Fault Geometric Characterization: Insights from Deep Learning Applied to High Resolution 3D Seismic Data

Summary This study applies deep learning techniques to detect and characterize fault geometries in very-high-resolution 3D seismic data. Using a 3D U-net fault probability model, we investigated a shallow fault array within the Bjarmeland Platform in the southern Barents Sea. Structural analysis was conducted across key formations -Naust, Kolmule, Kolje, Stø, and Snadd-. The deep learning model enhanced the extraction of realistic fault shapes and the quantification of their geometric attributes—length, segmentation, and throw—in three dimensions. Results revealed a high degree of compartmentalization within the three deepest stratigraphic horizons. The fault array in this interval is characterized by predominantly orthogonal (∼90°) crosscutting relationships. While most faults exhibit relatively straight strike traces, some display more curvilinear geometries. Fault segmentation increases toward the base, and throw decreases toward the tip lines, with additional variability observed between specific stratigraphic levels. These findings underscore the value of data-driven approaches in minimizing interpretation bias, enhancing efficiency, and advancing our understanding of fault architecture and evolution in sedimentary basins.