The oil and gas industry faces the dual challenge of improving recovery from mature reservoirs while reducing carbon emissions during the energy transition. Understanding micro-scale mechanisms (phenomena spanning from molecular and nano-pore fluid behavior to pore-scale multi-phase flow) is essential for advancing carbon dioxide (CO 2 )-enhanced oil recovery (CO 2 -EOR) and carbon capture, utilization, and storage (CCUS) technologies. Digital imaging and numerical simulation methods, while foundational, face inherent limitations, including the resolution-field of view (FoV) trade-offs and prohibitive computational costs that hinder their application in optimizing sustainable recovery strategies. This review provides a comprehensive assessment of how data-driven methods are transforming micro-scale petroleum research across three aspects. At the imaging level, we examine deep learning advances in automated segmentation, super-resolution reconstruction, and 3D pore structure generation. At the pore level, we analyze neural network approaches for rapid transport-property prediction and physics-informed methods balancing computational efficiency with physical rigor. At the molecular level, we evaluate machine learning interatomic potentials achieving near-quantum mechanical accuracy at classical simulation costs and their applications in nano-scale fluid-surface interactions. Our analysis indicates that while data-driven methods offer substantial computational acceleration, critical challenges remain in training data quality, model transferability, physical consistency, and pore-to-reservoir-scale bridging. This review aims to provide researchers and engineers with a systematic framework for evaluating both the transformative potential and current limitations of these approaches in sustainable oil and gas development.
Li et al. (Thu,) studied this question.