A Framework for Predicting Reservoir Recovery Using Machine Learning Models

Thesis Abstract

Reservoir management and hydrocarbon recovery predictions are critical components of optimizing oil field productivity, yet existing models often lack accuracy and adaptability to diverse reservoir conditions. This study aims to develop and validate a comprehensive framework that employs advanced machine learning models to predict reservoir recovery rates with higher precision and robustness. The specific objectives include identifying key geological, petrophysical, and operational variables influencing recovery, designing an integrated machine learning-based modeling approach, and evaluating the predictive performance of selected algorithms in real-world reservoir scenarios. The research adopts a quantitative, exploratory research design to facilitate rigorous modeling and statistical analysis. The study population comprises reservoir datasets from 150 mature oil fields, extracted from integrated subsurface data repositories. A stratified random sampling technique was employed to select a representative sample of 60 reservoirs with comprehensive well, core, seismic, production, and completion data. Data collection relied on secondary sources, including well logs, core analysis reports, production history, and reservoir simulation outputs, ensuring data consistency and quality. Instrumentation involved data extraction tools and standardized data validation procedures to ensure accuracy and completeness. Methodologically, the study employs data preprocessing techniques such as normalization, feature selection via principal component analysis, and data balancing methods to address class imbalance. Several machine learning algorithms—namely, random forest regression, support vector machines, gradient boosting machines, and artificial neural networks—are trained and optimized using grid search and cross-validation. The models’ performance is evaluated based on R-squared, mean absolute error, and root mean squared error metrics through a hold-out validation dataset. To underpin model development, the study leverages the Theory of Statistical Learning and the Reservoir Engineering Theory of Sweep Efficiency to inform variable selection and model interpretation. Preliminary findings are expected to demonstrate that ensemble-based algorithms, particularly gradient boosting and random forest models, outperform traditional deterministic models in predicting reservoir recovery factors with improved accuracy. It is anticipated that variable importance analyses will reveal that parameters such as net pay thickness, porosity, permeability, water saturation, and production rates significantly influence recovery predictions. The models are likely to exhibit robust performance across different reservoir types, suggesting their utility in operational decision-making. This research contributes novel insights into the applicability of machine learning in reservoir engineering, providing a systematic framework for integrating diverse geological and operational data into predictive models. It advances existing knowledge by offering a validated, adaptable, and scalable framework that enhances the accuracy of recovery predictions, thereby supporting more informed reservoir management strategies. Furthermore, the study demonstrates the potential of machine learning to overcome the limitations of traditional empirical and physics-based models, promoting data-driven decision processes in the petroleum industry. The study concludes that machine learning models, when properly trained and validated, can serve as reliable predictive tools for reservoir recovery estimation. It recommends the integration of this framework into reservoir management workflows and advocates for continued advancements in data acquisition and processing techniques to further improve model robustness. Future research should explore the incorporation of unsupervised learning methods for feature extraction and the application of the framework to unconventional reservoirs, such as shale and tight formations, to enhance its generalizability and impact.