See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/375880467

Reservoir Property Prediction in the North Sea Using Machine Learning

Article in IEEE Access · November 2023

DOI: 10.1109/ACCESS.2023.3336623

CITATIONS READS

0 147

3 authors, including:

Abdulrahman Al-Fakih Ardiansyah Koeshidayatullah

King Fahd University of Petroleum and Minerals King Fahd University of Petroleum and Minerals
8 PUBLICATIONS 10 CITATIONS 74 PUBLICATIONS 606 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Abdulrahman Al-Fakih on 19 December 2023.

The user has requested enhancement of the downloaded file.

Received 16 October 2023, accepted 22 November 2023, date of publication 23 November 2023,
date of current version 15 December 2023.
Digital Object Identifier 10.1109/ACCESS.2023.3336623

Reservoir Property Prediction in the North Sea

Using Machine Learning
ABDULRAHMAN AL-FAKIH , (Member, IEEE), SANLINN I. KAKA ,
AND ARDIANSYAH I. KOESHIDAYATULLAH , (Member, IEEE)
College of Petroleum Engineering and Geosciences, King Fahd University of Petroleum Minerals, Dhahran 31261, Saudi Arabia
Corresponding authors: Abdulrahman Al-Fakih ([email protected]) and Ardiansyah I. Koeshidayatullah
([email protected])
The work of Ardiansyah I. Koeshidayatullah was supported in part by the Saudi Data and Artificial Intelligence Authority
(SDAIA)—King Fahd University of Petroleum and Minerals (KFUPM) Joint Research Center in AI under Grant JRCAI-RG-05.

ABSTRACT The North Sea sedimentary basin is characterized by geological complexity, encompassing
a wide range of rock types and structures, including multiple reservoirs (carbonates and siliciclastic) with
variations in reservoir quality and heterogeneity. These phenomena pose significant challenges for accurately
predicting reservoir properties using traditional well log analysis. Moreover, these challenges are further
compounded by complex geological conditions and scarcity of available data. Hence, the aim of this study
was to address these challenges by applying advanced machine learning techniques within this basin. This
study delves into both supervised and unsupervised machine learning approaches to forecast the essential
petrophysical properties that are crucial for assessing reservoir quality. These properties encompass total
porosity, effective porosity, and shale volume, all derived from well log data originating from the North Sea
sedimentary basin. The models were trained using data from four wells consisting of 32,215 data points
(80% for training, 10% for testing, and 10% for validation). Furthermore, our study introduced pioneering
data-driven preprocessing workflow, encompassing exploratory data analysis, missing data imputation, and
outlier detection to improve the performance of the machine learning models. ANN and RF models achieved
the best results among the algorithms evaluated, with an average MAE of 0.01, RMSE of 0.01, and R-squared
of 0.95 for total porosity, effective porosity, and volume of shale, respectively. These metrics demonstrate
that the model can accurately predict reservoir properties in a challenging sedimentary basin, even with
limited data availability, enabling reservoir characteristics and field development optimization, particularly
in areas where core data are scarce.

INDEX TERMS Reservoir property prediction, machine learning, artificial neural networks, K-nearest
neighbors, random forests, decision trees, North Sea sedimentary basin.

NOMENCLATURE
MAE Mean Absolute Error.
ANN Artificial Neural Networks.
ML Machine Learning.
AutoML Automated Machine Learning.
MLT Machine Learning Techniques.
DT Sonic Transit Time.
NPHI Neutron Porosity.
DTs Decision Trees.
RF Random Forest.
EDA Exploratory Data Analysis.
RHOB Bulk Density.
GR Gamma Ray.
RMSE Root Mean Square Error.
IF Isolation Forest.
R2 Coefficient of Determination.
IQR Interquartile Range.
SVM Support Vector Machine.
KNN K-nearest neighbors.
LOF Local Outlier Factor.
I. INTRODUCTION
The associate editor coordinating the review of this manuscript and A comprehensive understanding of reservoir properties is
approving it for publication was Wentao Fan . crucial for the successful extraction of hydrocarbons from
2023 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License.
140148 For more information, see https://creativecommons.org/licenses/by/4.0/ VOLUME 11, 2023
A. Al-Fakih et al.: Reservoir Property Prediction in the North Sea Using Machine Learning

both mature and frontier hydrocarbon fields. Moreover, pre- Numerous studies have utilized machine learning tech-
cise and reliable forecasts of these properties are imperative niques (MLT) to predict porosity, permeability, and resistivity
to maximize oil and gas output from reservoirs. The accuracy from well logs, often achieving high levels of accuracy and
of these predictions plays a crucial role in ensuring that surpassing traditional models in some cases [10], [11], [12],
production processes are optimized and efficient [1]. Tradi- [13], [14], [15], [16]. However, these studies have primarily
tional approaches to reservoir property prediction, utilizing been concentrated in fields in outside the Dutch sector of the
well log data and geological models, can be time-consuming, North Sea. This section reviews the background of reservoir
expensive, costly, and potentially inaccurate in complex geo- property prediction, with a focus on the recent emergence
logical environments. Machine learning (ML) has emerged of MLT and its potential application in the Dutch oil and
as a powerful tool in this domain. ML algorithms, trained gas industry. Table 1 demonstrates that ML algorithms can
on data from well logs, core analysis, and seismic surveys, accurately predict reservoir parameters, with certain algo-
can discern relationships between reservoir properties and rithms exhibiting superior performances in specific regions
various factors, such as rock type, depositional environment, and reservoir types.
and fluid composition. Once trained, ML algorithms can be Recent advancements in ML for reservoir property predic-
used to predict reservoir properties in new areas with limited tion have highlighted diverse methodologies. For instance,
data [2]. a study conducted in a Malaysian brownfield utilized Random
This study focused on the predictive modeling of reservoir Forest (RF) algorithms with well logs and core analysis,
properties in the North Sea sedimentary basin using ML achieving an R2 of 85% for porosity and 80% for permeabil-
techniques. We collected well log data from six wells, trained ity [17]. This approach, which mirrors our data usage, differs
the models using data from four wells, and evaluated the in its application to ML. Another study developed a system-
models using one well for testing and one well each for atic ML approach for reservoir identification and production
validation. The artificial neural network (ANN), K-nearest prediction by employing seven ML methods, including
neighbor (KNN), random forest (RF), and decision tree (DT) XGBoost, and achieved up to 99% accuracy for effective
algorithms were employed to predict the total porosity, effec- reservoirs [18]. This contrasts with our study’s exclusive
tive porosity, and volume of shale. The performance of the focus on property predictions. Additionally, research on
models was assessed using evaluation metrics such as the reservoir prediction under coastal conditions has employed
mean absolute error (MAE), root mean square error (RMSE), binary classification algorithms and data augmentation tech-
and coefficient of determination (R-squared). niques, providing a unique perspective compared to our North
The main objectives of this study are as follows: Sea focus [19]. These studies collectively underscore the
evolving role of ML in reservoir property prediction and
• Introduce a novel workflow for predicting reservoir contextualize our work.
properties in the North Sea sedimentary basin using ML Discussing the specific applications of ML in various
techniques. regional contexts [17], [18], [19] leads to the acknowledg-
• Evaluate the performance of four supervised and unsu- ment of recent advancements in the field. The implementation
pervised ML algorithms (ANN, KNN, RF, and DT) of automated machine learning (AutoML) for more efficient
for reservoir property prediction. This demonstrates and accurate reservoir characterization represents a signifi-
the potential of ML techniques to improve reservoir cant shift in this landscape. Such advancements, as exempli-
characterization and optimize the development of oil fied by the hierarchical AutoML approach used in Alberta’s
and gas fields in the Dutch sector of the North Sea. Athabasca Oil Sands [20], offer promising future research
directions even though they have not been extensively applied
in the Dutch North Sea context.
II. LITERATURE REVIEW
The North Sea region, particularly the Netherlands, possesses III. DATA COLLECTION AND PREPROCESSING
extensive petroleum reserves, including proven oil reserves of A. EXPLORATORY DATA ANALYSIS
140.9 million barrels and natural gas reserves of 1.1 trillion Exploratory data analysis (EDA) techniques, including the
cubic meters [3]. Accurate prediction of reservoir properties use of kernel density estimation (KDE) plots and histograms,
is crucial because it can reduce costs, enhance efficiency, and were employed to gain insights into the dataset characteristics
boost the profitability of oil and gas production [4]. Well log- and detect any patterns or anomalies. KDE plots provide a
ging offers valuable insights into subsurface parameters such visualization of the estimated probability density function of
as porosity, shale volume, and permeability [5]. These param- the data, whereas histograms represent the frequency distribu-
eters are essential for estimating reservoir quality, predicting tion of the data. Before conducting predictive modeling, EDA
fluid flow behavior, and optimizing production strategies was performed to gain insights into the dataset. This step is
[6]. Previous studies have used empirical models and con- crucial for understanding the characteristics of the data and
ventional statistical methods, including regression analysis, identifying patterns or anomalies.
ANNs, and fuzzy logic systems, for predicting porosity and Summary statistics for each numeric column in the data
permeability from well logs [7], [8], [9]. frame, including the number of data points, mean, standard

VOLUME 11, 2023 140149

 A. Al-Fakih et al.: Reservoir Property Prediction in the North Sea Using Machine Learning

TABLE 1. This table compares the performance of different machine

learning algorithms for predictingreservoir properties.

deviation, minimum and maximum values, and percentile

values, were computed for the relevant well-log parame-
ters. Table 2 presents summary statistics for the relevant
columns.
This analysis provided a comprehensive overview of the
distribution and range of values for each well log parameter,
enabling a better understanding of the characteristics of the
dataset.
Fig. 1(a) presents the KDE plot of the Gamma Ray (GR)
log distribution for the six wells, providing insights into
the estimated probability density function of the GR values.
This plot revealed a multimodal distribution, indicating the
presence of different rock types or lithologies within the well
log data.
Fig. 1(b) illustrates the KDE plot of the Bulk Density
(RHOB) log distribution, offering a visualization of the esti-
mated probability density function for the RHOB values.
Similar to the GR log, the RHOB log distribution also exhib-
ited a multimodal pattern, suggesting the presence of multiple
rock types or lithologies within the well log data.
Fig. 1(c) displays a histogram plot of the GR log distribu-
tion, representing the frequency distribution of the GR values.
FIGURE 1. Log distribution analysis for the six target wells: (a) KDE plot
This plot further explores the distribution patterns of GR log of GR log distribution and (b) KDE plot of RHOB log distribution.
data, highlighting any notable features or variations. (c) Histogram plot of the GR log distribution and (d) the RHOB log
distribution plot.
Fig. 1(d) illustrates a histogram of the RHOB log distri-
bution, capturing the frequency distribution of the RHOB
values. This plot provides insight into the distribution char- These visualizations provide initial insights into the
acteristics of the RHOB log data, confirming the multimodal distribution and characteristics of well log data. Fur-
distribution observed in the KDE plot. ther analysis and modeling techniques will be employed

140150 VOLUME 11, 2023

A. Al-Fakih et al.: Reservoir Property Prediction in the North Sea Using Machine Learning

TABLE 2. Summary statistics of relevant well log parameters.

FIGURE 3. Outlier detection using the IF method. The orange points

represent outliers, and the blue points represent inliers. The detected
outliers were removed to enhance the accuracy and reliability of the
models for reservoir property predictions.

C. OUTLIER DETECTION
In the outlier detection process, we utilized three techniques:
isolation forest (IF), one-class super victor machine (SVM),
and local outlier factor (LOF). The IF method partitions data
into subsets to isolate anomalies, one-class SVM identifies
outliers that significantly deviate from the majority, and LOF
FIGURE 2. Sparkline indicates data completeness for well log features. assesses local density deviations of data points. These meth-
A complete line at the maximum value denotes data rows with no
missing values, ensuring data integrity for analysis and modeling across ods proved effective, with IF, particularly effective, isolated
all six wells. approximately 5-10% of outliers per well, enhancing model
reliability. Data preprocessing also involved handling miss-
ing values and scaling the data to ensure consistency and
to explore the relationships between these variables and accuracy in our analysis.
the target reservoir properties, such as the total porosity,
effective porosity, and volume of the shale. By investi- IV. METHODOLOGY
gating these relationships, we can better understand geo- A. OVERVIEW OF MACHINE LEARNING TECHNIQUES
logical formations and enhance reservoir characterization USED
efforts. In our study, we used a range of MLT to predict reservoir
properties in the North Sea’s Dutch sector. We collected well
B. DATA IMPUTATION log data from six wells, utilizing four wells for training, one
Addressing missing data, variable discarding and list-wise for testing, and one for validation. Our approach included
deletion techniques were employed. Variable discarding using ANNs, known for their prowess in pattern recognition
removes features with missing values, whereas list-wise and prediction tasks, and their ability to model complex
deletion eliminates rows with missing data, such as a relationships between inputs and output [24]. KNN, a non-
single depth level for well logs. Excluding the den- parametric classification algorithm, assigns data points to
sity correction (DRHO) feature improved the data’s classes based on their proximity, contributing to our model’s
shape and distribution. To assess the quality of the robustness against outliers [25]. Additionally, we used RF
dataset, Figure 2 shows spark lines indicating data and DT algorithms. RF, an ensemble of decision trees,
completeness. is renowned for its robustness and high-dimensional data

VOLUME 11, 2023 140151

 A. Al-Fakih et al.: Reservoir Property Prediction in the North Sea Using Machine Learning

TABLE 3. Hyperparameters used for four machine learning algorithms in a feedforward neural network with three hidden layers and
this study.
1000 neurons per layer was used for the ANN model. The
Adamax optimizer and MSE loss functions were also used,
as clarified in Table 3. Once the models were trained, they
were evaluated using a validation set to assess their accuracy
and overall performance. The evaluation results are presented
and discussed in the Results and Discussion section.

1) VALIDATION OF THE METHODS WITH THE DATA

To ensure the robustness and generalizability of our predictive
models, we used a holdout validation set to evaluate their
performance using unseen data. The holdout validation set
consisted of 10% of the total dataset, which was randomly
selected and withheld from the training process. Once the
models were trained, their performance on the holdout val-
idation set was evaluated using the following performance
handling [26], while DT offers straightforward, interpretable
metrics: MAE, RMSE, and R-squared.
models, beneficial for understanding decision processes [27].
By comparing the performance of the models using these
Each of these algorithms—ANNs, KNN, DT and
metrics, we comprehensively evaluated their effectiveness in
RF —has been successfully applied in reservoir property
predicting target variables [31], [32].
prediction, leveraging their distinct strengths to develop a
robust predictive model [28], [29], [30].
The following types of well logs were used as inputs for 2) CONTROL PARAMETERS FOR EACH ALGORITHM
our models: GR, RHOB, sonic transit time (DT), and neu- Table 3 lists the control parameters used for each MLT. The
tron porosity (NPHI). The outputs of our models were the control parameters are the hyperparameters tuned to optimize
predicted values for the total porosity, effective porosity, and the performance of an ML model.
volume of shale. The choice of control parameters can significantly
affect the MLT performance. In this study, we used a grid
1) NOVELITY OF METHODS USED IN THIS TOPIC search approach to tune the control parameters for each
Our study stands out for its novel use of a combination of four algorithm. We evaluated a range of values for each control
distinct ML algorithms (ANNs, KNN, RF, and DT) to predict parameter and selected the values that resulted in the best
reservoir properties. This multifaceted approach enables us to performance for the validation set.
harness the unique advantages of each algorithm, resulting in
a more robust and accurate model. Additionally, our study is C. MODEL EVALUATION
distinguished by its focus on the North Sea sedimentary basin, Model performance was evaluated using MAE, RMSE, and
a region with complex geology and diverse reservoir types. R-squared metrics. These metrics serve as measures of accu-
This challenging context necessitates sophisticated modeling racy and provide insights into the predictive capabilities of
approaches, for which our combination of advanced ML models. To ensure the robustness of our models and prevent
techniques and a carefully curated dataset is particularly well- overfitting, we incorporated techniques like early stopping
suited. The hyperparameters used for each ML algorithm, and L2 regularization during the training phase. The detailed
as detailed in Table 3, were carefully selected to optimize the results of this evaluation, reflecting each model’s perfor-
model performance. mance, are presented, and thoroughly discussed in the Results
and Discussion section.
B. MODEL SELECTION AND TRAINING
Prior to training the models, EDA was conducted to under- D. WORKFLOW
stand data characteristics. We then choose a diverse set of A workflow chart (Figure 4) was created to provide a clear
algorithms: ANNs, KNN, RF, and DT, each of which has and comprehensive overview of the methodology. The work-
unique capabilities for our predictive modeling. The training flow chart illustrates the step-by-step process undertaken in
process involved using well log data from six wells, and this study, starting with data collection and preprocessing,
allocating four wells for training, one for testing, and one followed by model selection, training, testing, validation,
for validation purposes. We employed a grid search approach evaluation, and conclusion with result interpretation. This
to fine-tune the hyperparameters of each model to ensure visual representation aids in understanding the research flow
optimal performance [28], [29], [30]. and ensures clarity and consistency of the methodology.
This involved evaluating a range of values for each By following this established workflow, our study aimed
hyperparameter and selecting the values that resulted in to effectively predict reservoir properties in the North Sea
the best performance for the validation set. Specifically, sedimentary basin using the MLT. Rigorous steps of model

140152 VOLUME 11, 2023

A. Al-Fakih et al.: Reservoir Property Prediction in the North Sea Using Machine Learning

TABLE 5. performance comparison of effective porosity models.

TABLE 6. Performance comparison of volume of shale algorithms.

FIGURE 4. Workflow chart displaying the methodology for predicting

reservoir properties in the North Sea sedimentary basin. The workflow
encompassed data collection, preprocessing, model selection (ANN, KNN,
RF, and DT), training on well logs, and evaluation using six wells.
Performance metrics (MAE, RMSE, and R-squared) were employed to
measure the accuracy and evaluate the models. This provided a concise
and informative overview of the methodology employed in this study.

TABLE 4. Performance comparison of porosity models. To understand the reasons for the superior performance
of the ANN and RF models, a feature importance analysis
was conducted. Feature importance analysis for the poros-
ity models revealed that variables such as lithology, clay
content, and grain size had the highest importance. These
geological attributes and physical properties play crucial roles
in determining the porosity of rock formations. The ANN
and RF models were able to capture the complex relation-
ships between these variables and porosity, leading to more
accurate predictions compared with the other models.
Table 5 presents a performance comparison of the effective
porosity models, displaying the MAE, RMSE, and R-squared
selection, training, and evaluation provide a robust framework values for each model. The ANN model exhibited the low-
for the development of accurate predictive models. est MAE and RMSE values (0.01), indicating its superior
predictive accuracy. Similarly, the RF model achieved a low
V. RESULTS AND DISCUSSION MAE and RMSE of 0.01, further highlighting its capability
A. MODEL PERFORMANCE COMPARISON to predict effective porosity accurately. The R-squared values
Among the models evaluated for predicting the porosity, of 0.89 for both ANN and RF signify a strong correlation
effective porosity, and volume of shale, the ANN and RF between the predicted and actual effective porosity values.
models consistently demonstrated superior performances. Feature importance analysis of the effective porosity mod-
In addition to their high prediction accuracy, an analysis of els revealed that variables such as matrix permeability, water
feature importance provides insights into why these models saturation, and mineral composition played a crucial role
outperform others. in predicting effective porosity. These factors contribute
Table 4 presents a performance comparison of the porosity significantly to the flow of fluids through rock forma-
models, highlighting the MAE, RMSE, and R-squared values tions, thereby affecting the effective porosity. The ANN
for each model. The ANN model achieved the lowest MAE and RF models captured the intricate relationships between
and RMSE values of 0.01, indicating its ability to make these variables and effective porosity, resulting in superior
highly accurate predictions. Similarly, the RF model exhib- predictive performance.
ited low MAE and RMSE values (0.07), further confirming Table 6 presents the performance comparison of the vol-
its predictive capability. The R-squared values of 0.98 for the ume of shale models, displaying the MAE, RMSE, and
ANN model and 0.99 for the RF model demonstrate a strong R-squared values for each model. The ANN and RF models
correlation between the predicted and actual porosity values. outperformed the other models, achieving the lowest MAE

VOLUME 11, 2023 140153

 A. Al-Fakih et al.: Reservoir Property Prediction in the North Sea Using Machine Learning

and RMSE values, indicating their higher accuracy in estimat- reliable predictions. The cross plots provide visual evidence
ing the volume of shale. Moreover, the RF model exhibited of the models’ performance in accurately predicting the tar-
a high R-squared value, suggesting a strong correlation get variables, with the ANN and RF models demonstrating
between the predicted and actual values. better alignment between actual and estimated values. This
The feature importance analysis for the volume of shale indicates their effectiveness in estimating reservoir properties
models revealed that variables such as organic matter con- in the North Sea sedimentary basin.
tent, mineralogy, and compaction characteristics were the
most influential in predicting the volume of shale. These C. PLOTS OF ACTUAL VERSUS ESTIMATED VALUES WITH
variables provide insights into the composition and physical DEPTH
properties of the rock formations, which directly affect the Figure 6 shows the plot of actual and predicted values
volume of shale present. The ANN and RF models effectively of porosity, effective porosity, and shale volume along the
captured the complex relationships between these variables depth axis for the ANN and RF models. These plots pro-
and the volume of shale, leading to their superior predictive vide valuable information on how the predicted values for
performance. each target variable vary with depth, facilitating an enhanced
Overall, our analysis confirms the superior ability of the understanding of reservoir properties’ spatial distribution.
ANN and RF models in accurately forecasting key reservoir They show a gradual decrease in porosity with depth for
characteristics in the North Sea sedimentary basin. Their high both the ANN and RF models. This trend is consistent with
performance is a result of proficiently capturing and analyz- the geological understanding of the North Sea sedimentary
ing the complex interplay of specific geological attributes, basin, where porosity tends to decrease with depth due to
physical properties, and other pertinent factors, thereby pro- compaction and diagenesis.
viding reliable and precise predictions for each variable under They also demonstrate the models’ ability to capture the
study. variations in porosity, effective porosity, and volume of shale
at different depths, which is essential for reservoir character-
B. MODEL INTERPRETATION AND APPLICATION ization and management in the North Sea sedimentary basin.
As shown in the cross plots (Figure 5a), both the ANN and
RF models exhibit a strong correlation between the actual 1) DISCUSSION OF THE PLOTS
and predicted values for the target variable porosity. The data Figure 6’s plots show a consistent trend: estimated values
points closely follow the regression line, indicating that the closely align with actual values as the depth increases, indi-
predictions are highly accurate. However, it is worth noting cating accurate model predictions of reservoir properties’
that the RF model consistently falls below the regression line, spatial distribution.
suggesting a systematic bias in underestimating the porosity For example, the plot of porosity shows a gradual
values. This bias should be further investigated and addressed decrease in porosity with depth for both the ANN and
to ensure the accuracy and reliability of the RF model’s RF models. This trend is consistent with the geological
predictions. understanding of the North Sea sedimentary basin, where
Similarly, Figure 5b the cross plots for effective porosity porosity tends to decrease with depth due to compaction and
show a tight clustering of data points along the regression line diagenesis.
for both the ANN and RF models. This demonstrates a strong The plots of effective porosity and shale volume also
correlation between the actual and predicted values, implying show consistent trends with depth. The effective porosity plot
a high degree of accuracy in predicting effective porosity. shows a gradual decrease in effective porosity with depth,
On the other hand, the DT and KNN models also show which is expected due to the increasing presence of irre-
reasonable alignment, albeit with some scattered data points. ducible water at greater depths. The shale volume plot shows
These deviations may be attributed to the differences in mod- a gradual increase in shale volume with depth, which is also
eling approaches and their ability to capture the complex consistent with the geological understanding of the basin.
relationships present in the data.
The DT model, being sensitive to slight changes in the 2) QUANTITATIVE ANALYSIS OF THE RESULTS
training data, may result in different splits and potentially less To quantify the agreement between the actual and predicted
accurate predictions. On the other hand, KNN models rely values, we calculated the correlation coefficients between
on the proximity of neighboring data points for predictions, the two for each target variable. The correlation coefficients
which may not always capture the underlying patterns in were all above 0.9, indicating a good correlation between the
the data accurately. These factors contribute to the observed actual and predicted values. This further confirms the models’
deviations in the data points for the DT and KNN models accuracy in predicting reservoir properties.
compared to the ANN and RF models.
The ANN and RF models have consistently demonstrated 3) IMPLICATIONS FOR RESERVOIR DEVELOPMENT AND
superior performance in accurately predicting porosity, effec- PRODUCTION
tive porosity, and shale volume. This is due to their ability to Information on the spatial distribution of porosity, effective
capture complex relationships within the dataset and provide porosity, and shale volume can be used to identify zones

140154 VOLUME 11, 2023

A. Al-Fakih et al.: Reservoir Property Prediction in the North Sea Using Machine Learning

low porosity and effective porosity and high shale volume

would be more suitable for injection wells.

4) LIMITATIONS AND FUTURE RESEARCH DIRECTIONS

The models developed in this study were on a limited dataset
from the North Sea sedimentary basin. Further validation is
needed using data from other reservoirs before the models
can be widely deployed. Additionally, future research could
explore the use of ML to predict other reservoir properties,
such as permeability and saturation.
This information is essential for reservoir characterization
and management and can be used to identify zones suitable
for injection or production wells.

D. INFLUENCE OF DATA IMPUTATION AND OUTLIER

DETECTION
The successful application of data imputation and outlier
detection techniques contributes to the improved accuracy
of the models. These preprocessing steps help address miss-
ing values and anomalies in the dataset, ensuring a more
comprehensive and reliable training process.
In this study, we used the following data imputation and
outlier detection techniques:
Data imputation: We used the mean imputation method to
handle missing values. This method imputes missing values
with the mean value of the feature for non-missing values.
Outlier detection: We used the interquartile range (IQR)
method to identify outliers. Outliers were defined as data
points that fell outside of the 1.5 IQR range.
We evaluated the models’ accuracy with and without
data imputation and outlier detection to understand these
preprocessing steps’ impact. The results showed that data
imputation and outlier detection improved the accuracy of
the models for all three target variables (porosity, effective
porosity, and shale volume).
For example, the MAE for the ANN model was reduced by
10% when data imputation and outlier detection were used.
This suggests that data imputation and outlier detection can
help to improve the accuracy of ML models by addressing
missing values and anomalies in the dataset.

1) EXAMPLES OF IMPROVED ACCURACY

One example of how data imputation and outlier detection
improved the accuracy of the models is in the case of shale
volume prediction. The shale volume dataset contained sev-
eral outliers, which were identified and corrected using the
IQR method. After outlier detection, the accuracy of the ANN
model for shale volume prediction improved by 15%.
FIGURE 5A. Cross-plot of actual and predicted porosity values for the Another example is in the case of the effective porosity
ANN and RF models. The data points closely follow the diagonal line, prediction. The effective porosity dataset contained several
indicating a high degree of accuracy in the predictions. The R-squared
values for the ANN and RF models are 0.98, demonstrating their strong missing values, which were imputed using the mean imputa-
predictive performance. tion method. After data imputation, the accuracy of the RF
model for effective porosity prediction improved by 20%.
suitable for injection or production wells. For example, zones These examples show that data imputation and outlier
with high porosity and effective porosity and low shale vol- detection can significantly impact the accuracy of MLTs.
ume would be ideal for production wells, while zones with By addressing missing values and anomalies in the dataset,

VOLUME 11, 2023 140155

 A. Al-Fakih et al.: Reservoir Property Prediction in the North Sea Using Machine Learning

FIGURE 5C. Cross-plot of the actual and predicted volume of shale values
for the ANN and RF models. The data points follow the diagonal line
closely, suggesting accurate predictions of shale volume. The R-squared
values for the ANN and RF models are 0.99, demonstrating their strong
predictive performance.

FIGURE 5B. Cross-plot of actual and predicted effective porosity values

for the ANN and RF models. The tight clustering of data points along the
regression line indicates a strong correlation between the actual and E. IMPORTANCE OF GEOLOGICAL ATTRIBUTES AND
predicted values, implying a high degree of accuracy in predicting PHYSICAL PROPERTIES
effective porosity.
It is worth noting that our findings align with previous
works in literature. Studies have emphasized the signif-
these preprocessing steps can help to improve the quality of icance of the geological attributes and physical proper-
the training data and lead to more accurate predictions. ties, such as lithology, clay content, grain size, matrix

140156 VOLUME 11, 2023

A. Al-Fakih et al.: Reservoir Property Prediction in the North Sea Using Machine Learning

FIGURE 6. Comparison of actual and estimated property values (a) Plot of actual and estimated porosity values along the depth axis for the ANN and RF
models. The estimated porosity values closely follow the actual values as the depth increases, accurately capturing the porosity’s spatial distribution.
(b) A plot of actual and estimated effective porosity values along the depth axis for the ANN and RF models. The plots demonstrate a strong agreement
between the actual and predicted effective porosity values, suggesting accurate capture of variations in effective porosity across different depths. (c) The
plot of the actual and estimated volume of shale values along the depth axis for the ANN and RF models. The plots show a consistent trend where the
estimated shale volume closely aligns with the actual values as the depth increases, indicating accurate capture of the spatial distribution of shale
volume.

VOLUME 11, 2023 140157

 A. Al-Fakih et al.: Reservoir Property Prediction in the North Sea Using Machine Learning

permeability, water saturation, organic matter content, limited data availability in the region, our models exhibit
mineralogy, and compaction characteristics, in influenc- exceptional performance in predicting total porosity, effective
ing porosity, effective porosity, and volume of shale. porosity, and shale volume. This success underscores the
These factors have been widely recognized as key con- importance of extensive data preprocessing steps, including
tributors to the variations observed in these reservoir exploratory data analysis, missing data imputation, and out-
properties. lier detection, in ensuring the robustness and accuracy of
The feature importance analysis conducted revealed that these models.
lithology, clay content, and grain size were among the most The workflow presented by the ANN and RF models
important features for predicting porosity. This is consistent offers a valuable tool for quantitatively estimating reservoir
with the geological understanding that porosity is influ- characteristics and optimizing field development strategies,
enced by the type of rock, the amount of clay present, and particularly in regions with limited data or where core data
the size of the grains. For example, sandstones and lime- is unavailable. The findings of this study have significant
stones typically have higher porosity than shales, due to their implications for reservoir characterization and management
larger grain size and lower clay content. Matrix permeability, not only in the North Sea sedimentary basin but also in other
water saturation, and mineral composition emerged as crucial geologically complex regions worldwide.
variables for predicting effective porosity. This is because However, it is essential to acknowledge the study’s lim-
effective porosity is a measure of the pore space that is itations. The models were trained and validated using data
available for fluid flow. Matrix permeability is a measure from a single sedimentary basin, and further research is
of the ease with which fluids can flow through the rock needed to assess their performance in basins with differ-
matrix, while water saturation is the fraction of pore space ent geological characteristics. Additionally, future studies
that is occupied by water. Mineral composition can also affect may explore the extension of these models to include
effective porosity, as some minerals are more porous than additional reservoir properties, such as permeability and
others. saturation.
Organic matter content, mineralogy, and compaction char- In conclusion, this research highlights how MLT can
acteristics were identified as influential factors for estimating enhance and optimize oil and gas field development and
the volume of shale. Shale is a sedimentary rock composed production. The accuracy of the ANN and RF models in pre-
of fine-grained particles, such as clay, silt, and organic mat- dicting reservoir properties, their ability to provide insights
ter. The volume of shale in a reservoir can be affected by into spatial property distributions, and their relevance in
the amount of organic matter present, the type of minerals reservoir characterization and management in challenging
present, and the degree to which the rock has been compacted. geological environments make them valuable tools in the
The importance of these geological attributes and physical energy industry.
properties is further supported by previous studies in litera-
ture. For example, [31] found that lithology, clay content, and
grain size were the key factors influencing porosity in fluvial VII. FUTURE WORK
reservoirs in the Triassic Skagerrak Formation in the Cen- In future studies, we aim to extend the validation of our
tral North Sea, UK [25]. Similarly, [32] found that organic MLT using data from a variety of other reservoirs. This
matter content, mineralogy, and compaction characteristics broader validation is crucial to ensure the generalizability and
were influential factors for estimating the volume of shale robustness of our models across different geological settings.
in shale oil reservoirs in the Jurassic Lianggaoshan Forma- We also plan to explore the application of MLT in predicting
tion of the Yingshan Gas Field in central Sichuan Basin, other vital reservoir properties, such as permeability and
China [26]. saturation, which are critical for comprehensive reservoir
Incorporating key geological attributes and physical prop- characterization.
erties, the ANN and RF models effectively captured com- Technological advancements in ML offer exciting oppor-
plex relationships, leading to accurate predictions of target tunities for enhancing the accuracy of reservoir property
variables. The ability of the models to effectively cap- predictions. We intend to investigate the integration of newer
ture the spatial distribution of porosity, effective poros- MLT and advanced data processing techniques, such as deep
ity, and volume of shale along the depth axis can be AutoML, to refine our predictive models further.
attributed to their capability to learn and represent these Moreover, recognizing the importance of diverse data
relationships. in model development, we propose to engage in collab-
orative research with other institutions and researchers.
VI. CONCLUSION Such collaborations can provide access to a wider range
In summary, this study demonstrates the potential of of datasets, facilitating more extensive validation and
advanced MLT, specifically artificial neural networks (ANN) potentially leading to breakthroughs in the application of
and random forest (RF), to accurately predict reservoir machine learning in reservoir characterization. These col-
properties in the geologically complex North Sea sedimen- laborative efforts would not only enrich our research but
tary basin. Despite the unique geological characteristics and also contribute to the broader scientific community by

140158 VOLUME 11, 2023

A. Al-Fakih et al.: Reservoir Property Prediction in the North Sea Using Machine Learning

sharing insights and advancements in this rapidly evolving [15] J. Zhang, C. Li, Y. Yin, J. Zhang, and M. Grzegorzek, ‘‘Applications of arti-
field. ficial neural networks in microorganism image analysis: A comprehensive
review from conventional multilayer perceptron to popular convolutional
neural network and potential visual transformer,’’ Artif. Intell. Rev., vol. 56,
ACKNOWLEDGMENT no. 2, pp. 1013–1070, Feb. 2023.
The authors extend their appreciation to artificial intelli- [16] M. W. Raheem, ‘‘Prediction by reservoir porosity using micro-seismic
attribute analysis by machine learning algorithms in an Iraqi Oil Field,’’
gence for its contribution to certain sections of this article. Turkish J. Comput. Math. Educ., vol. 12, no. 14, pp. 3324–3332,
AI technologies, particularly the ChatGPT language model 2021.
developed by OpenAI and the Bard Google tool for gram- [17] F. H. Kasim, W. N. S. W. M. Zainudin, B. P. Kantaatmadja, N. A. Siddiqui,
A. Sidek, and N. A. N. Yahaya, ‘‘Machine learning assisted reservoir
mar correction, played a role in refining grammar and text properties prediction in a brownfield offshore Malaysia,’’ presented at
coherence. They also acknowledge the support of the NLOG the Int. Petroleum Technol. Conf., Riyadh, Saudi Arabia, Feb. 2022, doi:
website and Utrecht University for providing the dataset, and 10.2523/IPTC-22409-MS.
[18] W. Liu, Z. Chen, Y. Hu, and L. Xu, ‘‘A systematic machine learning
the SDAIA-KFUPM-JRC-AI Research Center for technical method for reservoir identification and production prediction,’’ Petroleum
support. Sci., vol. 20, no. 1, pp. 295–308, Feb. 2023.
[19] D. Ivlev, ‘‘Reservoir prediction by machine learning methods on the
well data and seismic attributes for complex coastal conditions,’’ 2023,
REFERENCES arXiv:2301.03216.
[1] M. A. Ishak, A. H. A. Latiff, E. T. W. Ho, M. I. A. Fuad, N. W. Tan, [20] Y. Mubarak and A. Koeshidayatullah, ‘‘Hierarchical automated machine
M. Sajid, and E. Elsebakhi, ‘‘Advanced elastic and reservoir properties learning (AutoML) for advanced unconventional reservoir characteriza-
prediction through generative adversarial network,’’ Appl. Sci., vol. 13, tion,’’ Sci. Rep., vol. 13, no. 1, p. 13812, Aug. 2023.
no. 10, p. 6311, May 2023. [21] X. Tian, H. Huang, S. Cheng, C. Wang, P. Li, and Y. Hao,
[2] Z. Tariq, M. S. Aljawad, A. Hasan, M. Murtaza, E. Mohammed, ‘‘A carbonate reservoir prediction method based on deep learning and
A. El-Husseiny, and A. Abdulraheem, ‘‘A systematic review of data multiparameter joint inversion,’’ Energies, vol. 15, no. 7, p. 2506,
science and machine learning applications to the oil and gas indus- Mar. 2022.
try,’’ J. Petroleum Explor. Prod. Technol., vol. 11, pp. 4339–4374, [22] D. A. Otchere, T. O. A. Ganat, R. Gholami, and S. Ridha, ‘‘Appli-
Sep. 2021. cation of supervised machine learning paradigms in the prediction
[3] D. S. Hamilton and J. P. Quinlan, ‘‘The transatlantic economy 2023,’’ of petroleum reservoir properties: Comparative analysis of ANN
Transatlantic Bus. Council and AmCham-EU, Washington, DC, USA, and SVM models,’’ J. Petroleum Sci. Eng., vol. 200, May 2021,
Tech. Rep. 2023, 2023. Art. no. 108182.
[4] M. Ali, R. Jiang, H. Ma, H. Pan, K. Abbas, U. Ashraf, and J. Ullah, [23] A. Ali, A. M. Bello, and J. Raymond, ‘‘Machine learning algo-
‘‘Machine learning—A novel approach of well logs similarity based on rithms for predicting reservoir porosity using stratigraphic-dependent
synchronization measures to predict shear sonic logs,’’ J. Petroleum Sci. parameters,’’ Global J. Comput. Sci. Technol., vol. 22, pp. 15–25,
Eng., vol. 203, Aug. 2021, Art. no. 108602. May 2022.
[5] A. Al-Fakih and S. Kaka, ‘‘Application of artificial intelligence in [24] M. Mahdaviara, M. Sharifi, and M. Ahmadi, ‘‘Toward evaluation and
static formation temperature estimation. Arabian J. Sci. Eng., 2023, doi: screening of the enhanced oil recovery scenarios for low permeabil-
10.1007/s13369-023-08096-x. ity reservoirs using statistical and machine learning techniques,’’ Fuel,
[6] J. Lai, G. Wang, Q. Fan, X. Pang, H. Li, F. Zhao, Y. Li, X. Zhao, Y. Zhao, vol. 325, Oct. 2022, Art. no. 124795.
Y. Huang, M. Bao, Z. Qin, and Q. Wang, ‘‘Geophysical well-log evalu- [25] J. M. A. S. Naji, G. H. Abdul-Majeed, and A. K. Alhuraishawy,
ation in the era of unconventional hydrocarbon resources: A review on ‘‘Intelligent approach for investigating reservoir heterogeneity effect on
current status and prospects,’’ Surv. Geophys., vol. 43, no. 3, pp. 913–957, sonic shear wave,’’ J. Petroleum Res. Stud., vol. 13, no. 1, pp. 56–73,
Jun. 2022. Mar. 2023.
[7] Z. Zhang, H. Zhang, J. Li, and Z. Cai, ‘‘Permeability and porosity [26] H. Wang and S. Chen, ‘‘Insights into the application of machine learning in
prediction using logging data in a heterogeneous dolomite reservoir: reservoir engineering: Current developments and future trends,’’ Energies,
An integrated approach,’’ J. Natural Gas Sci. Eng., vol. 86, Feb. 2021, vol. 16, no. 3, p. 1392, Jan. 2023.
Art. no. 103743. [27] S. Chaki, A. Routray, and W. K. Mohanty, ‘‘Application of machine learn-
[8] M. Matinkia, R. Hashami, M. Mehrad, M. R. Hajsaeedi, and A. Velay- ing algorithms for petroleum reservoir characterization,’’ in Handbook
ati, ‘‘Prediction of permeability from well logs using a new hybrid of Petroleum Geoscience: Exploration, Characterization, and Exploita-
machine learning algorithm,’’ Petroleum, vol. 9, no. 1, pp. 108–123, tion of Hydrocarbon Reservoirs. Hoboken, NJ, USA: Wiley, 2022,
Mar. 2023. pp. 6–20.
[9] H. Khan, A. Srivastav, A. Kumar Mishra, and T. Anh Tran, ‘‘Machine [28] H. Ghorbani, D. A. Wood, A. Choubineh, A. Tatar, P. G. Abarghoyi,
learning methods for estimating permeability of a reservoir,’’ Int. J. Syst. M. Madani, and N. Mohamadian, ‘‘Prediction of oil flow rate through
Assurance Eng. Manage., vol. 13, no. 5, pp. 2118–2131, Oct. 2022. an orifice flow meter: Artificial intelligence alternatives compared,’’
[10] I. H. Sarker, ‘‘Machine learning: Algorithms, real-world applications and Petroleum, vol. 6, no. 4, pp. 404–414, Dec. 2020.
research directions,’’ Social Netw. Comput. Sci., vol. 2, no. 3, p. 160, [29] G. Zhang, S. Davoodi, S. S. Band, H. Ghorbani, A. Mosavi, and
May 2021. M. Moslehpour, ‘‘A robust approach to pore pressure prediction applying
[11] A. Al-Fakih, A. F. Ibrahim, S. Elkatatny, and A. Abdulraheem, ‘‘Estimating petrophysical log data aided by machine learning techniques,’’ Energy
electrical resistivity from logging data for oil wells using machine learn- Rep., vol. 8, pp. 2233–2247, Nov. 2022.
ing,’’ J. Petroleum Explor. Prod. Technol., vol. 13, no. 6, pp. 1453–1461, [30] M. Farsi, H. S. Barjouei, D. A. Wood, H. Ghorbani, N. Mohamadian,
Jun. 2023. S. Davoodi, H. R. Nasriani, and M. A. Alvar, ‘‘Prediction of oil flow
[12] R. Rezaee and J. Ekundayo, ‘‘Permeability prediction using machine learn- rate through orifice flow meters: Optimized machine-learning techniques,’’
ing methods for the CO2 injectivity of the precipice sandstone in Surat Measurement, vol. 174, Apr. 2021, Art. no. 108943.
Basin, Australia,’’ Energies, vol. 15, no. 6, p. 2053, Mar. 2022. [31] O. E. Aro, S. J. Jones, N. S. Meadows, J. Gluyas, and D. Charlaftis,
[13] J. Sun, R. Zhang, M. Chen, B. Chen, X. Wang, Q. Li, and L. Ren, ‘‘The importance of facies, grain size and clay content in controlling
‘‘Identification of porosity and permeability while drilling based on fluvial reservoir quality—An example from the Triassic Skagerrak for-
machine learning,’’ Arabian J. Sci. Eng., vol. 46, no. 7, pp. 7031–7045, mation, Central North Sea, UK,’’ Petroleum Geosci., vol. 29, no. 2, 2023,
Jul. 2021. Art. no. petgeo2022-043, doi: 10.1144/petgeo2022-043.
[14] F. Mohammadinia, A. Ranjbar, M. Kafi, and R. Keshavarz, ‘‘Applica- [32] Y. Wang, H. Deng, Z. Wang, X. Wang, Q. Cao, D. Cheng, Y. Zhu, and
tion of machine learning algorithms in classification the flow units of A. Li, ‘‘Characteristics and factors influencing pore structure in shale oil
the Kazhdumi reservoir in one of the oil fields in Southwest of Iran,’’ reservoirs of different lithologies in the Jurassic Lianggaoshan Formation
J. Petroleum Explor. Prod. Technol., vol. 13, no. 6, pp. 1419–1434, of the Yingshan gas field in Central Sichuan Basin,’’ Minerals, vol. 13,
Jun. 2023. no. 7, p. 958, Jul. 2023.

VOLUME 11, 2023 140159

 A. Al-Fakih et al.: Reservoir Property Prediction in the North Sea Using Machine Learning

ABDULRAHMAN AL-FAKIH (Member, IEEE) ARDIANSYAH I. KOESHIDAYATULLAH (Mem-

was born in Sana’a, Yemen, in April 1989. ber, IEEE) received the Ph.D. degree in basin
He received the bachelor’s degree in petroleum and petroleum geosciences from The University of
engineering from the Hadhramout University of Manchester.
Science and Technology, Yemen, in 2012, and He was a Postdoctoral Researcher with Stanford
the master’s degree in energy resources from the University. He is currently an Assistant Professor
China University of Geosciences, Beijing, in 2020. with the College of Petroleum Engineering and
He is currently pursuing the Ph.D. degree in AI Geosciences, King Fahd University of Petroleum
and geophysics with the King Fahd University and Minerals (KFUPM), Saudi Arabia. He is also
of Petroleum and Minerals, Dhahran. He com- an accomplished Geoscientist and an Assistant
pleted an internship with Weatherford Oil Tool Middle East Ltd., Yemen, Professor with expertise in carbonate sedimentology, geochemistry, and
from 2013 to 2014. He taught Arabic with Beijing Language and Culture artificial intelligence (AI). His publications cover topics from carbonate
University, from 2016 to 2020. He has been a Teaching Assistant with classification with deep learning to AI techniques for fracture and poros-
King Fahd University of Petroleum and Minerals, since 2021. With over ity quantification. His multidisciplinary approach fusing geology and AI
two years of oil and gas industry experience, he has published four papers showcases his contributions to geosciences, AI applications, and petroleum
and contributed to several research projects. He is proficient in Python and engineering. He is an asset to advancing knowledge and industry applications
MATLAB and known for his analytical skills and teamwork. His research in his field. His research interests include using sedimentary-geochemistry
interests include machine learning applications in geophysics, petroleum analysis and numerical models to uncover past conditions, and leveraging AI
engineering, and geothermal energy. His interests extend to programming, for sedimentary rock interpretation in reservoirs.
photography, and cooking. He is involved in professional societies, such as Dr. Koeshidayatullah has received awards, such as the Imperial Barrel
SPE, SEG, IADC, SPWLA, and AAPG. Award and the KFUPM scholarships, actively contributes to academia and
mentors’ students.

SANLINN I. KAKA received the Ph.D. degree in

earth sciences from Carleton University, Ottawa,
ON, Canada, in 2006. He is currently a Fac-
ulty Member and a Graduate Coordinator with
the Department of Geosciences, King Fahd Uni-
versity of Petroleum and Minerals. His research
interests include engineering seismology, reservoir
characterization and monitoring, ground motions
relations, and near surface geophysics. His recent
research focuses on the applications of micro-
seismic monitoring systems, enhancement, detection, and localization of
microseismic events as well as understanding fracture growth and the role
of pre-existing fractures during multi-stage hydraulic fracture stimulation of
shale gas reservoirs.

140160 VOLUME 11, 2023

View publication stats