Journal of Petroleum Exploration and Production Technology (2023) 13:2197–2215

https://doi.org/10.1007/s13202-023-01668-5

ORIGINAL PAPER-EXPLORATION GEOPHYSICS

Electro‑facies classification based on core and well‑log data

Reda Al Hasan1 · Mohammad Hossein Saberi2 · Mohammad Ali Riahi3 · Abbas Khaksar Manshad4

Received: 16 July 2022 / Accepted: 18 June 2023 / Published online: 13 July 2023
© The Author(s) 2023

Abstract
Facies studies represent a key element of reservoir characterization. In practice, this can be done by making use of core and
petrophysical data. The high cost and difficulties of drilling and coring operations coupled with the time-intensive nature
of core studies have led researchers toward using well-log data as an alternative. In the Teapot Dome Oilfield, where core
data are limited to those from only a single well, we used well-log data for reservoir electro-facies (EF) studies via two
unsupervised clustering methods, namely multi-resolution graph-based clustering (MRGC) and self-organizing map (SOM).
Satisfactory results were obtained with both methods, distinguishing seven electro-facies from one another, where MRGC
had the highest discriminatory accuracy. The best reservoir quality was exhibited by electro-facies 1, as per both methods.
Our findings can be used to avoid some time-intensive steps of conventional reservoir characterization approaches and are
useful for prospect modeling and well location proposal.

Keywords Multi-resolution graph-based clustering (MRGC) · Self-organizing map (SOM) · Electro-facies · Lithofacies

Abbreviations H-D Higher-dimensional

KNN The k-nearest neighbor
Abbreviations
LLD Resistivity log
AHC Ascendant hierarchical clustering
LF Lithofacies
ANN Artificial neural network
L-D Low-dimensional
ANNs Artificial neural networks
MRGC​ Multi-resolution graph-based clustering
BMU The best matching unit
NPHI or PHIN Neutron log
DC Dynamic clustering
PE Photoelectric log
DEN Density log
PRB The Powder River Basin
DT Sonic log
RHOZ Density log
GR Gamma ray log
SOM Self-organizing map
TSF Tensleep sandstone formation
* Mohammad Hossein Saberi
mh.saberi@semnan.ac.ir List of symbols
Reda Al Hasan D (x, y) Distance function
realhasan@albaath-univ.edu.sy Dij Distance between input vector and com-
Mohammad Ali Riahi munication weights
mariahi@ut.ac.ir Dist (X, Y) Euclidean distance between two samples
Abbas Khaksar Manshad x and y
khaksar@put.ac.ir KRI Kernel representative index
1
M (x, y) Neighborhood function
Master of Petroleum, Faculty of New Sciences
N Total number of data points
and Technologies, Department of Petroleum Engineering,
Semnan University, Semnan, Iran NI or I Neighboring index function
2 P Total number of properties
Petroleum Engineering Department New Technology
Campus, Semnan University, Semnan, Iran s(x) Sum of the weighted ranks for a given
3 measurement at point x
Institute of Geophysics, University of Tehran, Tehran, Iran
4
T Maximum number of iterations or train-
Petroleum Engineering at the Petroleum University
ing length
of Technology, Ahvaz, Iran

13
Vol.:(0123456789)
2198 Journal of Petroleum Exploration and Production Technology (2023) 13:2197–2215

t Iterations number of the training evolutions (Serra and Abbott 1982), not to mention reser-
algorithm voir properties distribution (Shoghia et al. 2020). LF studies
Wij Communication weights between the have been done by core, well-log, and seismic data. Core
map unit j and the input sample i data can provide accurate information about sedimentary
x  mTh nearest neighbor of y (m ≤ N − 1) processes, but the cost and time intensiveness of coring and
Xk Input vector core data analysis limits their availability in many cases. As
𝛼(t) Learning rate a workaround, core data have been combined with traditional
𝛼0 Initial training rate well-log data to obtain the best description of the LFs (Jarvie
a Smoothing parameter greater than zero et al. 2007; Loucks and Ruppel 2007; Dong et al. 2015).
This hybrid approach with well-log and core data has been
proven to be effective (Bishop 1995; Lim and Kim 2004;
Introduction Wang 2012; Bhattacharya et al. 2016; Schlanser et al. 2016).
Well logs represent many rock characteristics that are well
Electro-facies (EF) study is a basic step of any reservoir correlated to core data for geological interpretation (Rider
characterization effort. Conventionally, it has been done and Kennedy 2011). Among other well-log data, gamma ray
based on a combination of core and petrophysical (i.e., (GR), density (RHOB), sonic (DT), porosity (NPHI), and
well-log) data. In the meantime, the relatively high cost of photoelectric index (PE) logs have been acknowledged as the
coring and difficulties of core study as the number of cores best to describe underground rocks (Davis et al. 1997; Qing
grow larger have boosted the importance of well-log data for and Nimegeers 2008). Many statistical algorithms have been
lithofacies studies. Well logs have been used for reservoir used for EF characterization; these include support vector
characterization and geological evaluation. Manual analysis machines (SVMs) (Vapnik et al. 1997; Smola and Schölkopf
of well-log data is however difficult to perform due to the 2004), artificial neural networks (ANNs) (Liu et al. 1992;
relatively large amount of such data. EF studies have been Dubois et al. 2007), and multi-resolution graph-based clus-
emerged to address this problem. An EF refers to a series tering (MRGC) (Ye and Rabiller 2000; Wu et al. 2020).
of log responses that characterize a particular rock unit as In Iranian carbonate reservoirs, researchers have used
is distinctive of other rock units (Serra et al. 1982). An EF various mathematical and statistical concepts for EF studies.
usually captures one or more reservoir properties as log An EF study of Darian Formation was performed by MRGC
responses are measurements of physical rock properties. The at 22 wells penetrating this formation. The reservoir quality
concept of EF was originally used by geologists to identify corresponding to different EFs was then evaluated using core
rock units of similar properties and hence maturate prospects data. 3D modeling was performed, and sequential simulation
of hydrocarbon, coal, minerals, etc. The term (i.e., EF) was was applied as a geostatistical approach. Results confirmed
however first coined by Gressly (1838). By definition, EF the consistency of the 3D models built by the petrophysical
refers to a specific set of properties for a sedimentary rock logs over the Darian Formation, indicating the effectiveness
unit. These characteristics initially included only geological of EF studies for reservoir characterization (Mehmandosti
and fossilogical features, including color, stratification, com- et al. 2017). EF classification in reservoir zones has been
position, texture, sedimentary structures, and fossil append- practiced by various researchers. For instance, Pabakhsh
ages. Later on, a better and more comprehensive definition et al. (2012) used the MRGC for estimating photomechanical
of EF was proposed by Selley (1976): A sedimentary EF properties and hence distinguish between different lithologi-
characterizes a complex of sediments or sedimentary rocks cal formations. Kumar and Kishore (2006) used an ANN
with particular lithology, geometry, fossil appendages, sedi- to estimate EFs in a carbonate/clastic reservoir. In another
mentary structures, and paleo-stream patterns, which can work, Bahar et al. (1999) introduced different clustering
be discriminated from other sediments. Lithological-based methods and used them to identify reservoir EF in a carbon-
facies are referred to as lithofacies (LF) (Selley 1986). Most ate reservoir. Ye and Rabiller (2000) used MRGC for clas-
geologists have asserted that sedimentary facies represent sification of EFs. Using different clustering methods includ-
geological units (Embry and Johannessen 1993; Odezulu ing MRGC, self-organizing map (SOM), dynamic clustering
et al. 2014; Sisinni et al. 2016; Tomassetti et al. 2018). LFs (DC), ascending hierarchical clustering (AHC), and ANN,
reflect significant reservoir parameters, with each LF char- Khoshbakht and Mohammadnia (2012) achieved acceptable
acterizing a distinguished stratum from the others (Kadhim permeability prediction accuracy. Mahmoudi et al. (2011)
et al. 2015). This highlights the role of LF classification in utilized multivariate cluster analysis at Western Salt Well
reservoir characterization (Avanzini et al. 2016). Previous No. 1 (Bandar Abbas, Iran) to identify different EFs and
research works have demonstrated that proper arrangement zonate the reservoir accordingly. Hemmati et al. (2016)
of LFs facilitates the interpretation of vertical distribution employed four clustering approaches to zonate a carbonate
of geological events such as sedimentary sequences and reservoir in southern Iran based on geological/petrophysical

13
Journal of Petroleum Exploration and Production Technology (2023) 13:2197–2215 2199

Fig. 1  Geologic column of the

Teapot Dome and the study for-
mation inside the red rectangle

13
2200 Journal of Petroleum Exploration and Production Technology (2023) 13:2197–2215

results presented and discussed in Section “Results and dis-

cussion.” Final conclusions are drawn in Section “Results
and discussion”.

Geological setting

The Teapot Dome is a faulted dome structure within the Salt

Creek Anticline in the southwestern sector of the Powder
River Basin (PRB), Natrona County, Wyoming. It is a part
of the Basin-Margin Anticline Play of the PRB petroleum
province (Dolton et al. 1995), where a Precambrian base-
ment is overlaid by Paleozoic strata composed of relatively
thin interbedded sequences of sandstones (dune/interdune
origin), dolomites, limestone, and evaporites with marine
origin. Sedimentary strata of the Cretaceous age range
from fluvial sandstones and shales to marine shales and
sandstones.
The top two hydrocarbon-producing reservoirs in the
Teapot Dome are the Shannon Sandstone and Second Wall
Creek Sands of the Upper Cretaceous Cody Shale and Fron-
tier formations, respectively. Hydrocarbons have also been
produced from other Upper and Lower Cretaceous forma-
tions including Niobrara and Steele Shale and also from
Lower Cretaceous non-marine sands of the Thermopolis
Fig. 2  An image of a core box from Well 48-X-28 showing the top of
B-dolomite, the base of B-sandstone, and C1-dolomite (Kamran Jafri
Shales, the Muddy Sandstone, and the Dakota Sandstone
et al., 2015) (Dennen et al. 2005).
Being a major hydrocarbon-producing horizon in the
Teapot Oilfield, Wyoming, the Pennsylvanian Tensleep
facies. The best clustering performance was obtained with Sandstone Formation (TSF) is partially eolian in origin.
AHC and MRGC although AHC marginally outperformed The TSF consists of marine carbonate/dolomite beds and
the MRGC. porous and permeable eolian cross-bedded sandstones of
In the Teapot Dome Oilfield, Natrona County, Wyoming, dune and interdune origin. The siliciclastic units comprise
the vast area of the field coupled with the large number the dominant part of the TSF with alterations of dolomite
of drilled wells complicate the study of facies using core (Dolton et al. 1995, Kamran Jafri et al. 2016). Figure 1
data. Accordingly, we alternatively opted for well logs to shows the geologic column of the Teapot Dome.
undertake EF assessments. To increase the reliability of
the results, several logs, rather than one, were used in the
present study. These included GR, DT, RHOB, and NPHI,
among others. Materials and methods
The main goal of this study was to find the best reservoir
quality in Teapot Dome Oilfield based on EF classification Details of the data used
on well-log data with the help of MRGC and SOM. The
results were then verified and compared to the correspond- The core data from Well 48-X-28 were only available
ing core data. In the proposed methodology, our findings for the depth range of 5300–5653 m from mean sea level
can be used to avoid some time-intensive steps of conven- (MSL)—i.e., only for TSF, and the data from this well
tional reservoir characterization approaches and are useful were subsequently used for testing the results. Core images
for prospect modeling and well location proposal, especially and descriptions were used to describe the A- and B-dolo-
when reservoir heterogeneity is not significant. mite and B-sandstone units in the TSF, as per a previous
The rest of this article is managed as follows. Sec- study by Kamran Jafri et al. (2016). Figure 2 shows an
tion “Geological setting” presents the geological setting of image of a core box containing core samples from top of
the study area. Section “Materials and methods” elaborates B-dolomite, base of sandstone-B and C1-dolomite at Well
on the used data and the proposed methodology, with the 48-X-28. Given the wide extension of the study area and

13
Journal of Petroleum Exploration and Production Technology (2023) 13:2197–2215 2201

shale reservoirs, rock source assessment, potash deposit

assessment, and geological correlations. A sonic (DT) log
serves as an indication of porosity by measuring the inter-
val transit time (t, delta t, or Dt) for a compressed acoustic
wave traveling through the formation along the axis of
the well, thus indicating the porosity in the rock. RHOB
log data contribute to identification of evaporite minerals,
gas-bearing zones, and hydrocarbon density, highlight-
ing sandstone reservoirs. NPHI log records porosity data
by measuring the hydrogen content of the formation. In a
clean (i.e., shale-free) formation where the pore space is
filled with water or oil, the NPHI log measures the fluid-
filled porosity (Asquith et al. 2004).

Clustering methods

Clustering has been acknowledged as a powerful approach to

data analysis. First introduced in 1935, it has been developed
Fig. 3  A 2D SOM: all nodes of the map are directly connected to the into numerous aspects with a handful of different applica-
input vector, with no inter-node connection (Kohonen 2000) tions. Multivariate analysis aims at understanding and
describing possible associations among multiple variables.
In presence of complex relationships among different param-
practical infeasibility of studying core data at the many eters, single-parameter statistics tend to ignore variations of
wells drilled into the Teapot Dome, well-log data were other variables. Hence, different methods have been devel-
focused in this work. In this study, a total of 45 wells were oped to handle multivariate data. Properties of univariate
selected where the required well logs for facies study were and bivariate datasets can be easily extracted by examining
available. Indeed, well selection was done in such a way to the corresponding 2D histogram/plot. A 3D representation
ensure the availability of resistivity (LLD), GR, RHOB, is however necessary to analyze a trivariate dataset, which
NPHI, and DT logs in most of the wells. LLD measure- can be developed by compiling multiple 2D histograms of
ments can be used to distinguish between hydrocarbon- data points. As the number of dimensions increases, these
bearing and water-bearing areas, not to mention its appli- simple solutions may no longer work and one needs to
cability for porosity and permeability evaluations. Among reduce the problem dimensionality before it can be solved
other applications, GR log data help determine the shale properly. Several methods have been used to group similar
(clay) volume (Vshale, Vclay) in sandstone reservoirs con- features (i.e., variables) considering the studied dataset, so
taining uranium minerals, potassium feldspar, mica, and/ as to achieve reduced dimensionality of the problem. Such
or glauconite, differentiation between radioactive and oil methods are generally referred to as clustering (Webb 2002).

Fig. 4  Workflow of a SOM

13
2202 Journal of Petroleum Exploration and Production Technology (2023) 13:2197–2215

The basic idea behind the clustering is that visual interpreta- 4. dij =||xk − wij|| to find the best matching unit (BMU).
tion of a large problem can be quickly achieved by clustering 5. Calculate the neighborhood radius around the BMU
the entire dataset into distinctive clusters. In this respect, a using the desired neighborhood function. The size of
cluster refers a set of objects with maximum similarity to this neighborhood decreases with increasing the algo-
one another and maximum distinctiveness to other clusters. rithm time.
Herein, the similarity can be measured by different criteria. 6. Adjust communication weights of the nodes in the BMU
An example of similarity criterion is the inverse distance neighborhood to make them closer to the BMU. For this
between different objects, so that the higher the inverse dis- purpose, further adjustment must be applied to closer,
tance of the objects, the higher their similarity. rather than farther, nodes to the BMU.
The followings are some of the conventional clustering 7. Repeat Steps 2 to 6 iteratively until the algorithm con-
models: verges (i.e., the weight vectors exhibit no significant
change).
1. Connectivity models (e.g., hierarchical clustering),
where model is built based on a distance criterion. BMU calculation is based on the Euclidean distance,
2. Centroid models (e.g., K-means clustering), where each as the similarity criterion, between the weight vectors of
cluster is represented by an average vector. the output nodes and the values of the input vectors. The
3. Distribution models, where clusters are modeled using neighborhood radius shrinks as the training algorithm
statistical distributions. proceeds, eventually encompassing the BMU alone. The
4. Density models, where clusters (e.g., areas) are distin- neighborhood is usually determined by a Gaussian or
guished by density. exponential function so that nodes closer to the winning
BMU are more affected than the farther nodes. The learn-
Clustering algorithms can be classified based on their ing rate is set by an exponential function to ensure SOM
clustering model. So far, more than 100 clustering algo- convergence (Cai et al. 2019).
rithms have been presented. The following subsections
t
( )
explain the two clustering algorithms used in this study. 𝛼(t) = 𝛼0 1 − (1)
T
In Eq. (1), 𝛼0 indicates learning rate, t is the iteration
Clustering by SOM
number of the training algorithm, and T is the maximum
number of iterations (i.e., training length).
SOM is an unsupervised ANN that produces a low-dimen-
Then in the mapping step, SOM automatically catego-
sional (L-D) graph called a “map.” It is distinctive of other
rizes unforeseen input vectors.
neural networks in that it uses a neighborhood function to
An ANN is applied in two stages:
maintain topological properties of the input space, reflect-
ing an L-D representation of high-dimensional (H–D) input
1. Determining the entries from the record dataset.
data. In other words, SOM maps an H–D input to the cor-
2. Designing an error backpropagation neural network with
responding L-D map. This is done by finding the node with
an appropriate training algorithm.
the closest weight vector to the input vector and assigning
the coordinates of that node on the map to the corresponding
An error backpropagation network is a training moni-
input vector. Similar to other neural networks, SOM works
toring tool that feeds inputs to the network and compares
in two phases: training and mapping (i.e., automatic clas-
the error between the target output and the generated out-
sification of unseen input vectors) (Fig. 3).
put by the training dataset.
The error of the network is then backpropagated, and
Mechanism of the SOM algorithm the weights are adjusted through multiple iterations. The
training process terminates the calculated output which
An SOM algorithm goes through an iterative process involv- is close enough to the designed output. In many cases,
ing vector measurements through the following steps: training algorithms have been optimized by updating the
weights and biases (Sefidari et al. 2014; Cai et al. 2019),
1. Initialize the weight for each output node. and Fig. 4 shows the workflow of a SOM.
2. Select a random vector from the training data and take The network efficiency criterion was set to be the cor-
it to the SOM. relation coefficient of the considered feature between the
3. Calculate the distance between the input vector and com- modeling results and expected data, which was supposed
munication weights of each output node as

13
Journal of Petroleum Exploration and Production Technology (2023) 13:2197–2215 2203

Fig. 5  GR, DT, RHOB, NPHI,

and LLD logs along Well
14-15-SX in Teapot Dome and
demonstration of the depth
matching

13
2204 Journal of Petroleum Exploration and Production Technology (2023) 13:2197–2215

Fig. 6  Histograms of GR, DT, RHOB, NPHI, and LLD logs for the studied wells

Fig. 7  Cross-plots of GR, DT,

and LLD logs for the studied
wells

13
Journal of Petroleum Exploration and Production Technology (2023) 13:2197–2215 2205

to be minimal (Kohonen 1998; Sefidari et al. 2014, Hem- still exponentially proportional to the number of data points
mtin et al. 2016; Cai et al. 2019, de Passos et al. 2020). has limited its application for clustering purposes. The
KNN classifier is usually based on the Euclidean distance
Clustering by MRGC​ between an experimental sample and a training sample. The
Euclidean distance between two samples x and y is defined
Conventional clustering algorithms suffer from a number of as follows:
limitations. First, the number of clusters to be distinguished √
by the algorithm must be known. Second, they are highly
√N
√∑
sensitive to initial conditions and distinctions among data Dist(X, Y) = √ (Xi − Yi )2 (2)
points. Third, they are practically not robust to data discrep- i=1

ancy (Mourot and Bousghiri 1993). As a modern-generation Then the samples are sorted by their distance to the
algorithm, MRGC has been shown to be free of such limi- neighbors and the k-nearest neighbors are identified.
tations and superior to conventional clustering algorithms. Graphical classification methods are known to be suitable
K-nearest neighborhood (KNN) is a clustering algorithm for analyzing low-dimensional small datasets. They are gen-
that considers a fixed (i.e., k) number of neighboring data erally robust to different batch sizes (Aghchelou et al. 2013).
points rather than a fixed neighborhood in space (Dubois MRGC is a nonparametric method that combines the
et al. 2005). This approach comes with particular advan- KNN method with the graphical approach to enjoy benefits
tages. Being easy to implement, this method allows you to of both for clustering datasets of any dimension and complex
record and examine clusters of small size and very different structure.
densities. The fact that the power of estimation by KNN is

Table 1  Results of SOM clustering for the studied wells

Facies Weight GR (GAPI) GRR (GAPI) LLD (OHMM) MSFL (OHMM) NPHI (V/V) RHOB (K/M3) DT (US/M)

1 2102 112.6 112.6 6.73 5.41 0.31 2446.56 292.02

2 986 107.66 107.66 8.46 7.14 0.27 2498.03 274.07
3 642 99.6 99.6 13.31 11.73 0.21 2576.58 244.24
4 1636 99.39 99.39 8.15 6.71 0.29 2483.24 292.6
5 342 90.03 90.03 12.71 10.41 0.23 2526.39 269.09
6 860 78.2 78.2 25.63 20.38 0.16 2600.23 234.9
7 1300 84.85 84.85 10.15 8.4 0.27 2506.88 292.29
8 871 71.32 71.32 19.91 15.25 0.2 2539.06 263.82
9 1295 58.16 58.16 47.52 32.71 0.13 2608.41 228.93

Fig. 8  Logs of the EFs identified by SOM clustering based on the input logs

13
2206 Journal of Petroleum Exploration and Production Technology (2023) 13:2197–2215

Fig. 9  The EFs identified by the SOM clustering after the merging process

MRGC: why and how? groups of equal size while groups of equal volume can be
obtained by the distance (for more information on KRI, see
Facies analysis is very important to determine reservoir Ye and Rabiller 2001). KRI is expressed as follows:
characteristics. Two nearby points along a wellbore may
render geologically far different from one another—this
KRI(x) = NI(x) × M × D (7)
issue has been referred to as dimension problem. MRGC Main advantages of MRGC are listed below:
offers good capabilities for identifying such geological (i.e.,
facies) contrasts. Unlike conventional clustering algorithms, • Ability to identify natural patterns within graph data,
this method does not need a previous knowledge of the data representing the facies arrangement.
structure and number of clusters. Indeed, the optimum num- • No need to previous knowledge of the dataset.
ber of clusters is herein determined automatically, and this • Automatic determination of the optimal number of clus-
task can be adjusted, according to specific requirements of ters (i.e., facies).
the problem at hand, by setting the neighbor index (NI) and • Capability of handling real-time data with complex struc-
kernel representative index (KRI) (Ye and Rabiller 2000). ture.
NI is based on the weighted classification of a given • With adjustable parameters, it produces consistent
measurement point x concerning all other measurement results.
points y. Accordingly, a high NI indicates easily distinguish- • No theoretical limitation in the number of dimensions,
able points (for more information on NI and KNN, see Ye points and categories (Tian et al. 2016; Shi et al. 2017;
and Rabiller 2001). The number of clusters (i.e., facies in dos Passos et al. 2020).
this work) can be easily determined as follows:
N−1
Results and discussion
∑
s(x) = exp(−m∕a) (3)
n=1
Well-log data clustering was performed by SOM and
MRGC in Geolog software. For this purpose, depth match-
{ }
Smin = Min S(xi ) (4)
i=1,N
ing was performed based on anomalous log readings and
necessary corrections were made. Figure 5 shows the GR,
DT, RHOB, NPHI, and LLD logs recorded at Well 14-15-
{ }
Smax = Max S(xi )
i=1,N (5)
sx in the Teapot Dome together with the depth matching
results. Once finished with the initial data verification,
S(x) − Smin
NI(x) = (6) facies clustering was practiced using the Facimage mod-
Smax − Smin ule in the Geolog. To this end, we opted for a well that
provided good information about the field. Subsequently,
The KRI combines NI (x) with a neighborhood function
feature (i.e., log) selection was performed to identify the
and a distance function. By NI (x), the kernel of a cluster
training data. Accordingly, GR, DT, RHOB, NPHI, and
can be identified. The number of neighbors, M, helps create

13
Journal of Petroleum Exploration and Production Technology (2023) 13:2197–2215 2207

Fig. 10  Cross-plots of the logs used for final SOM clustering

LLD logs were selected and their histograms were checked indicating good reservoir quality when GR readings are
to see the data distribution and frequency range (Fig. 6). small.
Results indicated the data normality and suitability for
training the algorithm. Results of SOM
Figure 7 shows cross-plots of different logs. On this fig-
ure, different colors indicate density of the data points. This The 2D SOM clustering method was applied to distin-
figure shows that particular logs exhibit similar trends that guish between different clusters in the data. In this method,
are well correlated to one another. According to this fig- horizontal and vertical axes are defined to determine the
ure, the highest density of data points was seen to exhibit number of clusters. In this work, input data to the SOM
GR values of 60–120 API and DT values of 200–350 μs/m, clustering included the horizontal and vertical coordinates

13
2208 Journal of Petroleum Exploration and Production Technology (2023) 13:2197–2215

Table 2  Results of MRGC for the studied wells

Facies Weight GR (GAPI) GRR (GAPI) LLD (OHMM) MSFL (OHMM) NPHI (V/V) RHOB (K/M3) DT (US/M)

1 404 26.29 26.29 560.95 248.75 0.05 2816.93 191.49

2 382 28.32 28.32 187.33 49.83 0.06 2534.12 202.22
3 1014 57.35 57.35 15.78 13.77 0.18 2455.48 270.87
4 672 72.81 72.81 31.04 32.78 0.1 2665.26 217.67
5 800 100.7 100.7 23.66 22.49 0.16 2696.15 202.82
6 950 94.82 94.82 8.48 8.08 0.21 2541.57 260.87
7 360 118.49 118.49 3.74 2.08 0.38 2252.92 295.22
8 2512 91.52 91.52 8.35 6.97 0.3 2521.24 305.81
9 2255 122.38 122.38 7.14 6.48 0.3 2488.05 288.97
10 685 87.22 87.22 5.07 2.3 0.34 2281.75 263.87

Fig. 11  The EFs identified from the MRGC after the merging process

of each group in the network space. A training algorithm and EF6 exhibited similar characteristics corresponding to
was developed to structure SOM ensembles in such way to medium reservoir quality. Ultimately, KNN clustering was
represent the entire dataset and associated weights at each deployed to generalize the results to all intervals of the well.
iteration. At each iteration, a horizontal vector was randomly
selected from the dataset and its distance to all weighted Results of MRGC​
vectors of the network was calculated. Therefore, after the
training phase, we ended up running the SOM to obtain a MRGC is an unsupervised clustering algorithm for identify-
model of 9 distinctive facies. Table 1 displays the results of ing similar areas, categories, or facies. Some unsupervised
SOM clustering at the studied wells (Fig. 8). clustering algorithms (e.g., SOM) require that the number of
In the next step, EFs of similar characteristics were final clusters is known. MRGC, however, works based on the
merged to prevent cluster overgrowth. After reviewing the number of dimensions of the search space, where data distri-
clustering results, it was found that EF1, EF2, EF3, and EF5 bution density determines the number of clusters. Upon clus-
are similar enough to be merged into a single EF. Follow- tering, the produced cross-plots indicate different clusters
ing this procedure, final number of EFs was reduced to 7. (i.e., facies) clearly, as shown by different colors. Afterward,
Figure 9 shows the identified EFs after the merging process, one can see the number of facies and distribution of different
and Fig. 10 shows cross-plot of the logs used for the final logs with respect to each facies (i.e., number of samples to
clustering. From the cross-plots of Fig. 10, it is evident that which each facies is assigned and average values of each log
EF1 and EF7 provide the best and worst reservoir quali- against that facies). Optimal models were obtained with 10,
ties, respectively. On the other hand, EF2, EF3, EF4, EF5, 14, 18, 20, and 23 facies. Given the geological setting of the

13
Journal of Petroleum Exploration and Production Technology (2023) 13:2197–2215 2209

Fig. 12  Cross-plots of the logs used after implementing the MRGC method

13
2210 Journal of Petroleum Exploration and Production Technology (2023) 13:2197–2215

Fig. 13  Comparing EF clustering outputs with petrophysical logs at Well 14-15-SX

13
Journal of Petroleum Exploration and Production Technology (2023) 13:2197–2215 2211

Fig. 14  a, b Boxplots of GR log for the EFs obtained from the SOM and MRGC algorithms, and c, d boxplots of DT log for the EFs obtained
from the SOM and MRGC algorithms

Fig. 15  Cross-plots of GR versus LLD for the EFs resulting from a MRGC and b SOM

13
2212 Journal of Petroleum Exploration and Production Technology (2023) 13:2197–2215

study area and previous studies, we ended up developing a

10-EF model. Table 2 displays the results of the MRCG for
the studied wells.
As in the SOM method, the EFs of similar characteristics
were merged to prevent cluster overgrowth. This led to a
reduction of the number of facies to 7. Figure 11 shows the
EFs after the merging process, and Fig. 12 shows cross-plots
of the logs used for the clustering. From this cross-plot, it
is evident that EF1 and EF2 refer to the best reservoir while
the worst reservoir quality is associated with EF6 and EF7.
On the other hand, EF3, EF4, and EF5 exhibited similar
characteristics corresponding to medium reservoir quality.
The KNN algorithm was then utilized to propagate the
results to the entire well column. In Fig. 13, one can see and
compare the set of identified EFs on petrophysical logs. The
leftmost column indicates the depth in m. The GR and DT
logs are visualized along Tracks 1 and 2, respectively, while
Track 3 hosts RHOB and NPHI logs. The two columns on
the right demonstrate the results of EF clustering through
MRGC and SOM algorithms, respectively.

Prioritization of EF logs based on reservoir quality

Reservoir quality is known to be controlled by reservoir

porosity, permeability, and shale volume. Accordingly, any
geological process that improves either of these parameters
can contribute to reservoir quality. Such processes can be
physical or chemical. The generated EF logs were examined
and prioritized in terms of reservoir quality. For analyzing
the EF logs, interpretations were made on each final EF
based on the number of EFs merged to obtain that final EF.
Separate analysis for EF can be done using box diagrams.
Figure 14 shows variations of GR and DT logs in different
EFs identified by the SOM and MRGC methods. Indeed,
the lower the GR reading and/or the higher the DT reading,
the higher the reservoir quality. Accordingly, EF1 and EF2
in Fig. 14a refer to the best reservoir quality, as indicated
by their low GR readings (i.e., small shale volume), while
EF7 marks the lowest reservoir quality due to its high GR
readings. In Fig. 14b, however, the best and worst reser-
voir qualities are shown by EF1 and EF7, respectively. On
Fig. 14c, DT values range from 200 to 300 US/M, while the
corresponding DT readings to Fig. 14d fall in the range of
150–300 US/M.
From the above, we find that EF1 exhibits the best reser-
voir quality by both methods, while EF6 and EF7 showed
the worst reservoir quality with the SOM and MRGC meth-
ods, respectively. In professional communities, it is usually
common to classify different facies as either good or bad
ones. To provide a reliable standard for drilling and verify
the results in terms of correctness and practicality, Fig. 15
Fig. 16  Comparing the MRGC- and SOM-derived EFs against petro- shows a cross-plot of GR versus LLD. This figure validates
physical logs at Well 48x-28 the results.

13
Journal of Petroleum Exploration and Production Technology (2023) 13:2197–2215 2213

Fig. 17  Photographs of core

box from Well 48x-28. On this
figure, lithofacies (a), (b), (c),
(d), (e), (f), and (g) correspond
to EF1, EF2, EF3, EF4, EF5,
EF6, and EF7, respectively

Figure 16 depicts the depth, well logs (GR, LLD, and and this can be clearly seen by matching the records, for
RHOB), and the results of MRGC and SOM at Well 48x- example the facies at a depth of 1694–1703 m.
28. By comparing the MRGC- and SOM-derived EFs
with the logs, one may see that the MRGC outperformed
the SOM. Since core data were available from the men- Conclusions
tioned well in a depth interval of 1615–1723 m, Fig. 17
presents photographs of the core box corresponding to dif- Data compilation is the basis of modeling and classifica-
ferent EFs. On this figure, facies (a) corresponds to EF1, tion algorithms. A clustering algorithm identifies differ-
which has the highest porosity coupled with low GR and ent clusters of similar properties in a large set of data and
is encountered in a wide range of depths (1289–1304 m, tries to maximize the similarity within each cluster while
1660–1757 m); facies (b) corresponds to EF2, exhibits a minimizing it between different clusters. The SOM algo-
GR close to facies (a), and is abundant in the depth ranges rithm maps an H–D input space to a L-D map space. The
of 870–885 m and 1319–1332 m; facies (c) corresponds MRGC algorithm is a nonparametric method that com-
to EF3, exhibits medium reservoir quality, and occurs in bines the KNN method with the graphical classification
depth ranges of 152–163 m and 1703–1704 m; facies (d) techniques. Based on the results, the following conclusions
corresponds to EF4 and refers to a rock of lower reser- were drawn:
voir quality than EF3, being found in the depth ranges
1308–1312 m, 1430–1559 m, and 1605–1615 m; facies (e) • Considering the conditions of the research problem, we
corresponds to EF5 with a medium-to-weak reservoir qual- focused on a particular set of well logs, including GR,
ity in terms of porosity and GR and occurs in the depth RHOB, DT, NPHI, and LLD.
ranges of 757–813 m, 887–930 m, and 1204–1303 m; • Necessary corrections were made to raw data and the
facies (f) corresponds to EF6 with poor reservoir quality Facimage module in the Geolog software was utilized
and is found in the depth ranges of 163–405 m, 609–632 m, to implement SOM and MRGC algorithms.
1176–1197 m, and 1398–1319 m; and facies (g) corre- • With both methods, we ended up with 7 EFs, with the
sponds to EF7 with extremely poor reservoir quality and results verified by cross-plots of GR versus RT for the
occurs in the depth ranges of 472–753 m, 828–869 m, identified EFs coupled with photographs of the cor-
931–1171 m, and 1398–1402 m. From Fig. 16, we find responding core boxes.
that the MRGC method gives a better result than SOM,

13
2214 Journal of Petroleum Exploration and Production Technology (2023) 13:2197–2215

• With both algorithms, EF1 showed the best reservoir Bhattacharya S, Carr TR, Pal M (2016) Comparison of supervised and
quality, as shown by low GR coupled with high NPHI unsupervised approaches for mudstone lithofacies classification:
case studies from the Bakken and Mahantango-Marcellus Shale,
readings. At the other end of the spectrum, EF6 and USA. J Nat Gas Sci Eng 33:1119–1133. https://​doi.​org/​10.​1016/j.​
EF7 were associated with the poorest reservoir quality, jngse.​2016.​04.​055
as indicated by high GR coupled with low NPHI read- Bishop CM (1995) Neural networks for pattern recognition. Oxford
ings. University Press, Oxford, p 482
Cai H, Wu Q, Ren H, Li H, Qin Q (2019) Pre-stack texture based
• Comparing the results of the two algorithms with well semi-supervised seismic facies analysis using global optimiza-
logs, it was found that MRGC outperformed the SOM tion. J Seism Explor 28:513–532
in terms of accuracy. Davis R, Fontanilla J, Biswas P, Saha S (1997) Lithology, lithofa-
• Using the core data, actual lithofacies corresponding cies, and permeability estimation in Ghawar Arab-D reservoir.
Middle East Oil Show, Society of Petroleum Engineers 37702.
to the identified EFs were delineated. Bahrain 15–18:15–17. https://​doi.​org/​10.​2118/​37702-​MS
• EF1 exhibited the best reservoir quality in terms of Dennen K, Burns W, Burruss R, Hatcher K (2005) Geochemical
porosity and GR and occurred in the depth ranges of analyses of oils and gases, naval petroleum reserve No. 3, Tea-
1289–1305 m and 1659–1452 m. However, EF7 ended pot Dome Field, Natrona County, Wyoming. US Geological
Survey Open-File Report, 1275, 69. ‫‏‬U sgs Open-File Report
up with the poorest reservoir quality and occurred 2005–1275. BiblioBazaar (2013)
in the depth ranges of 472–753 m, 827–869 m, 931– Dolton GL, Fox JE (1995) Powder River Basin Province (033). In:
1172 m, and 1397–1402 m. Gautier DL, Dolton GL, Takahashi KI, Varnes KL
Dong T, Harris NB, Ayranci K, Twemlow CE, Nassichuk BR (2015)
Porosity characteristics of the Devonian Horn River shale, Can-
ada: insights from lithofacies classification and shale composi-
tion. Int J Coal Geol 141:74–90. https://​doi.​org/​10.​1016/j.​coal.​
2015.​03.​001
Funding This research does not have any fund from the government Dos Passos FV, Braga MA, Carelli TG, Plantz JB (2020) Electrofa-
sector and it was done only with the support of Semnan University. cies classification of ponta grossa formation by multi-resolu-
tion graph-based clustering (MRGC) and self-organizing maps
Open Access This article is licensed under a Creative Commons Attri- (SOM) methods. Braz J Geophys 38(1):52–61. https://​doi.​org/​
bution 4.0 International License, which permits use, sharing, adapta- 10.​22564/​rbgf.​v38i1.2
tion, distribution and reproduction in any medium or format, as long Dubois MK, Bohling GC, Chakrabarti S (2005) Comparison of four
as you give appropriate credit to the original author(s) and the source, approaches to a rock facies classification problem. Comput Geo-
provide a link to the Creative Commons licence, and indicate if changes sci 33:599–617. https://​doi.​org/​10.​1016/j.​cageo.​2006.​08.​011
were made. The images or other third party material in this article are Dubois MK, Bohling GC, Chakrabarti S (2007) Comparison of four
included in the article's Creative Commons licence, unless indicated approaches to a rock facies classification problem. Comput Geo-
otherwise in a credit line to the material. If material is not included in sci 33:599–617. https://​doi.​org/​10.​1016/j.​cageo.​2006.​08.​011
the article's Creative Commons licence and your intended use is not Embry AF, Johannessen EP (1993) T–R sequence stratigraphy, facies
permitted by statutory regulation or exceeds the permitted use, you will analysis and reservoir distribution in the uppermost Triassic–
need to obtain permission directly from the copyright holder. To view a Lower Jurassic succession, western Sverdrup Basin, Arctic
copy of this licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/. Canada. Elsevier, Amsterdam. https://​doi.​org/​10.​1016/​B978-
0-​444-​88943-0.​50013-7
Gressly A (1838) Observations geologiques sur le Jura Soleurois.
Neue Denkschr. Allg Schweiz. Ges Naturw 2:1–112
References Hemmati N, Nazari F, Tabatabai R, Hashem S (2016) Determina-
tion of electro-facies using clustering methods in one of the
Aghchelou M, Hemmati Ahoei HR, Nabi-Bidhendi M, Rahimi Bahar carbonate reservoirs in southern Iran. Sci Mon Oil Gas Explor
AA (2013) Lithofacies estimation by multi-resolution graph-based Prod 137:67–77. https://​doi.​org/​10.​1016/j.​marpe​tgeo.​2018.​06.​
clustering of petrophysical well logs: a case study from one of the 004.​inPer​sian
Persian Gulf’s gas fields, Iran. Iran J Geophys 7(4):11–30. https://​ Jafri MK, Lashin A, Ibrahim E, Naeem M (2016) Petrophysical evalu-
doi.​org/​10.​2118/​162991-​MS ation of the Tensleep Sandstone formation using well logs and
Asadi Mehmandosti E, Mirzaee S, Moallemi SA, Arbab B (2017) limited core data at Teapot Dome, Powder River Basin, Wyo-
Study and three-dimensional modeling of the Dariyan Formation ming, USA. Arab J Sci Eng 41(1):223–247. https://​doi.​org/​10.​
Electrofacies by using Geostatistics, in one of the Persian Gulf 1007/​s13369-​015-​1741-7
Oilfields Kharazmi. J Earth Sci 3(1):25–44 Jarvie DM, Hill RJ, Ruble TE, Pollastro RM (2007) Unconventional
Asquith GB, Krygowski D, Gibson CR (2004) Basic well log analy- shale-gas systems: the Mississippian Barnett Shale of north-cen-
sis, vol 16. American Association of Petroleum Geologists, tral Texas as one model for thermogenic shale-gas assessment.
Tulsa Am Asso Petrol Geol Bull 91:475–499. https://​doi.​org/​10.​1306/​
Avanzini A, Balossino P, Brignoli M, Spelta E, Tarchiani C (2016) 12190​606068
Lithologic and geomechanical facies classification for sweet spot Kadhim F, Samsuri A, Alwan H (2015) Determination of lithology,
identification in gas shale reservoir. Interpretation 4:SL21–SL31. porosity and water saturation for Mishrif carbonate formation. Int
https://​doi.​org/​10.​1190/​INT-​2015-​0199.1 J Geol Environ Eng 9:1025–1031
Bahar M, Johnstone AH, Hansell MH (1999) Revisiting learning dif- Khoshbakht F, Mohammadnia M (2012) Assessment of clustering
ficulties in biology. J Biol Educ 33(2):84–86. https://​doi.​org/​10.​ methods for predicting permeability in a heterogeneous carbonate
1080/​00219​266.​1999.​96556​48

13
Journal of Petroleum Exploration and Production Technology (2023) 13:2197–2215 2215

reservoir. J Pet Sci Technol 2(2):50–57. https://​doi.​org/​10.​3997/​ intelligent systems. Faculty of Geology, Campus of Sciences, Uni-
2214-​4609.​20145​462 versity of Tehran, 23(75):117–130. (in Persian) https://​www.​sid.​
Kohonen T (1998) The self-organizing map. Neurocomputing 21(1– ir/​fA/​Journ​al/​ViewP​aper.​aspx?​id=​231761.
3):1–6. https://​doi.​org/​10.​1016/​S0925-​2312(98)​00030-7 Selley RC (1976) An introduction to sedimentology. Academic Press,
Kohonen T, Kaski S, Lagus K, Salojärvi J, Honkela J, Paatero V, London, p 408
Saarela A (2000) Self-organization of a massive document col- Selley RC (1986) Ancient sedimentary and environment and their-
lection. IEEE Trans Neural Netw 11:574–585. https://​doi.​org/​10.​ subsurface diagnosis. Chapman and hall, London. 3rd ed. p 317
1109/​72.​846729 Serra O, Abbott HT (1982) The contribution of logging data to sedi-
Kohonen T (2000) The self-organizing map, https://​www.​saeds​ayad.​ mentology and stratigraphy. Soc Pet Eng J 22(1):117–131. https://​
com/​clust​ering_​som.​htm doi.​org/​10.​2118/​9270-​PA
Kumar B, Kishore M (2006) Electrofacies classification–a critical Shi X, Cui Y, Guo X, Yang H, Chen R, Li T, Meng L (2017) Logging
approach. In: 6th international conference & exposition on petro- facies classification and permeability evaluation: multi-resolution
leum geophysics, New Delhi, India, pp 822–825 graph based clustering. In: SPE annual technical conference and
Lim JS, Kim J (2004) Reservoir porosity and permeability estimation exhibition. OnePetro.‫ ‏‬https://​doi.​org/​10.​2118/​187030-​MS
from well logs using fuzzy logic and neural networks. In: society Shoghi J, Bahramizadeh-Sajjadi H, Nickandish AB, Abbasi M (2020)
of petroleum engineers asia pacific oil and gas conference and Facies modeling of synchronous successions-A case study
exhibition, Perth, Oct. 18–20. https://​doi.​org/​10.​2118/​88476-​MS from the mid-cretaceous of NW Zagros, Iran. J Afr Earth Sci
Liu RL, Zhou CD, Jin ZW (1992) Lithofacies sequence recognition 162:103696. https://​doi.​org/​10.​1016/j.​jafre​arsci.​2019.​103696
from well logs using time-delay neural networks. 33rd Annual Sisinni V, McDougall N, Guarnido M, Vallez Y, Estaba V (2016)
Logging Symposium of Society of Petrophysicists and Well Log Facies modeling described by probabilistic patterns using Multi-
Analysts, Oklahoma, Jun. 14–17, ID: SPWLA-1992-L point statistics an application to the kfield. Libya Am Assoc Petrol
Loucks RG, Ruppel SC (2007) Mississippian Barnett Shale: litho facies Geol. https://​doi.​org/​10.​1190/​ice20​16-​63201​06.1
and depositional setting of a deep-water shale-gas succession in Smola AJ, Schölkopf B (2004) A tutorial on support vector regression.
the Fort Worth Basin, Texas. Am Asso Petrol Geol Bull 91:579– Stat Comput 14:199–222
601. https://​doi.​org/​10.​1306/​11020​606059 Tian Y, Xu H, Zhang XY, Wang HJ, Guo TC, Zhang LJ, Gong XL
Mahmoudi HS, Harami M, Mehrgini B (2011) Identification and (2016) Multi-resolution graph-based clustering analysis for
zoning of EF using multivariate cluster analysis method in the lithofacies identification from well-log data: case study of intra-
western salt well 1 of Bandar Abbas region. In the first national platform bank gas fields, Amu Darya Basin. Appl Geophys
conference of Iranian geology, https://​civil​ica.​com/​doc/​117628/. 13(4):598–607
(in Persian( Tomassetti L, Petracchini L, Brandano M, Trippetta F, Tomassi A
Mourot G, Bousghiri S, Ragot J, (1993) Pattern recognition for diagno- (2018) Modeling lateral facies heterogeneity of an upper Oligo-
sis of technological systems: a review: international conference on cene carbonate ramp (Salento, southern Italy). Mar Petrol Geol
systems IEEE, pp 275−281. https://​doi.​org/​10.​1109/​TAES.​2009.​ 96:254–270. https://​doi.​org/​10.​1016/j.​marpe​tgeo.​2018.​06.​004
52591​78 Vapnik V, Golowich SE, Smola AJ (1997) Support vector method for
Odezulu CI, Olawole S, Saikia K, Mento D (2014) Effect of sequence function approximation, regression estimation and signal process-
stratigraphybased facies modeling for better reservoir characteri- ing. In: Mozer, M.C., Jordan, M.I., and Petsche, T. eds Proceed-
zation: a case study from powder river basin. In: paper presented ings of the 1997 conference on Advances in neural information
at the abu dhabi international petroleum exhibition and confer- processing systems. MIT Press, Cambridge, pp 281–287
ence. UAE, Abu Dhabi 2014/11/10/. https://​doi.​org/​10.​2118/​ Wang G (2012) Black Shale lithofacies prediction and distribution pat-
171746-​MS tern analysis of Middle Devonian Marcellus Shale in the Appa-
Pabakhsh M, Ahmadi K, Riahi MA, Abbaszade Shahri A (2012) Pre- lachian basin, Northeastern USA. Ph.D. Thesis, West Virginia
diction of PEF and LITH logs using MRGC approach. Life Sci University, Morgantown, USA, p 216
J 9(4):974–982 Web AR (2002) Statistical pattern recognition. Wiley, New Jersey
Qing H, Nimegeers AR (2008) Lithofacies and depositional history of Wu X, Geng Zh, Shi Y, Pham N, Fomel S, Caumon G (2020) Build-
Midale carbonate-evaporite cycles in a Mississippian ramp set- ing realistic structure models to train convolutional neural net-
ting, Steelman-Bienfait area, southeastern Saskatchewan Canada. works for seismic structural interpretation. GEOPHYSICS
Bull Can Pet Geol 56:209–234. https://d​ oi.o​ rg/1​ 0.2​ 113/g​ scpgb​ ull.​ 85:WA27–WA39
56.3.​209 Ye SJ, Rabiller P (2000) A new tool for electrofacies analysis: Multi-
Rider MH, Kennedy MK (2011) The Geological Interpretation of Well Resolution Graph Based Clustering: SPWLA 41st Annual Log-
Logs (3rd edition). Rider-French Consulting Limited, Scotland, ging Symposium, Dallas, Texas, USA, Jun 4−7
p 440 Ye SJ, Rabiller PJ (2001) U.S. Patent No. 6,295,504. Washington, DC:
Schlanser K, Grana D, Campbell-Stone E (2016) Lithofacies classifi- U.S. Patent and Trademark Office
cation in the Marcellus Shale by applying a statistical clustering
algorithm to petrophysical and elastic well logs. Interpretation Publisher's Note Springer Nature remains neutral with regard to
4:SE31–SE49. https://​doi.​org/​10.​1190/​INT-​2015-​0128.1 jurisdictional claims in published maps and institutional affiliations.
Sefidari I, Kadkhodai A, Sharifi M (2014) Comparison of self-organ-
izing map methods and cluster analysis to evaluate the amount
of organic carbon in hydrocarbon-containing formations using

13