1

Photoelectric Factor Prediction Using Automated Learning and Uncertainty

Quantification
Khalid L. Alsamadony1*, Ahmed Ibrahim1,2**, Salaheldin Elkatatny1,2, Abdulazeez Abdulraheem1,2
1
Petroleum Engineering Department, King Fahd University of Petroleum & Minerals, Saudi Arabia.
2
Center for Integrative Petroleum Research, King Fahd University of Petroleum & Minerals, Saudi
Arabia.
*E-mail: [email protected] **E-mail: [email protected]

Abstract
The photoelectric factor (PEF) is an important well logging tool to distinguish between different types of
reservoir rocks because PEF measurement is sensitive to elements with high atomic number. Furthermore,
the ratio of rock minerals could be determined by combining PEF log with other well logs.
However, PEF log could be missing in some cases such as in old well logs and wells drilled with barite-
based mud. Therefore, developing models for estimating missing PEF log is essential in those
circumstances. In this work, we developed various machine learning models to predict PEF values using
the following well logs as inputs: bulk density (RHOB), neutron porosity (NPHI), gamma ray (GR),
compressional and shear velocity.
The predictions of PEF values using adaptive-network-fuzzy inference system (ANFIS) and artificial
neural network (ANN) models have errors of about 16% and 14% average absolute percentage error
(AAPE) in the testing dataset, respectively. Thus, a different approach was proposed that is based on the
concept of automated machine learning. It works by automatically searching for the optimal model type
and optimizes its hyperparameters for the dataset under investigation.
This approach selected a Gaussian process regression (GPR) model for accurate estimation of PEF
values. The developed GPR model decreases the AAPE of the predicted PEF values in the testing dataset
to about 10% AAPE. This error could be further decreased to about 2% by modeling the potential noise
in the measurements using the GPR model.

Key Words: Machine Learning Algorithms; Well Logging; Photoelectric Factor; Automated Learning;
Uncertainty Quantification; Gaussian process regression; fuzzy logic; artificial neural network (ANN)

1 Introduction
The photoelectric factor (PEF) log is a robust tool for identifying lithology. This is because PEF
measurement is mainly affected by rocks with high atomic number, while being insensitive to variations
in porosity, fluid type and saturation. Therefore, PEF log capability to identify the lithology is not
diminished in gas-bearing zones, where it may be challenging to distinguish between lithologies using
only formation density and neutron logs. One more advantage of PEF log is that it can distinguish between
clean (i.e., low shale content) sandstone and clean limestone, while gamma ray log cannot clearly
demystify the difference between them. PEF (𝑃𝑒 ) is proportional to the atomic number as shown in the
next empirical relationship (Eq.1), where 𝑧 is the atomic number (Atlas, 1982; Bassiouni & others, 1994;
Ellis & Singer, 2007).

𝑍 3.6
𝑃𝑒 = (10) (1)

PEF log may be missing in some cases because of instrument failure, old logs missing this tool, and
economic constraints. Moreover, PEF log could be mismeasured and misinterpreted in wells drilled with
barite and in wells that have heavy elements even in small quantity. Remeasuring PEF log is often
impractical because of well casing or limited budget. Therefore, previous studies utilized machine learning
2

techniques to predict PEF log due to its importance in improving formation evaluation (Akinnikawe et al.,
2018; Asoodeh & Shadizadeh, 2015).
1.1 PEF Logs as Inputs for Machine Learning Models.
PEF log is an important input parameter for many machine learning models to classify lithology and
predict other valuable parameters such as density, porosity, water saturation, sonic and geochemical logs.
PEF was used for training classifiers (k-nearest neighbors (KNN), support vector machine (SVM), random
forest (RF), decision tree (DT), XGBoost, LightGBM, artificial neural network (ANN) and rough set
theory rules) to predict lithofacies and electrofacies (Amir et al., 2020; Hossain et al., 2020; Merembayev
et al., 2021). Moreover, it is an input for several models (ANN, SVM, Super Learner, XGBoost, CatBoost,
LightGBM and AdaBoost) to estimate water saturation (Hadavimoghaddam et al., 2021; Miah et al.,
2020).
PEF log was utilized to generate sonic logs (compressional and shear velocity) using ANN, recurrent
neural networks (RNNs), RF, ensemble method (combination of several models) and LightGBM, where
data preprocessing and feature engineering had a significant impact on the accuracy (Tatsipie & Sheng,
2021; Yu et al., 2021). Another study proposed fast fuzzy modeling method (FFMM) to predict sonic and
density logs using PEF log as an input, where the proposed method had similar accuracy to ANN and
fuzzy logic (FL), but with less computational and storage cost (Bahrpeyma et al., 2013).
PEF log is an important input in predicting geochemical logs (Al, Ca, Fe, Na, and Si) using the
following methods: SVM, ANN, RF, AdaBoost and XGBoost (Blanes de Oliveira & de Carvalho
Carneiro, 2021). Moreover, it is one of the inputs for long short-term memory (LSTM) model to output
neutron porosity and resistivity logs (Tatsipie & Sheng, 2021).
1.2 Prediction of PEF Logs
Tatsipie and Sheng trained LSTM model to predict PEF log using the following inputs: gamma ray, bulk
density, depth and well coordinates. PEF log was predicted with high errors relative to other predicted
logs in this study, which was justified that the trained LTSM model could not capture the complex relation
between PEF, gamma ray and bulk density (Tatsipie & Sheng, 2021).
Rostamian et al., predicted PEF, bulk density and neutron porosity logs using several machine learning
methods: multilinear regression, deep neural network (DNN), DT, KNN, RF, and XGBoost. These models
receive corrected gamma ray (CGR), caliper (CALI), deep latero log (LLD), shallow latero log (LLS),
micro-spherical focused log (MSFL) and sonic transit time, where all the well logs were normalized (Z-
score transformation). XGBoost, RF and KNN generated the best results, in addition their
hyperparameters and selected inputs were optimized by genetic algorithm. This optimization slightly
improved the accuracy and allowed reducing the required number of inputs, while predicting the missing
logs with acceptable accuracy (Rostamian et al., 2022).
1.3 Novelty
The objective of this study is to develop machine learning models for estimating PEF values for a
heterogenous dataset that was collected from a field in the middle east. In our dataset, the PEF values
range from (2.1) to (13.4), while they range from about (1) to (7.8), (2.5) to (5.5) and (1) to (4.9) in the
datasets of the previous studies (Rostamian et al., 2022; Tatsipie & Sheng, 2021; Wu et al., 2021).
The previous studies have predicted PEF values with relatively high error. Rostamian et al., predicted
PEF values that have about 10 times more error than the prediction of neutron porosity (NPHI) according
to the AAPE (Rostamian et al., 2022). Tatsipie and Sheng concluded that the predicted PEF log has the
worst accuracy compared to the other predicted well logs (Tatsipie & Sheng, 2021). Our results, despite
being based on a different dataset, confirm that predicting PEF log accurately could be challenging, hence
this seems related to the prediction of PEF logs regardless of the dataset. For this reason, a method based
on automated machine learning is proposed for better accuracy and efficient implementation of machine
learning tools through removing the time-consuming manual steps. Additionally, we utilize uncertainty
quantification to enhance the reliability of machine learning methods by modeling the potential noise that
is usually in the measurements of actual cases at varying levels.
 3

2 Methodology
In this work, we developed three machine learning models to predict PEF log, which are adaptive neuro-
fuzzy inference system (ANFIS), artificial neural network (ANN) and Gaussian process regression (GPR).
The inputs of these models are 5 logs: bulk density (RHOB), neutron porosity (NPHI), gamma ray (GR),
compressional and shear velocity. The accuracy of the trained models is evaluated using the common
metrics: average (mean) absolute percentage error (AAPE) (Eq.2) and Pearson correlation coefficient (R)
(Eq.3) between the predicted and the actual PEF values. For the uncertainty quantification (UQ) model,
AAPE refers to the percentage of the predicted PEF values that are outside the confidence interval (CI).

100 𝑎𝑐𝑡𝑢𝑎𝑙 𝑃𝐸𝐹𝑖 −𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 𝑃𝐸𝐹𝑖

AAPE = × ∑𝑛𝑖=1 | | (2)
𝑛 𝑎𝑐𝑡𝑢𝑎𝑙 𝑃𝐸𝐹𝑖

∑𝑛
𝑖=1 (𝑥𝑖 −𝑥̅ )(𝑦𝑖 −𝑦
̅)
R= (3)
√∑𝑛 2 𝑛 ̅)2
𝑖=1 (𝑥𝑖 −𝑥̅ ) √∑𝑖=1 (𝑦𝑖 −𝑦

where 𝑛 is the number of datapoints, 𝑥𝑖 is the actual PEF value, 𝑥̅ is the average of the actual PEF
values, 𝑦𝑖 is the predicted PEF value, and 𝑦̅ is the average of the predicted PEF values.
2.1 Data Description
The dataset consists of 9399 datapoints from a field in the middle east such that 4922 datapoints were used
for training the models, 895 datapoints for the initial testing, and 2603 datapoints for the final validation
of the models.
Statistical analysis of the dataset is presented in Table 1, where the heterogeneity of the dataset can be
observed. Specifically, the minimum and the maximum values of GR (3, 381) and NPHI (0, 0.44) indicate
that the dataset contains different types of zones with distinct properties. The kurtosis values of GR (9.75)
and NPHI (4.67) show that they have many extreme values. R-values depict that RHOB is the most
linearly correlated to PEF (-0.37), that is followed by compressional velocity (0.18). R value measures the
linear correlation between two variables, where value of one indicates a perfect proportional linear
correlation, value of negative one indicates a perfect inversely proportional linear correlation, and zero
indicates that there is no linear correlation between the two variables.

Table 1 Statistical description of the dataset

Statistical
RHOB NPHI GR Compressional Shear PEF
Analysis
Min 2.18 -0.01 3.17 40.31 60.82 2.14
Max 3.00 0.44 381.90 83.04 147.35 13.39
Mean 2.71 0.08 55.80 53.34 93.50 5.77
Mode 2.85 0.02 19.97 43.69 82.84 4.21
Median 2.59 0.22 192.53 61.67 104.08 7.76
Standard
0.17 0.06 49.14 8.86 11.68 2.05
deviation
Coef. of
0.06 0.80 0.88 0.17 0.12 0.36
variation
Skewness -0.43 1.21 2.29 0.75 1.23 0.83
Kurtosis 1.98 4.67 9.75 2.49 4.46 3.21
R (PEF) -0.37 0.10 0.03 0.18 0.13 1.00

Before the training process, the inputs and the PEF values were normalized to be between zero and one
using their minimum and maximum values (Eq.4). However, the predicted PEF values were restored to
4

the original scale (denormalized), so we can evaluate the accuracy of the models using AAPE and R. The
dataset was checked for erroneous values such as zeros, negatives, and very extreme unreasonable values.
Moreover, we attempted to remove the outliers using common statistical rules such as 1.5 IQR and 3 sigma
(standard deviation). However, the accuracy of ANN model did not improve, but it slightly decreased.
This probably due to the heterogenous nature of our dataset, where these extreme values represent some
layers. Thus, removing them hinders the predictive capability of the artificial intelligence models.
Additionally, numerous feature engineering and data transformation methods were applied, but they were
not effective at reducing the error, as they only reduced AAPE by about 1% in the best case. As a
consequence, a different approach was implemented utilizing automated learning and uncertainty
quantification.
𝑥−𝑚𝑖𝑛(𝑥)
𝑁𝑜𝑟𝑚𝑎𝑙𝑖𝑧𝑒𝑑 (𝑥) = 𝑚𝑎𝑥(𝑥)−𝑚𝑖𝑛(𝑥) (4)

2.2 Models Description

In this section, we briefly introduce the three machine learning methods that have been used to generate
the results in our work, which are ANFIS, ANN and GPR. Moreover, we cite some references that explain
the theory of these methods in detail and their different applications.
Adaptive-network-based fuzzy inference system (ANFIS) can be used as a supervised learning method
since it is capable of modeling non-linear functions. Its prediction (inference) is based on Fuzzy rules (if-
then rules) to map inputs of the system to outputs, thus it can resemble human knowledge and reasoning.
These rules (if-then) are implemented using membership functions. This is the main difference between
ANFIS and ANNs (Jang, 1993). Several studies used ANFIS to predict rock properties and well logs
(Desouky et al., 2021; Gamal et al., 2021; Gowida et al., 2021).
ANNs have been widely used for classification and regression tasks. The basic component of ANN is
the neuron, which forms the hidden layers. The hidden layers map the input layer to the output layer, and
each hidden layer is followed by a transfer function to learn non-linear features. The neurons are associated
with weights that are updated during the training process according to the backpropagated errors and the
training algorithm e.g., gradient descent (Gurney, 1997). Several studies used ANN to predict rock
properties and well logs (Asante-Okyere et al., 2018; Desouky et al., 2021; Gamal et al., 2021; Gowida et
al., 2021; Hiba et al., 2022).
The general concept of GPR assumes that datapoints with similar input values are expected to have
similar output values. This concept is applied using covariance (kernel) functions that model these
relations of similarity (Rasmussen & Williams, 2006). Due to this, kriging that is the well-known method
used in geo-statistics to interpolate the missing values based on the established spatial covariance or
variogram, is considered a GPR method. GPR is less commonly used than ANN to predict rock properties
and well logs (Asante-Okyere et al., 2018; Onalo et al., 2020).

3 Results
3.1 ANFIS Results
To develop the ANFIS model, we used subtractive clustering (Sugeno fuzzy) algorithm that searches for
data clusters, from which membership functions and rules are constructed (Chiu, 1994; Yager & Filev,
1994). The main hyperparameter of the subtractive clustering method is the radius of the clusters. During
the training phase, the ANFIS model was optimized by varying the radius values to identify the optimum
value that results in the highest R and lowest AAPE for the testing dataset. The optimized ANFIS model
has a radius of 0.139 and was trained for 200 epochs. The accuracy of the developed ANFIS model in
predicting the PEF values is 19.19% AAPE in the training data and 16.04% AAPE in the testing data.
Moreover, the correlation coefficient (R) between the predicted and the actual PEF values is 0.715 in the
training data and 0.77 in the testing data. Figure 1 presents the cross plots between the predicted and the
actual PEF values. We observe that the ANFIS model underfits the training data because it has higher
errors in predicting the PEF values in the training data than in the testing data.
 5

Figure 1: Cross plots between the actual and the predicted PEF values from the ANFIS model.

3.2 ANN Results

The accuracy of ANN models is affected by their hyperparameters such as the number of neurons and
layers, the transfer (activation) function and the training algorithm that minimizes the loss function by
updating the weights of the network. The studied search space of the ANN hyperparameters are the
following: the number of hidden layers (1-3), the number of neurons (5-35), the transfer functions
(hyperbolic tangent sigmoid, log sigmoid, hard limit, linear, radial basis, and soft max), and the training
algorithms (Levenberg–Marquardt, BFGS quasi-Newton, Bayesian, and conjugate gradients).
The developed ANN model was optimized by selecting the optimum combination of those
hyperparameters to generate the best accuracy for the testing dataset. The best hyperparameters are 2
hidden layers with 16 and 20 neurons respectively, log sigmoid transfer function, and Bayesian
regularization backpropagation. The accuracy of the ANN model in predicting the PEF values is
(AAPE=13.32% and R=0.856) in the training data, and (AAPE=14.09% and R=0.815) in the testing data.
Figure 2 shows the cross plots between the actual PEF and the predicted PEF values. Although, the
accuracy of the ANN model is better than the ANFIS model, it is still not perfectly accurate.

Figure 2: Cross plots between the actual and the predicted PEF values from the ANN model.
6

3.3 GPR Results

Automated machine learning searches for the optimal algorithm (model) for the dataset of interest, by
applying different types of methods, while optimizing their hyperparameters for the best results (Feurer
et al., 2015). Thus, it increases the efficiency of applying artificial intelligence tools through eliminating
the time wasted on manual steps.
The automated learning code compares different models including multi-linear regression, Gaussian
kernel regression, decision tree, GPR, ensemble learning (Least-squares gradient boosting and Bootstrap
aggregation (random forest)), ANNs and support vector machine (SVM). The code utilizes Bayesian
optimization to select the optimal hyperparameters for the models and find the most accurate model for
the prediction task (Snoek et al., 2012). The results show that the best model is a GPR model. One
additional unique advantage of GPR is that it can incorporate uncertainty, hence accounting for the
potential noise in the PEF measurements. This advantage is extremely critical in zones where the PEF
values cannot be predicted accurately because it reduces the probability of misinterpreting the predicted
PEF log by geoscientists and engineers.
The optimum GPR model has these hyperparameters: sigma=0.032, kernel scale=0.602, constant basis
function, and exponential kernel function. It is evident that the GPR model outperforms the ANFIS and
the ANN models as summarized in Table 2. The accuracy of the GPR model in predicting the PEF values
is (AAPE=0.01% and R=0.999) in the training data, and (AAPE=10.32% and R=0.889) in the testing data.
Figure 3 is the cross plots between the actual PEF and the predicted PEF values. Furthermore, the error
of the GPR model in the testing dataset could be further decreased to about (4.6% - 1.9%) using (95% -
99%) confidence interval to model the noise of the dataset.
We observe that the GPR model overfits the training data, which is common to some GPR methods
(Mohammed & Cawley, 2017). However, this probably does not reduce its accuracy as can be observed
by comparing the GPR model accuracy against the accuracy of the ANN and the ANFIS models using the
testing data and the final validation data (next section). In general, weak learners tend to underfit the data
(e.g., in our case the ANFIS model), while strong learners tend to overfit the data.

Table 2 Compares the accuracy of the developed models

Method R (train) AAPE (train), % R (test) AAPE (test), %

ANFIS 0.715 19.19 0.77 16.04
ANN 0.856 13.32 0.815 14.09
GPR 0.999 0.01 0.889 10.23

Figure 3: Cross plots between the actual and the predicted PEF values from the GPR model.
 7

4 Validation
In this section, the best model (GPR + UQ) is further tested on larger unseen dataset for final validation.
This dataset consists of 2603 datapoints. The cross plot between the predicted and the actual PEF values
in the validation data is shown in Figure 4. The GPR predicts the PEF values with the highest accuracy
(AAPE= 9.5% and R=0.894) in the validation data (Table 3), where estimating the uncertainty using 95%
confidence (prediction) interval decreases the error to (4.8%). The importance of uncertainty
quantification is noticeable in zones where the actual PEF values deviate from the predictions, e.g., depth
index from 1080 to 1090 (Figure 5). The uncertainty model could also consider the potential noise in the
inputs that can be observed from the depth index (1035-1040) in Figure 5, where the PEF values are
similar but have different confidence estimation because of having different inputs.
Despite that the error of the uncertainty could be further decreased to about 2% using 99% confidence
interval, we selected 95% confidence interval as it generates good accuracy with reasonable prediction
width. However, in other cases, a different confidence interval (lower or higher) may be selected
depending on the knowledge about the predicted parameter, the acceptable range of error and the width
of confidence interval.

Figure 4: Cross plot between the actual and the predicted PEF values from the GPR model (final validation data).

Table 3 Compares the accuracy of the developed models using the final validation dataset (2603 datapoints)

Method R (validation) AAPE (validation), %

ANFIS 0.761 15.51
ANN 0.810 13.32
GPR 0.894 9.55
8

Figure 5: Depth VS the predicted PEF values from the GPR+UQ model (final validation data).

5 Conclusion
In this work, we developed various models to predict PEF log. The results show that predicting PEF logs
can be challenging because they cannot be predicted accurately for some zones by the common machine
learning methods. Similar results have been reported by previous studies. To improve the accuracy of the
models, we attempted several data preprocessing and transformation methods, but they were not effective
at reducing the prediction error.
Therefore, automated machine learning and uncertainty quantification were proposed for robust
prediction of the PEF values in our dataset. The automated learning code recommended the GPR model
which outperformed the other machine learning methods in predicting the PEF values. Moreover,
uncertainty quantification has allowed further reduction of error by considering the noise in the data, which
is crucial for the troublesome reservoir layers. It could be used for risk assessment, hence decreasing the
possibilities of misinterpreting the predicted PEF values.
The GPR model generates the most accurate PEF predictions for the dataset under investigation, but
could this finding be generalized to other well logs, drilling and geo-mechanical parameters? More
investigations may be required to confirm the reliability of the GPR model in predicting PEF logs for
different wells, fields, data ranges, and geological layers. Additionally, the causes of the high error in the
predicted PEF values using machine learning methods could be studied. This error may be due to the
heterogeneity of the dataset, some special reservoir rocks, well (borehole) condition, heavy elements, and
noise from the well logging tools (PEF tool).
 9

References
Akinnikawe, O., Lyne, S., & Roberts, J. (2018). Synthetic Well Log Generation Using Machine
Learning Techniques. Proceedings of the 6th Unconventional Resources Technology Conference.
https://doi.org/10.15530/urtec-2018-2877021
Amir, S. M., Khan, M. R., Panacharoensawad, E., & Kryvenko, S. (2020, October 26). Integration of
Petrophysical Log Data with Computational Intelligence for the Development of a Lithology
Predictor. Day 2 Tue, October 27, 2020. https://doi.org/10.2118/202047-MS
Asante-Okyere, S., Shen, C., Ziggah, Y. Y., Rulegeya, M. M., & Zhu, X. (2018). Investigating the
predictive performance of Gaussian process regression in evaluating reservoir porosity and
permeability. Energies, 11(12). https://doi.org/10.3390/en11123261
Asoodeh, M., & Shadizadeh, S. R. (2015). The Prediction of Photoelectric Factor, Formation True
Resistivity, and Formation Water Saturation from Petrophysical Well Log Data: A Committee
Neural Network Approach. Energy Sources, Part A: Recovery, Utilization, and Environmental
Effects, 37(5), 557–566. https://doi.org/10.1080/15567036.2011.594859
Atlas, D. (1982). Well logging and interpretation techniques: The course for home study. Dresser Atlas.
Bahrpeyma, F., Golchin, B., & Cranganu, C. (2013). Fast fuzzy modeling method to estimate missing
logsin hydrocarbon reservoirs. Journal of Petroleum Science and Engineering, 112, 310–321.
https://doi.org/10.1016/j.petrol.2013.11.019
Bassiouni, Z., & others. (1994). Theory, measurement, and interpretation of well logs (Vol. 4). Henry L.
Doherty Memorial Fund of AIME, Society of Petroleum Engineers~….
Blanes de Oliveira, L. A., & de Carvalho Carneiro, C. (2021). Synthetic geochemical well logs
generation using ensemble machine learning techniques for the Brazilian pre-salt reservoirs.
Journal of Petroleum Science and Engineering, 196, 108080.
https://doi.org/10.1016/j.petrol.2020.108080
Chiu, S. L. (1994). Fuzzy Model Identification Based on Cluster Estimation. Journal of Intelligent and
Fuzzy Systems, 2(3), 267–278. https://doi.org/10.3233/IFS-1994-2306
Desouky, M., Tariq, Z., Aljawad, M. S., Alhoori, H., Mahmoud, M., & Abdulraheem, A. (2021).
Machine Learning-Based Propped Fracture Conductivity Correlations of Several Shale Formations.
ACS Omega, 6(29), 18782–18792. https://doi.org/10.1021/acsomega.1c01919
Ellis, D. v, & Singer, J. M. (2007). Well logging for earth scientists (Vol. 692). Springer.
Feurer, M., Klein, A., Eggensperger, K., Springenberg, J., Blum, M., & Hutter, F. (2015). Efficient and
Robust Automated Machine Learning. In C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, & R.
Garnett (Eds.), Advances in Neural Information Processing Systems (Vol. 28). Curran Associates,
Inc. https://proceedings.neurips.cc/paper/2015/file/11d0e6287202fced83f79975ec59a3a6-Paper.pdf
Gamal, H., Abdelaal, A., & Elkatatny, S. (2021). Machine Learning Models for Equivalent Circulating
Density Prediction from Drilling Data. ACS Omega, 6(41), 27430–27442.
https://doi.org/10.1021/acsomega.1c04363
Gowida, A., Elkatatny, S., & Gamal, H. (2021). Unconfined compressive strength (UCS) prediction in
real-time while drilling using artificial intelligence tools. Neural Computing and Applications,
33(13), 8043–8054. https://doi.org/10.1007/s00521-020-05546-7
Gurney, K. (Kevin N. ). (1997). An introduction to neural networks. UCL Press.
Hadavimoghaddam, F., Ostadhassan, M., Sadri, M. A., Bondarenko, T., Chebyshev, I., & Semnani, A.
(2021). Prediction of Water Saturation from Well Log Data by Machine Learning Algorithms:
Boosting and Super Learner. Journal of Marine Science and Engineering, 9(6), 666.
https://doi.org/10.3390/jmse9060666
Hiba, M., Ibrahim, A. F., Elkatatny, S., & Ali, A. (2022). Application of Machine Learning to Predict
the Failure Parameters from Conventional Well Logs. Arabian Journal for Science and
Engineering. https://doi.org/10.1007/s13369-021-06461-2
Hossain, T. M., Watada, J., Aziz, I. A., & Hermana, M. (2020). Machine Learning in Electrofacies
10

Classification and Subsurface Lithology Interpretation: A Rough Set Theory Approach. Applied
Sciences, 10(17), 5940. https://doi.org/10.3390/app10175940
Jang, J.-S. R. (1993). ANFIS: adaptive-network-based fuzzy inference system. IEEE Transactions on
Systems, Man, and Cybernetics, 23(3), 665–685. https://doi.org/10.1109/21.256541
Merembayev, T., Kurmangaliyev, D., Bekbauov, B., & Amanbek, Y. (2021). A Comparison of Machine
Learning Algorithms in Predicting Lithofacies: Case Studies from Norway and Kazakhstan.
Energies, 14(7), 1896. https://doi.org/10.3390/en14071896
Miah, M. I., Zendehboudi, S., & Ahmed, S. (2020). Log data-driven model and feature ranking for water
saturation prediction using machine learning approach. Journal of Petroleum Science and
Engineering, 194, 107291. https://doi.org/10.1016/j.petrol.2020.107291
Mohammed, R., & Cawley, G. (2017). Over-Fitting in Model Selection with Gaussian Process
Regression (pp. 192–205). https://doi.org/10.1007/978-3-319-62416-7_14
Onalo, D., Adedigba, S., Oloruntobi, O., Khan, F., James, L. A., & Butt, S. (2020). Data-driven model
for shear wave transit time prediction for formation evaluation. Journal of Petroleum Exploration
and Production Technology, 10(4), 1429–1447. https://doi.org/10.1007/s13202-020-00843-2
Rasmussen, C., & Williams, C. (2006). Gaussian processes for machine learning.,(MIT press:
Cambridge, MA).
Rostamian, A., Heidaryan, E., & Ostadhassan, M. (2022). Evaluation of different machine learning
frameworks to predict CNL-FDC-PEF logs via hyperparameters optimization and feature selection.
Journal of Petroleum Science and Engineering, 208, 109463.
https://doi.org/10.1016/j.petrol.2021.109463
Snoek, J., Larochelle, H., & Adams, R. P. (2012). Practical bayesian optimization of machine learning
algorithms. Advances in Neural Information Processing Systems, 25.
Tatsipie, N. R. K., & Sheng, J. J. (2021). Generating pseudo well logs for a part of the upper Bakken
using recurrent neural networks. Journal of Petroleum Science and Engineering, 200, 108253.
https://doi.org/10.1016/j.petrol.2020.108253
Wu, L., Dong, Z., Li, W., Jing, C., & Qu, B. (2021). Well-logging prediction based on hybrid neural
network model. Energies, 14(24). https://doi.org/10.3390/en14248583
Yager, R. R., & Filev, D. P. (1994). Generation of fuzzy rules by mountain clustering. Journal of
Intelligent & Fuzzy Systems, 2(3), 209–219.
Yu, Y., Xu, C., Misra, S., Li, W., Ashby, M., Pan, W., Deng, T., Jo, H., Santos, J. E., Fu, L., Wang, C.,
Kalbekov, A., Suarez, V., Kusumah, E. P., Aviandito, M., Pamadya, Y., & Izadi, H. (2021).
Synthetic Sonic Log Generation With Machine Learning: A Contest Summary From Five Methods.
Petrophysics – The SPWLA Journal of Formation Evaluation and Reservoir Description, 62(4),
393–406. https://doi.org/10.30632/PJV62N4-2021a4