SPE-221074-MS

Enhancing Real-Time Drilling Efficiency: Mechanism of ROP Prediction

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/24ATCE/24ATCE/D021S026R005/3604234/spe-221074-ms.pdf by Schlumberger Oilfield UK Plc user on 22 March 2025

Models and Novel Optimization Strategies in Chinese Oilfields

X. Z. Song, China University of Petroleum, Beijing, College of Artificial Intelligence, Beijing, China / National Key
Laboratory of Petroleum Resources and Engineering, China University of Petroleum, Beijing, Beijing, China;
R. Zhang, China University of Petroleum, Beijing, College of Artificial Intelligence, Beijing, China / National Key
Laboratory of Petroleum Resources and Engineering, China University of Petroleum, Beijing, Beijing, China;
Z. P. Zhu, China University of Petroleum, Beijing, College of Artificial Intelligence, Beijing, China / National Key
Laboratory of Petroleum Resources and Engineering, China University of Petroleum, Beijing, Beijing, China; Y. Wu,
CNOOC Research Institute Co., Ltd, Beijing, China; Z. Y. Pang, China University of Petroleum, Beijing, College
of Artificial Intelligence, Beijing, China / National Key Laboratory of Petroleum Resources and Engineering, China
University of Petroleum, Beijing, Beijing, China / CNOOC Research Institute Co., Ltd, Beijing, China; G. S. Li,
China University of Petroleum, Beijing, College of Artificial Intelligence, Beijing, China / National Key Laboratory
of Petroleum Resources and Engineering, China University of Petroleum, Beijing, Beijing, China; C. K. Zhang,
China University of Petroleum, Beijing, College of Artificial Intelligence, Beijing, China / National Key Laboratory of
Petroleum Resources and Engineering, China University of Petroleum, Beijing, Beijing, China

Copyright 2024, Society of Petroleum Engineers DOI 10.2118/221074-MS

This paper was prepared for presentation at the SPE Annual Technical Conference and Exhibition held in New Orleans, Louisiana, USA, 23 - 25 September 2024.

This paper was selected for presentation by an SPE program committee following review of information contained in an abstract submitted by the author(s). Contents
of the paper have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect
any position of the Society of Petroleum Engineers, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written
consent of the Society of Petroleum Engineers is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300 words; illustrations may
not be copied. The abstract must contain conspicuous acknowledgment of SPE copyright.

Abstract
Drilling parameter optimization constitutes a pivotal technique for expediting the intricate processes
involved in oil and gas drilling operations. Presently, optimization strategies predominantly hinge on
statistical data analysis or data-driven models and optimization algorithms. However, these methods
often overlook the real-time changes occurring in the downhole rock-breaking environment. Additionally,
achieving high precision with data-driven rate of penetration (ROP) models proves challenging, as they
may struggle to accurately reflect the intricate mechanisms involved in rock-breaking. Moreover, the
inherent randomness and uncertainty associated with optimization algorithms pose difficulties in ensuring
the consistent acceleration of the drilling process.
This study introduces areal-time targets-based intelligent optimization and decision-making system
tailored specifically for field drilling operations.The system workflow comprises data preprocessing,
predictive optimization, and interpretive decision-making stages. In the crucial modeling and prediction
phase, we developed a highly accurate ROP prediction model using advanced machine learning techniques
and mechanistic empirical constraints. By employing three model updating mechanisms alongside a 3D
heat map visualization technique, the most robust current ROP model can be selected, thereby enhancing
2 SPE-221074-MS

the adaptability of the model to dynamic downhole environment. This process establishes a solid foundation
for achieving specific ROP targets.
In the optimization decision stage, this work proposes the optimal path decision process with specific
target parameter optimization and small parameter fluctuation. The process begins by determining the ROP
enhancement target for the next stage based on the average ROP of the current formation. Subsequently, the
eligible combinations of drilling parameters undergo a preliminary screening. Afterwards, the optimal path
decision-making process is informed by the consideration of parameter fluctuation variance and mechanical

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/24ATCE/24ATCE/D021S026R005/3604234/spe-221074-ms.pdf by Schlumberger Oilfield UK Plc user on 22 March 2025

specific energy as a combined evaluation indicator. This approach ensures that drilling parameters are
adjusted with utmost efficiency, all while meeting the predetermined ROP enhancement target. Moreover,
it facilitates stable speed enhancement, minimizes drilling parameter fluctuations, and effectively manages
complex working conditions.
The application of the system in the Huizhouoilfield in China demonstrates the stability and effectiveness
of the method. The optimized drilling parameters exhibit low fluctuations, with a maximum reductionin
parameter variance of 63.8%compared to the original tuned parameters, ensuring robust real-time
operability. Notably, the method achieves asmooth speed-up of 15% to 25%, while minimizing mechanical
specific energy by a maximum of 19% and reducing bit energy loss. This study contributes novel insights
into real-time drilling parameter optimization, with a dynamic adaptation mechanism poised to incorporate
emerging drilling data streams.

Introduction
Improving drilling efficiency is crucial for reducing overall drilling costs and time, as well as maximizing
equipment utilization (Mal et al. 2023).One of the most important methods for improving drilling efficiency
is the optimization of drilling parameters. The key to this approach lies in developing accurate drilling
models to characterize the drilling process and employing optimization methods to make optimal decisions
regarding drilling parameters (Zhang et al. 2022; Pei et al. 2023).
In the drilling process, the rate of penetration (ROP) serves as a direct measure of drilling efficiency
and a primary parameter reflecting the interaction with geological formations, making it a focal point in
drilling engineering research. Numerous theoretical analyses and laboratory experiments have aimed to
develop ROP prediction models by incorporating factors such as bit properties, hydraulic parameters, and
engineering parameters. This has led to the establishment of several typical ROP equations (Bourgoyne and
Young 1974; Walker et al. 1986). However, many of these models rely on empirical parameters derived
from laboratory settings or simplified parameters, often resulting in prediction accuracies that fail to meet
field requirements.
With the rise of artificial intelligence (AI) technology across various fields, scholars worldwide have
increasingly turned to machine learning methods within artificial intelligence to develop accurate and
efficient intelligent ROP prediction models. These models include Support Vector Machine (SVM)(Ren,
Huang, and Gao 2022), Random Forests (Hegde et al. 2017), Artificial Neural Networks (ANNs) (Ashrafi
et al. 2019; Zhao et al. 2020), Gradient Boosting Trees (Zhou et al. 2022), and Long Short-Term
Memory networks (LSTM) (Ding et al. 2023; Safarov, Iskandarov, and Solomonov 2022). Comparative
analyses of these intelligent models have been conducted to assess their accuracy and stability in ROP
prediction.Overall, while physical ROP equations adhere to drilling mechanisms, they often lack the
capability to represent complex relationships accurately. On the other hand, high-precision static intelligent
ROP prediction models struggle to adapt to the real-time fluctuations of the downhole rock-breaking
environment, posing significant challenges to their stability, real-time applicability, and generalization.
Zhang et al. (2023) proposed to update the ROP prediction model in real time using the sliding window
method, which significantly improved the effect compared to the initial model. Ding et al. (2023)utilized a
transfer learning approach to fine-tune the ROP prediction model,enhancing the model's ability to generalize
SPE-221074-MS 3

across wells.The primary challenge in constructing an effective system lies in developing an ROP prediction
model that balances real-time performance, accuracy, and adherence to drilling mechanisms.
In drilling parameter optimization,numerous companies and universities, both domestically and
internationally, have initiated research on AI-based drilling parameter optimization systems (Li et al.
2022). The current approach involves establishing various optimization objectives and using optimization
algorithms to determine the best parameter combinations. These methods can be categorized into two types
based on the optimization goals: single-objective optimization and multi-objective optimization.Single-

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/24ATCE/24ATCE/D021S026R005/3604234/spe-221074-ms.pdf by Schlumberger Oilfield UK Plc user on 22 March 2025

objective optimization typically aims to maximize the ROP or minimize drilling costs. Some researchers
have also coupled ROP with mechanical specific energy (MSE) to form a comprehensive evaluation metric
for optimization (Hassan, Elkatatny, and Al-Majed 2020). On the other hand, multi-objective optimization
methods often target both ROP and MSE. For example, Zhang et al. (2023) aimed to minimize MSE
and maximize ROP, employing the Non-dominated Sorting Genetic Algorithm-II and TOPSIS algorithms
for multi-objective optimization and using real-time data streams to update the models.These parameter
recommendation methods primarily rely on the search capabilities of optimization algorithms. However, the
randomness and uncertainty inherent in these algorithms can hinder the effective acceleration of the drilling
process. Specifically, single-objective optimization focusing solely on ROP often results in highly variable
recommended drilling parameters, leading to complex downhole conditions. Multi-objective optimization
methods, influenced by multiple unbalanced goals, typically achieve only marginal improvements in speed,
or none at all, making existing drilling parameter optimization models less applicable in the field.Therefore,
improving current optimization strategies to achieve stable and significant speed enhancements is another
critical challenge in the field of drilling.
This study establishes an intelligent real-time optimization system for drilling parameters, achieving
dynamic characterization of the ROP and adaptive real-time optimization of drilling parameters
through mechanism-data fusion. This provides a technological foundation for the rational and
scientific optimization of drilling parameters.Specifically, in terms of dynamic parameter prediction, a
mechanistically optimized ROP prediction model was developed. For drilling parameter optimization, an
adaptive real-time optimization strategy based on a stepwise acceleration approach was employed. This
strategy comprehensively considers parameter fluctuations and energy consumption to enhance drilling
efficiency.The research provides effective technical support for accurate and efficient ROP prediction and
offers practical guidance for field engineers to adjust drilling parametersefficiently and scientifically.

Real-time Optimization System Workflow

The adaptive real-time optimization system for drilling parameters is primarily composed of four modules:
data service, modeling and prediction, model updating, and drilling parameter optimization decision-making
(Fig. 1). Each module performs specific functions to enable the interactive transfer of data streams.
4 SPE-221074-MS

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/24ATCE/24ATCE/D021S026R005/3604234/spe-221074-ms.pdf by Schlumberger Oilfield UK Plc user on 22 March 2025

Figure 1—The structure of drilling parameters autonomous real-time optimization system.

Data Service
The data service module encompasses three main functions: dataacquisition,data processing and storage.
The data acquisitionand processing function are responsible for integrating real-time data from on-site
drilling operations and systematically processing it. This involves parsing and processing the data based on
oilfield transmission protocols, selection of parameters in the drilling state using logical criteria, followed by
missing value filling, outlier identification, and generation of clean datasets. To mitigate the adverse effects
of sensor signal noise on model development, this module also employs filtering techniques to reduce noise
in the raw signals. The data storage function collects information from neighboring wells within a specific
block, including geological and drilling designs, drilling, mud logging and well logging data, to prepare for
the development of intelligent models.

Modeling and Prediction

First, the collected data from multiple wells are used to classify the formations and establish ROP
prediction models for different formations to reveal the rock-breaking relationships under specific formation
conditions. Next, a knowledge-guided intelligent ROP prediction model is constructed, along with a MSE
calculation model. During model training, a loss function is developed based on the original mean square
error, incorporating sample error and ROP equation constraints to enhance model stability and alignment
with drilling mechanism. Additionally, real-time monitoring of prediction errors triggers the adaptive
activation of the model updating module.

Model Updating
Considering the complex and variable nature of formations during drilling, static intelligent models may not
be suitable for different formation conditions. Thus, the updating module establishes three mechanisms for
model updates. Using processed real-time data streams, the ROP prediction model is dynamically updated
to improve robustness against varying lithologies. Specifically, 3D visualizations ofWeight on Bit (WOB),
Rotary Speed (RPM), ROP variations under three methods are employed to refine the model updating
mechanism. This approach excludes models that perform well locally but have multiple global extrema,
thereby enhancing the extrapolation capability of the intelligent models.
SPE-221074-MS 5

Real-time Optimization
The real-time optimization module leverages the established a real-time accurate prediction modelfor
ROP to achieve stable ROP improvement, addressing the limitations of traditional parameter optimization
methods that may result in minimal or no speed gains. By automatically setting acceleration intervals
and utilizing an improved optimization algorithm, this module ensures a steady increase in ROP while
minimizing parameter fluctuations and optimizing MSE, thereby ensuring safe and efficient drilling
operations.

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/24ATCE/24ATCE/D021S026R005/3604234/spe-221074-ms.pdf by Schlumberger Oilfield UK Plc user on 22 March 2025

Model Building and Updating
AI Model of ROP Prediction
In recent years, many researchers have extensively utilized machine learning methods to develop ROP
prediction models. Soares and Gray (2019), Bajolvand et al. (2022), and Zhang (2023) have compared
the accuracy and stability of Random Forests, ANN, and SVM for real-time ROP prediction. The main
advantage of Random Forests is their prediction stability, as their output consistently remains within a
controllable range, though their accuracy is lower compared to neural networks. Neural networks offer
higher prediction accuracy but often produce unreasonable anomaly predictions (Guo et al. 2021).
To address these issues, this section introduces a Decision Tree Network under Physical Constraints (PC-
DsNet) aimed at achieving accurate real-time ROP prediction. This model effectively combines the stability
of tree models with the representational capability of deep neural networks (Fig. 2). Each layer of PC-DsNet
can be viewed as a basic decision tree, corresponding to different steps of a fundamental decision tree,
including feature selection, threshold judgment, and conditional judgment.

Figure 2—The model structure of PC-DsNet.

The entire model is a weighted combination of multiple basic decision trees, with each tree selecting
different features as branching criteria. Through a series of mask layers, feature encoders, and activation
functions, the model ultimately outputs high-dimensional vectors. Each layer's high-dimensional vector
represents the weighted feature values under different drilling condition decisions that impact the final ROP
prediction results. Over multiple iterations, the model parameters are progressively optimized, integrating
the optimization strategy of gradient boosting decision tree.

Model Developmentunder Physical Constraints

Model Training. When using intelligent models for prediction, it is challenging to avoid significant
prediction biases due to the inherent weaknesses in their mechanistic representation and generalization
abilities. This section addresses these issues by constructing a knowledge-guided intelligent prediction
model that accounts for prediction errors under different conditions. The total training loss of the model
comprises the mean squared error loss, physical constraint loss, and weight loss.
By examining the rock-breaking drilling equations, specifically using a modified Young's ROP equation
(as shown in Eq. 2), we observe a clear positive relationship between WOB, RPM and ROP. Based on
6 SPE-221074-MS

this, we establish inequality constraints. To better integrate mechanistic knowledge into the intelligent
model, we transform the nonlinear programming problem with inequality constraints into an unconstrained
nonlinear programming problem, embedding the rock-breaking mechanism into the model training process,
as illustrated in Eq.4.
The weight loss term assigns adaptive weights to each sample (see Eq. 5), giving higher weights to
samples with larger errors during model training. This approach ensures that the model pays more attention
to such samples, improving its overall accuracy and robustness.

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/24ATCE/24ATCE/D021S026R005/3604234/spe-221074-ms.pdf by Schlumberger Oilfield UK Plc user on 22 March 2025

(1)
(2)

(3)

(4)

(5)

(6)

where Ltotal is the total loss, Lmodel is the mean squared error loss, Lknowledge is the mechanistic equation loss,
Lweight is the sample prediction error loss, ROPi is the predicted value of the model output and ROPi is the
true value.
Model Updating. In terms of model updating, continuously updating the ROP prediction model can
enhance its predictive accuracy and generalization ability, enabling the intelligent model to dynamically
adapt to real-time downhole environments and rock-breaking conditions. This study establishes three
distinct model updating mechanisms to ensure the ROP intelligent prediction model maintains high
accuracy. The three updating mechanisms are as follows:
1. Fine-tuning Update: The model undergoes offline training based on historical formation data and is
fine-tuned using real-timedrilling data.
2. Incremental Update: The model is incrementally trained and updated by continuously incorporating
real-time drilling data.
3. Online Update: The model is trained and updated online using a fixed amount of real-time drilling
data.
Fig. 3 shows the process of the model utilizing the real-time data stream to update the model for the
prediction task. Based on this module, the system can effectively utilize the real-time data flow, so that
the model performance is always kept in the optimal state.Notably, these model updating mechanisms are
activated when significant performance deviations are detected, ensuring that the updated models avoid local
overfitting and multiple global extrema, thereby maintaining global mechanistic consistency. By employing
3D visualization methods to plot trend graph ofWOB-RPM-ROP, we can select the updated models that best
align with drilling mechanistic experience. This approach lays the foundation for the rational optimization
of drilling parameters.
SPE-221074-MS 7

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/24ATCE/24ATCE/D021S026R005/3604234/spe-221074-ms.pdf by Schlumberger Oilfield UK Plc user on 22 March 2025

Figure 3—Prediction process under the dynamic updating method.

Fig. 4 illustratesthe 3D cloud diagram of WOB-RPM-ROP to observe the relationship among ROP, WOB,
and RPM. By analyzing the rock-breaking mechanism in drilling, we exclude 3D models with multiple
local extremum points, thus avoiding local optimums in parameter recommendation and ensuringthat the
parameter optimization is more reasonable and effective. It can be seen that the ROP predictive model
under the incremental updating mechanism exhibits stronger mechanistic fidelity, with ROP showing an
approximately proportional relationship to WOB and RPM.

Figure 4—Parameter optimization results for well B. (a) Fine-tuning update, (b) Incremental update,(c) Online update.

In contrast, the other two updating models demonstrate certain limitations.The online updating method
tends to model and fit the current real-time data, resulting in a relatively small data space. As seen in Fig.4(c),
the 3D plot displays multiple troughs and model failures at high RPM, indicating that the current model
fails to accurately capture the true mapping relationship of the ROP. This leads to scenarios where high
prediction accuracy does not translate to reasonable parameter recommendations.Compared to the online
updating method, the fine-tuning updating method shows some improvement in capturing the relationships
among multiple variables. However, it still exhibits several local extremum points and is less sensitive to
boundary conditions. Anomalies such as instances where WOB is zero while ROP is greater than zero occur
at high RPM.Overall, the 3D cloud chart of the ROP predictive model based on the incremental updating
method aligns more closely with mechanism. This approach holds significant practical application value
for the optimization of drilling operations.
Evaluation indicators. In terms of assessing and comparing the performance of the model, two common
error functions are chosen for the experimental evaluation metrices: the mean absolute percentage error
(MAPE) and Empirical Correlation Coefficient (COR), as shown in Eq. (7)-(8). Where MAPE is used to
measure the average relative error of n samples, COR indicates the correlation between predicted values
and actual values, with values closer to 1 indicating better performance. The model is favorably evaluated
by considering its performance comprehensively across these metrics.
8 SPE-221074-MS

(7)

(8)

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/24ATCE/24ATCE/D021S026R005/3604234/spe-221074-ms.pdf by Schlumberger Oilfield UK Plc user on 22 March 2025

where yi is the measured value ypre is the predicted value of the model output and is the mean value of
the parameters.

Real-time Optimization
This study uses WOB and RPM as decision parameters and employs an adaptive interval updating strategy
to reduce parameter fluctuations and lower MSE. Based on an intelligent drilling model, a multi-objective
optimization framework for drilling parameters under a stepwise acceleration strategy is constructed (see
Fig.5). The specific process is as follows:
1. Establish Optimization Functions and Goals: A real-time ROP prediction model is developed, and
the obtained ROP predictions are input into the MSE calculation model for parameter coupling. An
acceleration target range is established and set as the optimization objective.
2. Optimization Boundary Constraints: To ensure the scientific validity of the prediction model,
feasible boundary conditions must be considered during constrained parameter optimization. The
WOB and RPM ranges should account for the load-bearing capacity of drilling equipment and the
safety-limited data space, ensuring the reliability and scientific validity of the optimization results.
Additionally, this study constructs a set of constraint conditions based on historical neighboring well
data, forming the initial optimization parameter space.
3. Autonomous Update of Acceleration Targets: Based on real-time data from the current drilled
section, continuously monitor system performance and adaptively adjust as necessary. Real-time data
feedback is used to dynamically update targets and algorithms, ensuring the system consistently
operates in an optimal state with appropriately set acceleration intervals.
4. Parameter Optimization: To address the common issues of local optima and low convergence
precision in existing optimization algorithms, this study introduces an adaptive weighting method
to improve the Whale Optimization Algorithm (WOA). This enhancement boosts the algorithm's
local search capabilities, enabling the rapid solution of the aforementioned optimization problem. The
improved population algorithm is utilized to solve the initial optimization parameter set from step 2,
generating a candidate set of drilling parameters as the basis for optimal decision-making.
5. Optimal Path Decision: By comprehensively considering MSE and the amplitude of drilling
parameter fluctuations, a weighted decision is made from multiple feasible solutions to determine the
optimal path, as shown in Eq.9.
(9)
where α represents the MSE weighting coefficient, and β represents the parameter distance weighting
coefficient. The Euclidean distance Pdis between optimized WOB, optimized RPM, and current WOB,
current RPM is calculated for each solution. The feasible parameter set obtained from step 4is then
used to decide the optimal parameter recommendation, achieving smooth and phased acceleration.
6. Continuous Learning and Dynamic Adjustment: After the current parameter recommendation is
implemented, the adjustment effects and prediction errors are recorded for feedback calibration. This
SPE-221074-MS 9

process allows real-time adaptation to the complex and changing field environment, maintaining the
model's optimal performance.

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/24ATCE/24ATCE/D021S026R005/3604234/spe-221074-ms.pdf by Schlumberger Oilfield UK Plc user on 22 March 2025

Figure 5—A multi-objective optimization framework for drilling parameters under stepped ramp-up strategy.

By incorporating these steps, the proposed adaptive real-time optimization system not only enhances
drilling efficiency and stability but also ensures that the optimization outcomes align closely with practical
field operations.

Application of Real-time Optimization System for Drilling Parameters

The system is applied to the drilling A and B wells in Huizhou oilfield to verify the rationality and validity
of the drilling parameter optimization system.

Real Time Data Processing

Based on the data acquisition and processing module, we collect 10 logging parameters, including
MeasuredDepth (MD), BitDepth (DB), WOB, Hook Load (HL), RPM, Torque (TOB),Mud Flow In (MFI),
Mud Density In (MDIA), Stand PipePressure (SPP), and ROP. Logical rules are applied to filter the normal
drilling data for predictive modeling. To avoid the impact of highly correlated parameters on the model's
predictions, we retained WOB as an input parameter and excluded HL. The processed subset of input data
is presented in Table 1.

Table 1—Model input parameters.

MD/ m WOB/ t RPM/ rpm TOB/ kN·m SPP/ MPa MFI/ L·min-1 MDIA/ g·cm-3

2694.26 9.94 150 19.88 17.34 4052.55 1.3

2694.34 9.41 150 22.95 17.00 4008.26 1.3
2694.43 9.89 151 20.15 17.04 4052.55 1.3
2694.54 9.64 150 21.83 17.47 4052.55 1.3
2694.61 9.98 151 21.79 17.73 4052.55 1.3
2694.7 9.98 150 23.19 17.21 4052.55 1.3
2694.77 9.98 151 21.27 16.88 4052.55 1.3
2694.88 11.26 149 23.82 17.35 4030.41 1.3
2694.97 11.38 150 23.67 17.3 4030.41 1.3
2695.04 11.55 150 23.77 17.18 4052.55 1.3
10 SPE-221074-MS

Case 1: Well A
During the field tracking of well A'sthe third section drilling from 3020m to 3688m, intermittent
optimization of drilling parameters was carried out. As shown in Fig.6, the COR of the mechanical
drilling speed prediction model reached 0.93, indicating that the model effectively established the mapping
relationship between input parameters and ROP. The average ROP increased by 15.0%, and MSE decreased
by 12.1%. The recommended average WOB increased from 13.5t to 15.5t.The optimized standard deviation
of drilling pressure was reduced from 1.3t to 0.8t, a reduction of 37.3%.

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/24ATCE/24ATCE/D021S026R005/3604234/spe-221074-ms.pdf by Schlumberger Oilfield UK Plc user on 22 March 2025

Figure 6—Parameter optimization results for well A.

SPE-221074-MS 11

At 3020m, the recommended WOB increased from 12t to 14t, with the optimized average ROP rising
from 27.9m/h to 32.8m/h, an improvement of 17.4%. As illustrated in Fig. 7, at 3265m, the system considers
the current depth of cut insufficient and recommendsWOB increased from 14t to 16t, with the average ROP
increasing from 19.5m/h to 23.5m/h, an improvement of 20.5%. In the 3592m-3622m section, field tracking
revealed that increasing the original average WOB from 13t to 15t resulted in the average ROP increasing
from 13.8m/h to 18.7m/h, consistent with the anticipated acceleration effect.

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/24ATCE/24ATCE/D021S026R005/3604234/spe-221074-ms.pdf by Schlumberger Oilfield UK Plc user on 22 March 2025

Figure 7—Parameter optimization results for 3245m-3292m well section.

Case 2 Well B
During the field monitoring of Well B's third section drilling from 3574m to 4036m, intermittent
optimization of drilling parameters was performed. As shown in Fig.8, the predictive model for the rate
of penetration (ROP) achieved theCOR of 0.90 and a MAPE of 18.3%, effectively mapping the input
parameters to the drilling rate. The average ROP increased from 10.5m/h to 12.9m/h, an improvement of
23%, while MSE decreased by 19.3%. The recommended average WOBwas increased from 14.5t to 17t,
and the recommended average RPM was increased from 72rpm to 94rpm. The optimized drilling pressure
variance decreased from 3.6t to 1.3t, a decrease of 63.8%. Since the parameter optimization in the well
section after 3888m was mainly aimed at increasing the rotational speed, the optimized rotational speed
variance increased from 13.1rpm to 16.7rpm, an increase of 27.1%.
12 SPE-221074-MS

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/24ATCE/24ATCE/D021S026R005/3604234/spe-221074-ms.pdf by Schlumberger Oilfield UK Plc user on 22 March 2025

Figure 8—Parameter optimization results for well B.

At 3643m, the recommended WOB increased from 13t to 18t, with the average ROP expected to rise
from 10.9m/h to 18.0m/h, and the MSE was reduced by 31.5%. Monitoring of the well from 3652m to
3682m showed that using an average WOB of 16t, the average ROP reached 17.4m/h, aligning with the
anticipated acceleration effect. From 3912m to 4036m, at a depth of 4006m, the recommended parameters
were adjusted from 16t, 60rpm to 17t, 120rpm, resulting in an increase in average ROP from 8.6m/h to
13.53m/h. Additionally, observations of a neighboring well in the same formation using a WOB of 15t and
anRPM of 120rpm achieved an ROP of 12.6m/h, validating the recommended parameters of this system.
Model Updating. At a depth of 3899m, a prediction deviation exceeding the set threshold was observed,
prompting the activation of the model updating module. This module adjusted the model's weight parameters
through the updating mechanisms to quickly adapt to the current formation conditions.
SPE-221074-MS 13

Conclusion
The adaptive real-time optimization system for drilling parameters based on dynamic mechanistic updates
proposed in this paper achieves dynamic and accurate characterization of ROP and rational scientific
optimization of drilling parameters under mechanistic updates, effectively enhancing drilling efficiency and
better aligning optimization results with field drilling practices.
Firstly, the system establishes a data connection protocol with the field database, completing real-

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/24ATCE/24ATCE/D021S026R005/3604234/spe-221074-ms.pdf by Schlumberger Oilfield UK Plc user on 22 March 2025

time data storage and cleaning. Subsequently, a real-time ROP prediction model constrained by rock-
breaking mechanics is constructed, which comprehensively considers mechanistic errors and loss errors.
By incorporating the proposed mechanistic optimization updating mechanism, the system achieves real-
time accurate ROP prediction, enhancing the model's mechanistic consistency, stability, and accuracy. The
stepwise acceleration-based optimization and decision-making method for drilling parameters mitigates
parameter adjustment fluctuations compared to traditional optimization methods, effectively reducing the
likelihood of complex downhole conditions. Field trials in the Huizhou oilfield demonstrate that the COR of
ROP prediction is 0.93 and 0.90, respectively, with acceleration effects consistently within the target range
of 15-25%. The system also achieves a maximum reduction in MSE of 19%, enabling stable acceleration
of drilling operations.
Future research will consider the impact of drilling risks and complex downhole conditions on drilling
parameter optimization, aiming to further refine the intelligent optimization decision system for drilling
parameters.

Acknowledgements
The authors gratefully acknowledge the support of National Science Fund for Distinguished Young Scholars
(52125401) andNational Key Research and Development Project of China (2019YFA0708300).

References
Ashrafi, Seyed Babak, Anemangely, M., Sabah, M. et al 2019. Application of Hybrid Artificial Neural Networks for
Predicting Rate of Penetration (ROP): A Case Study from Marun Oil Field. Journal of Petroleum Science and
Engineering 175 (April):604–623. https://doi.org/10.1016/j.petrol.2018.12.013.
Bajolvand, M., Ramezanzadeh, A., Mehrad, M.et al 2022. Optimization of Controllable Drilling Parameters Using a Novel
Geomechanics-Based Workflow. Journal of Petroleum Science and Engineering 218 (November):111004. https://
doi.org/10.1016/j.petrol.2022.111004.
Bourgoyne, A.T., and Young, F.S. 1974. A Multiple Regression Approach to Optimal Drilling and Abnormal Pressure
Detection. Society of Petroleum Engineers Journal 14 (04): 371–384. https://doi.org/10.2118/4238-PA.
Ding, J.X., Zhang, R., Wen, X. et al 2023. Interpretable Feature Construction and Incremental Update Fine-Tuning Strategy
for Prediction of Rate of Penetration. Energies 16 (15): 5670. https://doi.org/10.3390/en16155670.
Guo, X.W., Quan, Y.H., Zhao, H. et al 2021. TabGNN: Multiplex Graph Neural Network for Tabular Data Prediction.
arXiv. http://arxiv.org/abs/2108.09127.
Hassan, A., Elkatatny, S. and Al-Majed, A. 2020. Coupling Rate of Penetration and Mechanical Specific Energy
to Improve the Efficiency of Drilling Gas Wells. Journal of Natural Gas Science and Engineering 83
(November):103558. https://doi.org/10.1016/j.jngse.2020.103558.
Hegde, C., Daigle, H., Millwater, H. et al 2017. Analysis of Rate of Penetration (ROP) Prediction in Drilling Using
Physics-Based and Data-Driven Models. Journal of Petroleum Science and Engineering 159 (November):295–306.
https://doi.org/10.1016/j.petrol.2017.09.020.
Li, G.S., Song, X.Z., Tian, S.C. et al 2022. Intelligent Drilling and Completion: A Review. Engineering 18
(November):33–48. https://doi.org/10.1016/j.eng.2022.07.014.
Mal, A., Oedegaard, S.I., Helgeland, S. et al 2023. Drilling Parameter Optimization in Real-Time. In Day 1 Tue, March
07, 2023, D011S007R005. Stavanger, Norway: SPE. https://doi.org/10.2118/212484-MS.
Pei, Z.J., Song, X.Z., Wang, H.T. et al 2023. Interpretation and Characterization of Rate of Penetration Intelligent
Prediction Model. Petroleum Science, October. https://doi.org/10.1016/j.petsci.2023.10.011.
14 SPE-221074-MS

Ren, C.J., Huang, W.J., and Gao, D.L. 2022. Predicting Rate of Penetration of Horizontal Drilling by Combining Physical
Model with Machine Learning Method in the China Jimusar Oil Field. SPE Journal, September, 1–24. https://
doi.org/10.2118/212294-PA.
Safarov, A., Iskandarov, V., and Solomonov, D. 2022. Application of Machine Learning Techniques for Rate of Penetration
Prediction. In SPE Annual Caspian Technical Conference, D021S013R002. Nur-Sultan, Kazakhstan: SPE. https://
doi.org/10.2118/212088-MS.
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 (January):934–959. https://doi.org/10.1016/

Downloaded from http://onepetro.org/SPEATCE/proceedings-pdf/24ATCE/24ATCE/D021S026R005/3604234/spe-221074-ms.pdf by Schlumberger Oilfield UK Plc user on 22 March 2025

j.petrol.2018.08.083.
Walker, B.H., Black, A.D., Klauber, W.P. et al 1986. "Roller-Bit Penetration Rate Response as a Function of Rock
Properties and Well Depth." SPE Annual Technical Conference and Exhibition. https://doi.org/10.2118/15620-MS.
Zhang, C.K., Song, X.Z., Liu, Z.H. et al 2023. Real-Time and Multi-Objective Optimization of Rate-of-Penetration
Using Machine Learning Methods. Geoenergy Science and Engineering 223 (April):211568. https://doi.org/10.1016/
j.geoen.2023.211568.
Zhang, C.K., Song, X.Z., Su, Y.N. et al 2022. Real-Time Prediction of Rate of Penetration by Combining Attention-Based
Gated Recurrent Unit Network and Fully Connected Neural Networks. Journal of Petroleum Science and Engineering
213 (June):110396. https://doi.org/10.1016/j.petrol.2022.110396.
Zhao, Y., Noorbakhsh, A., Koopialipoor, M. et al 2020. "A New Methodology for Optimization and Prediction of Rate
of Penetration during Drilling Operations." Engineering with Computers 36 (2): 587–595. https://doi.org/10.1007/
s00366-019-00715-2.
Zhou, F., Fan, H.H., Liu, Y.H. et al 2022. Application of Xgboost Algorithm in Rate of Penetration Prediction with
Accuracy. In International Petroleum Technology Conference, D012S111R002. Riyadh, Saudi Arabia: IPTC. https://
doi.org/10.2523/IPTC-22100-MS.