Petroleum Research 8 (2023) 167e182

Contents lists available at ScienceDirect

Petroleum Research
journal homepage: http://www.keaipublishing.com/en/journals/
petroleum-research/

Review of application of artificial intelligence techniques in petroleum

operations
Saeed Bahaloo a, Masoud Mehrizadeh b, Adel Najafi-Marghmaleki c, *
a
Department of Petroleum Engineering, Amirkabir University of Technology, Tehran, Iran
b
Department of Petroleum Engineering, School of Science and Engineering, Khazar University, Baku, Azerbaijan
c
Department of Petroleum Engineering, Ahwaz Faculty of Petroleum, Petroleum University of Technology, Ahwaz, Iran

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

Article history: In the last few years, the use of artificial intelligence (AI) and machine learning (ML) techniques have
Received 22 July 2021 received considerable notice as trending technologies in the petroleum industry. The utilization of new
Received in revised form tools and modern technologies creates huge volumes of structured and un-structured data. Organizing
7 July 2022
and processing of these information at faster pace for the performance assessment and forecasting for
Accepted 8 July 2022
Available online 16 July 2022
field development and management is continuously growing as an important field of investigation.
Various difficulties which were faced in predicting the operative features by utilizing the conventional
methods have directed the academia and industry toward investigations focusing on the applications of
Keywords:
Artificial intelligence
ML and data driven approaches in exploration and production operations to achieve more accurate
Machine learning predictions which improves decision-making processes. This research provides a review to examine the
Upstream operation use cases and application of AI and ML techniques in petroleum industry for optimization of the up-
Oil and gas industry stream processes such as reservoir studies, drilling and production engineering. The challenges related to
Petroleum systems routine approaches for prognosis of operative parameters have been evaluated and the use cases of
Decision-making performance optimizations through employing data-driven approaches resulted in enhancement of
decision-making workflows have been presented. Moreover, possible scenarios of the way that artificial
intelligence will develop and influence the oil and gas industry and how it may change it in the future
was discussed.
© 2022 The Authors. Publishing services provided by Elsevier B.V. on behalf of KeAi Communication Co.
Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/
by-nc-nd/4.0/).

1. Introduction production, the implemented methods fall into three categories

which are mathematic-based, physic-based, and empirical ones.
Newer reservoirs for exploration of hydrocarbons are encoun- While the mathematical methods depend on mass, momentum
tered in deeper depths and in farthest locations. As discussed by and energy conservation equations, the empirical approaches
Lantham (2019), the need for petroleum related materials is depend on past observations and experiments. Due to drawbacks of
increasing daily. Hence, the companies require to take actions for the mathematical methods at different operational situations and
production optimization, cost reduction, and environmental im- inaccurate results of empirical methods, the use of these ap-
pacts of hydrocarbon production. It is not possible to achieve these proaches needs several simplifying assumptions, which results in
objectives through using conventional approaches for exploration, their improper performance in handling complicated relationships,
production and management of hydrocarbon resources. However, noises, and incomplete data. In exploration and production oper-
with new emerging strategies and modeling methods, companies ations, huge volumes of data are created in various daily processes.
can gain more profit by utilizing data-driven technologies. Balaji These databases can be used in data driven methodologies and big
et al. (2018) sated that in classical approaches of hydrocarbon data interpretations, to achieve useful decision-making strategies.
The profit of using these models is enhancement and optimization
of hydrocarbon production (Hamzeh, 2016).
The use of soft computing methodologies has demonstrated to
* Corresponding author.
be hopeful in handling various difficult problems for many oilfield
E-mail addresses: [email protected] (S. Bahaloo), Mmehrizadeh@khazar.
org (M. Mehrizadeh), adelnajafi[email protected] (A. Najafi-Marghmaleki). related processes. Various studies have proved that the AI and ML

https://doi.org/10.1016/j.ptlrs.2022.07.002
2096-2495/© 2022 The Authors. Publishing services provided by Elsevier B.V. on behalf of KeAi Communication Co. Ltd. This is an open access article under the CC BY-NC-ND
license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
S. Bahaloo, M. Mehrizadeh and A. Najafi-Marghmaleki Petroleum Research 8 (2023) 167e182

methods grasp useful answers for operation related difficulties

currently being faced in the industrial operations and can effec-
tively assist in addressing the problems associated with processing
and interpretation of huge databases (Mohaghegh, 2016; Amini and
Mohaghegh, 2019).
Mohammadpoor and Torabi (2018) conducted an investigation
and recognized that more than 80% of the value chain employment
of big data is mainly in petroleum operations. Saputelli et al. (2003)
and (Saputelli et al., 2003) pointed out that big data is growingly
becoming a necessary tool for oil and gas industry. He reviewed the
use cases of AI and ML methods in operations resulted in better
decision makings for petroleum related operations. In the present
work, various ML approaches of relevance to petroleum industry
will be reviewed. Then, the use cases of these algorithms in
different sections of petroleum industry and the ways they can be
utilized to extract useful decision-making information from avail-
able databases will be discussed. We will review the application of
these methods within prediction of petrophysical features and Fig. 1. Representation of single neuron operation process (Rahmanifard and Plaksina,
reservoir studies, drilling and production optimization. 2019).

2. Artificial intelligence methods

full appearance of all the training datasets to the model within the
2.1. Artificial neural network (ANNs) training phase is named an iteration or epoch (Haykin and
Lippmann, 1994). The main aim of the training process of this
ANNs have been developed to reproduce the operation of the network is the optimization of the cost function. In another word,
processing neurons in the brain of humans. Their core characteristic the cost function should be minimized during the training process.
is that they are able to recognize and handle extremely complicated The schematic representation of the framework of MLP is depicted
nonlinear relations among various parameters (Ali, 1994). In addi- in Fig. 2. Various training algorithms are proposed including
tion, they are able to gain knowledge and extend it in situations LevenbergeMarquardt, back-propagation (BP), resilient propaga-
that they exhibit acceptable outputs in the unseen testing data. In tion, etc. (Haykin and Lippmann, 1994), to facilitate the training
addition, they are no limitations on ANNs in terms of hypothesizes phase of these networks. BP is the most favorite one for using MLP
of the original method. Overall, they exhibit robust performance in networks. It befits from the chain rule to characterize the impact of
addressing various forms of tasks and problems, such as approxi- individual weight terms within the structure of the network toward
mation of non-linear functions, recognition and classification of to the cost function's derivative. Its execution process initiates
existing patterns, optimizing the problem solution and automatic through calculating the cost function derivatives in the most pre-
control. The main component of these networks is the units called ceding network layer and next it propagates the computed de-
neurons (Hush and Horne, 1993). The neurons reproduce the rivatives in reverse direction toward the input layer (Haykin and
operation of the neurons within the brain. The schematic repre- Lippmann, 1994). However, this method exhibits various disad-
sentation of a neuron is displayed in Fig. 1. There are various neu- vantages along with itself, including the slow convergence, stuck on
rons in the skeleton of ANN associated together by influent local minima and the choice of proper training speed. A solution
directions (lines) (Hush and Horne, 1993). Their structure should be suggested to alleviate the training rate issue during this algorithm
trained in order to learn to execute different tasks by tuning is to use a parameter called momentum to update the weight terms.
influent directions between the neurons. Multilayer perceptron Hence, despite changing the weight terms during the training
(MLP) networks are the widely utilized sub-branch of ANNs. A MLP process by using the derivative of cost function, the previous
have a structure composed of parallel layers that each layer holds weight terms also contribute to determine the values of new
several neurons, and individual layers are totally linked to subse- weights as training process proceeds. The main drawback of this
quent layers. Depending on the type of connection among layers,
MLP networks are categorized into feed-forward and feedback
types. Feed-forward networks consists of layers which pass data in
a one-sided path from input to output layer while feedback types
allow interconnection of data in both internal and external di-
rections, which is possible through establishing specific loops
within the structure of the model. In addition, in another classifi-
cation, MLP networks fall within supervised and unsupervised
frameworks according to the type of their training algorithm (Hush
and Horne, 1993). Overall, there are three major layers in MLP
structure named input, hidden, and output ones. The input layer
feeds the required data into the network. The task of hidden layer
that is the most crucial layer, is to map the received data from input
layer to output layer. The process terminates in the output layer
that determines the network's output. The core stage in the
execution of ANNs is the training phase, where the variables of the
network (weight terms and biases) are adjusted within an iterative
procedure of adaption with the domain that the model is applied.
The training phase lasts until a minimum error is attained. A single Fig. 2. Schematic representation of ANN structure.

168
S. Bahaloo, M. Mehrizadeh and A. Najafi-Marghmaleki Petroleum Research 8 (2023) 167e182

process is that it needs to optimize two variables that are the rate of SVM has been implemented in various areas of research and sci-
training and the momentum term. In conventional adaptive-based ence. Consider the problem is associated with a dataset of input and
training techniques, the training rate and derivates impact on the output parameters. In SVM, the hyperplanes that split different
magnitude of the weight. Although the adaptation process of the classes are calculated in a way to maximize the distance of the
training rate is done precisely, it is considerably influenced with the hyperplane from the closest points of different classes. When the
unpredictable performance of the derivative itself. Riedmiller and splitting hyperplane is non-linear in the initial input domain, it is
Braun (1993) proposed a new technique for training process of possible to generate linearized hyperplanes in the multi-
these networks. The method directly adjusts the weight step ac- dimensional feature domain by applying specific functions known
cording to locally calculated gradients, where the optimization as kernels to input domain. The conceptual visualization of SVM is
procedure is carried out according to the order of dimensional represented in Fig. 3. In order to conduct regression analysis by
derivative signs in weight domain. This technique differs from SVM, the following equation is constructed by the model (Cortes
other techniques in that the weight adjustment is independent to and Vapnik, 1995):
the gradient size. This approach obtains the optimum weight
through utilizing a separate adjustment parameter Dij, which y ¼ wT 4ðxÞ þ b (1)
specifically evaluates the size of the updated weight. The Lev-
In which w and b denote the weight terms and bias values. In
enberge Marquardt (LM) algorithm was originated based on the
order to obtain w values, the following function should be mini-
Newton's optimization approach (Riedmiller and Braun, 1993;
mized subject to inequality constraints:
Mehrizadeh et al., 2020). This technique differs from BP method in
the way that it uses the derivatives to re-calculate weight terms. ( )
1 2 XN 
*
The main idea of Newton's approach is: min w þ C xi þ xi (2)
Newton's approach uses an expansion of the second-order 2 i¼1
Taylor series. Ak denotes the second derivative matrix of the cost
function. Newton's approach usually provides faster convergence n o
compared to the steepest descent approach. However, calculation yi  wT 4ðxi Þ þ b  j þ xi i ¼ 1; 2; :::; N
(3)
of the derivative matrix for neural networks is challenging and xi ; x*i 0 i ¼ 1; 2; :::; N
computation demanding. The LM approach was proposed to alle-
viate this problem by by-passing the need for determining the In which j is the function approximation term, xi , x*i are slack
derivate matrix for Newton approach while maintaining the speed variables and C is an adjustable parameter. Finally, the SVM solves
of training as before (Hagen et al., 1996). The major drawback of the the regression equation by introducing Lagrange multipliers (di d*i )
LM approach is the requirement for keeping large matrixes of
as follows:
adjustable weights and parameters during its calculations and
convergence process. However, a method was proposed which al- X
N  
lows LM not to compute and store the complete matrix to alleviate yðxÞ ¼ di  d*i kðx; xi Þ þ b (4)
the storage problem. The databases for implementation of ANNs i¼1
are usually categorized into train, validation, and test sets. The aim
In which k (x, xi) is called a kernel type function.
of train portion of data is training the model and optimize the
Least square SVM (LSSVM) is another type of SVM which utilizes
structure of the model. The portion of data used for validation aims
square form of errors in the formulation of cost function and con-
to control the model performance within the training phase. These
straints with equality type rather than positive error terms and
data are not considered during the optimization of the structure of
inequality type of constraints. Hence, LSSVM regression is
the network. The portion of data used for test set aims to check the
performance of fully developed network in prediction of the un-
seen data. These data are utilized as an index to evaluate the
generalization capacity of the completed model when it is facing
new input data. They are generally utilized to check how well the
network can reproduce the overall trend of data and avoids
memorization of data points and overfitting. Two methods can be
utilized during the training process of these models which are
batch and stochastic (sequential) approaches. During stochastic
method, the optimization of network parameters is done when a
portion of training data is introduced to the network. In batch
method, the optimization of weight terms occurs when the total
training set is introduced to the network in individual training it-
erations. The sequential method offers faster performance because
it needs less available space for storing the weight terms. However,
the batch method offers better preservation of convergence criteria
during optimization process. Moreover, various training techniques
such as use the batch method during their computation process.

2.2. Support vector machines (SVMs)

SVM is a branch of the ML methods and is a helpful technique

implemented in data regression and classification by mapping
them to disjointed future-spaces in both the existing input domain
in linear problems and multi-dimensional feature domains for
Fig. 3. Visual representation of SVM operation (Choubey and Karmakar 2020).
nonlinear regression and classification problems. In recent years,
169
S. Bahaloo, M. Mehrizadeh and A. Najafi-Marghmaleki Petroleum Research 8 (2023) 167e182

conducted through linearizing the problem instead of using  

quadratic programming (Kumar et al., 2018). Recently, the LSSVM k xi ; xj ¼ m1 kRBF þ m2 kPoly þ m3 kline (12)
has gained popularity and successfully being utilized in various
In which mi is an adjustable factor associated with RBF, Poly and
applications (Jiang et al., 2018; Kumar et al., 2018; Li et al., 2020;
linear kernels which are between 0 and 1. By adjusting mi values,
Song et al., 2020).
this type of combined kernel function provides optimized features
Considering the training data as (xi, yi) | xi 2 Rd, yi 2 R, in which
for various input parameters. The use of this type of integrated
i ¼ 1, 2, …, N, N denotes the number of data points and d is the
kernel functions the model will be able to better handle the
number of input parameters, xi and yi denote the input and target
nonlinear elements and obtains the solution of linearized equations
arrays. Utilizing a nonlinear function 4(x), a higher dimension
more efficiently.
domain will be considered to project the inputs onto it and perform
the regression computations. The mathematical formulation of this
regression analysis is expressed as follows: 2.3. Fuzzy logic (FL)

y ¼ wT 4ðxÞ þ b (5) Fuzzy logic is identical to daily experiences because it deals with
decisions made by human. The concept of fuzzy inference system
In which w and b denote the weight terms and bias values, being (FIS) is quite useful in balancing and controlling the incomplete
predicted within the algorithm computations in Eq. (6). observations (Bello et al., 2016). The advantages and drawbacks of
FL are as follows:
1 1 XN
C ¼ wT w þ g e2 (6)
2 2 i¼1 i (a) Advantages: The mathematical tools of FL are helpful in
translation and regulating the human decision-making into
appropriate expressions which could effectively be applied
yi ¼ wT 4ðxi Þ þ b þ ei utilizing computer; FL networks are able to process arbitrary
non-linear and complex problems; FL techniques are
In which g > 0 is a term called penalty multiplier and ei repre-
important in solving complicated sequential problems that
sents the error values.
are difficult or impossible to be addressed by even advanced
To make the computation process more straightforward,
mathematical tools; FL techniques are effective and gain
Lagrange formulation is developed as below:
popularity in incomplete knowledge situations; FL methods
are proved to be effective in addressing reasoning problems.
1 1 XN XN  
Lðw; b; e; aÞ ¼ kwk2 þ g e2i  ai wT 4ðxi Þ þ ei þ b  yi (b) Disadvantages: FL has several drawbacks including the spe-
2 2 i¼1 i¼1 cialists in this field who contribute in handling problems by
(7) FL, face difficulties to organize their judgements as the
problems are associated with various uncertainty and biases;
Where ai terms denote the Lagrange factors. By using the partial another issue with FL is that increasing the fuzzy subsets for
derivatives of u, e, b and a, and elimination of u and e terms, the individual input parameters results in considerable increase
computation outcomes are as below: of the required rules. This can cause increase in uncertainty
of network. Despite the existing disadvantages, FL method
X
N has been efficiently utilized for solving complex problems in
yðxÞ ¼ wT 4ðxÞ þ b ¼ ai kðx; xi Þ þ b (8) various research fields (Park et al., 2010).
i¼1
The drawbacks of FL and other AI methods resulted in more
In which k (x, xi) is called a kernel type function with positive
attempts in finding ways to decrease the influence of these draw-
values.
backs. One important achievement was development of hybridized
It has been shown that the prediction and globalization capa-
methods. Hybridized approach benefits from integration of several
bility of LSSVM greatly is influenced by the type of kernels and their
soft computing techniques. These methods attempt to compensate
associated adjustable variables (Sun et al., 2020). Various kernel
the drawbacks of one method with the robustness of other ones.
types such as RBF, Poly and linear are available to be used in
Recently, the hybridized methods gained considerable attention
modeling works and are mathematically represented as follows:
and has been effectively applied to various non-linear and complex
    2 . 2  problems (Acampora and Loia, 2005). FL has been increasingly
kRBF xi ; xj ¼ exp  xi  xj  2s (9) considered in hybridized models. Hybridized approaches using
combinations of FL with AI approaches including neural nets, GA
   d and SVM have been considerably utilized in academic and indus-
kPoly xi ; xj ¼ xi :xj þ 1 (10) trial applications. Popular methods are neuro fuzzy methods, SVM-
fuzzy techniques, and GAeFL hybridized approach. The integration
  of FL and neural nets is named neuro-fuzzy and benefits from the
kline xi ; xj ¼ xi :xj (11)
merits of both methods. Neuro-fuzzy approach integrates the self-
The radial basis function (RBF) is the most used kernel that can learning capabilities of ANNs and the unique knowledge demon-
perform locally approximations with excellent performance, while stration and inference abilities of FL. This technique offers various
polynomial type kernels (Poly) have good overall approximation features such as pattern recognition, classification, regression and
capacity, which offers acceptable capability in handling nonlinear scientific modeling. It has been used in various petroleum related
inputs. Some investigations combine different types of kernel applications. The tendency for gaining more accurate predictions of
functions (RBF, Poly and linear) by using weight terms to benefit actual data is the main concern of hybridized techniques. However,
from strengths of all functions and to increase the estimation the more crucial point is not to use the integrated and hybridized
capability of the LSSSVM model. The formulation of this type of techniques only for improving the performance of the network for a
kernel function can be represented as below: specific dataset but instead to use the power of integration and
170
S. Bahaloo, M. Mehrizadeh and A. Najafi-Marghmaleki Petroleum Research 8 (2023) 167e182

hybridization of models to help the technologies and contribute to

provide universally accurate models for real world problems.
Fuzzy c-means (FCM) is a clustering method that categorizes
individual data points into different clusters based on a specified
degree which is determined through membership functions (MFs).
Bezdek (2013, 2013) proposed the method for the first time as an
enhancement to previous clustering algorithms. The algorithm
demonstrates the process of grouping data in a way to separate the
multidimensional domain to certain clustering sets. FCM is a pop-
ular fuzzy-based clustering method that extends the conventional
c-means technique that is applicable in situations with ambiguity
in determining the exact number of required clusters for catego-
rizing the data (Chiu, 1994). FCM is an important method having the
ability of recognizing patterns in huge historical data from petro-
leum related processes. Identifying and clustering these patterns
will be useful to obtain better understandings of the past, present
and future performance of oil and gas fields that will result in
effective and cost saving field development and production
planning.

Fig. 4. Representation of a simple LR (Choubey and Karmakar, 2020).

2.4. Linear regression

This is a statistics-based modeling approach which builds a X

linear relationship between a parameter and one or several other
GINIðtÞ ¼ 1  ½pðijtÞ2 (15)
independent parameters within a set of data (xi, yi), in which i ¼ l. 2, X
3, …, n. A regression that predicts a parameter (y) based on an in- EntropyðtÞ ¼  pðijtÞlog 2 pðijtÞ (16)
dependent parameter (x) is usually known as Linear Regression
(LR). In situations in which the parameter (y) is predicted based on
ErrorðtÞ ¼ 1  maxpðijtÞ (17)
several independent parameters (x1, x2, x3, …, xn), the regression is
known as Multiple Linear Regression (MLR). The mathematical In which, t denotes a specific node and p(i|t) represents the class
formulations for these terms can be represented as below: probability. Fig. 5 shows a representation of a yes or no based DT to
select one of A and B terms (to decide to select the A or B). The main
Simple LR : y ¼ b0 þ b1 x þ ε (13) purpose of DT is to continuously divide the data to assign them
within their suitable class (as is displayed in Fig. 5) or until there is
Multiple LR : y ¼ b0 þ b1 x1 þ b2 x2 þ ::: þ bn xn þ ε (14) no improvement in the tree structure and accuracy compared to
previous ones. This technique is flexible to alter the training data
In which b0, b1, b2, …, bn denote the variables, and ε represents during operation, however the final output is sensitive to the
the regression error which needs to be determined by utilizing the amount of data in each class. This problem can be resolved by
data points. Maucec and Garni (2019) integrated the LR with the utilizing a well-known approach named Random Forest (RF), which
concept of Analysis of Variance (ANOVA) to describe a new enables the method to analyze multiple trees in parallel by using
modeling approach named generalized linear modelling. Fig. 4 the same dataset. RFs are effective in adjusting the model, con-
shows a basic representation for LR in which the points are trolling the noise, and handling large databanks with various at-
regressed by AB line to predict the Y parameter based independent tributes. This is because in this approach each tree does not pass its
parameter X (the black points represent pairs of (xi, yi) data for i ¼ l, error to other trees. The output of RF can be determined by
2, 3, .., n). comparing the error of various trees. The fittest tree is considered
as the model output.

2.5. Decision tree (DT) 2.6. Bayesian belief networks (BBN)

A DT can be used of to address problems related to regression The concept of BBN method originates from Bayes' theorem
and classification. This method applies Boolean approach to beginning from the query that in what way the likelihood of inci-
represent a single or multiple attributes. There is a specific node dent A is interconnected with an existing incident B (Glickman and
called the root or parental node in the structure of the method. This Van Dyk, 2007). The mathematical representation of this concept
node is connected to several other nodes which are usually known can be expressed by the P(A|B) symbol, which can be re-calculated
as child nodes and operational nodes. This approach is usually through new instances of B. The needed variables are the past
represented in a graphical tree structure to better visualize and likelihood in happening of incident A (P(A)), incident B (P(B)) and
identify the outcomes of the decisions. The algorithm selects the incident B with the condition-dependent likelihood of A (P(B|A) or
parent node from the best attributes and then starts the operation the probability of occurrence under the tested hypothesis (Ji and
process. Then, the remained attributes will be selected and the Marefat, 1995; Kruschke, 2015). Hence, the mathematical repre-
algorithm will continue the operation until there is no attribute, or sentation of the Bayes’ theorem is as below:
when the operation converges to optimum state. The criterion for
selecting the fittest attribute is based on an index. This index can be PðBjAÞ:PðAÞ
any of the three parameters namely Gini index, entropy or classi- PðAjBÞ ¼ (18)
PðBÞ
fication error which are formulated as follows:
171
S. Bahaloo, M. Mehrizadeh and A. Najafi-Marghmaleki Petroleum Research 8 (2023) 167e182

applying the Bayesian concept (Guo and Hsu, 2002).

Netica2 is a tool distributed for generation of Bayesian-based
structures on operating systems. This tool is free for performing
calculations up to 15 parameters. This tool is able to assign
continuous parameters to relevant ranges with designed frame-
works, rather than using non-continuous forms for the input pa-
rameters. The software uses three kinds of nodes named nature,
decision and utility-based nodes. The software generates the tables
of conditional probability during detection of the structure of the
BBN. BBN, uses non-continuous iterations, to determine adequate
number of fittest likelihoods for unseen parameters using available
data (Sample et al., 2016b). First, a comparison between absent
values and logarithm of probabilities will be carried out, next the
performance of the model will be maximized by training the initial
model. This process will be repeated until the convergence of all
likelihoods and achieving stable conditions for logarithm of prob-
abilities. The visualization of this model is depicted in Fig. 6, by
Fig. 5. A simple DT workflow (Choubey and Karmakar, 2020). utilizing a basic BBN framework.

The incident B is named a parental node or a factorization of

incident A, while incident A is named a child node of incident B 2.7. Principal component analysis (PCA)
(Kragt, 2009). The parental nodes in a BBN require to deliver prior
likelihoods, a single or more child nodes and a table for conditioned PCA is an approach for handling and reducing the number of
likelihood showing the linkage between parent nodes and their inputs for problems with large number of input data, to facilitate
children. By implementing the chain rule, a central table of their use and interpretation by other modeling algorithms. The
conditioned likelihood that consists of n parameters, can be ob- techniques for reducing the dimension of a problem select the
tained. In another word, the probability is calculated by taking the suitable features or remove un-suitable ones by linearly combining
values of P(node | parent(node)) and calculating their products using the primary features. PCA linearly transforms the data through
the individual network nodes (X1, …, Xn) as is formulated in Eq. (19) rotation of feature domain, where alignment of data along the di-
(Choubey and Karmakar, 2020): rection of maximum variance is more than other directions.
Kormaksson et al. (2015) provided a review of PCA and applied the
Y
n Y
n method for identifying sweet spots by utilizing well logging infor-
PðX1 ; :::; Xn Þ ¼ PðXi jX1 ; :::Xi1 Þ ¼ PðXi jParentðXi ÞÞ (19) mation. This algorithm transforms an n-dimensional matrix into a
i¼1 i¼1 k-dimensional one where k is lower than n. The transformation is
In which, (Xi) represent parental nodes of Xi parameters. For carried out by building a matrix from the foremost k Eigen vectors
instance, by n parameters, there should be 2n likelihoods. For more based on the covariance Cx values of the original matrix as is
complicated systems, the required likelihoods will drastically rise formulated in Eq. (20). Next, the covariance Cy values of converted
to considerably huge numbers, mainly due to occurrence of mul- matrix is determined by Eq. (21) (Kormaksson et al., 2015; Choubey
tiple states for different parameters (Doolittle et al., 2006). This and Karmakar, 2020):
disadvantage can be alleviated through extracting the feasible re-
lations between parameters by quantitatively measuring control- ðX  XÞðX  XÞT
Cx ¼ (20)
ling alternative node states. However, conversion of qualitative and n1
expert knowledge-based parameters to probability terms is a
difficult and challenging task (Jakoby et al., 2014). By visualizing the
Cy ¼ P:Cx :P T (21)
probabilities in reason-consequence graphical representations
subject to uncertain states, a BBN holding nodes and linkages is able
to utilize expert knowledge to transform different relations into
probabilistic sets. The model permits to arrange existing concepts,
to evaluate feasible possibilities during initialization, execution and
re-calculation stages. The linkages denote the relevant paths and
interrelations of nodes. There is no incoming connection to the
parental nodes and the last child have no output connection
(Sample et al., 2016a). In another word, these techniques owe two
branchesd(i) an acyclic graph with is directed (consists of nodes
and linkages) where the nodes denote the parameters and linkages
denote the tree-based connections between the parameters and,
(ii) Restricted probability distribution lists for evaluation of the
probability-related association between parameters. These pa-
rameters can be non-continuous or continuous. In order to name
this kind of structure as BBN, one need to initialize it by assigning
the prior probabilities to parental nodes and table of conditional
probabilities to child ones. These tables are the core processing
units of BBN for determining the final output values of posterior
likelihoods based on existing parent-child relations through Fig. 6. Illustration of the BBN method (Choubey and Karmakar, 2020).

172
S. Bahaloo, M. Mehrizadeh and A. Najafi-Marghmaleki Petroleum Research 8 (2023) 167e182

2.8. Gradient boosted machine (GBM) and gas industry.

This method is identical to random forest (RF) in that both of 3. Application of AI in different fields
them try to enhance the performance of DTs. However, the
evolvement of trees in RF approach is done in parallel, whereas in 3.1. Reservoir studies
GBM the growth of trees is carried out through received signals of
predecessors, and every tree lies within the updated type of the The process of reservoir characterization involves the quanti-
initial data. In order to recompense the shortages induced through tative allocation of corresponding parameters including porosity,
the undesirable gradients, the tree's cost function is controlled in permeability, fluid property and other characteristics to the reser-
subsequent sequences. Despite slower learning process for boosted voir to gain better understanding about it (Mohaghegh et al., 1994;
trees, they are less likely to suffer from overfitting problems thanks Larki et al., 2018). This process also includes identifying geological
to their reduced size in comparison with RF (Maucec and Garni, structures and uncertainties as well as spatial changes in geological
2019; Choubey and Karmakar, 2020). features. Appropriate execution of characterization process is
important to optimize the production of hydrocarbons. There are
2.9. Big data methods and tools (BDMT) different approaches in characterization of reservoirs in the pe-
troleum industry, but the widely used and best practice approach is
Mohammadpoor and Torabi (2018) as well as Choubey and to take cores from reservoir sections. The data obtained from
Karmakar (2020), provided a detailed review of well-known analyzing the core data are absolute and relative permeability,
methods and main tools for handling of big data which will be porosity, fluid saturations, composition of rocks, size of particles
briefly presented here: and geo-mechanical information. While most of the features help
to examine the reservoir potential for storing and producing fluids,
2.9.1. Hadoop some of them including the composition of rock and geo-
Hadoop is a freely available software which integrates the net- mechanical information aim to conduct the drilling and produc-
works of a large number of processors. It splits large files into sub- tion processes properly. Hu et al. (2017) categorized petroleum
sections and maintains them in disseminated storage. This provides related minerals into minerals susceptible to water, minerals sus-
the option of processing the high-dimensional information in ceptible to salinity, minerals susceptible to acid, minerals suscep-
parallel to perform tasks such as interpretation of logging infor- tible to alkaline and minerals susceptible to velocity of flowing
mation and detection of anomalies in tools functioning using sen- fluid. For example, understanding the composition of mineral,
sors in the petroleum industry. mainly the type of clay and distribution of the desired zone will
help to formulate safer fluids for drilling operations as well as aim
2.9.2. Mongo DB to design practical treatments for stimulation or fracturing opera-
This is a NoSQL and document-orientated database tool. Its main tions. In addition, knowledge of geo-mechanical features of rocks
usage is in cases needing active queries and also in cases in which will help in designing the true density values of fluids in drilling,
defining indexes are preferred. Its main usage is for storage of dy- better control of stability conditions for wellbore and better control
namic data such as drilling rig information during exploration or of sand production. Hence, extracted knowledge from interpreta-
production in onshore or offshore conditions. tion of core data will help to better characterize a reservoir and will
be useful for optimization of drilling and production processes.
2.9.3. Cassandra Currently, it is not possible and practical to conduct coring surveys
Similar to Mongo DB, this is also a distributed and NoSQL in all drilled wells because taking core from reservoirs is time
database tool. It is mainly utilized in situations where large data- demanding and costly. This induces uncertainties for development
bases need to be processed and stored on servers. When gathering of a reliable three-dimensional static model of the reservoir as the
information from huge arrays of sensors, it is required to store the available core data are considerably fewer than the existing wells.
information in several servers. This tool facilitates the process of Characterization of a reservoir is a difficult and complex task
categorizing several databases and simultaneously storing them because of the existing non-linearity and heterogeneity in various
from different sensors in various servers. parameters. These challenges can effectively be handled efficiently
and precisely through utilizing the soft computing methods such as
2.9.4. R programming ANNs, FL, SVM, GA, etc. Over the past few years, these techniques
R is a new functional-based language for programmers that were successfully implemented within the petroleum industry and
makes it easy to perform complicated statistics-based computa- it was shown that they are able to provide accurate predictions of
tional processes as well as generating good-quality visualizing uncertain parameters including permeability and facies in wells
outputs. It is mainly utilized for statistically analyzing the data and with log data and without core data (Cuddy 2000; Hambalek and
software development for data driven interpretation. Gonzalez, 2003; Khoshmardan et al., 2021).
However, soft computing methods do not need any simplifica-
2.9.5. Datameer tions and has shown to perform well to identify various rock types.
It is a spreadsheet-based programming technology, which offers The offered regression techniques by soft computing and AI ap-
inherent tools for data interpretation, manipulation, and modeling proaches have been shown to be better than the classical regression
facilities. It has the ability to combine, interpret, and display data operations mainly in situations where highly uncertain data are
irrespective of source, configuration, and size. dealt with. Chen et al. (1995) used a regression approach using FL to
evaluate the variables of the Archie equation in a situation where
2.9.6. BigSheets the observed and modeled data differences were attributed to
This is a spreadsheet free web app which is identical to Data- various sources including errors in experimentations and ambigu-
meer in functionality. The IBM developed this tool which enables ities of parameters in the network (Chen et al., 1995). Fang and Chen
users to split large data into useable commercial outputs. It is (1997) used FL to estimate porosity and permeability of sandstone
frequently applied in cleaning, interpreting, and displaying large rocks by using 5 composition and texture related input. They
databases. Data like sales by year by non-technical users in the oil clustered the experimental data into various categories by FCM. The
173
S. Bahaloo, M. Mehrizadeh and A. Najafi-Marghmaleki Petroleum Research 8 (2023) 167e182

sandstone features of individual clusters were utilized to extend into comparison with other classical approaches namely expo-
the fuzzy rules. There were a single if-then rule and five input pa- nential and multilinear regression approaches. The exponential
rameters associated with individual clusters. The rules were approach gave a considerably small R2 value of 0.43 for estimation
applied to conduct estimations in a linguistic manner through of core permeability data which indicates that the method is not
integrating the output of individual rules. Another field of reservoir applicable for the heterogenous reservoirs. Multilinear regression
characterization through which the AI approaches have been uti- approach gave a more accurate performance compared to the
lized is the classification of reservoir anisotropic features. Lian et al. exponential approach with R2 value of 0.89 for porosity and 0.53 for
(1998) integrated FL and geo-statistical techniques and obtained permeability. However, the method overpredicts high data and
the existing spatial relation between reservoir parameters. They underpredicts the small data. Owing to the nonlinearity and het-
utilized the outcomes to estimate the distribution of permeability erogeneity of data points the conventional methods were not
within the reservoir. Nikravesh et al. (2001) utilized k-means effective in prediction of actual core data. Finally, the modeling task
clustering, neural nets and FCM approaches for field characteriza- was effectively handled by utilizing ANN with R2 value of 0.99 for
tion. They utilized the approaches to cluster three dimensional both porosity and permeability data. The merit of soft computing
seismic features and find reasonable relationships among various methods over conventional approaches in characterization of res-
clusters and the production log. They succeeded to find the best ervoirs is that they are able to handle limited and incomplete da-
positions for planned wells according to the useful information tabases that is an important difficulty during the primary phases of
extracted from individual clusters. Nikravesh and Aminzadeh reservoir exploration. Soft computing methods do not need any pre
(2001a) as well as Nikravesh and Aminzadeh (2001a), utilized supposition on the complexities of the reservoir to develop an
neuro-fuzzy approach to optimally propose a number of rules for effective model from measured data (Shokir and Engineering,
nonlinear correlation between permeability and several other pa- 2006). AI and computer-based models are important tools that
rameters. They extracted the rules by utilizing training data and offer the potential of integrating a huge amount of input data and
implemented the model to estimate permeability values by using being utilized in reservoirs having many drilled wells in consider-
new databases. They obtained accurate predictions by the proposed ably fast computation process with nearly automatic execution
model. The work also revealed that the model detected existing process. Reservoir characterization is an important way of
relations between input data. Cuddy (2000) focused on the appli- describing hydrocarbon reservoirs in order to obtain the highest
cation of FL in extracting lithofacies information and estimation of profit within their production lifetime. Obtaining the highest profit
permeabilities in different wells without core data by utilizing a requires to take into consideration the intrinsic error related to all
network trained on wells with core data. They compared the results measurement, recognitions of uncertainty coming from the heter-
of FL and ANN and observed better performance for the FL model. ogenous reservoirs as well as the capability to handle gaps and
Saggaf et al. (2003) predicted the distribution of porosity from deficits in available information. AI-based techniques have per-
seismic features in a field by utilizing single hidden layer ANN with formed efficiently robust regarding these necessities. However,
regularized BP training algorithm. This technique could overcome future works need to be directed toward the application of com-
the overfitting problem that is usually encountered in conventional bination of AI methods to compensate the shortcomings of in-
BP algorithms. The approach provided accurate prediction and dividuals techniques with the robustness of other AI techniques
fewer adjustment of network variables at the expense of some including ANN, FL, SVM, GA, etc. Despite implementation of several
computational difficulty. Saggaf and Nebrija (2003) summarized hybrid models, their capabilities in identifying helpful relations
the merits of FL in identifying different rock types by using well log among different types of reservoir parameters and conducting ac-
data. Their analysis revealed that FL can estimate flow units and curate extensions far from drilled wells have been rarely studied.
rock types in wells with no core data with high precision in com- Considering the huge volume of associated databases in petroleum
parison with classical approaches. This approach can greatly save industry which has been gathered during many years, in-
coring expenses and time for the industrial operations. Ilkhchi et al. vestigations on in what way to consider this information in current
(2006) utilized data from real offshore gas wells to build FL and future applications by utilizing hybridized AI techniques that
frameworks for the studied field. They used FCM for rock typing benefits from various soft computing methods will reduce the in-
analysis by utilizing core porosity and permeability information. vestment risks and will offer clearer insights for enhanced char-
They used TakagieSugeno-Kang type fuzzy inference system (FIS) acterization of reservoirs for in various fields.
to estimate the permeability data. Their outcomes revealed that FL Soumi et al. (2018) and Chaki et al. (2018) utilized various pre/
method can be successful applied for estimation of core derived post processing signals for the estimation of reservoir lithological
permeability data. Al-Bulushi et al. (2009) utilized ANN framework characteristics. They applied three types of AI methods including
to predict water saturation and distribution of fluids in a reservoir. ANN, ANFIS, and SVR to model the properties by integration of
The input data for development of the model were a set of well log seismic and well log data. Table 1 provides a more detailed review
data. It was observed that the highest influential parameter in of application of various AI and ML techniques in reservoir studies.
prediction of water saturation was the resistivity log. The highest
influential parameter in prediction of fluid distributions was the 3.2. Production
distribution of pressure. The distribution of pressure was obtained
from core information. They achieved a R2 and RMSE value of 0.91 Hydrocarbon production prediction is becoming very crucial
and 0.025 for prediction of water saturation, and for distribution of with respect to planning and development of projects involving
fluids, they obtained RMSE value of 0.05 for well log data and 0.046 environment and financial regulations, as well as commissioning
for core data, respectively. Hamidi et al. (2010) suggested that the and de-commissioning of facilities. The precise prediction of the
available approaches for classification of rocks exhibit sudden performance of producing wells can deliver roadmaps for produc-
transitions that sometimes do not exist in many situations. tion optimization, artificial lifts usages, detecting the requirement
Moghadam et al. (2011) predicted the porosity and permeability for work-over processes, stimulation of wells, design of facilities,
in a reservoir by utilizing ANN and candidate well log data. Six log and planning the enhanced oil recovery processes (Najafi-
parameters were used for their work according to feature extrac- Marghmaleki et al., 2018). Oilfield production optimization is gov-
tion methods, regression analysis, and analysis of correlating pa- erned through the nodal analysis process to evaluate the optimum
rameters. ANN outcomes based on the well log information was put operational conditions. Large volumes of data are produced during
174
S. Bahaloo, M. Mehrizadeh and A. Najafi-Marghmaleki Petroleum Research 8 (2023) 167e182

Table 1
Summary of AI and ML applications in reservoir studies.

ML Algorithm Reference Size of Data Predicted Parameter Result

ANFIS coupled Huang et al. 563 data points from 4 wells Permeability The GA helped to improve the accuracy of ANFIS with R2
with GA (2001) value of 0.71
ANN Huang et al. Not Given CO2/oil MMP The model provided accurate outcomes with R2 value of 0.99
(2003) for training and 0.936 for testing
ANFIS Lim and Kim 47 data points of an offshore well Porosity and Permeability The model predicts the core data better than conventional
(2004) approaches with overall R2 value of 0.9993 and 0.9998 for
porosity and permeability data, respectively.
ANFIS Shokir (2006) 504 data points from two wells Permeability The model performed well with R2 value of 0.985
ANN, SVM El-Sebakhy 782 data points PVT properties (bubble point The SVM provided better prediction results compared to ANN
(2009) pressure and oil formation and empirical correlations
volume factor)
ANN Al-Bulushi 13 data points for water saturation and Water saturation, The model successfully predicted data with R2 and RMSE of
et al. (2009) saturation height functions of 53 core Fluid distribution 0.91, 0.025 for water saturations and RMSE of 0.046 for fluid
plugs for fluid distribution distribution
GA Chen et al. Simulation model of a reservoir Parameters of WAG process By using the newly developed model, oil recovery can be
(2010) increased by 11.4%
ANFIS Yetilmezsoy 258 data points Water in oil emulsion The model provided predictions with RMSE of 2.0907 and R2
et al. (2011) formation of 0.967
ANN Moghadam 54 data points from two wells Porosity and Permeability Predicted the porosity data with overall R2 value of 0.993 and
et al. (2011) the permeability data with overall R2 value of 0.991
Hybrid of ANN, FL Anifowose 723 porosity data points and 1219 Porosity and Permeability The combination of ANN, FL and SVM provided the most
and SVM and permeability data points from three accurate predictions with maximum R2 value of 0.96 for
Abdulraheem wells porosity data and 0.95 for permeability data
(2011)
ANN, MNN (Tahmasebi 34 core permeability data points Permeability The MNN model performed better than the ANN model. The
and R2 value of ANN and MNN models were 0.94 and 0.99,
Hezarkhani, respectively
2012)
ANN, FL, ANFIS Kenari and Not Given Water saturation ANFIS predictions are better than other models
Mashohor
(2013)
ANN-Fuzzy Wang et al. 231 data point from three wells Porosity The model predicted core data with overall R of 0.9504
(2013)
ANN, LSSVM Chamkalani 4756 data points Gas compressibility factor The LSSVM model presented better outcomes with MSE and
coupled with et al. (2013) R2 values of 0.000014 and 0.999, respectively
CSA
ANN coupled with Zendehboudi 169 data points Recovery factor and cumulative The model showed average error less than 7% and better
PSO et al. (2014) steam to oil ratio in SAGD performance compared to conventional ANN
process
ANFIS coupled Afshar et al. 153 data points Oil bubble point pressure The GA helped to improve the accuracy of ANFIS model and
with GA (2014) provided an overall R2 value of 0.9909
ANN-Fuzzy Zerrouki et al. 420 data points of a single well Natural fracture porosity The model exhibited a R2 value of 0.965. Applying the model
(2014) for prediction of data in another well resulted in R2 value of
0.878
ANN coupled with Xue et al. 61 data points Fracture in core samples The model provided well-stablished results for prediction of
GA (2014) fractures in core samples
ANFIS Aïfa et al. 690 data points Porosity and permeability The model exhibited accurate results with R2 values of 0.99
(2014) and 0.97 for porosity and permeability, respectively
ANN coupled with Amiri et al. 2200 data points from 12 wells Water saturation Using ICA improved the performance of ANN model and
ICA (2015) increase R2 value from 0.92 to 0.96
ANN Singh et al. Not Given (data from Kansan gas field) Porosity Acceptable prediction of porosity data with overall R2 value of
(2016) 0.9714.
ANN Zhang et al. 64 data points Wettability The model exhibited accurate predictions with average error
(2016) less than 5%
ANN Gholanlo et al. 564 data points from an oilfield Water saturation ANN predictions are better than the empirical approaches
(2016)
ANN, ANFIS Khan et al. 150 data points Water saturation ANFIS predictions with testing R2 of 0.956 are better than
(2018) ANN
SVM coupled with Zhong and 94 data points of six wells Porosity The model presented robust performance with R2 value of
PSO Carr (2019) 0.914
ANN Bruyelle and Simulation model of Brugge field Dynamic reservoir parameters The ANN showed much predictive capability than other
Gue
rillot such as pressure, oil and water methods
(2019) rates
ANN Amini and CO2 sequestration simulation model Dynamic reservoir parameters The proposed method can produce target data with less
Mohaghegh including pressure, saturation computational cost and time compared to the numerical
(2019) and CO2 mole fraction simulator
ANN, LSSVM Miah et al. 182 data points Water saturation The ANN trained by LM method and the LSSVM coupled with
integrated with (2020) CSA and RBD kernel provided the most accurate predictions
CSA
ANN, SVM and Zhang et al. 253 core permeability data points Porosity, FZI, FZI* The hybrid PSO-SVM method provided the most accurate
their (2021) predictions compared to other methods
integration
with GA and
PSO
(continued on next page)

175
S. Bahaloo, M. Mehrizadeh and A. Najafi-Marghmaleki Petroleum Research 8 (2023) 167e182

Table 1 (continued )

ML Algorithm Reference Size of Data Predicted Parameter Result

ANN, MLR, Chai et al. A full physics reservoir model under History matching and The applied methods provided improved history matching
XGBoost, GA, (2021) waterflooding optimization of field and field development workflows
PSO development

monitoring the conditions of producing wells in real time mode. gathered data from temperature and pressure measurements,
These data include real-time response of pressure and flowing rate completions, and production logs for prediction of gas production
data which are utilized for graphical and analysis purposes for in a horizontal well after stimulation operations. They utilized ANN,
improved decision-makings. Generally, regression and simulation SVM and RF methods and found that the RF was the fittest algo-
approaches are utilized to estimate future production of wells ac- rithm with faster computation time and convergence. Wang and
cording to their pre-existing observed production rates, which is a Chen (2019) investigated the production conditions of wells being
time consuming and computational demanding task. On the his- hydraulically fractured by using soft computing methods. They
torical production data. To implement this process in more effort- utilized various AI techniques to predict first year production data
less and cost saving ways, AI approaches can be used. This section of wells. They concluded that among various methods, the RF
focuses on the application of AI approaches in production optimi- approach is more accurate thanks to offering more suitable statis-
zation processes within petroleum industry. Alimonti and Falcone tical performance measure parameters. Mahmoud et al. (2019)
(2004) used FL for monitoring the production rate of fluids and provided an ANN technique for optimization of the number of
composition of steam during multiphase flow process. In the work separation steps, pressure and temperature in multiphase sepa-
done by Li et al. (2013), the DT approach was utilized to estimate the rator equipment according to the composition of flowing fluid. The
fluid production rates through taking into account interrelations results revealed that the developed method can be utilized to
between the input parameters. They compared the performance of predict the operational conditions of separators to enhance the
DT and methods and observed that both methods were reliable and resulted fluid quality achieved from surface separators. The work
accurate. However, further analysis of statistical parameters done by Liu et al. (2020) highlighted application of ANN and SVM
revealed the better classification performance of the DT approach. methods for prediction of oil production. Their results showed that
Chakra et al. (2013) used a branch of ANN to estimate cumulative the models were able to reproduce the target data with acceptable
fluid production of an oilfield with few numbers of train data. The accuracy. Crnogorac et al. (2020) utilized a FL optimization tech-
results showed that the implemented method is reliable for pre- nique to identify the best artificial lift scenario for several wells. The
diction of production data for short and long run plans. The out- appropriate artificial lift scenarios were identified and classified for
comes of method according to the performance measure selected wells by conducting sensitivity analysis study. Table 2
parameters were in consistency with the results achieved from the provides a more detailed review of application of various AI and
simulation works. However, the approach needs further improve- ML techniques in production optimizations.
ment to enhance the precision and robustness by involving the
characteristic variables of the reservoir. Yu et al. (2013) proposed 3.3. Drilling operations
SVM approach for detection of failures in submersible pump
operation by using performance-based variables. The developed The process of well drilling is a difficult job as it includes few
method performed well through reducing the rate of misjudgment previous knowledge of the characteristics of the sub e surface
and excellent pattern recognition capability. Kamari et al. (2014) settings and the challenge boosts with progress of drilling to higher
utilized LSSVM method coupled with CSA to estimate unloading depths or deviation of the well path from vertical trajectory. In
gradient pressures in wells under gas lift operation to evaluate the addition, presence of drilling events including differential pipe
optimum rates for production and injection of fluids. Khamehchi sticking, lost circulation, dogleg severity, etc., increases the
et al. (2014), utilized ANN coupled with PSO to estimate the sand complexity and challenges of the process. Recently, AI techniques
production onset. Their results showed that coupling the ANN with were gained popularity for addressing these problems.
PSO increases it performance accuracy for prediction of actual data Wu et al. (1997) conducted a comparison between FL and sta-
compared to back propagation algorithm. Li et al. (2015) utilized a tistical techniques to predict the recovery during the initial stages
soft computing approach for fault detection in wells under pro- of production, primary water injection and during drilling of infill
duction by sucker rod pumps to choose clustering index and the wells. They observed that FL recovery prediction methods con-
highest appropriate scale variable. Choubineh et al. (2017) showed structed by utilizing extracted parameters from statistical method
that hybridized ANN approach with six input variables of specific gives the most accurate outcomes. Garrouch and Labbabidi (2003)
gravity of oil and gas, gas to oil ratio, size of choke, pressure and utilized FL in selection of underbalanced drilling (UBD) candidate
temperature at wellhead estimated the choke critical flow rates of wells. They achieved reliable outcomes by studying two real cases,
liquid at installed at wellhead with acceptable correctness as i.e., the first one was a reservoir under extreme depletion and
inferred from statistical performance measure parameters. Okwu damage and the second one was a fractures reservoir with low
and Adetunji (2018) performed an investigation to optimize the pressure. Malallah and Nashawi (2005) developed an ANN method
distribution cost of several products to different locations by uti- for prediction of fracture gradient of formations during drilling
lizing a hybrid method called adaptive neuro-fuzzy inference sys- operations. They utilized data from 16 wells located in various
tem (ANFIS). Liu et al. (2018) implemented an ANFIS framework geological settings. They concluded that the ANN model performs
coupled with the PSO algorithm to estimate the weight percentage better than other literature models in prediction of the fracture
of unstable asphaltene. The developed method was trained and gradient. Garrouch and Lababidi (2005) used FL in formations
tested utilizing the data points gathered from literature available producing heavy oil from multilaterally completed wells. They
works and the accuracy was monitored by statistical parameters. implemented FL to design geometry of the wellbore and comple-
Bhattacharya et al. (2019) utilized AI approaches coupled with tion methods of the wells. They concluded that in cases with
176
S. Bahaloo, M. Mehrizadeh and A. Najafi-Marghmaleki Petroleum Research 8 (2023) 167e182

Table 2
Summary of AI and ML applications in production optimization.

ML Algorithm Reference Size of Data Predicted Parameter Result

BBN Ghoraishy et al. Data from 59 wells Gel treatment The developed method was able to predict the target data with accuracy over 75%
(2008) performance
ANN, DT Li et al. (2013) 320 oil production Oil production The models are able to acceptable classify the production data with high
data classification classification rate
ANN Chakra et al. Historical production Oil production The model has high capability of predicting the cumulative oil production even
(2013) data with limited data
SVM Yu et al. (2013) Data from a pumping ESP failure The model provides accurate performance recognition capability
unit
LSSVM coupled Kamari et al. 87 data points unloading gradient The model average error in predictions was 1.084% with R2 value of 0.9994
with CSA (2014) pressure
ANN coupled Khamehchi et al. 20 data points onset of sand production Coupling the ANN with PSO increased its performance in prediction of target data
with PSO (2014)
SVM Guo et al. (2015) Data from 100 wells ESP failure The model exhibits acceptable modeling capability and efficiently monitor and
predict the working conditions of wells under ESP production
ANN Jia and Zhang Data from a Production decline The model gives accurate predictions with overall R2 value of 0.9987
(2016) production well analysis
SVM coupled Ebrahimi and Data from a reservoir Gas lift optimization The SVM model performed effectively in comparison with a commercial
with PSO Khamehchi (2016) simulation model simulator
LSSVM coupled Qiao et al. (2017) Historical production Production rate Results indicated that the model has acceptable convergence, robust estimation
with PSO data of an oilfield capability and fast computation process
ANN Choubineh et al. 113 data points Choke critical frow rate The model presented predictions with R2 value of 0.981 and average errors less
(2017) than 2.09%
LSSVM coupled Panja et al. (2018) Simulation data of a Oil and gas recovery, The developed models have very precise recovery prediction capability at various
with PSO, ANN share reservoir Producing gas-oil ratio operational conditions
ANFIS coupled Liu et al. (2018) 428 data points Asphaltene weight The model represented acceptable results with RMSE of 0.49 and R2 value of 0.96
with PSO percentage
ANN, SVM, RF Bhattacharya et al. Production data from Daily gas production rate The RF presented most accurate predictions accuracy of 96%. The ANN model was
(2019) a horizontal well also better than SVM
ANN Mahmoud et al. Data from multistage multistage separator The model can be effectively utilized for prediction of separator operating
(2019) separator parameters conditions
FL Crnogorac et al. Data from different Artificial lift selection The model presented acceptable performance with acceptable matches with
(2020) oilfields previous results of the same oilfield

constraints on operational processes, the fuzzy FL could assist available data from more than 21,000 wells. He concluded that the
drillers to properly decide on the most suitable drilling choices to use of FL for identifying the horizontal wells increased the field
select. Zarei et al. (2008) used the neuro-fuzzy technique in opti- production potential. This increase was explained through the
mizing the drilling locations of wells. The work revealed that the capability of FL to concentrate on locations with high production
use of neuro-fuzzy method effectively saves the interpretation time capacity. Yin et al. (2014) developed an ANN model for early
which is needed for implementation of the technique. warning of kick by using various input parameters. They trained
Drillers usually face several drilling difficulties including dif- their model based on available data of a specific portion of a single
ferential pipe sticking. Prevention of these issues could reduce the well and then applied their model to predict the kick occurrence in
cost of drilling and optimize the process of drilling. Murillo et al. other portions of other wells. They obtained acceptable and reliable
(2009) integrated ANN and FL to examine the risks associated results with learning rates of 91.4% and 86.7%. Toreifi and Rostami
with pipe stuck within the well planning phase and the real drilling (2014) developed an ANN model to estimate the filtrate loss dur-
process. Alireza et al. (2011) applied ANN for prediction of lost ing drilling operations. The developed model utilized different
circulation during drilling operations. They used 18 input variables drilling and mud information (e.g., location of well, depth, pene-
and real data from 32 wells to build and train the network. They tration rate, lithology of formation, rate of pump, pressure of mud,
obtained satisfactory results with correlation coefficient of 0.95, and rheology of mud) as input variables. The model was unfortu-
0.76 and 0.82 for training, validation and testing data sets. nately just applicable for certain reservoirs because of its de-
Shadizadeh et al. (2010) developed an ANN method for forecasting pendency on the location of well and type of formation. Fang et al.
the pipe sticking in drilling wells based on input data for static and (2017) developed an ANN model for estimation of fracture density
dynamic conditions. They observed an accuracy more than 90% for by using several well logs and obtained accurate predictions. They
the implemented method. Jahanbakhshi et al. (2012) also devel- also utilized their network to predict the fracture density in three-
oped ANN and SVM frameworks for prediction of pipe sticking in dimensional model by first propagating the input parameters of the
drilling operations by using 12 input variables and 214 data points. network within the model and then applying the network for
They obtained better predictive performance for the SVM model in prediction of fracture density. Roy et al. (2018) implemented
comparison with the ANN approach. various algorithms such as ANN, FL, ANFIS and multiple linear
In incomplete depleted fields, AI techniques has been used in regression for prediction of fracture gradient using limited data
operations related to the drilling of infill wells. Drilling infill wells is points. They observed that the ANFIS model provide more accurate
useful to help in producing the remaining hydrocarbons and can be result compared to other methods. Liang et al. (2019) employed
subject to optimization methodologies when the natural reservoir conventional and GA trained ANNs for kick detection based on
uncertainty in reservoir is involved using AI methods. Popa (2013) available information from drilled wells. They observed better
examined a scenario in which the FL method was utilized to performance for GA trained network compared to the conventional
identify horizontal wells through utilizing a complicated three- network. Anemangely et al. (2018) conducted a comparison
dimensional fully constructed field model constructed with investigation between ANN trained by PSO and Cuckoo

177
S. Bahaloo, M. Mehrizadeh and A. Najafi-Marghmaleki Petroleum Research 8 (2023) 167e182

Table 3
Summary of AI and ML applications in drilling operations.

ML Algorithm Reference Size of Data Predicted Result

Parameter

ANN Yilmaz et al. 55 data points Bit type The model exhibited satisfactory results in both training and testing
(2002) scenarios
ANN Malallah and 21513 data points of 16 wells Fracture gradient The predictions of the ANN model were better than other correlations
Nashawi (2005) coefficient
ANN Shadizadeh et al. 195 data for dynamic and Pipe sticking The model provided acceptable predictions with average errors less than
(2010) 231 data for static conditions 5%
ANN Mohammadpoor 167 data points Bottom hole The model can predict data with average error of 2.06% and better
et al. (2010) pressure performance compared to existing correlations
ANN coupled with GA and Ashena and 160 data points Bottom hole It was observed that both GA and ACO are able to improve the
ACO Moghadasi pressure performance of ANN
(2011)
ANN Alireza et al. Not Given Lost circulation The model provided precise outcomes with train, test and validation R2
(2011) values of 0.95, 0.76, 0.82, respectively
ANN, SVM Jahanbakhshi 214 data points Pipe sticking The predictions of the SVM model were better than the ANN
et al. (2012)
ANN Toreifi and 1630 data points from 38 Lost circulation The model provided acceptable predictions with R2 value of 0.94
Rostami (2014) wells
PSO Humphries et al. A reservoir simulation model Well placement The model presented acceptable performance in addressing the well
(2014) placement optimization problem.
ANN (Chao et al., 2015) Not Given Fracture pressure The model showed predictions with average errors less than 10%
ANN Vega et al. (2016) 5000 data points Equivalent The proposed method offered a better and faster control and monitoring
circulation density of equivalent circulation density compared to classical methods
GA Farshi (2008) Not Given Well placement The GA method was modified to better detect the optimum well
placement locations
ANN Fang et al. (2017) Not Given Fracture density The model performed well in identifying the existing fractures
ANN Hosseini (2017) Not Given Lost circulation The model provides predictions with R2 value of 0.65
ANN coupled with GA Xie et al. (2018) 1440 data points Kick formation The GA improved the prediction performance and learning time of the
ANN model
ANN, FL, ANFIS Roy et al. (2018) 46 data points Fracture toughness The ANFIS model provided the most accurate predictions
ANFIS Agin et al. (2018) 2400 data points from 61 Lost circulation The model provided reliable predictions with train, test, and validation
wells RMSE of 0.08, 0.09 and 0.15, respectively
ANN Gowida et al. 515 data points Rheological The model predicted target data with R2 value of 0.97 and errors less
(2019) properties of than 6.1%
drilling fluid
ANN coupled with GA Liang et al. 120 data points Drilling overflow The GA helped to improve the prediction performance of ANN
(2019)
DT, ANN, ANFIS Sabah et al. 1900 data points of 61 wells Lost Circulation Although all models exhibited good performance, the DT model
(2019a) provided better predictions.
ANN, SVM Abbas et al. 1120 data points Lost Circulation The performance of ANN was better than the SVM approach
(2019)
ANN Ahmed et al. 245 data points Formation pressure The model performed well in prediction of pore pressure data
(2019a) during drilling
ANN, SVM, FL Ahmed et al. 245 data points Formation pressure All developed methods exhibited R2 values greater than 0.99 and
(2019b) during drilling average errors less than 0.4%
ANN, ANFIS Abdelgawad et al. 2376 data points Equivalent The models were able to predict the data with R2 value of 0.99 and
(2019) circulation density average errors less than 0.22%
ANN Alkinani et al. 100000 data points from Equivalent The model showed acceptable accuracy in prediction of output
(2020) more than 2000 wells circulation density parameter
ANN coupled with PSO, Ashrafi et al. 1000 data 500 m of drilling Rate of penetration Prediction of the ANN coupled by PSO with RMSE of 1.12 were more
GA, ICA (2019a) depth accurate than other methods
ANN Agwu et al. 676 data points Drill cutting settling The ANN provided accurate predictions with train and testing RMSE
(2019) velocity values of 0.03 and 0.09, respectively
ANN, SVM, RF Soares and Gray 7415 data points Rate of penetration The developed models where better than the existing literature models
(2019) and the most accurate model was RF
LSSVM, SVM, ANN Mehrad et al. 2096 data points from two Rate of penetration While all developed models performed well, the LSSVM coupled with
coupled with GA, PSO (2020) wells COA was superior to all of them
and COA
ANN Hou et al. (2020) data points from 50 wells Lost circulation The model provided accurate predictions with accuracy over 90%

Optimization Algorithm (COA) optimization algorithms with a (2019a) investigated the performance of three soft computing
conventional ANN framework for prediction of the rate of pene- models named ANN, ANFIS and decision tree (DT) for prediction of
tration (ROP) and concluded that training the ANN with COA pro- lost circulation during drilling operations by using 1900 real field
vide more efficient solutions to obtain the most precise ROP model. data from 61 wells. Their results showed that by using fewer input
Abbas et al. (2019) utilized operational and geological input pa- variables, ANN and ANFIS provided better performance compared
rameters to develop ANN and SVM models for prediction of lost to DT, however, by increasing the number of input variables, the
circulation using 1120 experimental data. They observed that the performance of DT improved considerably and performed better
SVM model performs slightly better than the ANN model. In predictions than ANN and ANFIS methods. Sabah et al. (2019b)
addition, they also concluded that the accuracy of developed compared ANN trained by PSO with simple ANN, SVM, DT and
models greatly depends on the range of the data which was random forest (RF) in prediction of ROP parameter. The outcomes
considered during the learning phase of the models. Sabah et al. showed that the ANN trained by PSO provides the best results.
178
S. Bahaloo, M. Mehrizadeh and A. Najafi-Marghmaleki Petroleum Research 8 (2023) 167e182

Ashrafi et al. (2019b) have focused on the applicability of MLP and applications for different pumps to develop predictive maintenance
radial basis function (RBF) neural networks in ROP prediction while tools and decide on the optimum operational regions to minimize
training them by various optimization algorithms including ICA, the costs of operation versus production. Most of the pumps have
PSO, GA, ICA, and BBO. The outcomes showed that the MLP model sensors for recording various parameters such as temperature,
trained by PSO provides the most reliable predictions. Elkatatny pressure, flow rates, vibrations, etc. There are various cases of using
(2019) tuned the structure of simple ANN by integrating it with data-aided models to determine the optimum operation regimes,
differential evolution algorithm. Then, proposed a novel ROP model failure prediction and prevention, and saving the time of mainte-
that exhibited more accurate outcomes than existing literature nance. In addition, the cost of investment and associated risks can
models. Table 3 provides a more detailed review of application of be significantly reduced by gathering the data from the wells which
various AI and ML techniques in drilling operations. are under production by various treatment jobs. There are various
reported researches on estimation of the hydraulic fracturing effi-
3.4. Future of AI in petroleum industry ciency and evaluation of injectivity problems by using ML tools
(Orlov and Koroteev, 2019). We think that future development of
Novel pattern identification methods built on the basis of deep methods according to mathematical and programming optimiza-
learning have gained popularity in seismic operations, speeding up tion will bring the potential for describing the entire job operation
the interpretation process with a factor of 10e1000 (Cunha et al., in a smart way. This smart system will enable to choose a specific
2020). It is less likely that the AI methods can hanlde the physical design for treatment of a well.
part (i.e., quantity, fee, and installation location of sensors) of the
seismic process. However, they are beneficial for optimization of 4. Conclusions
the secondary surveys. The mathematical processing systems and
prediction abilities of ML methods will provide better suggestions The AI and ML methods are increasingly gaining application in
on having cheaper secondary surveys with the lowest damage to various sections of petroleum industry as effective approaches
the value of gathered data. In this regard, AI-driven approaches are substituting the conventional methods. Different works have
an efficient method to speed up and, more importantly, to elimi- proven that these methods exhibit predictive performance over
nate the subjective nature of the interpretation procedure (Portugal 90% according to statistical quality assessing evaluations. These
et al., 2018). In addition, apart from future application of ML in methods have been capable of predicting lithological and strati-
seismic processes, a notable section in the well logging data could graphic features, lithofacies, and shale sweet spots as well as
be efficiently generated with ML. In future, this would allow the detection of seismic horizons. The drilling variables such as ROP,
usage of ML in well logging operations. Systems enabling the oil velocity of settled cuttings and differential sticking are efficiently
corporates allocate less resources to the operational section of well predicted and optimized and drilling challenges including lost cir-
logging. Similarly, core analysis operations can also be accelerated culation volume estimates or identification of drilling hazardous
with a similar approach. There are three main areas for using AI in events were minimized. The use of ML and AI methods to improve
reservoir engineering. First one is associated with calculations different sections of reservoir studies by estimating reservoir
performed with conventional reservoir simulators. The simulators characteristic variables including porosity, permeability, water
conduct numerical computations of partial differential equations saturation, bubble point pressure, and recovery factor have been
related to the physics of flow within reservoirs. The calculations are evaluated. In addition, log generation, history matching and flow
done on the three-dimensional grid comprising of from 1 million to rate estimations, have also been addressed in the present work. The
several billions of cells. The calculations are usually time use cases of AI and ML methods in optimization of fluid production
consuming, despite using updated workstations and servers, which by estimation of oil and gas production rates, well treatment and
restricts the possibility of performing many runs. Hence, this re- artificial lift planning which improves the production related de-
duces the optimization capability for suitable planning for devel- cision makings have also been discussed. In addition, possible
opment of the field. One of the most important roles of AI criteria of the way in which artificial intelligence can develop and
techniques is speeding up the reservoir modeling computations. spread within the oil and gas industry in the future was evaluated.
New surrogate models for various reservoirs with modern calcu- The AI and ML approaches have considerably saved operation and
lations engine according to deep neural networks reduces the computation time and reduced the associated expenses by
dimension of the mathematical problem and predicts the time providing efficient predictions and supplying effective solutions for
dependent variables in an order 100 to 1000 times faster than the various operations in the petroleum industry.
common simulators. Hence, AI methods are capable of speeding up
the calculations while maintaining the same functionality. The next
Declaration of competing interest
area is using AI in upscaling (i.e., converting the data obtained from
different scales of geophysical analysis to a single model for
The authors declare that they have no known competing
geological and reservoir studies). The process of upscaling needs a
financial interests or personal relationships that could have
significant level of creativity to be done. There is no unique scien-
appeared to influence the work reported in this paper.
tific procedure to conduct upscaling, and usually reservoir engi-
neers consider several tricks to complete it in such that it sounds
right to them. It brings a significant bias into the simulation model. References
Since there is no unique and correct framework for this process, a Abbas, A.K., Al-haideri, N.A., Bashikh, A.A., 2019. Implementing artificial neural
reservoir engineer may think of enhancing objectiveness by relying networks and support vector machines to predict lost circulation. Egypt J.
on experience and benefiting from a smart system. This can be Petrol. 28, 339e347.
Abdelgawad, K.Z., Elzenary, M., Elkatatny, S., Mahmoud, M., Abdulraheem, A.,
carried out correctly with a deep learning algorithm which trained
Patil, S., 2019. New approach to evaluate the equivalent circulating density
based on several manual upscaling cases. This brings the advan- (ECD) using artificial intelligence techniques. J. Pet. Explor. Prod. Technol. 9,
tages of enhanced objectiveness as well as speeding up and 1569e1578.
reducing the time of upscaling process. Acampora, G., Loia, V., 2005. Using FML and fuzzy technology in adaptive ambient
intelligence environments. Int. J. Comput. Intell. Res. 1, 171e182.
Reservoirs which are under production and green fields are also Afshar, M., Gholami, A., Asoodeh, M., 2014. Genetic optimization of neural network
appealing for AI-driven applications. There are reported ML and fuzzy logic for oil bubble point pressure modeling. Kor. J. Chem. Eng. 31,

179
S. Bahaloo, M. Mehrizadeh and A. Najafi-Marghmaleki Petroleum Research 8 (2023) 167e182

496e502. J. Can. Petrol. Technol. 49, 75e82.

Agin, F., Khosravanian, R., Karimifard, M., Jahanshahi, A., 2018. Application of Chiu, S.L., 1994. Fuzzy model identification based on cluster estimation. J. Intell.
adaptive neuro-fuzzy inference system and data mining approach to predict Fuzzy Syst. 2, 267e278.
lost circulation using DOE technique (Case study: maroon oilfield). Petroleum 6, Choubey, S., Karmakar, G., 2020. Artificial intelligence techniques and their appli-
423e437. cation in oil and gas industry. Artif. Intell. Rev. 1e19.
Agwu, O.E., Akpabio, J.U., Dosunmu, A., 2019. Artificial neural network model for Choubineh, A., Ghorbani, H., Wood, D.A., Moosavi, S.R., Khalafi, E., Sadatshojaei, E.,
predicting drill cuttings settling velocity. Petroleum. 2017. Improved predictions of wellhead choke liquid critical-flow rates:
Ahmed, A., Elkatatny, S., Ali, A., Abdulraheem, A., 2019a. Comparative analysis of modelling based on hybrid neural network training learning based optimiza-
artificial intelligence techniques for formation pressure prediction while dril- tion. Fuel 207, 547e560.
ling. Arabian J. Geosci. 12, 1e13. Cortes, C., Vapnik, V., 1995. Support-vector networks. Mach. Learn. 20, 273e297.
Ahmed, A., Elkatatny, S., Ali, A., Mahmoud, M., Abdulraheem, A., 2019b. New model Crnogorac, M., Tanasijevi
c, M., Danilovi
c, D., Karovi

c Mari ci

c, V., Lekovi

c, B., 2020.
for pore pressure prediction while drilling using artificial neural networks. Selection of artificial lift methods: a brief review and new model based on fuzzy
Arabian J. Sci. Eng. 44, 6079e6088. logic. Energies 13, 1758.
Aïfa, T., Baouche, R., Baddari, K., 2014. Neuro-fuzzy system to predict permeability Cuddy, S., 2000. Litho-facies and permeability prediction from electrical logs using
and porosity from well log data: a case study of Hassi R‫ ׳‬Mel gas field, Algeria. fuzzy logic. SPE Reservoir Eval. Eng. 3, 319e324.
J. Petrol. Sci. Eng. 123, 217e229. Cunha, A., Pochet, A., Lopes, H., Gattass, H., 2020. Seismic fault detection in real data
Al-Bulushi, N., King, P.R., Blunt, M.J., Kraaijveld, M., 2009. Development of artificial using transfer learning from a convolutional neural network pre-trained with
neural network models for predicting water saturation and fluid distribution. synthetic seismic data. Comput. Geosci. 135, 104344.
J. Petrol. Sci. Eng. 68, 197e208. Doolittle, J., Jenkinson, B., Hopkins, D., Ulmer, M., Tuttle, W., 2006. Hydropedological
Ali, J., 1994. Neural Networks: a New Tool for the Petroleum Industry? Society of investigations with ground-penetrating radar (GPR): estimating water-table
Petroleum Engineers. depths and local ground-water flow pattern in areas of coarse-textured soils.
Alimonti, C., Falcone, G., 2004. Integration of multiphase flowmetering, neural Geoderma 131, 317e329.
networks, and fuzzy logic in field performance monitoring. SPE Prod. Facil. 19, Ebrahimi, A., Khamehchi, E., 2016. Developing a novel workflow for natural gas lift
25e32. optimization using advanced support vector machine. J. Nat. Gas Sci. Eng. 28,
Alireza, M., Mohammad, N., Ahmad, A., 2011. Reducing consumed energy while 626e638.
drilling an oil well through a deep rig time analysis. Adv. Petrol. Explor. Dev. 1, El-Sebakhy, E.A., 2009. Forecasting PVT properties of crude oil systems based on
22e31. support vector machines modeling scheme. J. Petrol. Sci. Eng. 64, 25e34.
Alkinani, H.H., Al-Hameedi, A.T.T., Dunn-Norman, S., Lian, D., 2020. Application of Elkatatny, S., 2019. Development of a new rate of penetration model using self-
artificial neural networks in the drilling processes: can equivalent circulation adaptive differential evolution-artificial neural network. Arabian J. Geosci. 12,
density be estimated prior to drilling? Egypt J. Petrol. 29, 121e126. 19.
Amini, S., Mohaghegh, S., 2019. Application of machine learning and artificial in- Fang, J., Chen, H., 1997. Fuzzy modelling and the prediction of porosity and
telligence in proxy modeling for fluid flow in porous media. Fluid 4, 126. permeability from the compositional and textural attributes of sandstone.
Amiri, M., Ghiasi-Freez, J., Golkar, B., Hatampour, A., 2015. Improving water satu- J. Petrol. Geol. 20, 185e204.
ration estimation in a tight shaly sandstone reservoir using artificial neural Fang, J., Zhou, F., Tang, Z., 2017. Discrete fracture network modelling in a naturally
network optimized by imperialist competitive algorithmeA case study. J. Petrol. fractured carbonate reservoir in the Jingbei oilfield, China. Energies 10, 183.
Sci. Eng. 127, 347e358. Farshi, M.M., 2008. Improving Genetic Algorithms for Optimum Well Placement.
Anemangely, M., Ramezanzadeh, A., Tokhmechi, B., Molaghab, A., Stanford University Stanford, CA.
Mohammadian, A., 2018. Drilling rate prediction from petrophysical logs and Garrouch, A., Lababidi, H., 2005. Implementation of a Fuzzy Expert System for
mud logging data using an optimized multilayer perceptron neural network. Multilateral Well CompletioneCase Studies (SPE93930). European Association
J. Geophys. Eng. 15, 1146e1159. of Geoscientists & Engineers.
Anifowose, F., Abdulraheem, A., 2011. Fuzzy logic-driven and SVM-driven hybrid Garrouch, A.A., Labbabidi, H., 2003. Using Fuzzy Logic for UBD Candidate Selection.
computational intelligence models applied to oil and gas reservoir character- Society of Petroleum Engineers.
ization. J. Nat. Gas Sci. Eng. 3, 505e517. Gholanlo, H.H., Amirpour, M., Ahmadi, S., 2016. Estimation of water saturation by
Ashena, R., Moghadasi, J., 2011. Bottom hole pressure estimation using evolved using radial based function artificial neural network in carbonate reservoir: a
neural networks by real coded ant colony optimization and genetic algorithm. case study in Sarvak formation. Petroleum 2, 166e170.
J. Petrol. Sci. Eng. 77, 375e385. Ghoraishy, S.M., Liang, J.-T., Green, D.W., Liang, H.C., 2008. Application of Bayesian
Ashrafi, S.B., Anemangely, M., Sabah, M., Ameri, M.J., 2019a. Application of hybrid Networks for Predicting the Performance of Gel-Treated Wells in the Arbuckle
artificial neural networks for predicting rate of penetration (ROP): a case study Formation. Kansas. Society of Petroleum Engineers.
from Marun oil field. J. Petrol. Sci. Eng. 175, 604e623. Glickman, M.E., Van Dyk, D.A., 2007. Basic bayesian methods. Topic Biostat.
Ashrafi, S.B., Anemangely, M., Sabah, M., Ameri, M.J., 2019b. Application of hybrid 319e338.
artificial neural networks for predicting rate of penetration (ROP): a case study Gowida, A., Elkatatny, S., Ramadan, E., Abdulraheem, A., 2019. Data-driven frame-
from Marun oil field. J. Petrol. Sci. Eng. 175, 604e623. work to predict the rheological properties of CaCl2 brine-based drill-in fluid
Balaji, K., Rabiei, M., Suicmez, V., Canbaz, H., Agharzeyva, Z., Tek, S., Bulut, U., using artificial neural network. Energies 12, 1880.
Temizel, C., 2018. Status of Data-Driven Methods and Their Applications in Oil Guo, H., Hsu, W., 2002. A Survey of Algorithms for Real-Time Bayesian Network
and Gas Industry. Society of Petroleum Engineers. Inference. Join Workshop on Real Time Decision Support and Diagnosis
Bello, O., Teodoriu, C., Yaqoob, T., Oppelt, J., Holzmann, J., Obiwanne, A., 2016. Systems.
Application of Artificial Intelligence Techniques in Drilling System Design and Guo, D., Raghavendra, C.S., Yao, K.-T., Harding, M., Anvar, A., Patel, A., 2015. Data
Operations: a State of the Art Review and Future Research Pathways. Society of Driven Approach to Failure Prediction for Electrical Submersible Pump Systems.
Petroleum Engineers. Society of Petroleum Engineers.
Bezdek, J.C., 2013. Pattern Recognition with Fuzzy Objective Function Algorithms. Hagen, M., Demuth, H., Beale, M., 1996. Neural Network Design. PWS Publishing Co.,
Springer Science & Business Media. Boston.
Bhattacharya, S., Ghahfarokhi, P.K., Carr, T.R., Pantaleone, S., 2019. Application of Hambalek, N., Gonzalez, R., 2003. Fuzzy Logic Applied to Lithofacies and Perme-
predictive data analytics to model daily hydrocarbon production using petro- ability Forecasting Case Study: Sandstone of Naricual Formation, El Furrial Field,
physical, geomechanical, fiber-optic, completions, and surface data: a case Eastern Venezuela Basin. Society of Petroleum Engineers.
study from the Marcellus Shale, North America. J. Petrol. Sci. Eng. 176, 702e715. Hamidi, J.K., Shahriar, K., Rezai, B., Bejari, H., 2010. Application of fuzzy set theory to
Bruyelle, J., Gue
rillot, D., 2019. Proxy Model Based on Artificial Intelligence Tech- rock engineering classification systems: an illustration of the rock mass
nique for History Matching-Application to Brugge Field. Society of Petroleum excavability index. Rock Mech. Rock Eng. 43, 335e350.
Engineers. Hamzeh, H., 2016. Application of Big Data in Petroleum Industry. Istanbul Sehir
Chai, Z., Nwachukwu, A., Zagayevskiy, Y., Amini, S., Madasu, S., 2021. An integrated University.
closed-loop solution to assisted history matching and field optimization with Haykin, S., Lippmann, R.J., 1994. Neural networks, a comprehensive foundation. Int.
machine learning techniques. J. Petrol. Sci. Eng. 198, 108204. J. Neural Syst. 5, 363e364.
Chaki, S., Routray, A., Mohanty, W.K., 2018. Well-log and seismic data integration for Hosseini, P., 2017. Estimation of lost circulation amount occurs during under
reservoir characterization: a signal processing and machine-learning perspec- balanced drilling using drilling data and neural network. Egypt J. Petrol. 26,
tive. IEEE Signal Process. Mag. 35, 72e81. 627e634.
Chakra, N.C., Song, K.-Y., Gupta, M.M., Saraf, D.N., 2013. An innovative neural fore- Hou, X., Yang, J., Yin, Q., Liu, H., Chen, H., Zheng, J., Wang, J., Cao, B., Zhao, X., Hao, M.,
cast of cumulative oil production from a petroleum reservoir employing higher- Liu, X., 2020. Lost circulation prediction in south China sea using machine
order neural networks (HONNs). J. Petrol. Sci. Eng. 106, 18e33. learning and big data technology. Offshore Technol. Conf.
Chamkalani, A., Zendehboudi, S., Chamkalani, R., Lohi, A., Elkamel, A., Chatzis, I., Hu, X., Hu, S., Jin, F., Huang, S., 2017. Physics of Petroleum Reservoirs. Springer.
2013. Utilization of support vector machine to calculate gas compressibility Huang, Y., Gedeon, T.D., Wong, P.M., 2001. An integrated neural-fuzzy-genetic-
factor. Fluid Phase Equil. 358, 189e202. algorithm using hyper-surface membership functions to predict permeability
Chen, H., Fang, J., Kortright, M., Chen, D., 1995. Novel. Approaches to the determi- in petroleum reservoirs. Eng. Appl. Artif. Intell. 14, 15e21.
nation of archie parameters II: fuzzy regression analysis. SPE Adv. Technol. 3, Huang, Y., Huang, G., Dong, M., Feng, G., 2003. Development of an artificial neural
44e52. network model for predicting minimum miscibility pressure in CO2 flooding.
Chen, S., Li, H., Yang, D., Tontiwachwuthikul, P., 2010. Optimal parametric design for J. Petrol. Sci. Eng. 37, 83e95.
water-alternating-gas (WAG) process in a CO2-miscible flooding reservoir. Humphries, T.D., Haynes, R.D., James, L.A., 2014. Simultaneous and sequential

180
S. Bahaloo, M. Mehrizadeh and A. Najafi-Marghmaleki Petroleum Research 8 (2023) 167e182

approaches to joint optimization of well placement and control. Comput. Petrol. Sci. Technol. 29, 2095e2112.
Geosci. 18, 433e448. Mohaghegh, S.D., 2016. Determining the main drivers in hydrocarbon production
Hush, D.R., Horne, B.G., 1993. Progress in supervised neural networks. IEEE Signal from shale using advanced data-driven analyticseA case study in Marcellus
Process. Mag. 10, 8e39. shale. J. Unconvention. oil Gas. Res. 15, 146e157.
Ilkhchi, A.K., Rezaee, M., Moallemi, S.A., 2006. A fuzzy logic approach for estimation Mohaghegh, S., Arefi, R., Ameri, S., Hefner, M., 1994. A Methodological Approach for
of permeability and rock type from conventional well log data: an example Reservoir Heterogeneity Characterization Using Artificial Neural Networks.
from the Kangan reservoir in the Iran Offshore Gas Field. J. Geophys. Eng. 3, Society of Petroleum Engineers.
356e369. Mohammadpoor, M., Torabi, F., 2018. Big Data Analytics in Oil and Gas Industry: an
Jahanbakhshi, R., Keshavarzi, R., Aliyari Shoorehdeli, M., Emamzadeh, A., 2012. Emerging Trend. Petroleum.
Intelligent prediction of differential pipe sticking by support vector machine Mohammadpoor, M., Shahbazi, K., Torabi, F., Qazvini Firouz, A.R., 2010. A New
compared with conventional artificial neural networks: an example of iranian Methodology for Prediction of Bottomhole Flowing Pressure in Vertical Multi-
offshore oil fields. SPE Drill. Complet. 27, 586e595. phase Flow in Iranian Oil Fields Using Artificial Neural Networks (ANNs). So-
Jakoby, O., Grimm, V., Frank, K., 2014. Pattern-oriented parameterization of general ciety of Petroleum Engineers.
models for ecological application: towards realistic evaluations of management Murillo, A., Neuman, J., Samuel, R., 2009. Pipe Sticking Prediction and Avoidance
approaches. Ecol. Model. 275, 78e88. Using Adaptive Fuzzy Logic Modeling. Society of Petroleum Engineers.
Ji, Q., Marefat, M., 1995. Bayesian approach for extracting and identifying features. Najafi-Marghmaleki, A., Kord, S., Hashemi, A., Motamedi, H., 2018. Experimental
Comput. Aided Des. 27, 435e454. investigation of efficiency of MEOR process in a carbonate oil reservoir using
Jia, X., Zhang, F., 2016. Applying data-driven method to production decline analysis Alcaligenes faecalis: impact of interfacial tension reduction and wettability
and forecasting. In: SPE Annual Technical Conference and Exhibition. Society of alteration mechanisms. Fuel 232, 27e35.
Petroleum Engineers. Nikravesh, M., Aminzadeh, F., 2001a. Mining and fusion of petroleum data with
Jiang, Y., Huang, G., Peng, X., Li, Y., Yang, Q.S., 2018. A novel wind speed prediction fuzzy logic and neural network agents. J. Petrol. Sci. Eng. 29, 221e238.
method: hybrid of correlation-aided DWT, LSSVM and GARCH. J. Wind Eng. Ind. Nikravesh, M., Adams, R.D., Levey, R.A., 2001. Soft computing: tools for intelligent
Aerod. 174, 28e38. reservoir characterization (IRESC) and optimum well placement (OWP).
Kamari, A., Bahadori, A., Mohammadi, A.H., Zendehboudi, S., 2014. Evaluating the J. Petrol. Sci. Eng. 29, 239e262.
unloading gradient pressure in continuous gas-lift systems during petroleum Okwu, M.O., Adetunji, O., 2018. A comparative study of artificial neural network
production operations. Petrol. Sci. Technol. 32, 2961e2968. (ANN) and adaptive neuro-fuzzy inference system (ANFIS) models in distribu-
Kenari, S.A.J., Mashohor, S., 2013. Robust committee machine for water saturation tion system with nondeterministic inputs. Int. J. Eng. Bus. Manag. 10,
prediction. J. Petrol. Sci. Eng. 104, 1e10. 1847979018768421.
Khamehchi, E., Kivi, I.R., Akbari, M., 2014. A novel approach to sand production Orlov, D., Koroteev, D., 2019. Advanced analytics of self-colmatation in terrigenous
prediction using artificial intelligence. J. Petrol. Sci. Eng. 123, 147e154. oil reservoirs. J. Petrol. Sci. Eng. 182, 106306.
Khan, M.R., Tariq, Z., Abdulraheem, A., 2018. Machine Learning Derived Correlation Panja, P., Velasco, R., Pathak, M., Deo, M., 2018. Application of artificial intelligence
to Determine Water Saturation in Complex Lithologies. Society of Petroleum to forecast hydrocarbon production from shales. Petroleum 4, 75e89.
Engineers. Park, H.-J., Lim, J.-S., Roh, J., Kang, J.M., Min, B.-H., 2010. Production-system opti-
Khoshmardan, M.A., Mehrizadeh, M., Zand, N., Najafi-Marghmaleki, A., 2021. Ac- mization of gas fields using hybrid fuzzy/genetic approach. SPE J. 15, 417e425.
curate prediction of viscosity of mixed oils. Petrol. Sci. Technol. 39, 351e361. Popa, A.S., 2013. Identification of Horizontal Well Placement Using Fuzzy Logic.
Kormaksson, M., Vieira, M.R., Zadrozny, B., 2015. A Data Driven Method for Sweet Society of Petroleum Engineers.
Spot Identification in Shale Plays Using Well Log Data. Society of Petroleum Portugal, I., Alencar, P., Cowan, D., 2018. The use of machine learning algorithms in
Engineers. recommender systems: a systematic review. Expert Syst. Appl. 97, 205e227.
Kragt, M.E., 2009. A beginners guide to Bayesian network modelling for integrated Qiao, Y., Peng, J., Ge, L., Wang, H., 2017. Application of PSO LS-SVM forecasting
catchment management. Landscape Logic. model in oil and gas production forecast. In: 2017 IEEE 16th International
Kruschke, J., 2015. Doing Bayesian Data Analysis, second ed. Elsevier, San Diego, CA. Conference on Cognitive Informatics & Cognitive Computing (ICCI* CC).
Kumar, L., Sripada, S.K., Sureka, A., Rath, S.K., 2018. Effective fault prediction model Rahmanifard, H., Plaksina, T., 2019. Application of artificial intelligence techniques
developed using least square support vector machine (LSSVM). J. Syst. Software in the petroleum industry: a review. Artif. Intell. Rev. 52, 2295e2318.
137, 686e712. Riedmiller, M., Braun, H., 1993. A Direct Adaptive Method for Faster Back-
Lantham, A., 2019. Why Explorers Are in the Grip of an Existential Crisis: Explo- propagation Learning: the RPROP Algorithm. IEEE International Conference on
ration's Role to Undergo Profound Change as Energy Transition Gains Pace. Neural Networks, pp. 586e591.
Society of Petroleum Engineers. Roy, D.G., Singh, T., Kodikara, J., 2018. Predicting mode-I fracture toughness of rocks
Larki, S.A., Banashooshtari, H., Shokrollahzadeh Behbahani, H., Najafi- using soft computing and multiple regression. Measurement 126, 231e241.
Marghmaleki, A., 2018. Effect of acid number of crude oil on oil recovery of Sabah, M., Talebkeikhah, M., Agin, F., Talebkeikhah, F., Hasheminasab, E., 2019a.
smart water coupled with silica nanoparticles. Petrol. Sci. Technol. 36, 343e349. Application of decision tree, artificial neural networks, and adaptive neuro-
Li, X., Chan, C., Nguyen, H., 2013. Application of the neural decision tree approach fuzzy inference system on predicting lost circulation: a case study from
for prediction of petroleum production. J. Petrol. Sci. Eng. 104, 11e16. Marun oil field. J. Petrol. Sci. Eng. 177, 236e249.
Li, K., Gao, X.-W., Zhou, H.-B., Han, Y., 2015. Fault diagnosis for down-hole condi- Sabah, M., Talebkeikhah, M., Wood, D.A., Khosravanian, R., Anemangely, M.,
tions of sucker rod pumping systems based on the FBHeSC method. Petrol. Sci. Younesi, A., 2019b. A machine learning approach to predict drilling rate using
12, 135e147. petrophysical and mud logging data. Earth Sci. Informatic. 12, 319e339.
Li, J., Chen, W., Han, K., Wang, Q., 2020. Fault diagnosis of rolling bearing based on Saggaf, M., Nebrija, E.L., 2003. A fuzzy logic approach for the estimation of facies
GA-VMD and improved WOA-LSSVM. IEEE Access 8, 166753e166767. from wire-line logs. AAPG (Am. Assoc. Pet. Geol.) Bull. 87, 1223e1240.
Liang, H., Zou, J., Liang, W., 2019. An early intelligent diagnosis model for drilling Saggaf, M.M., Tokso € z, M.N., Mustafa, H.M., 2003. Estimation of reservoir properties
overflow based on GAeBP algorithm. Cluster Comput. 22, 10649e10668. from seismic data by smooth neural networks. Geophysics 68, 1969e1983.
Lim, J.-S., Kim, J., 2004. Reservoir Porosity and Permeability Estimation from Well Sample, J.E., Baber, I., Badger, R., 2016a. A spatially distributed risk screening tool to
Logs Using Fuzzy Logic and Neural Networks. Society of Petroleum Engineers. assess climate and land use change impacts on water-related ecosystem ser-
Liu, Z., Zhang, D., Peng, W., 2018. A Novel ANFIS-PSO Network for forecasting oil vices. Environ. Model. Software 83, 12e26.
flocculated asphaltene weight percentage at wide range of operation condi- Sample, J.E., Baber, I., Badger, R., 2016b. A spatially distributed risk screening tool to
tions. Petrol. Sci. Technol. 36, 1044e1050. assess climate and land use change impacts on water-related ecosystem ser-
Liu, W., Liu, W.D., Gu, J., 2020. Forecasting oil production using ensemble empirical vices. Environ. Model. Software 83, 12e26.
model decomposition based Long Short-Term Memory neural network. Saputelli, L., Nikolaou, M., Economides, M., 2003. Self Learning Reservoir Manage-
J. Petrol. Sci. Eng. 189, 107013. ment. SPE Annual Technical Conference and Exhibition, Denver, Colorado, USA.
Mahmoud, M., Tariq, Z., Kamal, M.S., Al-Naser, M., 2019. Intelligent prediction of Shadizadeh, S., Karimi, F., Zoveidavianpoor, M., 2010. Drilling stuck pipe prediction
optimum separation parameters in the multistage crude oil production facil- in iranian oil fields: an artificial neural network approach. Iranian J. Chem. Eng.
ities. J. Pet. Explor. Prod. Technol. 9, 2979e2995. 7, 29e41.
Malallah, A., Nashawi, I.S., 2005. Estimating the fracture gradient coefficient using Shokir, E.M., 2006. A novel model for permeability prediction in uncored wells. SPE
neural networks for a field in the Middle East. J. Petrol. Sci. Eng. 49, 193e211. Reservoir Eval. Eng. 9, 266e273.
Maucec, M., Garni, S., 2019. Application of Automated Machine Learning for Multi- Singh, S., Kanli, A.I., Sevgen, S., 2016. A general approach for porosity estimation
Variate Prediction of Well Production. Society of Petroleum Engineers. using artificial neural network method: a case study from Kansas gas field.
Mehrad, M., Bajolvand, M., Ramezanzadeh, A., Neycharan, J.G., 2020. Developing a Studia Geophys. Geod. 60, 130e140.
new rigorous drilling rate prediction model using a machine learning tech- Soares, C., Gray, K., 2019. Real-time predictive capabilities of analytical and machine
nique. J. Petrol. Sci. Eng. 192, 107338. learning rate of penetration (ROP) models. J. Petrol. Sci. Eng. 172, 934e959.
Mehrizadeh, M., Nikbin-Fashkacheh, H., Zand, N., Najafi-Marghmaleki, A., 2020. Song, X., Zhao, J., Song, J., Dong, F., Xu, L., Zhao, J., 2020. Local demagnetization fault
Accurate prediction of kinematic viscosity of biodiesels and their blends with recognition of permanent magnet synchronous linear motor based on S-
diesel fuels. J. Am. Oil Chem. Soc. 97, 1083e1094. transform and PSOeLSSVM. IEEE Trans. Power Electron. 35, 7816e7825.
Miah, M.I., Zendehboudi, S., Ahmed, S., 2020. Log data-driven model and feature Sun, S., Fu, J., Zhu, F., Du, D., 2020. A hybrid structure of an extreme learning ma-
ranking for water saturation prediction using machine learning approach. chine combined with feature selection, signal decomposition and parameter
J. Petrol. Sci. Eng. 194, 107291. optimization for short-term wind speed forecasting. Trans. Inst. Meas. Control
Moghadam, J.N., Salahshoor, K., Kharrat, R., 2011. Intelligent prediction of porosity 42, 3e21.
and permeability from well logs for an Iranian fractured carbonate reservoir. Tahmasebi, P., Hezarkhani, A., 2012. A fast and independent architecture of artificial

181
S. Bahaloo, M. Mehrizadeh and A. Najafi-Marghmaleki Petroleum Research 8 (2023) 167e182

neural network for permeability prediction. J. Petrol. Sci. Eng. 86, 118e126. Yin, H., Si, M., Wang, H., 2014. The warning model of the early kick based on BP
Toreifi, H., Rostami, H., 2014. New method for prediction and solving the problem of neural network. Pakistan J. Statis. 30.
drilling fluid loss using modular neural network and particle swarm optimi- Yu, D., Zhang, Y., Bian, H., Wang, X., Qi, W., 2013. A new diagnostic method for
zation algorithm. J. Pet. Explor. Prod. Technol. 4, 371e379. identifying working conditions of submersible reciprocating pumping systems.
Vega, M., Vieira, F., Fernandes, L., Freitas, M., Russano, E., Martins, A., 2016. Smart Petrol. Sci. 10, 81e90.
monitoring and decision making for regulating annulus bottom hole pressure Zarei, F., Daliri, A., Alizadeh, N., 2008. The Use of Neuro-Fuzzy Proxy in Well
while drilling oil wells. Braz. J. Chem. Eng. 33, 969e983. Placement Optimization. Society of Petroleum Engineers.
Wang, S., Chen, S., 2019. Insights to fracture stimulation design in unconventional Zendehboudi, S., Rajabzadeh, A.R., Bahadori, A., Chatzis, I., 2014. Connectionist
reservoirs based on machine learning modeling. J. Petrol. Sci. Eng. 174, model to estimate performance of steam-assisted gravity drainage in fractured
682e695. and unfractured petroleum reservoirs: enhanced oil recovery implications. Ind.
Wang, B., Wang, X., Chen, Z., 2013. A hybrid framework for reservoir characteriza- Eng. Chem. Res. 53, 1645e1662.
tion using fuzzy ranking and an artificial neural network. Comput. Geosci. 57, Zerrouki, A.A., Aifa, T., Baddari, K., 2014. Prediction of natural fracture porosity from
1e10. well log data by means of fuzzy ranking and an artificial neural network in
Wu, C., Lu, G., Gillespie, W., Yen, J., 1997. Statistical and Fuzzy Infill Drilling Recovery Hassi Messaoud oil field, Algeria. J. Petrol. Sci. Eng. 115, 78e89.
Models for Carbonate Reservoirs. Society of Petroleum Engineers. Zhang, Y., Chen, M., Jin, Y., Lu, Y., 2016. Experimental study and artificial neural
Xie, H., Shanmugam, A.K., Issa, R.R., 2018. Big data analysis for monitoring of Kick network simulation of the wettability of tight gas sandstone formation. J. Nat.
Formation in complex underwater drilling projects. J. Comput. Civ. Eng. 32, Gas Sci. Eng. 34, 387e400.
04018030. Zhang, Z., Zhang, H., Li, J., Cai, Z., 2021. Permeability and porosity prediction using
Xue, Y., Cheng, L., Mou, J., Zhao, W., 2014. A new fracture prediction method by logging data in a heterogeneous dolomite reservoir: an integrated approach.
combining genetic algorithm with neural network in low-permeability reser- J. Nat. Gas Sci. Eng. 86, 103743.
voirs. J. Petrol. Sci. Eng. 121, 159e166. Zhong, Z., Carr, T.R., 2019. Application of a new hybrid particle swarm optimization-
Yetilmezsoy, K., Fingas, M., Fieldhouse, B., 2011. An adaptive neuro-fuzzy approach mixed kernels function-based support vector machine model for reservoir
for modeling of water-in-oil emulsion formation. Colloids Surf., A 389, 50e62. porosity prediction: a case study in Jacksonburg-Stringtown oil field. West
Yılmaz, S., Demircioglu, C., Akin, S., 2002. Application of artificial neural networks to Virginia, USA Interpretation 7, T97eT112.
optimum bit selection. Comput. Geosci. 28, 261e269.

182