Improving Relative Geological Time Cubes with Contrastive Self-Supervised Learning Clustering

Abstract Relative Geological Time (RGT) cubes are essential for interpreting subsurface geological structures, as they assign relative ages to seismic data and help delineate facies and geological units. However, the accuracy of RGT cubes is often compromised in the presence of major faults and discontinuities, necessitating labor-intensive manual corrections. This paper introduces a novel approach that integrates RGT priors into a classical two-step self-supervised learning algorithm, eliminating the need for manual labeling. By leveraging a puzzle mask derived from fault detection and RGT horizons, the method employs a contrastive N-pair-loss in the autoencoder training and incorporates RGT values directly into the clustering step using Gaussian Mixture Models (GMM). The approach is validated on the challenging Parihaka dataset, demonstrating significant improvements in clustering quality and alignment of geological units. The final rendering process further refines the stratigraphic consistency of the results, providing a robust, automated alternative for RGT correction in complex seismic environments. This advancement paves the way for more efficient and accurate seismic interpretation workflows, with promising preliminary results for 3D applications.