Detailed 3D characterization of fault geometric attributes: Insights from deep learning-based fault imaging in seismic data

Generation of 3D fault network in conventional seismic interpretation involves creation of structural models by manually tracking the same discontinuity across seismic horizons on several adjacent vertical profiles. This process is often time-consuming, susceptible to human bias and introduces uncertainties in the characterization of fault geometric attributes-such as length, width and throw-that cannot be reliably quantified. In this study, we adopt a data-driven approach and extract the geometric attributes of the individual segments constituting the 3D seismic fault zone architecture in remarkable detail, that provides insights beyond the capabilities of conventional seismic attributes. This is achieved through the application of 3D Deep Learning (DL) Networks trained on 3D synthetic seismic datasets on a seismic survey from the eastern flank of Polhem Subplatform, SW of Norwegian Barents Sea to automatically create probabilistic fault volumes. Detailed extraction of individual fault segments from the fault probability volume allows us to investigate the seismic fault zone architecture at multiple scales. Our results show that the maximum segment length of all the studied faults is present near the upper tips and reduces towards the lower tips. The fault zone widths of individual segments measured on E-W-oriented vertical scanplanes orthogonal to the strike of the dominant fault set mostly vary between 10-20m. They exhibit higher values at locations where fault segments link laterally/vertically or interact with antithetic fault sets with similar strike of the dominant westward dipping fault set. Throw estimates and the total accumulated fault zone widths are higher towards the lower tips of the faults. Thus, the application of Deep Learning enables data-driven, high-resolution visualization of the seismic fault zone architecture and facilitates comprehensive extraction of fault geometric attributes, providing a more robust complement to traditional interpretation techniques. • Detailed fault geometric attributes in 3D are extracted from deep-learning-based fault imaging of seismic data. • Central portion of the fault zones exhibits most continuous segments. • Width of most segments in the studied fault zones are 10–20 m. • Total accumulated width of segments increases towards the lower tips of fault zones. • Throw estimated below the Upper Barremian Unconformity is higher than above it.