The reliable identification of subsurface sedimentary structures (i.e., geologic heterogeneity) is critical in various earth and environmental sciences, petroleum reservoir engineering, and other porous media-related application. The application includes some important and societally relevant problems such as contaminated aquifer remediation, enhanced oil recovery, geological carbon storage, geological hydrogen storage, radioactive waste disposal, and contaminant fate and transport modeling. An inaccurately estimated subsurface sedimentary structure may introduce a larger bias into simulation results than inappropriate model parameters. Research on the development of subsurface sedimentary structure identification methods has recently witnessed increasing interest in deep learning (DL)-based methods. Such methods allow structure identification in a considerably different manner compared to traditional methods (e.g., covariance-based (co)kriging, multi-point statistics). The DL-based methods achieve significantly higher efficiency and accuracy. This review describes how DL-based methods have been utilized for subsurface sedimentary structure identification from the viewpoint of different identification approaches (direct and data assimilation-based modeling). Differences between DL-based and traditional methods are discussed. Furthermore, the limitations and challenges of existing DL-based methods are summarized. This includes training data acquisition, comparison of different algorithms, and limitations on accuracy and efficiency. Finally, future research directions are explored, including coupling DL-based and traditional methods, development of benchmark databases, DL-based methods driven by both data and theory, and applications of meta- and transfer learning. Effective solutions to these problems can provide numerous opportunities for DL-based methods to realize advances in subsurface sedimentary structure identification, thereby enabling a deeper scientific understanding of subsurface sedimentary structures.

地下沉积结构（即地质非均质性）的可靠识别在各种地球和环境科学、石油储层工程和其他多孔介质相关应用中至关重要。该应用程序包括一些重要的和社会相关的问题，例如受污染的含水层修复、提高石油采收率、地质碳储存、地质储氢、放射性废物处理以及污染物归宿和传输建模。与不适当的模型参数相比，不准确估计的地下沉积结构可能会给模拟结果带来更大的偏差。最近，对地下沉积结构识别方法发展的研究见证了人们对基于深度学习 (DL) 的方法越来越感兴趣。与传统方法（例如，基于协方差的克里金法、多点统计）相比，此类方法允许以截然不同的方式进行结构识别。基于深度学习的方法显着提高了效率和准确性。本综述从不同的识别方法（直接和基于数据同化的建模）的角度描述了如何将基于 DL 的方法用于地下沉积结构识别。讨论了基于 DL 的方法和传统方法之间的差异。此外，总结了现有基于 DL 的方法的局限性和挑战。这包括训练数据采集、不同算法的比较以及准确性和效率的限制。最后，探索了未来的研究方向，包括将基于 DL 的方法与传统方法相结合，基准数据库的开发、由数据和理论驱动的基于 DL 的方法，以及元学习和迁移学习的应用。这些问题的有效解决方案可以为基于 DL 的方法提供大量机会，以实现地下沉积结构识别的进步，从而使对地下沉积结构有更深入的科学理解。