Scalable and adaptive injection-production control in reservoirs via a multi-agent reinforcement learning approach

The growing global energy demand necessitates more efficient reservoir management, highlighting the need for advanced injection-production optimization strategies. Traditional methods are computationally expensive and scale poorly, especially when new wells are added and nearby agents must be retrained. These limitations hinder their application in large, heterogeneous reservoirs and reduce their effectiveness for adaptive decision-making. To address these challenges, this study introduces an enhanced multi-agent reinforcement learning (MARL) framework with three key innovations. First, an online-updating surrogate model, based on a simplified U-Net architecture, partially replaces costly numerical simulations, significantly reducing computational overhead. Second, a regional observation construction method encodes relative well positions to capture inter-well dependencies and enhance local decision-making. Third, a self-adaptive, graph-based unified agent design eliminates the need for retraining when new wells are added, ensuring scalability. The proposed framework was validated using both a synthetic “Three-Channel” model and a real oilfield case. Experimental results show substantial improvements in displacement efficiency, delayed water breakthrough, and increased Net Present Value (NPV). Additionally, the framework adapts seamlessly to the addition of new wells without retraining, maintaining high computational efficiency. These results underscore the practical potential of the MARL-based approach as a robust and flexible solution for real-time reservoir management in dynamically evolving oil fields. • A multi-agent RL framework is developed for multi-well control optimization in complex reservoir environments. • A U-Net-based surrogate model is integrated to accelerate simulation and training, enabling efficient policy learning. • A graph-based, adaptive unified agent architecture was designed to capture inter well connectivity and ensure scalability. • The method shows strong performance compared with PSO and GA, achieving higher NPV and more stable control strategies. • Discussion on model generalizability, surrogate limitations, and potential extension to multi-objective optimization.