Machine Learning‐Based Seismic Subsurface Characterization: The State of the Art and Future Perspectives

Seismic subsurface characterization involves interpreting seismic records to delineate geological features and estimate the physical properties of subsurface rocks (such as lithofacies, porosity, and fluid saturation). As an effective remote sensing technology to characterize the large‐scale spatial distribution of subsurface properties, seismic surveys play a crucial role in many fields of Earth, Energy, and Environmental Sciences, including hydrocarbon exploration and development, geological sequestration of CO 2 , underground water management, and geothermal energy exploitation. With the growing need for characterizing geological heterogeneity, the application of machine learning (ML)—capable of nonlinear mapping and feature extraction—has become essential for effective seismic subsurface characterization. To clarify the current state of ML‐based subsurface characterization and promote its application to complex geological formations, we review conventional and machine learning workflows, along with the challenges they face. Subsequently, we delve into ML and deep learning concepts, summarizing recent advances and case studies in seismic subsurface characterization using supervised, semi‐supervised, unsupervised, and transfer learning approaches. Given the non‐unique nature of subsurface inference, we emphasize the importance of physical and geological consistency and of uncertainty quantification, and discuss three actively researched topics: deep generative geomodeling, physics‐guided (or geology‐guided) ML, and uncertainty quantification. Other ML‐based characterization approaches, including large language models, multi‐task learning, multi‐source and multi‐modal data fusion, etc., are further discussed. We also compile openly available data sets and codes, emphasizing their importance for future research. In summary, our objective is to provide the geosciences community with an overview of current research and future perspectives in ML‐based subsurface characterization.