AI-Powered Fault Detection and Interpretation: From Neural Networks to Ready-to-Use Fault Surfaces

We present a comprehensive automated solution for 3D seismic fault detection and interpretation that combines deep learning with advanced geometric post-processing. The method integrates a 3D U-Net neural network trained on synthetic data with normalized distance function targets and an integrated post-processing pipeline including planarity filtering, Hessian-based faultness extraction, time-slice fault tracing, and a novel mutual neighbor confirmation clustering algorithm. The core contribution lies in system-level integration of known methods into a fully automated end-to-end workflow with minimal parameterization, where critical parameters are determined automatically. Time-slice tracing with Hessian-based orientation provides azimuth-agnostic detection; the method is most effective for high-angle faults (dip > 45°). Internal validation on proprietary datasets achieved F1-scores of 0.82–0.87. In expert qualitative assessment, the majority of faults required no or minimal manual correction. The result is a production-ready system integrated into commercial software, significantly accelerating structural seismic interpretation.