PRISM: A Pre-Trained Foundation Model for Rapid Seismic Interpretation

While robust seismic interpretation and modeling is essential in successful subsurface mapping and characterization, it remains challenging due to its strong dependency on data quality, domain knowledge, and expert supervision. In real scenarios of noisy seismic, correspondingly, manual efforts could be intensive, and thus significant progress is anticipated in automation and efficiency improvement. The recent success of deep learning (DL) in natural image processing and pattern recognition has led to the development of many more (semi-) automated solutions to resolve various seismic interpretation challenges, such as saltbody detection, fault picking and horizon tracking. However, each of them is defined as a standalone process for completing one or a few specific tasks only and requires repeating model training in case of any changes in data and/or labels, causing computational redundancy. Another major drawback is their failure to integrate with non-seismic data for developing multidisciplinary applications. Resolving both challenges becomes feasible through pretraining a large seismic foundation model (FM) that provides rich and reliable representation of diverse seismic patterns observed on post-stack seismic images and adopting it into multiple downstream workflows. Therefore, pretrained seismic FMs are considered of great potential in renovating how seismic interpretation can be more automated and efficient. This paper presents a self-supervised approach for training such a seismic FM and demonstrates three scenarios of leveraging it for seismic data conditioning, interpretation, and inversion through six real use cases.