This paper proposes a method for rapid and interactive mapping of geological structures in 3D seismic volumes using self-supervised learning and content-based image retrieval (CBIR). The method makes it possible to label interactively, meaning that the results will appear in near-real time as the user is labelling. The method consists of a Vision Transformer (ViT) model pre-trained in a masked autoencoder setting on seismic data. We extend the standard 2D ViT to 3D by using multiple 2D slices which allows for 3D context without excessive computational requirements. The pretrained ViT is used to extract features prior to labelling. A simple metric is used to measure how close any point in the cube is to the means of positive and negative labelled points. The method is demonstrated on a complex injectite system, showing efficient mapping of the structure. Future research will focus on enhancing the pre-training and similarity search processes to generate more discriminative features and increase the precision of the labelling process.

‍