A Framework for Predicting Enhanced Oil Recovery Performance Using Machine Learning

Thesis Abstract

Enhancing oil recovery efficiency remains a critical challenge within the petroleum industry, primarily due to the complex and nonlinear relationships between reservoir properties, recovery methods, and operational parameters. Traditional predictive techniques often rely on deterministic models that lack adaptability and may fail to accurately forecast performance under varying reservoir conditions. This study aims to develop a robust, data-driven framework leveraging machine learning algorithms to predict enhanced oil recovery (EOR) performance with higher precision and operational relevance. The specific objectives include identifying key reservoir and operational parameters affecting EOR outcomes, training and validating machine learning models for performance prediction, and establishing a practical decision-support system to optimize EOR strategies. The research adopts a quantitative, exploratory research design grounded in supervised learning methodologies. The population comprises 350 reservoir performance datasets collected from offshore and onshore fields with diverse geological characteristics and EOR techniques, such as chemical flooding, thermal recovery, and gas injection. A stratified random sampling technique is employed to select a representative sample of 200 datasets, ensuring adequate variation across reservoir types, fluid properties, and EOR methods. Data collection involves leveraging historical production records, geophysical logs, petrophysical measurements, and operational reports obtained from industry databases and field operators. The primary data collection instruments include structured data extraction forms and validated digital data repositories, with additional validation through cross-referencing multiple data sources. Model development involves preprocessing procedures such as normalization, feature selection using recursive feature elimination, and data balancing techniques to address class imbalance issues. Several machine learning algorithms—namely Random Forests, Gradient Boosting Machines, and Support Vector Machines—are trained using cross-validation to prevent overfitting and to optimize hyperparameters via grid search. Model evaluation employs performance metrics including mean squared error (MSE), R-squared, and area under the receiver operating characteristic curve (AUC-ROC) to determine predictive accuracy and robustness. The analytical framework is anchored in the Reservoir Engineering Decision Theory and the System Dynamics Theory, which provide the conceptual basis for integrating reservoir properties with operational data into a dynamic predictive model. The study also explores the application of explainable AI techniques, such as SHAP (SHapley Additive exPlanations), to interpret model predictions and identify influential parameters. The expected outcome of this research is a validated machine learning-based predictive model capable of accurately estimating EOR performance for different reservoir contexts, thus assisting engineers in strategic decision-making. It is anticipated that the models will highlight critical parameters—such as reservoir heterogeneity, fluid saturation levels, injection rates, and chemical concentrations—that significantly influence recovery efficiency. Furthermore, the study aims to demonstrate the model’s superior predictive capability over conventional empirical or semi-empirical methods, as evidenced by improved performance metrics (e.g., a minimum of 0.85 R-squared value and 0.90 AUC-ROC). This research contributes to the existing body of knowledge by integrating advanced machine learning techniques within the domain of petroleum reservoir management, facilitating a shift toward more intelligent, data-centric EOR planning. The created framework provides a foundation for future development of real-time predictive tools and decision support systems, thus promoting reservoir management strategies that optimize recovery while minimizing operational risks and costs. Based on the findings, recommendations include adopting the predictive framework for routine reservoir evaluation, further expanding the dataset with real-time monitoring data, and exploring hybrid models that combine machine learning with physics-based simulations for enhanced predictive power. The study concludes that an evidence-based, machine learning-driven framework offers significant advantages for optimizing enhanced oil recovery performance, ultimately contributing to more sustainable and economically viable petroleum extraction practices.