://www.ijcir.info>.
Goda, H. M., Maier, & Behrenbruch, H. R. (2007). Use of artificial intelligence
techniques for predicting irreducible water saturation – Australian
Hydrocarbons Basins. In Society of Petroleum Engineers Asia Pacific Oil & Gas
Conference and Exhibition, Jakarta, Indonesia, 2007.
Guojie, L. (2004). Radial basis function neural network for speaker verification. A
Master of Engineering Thesis. The Nanyang Technological University.
He, J. (2008). Computational intelligence. Research interests, <http://
www.cs.bham.ac.uk/~jxh/hejunrs.html>.
Hosmer, D., & Lemeshow, S. (2000). Applied logistic regression (p. 392). John Wiley
and Sons.
Jian, H., & Wenfen, H. (2006). Novel approach to predict potentiality of enhanced oil
recovery. In Society of petroleum engineers intelligent energy conference and
exhibition, Amsterdam, The Netherlands, 2006.
K.S. University. (2007). Electronic document Archive, docs.ksu.edu.sa/DOC/
Articles49/Article490822.doc.
Karnik, N. N., & Mendel, J. M. (1999). Type-2 fuzzy logic systems. IEEE Transactions
on Fuzzy Systems, 7(6), 5.
Eitel Lauría, J. M., & Peter Duchessi, J. (2006). A Bayesian belief network for it
implementation decision support. In Decision support systems and international
federation for information processing international conference on decision support
systems, 2006.
 T. Helmy et al. / Expert Systems with Applications 37 (2010) 5353–5363
Li Sun, Z., Au, K., & Choi, T. (2007). A neuro-fuzzy inference system through
integration of fuzzy logic and extreme learning machines. IEEE Transactions on
Systems, Man, and Cybernetics – Part B: Cybernetics, 37(5), 1321–1331.
Lim, J. (2005). Reservoir properties determination using fuzzy logic and neural
networks from well data in offshore Korea. Journal of Petroleum Science and
Engineering, 49, 182–192.
Littman, W. T. A. (2003). Introduction to support vector machines. Machine learning
course 536. In Tom Mitchell, Text book: Machine learning, McGraw Hill, 1997.
Department of Computer Science, Rutgers University, The State University of
New Jersey, USA. <http://www.cs.rutgers.edu/~mlittman/courses/ml03/>.
Mendel, J. M. (2003). Type-2 fuzzy sets: Some questions and answers. IEEE
Connections, Newsletter of the IEEE Neural Networks Society, 1, 10–13.
Munakata, T. (2008). Fundamentals of the new artificial intelligence neural,
evolutionary, fuzzy and more (2nd ed.). Springer Science + Business Media.
Nong, Y. (2003). The handbook of data mining pp. 561, 617. Lawrence Erlbaum
Associates.
Peddabachigari, S., Abrahamb, A., Grosan, C., & Thomas, J. (2007). Modeling
intrusion detection system using hybrid intelligent systems. Journal of
Network and Computer Applications, 30, 114–132.
Phillips-Wren, G., Ichalkaranje, N., & Jain, C. C. (2008). Intelligent decision making: An
AI-based approach, studies in computational intelligence (Vol. 97, pp. 88, 102,
355). Berlin, Heidelberg: Springer-Verlag.
5363
Saemi, M., Ahmadi, M., & Varjani, A. Y. (2007). Design of neural networks using
genetic algorithm for the permeability estimation of the reservoir. Journal of
Petroleum Science and Engineering, 59, 97–105.
Salah, A., Rahman, S., & Nath, K. (2005). An enhancement of K-nearest neighbor
classification using genetic algorithm. In Midwest instruction and computing
symposium.
Shigeo, A. (2004). Fuzzy LP-SVMs for multiclass problems. In Proceedings of the
European symposium on artificial neural networks, Bruges, Belgium, pp. 429–434.
Taboada, J., Matías, J. M., Ordóñez, C., & García, P. J. (2007). Creating a quality map of
a slate deposit using support vector machines. Journal of Computational and
Applied Mathematics, 204, 84–94.
Triantaphyllou, E., & Felici, G. (2006). Data mining and knowledge discovery
approaches based on rule induction techniques. Springer Science-i-Business
Media. 48.
Wang, L., & Fu, X. (2005). Data mining with computational intelligence (pp. 229).
Berlin, Heidelberg: Springer-Verlag.
Zarandi, M. H. F., Rezaee, B., Turksen, I. B., & Neshat, E. (2007). A Type-2 fuzzy rule-
based expert system model for stock price analysis. Expert Systems with
Applications, 9–34. doi:10.1016/j.eswa.2007.09.034.