Explainable AI for subsurface carbon capture, utilization, and storage systems: A review

초록

The high-stakes nature of carbon capture, utilization, and storage (CCUS) demands artificial intelligence (AI)driven models that are not only accurate but also transparent, auditable, and trustworthy. While black-box machine learning offers predictive power, its opacity undermines regulatory acceptance, risk assessment, and operational reliability in subsurface applications. Explainable AI (XAI) addresses this disparity by providing interpretable and comprehensible decision-making assistance. This review presents the first systematic examination of XAI specifically within CCUS, synthesizing advances in interpretable modeling, post hoc explanation (PHE) techniques, and surrogate-based approaches across reservoir characterization, carbon dioxide (CO2)enhanced oil recovery (CO2-EOR), storage integrity monitoring, and site selection. Hybrid and ensemble XAI methods, particularly those combining domain knowledge with multi-task learning (MTL), achieve superior trade-offs between accuracy and interpretability. The analysis, supported by bibliometric evidence, reveals growing research activity but limited real-world validation and standardized evaluation protocols. Case studies demonstrate XAI's capacity to clarify CO2 plume behavior, quantify uncertainty, and support transparent history matching. The principal outcome of this review is that XAI is indispensable for building credible, socially accepted CCUS systems. Nevertheless, its full potential hinges on robust validation frameworks, data transparency, and tighter coupling between explainability, physical constraints, and regulatory requirements. This review establishes a foundation for deploying human-centered, trustworthy AI in next-generation CCUS operations.

키워드