Encoding the Subsurface in 3D with Seismic

Abstract

This article presents a self-supervised generative AI approach to seismic data processing and interpretation using a Masked AutoEncoder (MAE) with a Vision Transformer (ViT) backbone. We modified the MAE-ViT architecture to process 3D seismic mini-cubes to analyze post-stack seismic data. The MAE model can semantically categorize seismic features, demonstrated through t-SNE visualization, much like large language models (LLMs) understand text. After we fine-tune the model, its ability to interpolate seismic volumes in 3D showcases a downstream application. The study's use of an open-source dataset from the "Onward - Patch the Planet" competition ensures transparency and reproducibility of the results. The findings are significant as they represent a step towards utilizing state-of-the-art technology for seismic processing and interpretation tasks.