A New Approach to Multi-Scale Characterization of the 3D Seismic Fault Zone Architecture using Deep Learning

Summary Conventional fault interpretation workflows in reflection seismic data is not only time-consuming but also subject to significant human-bias, which is difficult to quantify reliably. In this study, we apply 3D Deep Learning (DL) networks trained on synthetic 3D seismic datasets, to a seismic survey from the Polhem Subplatform, SW Barents Sea for automatic generation of probabilistic fault volumes. The detailed extraction of fault segments from the prediction of DL algorithms enables a multiscale characterization of the seismic fault zone architecture in 3D. These fault zones are affected by the presence of unconformities. Our results indicate that the maximum segment lengths of the faults occur near their upper tips and progressively decrease towards the lower tips. Fault zone widths, measured on E–W-oriented scanlines-trending orthogonal to the dominant fault-strike at various depth intervals, generally range between 10–20m. It increases where fault segments either link laterally/vertically or interact with antithetic fault sets with similar strike as that of the dominant westward-dipping fault set. Fault zone widths and throw are greater towards the lower tips of faults. Thus, the DL-based approach utilized in this study enables the extraction of reliable, quantitative insights into the seismic fault zone architecture, while significantly reducing interpretation-based uncertainty.