Journal of Petroleum Science and Engineering 112 (2013) 17–23

Contents lists available at ScienceDirect

Journal of Petroleum Science and Engineering

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

A comparison study of using optimization algorithms and artificial

neural networks for predicting permeability
Hossein Kaydani, Ali Mohebbi n
Department of Chemical Engineering, College of Engineering, Shahid Bahonar University of Kerman, Kerman, Iran

art ic l e i nf o a b s t r a c t

Article history: This paper presents a novel approach of permeability prediction by combining cuckoo, particle swarm
Received 22 May 2012 and imperialist competitive algorithms with Levenberg–Marquardt (LM) neural network algorithm in
Accepted 1 November 2013 one of heterogeneous oil reservoirs in Iran. First, topology and parameters of the Artificial Neural
Available online 21 November 2013
Network (ANN) as decision variables were designed without the optimization method. Then, in order to
Keywords: improve the effectiveness of forecasting when ANN was applied to a permeability predicting problem,
permeability the design was performed using Cuckoo Optimization Algorithm (COA) algorithm. The validation test
neural network result from a new well data demonstrated that the trained COA–LM neural model can efficiently
cuckoo optimization algorithm accomplish permeability prediction. Also, the comparison of COA with particle swarm optimization and
particle swarm optimization
imperialist competitive algorithms showed the superiority of COA on fast convergence and best optimum
imperialist competitive algorithm
solution achievement.
well logs data
& 2013 Elsevier B.V. All rights reserved.

1. Introduction
Permeability prediction is one of the main challenges in
reservoir engineering. On the other hand, suggesting engineering
methods to solve reservoir modeling and managements is impos-
sible without the knowledge of the actual permeability values
(Biswas et al., 2003). The reservoir rock permeability cannot be
measured directly with exception of using core plugs as direct
measurement. However, direct measuring methods are expensive
and time-consuming (Mohaghegh et al., 1997). In recent years
intelligent techniques such as Artificial Neural Networks (ANNs)
have been increasingly applied to predict reservoir properties
using well log data (Mohaghegh et al., 1994). Moreover, previous
investigations have indicated that neural networks can predict
formation permeability even in highly heterogeneous reservoirs
using geophysical well log data with good accuracy (Mohaghegh
et al., 1996; Aminian et al., 2000).
In spite of the wide range of applications, a significant amount of
time and effort has been expended to find the optimum or near
optimum structure for a neural network for the desired task. To
mitigate these deficiencies, design of neural networks using optimi-
zation algorithms such as Genetic Algorithm (GA) and Particle Swarm
Optimization (PSO) have been proposed (Boozarjomehry and Svrcek,
2001; Wang, 2005; Zhang et al., 2007; Kaydani et al., 2011).
Recently, a new evolutionary algorithm has been proposed by
Rajabioun (2011), which was inspired by special lifestyle of cuckoo
n
Corresponding author.
E-mail addresses: amohebbi2002@yahoo.com, amohebbi@uk.ac.ir (A. Mohebbi).
birds and was called Cuckoo Optimization Algorithm (COA). COA
mimics the breeding behavior of cuckoos, where each individual
searches for the most suitable nest to lay an egg in order to
maximize the egg's survival rate, which is an efficient search pattern.
Application of the COA in different optimization problems has
proven its capability to deal with difficult optimization problems,
especially multi-dimensional problems (Rajabioun, 2011).
In this study, the COA optimization and the LM algorithm were
combined in order to form a two-stage learning algorithm for
optimizing the parameters of the feed forward neural network.
The proposed method was applied in permeability prediction in
one of the heterogeneous oil reservoirs in Iran. Then, the results of
the hybrid model were compared with experimental measure-
ments from a new well data to find the efficiency of the present
algorithm. Moreover, the optimal network design was done with
particle swarm optimization and Imperialist Competitive Algo-
rithm (ICA) and their results were compared with cuckoo optimi-
zation algorithm on designing neural network structure with LM
training algorithm.
2. Artificial neural networks
Artificial neural network (ANN) is an especially efficient algo-
rithm to approximate any function with finite number of disconti-
nuities by learning the relationships between input and output
vectors (Ganguly, 2003; Richon and Laugier, 2003). Feed forward
type neural networks have an input, an output and, in most
applications, have one hidden layer. The number of inputs and
outputs of the neural networks are determined by considering the

0920-4105/$ - see front matter & 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.petrol.2013.11.009
18 H. Kaydani, A. Mohebbi / Journal of Petroleum Science and Engineering 112 (2013) 17–23

characteristics of the application. Each neuron of a layer is
generally connected to the neurons in the proceeding layer. A
neuron has two components: (1) a weighted sum, which performs
a weighted summation of its inputs with components (X1, X2, X3,
…, Xn), i.e., S ¼∑WiXi þb, where b is the bias of the networks; and
(2) a linear, nonlinear or logic function, which gives an output
corresponding to S. In general, the output from neural j in layer k
can be calculated by the following equation:
! Prof it ¼  costðhabitatÞ ¼  f c ðX 1 ; X 2 ; …; X Nvar Þ
Nk  1
ujk ¼ F k ∑ wijk uiðk  1Þ þ bjk Þ
i¼1
Coefficients wijk and bjk are the connection weight and bias of
the network, respectively; they are the fitting parameters of the
model and Fk is the transfer function of layer k. The output of the
network is created at the output layer. Then the network's output
pattern is compared with the target vector according to Mean
Squared Error (MSE). The MSE of the neural network can be
defined as follows:
∑ni¼ 1 ðOi  T i Þ2
MSE ¼
n
where Oi is the desired output, Ti is the network output, and n is
the number of data in the training data set or the cross validation
data set (Demuth et al., 2006).
For neural networks based on supervised learning, the data are
usually split into a training set, validation set, and a test set.
Various networks are trained by minimization of an appropriate
error function defined with respect to the training set. Perfor-
mance of the networks is then compared by evaluating the error
function using the training and validation set, and the network
having the smallest error with respect to the validation set is
selected (Niculescu, 2003).
The profit of a habitat is obtained by evaluation of profit
function fp at a habitat, so
prof it ¼ f p ðhabitatÞ ¼ f p ðX 1 ; X 2 ; …; X Nvar Þ ð4Þ
from this point of view, the entire algorithm searches to maximize
a profit function. To use COA in cost minimization problems, one
can easily maximize the following profit function:
ð5Þ
ð1Þ
By nature, each cuckoo lays from 5 to 20 eggs. These values are
used as the upper and lower limits of egg dedication to each
cuckoo at different iterations. Another habit of real cuckoos is that
they lay eggs within a maximum distance from their habitat,
which is called Egg Laying Radius (ELR) and defined as (Rajabioun,
2011)
Number of current cuckoo0 s eggs
ELR ¼ α  ðvar hi  var low Þ ð6Þ
Total number of eggs
where varhi and varlow are the upper limit and lower limit of
ð2Þ variables respectively, and α is an integer, supposed to handle the
maximum value of ELR.
Each cuckoo starts laying eggs randomly in some other host
birds' nests within her ELR. Some of these eggs, which are more
similar to the host bird's eggs have the opportunity to grow up and
become a mature cuckoo. Other eggs are detected by host birds
and are killed. When young cuckoos grow and become mature,
they immigrate to new and better habitats. After the cuckoo
groups are formed in different areas with the K-means clustering
method, the society with the best profit value is selected as the
goal point for other cuckoos to immigrate. Fig. 1 shows the
movement of cuckoos towards the destination habitat. In this
movement, λ and φ are the random numbers, which are defined by
(Rajabioun, 2011)

3. Optimization algorithms λ  Uð0; 1Þ ð7Þ

In recent years, Evolutionary Computation (EC) and metaheur- φ  Uð  ω; ωÞ ð8Þ

istis optimization algorithms have been a noticeable part for
where ω is a parameter that constrains the deviation from goal
solving real mathematics and engineering problems. In the next
habitat and a value of about π/6 (Radian) for ω results in a good
sections, we describe some of optimization algorithms, which
convergence of countries to the global minimum.
were used in this study.
There is always equilibrium in birds' population so a number
Nmax controls and limits the maximum number of live cuckoos in
3.1. Cuckoo optimization algorithm
the environment, which can be the result of food limitations, and
other environmental conditions. After some iterations, all the
Cuckoo Optimization Algorithm (COA) is a new evolutionary
cuckoo population move to one best habitat with maximum
optimization method, which is inspired by the life of a bird family,
similarity of eggs to the host birds and also with the maximum
called Cuckoo. Special lifestyle of these birds and their character-
food resources. This habitat will produce the maximum profit ever.
istics in egg laying and breeding has been the basic procedure of
There will be least egg losses in this best habitat. Convergence of
this evolutionary optimization algorithm. Like other evolutionary
more than 95% of all cuckoos to the same habitat puts an end to
algorithms, it starts with an initial population including mature
cuckoos and their eggs. The effort to survive among cuckoos
constitutes the basis of COA. During the survival competition
some of the cuckoos or their eggs, are destroyed. The survived
cuckoo societies immigrate to a better environment and start
reproducing and laying eggs. Cuckoos' survival effort hopefully
converges to a state in which there is only one cuckoo society, all
with the same profit values.
Each variable in this algorithm can be interpreted as a habitat,
which is equivalent to chromosome in GA or particle in PSO
algorithm. In N variable optimization problem, the habitat array
is defined as follows, which represents the current position of
variables (Rajabioun, 2011):

habitat ¼ ½X 1 ; X 2 ; …; X Nvar  ð3Þ Fig. 1. Immigration of a sample cuckoo toward goal habitat (Rajabioun, 2011).
 H. Kaydani, A. Mohebbi / Journal of Petroleum Science and Engineering 112 (2013) 17–23 19

Cuckoo Optimization Algorithm. More details of COA are available

in literature (Rajabioun, 2011).

3.2. Imperialist competitive algorithm

Imperialist Competitive Algorithm (ICA) is an evolutionary

optimization method, which is inspired by imperialistic competi-
tion. Like other evolutionary algorithms, it starts with an initial
population, which is called country and is divided into two types:
the colonies and the imperialists, which together form empires. All
the colonies of initial countries are divided among the mentioned
imperialists based on their power. The power of each country, the
counterpart of fitness value in the GA, is inversely proportional to
its cost. Imperialistic competition among these empires forms the
proposed evolutionary algorithm (Atashpaz-Gargari and Lucas,
2007).
Each variable in the country can be interpreted as a socio-
political characteristic of a country. From this point of view, the
entire algorithm searches for the best country that has the best
combination of socio-political characteristics. From optimization
point of view this leads to finding the optimal solution of the
problem, the solution with the least cost value. More details of ICA
are available in literature (Atashpaz-Gargari and Lucas, 2007).
Fig. 2. Locations of wells in mentioned field.

3.3. Particle swarm optimization

Particle Swarm Optimization (PSO) is an evolutionary compu-

tation technique with the mechanism of individual improvement,
population cooperation and competition, which is based on the
simulation of simplified social models, such as bird flocking, fish
schooling and the swarming theory (Kennedy and Eberhart, 1995).
PSO starts with the random initialization of a population (swarm)
of individuals (particles) in the search space and works on the
social behavior of the particles in the swarm. Therefore, it finds the
global best solution by simply adjusting the trajectory of each
individual towards its own best location and towards the best
particle of the swarm at each time step (i.e. generation). However,
the trajectory of each individual in the search space is adjusted by
dynamically altering the velocity of each particle, according to its
own flying experience and the flying experience of the other
particles in the search space. The process is repeated until a user
defined stopping criterion is reached. More details of PSO are
available in literature (Kennedy and Eberhart, 1995).
Fig. 3. Feed-forward multilayer ANN scheme with network inputs data.

4. Case study 4.2. Data analysis

The prototype oil reservoir is located at the southwest of Iran.
With the data acquired (NISOC, 2006), the reservoir can be
classified as a structural complex and heterogonous reservoir.
4.1. Reservoir history
All the 68 wells of the mentioned reservoir that were drilled
until 2012 have been logged. Therefore, the petrophysical data of
almost all the wells were available. In this field only six wells were
cored in the reservoir layer (see Fig. 2). It should be noted that core
data are often only available from few wells in a reservoir while
the well logs are available from the majority of the wells. Thus,
evaluation of permeability from the well log data represents a
significant technical as well as economical advantage (NISOC,
2006).
In this study, six wells with suitable distances from each other
are selected, where all of the essential information including well
logs and core permeability are available. Among the available
petrophysical data, the inputs to the network are Spectral Gamma
Ray (SGR), Water Saturation (SWT), Total Porosity (PHIT), Bulk
Density (RHOB), Sonic porosity (DT), Neutron Porosity (NPHI) and
depth, while the output is permeability values from the core
analysis. SGR is a valid indicator for shale regions and SWT shows
a permeable region. Therefore, we include them in inputs vector.
Also, porosity logs, namely PHIT, RHOB, DT and NPHI were selected
as inputs according to a previous study (Biswas et al., 2003) that
observed the correlation between porosity and permeability.
Furthermore, because of considering the effect of the vertical
indicator and overburden pressure on permeability, the depth of
each of the cores was selected. In Fig. 3 inputs and output of the
network are shown graphically.
20 H. Kaydani, A. Mohebbi / Journal of Petroleum Science and Engineering 112 (2013) 17–23

Table 1
Correlation coefficients between parameters before and after normalizing.

Depth DT-FL NPHI-COR RHOB-COR SGR-COR PHIT (%) SWT

Permeability 0.043 0.43 0.388  0.393  0.154 0.402  0.291

N.Depth N.DT NPHI-COR N.RHOB N.SGR N.PHIT N.SWT
Normalized permeability 0.129 0.497 0.461  0.449  0.181 0.525  0.274

N: Normalized.

In most cases, data should be scaled, normally linearly, to a Table 2

smaller interval in order to increase the training speed, facilitate
modeling and predicting. In this work, each group of training data
including logs and core measurements are normalized to [0.2, 1] as
follows:
X  X min
X n ¼ 0:8  þ 0:2
X max  X min
where, X is the original data; Xmin is the minimum of X; Xmax is the
maximum of X; and Xn is the result of normalization. The
correlation coefficient between the dependent and independent
parameters can be improved in normalized distribution data.
Table 1 compares the correlation coefficient between normalized
and non-normalized data.
The basis for well logging data is the mean sea level and
measuring of depth is according to the length of wire driven into
the well. In coring operation, the depth basis for extracted cores is
the rotary table elevation of drilling rig and depth measuring is
according to the length of drilling strings driven into the well.
Therefore, to use well logging data and comparing with laboratory
core data, depth matching operation appears to be necessary. The
details of depth matching are given elsewhere (Kaydani et al.,
2011).
5. ANN designing by using COA algorithms
The LM feed-forward neural network is one of the main
algorithms for understanding neural networks, which has some
drawbacks in petroleum engineering applications (Huang et al.,
2003). To address some of these drawbacks, the Cuckoo Optimiza-
tion Algorithm (COA) and the Levenberg–Marquardt (LM) algo-
rithm are combined in order to form a hybrid learning algorithm
with the advantage of both algorithms to optimize the weights
and threshold of the feed forward neural network. First, the COA is
employed to search in the solution space and find near optimal
solutions. Then, the LM training, which can provide the second
order of convergence rate, is applied according to these near
optimal weights to find the best solution around the global
optimum.
When the COA evolves the weights and thresholds of the feed
forward neural network, every habitat represents a set of weights
and thresholds. In the encoding strategy, every habitat is encoded
as an array to represent all weights and thresholds of the neural
network's structure.
The goal in this algorithm is to minimize a cost function. The
Mean Square Error (MSE) was used as a cost function in ICA
optimization, which can be calculated according to Eq. (2). The
searching mechanism of the implemented hybrid strategy can be
succinctly described as follows:
Step 1: initialize a feed forward neural network structure;
initialize the parameters in the COA.
Step 2: set up encoding relationship according to Eq. (4)
between the neural network structure and the COA parameters.
Step 3: initialize cuckoo habitants with some random point on
the profit function (Rajabioun, 2011).
Parameters of COA optimization approach.
Parameter Value
Maximum number of cuckoos 200
at same time
Number of k-means clusters 4
ð9Þ
Maximum number of iteration 300
Minimum number of eggs for 5
each cuckoo
Maximum number of eggs for 10
each cuckoo
Step 4: dedicate some eggs and ELR to each cuckoo.
Step 5: let cuckoos lay eggs inside their corresponding ELR.
Step 6: kill some eggs by host birds; lets other eggs hatch and
chicks grow.
Step 7: evaluate the habitant of each newly grown cuckoo.
Step 8: limit cuckoos maximum number in environment and
kill those who live in worst habitats.
Step 9: cluster cuckoos and find best group and select goal
habitat; immigrate new cuckoo population toward this point.
Step 10: if stop conditions satisfied, go to step 11; if not go to 3
(Rajabioun, 2011).
Step 11: decode near optimal solution to neural networks
structure.
Step 12: train the network by LM up to desired MSE or up to
maximum number of iterations.
Step13: display the performance by presenting testing para-
meters.
In order to optimize the network structure, the parameters of
the network, such as number of neurons in the hidden layers,
weights and biases of layers, are treated as variables, which must
be optimized with COA. Here, the numbers of neurons in the
hidden layer are allowed to take only integer values, which
bounded in a range between a lower bound of L/2, where L is
the number of input variables, and an upper bound of 3L
(NeuroDimension Software Inc., 2005). The other parameters,
weights and biases are coded and allowed to take real values,
which had a lower bound of 2 and an upper bound of þ2. The
COA parameters are initially specified, as given in Table 2. For
implementation of hybrid algorithms for permeability prediction,
a computer program has been written under MATLAB R2009
software, which had been produced by MathWorks, Inc.
6. Results and discussion
Well logs are inputs to the network and core permeability
values are the outputs. The optimal number of the neurons of a
hidden layer network was obtained first by using the trial and
error method and then by optimization algorithms with LM
training algorithm, which form a two-stage learning algorithm.
All the 1260 available data points were organized from all the
selected wells as input and output. However, the data from
one well was left as new well data set from the ANN training
 H. Kaydani, A. Mohebbi / Journal of Petroleum Science and Engineering 112 (2013) 17–23 21

0.025 60

Training
0.02
Average of Min MSEs

Cross validation
50
0.015 R=0.99

1 mD= 9.869 E-16 m2

Predicted permeability (mD)

0.01
40

0.005

30
0
5 7 9 11 13 15 17
Neurons number in hidden layer
20
Fig. 4. Performance of ANNs with different neurons in their middle layer.

10
Table 3
Performance of ANN for test data set trained with-
out optimization methods.
0
Parameter Best network 0 10 20 30 40 50
Core permeability (mD)
AAD 0.155
NMSE 0.026 Fig. 5. Hybrid model predictions versus core measurements permeability for
R 0.831 training data set.
R2 0.691

procedure in order to indicate capability prediction of the best Table 4

network and comparison between optimization methods. Performance of COA–LM neural model for new
data set.

6.1. Training without optimization Parameter Best network

Well log responses were as inputs to the network, while
permeability values of core data were the outputs. The database
to be introduced to the neural network training were broken down
randomly into three groups: training, cross-validation, and testing.
Typically, 80% of the data is used for the training process and the
other 20% of the data is categorized as validation and testing. The
number of hidden layers and the number of input and output
neurons in the ANN model were defined as 1, 7 and 1, respectively.
Fig. 4 shows the optimal number of the neurons of a single hidden
layer network using this method. The training results for the ANN
on the cross-validation data showed the lowest MSE when the
number of hidden neurons was 11. It was found that the 7–11–1
architecture was the best model in terms of MSE, which means 7,
11, and 1 neurons in the input, hidden and output layers respec-
tively. More than a single hidden layer ANNs were also used
throughout this stage, but the result was unsatisfactory.
Table 3 gives the efficient ANNs performances in terms of
Normalized Mean-Square Error (NMSE) (Wei et al., 2010), average
absolute deviation (AAD), the linear correlation coefficient (R) and
squared linear correlation coefficient (R2) between the core per-
meability and neural network output in new well data set. In brief,
the ANN predictions are optimum if R2, R, AAD and NMSE are
found to be close to 1, 1, 0 and 0 respectively.
6.2. COA optimization method
The number of hidden layers and the number of input and
output neurons in the COA–LM model were defined as 1, 7, and 1,
respectively. The Mean Square Error (MSE) was used as a criterion
for choosing the best design in the COA–LM optimization model,
which can be calculated according to Eq. (2). It was found that the
best ANN model of the COA–LM was obtained in a structure with
the 7–11–1 architecture in terms of the least cost function
(minimum MSE), which means 7, 11, and 1 neurons in the input,
hidden and output layers respectively. Fig. 5 shows the hybrid
AAD 0.099
NMSE 0.012
R 0.978
R2 0.957
model predictions versus core measurements permeability for a
tainting data set. The ANNs with more than one single hidden
layer were also used, but the result was not improved.
In this case, to evaluate the hybrid model generalization on the
chosen data set or new well data points, the performance of the
model is reported in Table 4. This table reveals agreement between
the predicted values of permeability by the COA–LM optimization
model and the permeability of the actual core measurements.
Fig. 6 compares the actual permeability values, which were
measured in laboratory (and never seen by the network during
the optimized training) with the network's prediction for each
sample. This figure shows that there is an acceptable agreement (i.
e. linear correlation coefficient of 98%) between the predicted
values and the experimental data.
Comparison of the results presented in Tables 3 and 4 reveals
that the correlation coefficient between core permeability and
output of ANN without COA is 0.831 while this parameter between
core permeability and the COA–LM model output is 0.978. There-
fore, the hybrid algorithm has higher accuracy than the design
without COA. To achieve the optimal solution in the first model (i.
e. ANN without COA), the LM training algorithms must be run
several times for every fixed hidden neurons number, but in the
COA–LM training model this solution is achieved by searching
ability of COA algorithm in all of the space solutions.
6.3. Comparison of COA with PSO and ICA
In order to do a comparison among the algorithms, COA, PSO and
ICA were applied to the optimum design of neural networks structure.
Fig. 7 shows the proposed flowchart of the optimized LM neural
22 H. Kaydani, A. Mohebbi / Journal of Petroleum Science and Engineering 112 (2013) 17–23

50 0.031
PSO

ICA
R=0.98 0.026
40
COA
1 mD= 9.869 E-16 m2
Prediction permeability (md)

0.021

Cost function (MSE)

30

0.016
20

0.011

10

0.006

0
0 15 30 45 0.001
Core permeability (md) 0 50 100 150 200
Iterations
Fig. 6. The hybrid model predictions versus core measurements permeability for
the new well data set. Fig. 8. Cost function minimization plot for three optimization algorithms in neural
network design.

START
Table 5
Performance of neural network optimum model by different methods.
Initialize a feed-forward neural network and optimization algorithm
Parameter COA–LM ICA–LM PSO–LM

AAD 0.099 0.104 0.119

Set up encoding and decoding relationship between the NMSE 0. 012 0.017 0.020
neural network and the optimization algorithm R 0.978 0.961 0.943
R2 0.957 0.924 0.889

Implementation of the optimization algorithm (PSO or ICA or COA)

Export near critical weights and biases to neural network
Train the network by LM up to desired MSE or up to
maximum number of iterations
Display the performance by presenting testing parameters
END
Fig. 7. Flowchart of the hybrid an optimization – LM neural network.
network modeling with optimization algorithms for permeability
prediction. For all three methods, the maximum number of iterations
is set to 200. Fig. 8 illustrates a sample cost minimization plot of MSE
function of all three algorithms. This figure shows that the COA
converges to a desirable tolerance (i.e. MSE¼ 0.001) by about 150
iterations, whereas ICA and PSO cannot converge to this stopping
criteria within 200 iterations. In a multi-dimensional problem like
optimization of neural network structure, the optimization function
has lots of local minima and is one of the difficult problems to solve. In
this type of problem, the final results are more important since they
reflect the ability of the algorithm in escaping from local optima and
achieving a near-global optimum. As one can see clearly from Fig. 8,
the COA has a faster convergence than the ICA and PSO. In other
words, the amount of computation (i.e. iterations) required to achieve
the global minimum for the ICA and PSO is much greater than that for
the COA. The progress of COA algorithm toward the near-global
minimum indicates the superiority of COA algorithm especially in
multi-dimension of problems over the other well-known heuristic
search methods. Table 5 shows R2, R, AAD and NMSE for all three
models. From this table, it can be seen that the COA–LM gives the best
results compared with the other methods. Therefore, the good
estimation of global minimum solution can be result from the
optimum LM neural network modeling.
7. Conclusions
In this study, an attempt was made to develop a methodology
for combining COA, ICA and PSO with LM feed forward neural
network to predict permeability for a heterogeneous oil reservoir.
The COA was used in order to minimize the effort required to find
the optimal architecture and parameters of the feed forward ANNs.
The performance of the network with respect to the predictions
made on the test sets showed that this new approach was able to
sufficiently estimate the permeability of the reservoir with high
correlation coefficient. Comparison of the prediction efficiency of
the COA optimized ANN model with that of the ANN model without
optimization showed that the COA–LM neural model produces a
higher accuracy than without the optimization method. Also,
comparison of COA with PSO and ICA showed the superiority of
COA in fast convergence and global optima achievement. Therefore,
the COA–LM neural network model is able to predict rock perme-
ability with sufficient and reliable geophysical well log data very
close to laboratory core permeability measurements.
Acknowledgments
The authors would like to gratefully thank National Iranian oil
company (NIOC) and National Iranian South Oilfields Company
(NISOC) for their help and financial support.
 H. Kaydani, A. Mohebbi / Journal of Petroleum Science and Engineering 112 (2013) 17–23
References
Aminian, K., Bilgesu, H.I., Ameri, S., Gil, E., 2000. Improving the Simulation of Water
Flood Performance with the Use of Neural Networks. SPE Paper 65630.
Atashpaz-Gargari, E., Lucas, C., 2007. Imperialist competitive algorithm: An algo-
rithm for optimization inspired by imperialistic competition. In: Proceedings of
the IEEE Congress on Evolutionary Computation, pp. 4661–4667.
Biswas, D., Suryanarayana, P.V., Frink, P.J., Rahman, S., 2003. An Improved Model to
Predict Reservoir Characteristics During Underbalanced Drilling. SPE Paper
84176.
Boozarjomehry, R.B., Svrcek, W.Y., 2001. Automatic design of neural network
structures. Comput. Chem. Eng. 25, 1075–1088.
Demuth, H., Beale, M., Hagan, M., 2006. Neural Network Toolbox User's Guide. the
Math Works Inc.
Ganguly, S., 2003. Prediction of VLE data using radial basis function network.
J. Comput. Chem. Eng. 27, 1445–1454.
Huang, Y.F., Huang, G.H., Dong, M.Z., Feng, G.M., 2003. Development of an artificial
neural network model for predicting minimum miscibility pressure in CO2
flooding. J. Pet. Sci. Eng. 37, 83–95.
Kaydani, H., Mohebbi, A., Baghaie, A., 2011. Permeability prediction based on
reservoir zonation by a hybrid neural genetic algorithm in one of the Iranian
heterogeneous oil reservoirs. J. Pet. Sci. Eng. 78, 497–504.
Kennedy, J., Eberhart, R.C., 1995. Particle swarm optimization. In: Proceedings of the
IEEE International Conference on Neural Networks in 1995, vol. 4, pp. 1942–
1948.
23
Mohaghegh, S., Arefi, R., Ameri, S., Rose, D., 1994. Design and Development of an
Artificial Neural Network for Estimation of Formation Permeability. SPE Paper
28237.
Mohaghegh, S., Ameri, S., Aminian, K., 1996. A methodological approach for
reservoir heterogeneity characterization using artificial neural networks.
J. Pet. Sci. Eng. 16, 263–274.
Mohaghegh, S., Balan, B., Ameri, S., 1997. Permeability Determination from Well Log
Data. SPE Paper 30978.
Niculescu, S.P., 2003. Artificial neural networks and genetic algorithms in QSAR.
J. Mol. Struct. Theochem. 622, 71–83.
NISOC Report, 2006. Geological Studies Report for Mansuri Oil Field Development
in Bangestan Formation.
NeuroSolutions for Excel Release 5 software help, Copyright 2005, NeuroDimen-
sion, Inc.
Rajabioun, R., 2011. Cuckoo optimization algorithm. Appl. Soft Comput. 11,
5508–5518.
Richon, D., Laugier, S., 2003. Use of artificial neural networks for calculating derived
thermodynamic quantities from volumetric property data. Fluid Phase Equilib.
210, 247–255.
Wang, L., 2005. A hybrid genetic algorithm–neural network strategy for simulation
optimization. J. Appl. Math. Comput. 170, 1329–1343.
Wei, H.L., Billings, S.A., Zhao, Y.F., Guo, L.Z., 2010. An adaptive wavelet neural network for
spatio-temporal system identification. Neural Netw. 23, 1286–1299.
Zhang, J.R., Zhang, J., Lok, T.M., Lyu, M., 2007. A hybrid particle swarm optimization
back-propagation algorithm for feed forward neural network training. J. Appl.
Math. Comput. 185, 1026–1037.