Machine learning for CO2 geological storage and CO2-enhanced oil recovery in hydrocarbon reservoirs: A critical review of methodologies, the sim-to-real gap, and a roadmap for field-scale deployment

• ML proxies for CO 2 -EOR/CGS often achieve R² > 0.95 with 10 3 -10 6 speedups over compositional simulators. • The sim-to-real gap, data leakage risks, and regulatory hesitance toward black-box models restrict field deployment. • PINNs and hybrid ML-physics workflows enhance physical consistency and improve generalization under geological uncertainty. • Explainable AI (XAI) enhances interpretability for reservoir engineering decisions and MVA compliance. Carbon capture, utilization, and storage (CCUS) in hydrocarbon reservoirs is a cornerstone technology for mitigating anthropogenic CO₂ emissions. Yet, its large-scale deployment is hindered by computational complexity, geological uncertainty, and stringent regulatory requirements. Machine learning (ML) has emerged as a surrogate modeling paradigm to accelerate reservoir simulation, history matching, uncertainty quantification, and optimization. This review critically synthesizes 93 empirical studies (2010–2025) on ML applications in CO₂ geological storage (CGS) and CO₂-enhanced oil recovery (CO₂-EOR). We categorize contributions by algorithmic families and petroleum engineering tasks, moving beyond descriptive reporting to a cross-study analysis of methodological trade-offs and systemic weaknesses that perpetuate the sim-to-real gap. Across applications, ML proxy models frequently report coefficients of determination (R²) exceeding 0.95 with computational speedups of 10³–10⁶ compared to full-physics simulators. However, our critical analysis reveals that this high performance often masks a critical "sim-to-real" gap: the field is over-reliant on synthetic datasets with limited validation against field data, lacks standardized benchmarking, and inadequately addresses out-of-distribution generalization. Furthermore, the prevalent use of black-box models without integrated physical constraints or robust uncertainty quantification (UQ) poses a significant barrier to risk-informed, regulatory-compliant reservoir management. We identify physics-informed neural networks (PINNs), hybrid ML–physics workflows, and explainable artificial intelligence (XAI) as key enablers for trustworthy industrial deployment. We conclude by proposing a pragmatic petroleum engineering roadmap that integrates ML-assisted CCUS workflows into field development planning, injection optimization, and long-term storage assurance, emphasizing the need for a shift from academic proof-of-concept to engineering-grade, deployable solutions.