Physics-Informed Machine Learning for Geological CO₂ Storage Optimization

Abstract

Geological CO₂ storage (GCS) is increasingly recognized as an indispensable pillar of net-zero emissions strategies, yet optimizing injection design, storage capacity, and containment safety across heterogeneous subsurface formations remains computationally and technically demanding. Conventional numerical reservoir simulation, while physically rigorous, is too computationally expensive to support the exhaustive scenario evaluation and real-time decision-making that large-scale, multi-site GCS deployment requires. Purely data-driven machine learning models, by contrast, can be fast but often violate fundamental physical laws when extrapolated beyond their training distribution, undermining confidence in high-consequence subsurface applications. Physics-informed machine learning (PIML) has emerged as a synthesis of these two paradigms, embedding governing equations—mass and energy conservation, Darcy flow, geomechanical equilibrium—directly into neural network architectures and loss functions to produce models that are both computationally efficient and physically consistent. This review examines the theoretical foundations, architectures, and field applications of PIML for GCS optimization, drawing on recent case studies from heterogeneous depleted gas reservoirs, geothermal-analogue reservoir characterization, and geomechanically informed wellbore stability assessment. It further considers methodological parallels with large-scale AI-driven optimization frameworks developed in public health and healthcare administration, which offer transferable lessons on deploying physics- and rule-constrained AI systems at national scale. The review concludes with a discussion of persistent challenges in training stability, multiphysics coupling, and uncertainty quantification, and proposes a research agenda for advancing PIML toward operational deployment in commercial-scale CO₂ storage projects.