Journal of Petroleum Science and Engineering 121 (2014) 87–102

Contents lists available at ScienceDirect

Journal of Petroleum Science and Engineering

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

A hybrid approach for litho-facies characterization in the framework

of sequence stratigraphy: A case study from the South Pars gas field,
the Persian Gulf basin
Ebrahim Sfidari a, Ali Kadkhodaie-Ilkhchi b,n, Hossain Rahimpour-Bbonab a, Behzad Soltani a
a
Department of Geology, College of Science, University of Tehran, Tehran, Iran
b
Department of Geology, Faculty of Natural Science, University of Tabriz, Tabriz, Iran

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

Article history: The Upper Dalan and Kangan formations with dominant lithology of limestone and dolomite associated
Received 26 July 2013 with anhydrite nodules and interbeds form the Permo-Triassic succession of South Pars gas field (SPGF)
Accepted 15 June 2014 and host the largest none-associated gas reservoir in the world. The current study focuses on preparing a
Available online 7 July 2014
comprehensive litho-facies model in the framework of sequence stratigraphy. For this purpose, Self-
Keywords: Organizing Map Neural Network (SOM-ANN) and hierarchical cluster analysis (HCA) were utilized as
litho-facies effective tools to prepare the preliminary data for litho-facies mapping. Neural networks (self-organizing
electro-facies maps) and hierarchical clustering approaches were applied to characterize litho-facies in un-cored
self-organizing maps but logged wells. Particularly, the powerful visualization tools of the SOM-ANN which provide more
hierarchical cluster analysis
information in comparison to HCA facilitate the task of establishing an order of priority between the
South Pars gas field
distinguished electro-facies groups. The mentioned method of SOM-ANN clustering algorithm showed a
good performance in petrophysical data clustering and litho-facies determination. Based on the porosity
and permeability maps at different depth levels, the target reservoir is ranked and classified into four
litho-facies and six electro-facies. They include litho-facies 3 with good reservoir quality (equivalent
of electro-facies 4–6), litho-facies 4 with moderate reservoir quality (equivalent of electro-facies 2) and
litho-facies 1 and 2 with poor to bad reservoir quality (equivalent of electro-facies 1 and 3).
The main litho-facies assemblages are indicative of deposition within tidal flat, lagoon, shoal and off-
shoal environments. The most shoal litho-facies with best reservoir quality occurs in the high energy
sub-environment within upper transgression position (HST) of the 3rd-order cycle in K4 and K2 reservoir
units. Distribution of the petrophysical characteristics was analyzed in detail in the framework of
electro-facies and sequence stratigraphy. The methodology is illustrated by using a case study from
SPGF, Iran.
& 2014 Elsevier B.V. All rights reserved.

1. Introduction With the development of computer technology, many algo-

Since flow properties are usually assigned using facies-specific
correlations, electro-facies and litho-facies classifications can support
reservoir characterization. In un-cored wells, sedimentary facies
are interpreted from petrophysical data, because they cannot be
observed directly. Some of the well logs, such as gamma spectrum,
density and neutron logs and also spontaneous potential logs, which
can show lithology variations, are known as fine indicators of facies.
n
Corresponding author. Tel./fax: þ 98 411 335 6027.
E-mail addresses: ebispidari@gmail.com (E. Sfidari),
kadkhodaie_ali@tabrizu.ac.ir (A. Kadkhodaie-Ilkhchi),
rahimpor@khayam.ut.ac.ir (H. Rahimpour-Bbonab),
b.soltani67@gmail.com (B. Soltani).
rithms have been used for log-facies determination. Two common
methods are artificial intelligence (e.g. artificial neural network
and fuzzy logic) and multivariate statistic (discriminant analysis,
cluster analysis, regression analysis and Bayesian analysis). A
review through literature shows increasing rate of application of
intelligent methods in reservoir characterization in recent decade.
Neural network clustering based on SOM has recently been
used in many applications such as image processing (Ji, 2000),
precipitation estimations (Sorooshian et al., 2000; Hong et al.,
2005, 2006), wind circulation (Cassano et al., 2006), seismic
interpretation (Strecker and Uden, 2002; Cole'ou et al., 2003),
total organic carbon estimation and litho-facies clustering (Sfidari
et al., 2012a). Also, many applications of hierarchical cluster
analysis in the litho-facies identification have been published by
Perez et al. (2003), Sfidari et al. (2012a) and Ma et al. (2014). There

http://dx.doi.org/10.1016/j.petrol.2014.06.013
0920-4105/& 2014 Elsevier B.V. All rights reserved.
88 E. Sfidari et al. / Journal of Petroleum Science and Engineering 121 (2014) 87–102

are various applications of artificial neural networks in facies
classification (Wolff and Pelissier-Combescure, 1982; Rogers et al.,
1992; Karpur et al., 2000; Saggaf and Nebrija, 2000; Tang and
Ji, 2006; Qi and Carr, 2006a, 2006b). Unfortunately, core samples
are restricted due to long time and expenses associated with a
coring program (Hong and Christopher, 2008). Therefore, is a time
consuming and expensive way to core study for most of the wells
in a hydrocarbon field. In this paper, an applied method based
upon unsupervised neural networks is discussed for determina-
tion of litho-facies and electro-facies and the result is compared
with HCA. Finally, electro-facies are analyzed and discussed in the
framework of sequence stratigraphy.
Bahrain) and Rub-Al-Khali in the U.A.E. (Alshahran and Nairn,
1997; Konert et al., 2001; Ziegler, 2001).
The Permian-Triassic successions of Upper Dalan and Kangan
formations are divided into K1–K4 reservoir units according to
their geological and reservoir characteristics. In the Iranian side,
the Kangan formation with Triassic age is divided into K1 and K2
zones, and the Upper Permian Dalan formation is divided into K3
and K4 zones. The mentioned reservoir layers are capped by

2. Geological setting

The Upper Dalan and Kangan formations (equivalent of Khuff in

the Arabian nomenclature) host part of the world's largest gas
reservoir in the Middle East region. This field along with its
southern extension, Qatari Dome, regionally called North Dome
field, is the largest offshore gas field in the world. The Qatar arch is
a north–northeast trending positive tectonic feature, extending
into Persian Gulf and effectively dividing it into two basins. The
eastern side of the Arch is a gently dipping monocline and the
western side is bounded by Kazerun fault as mentioned and
mapped by several authors (e.g. Alshahran and Nairn, 1997;
Konert et al., 2001; Ziegler, 2001). SPGF lies on the north–north-
east plunge of this Arch. The Qatar structural Arch has had a
fundamental influence on the tectonic pattern of the Persian Gulf
and separates its trough in the central part of the Persian Gulf
basin (Saudi Arabia, northern branch of the Persian Gulf and Fig. 2. Work flow illustrating major steps of integrated approach used in this study.

Fig. 1. Location map of South Pars gas field in the Persian Gulf (right) and stratigraphic position of the studied formations (left).
 E. Sfidari et al. / Journal of Petroleum Science and Engineering 121 (2014) 87–102 89

Fig. 3. Main litho-facies assemblages indicative of depositional setting of SPGF.

Fig. 5. Correlation matrix containing Pearson's correlation coefficients for each pair
of variables. Light colors represent strong positive linear correlations, whereas dark
colors document negative correlations. Orange suggests no correlation. (For
interpretation of the references to color in this figure legend, the reader is referred
to the web version of this article.)

Table 1
Cophenetic correlation coefficient for various distance and linkage functions.

Distance function Linkage function

Average Centroid Complete Median Single Ward

Correlation 0.8575 0.8567 0.7232 0.8474 0.8129 0.7073

City block 0.8408 0.8405 0.7168 0.7194 0.8023 0.7056
Euclidean 0.9307 0.9304 0.9304 0.9291 0.9280 0.9291
Fig. 4. Box and Whisker plots of core porosity (a) and core permeability (b) for
different litho-facies assemblage of Kangan-Dalan formations. Note: The underlined element represents the highest Cophenetic correlation
achieved by "Euclidean" distance function and "average" linkage method.
a massive carbonate-evaporate formation regionally called Dash-
tak (Sudair in Arabian nomenclature). It has been postulated that 3. Methodology
the Silurian shales are the source rocks of the huge Paleozoic oil
and gas reserves found in the Middle East sedimentary basin as The main objective of this study is to develop an integrated
well as Dalan and Kangan formations (Kamali and Rezaee, 2003; geological and computational approach for characterization of litho-
Aali et al., 2006) (Fig. 1). facies and electro-facies in the framework sequence stratigraphy
90 E. Sfidari et al. / Journal of Petroleum Science and Engineering 121 (2014) 87–102

in the Kangan-Dalan successions of SPGF. For this purpose, a four

step approach is proposed as follows. Firstly, macroscopic and
microscopic core studies are carried out with the aim of litho-
facies mapping in cored intervals. Secondly, self-organizing maps
and multivariate statistical methods are examined to characterize
litho-facies in cored wells. Thirdly, core to log calibration is carried
out to propagate litho-facies in un-cored but logged wells. Finally,
reservoir quality of litho-facies is analyzed in the framework of
sequence stratigraphy. The workflow of the procedure used in this
study is shown schematically in Fig. 2.

3.1. Data preparation

Fig. 6. Dendrogram (cluster tree) extracted by hierarchical clustering analysis for
dataset from well A.
The data presented in this analysis came from two cored and
logged key-wells together with 9 logged wells of SPGF. Well log
data were controlled in terms of quality and data from wash out
intervals were not considered in models. The depth matching
is performed to ensure the correct reading of well logs against
corresponding core data (thin-sections, measured porosity and
permeability data). Then, the log data were normalized to unify
the data ranges. Finally, with consideration of data quality and

Fig. 7. Graphical illustrations showing results of HCA on well A, (a) NPHI-RHOB crossplot, (b) DT-RHOB crossplot (d) DT-PEF crossplot, (d) boxplot of NPHI for each individual cluster.
 E. Sfidari et al. / Journal of Petroleum Science and Engineering 121 (2014) 87–102 91

field-wide availability, a suite of well logs including gamma ray
(GR), acoustic transmit-time (DT), neutron (NPHI), density (RHOB)
and photoelectric log (PEF) were selected for electro-facies
analysis.
3.2. Litho-facies analysis
In this study, to reveal the relation between electro-facies and
litho-facies, an integration of core and thin-section studies was
performed. For this purpose, litho-facies analysis was performed
in the framework of main diagenetic overprints and porosity
changes.
Litho-facies were classified and documented based on texture,
grains size and type (such as ooids, peloids, shells and other
diagnostic allochems), lithology, sedimentary structure and other
features, for the Upper Da lan-Kangan reservoirs of SPGF. To assess
their depositional settings, litho-facies were compared with mod-
ern and ancient analogs that are well documented in the literature.
Finally, four main litho-facies assemblages indicative of deposition
within tidal flat, lagoon, shoal and off-shoal environments were
identified (Fig. 3).
3.2.1. Reservoir quality of litho-facies assemblage
Available core porosity and permeability data were used to
examine petrophysical response of the litho-facies assemblage.

Fig. 8. Graphs of input well logs used for cluster analysis and corresponding electro-facies, lithology, porosity and permeability.
92 E. Sfidari et al. / Journal of Petroleum Science and Engineering 121 (2014) 87–102

Analysis of core-measured porosity and permeability of different
litho-facies assemblage shows that the high porosity and perme-
ability values are generally associated with the high energy
facies such as shoal litho-facies assemblage. The shoal litho-
facies assemblage is characterized by high interparticle porosity
and permeability which is represented as the best reservoir quality
(Figs. 3 and 4).
3.3. Cluster analysis
3.3.1. Agglomerative hierarchical clustering method
Hierarchical clustering classifies data over a variety of scales by
creating a cluster tree or dendrogram. The tree is not a single set of
clusters, but rather a multilevel hierarchy, where clusters at one
level are joined as clusters at the next level. This allows us to
decide the level or scale of clustering that is most appropriate for
our application. For this case study, the well log data for electro-
facies determination are chosen from those related to reservoir
lithology and porosity. They include GR, DT, NPHI, RHOB and
PEF logs. It is worth mentioning that in most of the clustering
cases well log data are not independent from each other and
performing a principle component analysis (PCA) or factor analysis
can provide inputs for other clustering methods. A successful
PCA requires linear correlations between variables. However, as
is seen from the correlation matrix plot of Fig. 5, we observed
no dependency between all of the used variables expect GR and
RHOB. Actually, principal component analysis is usually used as an
approach to reduce the dimension of problem with a large number

Fig. 9. Component planes of SOM model for 5 input variables used. Each map corresponds to one variable (component). Hexagons in the same place on different component
planes correspond to the same map unit. Colors indicate value of component in weight vector of each unit from map according to color bars on right. (For interpretation of
the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 10. Graphical illustrations showing (a) well log data hits, (b) U-Matrix, (c) SOM weight positions for dataset from well A.
 E. Sfidari et al. / Journal of Petroleum Science and Engineering 121 (2014) 87–102
of variables. In this study, we employed only 5 variables derived
from well logging data. For this reason, PCA was not necessary to
reduce the dimension of the variables and no linear correlations
are seen between variables.
Following steps were performed to do cluster analysis of well
log data. Firstly, by using the various distance functions the
similarity or dissimilarity was calculated between every pair of
multi-dimensional objects in the original dataset for each well.
The distance function supports many different ways to compute
this measurement. As is seen in Table 1, the Euclidean measure is
more compatible for data of this case study. For a dataset which is
made up of m objects, there are m  (m  1)/2 pairs in the distance
dataset. The result of this computation is commonly known as a
distance or dissimilarity matrix.
Once the proximity between objects in the dataset has been
computed, linking between distance data can determine how
objects in the dataset should be grouping into clusters. In the
next step, the objects with minimum distance were linked to form
new clusters. These newly formed clusters link to each other and
to other objects to create bigger clusters until all the objects in
the original dataset are linked together in a hierarchical tree. As
there are many ways to link one object to others, a Cophenetic
Fig. 11. Graphical illustrations showing results of SOM clustering on well A, (a) NPHI-RHOB crossplot, (b) DT-RHOB crossplot, (c) DT-PEF crossplot, and (d) boxplot of NPHI for
each individual cluster.
93
correlation coefficient (CPCC) was used to determine the appro-
priate clustering method. CPCC is the correlation between the
entries of the linkage matrix and the primary dissimilarity matrix
and also is a standard measure of how well a hierarchical
calculating (of a particular type) fits the data. One of the most
common uses of this measure is to evaluate which type of
hierarchical clustering is optimal for a particular type of data.
According to the results presented in Table 1, the Euclidean
distance measure along side with the Average linkage func-
tion provides a more valid clustering with Cophenetic correlation
of 0.93.
In the next step, the clusters (electro-facies) were generated
based on the cluster tree or dendrogram of data (Fig. 6).
A dendrogram consists of many U-shaped lines connecting objects
in a hierarchical tree. The height of each U represents the distance
between two objects being connected. In the cluster tree, two
objects with minimum distance link together to create a new and
larger cluster. This pattern would continue until finally a single
cluster remains.
The method discussed above was applied for the dataset from
all available wells. Since there is not enough space to illustrate
all results, we will confine our discussion to illustrate the results
94 E. Sfidari et al. / Journal of Petroleum Science and Engineering 121 (2014) 87–102

of clustering on dataset from well A. As shown in the graphical
illustrations of Fig. 7, performing HCA on data from well A
leads to identification of six electro-facies. These electro-facies
are shown in NPHI, RHOB and DT crossplots of Fig. 7. Graphs of
input well logs used for clustering process and the corresponding
electro-facies, lithology, porosity and permeability are displayed
in Fig. 8.
3.3.2. Self-organizing map clustering
In this section, a self-organizing map neural network was
designed for determination of electro-facies from the same set
of well log data used in HCA. The SOM implements an ordered
dimensionality-reducing mapping of the training data. That is, it
provides projection of the multidimensional data into a 2-dimen-
sional map preserving the topology of the input data space.
The SOM model designed in this study is a set of neurons that
are arranged on a regular 2-dimensional hexagonal grid. Each
neuron is symbolized by a weight or primitive vector, which has
as many components as the dimension of the input space (i.e. as
the number of input variables). The Batch training algorithm of
Kohonen (1998, 2001) was used to train the SOM. In the Batch
training, the sample vectors of the input dataset are presented to
the SOM and the new weight vectors are weighted averages of the
data vectors (Vesanto et al., 1999). The iterative process involves
calculating and comparing the Euclidian distances between each
sample vector and all weight vectors of SOM. For each particular
input vector, the neuron whose weight vector is the best match

Fig. 12. Graphs of input well logs used for SOM clustering and corresponding electro-facies, lithology, porosity and permeability.
 E. Sfidari et al. / Journal of Petroleum Science and Engineering 121 (2014) 87–102
(minimum distance) is chosen and called the Best-Matching Unit
(BMU). In each training step, the weight vectors are updated in
such a way that the new weight vectors are weighted averages of
the input data vectors. During this iterative training, the SOM
behaves like a flexible net that folds onto the “cloud” formed by
input data and at the end of training, BMUs of identical data
samples will be close to each other on the final map grid obtained.
For interpreting the results, the SOM visualization process
begins with the map itself that shows what unit is the BMU for
each case of the input data. Samples within the same unit will be
most similar according to the variables considered, while samples
very different from each other are expected to be distant in the
map. There is a weight plane for each element of the input vector
(two, in this case). They are visualizations of the weights that
connect each input to each of the neurons.
The component planes of the SOM model for the 5 input
variables used in this study are displayed in Fig. 9. Each map
corresponds to one variable (component). Hexagons in the same
place on different component planes correspond to the same map
unit. Colors indicate the value of the component in the weight
vector of each unit from the map, according to the color bars on
the right. The darker colors of the SOM architecture represent low
weights, so, we can assume that the inputs were highly correlated
if the connection patterns of two inputs are very similar. As shown
Fig. 13. Box plots showing core porosity and permeability range within each electro-facies extracted from HCA (11a, c) and SOM (11b, d).
95
in Fig. 9, input 1 (NPHI) has connections that are very similar to
input 3 (DT). On the other hand, the connections between input 1
(NPHI) and input 5 (PEF) are very different. As shown in Fig. 10, the
other SOM outputs consist of U-Matrix, SOM data hits and weight
positions. The plot of SOM weight positions illustrates data points
and weight vector locations. After 200 iterations of the Batch
algorithm, the map is well distributed through the input space.
When the input space is high dimensional, visualization of all the
weights at the same time is not possible. In this case, the U-Matrix
is useful to indicate the distances between neighboring neurons.
The colors of the U-Matrix are interpreted as follows.
The blue hexagons represent the neurons, the red lines connect
neighboring neurons, colors in the regions that containing red
lines show distances between neurons, the darker colors represent
larger distances and the lighter colors indicate smaller distances.
The SOM data hits illustrate how many data points are associated
with each neuron. It is best if the data are evenly distributed across
the neurons.
As shown in the graphical illustrations of Fig. 11, performing
the SOM clustering on data from well A leads to identification of
six electro-facies. Fig. 11 illustrates how these electro-facies are
separated on 2D crossplots between NPHI, RHOB, DT and PEF logs.
Comparing plots shown in Fig. 12, using the SOM approach a very
good separation of electro-facies is achieved.
96 E. Sfidari et al. / Journal of Petroleum Science and Engineering 121 (2014) 87–102

Fig. 14. Core to log calibration for key-well A.

 E. Sfidari et al. / Journal of Petroleum Science and Engineering 121 (2014) 87–102 97

Normally, high reservoir quality zones are characterized
by high porosity and permeability. From petrophysical point
of view, high reservoir quality electro-facies correspond
to high quality reservoir beds. They show higher sonic
transit time, lower density, higher neutron porosity and lower
gamma ray. Fig. 12 shows a good agreement between reservoir
electro-facies, lithology, well log responses, core porosity and
permeability.
Box plots of porosity and permeability for each individual
electro-facies extracted from the both clustering approaches are
displayed in Fig. 13. As shown in Fig. 13a and c, centers of clusters
obtained from the HCA are too close to each other, while the SOM
clustering represents a better separation between electro-facies
classes. Accordingly, the results obtained from the SOM model
were utilized to analyze and propagate reservoir electro-facies in
all wells from the studied field.

Fig. 15. Core to log calibration for key-well B.

98 E. Sfidari et al. / Journal of Petroleum Science and Engineering 121 (2014) 87–102
4. Core to log calibration
Because of the non-linear relationship between litho-facies and
wire-line log data, transforming electro-facies to litho-facies or
predicting them from well line logs in un-cored wells is a very
challenging issue. Application of neural network to predicting
litho-facies from conventional logs is documented in numerous
publications (Saggaf and Nebrija, 2003; Qi and Carr, 2006a, 2006b;
Dubois et al., 2007; Al-Anazi and Gates, 2010; Sfidari et al., 2012b).
The approach used in this study is based on a set of well logs
which has to be fed as inputs to the neural network model for each
individual litho-facies assemblage. Nevertheless, extracting litho-
facies in un-cored wells is carried out from well logs with
a successful electro-litho facies calibration. A good agreement
between the litho-facies and electro-facies in the key-wells stu-
died is graphically illustrated in Figs. 14 and 15. The results of
calibration reveal a good correlation between litho-facies 3 with
good reservoir quality and electro-facies 4–6. Litho-facies 4 with
moderate reservoir quality is in connection with electro-facies 2.
Litho-facies 1 and 2 with poor to bad reservoir quality correspond
to electro-facies 1 and 3.
5. Reservoir quality analysis of electro-facies
In general, carbonate reservoirs contain more variable porosity
and permeability than their siliciclastic counterparts that make
their lateral and vertical properties prediction complicated.
The porosity and permeability of carbonate reservoirs are affected
by depositional environment, diagenetic overprints and their
sequence stratigraphic position.
Porosity and permeability crossplot for well A from the studied
field is shown in Fig. 16. The plot shows no distinct relation
between porosity and permeability and data fall into all Lucia rock
fabric classes (Lucia, 1999). This heterogeneity arises from the
severe diagenetic processes acted on the heterogeneous carbonate
rocks of the Kangan and Dalan formations. The SOM method is
adapted to decrease this heterogeneity and non-linearity between
the porosity and permeability data. Lucia porosity and perme-
ability crossplots for electro-facies derived from the SOM method
are shown in Fig. 17. The electro-facies classification process has
Fig. 16. Porosity and permeability crossplot for well A from studied field. Plot
shows no distinct relation between porosity and permeability and data are
scattered in all Lucia rock fabric classes.
improved the complex relationship between porosity and perme-
ability. Accordingly, it is possible to assess the reservoir quality of
each electro-facies in relation to sedimentary facies and Lucia
classification scheme.
On the basis of Lucia classification scheme (Lucia, 1999),
facies 5 with the best reservoir quality falls in the fields of
dolomitized and non-dolomitized grainstones, ooid grainstones
and grain-dominated packstones. Facies 6 with good reservoir
quality comprises wackestones, mud-dominated packstones and
tight grainstones. Facies 4 with good reservoir quality is found
in a wide range of porosity–permeability and contains tight
mudstones to wackestones, anhydrite cemented grainstones and
highly cemented dolomites. Facies 3 with moderate to bad
reservoir quality mainly consists of anhydrite and tight mudstones,
highly cemented dolomites and minor grainstones. Facies 1 and
2 show poor to bad reservoir quality and include anhydrites, tight
wackestone/mudstones and anhydrite cemented grainstones.
6. Electro-facies analysis in the framework of
sequence stratigraphy
In this section, the SOM derived electro-facies were analyzed in
the framework of sequence stratigraphy. In the South pars gas
field, a third order sequence is interpreted based on interpretation
of core, log and seismic data. The most common depositional
environments for carbonates development are ramps, shelves,
reefs and lagoons. Relative sea level change is a key process in
such environments affecting porosity and permeability of the
carbonates. During the highstand level, the land is immersed by
water, while the energy is low on the ramp and medium to high in
the shelf margin, so in the reef sub-environment some biogenic-
dominated facies with high porosity/permeability values are
developed. The facies developed here in this stage of sea level
change (electro-facies 5 and 6, litho-facies 3) fall in classes 1 and
2 of Lucia and are marked by low RHOB, high DT and high NPHI on
well log responses. The facies developed in the sabkha and lagoon
sub-environments during the highstand (electro-facies 1 and 3,
litho-facies 1 and 3) are characterized by low porosity and
permeability along with high RHOB, low NPHI, low DT and high
GR. As seen in Figs. 18 and 19, electro-facies results show a good
agreement with litho-facies and sequence stratigraphy studies.
Average log properties and core porosity and permeability for the
studied electro-facies are listed in Table 2. Eventually, we intro-
duce six electro-facies in connection with sequence stratigraphy
and reservoir quality as follows.
EF1 contains low porosity and permeability values. The main
sedimentary facies include anhydrite, tight wackstone/mud-
stone and anhydrite cemented grainstone. It is frequently
distributed in the lower K4, K3 and K2 zones of the Kangan
and Dalan formations (Figs. 18 and 19). From the sequence
stratigraphic point of view it corresponds to LST and rarely HST.
It is characterized by high RHOB, low NPHI, low DT and
high GR.
EF2 is characterized by the lowest porosity and permeability
values. The main sedimentary facies comprise anhydrite, tight
wackstone/mudstone and anhydrite cemented grainstone. It
shows high RHOB, low NPHI, low DT and high GR on the log
responses.
EF3 has a wide range of porosity and permeability values. The
main sedimentary facies are dolomitized and non-dolomitized
grainstone, ooid grainstone, grain-dominated packstone, tight
doloston and anhydrite. Based on the Lucia rock fabric classi-
fication it falls in classes 1–3 and also in the field of anhydrite
 E. Sfidari et al. / Journal of Petroleum Science and Engineering 121 (2014) 87–102 99

Fig. 17. Lucia porosity and permeability crossplots for electro-facies derived from SOM method.

and tight dolostone facies. It is characterized by moderate
RHOB, DT and NPHI on the log responses.
EF4 includes facies with low to moderate porosity and high
permeability. The main sedimentary facies are dolomitized and
non-dolomitized grainstone, tight doloston with fractures and
dolomite bearing anhydrite cement. It is characterized by low
to moderate NPHI and DT and high to moderate RHOB on the
log responses.
EF5 has the highest porosity and permeability in comparison
to the other electro-facies. It s formed in the highstand
system tract in the stage of sea level rise when reefs move
towards the continent. As can be seen in Fig. 18, it is well
distributed in K4 zone where the best reservoir conditions
for gas accumulations exist. The main sedimentary facies
are dolomitized and non-dolomitized grainstone, ooid grain-
stone and grain-dominated packstones. Based on the Lucia rock
fabric classification it falls in class 1 showing good reservoir
conditions. It is characterized by low RHOB, high DT and
high NPHI.
EF6 includes facies with high porosity and permeability. As with
the EF5, it is formed in the HST when reefs move towards the
continent and in the some cases it shows the lowstand system
tract properties. The main sedimentary facies are wackestone,
mudstone, mud-dominated packstone and tight grainstone. Based
100 E. Sfidari et al. / Journal of Petroleum Science and Engineering 121 (2014) 87–102

Fig. 18. Electro-facies representation versus system tracts, reservoir zones, lithology, core data and depositional models for well A.

on the Lucia rock fabric classification it falls in class 2. It is
characterized by low RHOB, high DT and high NPHI.
7. Conclusions
Reservoir zonation is the most important job in oil industry
which dominantly relies on the main properties of rocks. In this
paper, a method is proposed for reservoir zonation and electro-
facies determination based on the statistical and intelligent
clustering methods. Comparing the results of electro-facies deter-
mination showed that the intelligent clustering method based on
the SOM neural networks provided more successful clustering.
The electro-facies derived from the SOM networks show a good
agreement with reservoir geological (litho-facies) and petrophysi-
cal data. Based on geological studies, six litho-facies are identified
for the carbonate rocks of Kangan and Dalan formations. Accord-
ingly, it is preferred to choose six electro-facies as inputs of
clustering algorithms utilized in this study. As is seen from the
 E. Sfidari et al. / Journal of Petroleum Science and Engineering 121 (2014) 87–102 101

Fig. 19. Correlation of electro-facies determined from well logs and sequence stratigraphic units in 9 wells of SPGF.

results presented in Figs. 13 and 14, selection of six EFs leads to a (equivalent of electro-facies 4–6), litho-facies 4 with moderate
better discrimination of reservoir from non-reservoir intervals. reservoir quality (equivalent of electro-facies 2) and litho-facies
The reservoir rocks were classified into four litho-facies assem- 1 and 2 with poor to bad reservoir quality (equivalent of electro-
blage including litho-facies 3 with good reservoir quality facies 1 and 3). Finally, the distribution of the petrophysical
102 E. Sfidari et al. / Journal of Petroleum Science and Engineering 121 (2014) 87–102

Table 2 Ji, C.Y., 2000. Land-use classification of remotely sensed data using Kohonen self-
Average log properties and core porosity and permeability for the studied electro- organizing feature map neural networks. Photogramm. Eng. Remote Sens. 66,
facies. 1451–1460.
Kamali, M.R., Rezaee, M.R., 2003. Burial history reconstruction and thermal
EFs NPHI RHOB DT GR PEF POR. LOG PERM. Color code modelling at Kuh-e Mond. SW Iran J. Pet. Geol. 26 (4), 451–464.
Karpur, L., Lake, L., Sepehronoori, K., 2000. Probability logs for facies classification.
in situ 24 (1), 57–78.
1 0.044 2.770 144.41 25.68 4.64 0.02  1.1 Red
Kohonen, T., 2001. Self-organizing maps, third extended ed.Springer Series in
2 0.043 2.793 143.39 25.13 4.57 0.03  1.8 Green
Information Sciences, vol. 30. Springer, Berlin (501 pp.).
3 0.08 2.668 158.60 29.27 4.81 0.07  0.2 Blue
Kohonen, T., 1998. The self-organizing map. Neurocomputing 21 (1–3), 1–6.
4 0.072 2.690 137.84 28.04 4.64 0.06 0.3 Light blue Konert, G., Afif, A.M., AL-Hajari, S.A., Droste, H., 2001. Paleozoic stratigraphy and
5 0.123 2.502 179.56 32.44 5.11 0.19 1.9 Purple hydrocarbon habitat of the Arabian Plate. GeoArabia 6 (3), 407–442.
6 0.11 2.554 172.66 36.01 4.81 0.11 1 Yellow Lucia, F.J., 1999. Carbonate Reservoir Characterization. Springer-Verlag, Berlin
(226 pp.).
Ma, Z.Y., Wang, H., Sitchler, J., Gurpinar, O., Gomez, E., Wang, Y., 2014. Mixture
decompositions and lithofacies clustering from wireline logs. J. Appl. Geophys.
properties is studied in detail in the framework of electro-facies 102, 10–20.
analysis and sequence stratigraphy. The results of electro-facies Perez, H.H., Datta-Gupta, A., Mishra S., 2003. The Role of Electrofacies, Lithofacies,
analysis in 9 wells from the South Pars gas field and their and Hydraulic Flow Units in Permeability Predictions from Well Logs: A
Comparative Analysis Using Classification Trees. SPE Annual Technical Con-
geological specifications show a good agreement with reservoir ference and Exhibition, Denver, Paper 84301-MS, pp. 1–11.
zones and sequence stratigraphic units. It is expected to use the Qi, L., Carr, T., 2006a. Neural network prediction of carbonate lithofacies from well
methodology presented in this study for a more successful logs, Big Bow and Sand Arroyo Creek fields, Southwest Kansas. Comput. Geosci.
32, 947–964.
reservoir characterization and modeling. Qi, L.S., Carr, T.R., 2006b. Neural network prediction of carbonate lithofacies from
well logs, Big Bow & Sand Arroyo Creek Fields, Southwest Kansas. Comput.
References Geosci. 32, 947–964.
Rogers, S.J., Fang, J.H., Karr, C.L., Stanley, D.A., 1992. Determination of lithology from
well logs using a neural network. AAPG Bull. 76, 731–739.
Aali, J., Rahimpour-Bonab, H., Kamali, M.R., 2006. Geochemistry and origin of the Saggaf, M.M., Nebrija, E.L., 2003. A fuzzy logic approach for the estimation of facies
world's largest gas field from Persian Gulf. Iran J. Pet. Sci. Eng. 50, 161–175. from wire-line logs. AAPG Bull. 87 (7), 1233–1240.
Al-Anazi, A., Gates, I.D., 2010. A support vector machine algorithm to classify Saggaf, M.M., Nebrija, E.L., 2000. Estimation of lithologies and depositional facies
lithofacies and model permeability in heterogeneous reservoirs. Eng. Geol. from wireline logs. AAPG Bull. 84 (10), 1633–1646.
114 (3–4), 267–277. Sfidari, E., Amini, A., Kadkhodaie, A., Ahmadi, B., 2012a. Electrofacies clustering and
Alshahran, A.S., Nairn, A.E.M., 1997. Sedimentary Basins and Petroleum Geology of a hybrid intelligent based method for porosity and permeability prediction in
the Middle East. Elsevier, Netherlands (843 pp.). the South Pars Gas Field, Persian Gulf. JGEOPE 2 (2), 11–23.
Cassano, E.N., Lynch, A.H., Cassano, J.J., Koslow, M.R., 2006. Classification of synoptic Sfidari, E., Kadkhodaie-Ilkhchi, A., Najjari, S., 2012b. Comparison of intelligent
patterns in the western Arctic associated with extreme events at Barrow, and statistical clustering approaches to predicting total organic carbon using
Alaska, USA. Clim. Res. 30, 83–97. intelligent systems. J. Pet. Sci. Eng. 86–87, 190–205.
Cole'ou, T., Poupon, M., Azbel, K., 2003. Unsupervised seismic facies classification: a Sorooshian, S., Hsu, K., Gao, X., Gupta, H.V., Imam, B., Braithwaite, D., 2000.
review and comparison of techniques and implementation. Lead. Edge 22, Evaluation of PERSIANN system satellite-based estimates of tropical rainfall.
942–953. Bull. Am. Meteorol. Soc. 81, 2035–2046.
Dubois, M.K., Bohling, G.C., Chakrabarti, S., 2007. Comparison of four approaches to Strecker, U., Uden, R., 2002. Data mining of 3D poststack seismic attribute volumes
a rock facies classification problem. Comput. Geosci. 33 (5), 599–617. using Kohonen self-organizing maps. Lead. Edge 21, 1032–1037.
Hong, Y., Hsu, K.L., Sorooshian, S., Gao, X.G., 2005. Self-organizing nonlinear output Tang, H., Ji, H., 2006. Incorporation of spatial characteristics into volcanic facies and
(SONO): a neural network suitable for cloud patch-based rainfall estima- favorable reservoir prediction. SPE Reserv. Eval. Eng. 9 (5), 565–573.
tion from satellite imagery at small scales (Article no. W03008). Water Resour. Vesanto, J., Himberg, J., Alhoniemi, E., Parhankangas, J., 1999. Self-organizing map in
Res. 41. Matlab: the SOM Toolbox. In: Proceedings of the Matlab DSP Conference. Espoo,
Hong, Y., Chiang, Y., Liu, Y., Hsu, K.L., Sorooshian, S., 2006. Satellite-based Finland: Comsol Oy, pp. 35–40.
precipitation estimation using watershed segmentation and growing hierarch- Wolff, M., Pelissier-Combescure, J., 1982. FACIOLOG: automatic electrofacies deter-
ical selforganizing feature mapping techniques. Int. J. Remote Sens. 27, mination. Society of Professional Well Log Analysts Annual Logging Sympo-
5165–5184. sium, pp. 6–9.
Hong, T., Christopher, D.W., 2008. Multivariate statistical log log-facies classification Ziegler, M., 2001. Late permian to holocene paleofacies evolution of the Arabian
on a shallow marine reservoir. J. Pet. Sci. Eng. 61 (2–4), 88–93. plate and its hydrocarbon occurrences. GeoArabia 6 (3), 445–504.