抄録
Accurate, efficient, and robust reservoir decision-making remains a major challenge as reservoir development becomes increasingly constrained by geological complexity, operational uncertainty, and computational demands. Recent advances in surrogate modeling, mechanism–data fusion, and reinforcement learning have accelerated the transition from conventional simulation-based workflows to intelligent reservoir decision-making. This review systematically synthesizes the theoretical foundations, methodological advances, and engineering applications of intelligent reservoir decision-making, with particular emphasis on mechanism–data fusion modeling and reinforcement learning–driven optimization. Surrogate and deep learning models substantially improve computational efficiency for high-dimensional reservoir prediction, while physics-informed learning enhances predictive accuracy, physical consistency, and cross-scenario generalization by integrating governing reservoir physics with data-driven models. Reinforcement learning further provides a unified framework for sequential optimization of production control, enhanced oil recovery, and integrated CCUS–EOR operations under dynamic uncertainty. The complementary strengths of physics-based and data-driven approaches indicate that their deep integration is emerging as the dominant paradigm for next-generation reservoir management. Remaining challenges include trustworthy model development, robust decision-making under uncertainty, scalable learning from sparse data, and large-scale field deployment. This review establishes an integrated analytical framework linking mechanism–data fusion with reinforcement learning, providing a systematic perspective for developing trustworthy, scalable, and intelligent reservoir decision-making systems.
