Accelerating multi-well placement optimization with machine learning-based proxy models for CO2 enhanced oil recovery and storage

The urgent need to mitigate greenhouse gas emissions, particularly CO 2 , has spurred extensive research into Geological CO 2 Storage. Utilizing depleted hydrocarbon reservoirs for both CO 2 -enhanced oil recovery and storage presents a multifaceted approach to addressing environmental, economic, and technological challenges. This paper focuses on the strategic placement of wells to optimize both hydrocarbon recovery and CO 2 storage in partially depleted oil reservoirs. Traditional well placement optimization is computationally intensive due to geological complexity, especially under multi-well configurations where each injector and producer is treated as a free decision variable rather than being fixed to predefined patterns, spacing, or areas. To address this challenge, this study introduces a novel data-driven well placement optimization framework that combines physics-guided intelligent data-driven models with evolutionary optimization algorithms. Quality Map-constrained random sampling is employed to generate training datasets, with input features including well geometries, petrophysical parameters, fluid properties, and time of flight that serves as a physics-based feature to improve model generalization. The outputs encompass cumulative fluids production over a fixed period. Various machine learning (ML) methods, including Multiple Perceptron (MLP), Extreme Gradient Boosting (XGBoost), and Linear Regression as a baseline model, are explored, yielding fast and accurate predictions. We investigate the strengths and limitations of ML algorithms across different dataset sizes, where MLP-based and XGBoost-based proxies performed desirably in terms of accuracy and computational efficiency, especially with larger datasets. Integrating these proxies trained upon 1,000 datapoints into a genetic algorithm (GA) effectively optimizes the injector and producer locations in a heterogeneous, three-dimensional reservoir, with net cash flow (NCF) as the objective function. MLP-GA and XGBoost-GA both outperform traditional simulator-based GA optimization in terms of computational efficiency while achieving accurate optimal well locations and NCF values.