Physics-Informed AI for Subsurface Natural Gas Storage

Abstract

Subsurface natural gas storage in depleted reservoirs, saline aquifers, and salt caverns requires accurate, computationally efficient modeling of multiphase flow, pressure evolution, and geomechanical response across repeated injection-withdrawal cycles. Conventional numerical reservoir simulation provides high physical fidelity but is computationally expensive to run across the many geologic realizations and operational scenarios needed for robust site characterization and operational optimization, while purely data-driven machine learning models, though fast, often violate fundamental conservation laws when applied beyond their training distribution—a particular concern for the strongly nonlinear, shock-forming multiphase flow behavior characteristic of gas injection and withdrawal. Physics-informed artificial intelligence (PIAI), encompassing physics-informed neural networks (PINNs) and related hybrid architectures that embed governing partial differential equations directly into neural network training, has emerged as a means of resolving this tension, delivering models that are simultaneously computationally efficient and physically consistent. This article presents a research synthesis of physics-informed AI applications for subsurface natural gas storage, examining foundational PINN formulations for two-phase flow and transport, domain decomposition strategies for sparse-data reservoir simulation, and the documented limitations of naive physics-informed approaches when applied to shock-forming hyperbolic flow problems characteristic of gas injection. It draws on case studies from heterogeneous depleted gas reservoir optimization, geothermal-analogue reactive heat exchanger engineering, and high-resolution petrophysical characterization, and further considers methodological parallels with large-scale, constraint-consistent AI frameworks developed for public health, healthcare administration, and healthcare economics. The article concludes by identifying persistent challenges in training stability, shock-capturing fidelity, and multiphysics integration, and proposes a research agenda for advancing physics-informed AI toward routine, trustworthy use in subsurface natural gas storage management as the field matures through 2026 and beyond.