Artificial Intelligence in Subsurface Energy Storage: A Critical Review of Characterization, Monitoring, Forecasting, and Risk Assessment

Abstract:

Subsurface storage of natural gas, hydrogen, thermal energy, and carbon dioxide is central to decarbonization strategies but requires forecasting coupled geological, hydrodynamic, geomechanical, and biogeochemical processes at field scale. This critical review evaluates the contribution of artificial intelligence to characterization, monitoring, forecasting, and risk assessment across four storage systems, covering classical ML, deep learning, operator learning, physics-informed neural networks (PINNs), reinforcement learning (RL), and Bayesian inference. A PRISMA-style search of Scopus, Web of Science, OnePetro, and Google Scholar returned 1, 658 records, of which 80 application-focused papers published on or before October 2024 were retained after screening, SANRA quality assessment, and per-study appraisal. The strongest field-demonstrated applications are CNN seismic interpretation, Bayesian history matching with ML proxies, DAS microseismic detection, tree-ensemble petrophysical regression under spatial cross-validation, and NRAP-IAM-class Bayesian risk frameworks for carbon dioxide storage. These are field-demonstrated rather than deployment-ready: uncertainty calibration, basin-transfer evidence, and operational integration remain uneven. The clearest over-claims concern plain PINNs for multi-scale coupled flow, RL policies for closed-loop control under geological uncertainty, and petrophysical R² values from random k-fold cross-validation, which inflate apparent generalization by 0.10-0.30 versus spatially honest evaluation. The dominant methodological weakness is undisclosed cross-validation strategy in spatially autocorrelated subsurface data. AI complements rather than replaces reservoir physics, adding value where data density and validation design constrain model behavior. Priority directions are mandatory cross-validation disclosure, operator-learning architectures for plume forecasting, and hydrogen-microbial coupled machine learning.