AI-Driven Uncertainty Quantification for Long-Term CO₂ Storage Security

Abstract

Long-term security of geological CO₂ storage (GCS) depends fundamentally on the ability to characterize and communicate uncertainty in predictions of plume migration, pressure evolution, and containment integrity across decadal to centennial project timescales. Point-estimate predictions from deterministic reservoir simulations or single-output machine learning models, however accurate on average, provide an insufficient basis for the risk-informed regulatory and operational decisions that long-term storage security demands, particularly given the deep geologic uncertainty inherent to subsurface characterization and the practical impossibility of directly validating century-scale containment predictions within a human planning horizon. Artificial intelligence-driven uncertainty quantification (UQ)—spanning Bayesian neural networks, ensemble-based deep learning surrogates, and simulation-based inference techniques—has emerged as an essential complement to point-prediction machine learning, enabling the propagation of both aleatoric uncertainty arising from inherent geologic variability and epistemic uncertainty arising from limited data and model knowledge into calibrated, decision-relevant probabilistic forecasts. This review examines the theoretical foundations and applications of AI-driven UQ for long-term CO₂ storage security, synthesizing recent advances in dimension-adaptive Bayesian neural networks, deep convolutional encoder-decoder surrogates, AI-enhanced data assimilation, and uncertainty-aware digital shadow frameworks. It draws on case studies from heterogeneous depleted gas reservoirs, geothermal-analogue petrophysical characterization, and geomechanically informed wellbore stability assessment, and further considers methodological parallels with large-scale AI-driven predictive analytics and economic impact frameworks developed for national healthcare systems. The review concludes by identifying persistent challenges in uncertainty calibration, computational scalability to three-dimensional and century-scale problems, and the translation of quantified uncertainty into regulatory decision criteria, and proposes a research agenda for advancing AI-driven UQ toward routine use in long-term CO₂ storage security assessment.