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.