Deep Learning for Geological Characterization and Modeling from Pore Scale to Reservoir Scale: State of the Art and Future Directions

Abstract

Accurate characterization and modeling of subsurface reservoirs necessitate the integration of multiscale information, ranging from pore-scale structures to reservoir-scale heterogeneity. Recent advances in deep learning have opened new avenues for subsurface representation and physical behavior prediction. Unlike conventional methods that depend on empirical assumptions and manual feature engineering, deep learning enables the automatic extraction of complex nonlinear geological patterns, implicit spatial correlations, and physics-informed constraints from multisource geophysical data. This review presents a comprehensive overview of deep-learning applications in four interconnected domains of reservoir characterization and modeling: digital core reconstruction, pore-scale flow and transport analysis, reservoir geological modeling, and fracture characterization. Digital core reconstruction from imaging data using deep learning methods is first elaborated. Building upon these reconstructed pore geometries, the focus then shifts from structural representation to physical response learning, where deep learning is employed to establish quantitative mappings between pore structures and transport behaviors through structure-to-property predictions and surrogate flow modeling. Scaling up to the reservoir level, deep learning is applied to reservoir geological reconstruction, where seismic data and well logs are integrated to infer large-scale structural and petrophysical distributions with sparse data. Beyond continuous reservoir representations, recent developments also extend these approaches to hydraulic fracture characterization, in which deep learning is used to model fracture networks and capture discontinuous geological features. This review highlights a holistic perspective on how deep learning is reshaping subsurface characterization and reservoir-scale modeling, from digital cores to reservoir-scale models.