Intelligent Identification of Sample-Adaptive Fracture Systems and Seismic Structure Analysis: A Case Study of the Hutubi Gas Storage Field in Xinjiang, China

Abstract Accurate identification of fracture systems is crucial for analyzing the causal mechanisms of induced earthquakes. Previous studies have been constrained by the low resolution of seismic reflection profile analyses, limiting the characterization of fracture zones with small displacements and short extensions, and hindering the identification of seismogenic structures. This study focuses on the Hutubi gas storage (HUGS) field in Xinjiang, China, a region prone to frequent small-magnitude earthquakes. Although most seismogenic structures in this region are understood, uncertainties remain due to a few earthquake clusters with unconfirmed structures, complicating seismic risk assessment. Utilizing high-resolution 3D seismic reflection data, we identified and explored the seismogenic structures of small-magnitude earthquakes in the peripheral region of HUGS. The causal mechanisms ranged from pre-existing faults to extensive fracturing at various scales around HUGS. We developed a deep-learning-based method for identifying multiscale faults and fractures and applied it to study the seismogenic structures. Large-scale faults and fractures were first identified, and the reliability of this intelligent identification was evaluated alongside manual interpretation. Sample-adaptive methods were then employed to resample the seismic dataset multiple times, identifying small- and medium-scale faults and fractures. Finally, we combined the multiscale identification results with precise earthquake locations to determine the seismogenic structures. The study found that stress perturbations within the HUGS field induce seismic activity along pre-existing faults and small-scale fractures, influencing the spatial distribution of shallow, small-magnitude seismic events. In addition, earthquake clusters on the southeastern margin of HUGS are likely due to stress concentration along the fold axis, leading to intralayer slip and activation of near-surface parallel fractures. These findings suggest that deep-learning-based multiscale fault and fracture identification offers a detailed perspective for studying seismogenic structures and assessing seismic hazards.