Machine learning for subsurface geological feature identification from seismic data: Methods, datasets, challenges, and opportunities

Identification of geological features from seismic data such as faults, salt bodies, and channels, is essential for studies of the shallow Earth, natural disaster forecasting and evaluation, carbon capture and storage, hydrogen storage, geothermal energy development, and traditional resource exploration. However, manual seismic interpretation is distinctly subjective and labor-intensive. With the advent and rise of 3D surveys, the size of seismic data has increased dramatically, making purely manual interpretation impractical. Since 1989, a large number of machine learning-based methods for identifying geological features have been proposed to address these challenges. To date, these methods have not been reasonably synthesized. Motivated by a progressive increase in applications, this review presents an overview of advances in the utilization of machine learning to identify geological features from seismic data. First, we classify these methods from five different perspectives. Second, we provide a comprehensive overview of 241 publications related to seismic geological feature identification and offer a detailed analysis of the development of these methods categorized by geological feature type. Third, 20 field and 12 synthetic seismic datasets, which are publicly available and relevant to the identification of faults, salt bodies, channels, caves, and horizons, are cataloged. Fourth, we discuss the issue of false positive identification caused by the limited geological features in the training dataset. To address the problems of false positives and insufficient labeled training datasets, we propose a simulation framework for generating 3D synthetic seismic data and corresponding geological labels that include a rich variety of geological features. To the best of our knowledge, this is the synthetic seismic dataset that contains the richest geological features. Finally, we discuss in depth the current challenges and future opportunities to inspire further relevant research.