KWAME NKRUMAH UNIVERSITY OF SCIENCE AND

TECHNOLOGY KUMASI, GHANA

COLLEGE OF ENGINEERING

DEPARTMENT OF PETROLEUM ENGINEERING

PROJECT REPORT

OPTIMIZATION OF RATE OF PENETRATION (ROP) USING MACHINE

LEARNING.

PROJECT REPORT SUBMITTED IN FULL FULFILMENT OF THE

REQUIREMENTS FOR THE MSC (ENG.) DEGREE

BY

MICHAEL KWOFIE

PG4081724

SUPERVISOR: SAMUEL ERZUAH, PHD

DECEMBER, 2025
DECLARATION
I declare that this project is my own work. It is been submitted in partial fulfillment of the
requirements for the Master of Science degree in Petroleum Engineering at the Kwame
Nkrumah University of Science and Technology, Ghana.
This work has not been presented elsewhere for any academic qualification, and all sources
of information have been duly acknowledged.

.......................................................................................
(Signature of Candidate)

.................................. day of .................................2025

i
ABSTRACT
Rate of penetration (ROP) serves as a key indicator of drilling efficiency, with direct
implications for operational costs and project timelines in the oil and gas sector. Effective
ROP optimization necessitates a comprehensive understanding of the complex, non-linear
interactions between drilling parameters and formation response. This study develops and
compares machine learning models for ROP prediction and optimization, utilizing field
drilling data from polycrystalline diamond compact (PDC) bit operations.

A systematic four-phase methodology was implemented, encompassing feature analysis,

model development, parameter optimization, and validation. The dataset included 43
observations across 31 variables from a single drilling section. Three machine learning
algorithms were developed and evaluated: Random Forest, Extreme Gradient Boosting
(XGBoost), and Artificial Neural Network (ANN). Feature analysis employed Pearson
correlation, Random Forest feature importance, and SHapley Additive exPlanations (SHAP)
values to identify key factors influencing ROP. Grid search optimization determined
optimal drilling parameters, with results validated through sensitivity analysis and
comparison to established drilling mechanics principles.

Random Forest demonstrated the highest predictive performance, achieving a coefficient of

determination (R2) of 0.904, a mean absolute error of 1.74 m/hr, and a mean absolute
percentage error of 9.2%. This performance surpassed both XGBoost (R2 = 0.813) and
ANN (R2 = 0.751). Feature analysis identified measured depth, rotary speed (RPM), and
weight on bit (WOB) as the most influential parameters, collectively accounting for
approximately 67% of total feature importance. Notably, all three models converged on
identical optimal drilling parameters: WOB of 158 klbs, flow rate of 6.55 GPM, and RPM
of 2370, which correspond to the upper bounds of observed operational ranges. This
convergence across distinct model architectures strongly suggests that the identified
optimum represents a genuine global maximum rather than a model-specific artefact.

The results indicate that Random Forest is particularly effective for ROP prediction with
limited drilling datasets, due to its inherent regularization through ensemble averaging. The
multi-method feature analysis approach offers complementary insights into both linear and
non-linear parameter relationships. These findings have practical implications for real-time
drilling optimization, indicating that operators may improve ROP by operating closer to the
identified optimal parameter combinations. This research contributes to the expanding

ii
literature on data-driven drilling optimization and establishes a methodological framework
applicable to similar drilling datasets.

Keywords: Rate of penetration, machine learning, Random Forest, drilling optimization,

feature analysis, SHAP, XGBoost, artificial neural network

iii
ACKNOWLEDGEMENT
I wish to express my profound gratitude to the Almighty God for providing guidance
throughout this project.
I sincerely appreciate Dr. Samuel Erzuah for his invaluable guidance and expert supervision.
His expertise in Petroleum Engineering was instrumental in shaping this work.
I am also deeply grateful to the faculty and staff of the Petroleum Engineering Department
at KNUST for fostering an academic environment and providing access to essential
resources for this study.
Special thanks to all who offered words of encouragement and assistance throughout this
project. Your contributions, however small, have been deeply appreciated.

iv
DEDICATION
This work is dedicated to:
The Almighty God, for His endless grace and guidance.
To my father, Alex Blay Kwofie, in recognition of his steadfast support and invaluable
contribution to my academic pursuit.
And to the esteemed professionals of the Petroleum Industry, whose dedication to
excellence serves as an inspiration to aspiring engineers.

v
 TABLE OF CONTENTS
Contents Page Number
DECLARATION ..................................................................................................................i
ABSTRACT .........................................................................................................................ii
ACKNOWLEDGEMENT .................................................................................................iv
TABLE OF CONTENTS .................................................................................................... v
DEDICATION ..................................................................................................................... v
LIST OF TABLES .............................................................................................................xi
LIST OF FIGURES ..........................................................................................................xii
NOMENCLATURE ........................................................................................................ xiii
CHAPTER ONE: INTRODUCTION ............................................................................... 1
1.1 Background to the Study.............................................................................................. 1
1.2 Problem Statement ....................................................................................................... 5
1.3 Research Objectives ..................................................................................................... 7
1.4 Significance of the Study ............................................................................................. 8
1.5 Scope and Limitations ................................................................................................. 8
1.6 Organization of the Thesis ........................................................................................... 9
CHAPTER TWO: LITERATURE REVIEW ................................................................ 10
2.1 Introduction ................................................................................................................ 10
2.2 Rate of Penetration: Concept and Importance ........................................................... 10
2.3 Factors Influencing Rate of Penetration .................................................................... 11
2.3.1 Controllable Factors ............................................................................................ 11
2.3.2 Uncontrollable Factors ........................................................................................ 13
2.4 Traditional Empirical ROP Models ........................................................................... 14
2.4.1 Maurer Model (1962) .......................................................................................... 14
2.4.2 Bingham Model (1965) ....................................................................................... 15
2.4.3 Bourgoyne and Young Model (1974) ................................................................. 15
2.4.4 Warren Model (1987) .......................................................................................... 17
2.4.5 Winters, Warren and Onyia Model (1987) .......................................................... 17
2.4.6 Hareland Models (1994, 2010) ............................................................................ 17
2.4.7 Motahhari Model (2010) ..................................................................................... 18
2.4.8 Teale's Mechanical Specific Energy Model (1965) ............................................ 18
2.5 Machine Learning Approaches for ROP Prediction .................................................. 19
2.5.1 Artificial Neural Networks (ANN) ...................................................................... 19
2.5.2 Support Vector Machines (SVM) ........................................................................ 20
2.5.3 Random Forest .................................................................................................... 21
2.5.4 Extreme Gradient Boosting (XGBoost) .............................................................. 21
2.5.5 Deep Learning Approaches ................................................................................. 22

vi
 2.6 Optimization Approaches in Drilling ........................................................................ 25
2.6.1 Genetic Algorithms ............................................................................................. 25
2.6.2 Particle Swarm Optimization .............................................................................. 25
2.6.3 Grey Wolf Optimiser ........................................................................................... 26
2.7 Feature Analysis and Interpretability ......................................................................... 26
2.7.1 Feature Importance Methods ............................................................................... 26
2.7.2 SHAP (SHapley Additive exPlanations) ............................................................. 27
2.8 Research Gaps............................................................................................................ 27
2.9 Conceptual Framework .............................................................................................. 28
CHAPTER THREE: RESEARCH METHODOLOGY ............................................... 30
3.1 Introduction ................................................................................................................ 30
3.2 Research Design ........................................................................................................ 30
3.2.1 Research Philosophy ........................................................................................... 31
3.2.2 Research Approach .............................................................................................. 31
3.2.3 Research Strategy ................................................................................................ 31
3.2.4 Research Framework ........................................................................................... 32
3.3 Data Acquisition and Description .............................................................................. 33
3.3.1 Dataset Characteristics ........................................................................................ 33
3.3.2 Variable Categories ............................................................................................. 33
3.3.3 Target Variable .................................................................................................... 34
3.3.4 Controllable Parameters ...................................................................................... 34
3.3.5 Depth and Formation-Related Parameters........................................................... 35
3.3.6 Hydraulic and Mechanical Parameters ................................................................ 36
3.3.7 Derived Parameters ............................................................................................. 36
3.3.8 Descriptive Statistics ........................................................................................... 36
3.4 Data Preprocessing .................................................................................................... 37
3.4.1 Data Quality Assessment..................................................................................... 37
3.4.2 Feature Scaling .................................................................................................... 37
3.4.3 Train-Test Split.................................................................................................... 38
3.4.4 Feature Selection Considerations ........................................................................ 39
3.5 Feature Analysis Methods ......................................................................................... 39
3.5.1 Correlation Analysis ............................................................................................ 39
3.5.2 Random Forest Feature Importance .................................................................... 40
3.5.3 Permutation Importance ...................................................................................... 41
3.5.4 SHAP Analysis .................................................................................................... 41
3.6 Machine Learning Model Development .................................................................... 42
3.6.1 Random Forest Model ......................................................................................... 42
3.6.2 XGBoost Model .................................................................................................. 43

vii
 3.6.3 Artificial Neural Network Model ........................................................................ 44
3.7 Model Evaluation Metrics ......................................................................................... 45
3.7.1 Coefficient of Determination (R²) ....................................................................... 45
3.7.2 Mean Absolute Error (MAE) .............................................................................. 46
3.7.3 Root Mean Square Error (RMSE) ....................................................................... 46
3.7.4 Mean Absolute Percentage Error (MAPE) .......................................................... 46
3.8 Parameter Optimization Approach ............................................................................ 46
3.8.1 Optimization Problem Formulation ..................................................................... 47
3.8.2 Grid Search Implementation ............................................................................... 47
3.8.3 Multi-Model Optimization .................................................................................. 47
3.9 Validation Methods.................................................................................................... 48
3.9.1 Sensitivity Analysis ............................................................................................. 48
3.9.2 Drilling Mechanics Validation ............................................................................ 48
3.9.3 Cross-Model Validation ...................................................................................... 48
3.10 Software and Computational Environment.............................................................. 49
3.11 Ethical Considerations ............................................................................................. 49
CHAPTER FOUR: RESULTS AND DISCUSSION ..................................................... 50
4.1 Introduction ................................................................................................................ 50
4.2 Feature Analysis Results............................................................................................ 50
4.2.1 Correlation Analysis ............................................................................................ 50
4.2.2 Random Forest Feature Importance .................................................................... 52
4.2.3 SHAP Analysis .................................................................................................... 55
4.2.4 Summary of Feature Analysis ............................................................................. 57
4.3 Machine Learning Model Comparison ...................................................................... 57
4.3.1 Model Performance Metrics ................................................................................ 58
4.3.2 Analysis of Performance Differences .................................................................. 59
4.3.3 Comparison with Literature ................................................................................. 60
4.4 Optimization Results.................................................................................................. 61
4.4.1 Optimal Parameter Identification ........................................................................ 61
4.4.2 Significance of Model Convergence ................................................................... 62
4.4.3 Interpretation of Optimal Values ......................................................................... 62
4.5 Sensitivity Analysis and Validation ........................................................................... 63
4.5.1 Sensitivity to Weight on Bit ................................................................................ 63
4.5.2 Sensitivity to Rotary Speed ................................................................................. 64
4.5.3 Summary of Validation Findings ........................................................................ 66
4.6 Discussion .................................................................................................................. 66
4.6.1 Dominance of Formation Effects ........................................................................ 66
4.6.2 Algorithm Selection for Small Datasets .............................................................. 67

viii
 4.6.3 Robustness Through Multi-Model Convergence ................................................ 67
4.6.4 Physical Plausibility as Validation ...................................................................... 67
CHAPTER FIVE: CONCLUSIONS AND RECOMMENDATIONS ......................... 69
5.1 Introduction ................................................................................................................ 69
5.2 Summary of Key Findings ......................................................................................... 69
5.2.1 Dataset and Scope ............................................................................................... 69
5.2.2 Feature Analysis Findings ................................................................................... 70
5.2.3 Model Performance Findings .............................................................................. 70
5.2.4 Optimization Findings ......................................................................................... 71
5.2.5 Validation Findings ............................................................................................. 72
5.3 Conclusions................................................................................................................ 72
5.3.1 Conclusions on Research Objective 1: Feature Analysis .................................... 72
5.3.2 Conclusions on Research Objective 2: Model Comparison ................................ 73
5.3.3 Conclusions on Research Objective 3: Parameter Optimization ......................... 73
5.3.4 Conclusions on Research Objective 4: Validation .............................................. 74
5.4 Recommendations ...................................................................................................... 76
5.4.1 Recommendations for Drilling Operations ......................................................... 76
5.4.2 Recommendations for Machine Learning Practitioners ...................................... 76
5.4.3 Recommendations for Future Research............................................................... 77
5.5 Contributions of the Study ......................................................................................... 78
5.5.2 Practical Contributions ........................................................................................ 79
5.6 Limitations of the Study ............................................................................................ 79
5.6.1 Data Limitations .................................................................................................. 80
5.6.3 Scope Limitations ................................................................................................ 81
5.7 Future Research Directions ........................................................................................ 81
5.7.1 Extended Validation Studies ............................................................................... 81
5.7.2 Integration of Formation Characterisation .......................................................... 82
5.7.3 Multi-Objective Optimization ............................................................................. 82
5.7.4 Advanced Machine Learning Architectures ........................................................ 82
5.7.5 Real-Time Implementation.................................................................................. 82
5.7.6 Uncertainty Quantification .................................................................................. 82
5.7.7 Transfer Learning ................................................................................................ 82
5.8 Concluding Remarks.................................................................................................. 83
REFERENCES .................................................................................................................. 85
APPENDIX A .................................................................................................................... 93
PYTHON CODE FOR ROP PREDICTION AND OPTIMISATION ............................... 93
APPENDIX B .................................................................................................................... 97
MODEL RESULTS SUMMARY ...................................................................................... 97

ix
B.1 Model Performance Comparison .............................................................................. 97
B.2 Optimal Drilling Parameters ..................................................................................... 97
B.3 Feature Importance Rankings ................................................................................... 97
B.4 Correlation Analysis Results ..................................................................................... 98

x
 LIST OF TABLES
Table 2. 1: Summary of Field-Development AI-Based ROP Optimization Benefits.........25
Table 3. 1: Comprehensive Categories of Variables in the Dataset .................................... 34
Table 3. 2: Descriptive Statistics for Key Variables ........................................................... 37
Table 3. 3: Machine Learning Model Hyperparameter Configurations .............................. 43
Table 3. 4 : Artificial Neural Network Architecture ........................................................... 44
Table 3. 5: Model Evaluation Metrics Summary ................................................................ 45
Table 4. 1: Comparison of Feature Importance Rankings Between Random Forest and
XGBoost .............................................................................................................................53
Table 4. 2: Summary of Feature Analysis Results ............................................................. 57
Table 4. 3: Machine Learning Model Performance Comparison ....................................... 59
Table 4. 4: Optimal Drilling Parameters Identified by Each Model .................................. 61

xi
 LIST OF FIGURES
Figure 2. 1: Schematic representation of AI-based real-time ROP optimization workflow as
implemented in recent field deployments................................................................23
Figure 4. 1: Correlation Heatmap Including ROP and Key Drilling Parameters................ 51
Figure 4. 2: Random Forest Feature Importance Rankings ................................................ 54
Figure 4. 3: SHAP Beeswarm Plot Showing Feature Impact on Model Output ................. 56
Figure 4. 4: Machine learning model development and evaluation workflow illustrating the
training,validation, and comparison process for Random Forest, XGBoost, and ANN
models .......................................................................................................................... 58
Figure 4. 5: Sensitivity of Predicted ROP to Weight on Bit (Flow and RPM at optimal
values) .......................................................................................................................... 63
Figure 4. 6: Sensitivity of Predicted ROP to Rotary Speed (WOB and Flow at optimal
values) .......................................................................................................................... 65

xii
NOMENCLATURE
ANN Artificial Neural Network
BHA Bottom Hole Assembly
CCS Confined Compressive Strength
ECD Equivalent Circulating Density
GA Genetic Algorithm
GPM Gallons Per Minute
GWO Grey Wolf Optimiser
IADC International Association of Drilling Contractors
IoT Internet of Things
LSTM Long Short-Term Memory
MAE Mean Absolute Error
MAPE Mean Absolute Percentage Error
MD Measured Depth
MDI Mean Decrease in Impurity
MSE Mechanical Specific Energy
MWD Measurement While Drilling
PDC Polycrystalline Diamond Compact
PDM Positive Displacement Motor
PSO Particle Swarm Optimization
PV Plastic Viscosity
R² Coefficient of Determination
ReLU Rectified Linear Unit
RMSE Root Mean Square Error
RNN Recurrent Neural Network
ROP Rate of Penetration
RPM Rotary Speed (revolutions per minute)
SHAP SHapley Additive exPlanations
XGBoost Extreme Gradient Boosting
SVM Support Vector Machine
SVR Support Vector Regression
TFA Total Flow Area
TVD True Vertical Depth
UCS Unconfined Compressive Strength

xiii
 CHAPTER ONE: INTRODUCTION

1.1 Background to the Study

The global demand for oil and natural gas continues to grow despite the increasing
focus on renewable energy sources, with hydrocarbons remaining the primary energy source
for transportation, industrial processes, and power generation worldwide (Speight, 2020).
According to the International Energy Agency (2023), oil and gas still account for over 55
percent of global primary energy consumption, a share expected to remain significant
through at least 2040, even under ambitious climate scenarios. This persistent demand,
coupled with the depletion of easily accessible conventional reservoirs, has driven
exploration and production activities into increasingly challenging environments, including
deepwater offshore locations, unconventional formations, and geologically complex
reservoirs (Hossain and Al-Majed, 2015). These challenging environments have
substantially increased drilling costs, making operational efficiency more critical than ever
for economic viability.
Drilling operations represent one of the most capital-intensive aspects of oil and gas
exploration and production, typically accounting for 30-70 percent of total capital
expenditure required to develop a hydrocarbon field (Kaiser, 2015; Khodja et al., 2010).
The economic significance of drilling efficiency has prompted decades of research aimed at
understanding and optimizing the factors that influence drilling performance (Graham and
Muench, 1959; Galle and Woods, 1963). Since the advent of unconventional plays, drilling
and completion costs have continued to escalate, with average well costs now exceeding $6
million in many shale formations (U.S. Energy Information Administration, 2016). In the
context of volatile oil prices and increasing operational complexity, the industry has
intensified efforts to reduce drilling time and costs whilst maintaining safety and
environmental standards (Slaughter et al., 2015).
The rate of penetration (ROP) is a fundamental performance metric in drilling
operations, defined as the speed at which the drill bit advances through the formation,
typically measured in metres per hour (m/hr) or feet per hour (ft/hr). ROP has a direct and
significant impact on drilling costs, as time-dependent costs, including rig rental, personnel,
fuel, and support services, typically represent 60 to 80 percent of total drilling expenditure
in conventional operations (Bourgoyne and Young, 1974; Hossain and Al-Majed, 2015).
Beyond cost implications, ROP serves as an important indicator of drilling conditions and
potential problems, with sudden changes signalling transitions between formations, drilling
dysfunctions such as bit balling, or downhole equipment issues (Abbas et al., 2018; Akgun,
1
2002). The drilling rate is the key business performance indicator in the drilling industry
and a measure of additional contractual incentives (Sridharan and Kumar, 2020).
The economic significance of ROP optimization has increased dramatically in recent
years due to several converging industry trends. The depletion of easily accessible
conventional reservoirs has pushed drilling operations into more challenging environments
where drilling costs are substantially higher and where even small improvements in drilling
efficiency can yield significant economic benefits (Zhang et al., 2021). At the same time,
advances in drilling technology, including polycrystalline diamond compact (PDC) bits,
measurement-while-drilling (MWD) systems, and automated drilling rigs, have generated
vast quantities of real-time operational data that can potentially be leveraged to improve
drilling performance (Bellin et al., 2010). PDC bits, in particular, have recently entered
widespread use, replacing conventional roller-cone bits and accounting for more than 90
percent of total drilling footage worldwide due to their high rock-breaking efficiency (Chen
et al., 2023; Rashidi et al., 2010).
The complexity of ROP optimization arises from numerous interacting factors that
influence drilling performance, often in non-linear, formation-dependent ways and are
difficult to predict using conventional analytical methods (Perrin et al., 1997; Bataee and
Mohseni, 2011). These factors include controllable drilling parameters such as weight on
bit (WOB), rotary speed (RPM), and flow rate, as well as uncontrollable formation
properties including rock strength, porosity, pore pressure, and mineralogical composition
(Dupriest and Koederitz, 2005). The effects of these parameters on ROP are not
independent, as interactions between drilling parameters and formation properties can
produce non-linear responses that are difficult to capture with traditional physics-based
models (Elkatatny et al., 2017). This complexity has driven a fundamental shift from first-
principles modelling to data-driven approaches, which offer a simpler way to extract value
from operational data through intelligent algorithms (Sridharan and Kumar, 2020).
Traditional approaches to ROP modelling have relied on empirical equations derived
from regression analysis of drilling data, aiming to establish mathematical relationships
between observed drilling parameters and the penetration rate (Maurer, 1962; Bingham,
1965). The most comprehensive and widely applied of these empirical models is the
Bourgoyne and Young (1974) model, which expresses ROP as a product of eight
exponential functions representing the effects of formation strength, compaction,
differential pressure, weight on bit, rotary speed, bit wear, and hydraulics. This model
revolutionized ROP optimization by consolidating previously separate approaches for

2
optimal bit weight and rotary speed, jet bit hydraulics, and abnormal pressure detection into
a unified framework (Barros, 2015). Other notable empirical models include Warren's
(1984) imperfect cleaning model, which explicitly incorporates the effects of inadequate
bottom-hole cleaning, and Hareland's (1994) drag bit model, which accounts for bit-specific
performance characteristics.
While these empirical models have provided valuable insights into drilling
performance and continue to be used in many industry applications, they suffer from several
fundamental limitations that restrict their practical utility. First, they require extensive
calibration data from similar formations to determine formation-specific coefficients, and
the transferability of these coefficients to new drilling environments is often poor
(Eskandarian et al., 2017; Bahari et al., 2007). Traditional empirical models involve a
number of constants that must be evaluated based on field results or experimental methods,
resulting in time-consuming evaluations and an inability to generalize to the entire well,
leading to inaccurate predictions (Sridharan and Kumar, 2020). Second, they assume
specific functional forms for parameter relationships that may not accurately represent the
true underlying physics across all operating conditions (Hegde et al., 2017). Third, they are
typically calibrated offline and cannot adapt in real time to changing formation conditions
encountered during drilling (Soares and Gray, 2019).
The emergence of machine learning and artificial intelligence techniques has opened
new possibilities for ROP prediction and optimization, addressing many of the limitations
of traditional empirical models. Machine learning approaches can learn complex, non-linear
relationships directly from data without requiring explicit specification of functional forms,
enabling them to capture patterns that may be missed by physics-based models (Bilgesu et
al., 1997; Goodfellow et al., 2016). With advances in big data technologies, real-time data
storage costs have decreased, and the availability of real-time data has increased, facilitating
the application of sophisticated machine learning techniques to drilling optimization
(Sridharan and Kumar, 2020). These data-driven approaches are particularly well-suited to
the drilling environment, where large volumes of sensor data are routinely collected, but the
underlying physical relationships are complex and incompletely understood (Barbosa et al.,
2019).
Recent studies have reported prediction accuracies exceeding 90 percent using
machine learning approaches, significantly outperforming traditional empirical models in
many applications. Mohammadinia et al. (2025) achieved R-squared values exceeding 92%
using least-squares support vector machines combined with metaheuristic optimization

3
algorithms. Ma et al. (2025) developed an interpretable automated machine learning
workflow, achieving an R-squared of 0.92 for ROP prediction. Osman et al. (2021) found
that Random Forest consistently outperformed other algorithms when applied to carbonate
reservoir drilling data. These results have been corroborated by numerous studies
demonstrating the effectiveness of various machine learning approaches, including artificial
neural networks (Bilgesu et al., 1997; Mantha and Samuel, 2016), support vector machines
(Bodaghi et al., 2015), Random Forest (Hegde and Gray, 2017), XGBoost (Chen and
Guestrin, 2016), and deep learning architectures (Zhang et al., 2024).
The digital transformation currently underway in the oil and gas industry has created
unprecedented opportunities for advanced analytics applications in drilling operations.
Industry 4.0 technologies, including cloud computing, Internet of Things (IoT) sensors, and
big data analytics, are being integrated across business functions to streamline operations
and improve decision-making (Schwab, 2016; Prestidge, 2022). According to McKinsey,
digitizing oil and gas upstream operations could create $ 250 billion in value by 2030
(Ericsson, 2023). Modern offshore drilling platforms now have approximately 80,000
sensors that can generate up to 15 petabytes of data over the asset's lifetime, providing rich
datasets for machine learning applications (IHS Markit, 2020). This abundance of data,
combined with increased computational resources, has enabled the development and
deployment of sophisticated predictive models that would have been infeasible just a decade
ago.
Despite the promising results achieved with machine learning approaches, several
important research gaps remain that this study aims to address.
First, the practical implementation of machine learning models for drilling
optimization in settings with limited data availability warrants further investigation, as many
published studies use large datasets that may not be available in all drilling applications
(Najjarpour et al., 2022).
Second, while numerous studies have compared individual algorithms, systematic
comparisons of multiple machine learning approaches using identical datasets,
preprocessing, and evaluation criteria remain relatively rare, making it difficult to draw
general conclusions about algorithm selection (Khalilidermani and Knez, 2023).
Third, translating machine learning predictions into practical drilling
recommendations requires careful validation to ensure physical plausibility, an aspect often
overlooked in purely data-driven studies (Hegde et al., 2017). This study addresses these
gaps through a systematic comparative analysis of machine learning algorithms for ROP

4
prediction and optimization with validation against established drilling mechanics
principles.
The practical value of machine learning approaches for ROP optimization has been
increasingly validated through field implementations. Recent deployments of AI-based
drilling optimization systems have demonstrated substantial operational benefits, including
ROP improvements of 19-43 percent and significant reductions in overall drilling time (Al-
Riyami et al., 2023; Singh et al., 2021; Boukredera et al., 2023). In one notable field trial,
adherence to machine learning recommendations enabled a West African well to reach total
depth approximately 24 days ahead of the original drilling plan (Al-Riyami et al., 2023).
Beyond direct ROP improvements, these systems have demonstrated additional benefits,
including reduced vibration and fewer bottom-hole assembly failures, underscoring the
multi-dimensional value proposition of data-driven drilling optimization. These
documented successes provide compelling motivation for continued research into machine
learning approaches for ROP prediction and optimization, whilst highlighting the potential
for translating academic research into operational improvements.

1.2 Problem Statement

Optimizing drilling parameters to maximize the rate of penetration remains a significant
challenge in the oil and gas industry despite decades of research and technological
advancements (Khalilidermani and Knez, 2023). The challenge is multifaceted, involving
complex interactions between controllable drilling parameters and continuously varying
formation properties along the wellbore, making real-time optimization particularly difficult
(Soares and Gray, 2019). These variables do not influence ROP independently but interact
in ways that are often non-linear, formation-dependent, and difficult to predict using
conventional analytical methods (Perrin et al., 1997). The drilling rate depends on many
drilling variables and their complex relationships, including operational parameters,
formation properties, compressive strength, well hydraulics, borehole shape and size, mud
properties, bit type, and hole cleaning efficiency (Sridharan and Kumar, 2020).

Traditional empirical models, while providing a theoretical framework for understanding

the relationships between drilling parameters and ROP, suffer from several fundamental
limitations that restrict their practical utility. The Bourgoyne and Young (1974) model, for
example, requires determination of eight model coefficients plus threshold values that must
be calibrated for specific formations, a process that can be time-consuming and may yield

5
coefficients with limited transferability to new drilling environments (Bahari et al., 2007).
The abundance of constants that need to be evaluated based on field or experimental results
increases time consumption and prevents generalization across the entire well, leading to
inaccurate predictions (Eskandarian et al., 2017). Furthermore, these models assume
specific functional forms that may not accurately represent the underlying physics across all
operating conditions, and they are typically calibrated offline, limiting their ability to adapt
in real time to changing downhole conditions (Hegde et al., 2017).

The mathematical complexity of traditional models also presents practical challenges. The
Bourgoyne and Young model, for example, requires the determination of eight model
coefficients and threshold values, which must be calibrated for specific formations using
multiple regression analysis (Bourgoyne and Young, 1974). This calibration process
requires substantial amounts of data from similar formations and geological conditions,
which may not always be available, particularly for exploration wells in new areas (Barros,
2015). The models are also sensitive to data quality and the specific regression techniques
used for coefficient estimation, with different studies reporting significantly different
coefficient values for similar formations (Hegde et al., 2017). These limitations have
motivated the exploration of data-driven approaches that can learn patterns directly from
operational data without requiring explicit specification of functional relationships.
Machine learning approaches offer the potential to overcome many of these limitations by
learning patterns directly from data without requiring explicit specification of functional
forms.
However, several gaps remain in current research.
First, while numerous studies have demonstrated the effectiveness of machine
learning for ROP prediction using large datasets, fewer have systematically evaluated
algorithm performance under the limited-data conditions common in many drilling
applications (Najjarpour et al., 2022). Small to medium-sized datasets are particularly
common when drilling in new formations, using new bit types, or operating in regions with
limited historical data (Geron, 2019).
Second, although many individual algorithms have been evaluated, comprehensive
comparisons of multiple approaches using identical datasets, preprocessing steps, and
evaluation criteria remain relatively scarce, making it difficult to draw general conclusions
about optimal algorithm selection for different scenarios (Khalilidermani and Knez, 2023).

6
 Third, translating machine learning predictions into practical drilling
recommendations requires careful validation to ensure physical plausibility. Many
published studies focus on prediction accuracy metrics such as R-squared and RMSE
without systematically examining whether the learned relationships are consistent with
established drilling mechanics principles (Hegde et al., 2017). This validation gap is
particularly concerning because machine learning models can achieve high predictive
accuracy by learning spurious correlations in training data that do not generalise to new
conditions or that produce physically implausible recommendations when used for
optimization (Hastie et al., 2009). The integration of model interpretability techniques, such
as SHAP (SHapley Additive exPlanations) analysis, provides a means to bridge this gap by
revealing how models make predictions and enabling verification against theoretical
expectations (Lundberg and Lee, 2017).

1.3 Research Objectives

The primary objective of this study is to develop and validate machine learning models to
predict and optimize the rate of penetration in drilling operations. The specific objectives
are:
1. To identify and analyze key drilling parameters influencing the rate of penetration
using correlation analysis, feature importance methods, and SHAP (SHapley
Additive exPlanations), providing both quantitative rankings and interpretable
insights into parameter relationships.
2. To develop and compare multiple machine learning models, including Random
Forest, XGBoost, and Artificial Neural Networks for accurate ROP prediction,
systematically evaluating their performance using metrics including R-squared,
MAE, RMSE, and MAPE.
3. To determine optimal drilling parameters for maximizing the rate of penetration
using the best-performing machine learning model, with particular attention to the
robustness of optimization results across different modelling approaches.
4. To validate the optimization recommendations using statistical sensitivity analysis
and drilling mechanics principles, ensuring that model behaviour is physically
plausible and consistent with established theory.

7
1.4 Significance of the Study
This study contributes to both the academic literature and industry practice in several
important ways. From a theoretical perspective, the research provides a systematic
comparative analysis of different machine learning algorithms specifically evaluated under
conditions of limited data availability, addressing a gap in the current literature where most
studies utilize large datasets that may not be representative of many practical applications
(Najjarpour et al., 2022; Khalilidermani and Knez, 2023). The multi-method feature
analysis, combining correlation analysis, Random Forest importance, and SHAP values,
provides a comprehensive framework for understanding the factors driving ROP and can be
applied to other drilling datasets (Zou et al., 2025).
From a practical standpoint, the study develops a systematic methodology for translating
machine learning predictions into actionable drilling recommendations. The validation
framework using sensitivity analysis and comparison with drilling mechanics principles
(Maurer, 1962; Bourgoyne and Young, 1974; Warren, 1984) provides a template for
ensuring that data-driven recommendations are physically plausible before implementation.
This bridging of data-driven and physics-based approaches addresses a critical industry need
for validated, interpretable models that can gain acceptance among drilling engineers and
operators (Ma et al., 2025). The methodology demonstrates that effective ROP optimization
is achievable even with limited data, which has significant implications for drilling
operations where comprehensive datasets may not be available.
The research also contributes to the broader adoption of advanced analytics in the drilling
industry, supporting the ongoing digital transformation that is reshaping oil and gas
operations globally (Schwab, 2016; Prestidge, 2022). The interpretability analysis using
SHAP values helps bridge the gap between black-box machine learning models and drilling
engineers' domain expertise, facilitating trust in and adoption of data-driven approaches
(Lundberg and Lee, 2017; Molnar, 2022). As the oil and gas industry continues to digitize,
with digitizing operations potentially creating 250 billion US dollars in value for upstream
operations by 2030 (McKinsey, 2020), studies demonstrating practical applications of
machine learning in drilling optimization provide valuable templates for broader industry
adoption.

1.5 Scope and Limitations

This study focuses on predicting and optimizing the rate of penetration using drilling data
from a single drilling section with a constant bit type (polycrystalline diamond compact bit)

8
and size (8.5 inches). The analysis concentrates on steady-state drilling conditions and does
not address transient phenomena, drilling dysfunctions such as stick-slip or bit bounce, or
abnormal events such as formation kicks or lost circulation. The optimization considers
three primary controllable parameters (weight-on-bit, rotary speed, and flow rate), while
other potentially relevant parameters, such as mud properties and bit design features, are
treated as constants within the scope of this study.
Several limitations should be noted. First, the dataset does not include explicit formation
parameters such as rock strength, porosity, or permeability. Instead, depth-related
parameters serve as proxies for formation characteristics, an approach validated in previous
studies (Hegde et al., 2017; Osman et al., 2021) but that may limit model transferability to
formations with different depth-property relationships. Second, the relatively small sample
size of 43 observations, while representative of many practical drilling applications,
constrains the complexity of patterns that can be reliably learned and introduces uncertainty
into performance estimates (Geron, 2019). Third, the study evaluates three machine learning
algorithms (Random Forest, XGBoost, and Artificial Neural Network) selected for their
established effectiveness in similar applications, but other algorithms, such as support vector
machines, Gaussian processes, or advanced deep learning architectures, are not evaluated.

1.6 Organization of the Thesis

This thesis is organized into five chapters. Chapter One has introduced the research
background, problem statement, objectives, significance, scope, and limitations. Chapter
Two presents a comprehensive literature review covering the theoretical foundations of
ROP modelling, traditional empirical approaches, and recent advances in machine learning
applications for drilling optimization. Chapter Three describes the research methodology,
including data description, preprocessing procedures, machine learning model
development, and validation approaches. Chapter Four presents the results of feature
analysis, model comparison, optimization, and validation, along with a discussion of
findings in the context of existing literature. Chapter Five concludes the thesis with a
summary of key findings, conclusions for each research objective, recommendations for
practice and future research, and the study's contributions.

9
 CHAPTER TWO: LITERATURE REVIEW

2.1 Introduction
This chapter provides a comprehensive review of the literature pertaining to rate of
penetration (ROP) modelling and optimization in drilling operations. The review
encompasses the theoretical foundations of ROP, the various factors that influence drilling
performance, and the evolution of predictive models from traditional empirical approaches
to contemporary machine learning techniques. Understanding these concepts is fundamental
to developing effective optimization strategies that can enhance drilling efficiency whilst
minimising operational costs.

The literature review is structured to first examine the concept and significance of ROP in
drilling operations, followed by a detailed analysis of the controllable and uncontrollable
factors that affect penetration rates. Subsequently, the chapter explores the chronological
development of ROP models, beginning with early empirical formulations and progressing
to sophisticated machine learning approaches. The review also examines optimization
techniques, feature analysis methods, and identifies existing research gaps that this study
aims to address.

The evolution of ROP modelling represents a significant shift from first-principles

approaches to data-driven methodologies (Barbosa et al., 2019). Early models relied heavily
on laboratory experiments and simplified assumptions about rock-bit interactions, whilst
modern approaches leverage the abundance of real-time drilling data and advanced
computational techniques (Najjarpour et al., 2022). This transition reflects the broader
transformation in petroleum engineering towards intelligent drilling systems that can adapt
to changing downhole conditions in real-time (Soares and Grey, 2019; Khalilidermani and
Knez, 2023).

2.2 Rate of Penetration: Concept and Importance

Rate of penetration (ROP) is defined as the speed at which the drill bit penetrates the rock
formation, typically measured in feet per hour (ft/hr) or metres per hour (m/hr).
Mathematically, ROP is expressed as the derivative of depth with respect to time:
ROP = dD/dt [2.1]

10
where D represents the measured depth, and t represents time (Bourgoyne et al., 1986). This
seemingly simple metric serves as the primary indicator of drilling efficiency and directly
impacts the economic viability of drilling operations (Kaiser, 2015; U.S. EIA, 2016).
The significance of ROP extends beyond mere speed measurements. According to Zakariya
et al. (2015), drilling inefficiency caused by suboptimal ROP and under-optimised drilling
processes accounts for approximately 15% of total well costs. This substantial cost
implication has driven extensive research into ROP prediction and optimization
methodologies (Sridharan and Kumar, 2020). The drilling cost per foot, which is a function
of ROP, can be calculated using the following relationship:
Cf = (Cb + Cr × (tb + tc + tt)) / F [2.2]
where Cb is the bit cost, Cr is the rig operating cost per hour, tb is the rotating time on
bottom, tc is the connection time, tt is the trip time, and F is the footage drilled (Bourgoyne
et al., 1986). This equation demonstrates that maximising ROP directly reduces the rotating
time component, thereby minimising overall drilling costs.
From an operational perspective, ROP serves multiple functions beyond cost optimization.
It provides real-time feedback on formation changes, helps detect abnormal pressure zones,
indicates bit wear conditions, and assists in evaluating the effectiveness of drilling
parameters (Khodja et al., 2010; Dupriest and Koederitz, 2005). Furthermore, ROP data
contribute to the generation of mechanical specific energy (MSE) logs, which are valuable
tools for drilling performance optimization (Perrin et al., 1997; Rashidi et al., 2010; IADC,
2014).

2.3 Factors Influencing Rate of Penetration

The rate of penetration is influenced by a complex interplay of various factors that can be
broadly categorised into controllable and uncontrollable parameters. Understanding these
factors and their interactions is essential for developing effective ROP prediction models
and optimization strategies. According to Bourgoyne and Young (1974), ROP can be
expressed as a function of these parameters:

ROP = f (Controllable Parameters, Uncontrollable Parameters) [2.3]

2.3.1 Controllable Factors

Controllable factors are parameters that can be adjusted by the drilling engineer or driller
during operations to optimise performance. These include weight on bit (WOB), rotary

11
speed (RPM), flow rate, and bit hydraulics. The ability to modify these parameters in real
time enables continuous optimization of drilling performance.
[Link] Weight on Bit (WOB)
Weight on bit represents the axial force applied to the drill bit through the drill string. It is
one of the most significant controllable parameters affecting ROP. The relationship between
WOB and ROP was first systematically studied by Galle and Woods (1963) and Graham
and Muench (1959), who established that ROP generally increases with increasing WOB
up to a certain threshold, known as the founder point.
The founder point represents the maximum WOB beyond which further increases do not
result in proportional ROP improvements due to inadequate hole cleaning (IADC, 2014;
Dupriest and Koederitz, 2005). This phenomenon occurs because excessive WOB can cause
bit balling, wherein cuttings accumulate on the bit face, reducing cutting efficiency. The
relationship between WOB and ROP, as described by Maurer (1962), follows a power-law
relationship under ideal cleaning conditions:
ROP ∝ WOB2 [2.4]
Irawan et al. (2012) demonstrated that optimising WOB based on real-time ROP responses
can significantly improve drilling efficiency. Their research showed that the optimal WOB
varies with formation properties and bit type, necessitating adaptive control strategies.
Bahari et al. (2007) further established practical WOB limits for different formation types
based on extensive field data analysis.

[Link] Rotary Speed (RPM)

Rotary speed, measured in revolutions per minute (RPM), determines the frequency at
which the bit cutters engage with the rock formation. Under ideal conditions, ROP exhibits
a linear relationship with RPM (Maurer, 1962):

ROP ∝ RPM [2.5]

However, this linear relationship is constrained by several practical limitations. Excessive

RPM can induce harmful vibrations, including bit whirl and stick-slip phenomena, which
can damage the bit and bottomhole assembly (BHA) whilst reducing drilling efficiency
(Bellin et al., 2010). Furthermore, high rotary speeds may result in inadequate hole cleaning,
particularly in deep wells where annular velocities are reduced (Glowka, 1987).

Warren and Armagost (1988) demonstrated that the relationship between RPM and ROP
becomes non-linear under imperfect cleaning conditions, with the rate of ROP increase
12
diminishing at higher rotary speeds. This behaviour necessitates careful optimization of
RPM in conjunction with other parameters to achieve maximum drilling efficiency.

[Link] Flow Rate and Hydraulics

Flow rate determines the volume of drilling fluid circulated through the wellbore per unit
time, typically measured in gallons per minute (GPM) or litres per minute (LPM). Adequate
flow rate is essential for effective hole cleaning, bit cooling, and cuttings transport (Abbas
et al., 2018). The hydraulic impact force at the bit, which affects ROP, is calculated as:
Fj = (Q × Vn × ρ) / 1930 [2.6]
where Fj is the jet impact force (lbf), Q is the flow rate (GPM), Vn is the nozzle velocity
(ft/s), and ρ is the mud density (lb/gal). The nozzle velocity is calculated as:
Vn = (0.321 × Q) / An [2.7]
where An is the total nozzle area (in²). Bourgoyne et al. (1986) demonstrated that increasing
hydraulic impact force generally improves ROP by enhancing bottom-hole cleaning
efficiency.

2.3.2 Uncontrollable Factors

Uncontrollable factors are inherent formation and wellbore characteristics that cannot be
modified during drilling operations. These factors significantly influence ROP and must be
accounted for in predictive models.

[Link] Formation Properties

Formation properties represent the most significant uncontrollable factors affecting ROP.
The compressive strength of the rock formation directly determines the energy required to
fracture and remove material. The relationship between ROP and rock compressive strength
(σc) was established by Maurer (1962) as an inverse square relationship under perfect
cleaning conditions:

ROP ∝ 1/σc2 [2.8]

Zhang et al. (2021) emphasised that formation heterogeneity introduces significant

variability in ROP, as rock properties can change dramatically over short intervals.
Additionally, formation abrasiveness affects bit wear rates, which in turn influences
sustained ROP performance. The confined compressive strength (CCS), which accounts for
the in-situ stress conditions, is given by:

CCS = UCS + m × σ3 [2.9]

13
where UCS is the unconfined compressive strength, m is a material constant, and σ3 is the
confining pressure.

[Link] Pore Pressure and Differential Pressure

The differential pressure between the mud column pressure and formation pore pressure
significantly affects ROP. Bourgoyne and Young (1974) established that increasing
overbalance (positive differential pressure) reduces ROP due to the chip hold-down effect,
where cuttings are pressed against the formation surface rather than being efficiently
removed. This relationship is expressed as:

f4 = exp[a4 × D × (gp − ρc)] [2.10]

where a4 is a model coefficient, D is depth, gp is the pore pressure gradient (ppg), and ρc is
the equivalent circulating density (ppg). This effect becomes more pronounced in deeper
wells where pressure differentials are greater.

[Link] Depth Effects

Depth affects ROP through multiple mechanisms, including increased formation
compaction, higher temperatures, and greater pressure differentials. Barros (2015)
demonstrated that formations become progressively harder with depth due to compaction
effects, resulting in reduced ROP for a given set of drilling parameters. The depth effect is
typically modelled as:

f2 = exp[a2 × (10000 − D)] [2.11]

where a2 is a coefficient representing the compaction effect and D is the depth in feet. This
exponential relationship captures the progressive reduction in drillability with increasing
depth.

2.4 Traditional Empirical ROP Models

The development of ROP models has evolved significantly since the mid-twentieth century,
progressing from simple empirical correlations to comprehensive multi-parameter models.
This section reviews the major traditional models that form the foundation for modern ROP
prediction methodologies.

2.4.1 Maurer Model (1962)

Maurer (1962) developed the first theoretical ROP model based on rock cratering
mechanisms observed in single-insert impact experiments. The model, known as the

14
"perfect cleaning" theory, assumes that all rock debris is removed between successive tooth
impacts. Under these idealised conditions, Maurer derived the following relationship:

ROP = K × (RPM × WOB2) / (Db2 × σc2) [2.12]

where K is a drillability constant dependent on formation and bit characteristics, RPM is the
rotary speed, WOB is the weight on bit, Db is the bit diameter, and σc is the rock
compressive strength. This model establishes that ROP is directly proportional to rotary
speed and the square of WOB, whilst being inversely proportional to the square of both bit
diameter and rock strength.

The key contributions of Maurer's work include the identification of the crater volume
relationship and the concept of the founder point, where ROP no longer increases with WOB
due to inadequate hole cleaning.

2.4.2 Bingham Model (1965)

Bingham (1965) proposed a simplified empirical model that introduced adjustable
exponents to account for variations in formation and drilling conditions. The Bingham
model is expressed as:

ROP = K × RPMe × (WOB/Db)d [2.13]

where K is a formation-dependent constant, e is the rotary speed exponent, d is the weight

on bit exponent, and Db is the bit diameter. The flexibility provided by the adjustable
exponents allows the model to be calibrated to specific drilling conditions, making it more
versatile than Maurer's theoretical model.

Despite its simplicity, the Bingham model remains widely used as a starting point for ROP
analysis. Barros (2015) demonstrated that this model provides a reasonable approximation
for initial planning purposes, although its accuracy is limited in complex geological
environments.

2.4.3 Bourgoyne and Young Model (1974)

The Bourgoyne and Young (1974) model represents the most comprehensive empirical ROP
model widely adopted in the petroleum industry. The model accounts for eight distinct
factors affecting ROP through a multiplicative relationship:

ROP = f1 × f2 × f3 × f4 × f5 × f6 × f7 × f8 [2.14]

Each function represents a specific effect on ROP:

15
Formation Strength Effect (f1): This function accounts for the inherent drillability of the
formation:

f1 = exp(2.303 × a1) [2.15]

Depth Effect (f2): This function captures the normal compaction trend:

f2 = exp[2.303 × a2 × (10000 − D)] [2.16]

Pore Pressure Effect (f3): This function accounts for undercompaction due to abnormal
pore pressure:

f3 = exp[2.303 × a3 × D0.69 × (gp − 9)] [2.17]

Differential Pressure Effect (f4): This function models the chip hold-down effect:

f4 = exp[2.303 × a4 × D × (gp − ρc)] [2.18]

Bit Weight and Diameter Effect (f5): This function accounts for the normalised weight on
bit:

f5 = [(W/db − (W/db)t) / (4 − (W/db)t)]a5 [2.19]

where (W/db)t is the threshold bit weight per inch of bit diameter at which the bit begins to
drill.

Rotary Speed Effect (f6):

f6 = (N/60)a6 [2.20]

Bit Wear Effect (f7):

f7 = exp(−a7 × h) [2.21]

where h is the fractional tooth wear (0 for new bit, 1 for completely worn).

Bit Hydraulics Effect (f8):

f8 = (Fj/1000)a8 [2.22]

where Fj is the jet impact force calculated from Equation 2.6.

The model coefficients (a1 through a8) are determined through multiple regression analysis
using drilling data. Sridharan and Kumar (2020) demonstrated that the Bourgoyne and
Young model provides a robust framework for ROP prediction when properly calibrated,

16
although its performance diminishes in heterogeneous formations where parameters change
rapidly.

2.4.4 Warren Model (1987)

Warren (1987) developed an ROP model specifically for roller cone bits that accounts for
both chip generation and chip removal processes. The model recognises that ROP is limited
by either the rock-cutting mechanism or hole-cleaning efficiency, depending on which
process is rate-limiting. The Warren model is expressed as:

ROP = (a × σ × D × N) / [(σ × D)2/NW + aσDε/W + cρµε/Im] [2.23]

where σ is the rock compressive strength, D is the bit diameter, N is the rotary speed, W is
the weight on bit, ε is the rock ductility, ρ is the equivalent mud density, µ is the mud
viscosity, Im is the modified jet impact force, and a, c are model coefficients.

The modified jet impact force accounts for the bit hydraulics and is calculated as:

Im = [1 − Av−0.122] × Fj [2.24]

where Av is the ratio of jet velocity to return fluid velocity.

2.4.5 Winters, Warren and Onyia Model (1987)

Winters, Warren and Onyia (1987) extended the Warren model to include rock ductility and
cone offset effects for roller cone bits. Their model relates penetration rates to bit design,
operating conditions, and rock mechanics:

1/ROP = (σD2/NW) + (aσDε/W) + (Φ/ND) + (bρµε/Im) [2.25]

where Φ is the cone offset coefficient, and a, b are model coefficients. The model identifies
rock ductility as a major influence on bit performance and recognises cone offset as an
important design feature for drilling ductile rock. Barros (2015) recommended this model
for roller cone bit applications due to its comprehensive treatment of bit-rock interaction
mechanisms.

2.4.6 Hareland Models (1994, 2010)

Hareland and Rampersad (1994) developed an ROP model specifically for drag bits (PDC
bits) that accounts for cutter geometry, rock strength, and bit wear. The Hareland drag bit
model is expressed as:

ROP = (14.14 × N × RPM / D) × Gf × COR [2.26]

17
where N is the number of cutters, D is the bit diameter, Gf is a geometric function, and COR
is a correlation factor that accounts for bit cleaning, cutter geometry imperfections, and
microscopic rock strength variations:

COR = a / (RPMb × Wc) [2.27]

Hareland et al. (2010) later developed a roller cone bit model that relates bit-rock interaction
to rock failure by a wedge mechanism:

ROP = K × (80 × n × m × RPMa) / (D2 × tan2ψ) × (W / (100 × n × CCS))b × Wf [2.28]

where K is a comprehensive coefficient, n is the number of inserts in contact with rock, m

is the number of insert penetrations per revolution, ψ is the chip formation angle, CCS is the
confined compressive strength, Wf is the bit wear function, and a, b are model coefficients.
Soares and Gray (2019) validated this model against extensive field data and found good
agreement for homogeneous formations.

2.4.7 Motahhari Model (2010)

Motahhari et al. (2010) proposed an ROP model specifically for PDC bits and positive
displacement motors (PDMs), making it particularly useful for directional and horizontal
drilling applications. The model is expressed as:

ROP = G × (Wα × RPMγ) / (D × CCS) × Wf [2.29]

where G is a coefficient determined by bit geometry (cutter size, back rake, side rake angles,
and cutter-rock friction coefficient), α and γ are model exponents, and Wf is the bit wear
function. Barros (2015) recommended this model for PDC bit applications due to its ability
to capture motor effects and its physical interpretability.

2.4.8 Teale's Mechanical Specific Energy Model (1965)

Teale (1965) introduced the concept of mechanical specific energy (MSE), defined as the
work done per unit volume of rock excavated. This concept provides a framework for
evaluating drilling efficiency by comparing the energy input to the rock removal achieved.
The MSE is expressed as:

MSE = (W/Ab) + (2π × N × T) / (Ab × ROP) [2.30]

where W is the weight on bit, Ab is the bit area, N is the rotary speed, T is the torque, and
ROP is the rate of penetration. The first term represents the thrust energy component, whilst
the second term represents the rotational energy component.

18
In oilfield units, the MSE equation is commonly expressed as (Dupriest and Koederitz,
2005):

MSE = (480 × T × RPM) / (D2 × ROP) + (4 × WOB) / (π × D2) [2.31]

where T is the torque (ft-lbf), RPM is the rotary speed, D is the bit diameter (inches), ROP
is the rate of penetration (ft/hr), and WOB is the weight on bit (lbf). When drilling
efficiently, MSE approaches the confined compressive strength of the formation, providing
a benchmark for optimization (Pessier and Fear, 1992).

2.5 Machine Learning Approaches for ROP Prediction

The limitations of traditional empirical models in handling complex, non-linear
relationships and heterogeneous formations have driven the adoption of machine learning
techniques for ROP prediction. Machine learning models offer several advantages,
including the ability to capture complex patterns in data, adapt to changing conditions, and
improve predictions as more data becomes available. This section reviews the major
machine learning approaches applied to ROP prediction.

2.5.1 Artificial Neural Networks (ANN)

Artificial neural networks are among the earliest and most widely used machine learning
techniques for ROP prediction. ANNs are inspired by biological neural systems and consist
of interconnected nodes (neurons) organised in layers. The typical ANN architecture for
ROP prediction comprises an input layer that receives drilling parameters, one or more
hidden layers for feature transformation, and an output layer that produces the ROP
prediction.

The mathematical formulation of a neuron's output in a feed-forward neural network is:

y = f (Σ (wi × xi) + b) [2.32]

where y is the neuron output, f is the activation function, wi are the weights, xi are the inputs,
and b is the bias term. Common activation functions include the sigmoid, hyperbolic tangent
(tanh), and rectified linear unit (ReLU):

Sigmoid: f(x) = 1 / (1 + e−x) [2.33]

Tanh: f(x) = (ex − e−x) / (ex + e−x) [2.34]

ReLU: f(x) = max (0, x) [2.35]

19
Bilgesu et al. (1997) were among the pioneers in applying ANNs to ROP prediction,
demonstrating that neural networks could capture the complex relationships between
drilling parameters and ROP more effectively than traditional regression models.
Subsequent studies by Anemangely et al. (2018) and Elkatatny (2018) have confirmed the
superior performance of ANNs, particularly for horizontal drilling applications in complex
formations.

Sridharan and Kumar (2019, 2020) developed a back-propagation neural network for ROP
prediction using real-time surface parameters including WOB, RPM, torque, standpipe
pressure, and flow rate. Their model achieved a coefficient of determination (R²) of 0.99 on
training data and demonstrated significant improvement over traditional Bourgoyne and
Young predictions. The model accuracy is measured using the root mean squared error:

RMSE = √ [Σ (ypredicted − yactual)2 / n] [2.36]

2.5.2 Support Vector Machines (SVM)

Support vector machines, developed by Vapnik (1995), represent a powerful supervised
learning approach that can handle both classification and regression problems. For ROP
prediction, support vector regression (SVR) is employed, which seeks to find a function that
deviates from the actual ROP values by no more than a specified margin (ε) whilst being as
flat as possible.

The SVR optimization problem can be formulated as:

min (1/2) ||w||2 + C × Σ (ξi + ξi*) [2.37]

subject to: yi − (w·xi + b) ≤ ε + ξi and (w·xi + b) − yi ≤ ε + ξi*, where w is the weight vector,
C is the regularisation parameter, ξi and ξi* are slack variables, and ε is the epsilon-
insensitive loss. The kernel function K (xi, xj) enables SVR to handle non-linear
relationships by mapping inputs to a higher-dimensional feature space. Common kernels
include linear, polynomial, and radial basis function (RBF):

K (xi, xj) = exp (−γ||xi − xj||2) [2.38]

Bataee et al. (2014) applied SVR to ROP prediction and found that it provided robust
performance, particularly when dealing with limited training data. The regularisation
mechanism inherent in SVM helps prevent overfitting, making it suitable for generalisation
to new drilling scenarios.

20
2.5.3 Random Forest
Random Forest, introduced by Breiman (2001), is an ensemble learning method that
constructs multiple decision trees during training and outputs the average prediction of
individual trees for regression tasks. The algorithm introduces randomness through
bootstrap sampling and random feature selection at each split, thereby reducing overfitting
and improving generalisation.

For a random forest with M trees, the ROP prediction is calculated as:

ROPpredicted = (1/M) × Σ Tm(x) [2.39]

where Tm(x) is the prediction of the mth tree for input vector x. Random Forest provides
inherent feature importance measures through the mean decrease in impurity or permutation
importance, which aids in understanding the relative contribution of drilling parameters to
ROP.

Hegde and Gray (2017) demonstrated that Random Forest models could achieve high
prediction accuracy whilst providing insights into feature importance, making them valuable
for both prediction and interpretation. Delavar et al. (2023) combined Random Forest with
multi-objective particle swarm optimization to simultaneously optimise drilling parameters
for maximum ROP and minimum cost.

2.5.4 Extreme Gradient Boosting (XGBoost)

Extreme Gradient Boosting (XGBoost), developed by Chen and Guestrin (2016), is an

optimised implementation of gradient boosting that has achieved state-of-the-art results in
numerous machine learning competitions. XGBoost builds on the gradient boosting
machine framework established by Friedman (2001), which constructs an additive model in
a forward stagewise fashion:

F(x) = Σ fm(x) [2.40]

where each fm is a weak learner (typically a decision tree) that is fitted to the negative
gradient of the loss function. XGBoost introduces several enhancements including
regularised learning objective, efficient handling of sparse data, and parallel processing
capabilities. The regularised objective function is:

L = Σ l(yi, ŷi) + Σ Ω(fm) [2.41]

21
where l is the loss function and Ω(f) = γT + (1/2)λ||w||² is the regularisation term, with T
being the number of leaves and w being the leaf weights. Agbaji (2024) applied XGBoost
to ROP prediction and achieved superior performance compared to other machine learning
methods, with the added benefit of interpretability through SHAP analysis.

2.5.5 Deep Learning Approaches

Deep learning extends traditional neural networks by incorporating multiple hidden layers,
enabling the automatic extraction of hierarchical features from raw data. Recent advances
in computational power and algorithmic improvements have made deep learning
increasingly practical for drilling applications.

Long Short-Term Memory (LSTM) networks, a type of recurrent neural network (RNN),
are particularly suited for time-series drilling data due to their ability to capture temporal
dependencies. The LSTM cell incorporates gating mechanisms that control information
flow:

ft = σ(Wf · [ht−1, xt] + bf) [2.42]

it = σ(Wi · [ht−1, xt] + bi) [2.43]

ot = σ(Wo · [ht−1, xt] + bo) [2.44]

Liu et al. (2024) developed a BiLSTM-SA-IDBO model that combines bidirectional LSTM
with self-attention mechanisms for enhanced ROP prediction. Their approach demonstrated
improved accuracy by capturing both forward and backward temporal dependencies in
drilling data.

22
Figure 2. 1: Schematic representation of AI-based real-time ROP optimization workflow
as implemented in recent field deployments
2.5.6 Field Applications and Practical Benefits of AI-Based ROP Optimization
While the preceding sections have established the theoretical foundations and predictive
capabilities of machine learning approaches for ROP modelling, the practical value of these
methods is ultimately demonstrated through field implementation. Recent years have
witnessed growing evidence from real-world deployments that validate the operational
benefits of AI-based drilling optimization systems. These field trials have documented
substantial improvements in drilling performance, including increased penetration rates,
reduced drilling times, and decreased equipment failures.

Al-Riyami et al. (2023) presented comprehensive field test results from a real-time ROP
optimization system deployed across multiple drilling campaigns. In a Malaysian test well,
following the machine learning system's parameter recommendations yielded a threefold
increase in ROP from 10.4 to 31 m/h over a 136-metre drilling interval. Perhaps more
significantly, a West Africa field trial demonstrated the cumulative impact of sustained
optimization, with total depth reached approximately 24 days ahead of the original drilling

23
plan and 2.4 days ahead of the expected technical limit. These results illustrate how machine
learning optimization can translate laboratory-scale model accuracy into substantial
operational time savings.

Singh et al. (2021) reported systematic field validation of a cloud-based machine learning
ROP optimization solution deployed in the Permian Basin. Their field trials demonstrated
ROP improvements ranging from 19 to 33 percent when comparing optimized drilling
intervals against observation and control footage within the same wells. When benchmarked
against qualified offset wells with similar target formations, bottom-hole assemblies, and
wellbore trajectories, ROP improvements ranged from 5 to 33 percent. Importantly, post-
run data analysis confirmed that the optimization recommendations did not induce
detrimental effects such as directional trend divergence or elevated axial and lateral
vibrations, addressing concerns about the safety implications of aggressive ROP
optimization.

Beyond direct ROP improvements, AI-based drilling optimization has demonstrated

significant benefits in mitigating downhole dysfunctions that can compromise both drilling
efficiency and equipment integrity. Boukredera et al. (2023) developed an integrated
machine learning and particle swarm optimization workflow that addressed the
interdependence between ROP optimization and vibration management. Their simulations
predicted a 43 percent increase in ROP whilst simultaneously improving torsional stability
through intelligent parameter selection. The practical implications were substantial:
optimized drilling parameters reduced projected drilling time from 66 to 31 hours whilst
maintaining the stick-slip index below acceptable thresholds, potentially eliminating the
need for premature bit trips associated with vibration-induced wear.
The impact of intelligent parameter optimization on bottom-hole assembly (BHA) reliability
represents another dimension of practical benefit. Uncontrolled drill string vibrations,
including axial, lateral, and torsional modes, constitute a primary cause of BHA tool failure
and premature wear. Recent implementations of AI-based BHA optimization systems have
reported ROP improvements of 15 to 20 percent alongside reductions in non-productive
time attributable to BHA failures. These multi-objective benefits, encompassing improved
penetration rates, reduced drilling time, enhanced equipment reliability, and decreased non-
productive time, collectively strengthen the economic case for deploying machine learning
optimization systems in operational drilling environments.

24
Table 2.1 summarises the documented performance improvements from selected field
implementations of AI-based ROP optimization systems.

Table 2. 1: Summary of Field-Documented AI-Based ROP Optimization Benefits

Study Location/Context ROP Improvement
Al-Riyami et al. (2023) Malaysia, West Africa 10.4 to 31 m/h (198%)
Singh et al. (2021) Permian Basin, USA 19–33%
Boukredera et al. (2023) Algeria (simulation) 43%

These field results provide empirical validation for the theoretical capabilities demonstrated
in laboratory studies and establish a compelling case for the practical deployment of
machine learning optimization systems in drilling operations. The consistency of benefits
across diverse geographic locations, formation types, and operational contexts suggests that
the advantages of AI-based optimization are robust and transferable across drilling
environments

2.6 Optimization Approaches in Drilling

Once a reliable ROP prediction model is established, optimization algorithms can be
employed to determine the optimal combination of controllable drilling parameters that
maximise ROP whilst satisfying operational constraints. This section reviews the major
metaheuristic optimization approaches applied to drilling parameter optimization.

2.6.1 Genetic Algorithms

Genetic algorithms (GA), introduced by Holland (1975) and popularised by Goldberg

(1989), are evolutionary optimization techniques inspired by natural selection. GAs
maintains a population of candidate solutions that evolve over generations through
selection, crossover, and mutation operations. Guria et al. (2014) applied multi-objective
genetic algorithms with NSGA-II to optimise drilling parameters, demonstrating the ability
to simultaneously optimise multiple conflicting objectives such as maximum ROP and
minimum drilling cost.

The fitness function for ROP maximisation is typically defined as:

Fitness = ROP (WOB, RPM, Q) subject to operational constraints [2.45]

2.6.2 Particle Swarm Optimization

Particle swarm optimization (PSO), developed by Kennedy and Eberhart (1995), is a
population-based stochastic optimization technique inspired by the social behaviour of bird

25
flocking or fish schooling. Each particle represents a potential solution and moves through
the search space guided by its own best-known position and the swarm's best-known
position.

The velocity update equation for particle i at iteration k+1 is:

vi(k+1) = w × vi(k) + c1 × r1 × (pbesti − xi(k)) + c2 × r2 × (gbest − xi(k)) [2.46]
where w is the inertia weight, c1 and c2 are acceleration coefficients for cognitive and social
components, r1 and r2 are random numbers in [0,1], pbesti is the particle's personal best
position, and gbest is the global best position. The position update is:
xi(k+1) = xi(k) + vi(k+1) [2.47]

Onwunalu and Durlofsky (2010) successfully applied PSO to well placement optimization,
whilst Sridharan and Kumar (2020) integrated PSO with ANN models to optimise drilling
parameters in real-time. Their ANN-PSO approach demonstrated the ability to achieve
optimal ROP within practical computational constraints suitable for rig-site implementation.

2.6.3 Grey Wolf Optimiser

The Grey Wolf Optimiser (GWO), introduced by Mirjalili et al. (2014), is a metaheuristic
algorithm inspired by the hunting behaviour and social hierarchy of grey wolves. The
algorithm mimics the leadership hierarchy (alpha, beta, delta, and omega wolves) and
hunting mechanisms (searching, encircling, and attacking prey).

The position update equations are:

X(t+1) = (X1 + X2 + X3) / 3 [2.48)]

where X1, X2, and X3 represent positions guided by the three best solutions (alpha, beta,
and delta wolves). Recent studies have demonstrated GWO's effectiveness in drilling
optimization due to its balance between exploration and exploitation.

2.7 Feature Analysis and Interpretability

While machine learning models offer superior predictive performance, their "black box"
nature can limit their acceptance in drilling operations where understanding the underlying
mechanisms is important for decision-making. Feature analysis and model interpretability
techniques address this limitation by providing insights into how models make predictions.

2.7.1 Feature Importance Methods

Feature importance quantifies the contribution of each input variable to the model's
predictions. For tree-based models such as Random Forest and XGBoost, importance can
26
be calculated using mean decrease in impurity (MDI) or permutation importance. Molnar
(2022) provides a comprehensive treatment of these methods and their interpretation.

Permutation importance measures the increase in prediction error when a feature's values
are randomly shuffled:

PI(j) = errorpermuted(j) − errororiginal [2.49]

where higher values indicate greater feature importance.

2.7.2 SHAP (SHapley Additive exPlanations)

SHAP, introduced by Lundberg and Lee (2017), provides a unified framework for
interpreting predictions based on Shapley values from cooperative game theory. The
theoretical foundations were established by Štrumbelj and Kononenko (2010, 2014), who
first proposed using Shapley values to explain machine learning predictions.

The Shapley value for feature i is calculated as:

φi = Σ [|S|!(|N|−|S|−1)! / |N|!] × [f(S∪{i}) − f(S)] [2.50]

where N is the set of all features, S is a subset of features excluding i, and f(S) is the model
prediction using only features in S. SHAP values satisfy several desirable properties
including local accuracy, consistency, and missingness.

The sum of all SHAP values equals the difference between the model prediction and the
expected (baseline) prediction:

f(x) = E[f(X)] + Σ φi(x) [2.51]

SHAP analysis provides both local explanations (individual predictions) and global
interpretations (overall feature importance), making it a valuable tool for understanding
ROP prediction models.

2.8 Research Gaps

Despite significant advances in ROP modelling and optimization, several research gaps
remain that limit the practical application of these techniques:

Model Transferability: Most machine learning models are trained on data from specific
fields or formations and may not generalise well to new drilling environments. The
development of transfer learning approaches that can adapt pre-trained models to new
scenarios with limited data remains an open challenge (Hai et al., 2025).
27
Real-Time Implementation: While sophisticated models demonstrate excellent offline
performance, their integration into real-time drilling optimization systems poses practical
challenges related to computational efficiency, data quality, and latency requirements.

Interpretability-Accuracy Trade-off: Complex models that achieve higher accuracy often

sacrifice interpretability, making it difficult for drilling engineers to trust and adopt their
recommendations. Developing inherently interpretable models that maintain high accuracy
is an ongoing research area.

Integration of Domain Knowledge: Most machine learning approaches treat ROP

prediction as a purely data-driven problem, potentially missing opportunities to incorporate
established drilling physics and engineering constraints into the modelling framework.

Multi-Objective Optimization: Practical drilling operations involve multiple, often

conflicting objectives beyond ROP maximisation, including cost minimisation, equipment
longevity, and safety considerations. Comprehensive frameworks for multi-objective
drilling optimization remain underdeveloped.

2.9 Conceptual Framework

Based on the literature review, this study adopts a conceptual framework that integrates
machine learning prediction with metaheuristic optimization for ROP enhancement. The
framework comprises three main components:

Data Processing Layer: This layer handles data acquisition, quality control, and feature
engineering. Raw drilling data is cleaned using outlier detection methods (Z-score analysis)
and smoothed using appropriate filters to remove noise whilst preserving signal integrity.

Prediction Layer: Multiple machine learning models (ANN, Random Forest, XGBoost)
are trained and validated using cross-validation techniques. Model performance is evaluated
using metrics including R², RMSE, and MAE. The best-performing model is selected for
integration with the optimization layer.

Optimization Layer: The trained prediction model serves as the fitness function for
metaheuristic optimization algorithms (PSO, GA). The optimiser searches for optimal
combinations of controllable parameters (WOB, RPM, flow rate) that maximise predicted
ROP whilst satisfying operational constraints.

28
The theoretical foundations supporting this framework draw from Bourgoyne and Young's
(1974) comprehensive ROP model, which establishes the functional relationships between
drilling parameters and penetration rate, and modern machine learning theory, which
provides the tools for capturing complex, non-linear patterns in drilling data. The integration
of SHAP analysis ensures that model predictions remain interpretable, facilitating adoption
by drilling engineers.

29
 CHAPTER THREE: RESEARCH METHODOLOGY

3.1 Introduction
This chapter presents a comprehensive description of the research methodology adopted to
achieve the objectives of this study. The methodology encompasses the research design and
philosophical approach, data acquisition and detailed description, data preprocessing
procedures, feature analysis techniques with their mathematical foundations, machine
learning model development including architecture specifications and hyperparameter
configurations, model evaluation metrics, the optimization approach for identifying optimal
drilling parameters, and validation methods for ensuring physical plausibility of results.
Each section provides sufficient detail to enable replication of the research and application
of the methodology to other drilling datasets.

The methodological framework follows a systematic, four-phase approach that integrates

data-driven analysis with validation using established drilling mechanics principles
(Bourgoyne et al., 1986; Sridharan and Kumar, 2020). Phase one involves comprehensive
feature analysis to understand the relationships between drilling parameters and the rate of
penetration. Phase two develops and trains multiple machine learning models using the
prepared dataset. Phase three applies the best-performing model to identify optimal drilling
parameters. Phase four validates the optimization results through sensitivity analysis and
comparison with theoretical expectations from drilling mechanics literature.

The research methodology was designed to address the specific challenges of working with
drilling data, including the relatively small sample size typical of drilling applications, the
presence of multicollinearity among input variables, the non-linear relationships between
drilling parameters and ROP, and the need for model interpretability to enable practical
implementation of results (Hegde and Gray, 2017). Each methodological choice was made
with consideration for these challenges and the goal of producing robust, interpretable, and
practically applicable results.

3.2 Research Design

This study adopts a quantitative research approach using secondary drilling data to develop,
compare, and validate machine learning models for rate of penetration prediction and
optimization. The research follows an empirical methodology that combines data-driven
model development with validation using established drilling mechanics principles,
representing an applied research design aimed at solving practical drilling optimization

30
problems while contributing to the academic understanding of machine learning
applications in drilling engineering.

3.2.1 Research Philosophy

The study is grounded in a positivist research philosophy, which holds that knowledge can
be derived from observable, measurable phenomena and that relationships between
variables can be identified through systematic observation and analysis (Creswell and
Creswell, 2018). This philosophical stance is appropriate for the current research because
drilling operations generate quantitative data that can be objectively measured and analysed,
the relationships between drilling parameters and ROP are governed by physical principles
that manifest as statistical patterns in the data, and the goal of developing predictive models
aligns with the positivist emphasis on explanation and prediction.

The positivist approach is complemented by elements of pragmatism, particularly in the

validation phase, where the practical utility and physical plausibility of results are evaluated
alongside statistical performance metrics. This pragmatic element recognises that the
ultimate value of the research lies in its ability to inform real drilling decisions, not merely
in achieving high prediction accuracy on test data.

3.2.2 Research Approach

The study employs a deductive research approach, moving from general theories of drilling
mechanics and machine learning to specific hypotheses about model performance and
parameter relationships that are tested using the drilling dataset. The deductive approach is
manifest in several aspects of the research design. First, the selection of input variables is
guided by established drilling theory, particularly the factors identified in the Bourgoyne
and Young (1974) model and subsequent literature. Second, the interpretation of feature
importance results is evaluated against theoretical expectations from drilling mechanics.
Third, the validation of optimization results uses sensitivity analysis to compare model
behaviour with theoretically expected relationships.

The research also incorporates inductive elements, particularly in allowing the machine
learning algorithms to identify patterns and relationships in the data that may not be
explicitly specified in traditional drilling models (Bilgesu et al., 1997). This combination of
deductive and inductive approaches leverages both domain knowledge and data-driven
pattern recognition.

3.2.3 Research Strategy

31
The research employs an experimental strategy using computational experiments with
machine learning algorithms applied to historical drilling data. This strategy involves
systematically manipulating model configurations, comparing multiple algorithms under
controlled conditions, and quantitatively evaluating outcomes. The computational-
experimental approach enables rigorous comparisons of different modelling approaches
while controlling for factors such as data preprocessing, train-test splitting, and evaluation
metrics.

The experimental design follows a comparative structure, in which three machine learning
algorithms (Random Forest, XGBoost, and Artificial Neural Network) are developed,
trained, and evaluated using identical data preprocessing and evaluation procedures. This
controlled comparison enables valid conclusions about relative algorithm performance for
the specific task of ROP prediction with small drilling datasets.

3.2.4 Research Framework

The overall research framework consists of four interconnected phases that correspond to
the four research objectives:

Phase 1 (Feature Analysis): Comprehensive analysis of drilling parameters using

correlation analysis, Random Forest feature importance, and SHAP values to identify the
key factors influencing ROP and understand their relationships.

Phase 2 (Model Development): Development and training of Random Forest, XGBoost,

and ANN models using standardised preprocessing and consistent train-test splitting to
enable fair comparison.

Phase 3 (Optimization): Application of each trained model to identify optimal drilling

parameters through grid search over the feasible parameter space.

Phase 4 (Validation): Validation of optimization results through sensitivity analysis,

comparison of results across models, and evaluation against drilling mechanics principles.

The phased approach ensures systematic progression from understanding the data (Phase 1)
through model development (Phase 2) and optimization (Phase 3) to validation (Phase 4),
with each phase building on the results of previous phases.

32
3.3 Data Acquisition and Description
The dataset used in this study comprises drilling data collected from a single drilling section
using a polycrystalline diamond compact (PDC) bit with a diameter of 8.5 inches. The data
were obtained from operational drilling records and represent actual field conditions
encountered during drilling operations. The focus on a single bit type and size eliminates
bit-related variability that could confound the analysis of drilling parameter effects,
allowing clearer identification of the relationships between controllable parameters and
ROP (Soares and Gray, 2019).

3.3.1 Dataset Characteristics

The dataset contains 43 observations across 31 variables, representing drilling data collected
at discrete depth intervals throughout the drilling section. Each observation corresponds to
a specific depth interval and includes measurements of drilling parameters, formation
response indicators, and the resulting rate of penetration. The relatively small sample size
of 43 observations is typical of many practical drilling optimization applications where
comprehensive data collection is challenging, expensive, or limited by operational
constraints (Anemangely et al., 2018).

This sample size presents both challenges and opportunities: it requires careful attention to
overfitting and model selection, but also makes the research findings directly applicable to
many real-world situations where drilling engineers must work with limited data. The
drilling section was drilled using consistent operational practices, equipment, and mud
system throughout, reducing the variability that could arise from changes in the drilling
programme or equipment.

3.3.2 Variable Categories

The 31 variables in the dataset can be organised into seven categories based on their nature
and role in the drilling process. Table 3.1 presents a comprehensive summary of these
categories and the specific variables included in each.

33
Table 3. 1: Comprehensive Categories of Variables in the Dataset

Category Variables Description

Target Variable ROP Rate of Penetration (m/hr)
WOB, RPM, FLOW, Surface Parameters adjustable by driller
Controllable Parameters
Torque, Standpipe Pressure during operations
Measured and true vertical depth
Depth Parameters MD, MD_IN, MD_OUT, TVD
indicators
Mud Weight, PV, YP, Flow
Hydraulic Parameters Drilling fluid and hydraulic properties
Rate, ECD
Differential Pressure, Annular Mechanical and pressure-related
Mechanical Parameters
Velocity, JET_IMPACT measurements
Bit Size, Nozzle Config, TFA,
Bit Parameters Bit geometry and operational history
Bit Hours
MSE, d-exponent, Corrected d- Calculated drilling performance
Derived Parameters
exponent indicators

3.3.3 Target Variable

The target variable for prediction is the rate of penetration (ROP), measured in metres per
hour (m/hr). ROP represents the speed at which the drill bit advances through the formation
and is the primary performance metric for drilling efficiency (Bourgoyne et al., 1986). In
this dataset, ROP values were calculated from the depth drilled during each observation
interval divided by the corresponding drilling time:

ROP = ΔD / Δt [3.1]

where ΔD is the depth drilled during the interval (difference between MD_OUT and
MD_IN) and Δt is the elapsed drilling time. The ROP values in the dataset range from
approximately 8 m/hr to 28 m/hr, reflecting the variation in drilling conditions and formation
characteristics encountered throughout the section.

3.3.4 Controllable Parameters

Controllable parameters are those that can be adjusted by the driller during operations to
influence drilling performance. The primary controllable parameters in the dataset include:

Weight on Bit (WOB): The downward force applied to the drill bit, measured in thousands
of pounds (klbs). WOB is controlled by adjusting the amount of drill string weight supported
by the drawworks versus the weight transferred to the bit. The WOB values in the dataset
reflect the operational range used during drilling of the section.

34
Rotary Speed (RPM): The rate of rotation of the drill string and bit, measured in
revolutions per minute. RPM is controlled through the top drive or rotary table and affects
the rate at which the bit cutters engage with the formation (Maurer, 1962).

Flow Rate (FLOW): The volume of drilling fluid pumped through the drill string and out
through the bit nozzles, measured in gallons per minute (GPM). Flow rate affects bottom-
hole cleaning, bit cooling, and hydraulic impact force (Abbas et al., 2018).

Surface Torque: The rotational force required to turn the drill string, measured at surface.
Torque provides an indication of the resistance encountered during drilling and can signal
drilling dysfunctions or formation changes.

Standpipe Pressure: The pressure in the standpipe manifold, representing the total system
pressure required to circulate drilling fluid through the entire flow path.

3.3.5 Depth and Formation-Related Parameters

Depth parameters serve as proxies for formation characteristics in the absence of explicit
geomechanical data. The depth-related variables in the dataset include:

Measured Depth (MD): The total length of the wellbore from surface to the current bit
position, measured along the wellbore path. MD is the primary reference for correlating
drilling data with geological information.

MD_IN and MD_OUT: The measured depths at the start and end of each observation
interval, respectively. These values define the depth range over which each ROP observation
was calculated.

True Vertical Depth (TVD): The vertical distance from surface to the bit position, which
may differ from MD in deviated wells. TVD is used for pressure calculations and correlation
with seismic data.

These depth parameters are significant because formation properties (strength, porosity,
pore pressure) typically vary systematically with depth due to compaction and diagenetic
processes (Barros, 2015). In the absence of direct formation measurements, depth serves as
a proxy that captures these depth-dependent property variations.

35
3.3.6 Hydraulic and Mechanical Parameters
The hydraulic parameters describe the drilling fluid system and its interaction with the
wellbore. Key hydraulic variables include mud weight (density of drilling fluid), plastic
viscosity (PV), yield point (YP), and equivalent circulating density (ECD). The mechanical
parameters include differential pressure between the wellbore and formation, annular
velocity of drilling fluid returning to surface, and jet impact force (JET_IMPACT)
calculated from the hydraulic horsepower delivered at the bit nozzles (Bourgoyne et al.,
1986):

Fj = 0.000516 × ρm × Q × Vn [3.2]

where Fj is the jet impact force in pounds-force, ρm is the mud density in pounds per gallon,
Q is the flow rate in gallons per minute, and Vn is the nozzle velocity in feet per second.

3.3.7 Derived Parameters

The dataset includes several derived parameters calculated from primary measurements.
Mechanical Specific Energy (MSE) represents the energy required to remove a unit volume
of rock (Teale, 1965; Dupriest and Koederitz, 2005):

MSE = (480 × T × N) / (db2 × ROP) + (4 × WOB) / (π × db2) [3.3]

where T is torque, N is RPM, db is bit diameter, and ROP is rate of penetration. The d-
exponent is a normalised drilling exponent that accounts for the effects of WOB and RPM
on ROP (Rehm and McClendon, 1971):

d = log (ROP/60N) / log(12WOB/1000db) [3.4]

The corrected d-exponent (dc) further normalises for mud weight effects and is commonly
used for pore pressure prediction:

dc = d × (ρn/ρa) [3.5]

where ρn is the normal pore pressure gradient (typically 8.5 ppg) and ρa is the actual mud
weight.

3.3.8 Descriptive Statistics

Table 3.2 presents the descriptive statistics for the key variables in the dataset, providing an
overview of the data distribution and range of values encountered.

36
Table 3. 2: Descriptive Statistics for Key Variables

Variable Mean Std Dev Min Max Unit

ROP 17.42 5.31 8.12 27.85 m/hr
WOB 125.6 28.4 85.0 158.0 klbs
RPM 142.5 24.8 90.0 180.0 rev/min
FLOW 5.85 0.72 4.50 6.55 GPM
MD 2450.3 185.6 2100.0 2780.0 m
TORQUE 8.45 2.12 5.20 12.80 kft-lbs

3.4 Data Preprocessing

Data preprocessing is a fundamental step in any machine learning workflow, preparing raw
data for model training by addressing issues such as missing values, outliers, and differences
in variable scales (Pedregosa et al., 2011). The preprocessing procedures applied in this
study were designed to ensure data quality while preserving the information content
necessary for accurate ROP prediction.

3.4.1 Data Quality Assessment

The first step in preprocessing involved a comprehensive assessment of data quality. The
dataset was examined for missing values, duplicate records, and obvious data entry errors.
The examination revealed no missing values in any of the 31 variables across the 43
observations, indicating complete data capture during drilling operations. No duplicate
records were identified based on comparison of depth intervals and timestamps. Visual
inspection of variable distributions and scatter plots against ROP did not reveal obvious
outliers or data entry errors that would require correction or removal.

The absence of missing values eliminated the need for imputation strategies, which can
introduce bias or reduce prediction accuracy if not carefully implemented (Little and Rubin,
2019). The completeness of the dataset is attributed to the automated data acquisition
systems used during drilling operations and the careful data validation procedures applied
during data extraction.

3.4.2 Feature Scaling

Machine learning algorithms, particularly neural networks and distance-based methods, can
be sensitive to the scale of input features (LeCun et al., 2012). Features with larger
magnitudes or wider ranges can dominate the learning process, potentially leading to

37
suboptimal model performance. To address this issue, feature scaling was applied to
standardise all input variables to a common scale.

Standardisation (z-score normalisation) was selected as the scaling method, which

transforms each feature to have zero mean and unit variance:

zi = (xi − μ) / σ [3.6]

where zi is the standardised value, xi is the original value, μ is the mean of the feature across
all observations, and σ is the standard deviation of the feature. Standardisation was chosen
over min-max normalisation because it is less sensitive to outliers and preserves the relative
distances between observations. The StandardScaler class from the scikit-learn library was
used to implement this transformation.

Importantly, the scaling parameters (mean and standard deviation) were calculated using
only the training data and then applied to both training and test data. This approach prevents
information leakage from the test set into the training process, which could lead to overly
optimistic performance estimates (Hastie et al., 2009):

ztrain = (xtrain − μtrain) / σtrain [3.7]

ztest = (xtest − μtrain) / σtrain [3.8]

The target variable (ROP) was not scaled, as this would complicate interpretation of model
predictions and error metrics.

3.4.3 Train-Test Split

The dataset was divided into training and testing subsets to enable unbiased evaluation of
model performance on data not used during training. An 80/20 split ratio was adopted,
resulting in 34 observations for training and 9 observations for testing:

ntrain = 0.80 × 43 = 34 observations [3.9]

ntest = 0.20 × 43 = 9 observations [3.10]

Random sampling with a fixed random seed (seed = 42) was used to perform the split,
ensuring reproducibility of results. The random seed allows other researchers to replicate
the exact train-test split and verify the reported results. The 80/20 ratio represents a common
practice in machine learning that balances the competing needs of having sufficient data for
training while reserving enough data for meaningful evaluation (Kohavi, 1995).

38
It is acknowledged that the test set of 9 observations is small, which introduces uncertainty
into performance estimates. However, this limitation reflects the reality of many drilling
applications where data availability is constrained. The convergence analysis across
multiple models provides an additional validation mechanism that partially compensates for
the limited test set size.

3.4.4 Feature Selection Considerations

The dataset contains 31 variables, some of which exhibit high multicollinearity (e.g., MD,
MD_IN, and MD_OUT are perfectly correlated). While multicollinearity does not
necessarily impair the predictive performance of machine learning models, it can complicate
interpretation of feature importance and increase computational requirements (Dormann et
al., 2013). The feature selection approach in this study was guided by the following
considerations.

First, all variables were initially included in the feature analysis phase to ensure
comprehensive examination of potential predictors. Second, the machine learning
algorithms selected (Random Forest and XGBoost) are inherently robust to multicollinearity
through their ensemble and regularisation mechanisms (Breiman, 2001; Chen and Guestrin,
2016). Third, SHAP analysis was used to interpret model behaviour even in the presence of
correlated features (Lundberg and Lee, 2017). This approach allows the data to inform
feature relevance through the feature importance analysis while leveraging the robustness
of ensemble methods to handle correlated inputs.

3.5 Feature Analysis Methods

Feature analysis was conducted using four complementary methods to provide a
comprehensive understanding of the relationships between drilling parameters and ROP.
Each method has different theoretical foundations and captures distinct aspects of feature
relevance, and using multiple methods enables validation of findings across approaches and
identification of both linear and non-linear relationships.

3.5.1 Correlation Analysis

Correlation analysis examines the linear relationships between pairs of variables. The
Pearson correlation coefficient was calculated for all pairs of variables to construct a
correlation matrix. For variables X and Y, the Pearson correlation coefficient r is defined
as:

39
rXY = Σi=1n(xi − x̄)(yi − ȳ) / √[Σi=1n(xi − x̄)2 × Σi=1n(yi − ȳ)2] [3.11]

where xi and yi are individual observations, x̄ and ȳ are the sample means, and n is the
number of observations. The correlation coefficient ranges from −1 (perfect negative
correlation) to +1 (perfect positive correlation), with 0 indicating no linear relationship.

The correlation analysis serves multiple purposes. First, it identifies variables with strong
linear relationships to ROP, which are candidates for important predictors. Second, it reveals
multicollinearity among input variables that may affect model interpretation. Third, it
provides a baseline against which non-linear feature importance measures can be compared.

The statistical significance of correlation coefficients was assessed using the t-test with the
null hypothesis that the true correlation is zero:

t = r × √(n − 2) / √(1 − r2) [3.12]

with n−2 degrees of freedom. Correlations with p-values less than 0.05 were considered
statistically significant.

3.5.2 Random Forest Feature Importance

Random Forest provides a natural measure of feature importance through the mean decrease
in impurity (MDI), also known as Gini importance (Breiman, 2001). When a feature is used
for splitting a node, the improvement in the split criterion (reduction in variance for
regression) is recorded. The MDI for feature j is the sum of all improvements achieved using
that feature, averaged across all trees:

MDI(j) = (1/B) × Σb=1B Σt∈Tb I(v(t) = j) × ΔI(t) [3.13]

where B is the number of trees, Tb is tree b, v(t) is the splitting variable at node t, I(·) is the
indicator function, and ΔI(t) is the improvement in the splitting criterion at node t.

MDI importance has several advantages: it is computationally efficient (calculated during

training), accounts for non-linear relationships, and naturally handles feature interactions.
However, it can be biased towards features with many categories or high cardinality, and it
measures importance for the training data rather than predictive importance for new data.

The normalised importance is calculated as:

Importancenorm(j) = MDI(j) / Σk=1p MDI(k) [3.14]

where p is the total number of features.

40
3.5.3 Permutation Importance
Permutation importance provides a model-agnostic measure of feature importance that
assesses how much the model's performance degrades when a feature's values are randomly
shuffled (Fisher et al., 2019). The permutation importance for feature j is calculated as:

PI(xj) = S − Sj,perm [3.15]

where S is the model's performance score on the original data and Sj,perm is the score after
randomly permuting feature j. A large positive value indicates that the feature is important
for predictions, while values near zero suggest the feature provides little predictive
information.

Permutation importance has several advantages over MDI: it is calculated on held-out data,
so it measures predictive importance rather than training importance; it is model-agnostic
and can be applied to any model; and it is not biased by feature cardinality. However, it can
be affected by correlated features, where permuting one feature may not significantly
degrade performance if correlated features compensate.

3.5.4 SHAP Analysis

SHAP (SHapley Additive exPlanations) values provide a unified framework for interpreting
model predictions based on concepts from cooperative game theory (Lundberg and Lee,
2017). The Shapley value for feature j in predicting observation i is the average marginal
contribution of that feature across all possible orderings of features:

φj(i) = ΣS⊆N\{j} [|S|!(|N|−|S|−1)!/|N|!] × [f(S∪{j}) − f(S)] [3.16]

where N is the set of all features, S is a subset of features not including j, and f(S) represents
the model prediction using only features in S.

SHAP values have several desirable properties. Local accuracy ensures that the sum of
SHAP values equals the difference between the prediction and the expected value:

Σj=1M φj(i) = f(x(i)) − E[f(X)] [3.17]

Consistency ensures that if a feature's contribution increases, its SHAP value does not
decrease. Missingness ensures that features with no impact receive zero SHAP values.

Global feature importance can be derived from SHAP values by averaging absolute SHAP
values across all observations:

Importance(xj) = (1/n) × Σi=1n |φj(i)| [3.18]

41
3.6 Machine Learning Model Development
Three machine learning algorithms were selected for ROP prediction: Random Forest,
XGBoost, and Artificial Neural Network. These algorithms represent different approaches
to learning from data and have been successfully applied to ROP prediction in previous
studies (Hegde and Gray, 2017; Agbaji, 2024). The use of multiple algorithms enables
comparison of performance and validation of optimization results through convergence
analysis.

3.6.1 Random Forest Model

[Link] Algorithm Description
Random Forest is an ensemble learning method that constructs multiple decision trees
during training and outputs the average of their predictions for regression tasks (Breiman,
2001). Each tree is trained on a bootstrap sample of the training data, and at each node, only
a random subset of features is considered for splitting. This dual randomisation reduces
correlation among trees and improves generalisation.

The prediction for a new observation x is the average of predictions from all trees:

ŷ = (1/B) × Σb=1B Tb(x) [3.19]

where B is the number of trees and Tb(x) is the prediction from tree b. The averaging reduces
variance compared to a single tree while maintaining low bias, making Random Forest
particularly effective for small to medium-sized datasets where individual trees might
overfit.

[Link] Hyperparameter Configuration

The Random Forest model was configured with hyperparameters selected based on
established best practices and preliminary experimentation. Table 3.3 summarises the
hyperparameter configurations for all three machine learning models used in this study.

42
Table 3. 3: Machine Learning Model Hyperparameter Configurations

Hyperparameter Random Forest XGBoost ANN

Number of
100 trees 100 trees 500 max epochs
estimators/epochs
Learning rate N/A 0.1 0.001
Max depth None (full) 6 N/A
Regularisation Bootstrap L2 (λ=1) Dropout (0.2)
Subsample ratio Bootstrap (≈63%) 0.8 N/A
Random state 42 42 42

3.6.2 XGBoost Model

[Link] Algorithm Description
XGBoost (Extreme Gradient Boosting) is a gradient boosting algorithm that builds trees
sequentially, with each new tree designed to correct the errors of the existing ensemble
(Chen and Guestrin, 2016). XGBoost optimises an objective function that combines a loss
term and a regularisation term:

Obj = Σi=1n L(yi, ŷi) + Σk=1K Ω(fk) [3.20]

where L is the loss function (squared error for regression), ŷi is the prediction for
observation i, K is the number of trees, and Ω is the regularisation term. The regularisation
term penalises model complexity:

Ω(f) = γT + (1/2)λ||w||2 [3.21]

where T is the number of leaves in tree f, w is the vector of leaf weights, γ controls the
penalty for adding more leaves, and λ is the L2 regularisation parameter for leaf weights.

Trees are added sequentially to minimise the objective. At iteration t, the model prediction
is:

ŷi(t) = ŷi(t−1) + η × ft(xi) [3.22]

where η is the learning rate (shrinkage parameter) and ft is the new tree. The new tree is fit
to the negative gradient of the loss with respect to the current predictions:

gi = ∂L(yi, ŷi(t−1)) / ∂ŷi(t−1) [3.23]

For squared error loss, the negative gradient is simply the residual (yi − ŷi), so each new
tree attempts to predict the errors of the previous model.

43
3.6.3 Artificial Neural Network Model
[Link] Network Architecture
A feedforward neural network architecture was adopted following established practices in
drilling applications (Bilgesu et al., 1997; Sridharan and Kumar, 2020). Table 3.4 presents
the detailed architecture of the neural network model.

Table 3. 4 : Artificial Neural Network Architecture

Layer Neurons Activation Regularisation

Input 30 None None
Dropout (0.2),
Hidden 1 64 ReLU
BatchNorm
Dropout (0.2),
Hidden 2 32 ReLU
BatchNorm
Dropout (0.2),
Hidden 3 16 ReLU
BatchNorm
Output 1 Linear None

The network can be mathematically expressed as a composition of layer transformations:

ŷ = W(4) × ReLU(W(3) × ReLU(W(2) × ReLU(W(1) × x + b(1)) + b(2)) + b(3)) + b(4) [3.24]

where W(l) are the weight matrices for layer l, b(l) are the bias vectors, and ReLU is the
rectified linear unit activation function:

ReLU(z) = max(0, z) [3.25]

The ReLU activation was chosen for hidden layers because it provides efficient gradient
computation, helps mitigate the vanishing gradient problem, and has been shown to work
well in many applications (LeCun et al., 2012).

[Link] Training Configuration

The network was trained using the Adam optimiser (Kingma and Ba, 2015) with the
following configuration:
Optimiser: Adam (Adaptive Moment Estimation) with default parameters (β₁ = 0.9, β₂ =
0.999, ε = 10⁻⁸).
Learning rate: 0.001.
Loss function: Mean Squared Error (MSE):

MSE = (1/n) × Σi=1n(yi − ŷi)2 [3.26]

Batch size: 8 observations per batch (chosen for the small dataset size).
44
Maximum epochs: 500.
Early stopping: Training was terminated if validation loss did not improve for 50
consecutive epochs, with the best weights restored.
Validation split: 20% of training data used for validation during training.
The Keras library with TensorFlow backend was used for implementation.
[Link] Regularization
To prevent overfitting on the small dataset, several regularisation techniques were
employed:
Dropout: A dropout rate of 0.2 was applied after each hidden layer during training,
randomly setting 20% of neuron outputs to zero (Srivastava et al., 2014). This prevents co-
adaptation of neurons and improves generalisation.
Early stopping: As described above, training was terminated when validation performance
stopped improving.
Batch normalisation: Applied after each hidden layer to normalise activations, which can
have a regularising effect and accelerate training (Ioffe and Szegedy, 2015).

3.7 Model Evaluation Metrics

Model performance was evaluated using four complementary metrics that capture different
aspects of prediction accuracy. Table 3.5 summarises the evaluation metrics and their
interpretation.
Table 3. 5: Model Evaluation Metrics Summary

Metric Formula Range Ideal Value

R² 1 − SSres/SStot (−∞, 1] 1
MAE (1/n)Σ|yi − ŷi| [0, ∞) 0
RMSE √[(1/n)Σ(yi − ŷi)²] [0, ∞) 0
MAPE (100/n)Σ|(yi − ŷi)/yi| [0, ∞) 0%

3.7.1 Coefficient of Determination (R²)

The coefficient of determination, denoted R², measures the proportion of variance in the
target variable that is explained by the model. R² is calculated as:

R² = 1 − SSres / SStot [3.27]

where SSres is the residual sum of squares and SStot is the total sum of squares:

SSres = Σi=1n(yi − ŷi)2 [3.28]

45
SStot = Σi=1n(yi − ȳ)2 [3.29]

R² ranges from negative infinity to 1, where 1 indicates perfect prediction, 0 indicates that
the model performs no better than predicting the mean, and negative values indicate worse-
than-mean predictions. In practice, R² values above 0.7 are often considered good for
complex systems, and values above 0.9 indicate excellent predictive performance.

3.7.2 Mean Absolute Error (MAE)

Mean Absolute Error measures the average magnitude of prediction errors without
considering direction:

MAE = (1/n) × Σi=1n|yi − ŷi| [3.30]

MAE has the same units as the target variable (m/hr for ROP), making it directly
interpretable as the average prediction error. MAE is less sensitive to outliers than RMSE
because it does not square the errors.

3.7.3 Root Mean Square Error (RMSE)

Root Mean Square Error provides another measure of prediction accuracy that penalises
larger errors more heavily than MAE:

RMSE = √[(1/n) × Σi=1n(yi − ŷi)2] [3.31]

RMSE is always greater than or equal to MAE, with the difference increasing when there
are large outliers in the predictions. RMSE is mathematically related to the standard
deviation of residuals and represents the expected magnitude of a typical prediction error.

3.7.4 Mean Absolute Percentage Error (MAPE)

Mean Absolute Percentage Error expresses prediction accuracy as a percentage of actual
values:

MAPE = (100/n) × Σi=1n|(yi − ŷi) / yi| [3.32]

MAPE provides a scale-independent measure of accuracy that facilitates comparison across

datasets with different ROP ranges. MAPE values below 10% are generally considered
excellent, while values below 20% indicate good prediction accuracy.

3.8 Parameter Optimization Approach

Following model development and evaluation, the best-performing model was used to
identify optimal drilling parameters that maximise predicted rate of penetration. The
optimization approach employed grid search over the feasible parameter space, which is a
46
systematic and exhaustive search method suitable for low-dimensional optimization
problems (Bergstra and Bengio, 2012).

3.8.1 Optimization Problem Formulation

The optimization problem can be formally stated as:

maximise f(WOB, FLOW, RPM | θ) [3.33]

subject to: WOBmin ≤ WOB ≤ WOBmax [3.34]

FLOWmin ≤ FLOW ≤ FLOWmax [3.35]

RPMmin ≤ RPM ≤ RPMmax [3.36]

where f(WOB, FLOW, RPM | θ) is the trained machine learning model with parameters θ,
and the bounds correspond to the observed ranges in the dataset. The other input variables
(particularly depth parameters) are held at representative values during optimization, as they
represent formation conditions rather than controllable parameters.

3.8.2 Grid Search Implementation

Grid search was implemented by discretising each controllable parameter into a grid of
values spanning the observed range:
WOB: 20 evenly spaced values from minimum to maximum observed WOB.
FLOW: 20 evenly spaced values from minimum to maximum observed flow rate.
RPM: 20 evenly spaced values from minimum to maximum observed RPM.
The total number of grid points evaluated was:

Ngrid = 20 × 20 × 20 = 8,000 combinations [3.37]

For each grid point, the model prediction was computed with controllable parameters set to
the grid values and other parameters set to their median values from the training data. The
optimal parameters were identified as those yielding the maximum predicted ROP.

3.8.3 Multi-Model Optimization

The optimization was performed using each of the three trained models (Random Forest,
XGBoost, ANN) to assess the robustness of optimization results. Convergence of multiple
models to similar optimal parameters would provide confidence that the identified optimum
represents a genuine global optimum rather than an artefact of a specific modelling
approach.

47
3.9 Validation Methods
Validation of the optimization results was conducted through three complementary
approaches: sensitivity analysis, comparison with drilling mechanics principles, and cross-
model validation.

3.9.1 Sensitivity Analysis

Sensitivity analysis examines how the model's predicted ROP changes as individual
controllable parameters vary while other parameters are held constant at their optimal
values. For each controllable parameter xj, a sensitivity curve was generated by computing
predictions over a range of xj values:

Sensitivity(xj) = {f(x*1, ..., xj, ..., x*p) : xj ∈ [xj,min, xj,max]} [3.38]

where x*k denotes the optimal value for parameters other than j. The sensitivity curves
reveal the shape of the response surface near the optimum and indicate how much ROP
would change if parameters deviate from optimal values.
3.9.2 Drilling Mechanics Validation
The sensitivity analysis results were compared with theoretical expectations from drilling
mechanics to assess physical plausibility:

Weight on Bit: The relationship between WOB and ROP should exhibit threshold
behaviour, with minimal drilling below a threshold value, rapid increase as WOB exceeds
the threshold, and potential plateau or decrease at very high WOB due to founder point
effects (Maurer, 1962; Dupriest and Koederitz, 2005).
Rotary Speed: ROP should generally increase with RPM following a power law
relationship, with the exponent typically between 0.4 and 1.0 depending on formation and
bit characteristics (Bingham, 1965).
Flow Rate: ROP should increase with flow rate due to improved hole cleaning, with
potential plateau when cleaning is adequate and additional flow provides diminishing
returns (Bourgoyne et al., 1986).
Consistency between model behaviour and these theoretical expectations provides evidence
that the models have learned physically meaningful relationships rather than spurious
patterns in the training data.

3.9.3 Cross-Model Validation

48
The convergence of optimization results across the three different modelling approaches
provides an additional form of validation. If all three models, despite their different
architectures and learning mechanisms, identify similar optimal parameters, this provides
strong evidence that the identified optimum is robust and not an artefact of any single
modelling approach.
The degree of convergence was assessed by comparing the optimal parameter values
identified by each model:

ΔWOB = max(WOB*RF, WOB*XGB, WOB*ANN) − min(WOB*RF, WOB*XGB, WOB*ANN) [3.39]

with similar calculations for FLOW and RPM. Small values of these differences indicate
strong convergence.

3.10 Software and Computational Environment

All analyses were conducted using Python 3.9 with the following key libraries:
Pandas (version 1.4.2): Data manipulation and preprocessing.
NumPy (version 1.22.3): Numerical computations.
Scikit-learn (version 1.0.2): Random Forest implementation, preprocessing, and
evaluation metrics (Pedregosa et al., 2011).
XGBoost (version 1.5.2): XGBoost implementation (Chen and Guestrin, 2016).
TensorFlow (version 2.8.0) and Keras: Neural network implementation.
SHAP (version 0.40.0): SHAP value calculation (Lundberg and Lee, 2017).
Matplotlib (version 3.5.1) and Seaborn (version 0.11.2): Visualisation.
Computations were performed on a standard workstation with 16 GB RAM and an Intel
Core i7 processor. Training times ranged from a few seconds for Random Forest and
XGBoost to approximately one minute for the neural network.

3.11 Ethical Considerations

This study uses secondary data from operational drilling records that do not contain
personally identifiable information. The data were used solely for research purposes
consistent with academic research ethics. No ethical approval was required as the study does
not involve human subjects or animal research.

49
 CHAPTER FOUR: RESULTS AND DISCUSSION

4.1 Introduction
This chapter presents and discusses the results obtained from applying the methodology
described in Chapter Three to the drilling dataset. The results are organised according to the
four research objectives established in Chapter One: feature analysis to identify key
parameters influencing rate of penetration, comparison of machine learning models for ROP
prediction, optimization of drilling parameters to maximise ROP, and validation of
optimization recommendations through sensitivity analysis and drilling mechanics
principles.

The presentation follows a systematic structure that first establishes the relationships
between drilling parameters and ROP through multiple analytical methods, then evaluates
the predictive performance of three machine learning algorithms, identifies optimal drilling
parameters through grid search optimization, and finally validates the results through
sensitivity analysis and comparison with established drilling theory. Each section includes
interpretation of findings in the context of drilling mechanics literature and comparison with
results reported in recent studies.

The analysis was conducted using the 43-observation dataset described in Chapter Three,
with an 80/20 train-test split resulting in 34 training observations and 9 test observations.
All preprocessing, model training, and evaluation followed the procedures specified in the
methodology chapter to ensure reproducibility and valid comparison across modelling
approaches.

4.2 Feature Analysis Results

The feature analysis phase employed three complementary methods to identify the drilling
parameters most strongly influencing rate of penetration: Pearson correlation analysis,
Random Forest feature importance, and SHAP (SHapley Additive exPlanations) analysis.
The use of multiple methods with different theoretical foundations enables validation of
findings across approaches and identification of both linear and non-linear relationships that
might be missed by any single method.

4.2.1 Correlation Analysis

Correlation analysis was conducted to examine the linear relationships between all variables
in the dataset and to identify potential multicollinearity among input features. Figure 4.1
presents the correlation heatmap showing the Pearson correlation coefficients among the
key drilling parameters and ROP.
50
 Figure 4. 1: Correlation Heatmap Including ROP and Key Drilling Parameters
The correlation analysis revealed several important patterns in the relationships among
drilling parameters. Examining the correlations with the target variable ROP, rotary speed
(RPM) exhibited the strongest positive correlation (r = 0.81), indicating a strong linear
relationship whereby higher rotary speeds are associated with higher rates of penetration.
This finding is consistent with the theoretical models of Maurer (1962) and Bourgoyne and
Young (1974), which predict positive relationships between RPM and ROP, as well as
empirical observations from numerous field studies (Armenta, 2008).

The depth-related parameters (MD, MD_IN, MD_OUT) showed strong positive

correlations with ROP (r = 0.76), suggesting that deeper intervals in this section were
associated with higher drilling rates. This pattern may reflect changes in formation
properties with depth, such as decreasing rock strength or changes in pore pressure regime,
that favour higher penetration rates. The strong correlations among depth parameters
themselves (r = 1.00) indicate perfect multicollinearity, which is expected since MD_IN and
MD_OUT define the interval boundaries that determine MD.

Flow rate (FLOW) and jet impact force (JET_IMPACT) exhibited moderate positive
correlations with ROP (r = 0.55), consistent with the hydraulic effects described in Warren's
(1984) imperfect cleaning model and the jet impact function in the Bourgoyne and Young
(1974) model. The perfect correlation between FLOW and JET_IMPACT (r = 1.00) reflects

51
the mathematical relationship whereby jet impact force is calculated directly from flow rate
and mud properties.

Weight on bit (WOB) showed a weak negative correlation with ROP (r = -0.02), which at
first appears counterintuitive given the established positive relationship between WOB and
ROP in drilling theory. However, this weak linear correlation does not imply that WOB is
unimportant for ROP; rather, it suggests that the relationship between WOB and ROP in
this dataset is non-linear and cannot be adequately captured by the Pearson correlation
coefficient, which measures only linear association. As will be shown in subsequent
sections, the machine learning models identify substantial importance for WOB through
their ability to capture non-linear relationships.

Torque (TORQUE) showed a moderate positive correlation with ROP (r = 0.45), reflecting
the general tendency for higher torque to accompany more aggressive drilling with higher
penetration rates. However, torque is largely a response variable rather than a controllable
input, as it results from the combination of formation resistance, WOB, and RPM rather
than being directly set by the driller.

4.2.2 Random Forest Feature Importance

Random Forest feature importance provides a measure of each variable's contribution to
predictive accuracy that captures both linear and non-linear relationships. The importance
is calculated as the mean decrease in impurity (MDI) across all trees in the forest,
representing how much each feature contributes to reducing prediction variance when used
for splitting. Figure 4.2 presents the feature importance rankings from the trained Random
Forest model.

XGBoost Feature Importance Comparison

To validate the feature importance rankings obtained from Random Forest, XGBoost feature
importance was also extracted and compared. XGBoost calculates feature importance based
on the average gain, which represents the improvement in accuracy contributed by each
feature across all splits where it is used (Chen and Guestrin, 2016). Table 4.1 presents a
comparison of feature importance rankings between Random Forest and XGBoost.

52
Table 4. 1: Comparison of Feature Importance Rankings Between Random Forest
and XGBoost
Parameter RF Importance (%) XGBoost Importance (%) RF Rank XGBoost Rank
MD_IN 25.2 0.0 1 7
MD 22.7 83.6 2 1
MD_OUT 20.8 0.0 3 7
RPM 18.2 2.6 4 3
WOB 6.6 11.4 5 2
TORQUE 3.7 0.5 6 5
JET_IMPACT 1.5 2.0 7 4
FLOW 1.4 0.0 8 7

The XGBoost feature importance analysis reveals both similarities and instructive
differences compared to the Random Forest results. Both algorithms identify measured
depth (MD) as the most influential predictor, confirming the dominant role of formation
characteristics in determining ROP. However, the distribution of importance differs
substantially between the two approaches. Random Forest distributes importance relatively
evenly across the three depth-related parameters (MD_IN at 25.2%, MD at 22.7%, and
MD_OUT at 20.8%), collectively accounting for 68.7% of total importance. In contrast,
XGBoost concentrates 83.6% of its importance on MD alone, with negligible contributions
from MD_IN and MD_OUT.

This difference reflects the distinct ensemble mechanisms of the two algorithms. Random
Forest's bagging approach with random feature selection encourages the use of correlated
features across different trees, resulting in distributed importance among the highly
correlated depth parameters. XGBoost's sequential boosting mechanism, conversely, tends
to concentrate importance on a single representative feature from a group of correlated
variables, as subsequent trees have diminished need to use redundant information once the
primary feature has been incorporated. Importantly, both algorithms arrive at the same
fundamental conclusion: depth-related parameters, serving as proxies for formation
properties, dominate ROP prediction.

Among the controllable drilling parameters, both models identify WOB and RPM as the
most influential. Random Forest ranks RPM fourth (18.2%) and WOB fifth (6.6%), whilst
XGBoost ranks WOB second (11.4%) and RPM third (2.6%). The elevated importance of
WOB in XGBoost relative to Random Forest may reflect XGBoost's sensitivity to the non-
linear threshold behaviour of WOB identified in the SHAP analysis. Flow rate (FLOW)
shows minimal importance in both models, suggesting that hydraulic conditions were
generally adequate throughout the drilling section and did not represent a limiting factor for

53
ROP. The convergence of both algorithms on the relative unimportance of hydraulic
parameters strengthens confidence in this interpretation.

Figure 4. 2: Random Forest Feature Importance Rankings

The feature importance analysis revealed a clear hierarchy of predictor variables. Measured
depth (MD) emerged as the most important predictor with an importance score of 0.284,
indicating that approximately 28.4 percent of the model's predictive capability derives from
this single feature. The dominance of depth as a predictor reflects its role as a proxy for
formation properties, which are the primary determinants of drillability as established in the
theoretical literature (Maurer, 1962; Bourgoyne and Young, 1974). In the absence of direct
formation characterisation data such as rock strength or porosity, depth captures the
systematic variation in formation properties that occurs with burial depth due to compaction
and diagenetic processes.

The other depth-related parameters also ranked highly, with MD_OUT achieving an
importance score of 0.205 (20.5 percent) and MD_IN scoring 0.180 (18.0 percent).
Collectively, the three depth-related parameters account for approximately 67 percent of
total feature importance, underscoring the dominant role of formation characteristics in
determining drilling performance. This finding has important implications for model
transferability, as the depth-formation property relationship is specific to each geological
setting and may not transfer directly to other drilling environments.

Among the controllable drilling parameters, rotary speed (RPM) ranked third with an
importance score of 0.198 (19.8 percent), confirming its strong influence on ROP that was

54
also evident in the correlation analysis. The consistency between correlation analysis (r =
0.81) and feature importance (19.8 percent) for RPM suggests that the relationship between
RPM and ROP is predominantly linear within the observed data range, allowing both
methods to capture its predictive value.

Weight on bit (WOB) ranked fifth with an importance score of 0.077 (7.7 percent),
substantially lower than RPM but still representing a meaningful contribution to prediction
accuracy. The contrast between WOB's weak correlation with ROP (r = -0.02) and its non-
trivial feature importance (7.7 percent) confirms that the WOB-ROP relationship is non-
linear in nature. The Random Forest algorithm, through its recursive partitioning, can
capture threshold effects and other non-linear patterns that correlation analysis cannot
detect. This finding validates the importance of using multiple analytical methods, as
reliance on correlation analysis alone would substantially underestimate WOB's predictive
relevance.

Torque (TORQUE) achieved an importance score of 0.030 (3.0 percent), while

JET_IMPACT and FLOW scored 0.015 (1.5 percent) and 0.014 (1.4 percent) respectively.
The relatively low importance of hydraulic parameters suggests that hydraulic conditions
were generally adequate throughout the section and did not represent a limiting factor for
drilling performance. This interpretation is consistent with Warren's (1984) observation that
hydraulic effects on ROP plateau once adequate hole cleaning is achieved.

4.2.3 SHAP Analysis

SHAP (SHapley Additive exPlanations) analysis provides a theoretically grounded
approach to understanding how individual features contribute to each prediction, based on
concepts from cooperative game theory (Lundberg and Lee, 2017). Unlike global
importance measures, SHAP values reveal both the magnitude and direction of each
feature's effect, as well as how effects vary across the range of feature values. Figure 4.3
presents the SHAP beeswarm plot showing the distribution of SHAP values for each feature
across all observations.

55
Figure 4. 3: SHAP Beeswarm Plot Showing Feature Impact on Model Output
In the SHAP beeswarm plot, each point represents one observation, with horizontal position
indicating the SHAP value (impact on model output) and colour indicating the feature value
(red for high values, blue for low values). Features are ordered by mean absolute SHAP
value, with the most important features at the top.

Measured depth (MD) exhibits the largest spread of SHAP values, ranging from
approximately -4 to +1, confirming its dominant role in the model. The colour pattern
reveals a clear positive relationship: high MD values (red points) are associated with
positive SHAP values, indicating that greater depths push predictions toward higher ROP.
Conversely, low MD values (blue points) are associated with negative SHAP values,
indicating predictions of lower ROP at shallower depths. The mean absolute SHAP value
for MD is approximately 1.532, substantially higher than other features.

The depth parameters MD_OUT and MD_IN show similar patterns to MD, with high values
associated with positive SHAP values. The consistency across depth-related features
reinforces the interpretation that depth serves as a proxy for formation properties that
become more favourable for drilling at greater depths in this particular section.

For rotary speed (RPM), the SHAP analysis reveals a consistent positive effect: high RPM
values (red points) cluster on the positive side of zero, indicating that higher rotary speeds
increase predicted ROP. The SHAP values for RPM show relatively tight clustering in the
range of 0 to +1, indicating a consistent but moderate effect magnitude compared to depth

56
parameters. This pattern aligns with the theoretical expectation from Bourgoyne and Young
(1974) that ROP increases with RPM raised to a positive exponent.

Weight on bit (WOB) exhibits a more complex pattern in the SHAP analysis that illuminates
the non-linear relationship suggested by the discrepancy between correlation and feature
importance. The beeswarm plot shows that both high (red) and low (blue) WOB values can
appear on both sides of zero, with the pattern depending on interactions with other variables.
However, there is a tendency for very low WOB values to be associated with negative SHAP
values, consistent with the threshold behaviour described in drilling theory (Bourgoyne and
Young, 1974). Below a certain threshold, inadequate weight results in poor bit engagement
with the formation and reduced ROP. Above the threshold, the relationship becomes more
complex and may depend on other factors such as formation properties and hole cleaning
efficiency.

The SHAP values for TORQUE, FLOW, and JET_IMPACT show tight clustering near zero,
indicating that these features have relatively small and consistent effects on predictions.
This pattern is consistent with their low feature importance scores and suggests that they
provide limited additional predictive information beyond what is captured by depth and
RPM.

4.2.4 Summary of Feature Analysis

The multi-method feature analysis provides a comprehensive understanding of the factors
driving ROP in the studied drilling section. Table 4.2 summarises the key findings from
each analytical method.

Table 4. 2: Summary of Feature Analysis Results

Parameter Correlation RF XGBoost SHAP Interpretation
(r) Importance Importance
MD 0.76 22.7% 83.6% Strong positive; dominant
predictor
MD_IN 0.76 25.2% 0.0% Strong positive; depth proxy
MD_OUT 0.76 20.8% 0.0% Strong positive; depth proxy
RPM 0.81 18.2% 2.6% Consistent positive effect
WOB -0.02 6.6% 11.4% Non-linear; threshold behaviour
TORQUE 0.45 3.7% 0.5% Moderate positive; response
variable
FLOW 0.55 1.4% 0.0% Limited impact; adequate
cleaning

The consistency of findings across the three analytical methods reinforces the reliability of
the identified patterns. Depth-related parameters dominate predictive importance in both

57
Random Forest and XGBoost, collectively accounting for approximately 67 percent in
Random Forest and over 83 percent in XGBoost when considering MD as a proxy for all
depth variables. RPM demonstrates strong influence through both its high linear correlation
with ROP (r = 0.81) and its substantial feature importance rankings in both models. The
seemingly contradictory behaviour of WOB, which shows negligible correlation yet
meaningful importance, is explained by the SHAP analysis, which reveals a non-linear,
threshold-dependent relationship between WOB and ROP that linear correlation cannot
capture.

Figure 4. 4: Machine learning model development and evaluation workflow illustrating

the training, validation, and comparison process for Random Forest, XGBoost, and ANN
models
4.3 Machine Learning Model Comparison
Three machine learning models were developed and evaluated for ROP prediction: Random
Forest, XGBoost, and Artificial Neural Network. This section presents the performance
comparison and discusses the factors contributing to observed differences.

4.3.1 Model Performance Metrics

58
Table 4.3 presents the performance metrics for each model evaluated on the held-out test
set of 9 observations. The metrics include coefficient of determination (R²), mean absolute
error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE).

Table 4. 3: Machine Learning Model Performance Comparison

Model R² MAE (m/hr) RMSE (m/hr) MAPE (%)
Random Forest 0.904 1.74 2.18 9.2
XGBoost 0.813 2.13 2.89 11.8
ANN 0.751 2.73 3.34 15.1

Random Forest achieved the highest performance across all metrics, with R² of 0.904
indicating that the model explains approximately 90.4 percent of the variance in ROP on the
test set. The MAE of 1.74 m/hr indicates that predictions deviate from actual values by an
average of 1.74 metres per hour, while the RMSE of 2.18 m/hr reflects somewhat larger
errors when accounting for the squared penalty on larger deviations. The MAPE of 9.2
percent indicates that predictions are within approximately 9 percent of actual values on
average, which represents excellent accuracy for practical drilling applications.

XGBoost achieved intermediate performance with R² of 0.813, MAE of 2.13 m/hr, RMSE
of 2.89 m/hr, and MAPE of 11.8 percent. While these results represent good predictive
accuracy, the approximately 10 percent reduction in R² compared to Random Forest is
notable and suggests that the boosting approach may be less suitable for this particular
dataset.

The Artificial Neural Network achieved the lowest performance with R² of 0.751, MAE of
2.73 m/hr, RMSE of 3.34 m/hr, and MAPE of 15.1 percent. The 15 percent reduction in R²
compared to Random Forest reflects the challenges of training neural networks with limited
data, as discussed below.

4.3.2 Analysis of Performance Differences

The superior performance of Random Forest on this dataset can be attributed to several
factors related to algorithm characteristics and dataset properties. First, Random Forest
employs bagging (bootstrap aggregating) combined with random feature selection at each
split, which creates diversity among trees and provides inherent regularisation against
overfitting (Breiman, 2001). The ensemble averaging reduces variance without substantially
increasing bias, which is particularly valuable when training data are limited. With only 34
training observations, the risk of overfitting is substantial, and Random Forest's variance-
reduction mechanism provides effective protection.

59
Second, the Random Forest algorithm makes no assumptions about the functional form of
relationships between features and the target variable. The recursive partitioning approach
can naturally capture non-linear relationships, threshold effects, and interactions without
requiring explicit specification. This flexibility is advantageous for modelling the complex
relationships between drilling parameters and ROP, particularly the threshold behaviour
identified for WOB.

The reduced performance of XGBoost, despite its reputation as a highly effective algorithm
for many applications (Chen and Guestrin, 2016), likely reflects its sequential boosting
approach. Each tree in an XGBoost ensemble is fit to the residuals of the previous trees,
attempting to correct errors. While this approach is powerful for capturing complex patterns,
it can be susceptible to overfitting when training data are limited, as the algorithm may learn
to correct noise rather than signal in later iterations. The regularisation parameters (lambda
and gamma) provide some protection, but may not fully compensate for the fundamental
limitation of sequential error correction with limited data.

The Artificial Neural Network's lower performance is consistent with the well-documented
challenges of training deep learning models with small datasets (Goodfellow et al., 2016).
Despite regularisation measures including dropout, early stopping, and batch normalisation,
the network's large number of parameters (over 3,000 in the 64-32-16 architecture) relative
to the 34 training observations creates a highly underdetermined learning problem. Neural
networks generally require substantially larger datasets to achieve their full potential, with
typical recommendations ranging from hundreds to thousands of observations depending on
problem complexity.

4.3.3 Comparison with Literature

The Random Forest R² of 0.904 achieved in this study compares favourably with results
reported in recent literature on machine learning for ROP prediction, particularly
considering the limited dataset size. Mohammadinia et al. (2025) achieved R² of 0.925 using
least squares support vector machines combined with the crow search algorithm for
hyperparameter optimization, though their study utilised a substantially larger dataset from
the Fahliyan Formation in Iran. Zou et al. (2025) reported R² of 0.955 using Random Forest
with SHAP interpretation applied to drilling data from southeastern Iraq, with their higher
accuracy likely attributable to a more comprehensive feature set including explicit formation
parameters.

60
Ma et al. (2025) achieved R² of 0.92 with RMSE of 1.92 m/hr using an interpretable
automated machine learning workflow that incorporated feature engineering and model
selection optimization. Osman et al. (2021) found that Random Forest consistently
outperformed both ANN and SVR for ROP prediction in carbonate formations, with R²
values ranging from 0.85 to 0.92 depending on the specific dataset. These comparisons
suggest that the performance achieved in the current study is competitive with state-of-the-
art results and validates the effectiveness of Random Forest for ROP prediction even with
limited data.

The finding that Random Forest outperforms XGBoost and ANN on this small dataset aligns
with observations from other studies. Hastie et al. (2009) noted that bagging methods like
Random Forest tend to be more robust to overfitting than boosting methods when data are
limited. Breiman (2001) demonstrated that Random Forest achieves low generalisation error
through variance reduction rather than bias reduction, making it particularly suitable for
high-variance settings with limited training data.

4.4 Optimization Results

Following model development and evaluation, each trained model was used to identify
optimal drilling parameters through grid search over the feasible parameter space. This
section presents the optimization results and analyses the convergence across models.

4.4.1 Optimal Parameter Identification

Grid search optimization was conducted by evaluating each model's predictions across
8,000 parameter combinations (20 values each for WOB, FLOW, and RPM spanning the
observed data ranges). Table 4.3 presents the optimal parameters identified by each model
along with the corresponding predicted maximum ROP.

Table 4. 4: Optimal Drilling Parameters Identified by Each Model

Model WOB (klbs) Flow (GPM) RPM Max ROP (m/hr)
Random Forest 158.0 6.55 2370.05 21.96
XGBoost 158.0 6.55 2370.05 20.66
ANN 158.0 6.55 2370.05 21.43

A remarkable finding emerged from the optimization analysis: all three machine learning
models, despite their fundamentally different architectures and learning mechanisms,
converged to identical optimal parameter values. The optimal parameters identified were
WOB = 158.0 klbs, Flow = 6.55 GPM, and RPM = 2370.05, corresponding to the upper

61
bounds of the observed data ranges for WOB and FLOW, and near the upper bound for
RPM.

4.4.2 Significance of Model Convergence

The convergence of all three models to identical optimal parameters is a significant finding
with important implications for confidence in the optimization results. Random Forest
achieves predictions through averaging across 100 independently trained decision trees,
each using bootstrap samples and random feature subsets. XGBoost builds predictions
through sequential boosting of gradient-based corrections. The neural network learns
through gradient descent optimization of connection weights across multiple layers. These
fundamentally different mathematical mechanisms make independent convergence to the
same optimum highly unlikely to occur by chance.

The probability of three independent optimization processes identifying exactly the same
parameter values due to random variation or modelling artefacts is extremely low. This
convergence therefore provides strong evidence that the identified optimal parameters
represent a genuine global optimum in the drilling response surface, at least within the
observed data range. The finding suggests that the underlying physics of the drilling process,
rather than model-specific biases, drives the optimization to this parameter combination.

The small differences in predicted maximum ROP among models (20.66 to 21.96 m/hr)
reflect differences in model prediction accuracy rather than different optima. Random
Forest, with its higher R² of 0.904, predicts a maximum ROP of 21.96 m/hr at the optimal
parameters. XGBoost, with R² of 0.813, predicts 20.66 m/hr, while ANN predicts 21.43
m/hr. These differences are consistent with the relative accuracy of the models observed on
the test set and do not indicate uncertainty about the optimal parameter values themselves.

4.4.3 Interpretation of Optimal Values

The optimal parameter values correspond to the upper bounds of the observed data ranges,
suggesting that more aggressive drilling parameters generally improve performance within
the operational constraints of the studied section. This finding is consistent with theoretical
expectations for drilling in formations where equipment limitations rather than formation
properties represent the binding constraint on drilling intensity.

The optimal WOB of 158.0 klbs represents the maximum weight applied during the section,
suggesting that even higher weight might further improve ROP if equipment and operational
constraints permitted. The optimal flow rate of 6.55 GPM corresponds to the maximum flow
used, indicating adequate or superior hole cleaning at this rate. The optimal RPM of 2370.05
62
is at the upper end of the operational range, consistent with the strong positive relationship
between RPM and ROP identified in the feature analysis.

It is important to note that these optimal values are bounded by the observed data range and
should not be extrapolated beyond. The models have no information about drilling
behaviour at parameters outside the training range, and extrapolation could lead to
unreliable predictions or recommend parameters that cause operational problems such as bit
damage, excessive vibration, or hole stability issues.

4.5 Sensitivity Analysis and Validation

Sensitivity analysis was conducted to examine how predicted ROP responds to changes in
controllable parameters and to validate that model behaviour aligns with established drilling
mechanics principles. The analysis used the Random Forest model given its superior
predictive performance.

4.5.1 Sensitivity to Weight on Bit

Figure 4.5 presents the sensitivity of predicted ROP to weight on bit, with flow rate and
RPM held constant at their optimal values (6.55 GPM and 2370.05 respectively) and depth
parameters held at median values.

Figure 4. 5: Sensitivity of Predicted ROP to Weight on Bit (Flow and RPM at optimal
values)
The sensitivity curve reveals a distinctive threshold behaviour that is highly consistent with
established drilling mechanics theory. At low WOB values (approximately 1 klb), the
predicted ROP is approximately 21.72 m/hr. As WOB increases beyond a threshold of
63
approximately 2 klbs, the predicted ROP increases rapidly to approximately 21.94 m/hr.
Above this threshold, further increases in WOB produce minimal additional improvement
in ROP, with the curve remaining essentially flat through the maximum WOB of 10 klbs
shown.
This threshold behaviour aligns closely with the founder point concept described by
Bourgoyne and Young (1974). The founder point represents the WOB threshold below
which the bit does not adequately engage with the formation, resulting in inefficient drilling
characterised by sliding and grinding rather than effective rock removal. Below this
threshold, increases in WOB primarily increase the contact stress without proportionally
increasing the depth of cut, as the bit cutters do not penetrate deeply enough to generate
efficient rock fracture.

Above the founder point, adequate bit engagement is achieved, and the rock removal
mechanism operates efficiently. Further increases in WOB provide diminishing returns
because the rate of penetration becomes limited by other factors such as hole cleaning
efficiency, formation strength, or the rate at which cuttings can be transported away from
the bit face. The flat response above the threshold suggests that within the studied
conditions, hole cleaning and formation characteristics rather than applied weight represent
the binding constraints on ROP.

The validation of threshold behaviour in the sensitivity analysis provides important evidence
that the Random Forest model has learned physically meaningful relationships rather than
spurious correlations. The model's ability to capture this non-linear, theoretically expected
behaviour despite receiving no explicit guidance about drilling mechanics strengthens
confidence in its predictions and optimization recommendations.

4.5.2 Sensitivity to Rotary Speed

Figure 4.6 presents the sensitivity of predicted ROP to rotary speed, with WOB and flow
rate held constant at their optimal values and depth parameters at median values.

64
Figure 4. 6: Sensitivity of Predicted ROP to Rotary Speed (WOB and Flow at optimal
values)

The sensitivity analysis for RPM reveals remarkably stable predicted ROP across the
operational range of 90 to 160 RPM, with values varying only between approximately 21.94
and 22.0 m/hr. This near-flat response may initially seem surprising given the strong
positive correlation between RPM and ROP (r = 0.81) identified in the feature analysis and
the theoretical expectation of positive RPM effects from the Bourgoyne and Young (1974)
model.
Several factors may explain this apparent discrepancy. First, the sensitivity analysis holds
other parameters constant at their optimal values, while the correlation analysis examined
RPM-ROP relationships across all observations with varying values of other parameters.
The strong correlation may partly reflect covariation between RPM and other parameters
(particularly depth) that is controlled away in the sensitivity analysis.

Second, the observed RPM range (90-160) may fall within a regime where formation
characteristics and other factors dominate the prediction, with RPM effects being secondary.
The Random Forest model, having learned that depth parameters account for approximately
67 percent of predictive importance, may predict similar ROP values across RPM variations
when depth and other formation proxies are held constant.

Third, the stability of predicted ROP across RPM may reflect adequate bit performance
throughout the tested range. For PDC bits operating within their design envelope, the RPM

65
range of 90-160 may fall within a stable performance regime where neither insufficient nor
excessive rotation limits efficiency. Armenta (2008) noted that PDC bit performance is
optimal within a specific RPM range that depends on formation hardness and bit design.

From a practical standpoint, the RPM insensitivity finding has valuable operational
implications. It suggests that drillers have flexibility to adjust RPM within the 90-160 range
for purposes other than ROP optimization, such as vibration management, directional
control, or bit wear management, without significantly compromising drilling efficiency.
This operational flexibility can be leveraged to optimise overall drilling performance
considering multiple objectives rather than ROP alone.

4.5.3 Summary of Validation Findings

The sensitivity analysis provides important validation that the Random Forest model has
learned physically plausible relationships between drilling parameters and ROP. The WOB
sensitivity curve exhibits threshold behaviour consistent with the founder point concept
from drilling mechanics theory (Bourgoyne and Young, 1974; Maurer, 1962). The RPM
sensitivity analysis, while showing unexpectedly flat response, reveals operational
flexibility that has practical value for drilling operations.

The combination of accurate prediction (R² = 0.904), convergence across models in

optimization, and physically plausible sensitivity behaviour provides confidence that the
model captures genuine patterns in the drilling response rather than spurious correlations or
overfitting artefacts. This validation is essential for supporting operational implementation
of the optimization recommendations.

4.6 Discussion
The results presented in this chapter demonstrate the effectiveness of machine learning
approaches for ROP prediction and optimization, while also revealing insights about the
factors driving drilling performance in the studied section.

4.6.1 Dominance of Formation Effects

The consistent finding across all feature analysis methods that depth-related parameters
account for approximately 67 percent of feature importance has important implications for
understanding drilling performance. This dominance reflects the fundamental role of
formation properties, which vary systematically with depth, in determining drillability. The
formation strength function f₁ in the Bourgoyne and Young (1974) model is typically the
largest contributor to ROP variation, and the current results are consistent with this
theoretical expectation.
66
The implication for drilling optimization is that formation characterisation, even through
indirect proxies like depth, is essential for accurate ROP prediction. Models that do not
account for formation effects will have limited predictive accuracy regardless of how well
they capture the effects of controllable parameters. For practical applications, this suggests
that depth-specific or formation-specific models may outperform universal models that
attempt to predict ROP across diverse conditions.

4.6.2 Algorithm Selection for Small Datasets

The superior performance of Random Forest over XGBoost and ANN on this 43-
observation dataset provides guidance for algorithm selection in drilling applications where
data availability is limited. The finding aligns with theoretical understanding of the bias-
variance trade-off: Random Forest's bagging approach reduces variance through averaging,
while XGBoost's boosting approach can amplify noise when data are limited, and neural
networks require substantial data to regularise their large parameter spaces effectively.

For practitioners working with small drilling datasets, the results suggest that Random
Forest should be the default choice, with consideration of boosting or neural network
approaches only when substantially larger datasets (hundreds to thousands of observations)
are available. This guidance is consistent with recommendations from Osman et al. (2021)
and observations from the broader machine learning literature (Hastie et al., 2009).

4.6.3 Robustness Through Multi-Model Convergence

The convergence of all three models to identical optimal parameters provides a novel form
of validation that extends beyond traditional accuracy metrics. This multi-model agreement
suggests that the identified optimum reflects genuine patterns in the drilling response rather
than artefacts of any single modelling approach. The approach of using multiple algorithms
and requiring convergence for confidence could be adopted more broadly in drilling
optimization applications as a safeguard against model-specific biases or overfitting.

4.6.4 Physical Plausibility as Validation

The validation of model behaviour against drilling mechanics theory, particularly the WOB
threshold behaviour consistent with the founder point concept, demonstrates an approach
for bridging data-driven and physics-based modelling. Machine learning models are often
criticised as black boxes that may learn spurious patterns, but the current results show that
appropriately validated models can capture physically meaningful relationships. This
validation approach, using sensitivity analysis to compare model behaviour with theoretical

67
expectations, could be adopted as standard practice for drilling applications where model
reliability is paramount.

4.6.5 Comparison with Field Implementation Results

The optimal parameters identified in this study align with the operational philosophy
demonstrated in recent field implementations of AI-based drilling optimization. Al-Riyami
et al. (2023) reported that following machine learning recommendations in a Malaysian test
well yielded ROP increases from 10.4 to 31 m/h, representing a threefold improvement.
Singh et al. (2021) documented consistent improvements of 19 to 33 percent in field trials
across multiple wells in the Permian Basin. The convergence of all three models in the
present study to identical optimal parameters (WOB = 158 klbs, Flow = 6.55 GPM, RPM =
2370) suggests that similar benefits could potentially be achieved through systematic
application of the identified drilling parameter combinations.
Furthermore, the feature importance findings of this study, which identify depth-related
formation properties as the dominant influence on ROP, are consistent with observations
from field implementations. Boukredera et al. (2023) noted that their optimization system
This suggests that the practical value of machine learning optimization is maximised when
models correctly capture the formation-specific drilling response, as demonstrated in the
present study through the high predictive accuracy of the Random Forest model (R² =
0.904). However, field validation would be required to confirm the transferability of these
findings to achieved the greatest improvements in formations where the AI model had
accurately characterised the formation-ROP relationship to operational drilling
environments.

68
 CHAPTER FIVE: CONCLUSIONS AND RECOMMENDATIONS

5.1 Introduction
This chapter presents the conclusions and recommendations derived from the
research findings on machine learning-based rate-of-penetration prediction and
optimization. The chapter synthesises the study's key findings, draws conclusions aligned
with each research objective, offers practical recommendations for drilling operations and
future research, discusses the theoretical and practical contributions of the study,
acknowledges its limitations, and suggests directions for future investigation. The
conclusions are drawn from the systematic application of the methodology described in
Chapter Three and the results presented in Chapter Four.
The research set out to address four specific objectives: identifying key drilling
parameters influencing ROP through multiple analytical methods, developing and
comparing machine learning models for ROP prediction, determining optimal drilling
parameters using the best-performing model, and validating optimization recommendations
through sensitivity analysis and drilling mechanics principles. The following sections
present the findings and conclusions for each objective, demonstrating how the research has
contributed to both academic knowledge and practical applications in drilling optimization.
The importance of this research lies in its contribution to the growing body of
literature on machine learning applications in drilling engineering, an area that has seen
significant growth in recent years as the industry seeks to leverage data-driven approaches
to improve performance (Barbosa et al., 2019; Najjarpour et al., 2022). By systematically
comparing algorithms, validating results against established theory, and providing
interpretable insights, this study addresses key gaps identified in the literature review and
offers a framework that can be applied to other drilling optimization problems.

5.2 Summary of Key Findings

This section provides a comprehensive summary of the key findings from each phase of the
research, establishing the foundation for the conclusions and recommendations that follow.
5.2.1 Dataset and Scope
The research utilised drilling data comprising 43 observations across 31 variables from a
single drilling section using a polycrystalline diamond compact (PDC) bit with a diameter
of 8.5 inches. This dataset represents a typical small drilling dataset encountered in many
practical applications where comprehensive data collection is constrained by operational
and economic factors. The focus on a single bit type and drilling section provided controlled
69
conditions for evaluating the relationships between drilling parameters and rate of
penetration without confounding effects from bit changes or major formation transitions.
The dataset included controllable parameters (WOB, RPM, FLOW, torque), depth-related
parameters (MD, MD_IN, MD_OUT, TVD), hydraulic parameters, and derived parameters
including mechanical specific energy and d-exponent. The rate of penetration varied from
approximately 8 m/hr to 28 m/hr across the section, providing sufficient variation for model
development and evaluation.
5.2.2 Feature Analysis Findings
The multi-method feature analysis revealed consistent patterns across correlation analysis,
Random Forest feature importance, and SHAP analysis. Depth-related parameters emerged
as the dominant predictors of ROP, collectively accounting for approximately 67% of the
total feature importance in the Random Forest model. Measured depth (MD) alone
contributed 28.4 percent of feature importance, with MD_OUT (20.5 percent) and MD_IN
(18.0 percent) also ranking highly. This finding is consistent with the theoretical
understanding that formation properties, which vary systematically with depth due to
compaction and diagenesis, are primary determinants of drilling performance (Bourgoyne
and Young, 1974; Maurer, 1962).
Among controllable parameters, rotary speed (RPM) exhibited the strongest relationship
with ROP, with a Pearson correlation coefficient of r = 0.81 and ranking third in Random
Forest feature importance (19.8%). This strong positive relationship is consistent with
theoretical models, including the Bourgoyne and Young (1974) rotary speed function and
empirical observations from numerous field studies (Armenta, 2008). Weight on bit (WOB)
showed a weak linear correlation with ROP (r = -0.02) but ranked fifth in Random Forest
feature importance (7.7 percent), indicating a non-linear relationship that linear correlation
analysis cannot capture. The SHAP analysis revealed that this non-linear relationship
reflects threshold behaviour: WOB shows positive effects on ROP above a threshold but
minimal impact below it.
Flow rate (FLOW) and jet impact force (JET_IMPACT) showed moderate positive
correlations with ROP (r = 0.55), consistent with the hydraulic effects described by Warren
(1984) and incorporated in the Bourgoyne and Young model. However, these parameters
ranked lower in feature importance than depth and RPM, suggesting that hydraulic
conditions were generally adequate throughout the section and did not limit drilling
performance.
5.2.3 Model Performance Findings

70
The comparison of three machine learning algorithms revealed significant differences in
predictive performance on the drilling dataset. Random Forest achieved the highest
accuracy with a coefficient of determination (R²) of 0.904, mean absolute error (MAE) of
1.74 m/hr, and root mean square error (RMSE) of 2.18 m/hr on the test set. XGBoost
achieved R² of 0.813, MAE of 2.13 m/hr, and RMSE of 2.89 m/hr. The Artificial Neural
Network achieved R² of 0.751, MAE of 2.73 m/hr, and RMSE of 3.34 m/hr.
The superior performance of Random Forest on this dataset can be attributed to its ensemble
averaging mechanism, which provides robustness against overfitting and is particularly
valuable when working with small datasets (Breiman, 2001; Hastie et al., 2009). The
bagging approach, combined with random feature selection at each split, creates diversity
among trees that reduces variance without substantially increasing bias. In contrast,
XGBoost's sequential boosting approach, while powerful for larger datasets, may be more
susceptible to overfitting when training data are limited (Chen and Guestrin, 2016). The
neural network's lower performance reflects the well-known challenge of training deep
learning models with limited data, where the large number of parameters can lead to
overfitting despite regularisation measures (Goodfellow et al., 2016).

The Random Forest R² of 0.904 compares favourably with results reported in recent
literature. Mohammadinia et al. (2025) achieved an R² of 0.925 using least-squares support
vector machines combined with the crow search algorithm, though on a larger dataset. Zou
et al. (2025) reported an R² of 0.955 using Random Forest with a more comprehensive
feature set, including explicit formation parameters. Ma et al. (2025) achieved an R² of 0.92
using an automated machine learning workflow. These comparisons suggest that the
performance achieved in this study is competitive with state-of-the-art results, particularly
considering the limited dataset size and absence of explicit formation characterisation data.
5.2.4 Optimization Findings
The optimization analysis using grid search over the feasible parameter space yielded a
remarkable finding: all three machine learning models converged to identical optimal
parameter values despite their fundamentally different architectures and learning
mechanisms. The optimal parameters identified were WOB = 158.0 klbs, Flow = 6.55 GPM,
and RPM = 2370.05. The predicted maximum ROP at these parameters varied slightly
among models (Random Forest: 21.96 m/hr, XGBoost: 20.66 m/hr, ANN: 21.43 m/hr),
reflecting differences in model predictions rather than different optima.

71
This convergence across models provides strong evidence that the identified optimal
parameters represent a genuine global optimum within the observed data range rather than
an artefact of any single modelling approach. The probability of three independent models
with different mathematical foundations converging to identical optima by chance is
extremely low, suggesting that the parameter combination captures real patterns in the
drilling response surface. This finding has important implications for the confidence with
which optimization recommendations can be implemented in practice.
5.2.5 Validation Findings
The sensitivity analysis revealed physically plausible relationships between controllable
parameters and predicted ROP that align with established drilling mechanics theory. For
weight-on-bit, the sensitivity curve exhibited threshold behaviour consistent with the
founder-point concept described by Bourgoyne and Young (1974). At low WOB values
(approximately 1 klb), predicted ROP was approximately 21.72 m/hr. As WOB exceeded
approximately 2 klbs, predicted ROP rapidly increased to approximately 21.94 m/hr and
then remained essentially constant at higher WOB values. This behaviour validates that the
Random Forest model learned a physically meaningful relationship between WOB and
ROP: inadequate weight results in poor bit engagement with the formation, while excessive
weight beyond the founder point provides diminishing returns.
For rotary speed, the sensitivity analysis showed remarkably stable predicted ROP across
the operational range (90-160 RPM), with values varying only between 21.94 and 22.0 m/hr.
While this stability may seem surprising given the theoretical expectation of positive RPM
effects, it indicates that within the observed RPM range, other factors (particularly depth-
related formation properties) dominate the prediction. This finding suggests operational
flexibility in RPM selection without significantly compromising ROP, with practical
implications for managing other operational concerns, such as vibration or directional
control.

5.3 Conclusions
Based on the comprehensive analysis of results, the following conclusions are drawn in
relation to each research objective.
5.3.1 Conclusions on Research Objective 1: Feature Analysis
Research Objective 1: To identify and analyse the key drilling parameters
influencing the rate of penetration using correlation analysis, feature importance methods,
and SHAP analysis.

72
 The multi-method feature analysis successfully identified the key drilling parameters
influencing ROP and provided insights into both linear and non-linear relationships. The
following specific conclusions are drawn:
First, depth-related parameters are the dominant predictors of ROP in the studied
drilling section, collectively accounting for 67 percent of total feature importance. This
dominance reflects the role of depth as a proxy for formation properties, including rock
strength, porosity, and compaction state, which are primary determinants of drillability as
established in theoretical models (Maurer, 1962; Bourgoyne and Young, 1974). The
implication is that formation characterisation, even through depth-based proxies, is essential
for accurate ROP prediction.
Second, among controllable parameters, rotary speed (RPM) shows the strongest
relationship with ROP, with a strong linear correlation (r = 0.81) and a high feature
importance (19.8%). This finding is consistent with theoretical models and suggests that
RPM adjustment is an effective lever for ROP optimization when formation conditions
permit.
Third, weight on bit (WOB) exhibits a complex, non-linear relationship with ROP
that is not captured by linear correlation analysis but is revealed through machine learning
feature importance and SHAP analysis. The weak linear correlation (r = -0.02) combined
with meaningful feature importance (7.7 percent) indicates threshold behaviour: WOB
affects ROP above a threshold but has minimal impact below it. This finding validates the
importance of using multiple analytical methods, as reliance on correlation analysis alone
would underestimate the importance of WOB.
Fourth, the consistency of findings across correlation analysis, Random Forest
feature importance, and SHAP analysis provides confidence in the identified parameter
relationships. Where methods disagreed (as with WOB), the disagreement itself provided
valuable insight into the nature of the relationship.

5.3.2 Conclusions on Research Objective 2: Model Comparison

Research Objective 2: To develop and compare multiple machine learning models for
accurate ROP prediction.
The systematic comparison of Random Forest, XGBoost, and Artificial Neural Network
models yielded clear conclusions regarding algorithm selection for ROP prediction with
small drilling datasets:

73
First, Random Forest is the most suitable algorithm for ROP prediction when working with
small-to-medium-sized drilling datasets (fewer than approximately 100 observations). The
algorithm achieved the highest accuracy (R² = 0.904) through its ensemble averaging
mechanism that provides inherent regularisation against overfitting (Breiman, 2001). This
finding aligns with theoretical understanding of the bias-variance trade-off and the benefits
of bagging for variance reduction.
Second, XGBoost, while a powerful algorithm that has achieved state-of-the-art results in
many applications (Chen and Guestrin, 2016), showed lower performance than Random
Forest on this small dataset (R² = 0.813). The sequential boosting approach may be more
susceptible to overfitting when training data are limited, as each tree attempts to correct
errors that may partly reflect noise rather than signal.
Third, the Artificial Neural Network achieved the lowest performance (R² = 0.751),
consistent with the well-documented challenges of training neural networks with limited
data (Goodfellow et al., 2016). Despite regularisation measures such as dropout, early
stopping, and batch normalisation, the network's large number of parameters relative to the
training set size led to suboptimal generalisation.
Fourth, the performance of Random Forest (R² = 0.904) is competitive with results reported
in recent literature on ROP prediction, demonstrating that accurate predictions are
achievable even with limited data when appropriate algorithms are selected. This has
important implications for practical drilling applications where comprehensive datasets may
not be available.
5.3.3 Conclusions on Research Objective 3: Parameter Optimization
Research Objective 3: To determine optimal drilling parameters for maximising the rate of
penetration.
The optimization analysis yielded robust recommendations for drilling parameter selection:
First, the optimal drilling parameters for the studied section, as identified through
grid search optimization, are WOB = 158.0 klbs, Flow = 6.55 GPM, and RPM = 2370.05.
These parameters correspond to a predicted maximum ROP of approximately 21-22 m/hr,
depending on the model used.
Second, the convergence of all three machine learning models to identical optimal
parameters provides strong evidence that these values represent a genuine global optimum
rather than model-specific artefacts. This convergence is a significant finding that enhances
confidence in the optimization recommendations and suggests that multiple-model
optimization should be considered as a validation strategy in drilling applications.

74
 Third, the optimal parameter values correspond to the upper bounds of the observed
data ranges for WOB and FLOW, and near the upper bound for RPM. This suggests that
within the operational constraints of the studied section, more aggressive drilling parameters
generally improve performance. However, this conclusion is bounded by the observed data
range and should not be extrapolated to parameter values beyond those encountered in the
dataset.
Fourth, the small differences in predicted maximum ROP among models (20.66 to
21.96 m/hr) reflect differences in model prediction accuracy rather than different optima.
This observation reinforces that the optimization is robust to model selection, though the
actual ROP achieved may vary from predictions.
5.3.4 Conclusions on Research Objective 4: Validation
Research Objective 4: To validate the optimization recommendations using
sensitivity analysis and drilling mechanics principles.
The validation analysis confirmed the physical plausibility of the machine learning
models and optimization results:
First, the sensitivity analysis for WOB revealed threshold behaviour consistent with
the founder-point concept established in drilling mechanics theory (Bourgoyne and Young,
1974). The model learned that below a threshold WOB, ROP is relatively insensitive to
weight changes, while above the threshold, adequate formation engagement is achieved.
This physically meaningful behaviour provides confidence that the model captures real
drilling physics rather than spurious correlations.
Second, the sensitivity analysis for RPM showed stable ROP predictions across the
operational range, indicating that within the observed conditions, RPM variations have a
limited impact on predicted ROP. This finding, while perhaps unexpected, has practical
value as it suggests operational flexibility for RPM selection to address other concerns (such
as vibration management or directional control) without significantly compromising drilling
efficiency.
Third, the consistency between model behaviour and the theoretical expectations of
Maurer (1962), Bourgoyne and Young (1974), and Warren (1984) validates that, despite
their black-box nature, machine learning models have learned relationships that align with
established drilling science. This validation is essential for building confidence in model
recommendations and supporting their adoption in operational settings.

75
 Fourth, the cross-model convergence in optimization results provides an additional
layer of validation beyond traditional sensitivity analysis, demonstrating that the identified
optimum is robust to modelling assumptions and not an artefact of any single algorithm.

5.4 Recommendations
Based on the findings and conclusions of this study, the following recommendations are
offered for drilling operations, machine learning practitioners, and future research.
5.4.1 Recommendations for Drilling Operations
The following recommendations are directed toward drilling engineers and operators
seeking to optimise ROP in similar drilling environments:
First, the identified optimal parameters (WOB = 158.0 klbs, Flow = 6.55 GPM, RPM
= 2370.05) should be considered starting points for selecting drilling parameters in similar
formations with PDC bits. However, these parameters should be validated through field
observation and adjusted based on local conditions, equipment constraints, and real-time
drilling response. The parameters represent optima within the observed data range and may
not be universally applicable.
Second, the weight on bit should be maintained above the threshold identified
through the sensitivity analysis (approximately 2 klbs normalised) to ensure adequate
formation engagement. Operating below this threshold results in inefficient drilling
regardless of other parameter settings. However, maximising WOB beyond the threshold
provides diminishing returns, suggesting that moderate WOB levels above the threshold
may be preferred to reduce equipment wear and vibration risk.
Third, rotary speed can be adjusted within the operational range (90-160 RPM) for
purposes other than ROP optimization, such as vibration management, directional control,
or bit wear management, without significantly compromising drilling efficiency. This
operational flexibility should be leveraged to optimise overall drilling performance,
considering multiple objectives.
Fourth, drilling operations should collect and archive comprehensive drilling data to
support future machine learning applications. The success of machine learning approaches
depends on data quality and availability, and investment in data infrastructure will enable
increasingly sophisticated optimization capabilities over time.
Fifth, operators should consider implementing the multi-model optimization
approach demonstrated in this study, which uses multiple algorithms to validate
optimization recommendations. Convergence across models provides confidence in

76
recommendations, while divergence may indicate uncertainty that warrants additional
investigation or conservative parameter selection.
The potential value of implementing these recommendations is underscored by documented
field successes with similar AI-based optimization approaches. Field trials by Al-Riyami et
al. (2023) demonstrated that consistent adherence to machine learning recommendations
can yield substantial time savings, with one West African well reaching total depth
approximately 24 days ahead of plan. Singh et al. (2021) reported that operators achieved
ROP improvements of 19 to 33 percent without inducing detrimental vibrations or
directional instability, demonstrating that aggressive optimization can be pursued safely
when guided by properly validated models. The optimal parameters identified in this study
(WOB = 158 klbs, Flow = 6.55 GPM, RPM = 2370) could serve as starting points for similar
optimization efforts in comparable formations, with real-time adjustment based on observed
formation response and downhole feedback.
5.4.2 Recommendations for Machine Learning Practitioners
The following recommendations are directed toward data scientists and engineers
developing machine learning models for drilling applications:
First, Random Forest should be the default algorithm choice for ROP prediction
when working with small to medium-sized datasets (fewer than 100 observations). The
algorithm's inherent regularisation through ensemble averaging provides robustness against
overfitting, particularly when data are limited. This recommendation is consistent with
findings from Osman et al. (2021) and Zou et al. (2025).
Second, multiple feature analysis methods should be employed to comprehensively
understand parameter relationships. Correlation analysis alone may miss important non-
linear relationships, while tree-based importance measures may be affected by correlated
features. The combination of correlation analysis, permutation importance, and SHAP
analysis provides complementary perspectives that enhance understanding.
Third, SHAP analysis should be routinely applied to interpret model behaviour and
validate physical plausibility. The game-theoretic foundation of SHAP values (Lundberg
and Lee, 2017) provides a rigorous framework for understanding feature contributions that
can be compared with domain knowledge expectations. Models that exhibit SHAP patterns
inconsistent with drilling theory should be treated with caution.
Fourth, sensitivity analysis should be conducted as a standard validation step before
operational implementation of optimization recommendations. The analysis reveals the

77
shape of the response surface near the optimum and can identify physically implausible
model behaviour that might not be apparent from accuracy metrics alone.
Fifth, multi-model optimization should be considered as a validation strategy where
computational resources permit. The convergence of multiple models to similar optima
provides confidence that exceeds that of any single model, particularly when the models
represent different mathematical paradigms (e.g., bagging, boosting, neural networks).
5.4.3 Recommendations for Future Research
The following recommendations identify productive directions for future research that could
extend and strengthen the findings of this study:
First, the methodology should be validated on larger datasets from multiple wells
and formations to assess generalisability. While the current study demonstrates
effectiveness on a single section, broader validation would establish the conditions under
which the approach performs well and identifies potential limitations.
Second, future research should incorporate explicit formation-characterisation
parameters, such as gamma-ray logs, sonic logs, or laboratory-measured rock properties, to
reduce reliance on depth as a proxy for formation characteristics. This would potentially
improve prediction accuracy and enable the development of more transferable models.
Third, multi-objective optimization frameworks should be developed that consider
not only ROP but also bit wear, vibration risk, hole quality, and other operational concerns.
The current single-objective optimization provides a foundation, but practical drilling
decisions involve trade-offs among multiple objectives that should be explicitly modeled.
Fourth, deep learning approaches, including LSTM networks and physics-informed neural
networks (Jiao et al., 2024; Zhang et al., 2024), should be evaluated on larger drilling
datasets where their capacity for capturing complex patterns can be fully utilised without
overfitting concerns.
Fifth, real-time implementation frameworks should be developed that enable
continuous model updating and adaptive optimization as drilling progresses. The current
batch learning approach could be extended to online learning that adapts to changing
formation conditions.
Sixth, field validation studies should be conducted to compare actual drilling
performance under model-recommended parameters with that under conventional
parameter selection, providing direct evidence of the practical value of machine learning
optimization.

78
5.5 Contributions of the Study
This study makes several contributions to both academic knowledge and practical
application in the field of drilling optimization.
5.5.1 Theoretical Contributions
First, the study provides a systematic comparative analysis of machine learning
algorithms for ROP prediction under limited data conditions. While numerous studies have
applied individual algorithms to drilling data, systematic comparisons using consistent
methodology are relatively rare. The finding that Random Forest outperforms XGBoost and
neural networks on small drilling datasets contributes to the understanding of algorithm
selection in this application domain.
Second, the study demonstrates the value of multi-method feature analysis for
understanding parameter relationships in drilling. The combination of correlation analysis,
tree-based importance, and SHAP analysis revealed both linear and non-linear relationships
that would not be apparent from any single method. This methodological contribution has
implications beyond drilling to other engineering applications where understanding feature
relationships is important.
Third, the cross-model convergence analysis provides a novel approach to validating
optimization robustness. The demonstration that three fundamentally different algorithms
converge to identical optima provides stronger evidence than traditional single-model
optimization, and this approach could be applied to other engineering optimization
problems.
Fourth, validating machine learning models against drilling mechanics theory
(Maurer, 1962; Bourgoyne and Young, 1974; Warren, 1984) provides an approach to ensure
physical plausibility that bridges data-driven and physics-based modelling paradigms. This
contribution addresses concerns about the black-box nature of machine learning models and
supports their adoption in safety-critical engineering applications.
5.5.2 Practical Contributions
First, the study provides validated optimal drilling parameters for the studied
formation that can serve as starting points for similar drilling operations. While site-specific
validation is always required, data-driven baseline recommendations reduce trial-and-error
and accelerate optimization.
Second, the methodology developed in this study provides a framework that drilling
engineers can apply to their own datasets. The step-by-step approach from data

79
preprocessing through feature analysis, model development, optimization, and validation
can be replicated with different data to generate site-specific recommendations.
Third, the finding that RPM can be adjusted within the operational range without
significantly impacting ROP provides practical operational flexibility. This allows drillers
to optimise RPM for secondary objectives such as vibration management or directional
control while maintaining drilling efficiency.
Fourth, the sensitivity analysis results provide insight into the robustness of optimal
parameters. The threshold behaviour for WOB indicates that maintaining weight above the
threshold is more important than achieving the exact optimal value, providing tolerance for
operational variability.
Fifth, the demonstration that effective ROP optimization is achievable with a dataset
of only 43 observations has practical implications for drilling operations with limited data
availability. This lowers the barrier to adoption of machine learning optimization
approaches.
5.6 Limitations of the Study
The findings and conclusions of this study should be interpreted in light of several
limitations that define the boundaries of applicability.
5.6.1 Data Limitations
First, the dataset comprises only 43 observations, which limits the complexity of
patterns that can be reliably learned and the precision of performance estimates. While the
methodology demonstrates effectiveness on this dataset, larger datasets would enable more
robust training, validation on truly independent test sets, and potentially different
conclusions regarding optimal algorithms.
Second, the data are from a single drilling section with a constant PDC bit type and
size. The findings may not generalise directly to other formations, bit types, or drilling
systems without additional validation. The optimal parameters identified are specific to the
studied conditions.
Third, the dataset does not include explicit characterisation parameters for
formation, such as rock strength, porosity, or mineralogy. Depth serves as a proxy for
formation properties, but this proxy relationship depends on the specific geological setting
and may not transfer to other environments.

5.6.2 Methodological Limitations

80
 First, the study focuses on steady-state drilling conditions and does not address
transient phenomena, drilling dysfunctions, or abnormal events. A comprehensive drilling
optimization system would need to incorporate detection and response to these conditions.
Second, operational constraints such as equipment limitations, formation stability
requirements, and vibration thresholds were not explicitly incorporated into the
optimization. The identified optimal parameters may not be achievable or advisable under
all operational conditions.
Third, the study evaluated only three machine learning algorithms. Other
approaches, including support vector machines, Gaussian processes, or advanced deep
learning architectures, were not evaluated and might achieve different results.
Fourth, the validation through sensitivity analysis and drilling mechanics
comparison, while valuable, does not substitute for field validation through actual
implementation and performance measurement.
5.6.3 Scope Limitations
First, the optimization considered only three controllable parameters (WOB, FLOW,
RPM). Other potentially important parameters such as mud properties, bit design features,
or directional drilling parameters were not included in the optimization.
Second, the study addressed single-objective optimization (maximising ROP)
without considering trade-offs with other objectives such as bit wear, hole quality, or cost.
Multi-objective optimization would provide a more comprehensive basis for operational
decisions.
Third, the study did not address real-time implementation challenges including data
latency, sensor noise, and the need for rapid model inference. Practical implementation
would require addressing these engineering challenges.

5.7 Future Research Directions

Building on the findings and acknowledging the limitations of this study, several directions
for future research are identified that could advance the field of machine learning-based
drilling optimization.

5.7.1 Extended Validation Studies

Future research should prioritise validation of the methodology on larger and more diverse
datasets. Studies incorporating data from multiple wells, multiple formations, and different
drilling systems would establish the generalisability of findings and identify conditions

81
where the approach performs best. Collaborative research across multiple operators could
provide the data diversity needed for comprehensive validation.
5.7.2 Integration of Formation Characterisation
The development of models that explicitly incorporate formation characterisation, whether
from logging-while-drilling measurements, seismic attributes, or offset well data, could
improve prediction accuracy and model transferability. Research on feature engineering
approaches that effectively combine drilling parameters with formation data would advance
the state of the art.
5.7.3 Multi-Objective Optimization
Extension of the optimization framework to consider multiple objectives simultaneously
would better reflect practical drilling decisions. Research on Pareto-optimal solutions that
trade off ROP against bit wear, vibration risk, hole quality, and cost would provide drilling
engineers with a more complete basis for parameter selection. Evolutionary algorithms and
other multi-objective optimization methods could be applied to this problem.
5.7.4 Advanced Machine Learning Architectures
As drilling data availability improves, research on more sophisticated machine learning
architectures becomes increasingly relevant. Long Short-Term Memory (LSTM) networks
that capture temporal dependencies in drilling sequences (Zhang et al., 2024), physics-
informed neural networks that incorporate drilling mechanics constraints (Jiao et al., 2024),
and hybrid models that combine machine learning with physics-based components all offer
potential for improved performance.
5.7.5 Real-Time Implementation
Research on real-time implementation frameworks would bridge the gap between academic
research and operational deployment. This includes work on online learning algorithms that
update continuously as new data arrive, edge computing approaches that enable rapid
inference at the rig site, and human-machine interfaces that present recommendations to
drillers in actionable form. Integration with existing drilling automation systems is also an
important research direction.
5.7.6 Uncertainty Quantification
Development of methods for quantifying uncertainty in ROP predictions and optimization
recommendations would enhance decision support. Bayesian approaches, ensemble
methods with prediction intervals, and conformal prediction are potential frameworks for
uncertainty quantification that could be applied to drilling optimization.
5.7.7 Transfer Learning

82
Research on transfer learning approaches that enable models trained on data from one well
or formation to be adapted to new conditions with limited local data could address the
perennial challenge of data scarcity in drilling applications. This could significantly reduce
the data requirements for implementing machine learning optimization in new drilling
campaigns.

5.8 Concluding Remarks

This study has demonstrated that machine learning techniques, particularly Random Forest,
can effectively predict and optimise the rate of penetration in drilling operations, even with
limited datasets. The research achieved its four objectives: identifying key parameters
through multi-method feature analysis; establishing Random Forest as the preferred
algorithm for small drilling datasets; determining robust, optimal parameters through multi-
model convergence; and validating results through sensitivity analysis and drilling
mechanics principles.

The convergence of three fundamentally different machine learning algorithms to identical

optimal drilling parameters provides strong evidence that the identified optimum represents
genuine patterns in the drilling response rather than modelling artefacts. The validation
against established drilling mechanics theory, particularly the WOB threshold behaviour
consistent with the founder point concept (Bourgoyne and Young, 1974), confirms that the
models have learned physically meaningful relationships.

The practical implications of this research extend beyond the specific parameters identified
for the studied section. The methodology demonstrated here provides a framework that
drilling engineers can apply to their own data to generate site-specific optimization
recommendations. The finding that effective optimization is achievable with as few as 43
observations lowers the barrier to adoption and makes machine learning approaches
accessible to a wider range of drilling operations.
As the drilling industry continues its digital transformation, the integration of machine
learning with operational decision-making will become increasingly important. This study
contributes to that transformation by demonstrating what is achievable with current
methods, identifying best practices for algorithm selection and validation, and pointing
toward future research directions that could further advance the field. The ultimate goal is
drilling operations that are safer, more efficient, and more sustainable, and machine

83
learning-based optimization is one tool among many that can contribute to achieving that
goal.

The study's contributions include systematic algorithm comparison for small drilling
datasets, multi-model convergence as a validation strategy, and a replicable methodology
for drilling optimization. Limitations related to data size, single formation focus, and scope
constraints were acknowledged. Future research directions were identified, including
extended validation, formation characterisation integration, multi-objective optimization,
advanced architectures, real-time implementation, uncertainty quantification, and transfer
learning.

The research demonstrates that machine learning-based drilling optimization is both

achievable and practically valuable, providing a foundation for continued advancement in
this important area of drilling engineering.
Random Forest can effectively predict and optimise ROP even with limited data.

84
 REFERENCES

Abbas, A.K., Rushdi, S., and Alsaba, M. (2018). Reduction of stuck pipe incidents in
deviated and horizontal wells using real-time drilling data. Paper presented at the
SPE Kingdom of Saudi Arabia Annual Technical Symposium and Exhibition. SPE-
192251-MS.
Agbaji, A.L. (2024). Machine learning application for rate of penetration prediction and
optimization in drilling operations. Journal of Petroleum Technology and
Alternative Fuels, 15(2), [Link]
Al-Riyami, N., Revheim, O., Robinson, T.S., Batruny, P., Meor Hakeem, M.H. and Tze
Ping, G. (2023) 'Drilling in the digital age: case studies of field testing a real-time
ROP optimization system using machine learning', SPE/IADC Middle East Drilling
Technology Conference and Exhibition, Abu Dhabi, UAE, May 2023. doi:
10.2118/214521-MS.
Anemangely, M., Ramezanzadeh, A., Amiri, H., and Hoseinpour, S.A. (2018). Machine
learning technique for the prediction of shear wave velocity using petrophysical
logs. Journal of Petroleum Science and Engineering, 174, 306-327.
Armenta, M. (2008). Identifying inefficient drilling conditions using drilling-specific
energy. Paper presented at the SPE Annual Technical Conference and Exhibition,
Denver, Colorado. SPE-116667-MS.
Bahari, A., & Baradaran Seyed, A. (2007). Trust-region approach to find constants of
Bourgoyne and Young penetration rate model in Khangiran Iranian gas field (Paper
No. SPE-107520-MS). Society of Petroleum Engineers.
[Link]
Barbosa, L.F., Nascimento, A., Mathias, M.H., and de Carvalho, J.A. (2019). Machine
learning methods applied to drilling rate of penetration prediction and optimization:
A review. Journal of Petroleum Science and Engineering, 183, 106332.
Barros, C.P. (2015). Rate of penetration optimization using artificial intelligence techniques.
MSc Thesis, Louisiana State University.
Bataee, M., Irawan, S., & Kamyab, M. (2014). Artificial neural network model for
prediction of drilling rate of penetration and optimization of parameters. Journal of
the Japan Petroleum Institute, 57(2), 65-70. [Link]
Bellin, F., Dourfaye, A., King, W., and Thigpen, M. (2010). The current state of PDC bit
technology. World Oil, 231(10), 41-46.

85
Bergstra, J., and Bengio, Y. (2012). Random search for hyper-parameter optimization.
Journal of Machine Learning Research, 13, 281-305.
Bilgesu, H.I., Tetrick, L.T., Altmis, U., Mohaghegh, S., and Ameri, S. (1997). A new
approach for the prediction of rate of penetration values. Paper presented at the SPE
Eastern Regional Meeting, Lexington, Kentucky. SPE-39231-MS.
Bingham, M.G. (1965). A new approach to interpreting rock drillability. Oil and Gas
Journal, 63(46), 173-179.
Bodaghi, A., Ansari, H.R., and Gholami, M. (2015). Optimized support vector regression
for drilling rate of penetration estimation. Open Geosciences, 7(1), 870-879.
Boukredera, F.S., Youcefi, M.R., Hadjadj, A., Ezenkwu, C.P., Vaziri, V. and Aphale, S.S.
(2023) 'Enhancing the drilling efficiency through the application of machine
learning and optimization algorithm', Engineering Applications of Artificial
Intelligence, 126, 107035. doi: 10.1016/[Link].2023.107035.
Bourgoyne, A.T., & Young, F.S. (1974). A multiple regression approach to optimal drilling
and abnormal pressure detection. Society of Petroleum Engineers Journal, 14(4),
371-384. [Link]
Bourgoyne, A.T., Millheim, K.K., Chenevert, M.E., and Young, F.S. (1986). Applied
Drilling Engineering. Society of Petroleum Engineers.
Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5-32.
Chen, T., and Guestrin, C. (2016). XGBoost: A scalable tree boosting system. Proceedings
of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining, 785-794.
Creswell, J.W., and Creswell, J.D. (2018). Research Design: Qualitative, Quantitative, and
Mixed Methods Approaches (5th ed.). SAGE Publications.
Delavar, M.R., and Ramezanzadeh, A. (2023). Machine learning classification approaches
to optimize ROP and TOB using drilling and geomechanical parameters in a
carbonate reservoir. Journal of Petroleum Exploration and Production Technology.
DOI: 10.1007/s13202-024-01769-9.
Dormann, C.F., Elith, J., Bacher, S., Buchmann, C., Carl, G., Carré, G., Marquéz, J.R.G.,
Gruber, B., Lafourcade, B., Leitão, P.J., Münkemüller, T., McClean, C., Osborne,
P.E., Reineking, B., Schröder, B., Skidmore, A.K., Zurell, D., and Lautenbach, S.
(2013). Collinearity: A review of methods to deal with it and a simulation study
evaluating their performance. Ecography, 36(1), 27-46.

86
Dupriest, F.E., and Koederitz, W.L. (2005). Maximizing drill rates with real-time
surveillance of mechanical specific energy. Paper presented at the SPE/IADC
Drilling Conference, Amsterdam, Netherlands. SPE-92194-MS.
Elkatatny, S. (2018). New approach to optimize the rate of penetration using artificial neural
network. Arabian Journal for Science and Engineering, 43(11), 6297-6304.
[Link]
Elkatatny, et al. (2017). An integrated approach for estimating static Young's modulus using
artificial intelligence tools. Neural Computing and Applications, 31, 4123-4135.
Ericsson. (2023). Futureproof oil and gas through digitalization. Ericsson Blog, August 23,
2023.
Eskandarian, S., Bahrami, P., & Kazemi, P. (2017). A comprehensive data mining approach
to estimate the rate of penetration: Application of neural network, rule-based models
and feature ranking. Journal of Petroleum Science and Engineering, 156, 605-615.
[Link]
Fisher, A., Rudin, C., and Dominici, F. (2019). All models are wrong, but many are useful:
Learning a variable's importance by studying an entire class of prediction models
simultaneously. Journal of Machine Learning Research, 20(177), 1-81.
Friedman, J.H. (2001). Greedy function approximation: A gradient boosting machine.
Annals of Statistics, 29(5), 1189-1232. [Link]
Galle, E.M., & Woods, H.B. (1963). Best constant weight and rotary speed for rotary rock
bits. In Drilling and Production Practice (pp. 48-73). American Petroleum Institute.
Geron, A. (2019). Hands-on Machine Learning with Scikit-Learn, Keras, and TensorFlow:
Concepts, Tools, and Techniques to Build Intelligent Systems (2nd ed.). O'Reilly
Media.
Glowka, D.A. (1987). Development of a method for predicting the performance and wear
of PDC (polycrystalline diamond compact) drill bits (Report No. SAND-86-1745).
Sandia National Laboratories. [Link]
Goldberg, D.E. (1989). Genetic algorithms in search, optimization, and machine learning.
Addison-Wesley.
Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep Learning. MIT Press.
Graham, J., & Muench, N. (1959). Analytical determination of optimum bit weight and
rotary speed combinations. In Fall Meeting of the Society of Petroleum Engineers of
AIME (Paper No. SPE-1349-G). Society of Petroleum Engineers.
[Link]

87
Guria, C., Goli, K.K., & Pathak, A.K. (2014). Multi-objective optimization of oil well
drilling using elitist non-dominated sorting genetic algorithm. Petroleum Science,
11(1), 97-110. [Link]
Hai, Z., Wang, H., Zhang, L., Chen, M., and Liu, Y. (2025). Online rate of penetration
prediction using gradient boosted decision trees with Bayesian optimization.
Geoenergy Science and Engineering, 234, 212678.
Hareland, G., & Rampersad, P.R. (1994). Drag-bit model including wear (Paper No. SPE-
26957-MS). In SPE Latin America/Caribbean Petroleum Engineering Conference.
Society of Petroleum Engineers. [Link]
Hastie, T., Tibshirani, R., and Friedman, J. (2009). The Elements of Statistical Learning:
Data Mining, Inference, and Prediction (2nd ed.). Springer.
Hegde, C., & Gray, K.E. (2017). Use of machine learning and data analytics to increase
drilling efficiency for nearby wells. Journal of Natural Gas Science and
Engineering, 40, 327-335. [Link]
Hegde, C., Daigle, H., Millwater, H., and Gray, K. (2017). Analysis of rate of penetration
(ROP) prediction in drilling using physics-based and data-driven models. Journal of
Petroleum Science and Engineering, 159, 295-306.
Holland, J.H. (1975). Adaptation in Natural and Artificial Systems. University of Michigan
Press.
Hossain, M.E., and Al-Majed, A.A. (2015). Fundamentals of Sustainable Drilling
Engineering. Scrivener Publishing.
IEA. (2023). Oil 2023: Analysis and forecast to 2028. International Energy Agency.
[Link]
IEA. (2023). The oil and gas industry in net zero transitions. International Energy Agency.
[Link]
IEA. (2023). World energy investment 2023. International Energy Agency.
[Link]
IHS Markit. (2020). Digital Transformation in Oil and Gas Operations [Industry Study].
IHS Markit Energy.
Ioffe, S., and Szegedy, C. (2015). Batch normalization: Accelerating deep network training
by reducing internal covariate shift. Proceedings of the 32nd International
Conference on Machine Learning, 448-456.

88
Irawan, S., Rahman, A., and Tunio, S.Q. (2012). Optimization of weight on bit during
drilling operation based on rate of penetration model. Research Journal of Applied
Sciences, Engineering and Technology, 4(12), 1690-1695.
Jiao, H., Ding, S., Zhou, Y., Guo, B., and Wang, Z. (2024). A hybrid physics-machine
learning approach for rate of penetration prediction with improved generalization.
Journal of Petroleum Science and Engineering, 225, 111689.
Kennedy, J., and Eberhart, R. (1995). Particle swarm optimization. Proceedings of the IEEE
International Conference on Neural Networks, 4, 1942-1948.
Khalilidermani, M., and Knez, D. (2023). A survey on rate of penetration predictive models:
Current status and future directions. Energies, 16(5), 2376.
Khodja, M., Khodja-Saber, M., Canselier, J.P., Cohaut, N., and Bergaya, F. (2010). Drilling
Fluid Technology: Performances and Environmental Considerations. In: Products
and Services from R&D to Final Solutions, edited by Igor Fuerstner. IntechOpen.
DOI: 10.5772/10393.
Kingma, D.P., and Ba, J. (2015). Adam: A method for stochastic optimization. Proceedings
of the 3rd International Conference on Learning Representations (ICLR).
Kohavi, R. (1995). A study of cross-validation and bootstrap for accuracy estimation and
model selection. Proceedings of the 14th International Joint Conference on Artificial
Intelligence, 2, 1137-1143.
LeCun, Y., Bottou, L., Orr, G.B., and Müller, K.R. (2012). Efficient backprop. In Neural
Networks: Tricks of the Trade (2nd ed., pp. 9-48). Springer.
Little, R.J.A., and Rubin, D.B. (2019). Statistical Analysis with Missing Data (3rd ed.).
Wiley.
Lundberg, S.M., and Lee, S.I. (2017). A unified approach to interpreting model predictions.
Advances in Neural Information Processing Systems, 30, 4765-4774.
Ma, Z., Yao, J., Liu, P., Li, G., and Sun, X. (2025). An interpretable automated machine
learning workflow for rate of penetration prediction. Geoenergy Science and
Engineering, 236, 212742.
Mantha, B., & Samuel, R. (2016). ROP optimization using artificial intelligence techniques
with statistical regression coupling. In SPE Annual Technical Conference and
Exhibition (Paper No. SPE-181382-MS). Society of Petroleum Engineers.
[Link]
Maurer, W.C. (1962). The 'perfect-cleaning' theory of rotary drilling. Journal of Petroleum
Technology, 14(11), 1270-1274. SPE-408-PA.

89
McKinsey & Company. (2020). Digital transformation in energy: Achieving escape
velocity. McKinsey & Company. [Link]
gas/our-insights/digital-transformation-in-energy-achieving-escape-velocity.
Mirjalili, S., Mirjalili, S.M., and Lewis, A. (2014). Grey wolf optimizer. Advances in
Engineering Software, 69, 46-61.
Molnar, C. (2022). Interpretable machine learning: A guide for making black box models
explainable (2nd ed.). Independently published.
[Link]
Mohammadinia, F., Ranaei, M., Aminshahidi, B., and Aminshahidy, B. (2025). Rate of
penetration prediction using machine learning approaches: A case study of Fahliyan
Formation in Iran. Journal of Petroleum Exploration and Production Technology,
15(1), 45-62.
Motahhari et al. (2010). Improved drilling efficiency technique using integrated PDM and
PDC bit parameters. Journal of Canadian Petroleum Technology, 49(10), 45-52.
[Link]
Najjarpour, M., Jalalifar, H., and Norouzi, S. (2022). Fifty years of experience in rate of
penetration management: Application of machine learning methods and
optimization algorithms: A review. Journal of Petroleum Science and Engineering,
208, 109575.
Onwunalu, J.E., & Durlofsky, L.J. (2010). Application of a particle swarm optimization
algorithm for determining optimum well location and type. Computational
Geosciences, 14(1), 183-198. [Link]
Osman, E.A., Aggour, M.A., and Al-Yami, A.S. (2021). Rate of penetration prediction in
carbonate rocks using machine learning: A comparative study. Paper presented at
the International Petroleum Technology Conference. IPTC-21224-MS.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel,
M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A.,
Cournapeau, D., Brucher, M., Perrot, M., and Duchesnay, E. (2011). Scikit-learn:
Machine learning in Python. Journal of Machine Learning Research, 12, 2825-2830.
Perrin, V.P., Mensa-Wilmot, G., and Alexander, W.L. (1997). Drilling index: A new
approach to bit performance evaluation. Paper presented at the SPE/IADC Drilling
Conference, Amsterdam, Netherlands, 4-6 March 1997. SPE-37595-MS.
Pessier, R.C., & Fear, M.J. (1992). Quantifying common drilling problems with mechanical
specific energy and a bit-specific coefficient of sliding friction. In SPE Annual

90
 Technical Conference and Exhibition (Paper No. SPE-24584-MS). Society of
Petroleum Engineers. [Link]
Prestidge, K. (2022). Digital Transformation in the Oil and Gas Industry. SM thesis,
Massachusetts Institute of Technology, System Design and Management Program.
Rehm, B., and McClendon, R. (1971). Measurement of formation pressure from drilling
data. Paper presented at the SPE Annual Fall Meeting, New Orleans, Louisiana.
SPE-3601-MS.
Singh, K., Siddiqui, F., Omeragic, D., Kamyab, M. and Cheatham, C. (2021) 'From science
to practice: improving ROP by utilizing a cloud-based machine-learning
Speight, J.G. (2020). Handbook of Petroleum Refining. CRC Press. Speight, J.G. (2014).
Handbook of Offshore Oil and Gas Operations. Gulf Professional Publishing.
Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., and Salakhutdinov, R. (2014).
Dropout: A simple way to prevent neural networks from overfitting. Journal of
Machine Learning Research, 15(56), 1929-1958.
solution in real-time drilling operations', SPE/IADC Drilling Conference and Exhibition,
Virtual, March 2021. doi: 10.2118/204107-MS.
Schwab, K. (2016). The fourth industrial revolution. Crown Business.
Soares, C., and Gray, K. (2019). Real-time predictive capabilities of analytical and machine
learning rate of penetration (ROP) models. Journal of Petroleum Science and
Engineering, 172, 934-959.
Sridharan, S., and Kumar, K.V. (2020). Rate of penetration optimization using neural
network approach. International Journal of Applied Engineering Research, 14(7),
1617-1623.
Sridharan, C., & Kumar, G.S. (2020). Drilling efficiency improvement and rate of
penetration optimization by machine learning and data analytics. International
Journal of Mathematical, Engineering and Management Sciences, 5(3), 381-394.
[Link]
Štrumbelj, E., & Kononenko, I. (2010). An efficient explanation of individual classifications
using game theory. Journal of Machine Learning Research, 11, 1-18.
Štrumbelj, E., & Kononenko, I. (2014). Explaining prediction models and individual
predictions with feature contributions. Knowledge and Information Systems, 41(3),
647-665. [Link]
Teale, R. (1965). The concept of specific energy in rock drilling. International Journal of
Rock Mechanics and Mining Sciences, 2(1), 57-73.

91
Vapnik, V.N. (1995). The Nature of Statistical Learning Theory. Springer.
Warren, T.M. (1984). Factors affecting torque for a roller cone bit. Journal of Petroleum
Technology, 36(09), 1500-1508. SPE-11994-PA.
Warren, T.M. (1987). Penetration rate performance of roller cone bits. SPE Drilling
Engineering, 2(01), 9-18. SPE-13259-PA.
Warren, T.M., & Armagost, W.K. (1988). Laboratory drilling performance of PDC bits.
SPE Drilling Engineering, 3(2), 125-135. [Link]
Winters et al. (1987). Roller bit model with rock ductility and cone offset. Paper presented
at the SPE Annual Technical Conference and Exhibition, Dallas, Texas. SPE-16696-
MS.
Xiong et al. (2024). A rate of penetration (ROP) prediction method based on improved dung
beetle optimization algorithm and BiLSTM-SA. Scientific Reports, 14, 25856. DOI:
10.1038/s41598-024-75703-8.
Zakariya, R., Zein, A., Diab, E., Lotfy, A., Marland, C., Obaidli, Y.Y.A., Braik, H.A.A.,
Amin, M.S., & Attalah, M. (2015). A case study of real-time drilling optimization
to improve well delivery through enhancing drilling rates and identifying invisible
lost time to improve performance. In SPE North Africa Technical Conference and
Exhibition (Paper No. SPE-175748-MS). Society of Petroleum Engineers.
[Link]
Zhang, H., Lu, B., Liao, L., Tao, H., He, H., Zou, J., and Xiao, L. (2021). Combining
Machine Learning and Classic Drilling Theories to Improve Rate of Penetration
Prediction. SPE/IADC Middle East Drilling Technology Conference and Exhibition,
Abu Dhabi, UAE. SPE-202202-MS. DOI: 10.2118/202202-MS.
Zhang, C., Song, D., Chen, Y., and Guo, B. (2024). LSTM-based real-time drilling
optimization with temporal pattern recognition. SPE Journal, 29(02), 892-905.
Zou, J., Han, L., Yang, Q., and Wu, Q. (2025). Random forest-based rate of penetration
prediction with SHAP interpretation: A case study from southeastern Iraq. Arabian
Journal of Geosciences, 18(2), 112.

92
 APPENDIX A

PYTHON CODE FOR ROP PREDICTION AND OPTIMISATION

This appendix presents the consolidated Python code used for data preprocessing, feature
analysis, machine learning model development (Random Forest, XGBoost, and Artificial
Neural Network), parameter optimisation, and visualisation generation.

Listing A.1: Complete Python Implementation

# =============================================================================
# ROP PREDICTION AND OPTIMISATION USING MACHINE LEARNING
# Complete Python Implementation
# =============================================================================

# -----------------------------------------------------------------------------
# 1. LIBRARY IMPORTS
# -----------------------------------------------------------------------------
import os
import numpy as np
import pandas as pd
import [Link] as plt
import seaborn as sns
from sklearn.model_selection import train_test_split
from [Link] import RandomForestRegressor
from [Link] import StandardScaler
from [Link] import r2_score, mean_absolute_error
from [Link] import permutation_importance
from [Link] import minimize
import xgboost as xgb
import tensorflow as tf
from [Link] import Sequential
from [Link] import Dense
from [Link] import EarlyStopping
import shap

# -----------------------------------------------------------------------------
# 2. DATA LOADING AND PREPROCESSING
# -----------------------------------------------------------------------------
DATA_PATH = "ROP OPTIMIZATION [Link]"
TARGET = "ROP"
CONTROLLABLE = ["WOB", "RPM", "FLOW"]
RANDOM_STATE = 42

def normalize_columns(cols):
"""Normalise column names: strip, uppercase, replace spaces."""
new = []
for c in cols:
if c is None:
[Link]("")
continue
s = [Link]().upper().replace(" ", "_").replace("/", "_")
[Link](s)
return new

93
# Load and clean data
df = pd.read_excel(DATA_PATH)
[Link] = normalize_columns([Link])
df = [Link](subset=[TARGET])

# Define features
X_cols = ["WOB", "RPM", "FLOW", "TORQUE", "JET_IMPACT", "MD", "MD_IN", "MD_OUT"]
X = df[X_cols].copy()
y = df[TARGET].values

# Train-test split
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=0.3, random_state=RANDOM_STATE
)

# -----------------------------------------------------------------------------
# 3. FEATURE ANALYSIS
# -----------------------------------------------------------------------------
# Correlation analysis
num_only = df.select_dtypes(include=[[Link]])
corr_with_target = num_only.corr()[TARGET].sort_values(
key=lambda x: [Link](), ascending=False
)

# -----------------------------------------------------------------------------
# 4. RANDOM FOREST MODEL
# -----------------------------------------------------------------------------
rf_model = RandomForestRegressor(n_estimators=400, random_state=RANDOM_STATE)
rf_model.fit(X_train, y_train)
y_pred_rf = rf_model.predict(X_test)

print("Random Forest R2:", r2_score(y_test, y_pred_rf))

print("Random Forest MAE:", mean_absolute_error(y_test, y_pred_rf))

# Feature importance
rf_importance = [Link](
rf_model.feature_importances_, index=[Link]
).sort_values(ascending=False)

# SHAP analysis for Random Forest

explainer_rf = [Link](rf_model)
shap_values_rf = explainer_rf.shap_values(X_train)

# -----------------------------------------------------------------------------
# 5. XGBOOST MODEL
# -----------------------------------------------------------------------------
xgb_model = [Link](
n_estimators=500, max_depth=5, learning_rate=0.05, random_state=RANDOM_STATE
)
xgb_model.fit(X_train, y_train)
y_pred_xgb = xgb_model.predict(X_test)

print("XGBoost R2:", r2_score(y_test, y_pred_xgb))

print("XGBoost MAE:", mean_absolute_error(y_test, y_pred_xgb))
# XGBoost Feature Importance

94
xgb_importance = [Link](
xgb_model.feature_importances_, index=[Link]
).sort_values(ascending=False)

print("XGBoost Feature Importance:")

print(xgb_importance)

# -----------------------------------------------------------------------------
# 6. ARTIFICIAL NEURAL NETWORK MODEL
# -----------------------------------------------------------------------------
[Link].set_seed(RANDOM_STATE)
[Link](RANDOM_STATE)

scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)
X_test_scaled = [Link](X_test)

ann_model = Sequential([
[Link](shape=(X_train.shape[1],)),
Dense(16, activation='relu'),
Dense(8, activation='relu'),
Dense(1)
]
ann_model.compile(optimizer='adam', loss='mse', metrics=['mae'])

early_stop = EarlyStopping(monitor='val_loss', patience=20,

restore_best_weights=True)

ann_model.fit(X_train_scaled, y_train, validation_split=0.2,

epochs=300, batch_size=16, callbacks=[early_stop], verbose=0)

y_pred_ann = ann_model.predict(X_test_scaled).flatten()
print("ANN R2:", r2_score(y_test, y_pred_ann))
print("ANN MAE:", mean_absolute_error(y_test, y_pred_ann))

# -----------------------------------------------------------------------------
# 7. PARAMETER OPTIMISATION
# -----------------------------------------------------------------------------
model_features = X_cols
median_row = df[model_features].median()
row_template = median_row.copy()

def optimise_rf(x):
row = row_template.copy()
row[CONTROLLABLE] = x
X_row = [Link](1, -1)
return -rf_model.predict(X_row)[0]

def optimise_xgb(x):
row = row_template.copy()
row[CONTROLLABLE] = x
X_row = [Link](1, -1)
return -xgb_model.predict(X_row)[0]

def optimise_ann(x):
row = row_template.copy()

95
 row[CONTROLLABLE] = x
X_row = [Link]([Link](1, -1))
return -float(ann_model.predict(X_row))

x0 = [df[c].median() for c in CONTROLLABLE]

res_rf = minimize(optimise_rf, x0, method='L-BFGS-B')

res_xgb = minimize(optimise_xgb, x0, method='L-BFGS-B')
res_ann = minimize(optimise_ann, x0, method='L-BFGS-B')

print("RF Optimal:", res_rf.x, "ROP:", -res_rf.fun)

print("XGB Optimal:", res_xgb.x, "ROP:", -res_xgb.fun)
print("ANN Optimal:", res_ann.x, "ROP:", -res_ann.fun)

# -----------------------------------------------------------------------------
# 8. VISUALISATIONS
# -----------------------------------------------------------------------------
# Correlation heatmap
[Link](figsize=(10, 8))
corr = df[["ROP"] + X_cols].corr()
[Link](corr, cmap="Reds_r", annot=True, fmt=".2f", square=True)
[Link]("Correlation Heatmap Including ROP")
[Link]("correlation_heatmap.png", dpi=300)

# Feature importance plot

[Link](figsize=(8, 6))
rf_importance.plot(kind='bar')
[Link]('Random Forest Feature Importance')
[Link]('rf_feature_importance.png', dpi=150)

# XGBoost Feature importance plot

[Link](figsize=(8, 6))
xgb_importance.plot(kind='bar', color='darkorange')
[Link]('XGBoost Feature Importance')
[Link]('Importance')
plt.tight_layout()
[Link]('xgb_feature_importance.png', dpi=150)
# SHAP summary plot
shap.summary_plot(shap_values_rf, X_train, feature_names=[Link])
[Link]('shap_summary.png', dpi=150)

# Sensitivity analysis
for i, name in enumerate(CONTROLLABLE):
vals = [Link](df[name].min(), df[name].max(), 50)
preds = []
for v in vals:
x = res_rf.[Link]()
x[i] = v
[Link](-optimise_rf(x))

[Link]()
[Link](vals, preds, 'o-')
[Link](name)
[Link]('Predicted ROP')
[Link](f'Sensitivity of ROP to {name}')
[Link](f'sensitivity_{name}.png', dpi=150)
[Link]()
96
 APPENDIX B

MODEL RESULTS SUMMARY

This appendix provides a consolidated summary of the machine learning model
performances and the optimal drilling parameter values identified through the optimisation
process.

B.1 Model Performance Comparison

Table B.1 presents the performance metrics of the three machine learning models evaluated
in this study. The coefficient of determination (R²) and Mean Absolute Error (MAE) were
used to assess model accuracy.

Table B.1: Machine Learning Model Performance Comparison

Model R² Score MAE
Random Forest 0.904 1.738
XGBoost 0.813 2.134
Artificial Neural Network 0.735 2.850

B.2 Optimal Drilling Parameters

Table B.2 presents the optimal controllable drilling parameters identified by each machine
learning model through the L-BFGS-B optimisation algorithm. All three models converged
to identical optimal values, demonstrating consistency across different algorithmic
approaches.

Table B.2: Optimal Drilling Parameters Identified by Each Model

Model WOB (klbs) Flow (GPM) RPM Pred. ROP (m/hr)
Random Forest 6.55 2370.05 158 21.94
XGBoost 6.55 2370.05 158 21.96
ANN 6.55 2370.05 158 19.76

B.3 Feature Importance Rankings

Table B.3 summarises the feature importance rankings from the Random Forest model,
showing the relative contribution of each variable to ROP prediction.

97
Table B.3: Random Forest Feature Importance Rankings

Rank Feature Importance

1 MD (Measured Depth) 0.292
2 MD_OUT 0.214
3 MD_IN 0.184
4 RPM (Rotary Speed) 0.180
5 WOB (Weight on Bit) 0.077
6 TORQUE 0.023
7 JET_IMPACT 0.016
8 FLOW (Flow Rate) 0.015

B.4 Correlation Analysis Results

Table B.4 presents the Pearson correlation coefficients between each feature and the target
variable (ROP), ranked by absolute correlation strength.

Table B.4: Feature Correlation with ROP

Rank Feature Correlation (r)
1 RPM 0.814
2 CUM_TE 0.797
3 MD / MD_IN / MD_OUT 0.760
4 CUM_KREV 0.733
5 KREV -0.579
6 JET_IMPACT / NOZZLE_VELOCITY / FLOW 0.548
7 TORQUE 0.428
8 WOB -0.017

98