Characterizing Mesozoic Strike-Slip Faults in China's Tahe Oilfield: A Multi-Method Comparison from Traditional Seismic Attributes to AI

Strike-slip faults critically control hydrocarbon migration in the Mesozoic clastic reservoirs of the Tahe Oilfield, NW China. However, their identification is challenged by weak seismic responses due to subtle impedance contrasts, steep dips, and small throws. This study conducts a systematic, multi-method comparison to optimize fault detection, evaluating both conventional seismic attributes and a novel deep learning (DL) approach.

We first applied structure-oriented filtering to enhance data continuity. Subsequently, key conventional attributes were computed: coherence and curvature to delineate major structural discontinuities and flexures, ant tracking to highlight fault pathways, and likelihood to map fault lineaments. The core of our DL approach involved a ResU-Net model, pre-trained on extensive datasets and refined via transfer learning using 65 manually interpreted fault traces from the target area. This process generated a high-resolution fault probability volume.

Results from the key T34 horizon demonstrate a clear performance hierarchy. While coherence and curvature effectively image major faults, they lack resolution for secondary networks. Ant tracking and likelihood show sensitivity to small-scale features but suffer from poor continuity and noise. In stark contrast, the AI probability volume integrates the strengths of these methods, simultaneously providing superior boundary clarity for major faults and enhanced detection of subtle, secondary strike-slip faults crucial for hydrocarbon migration. It presents a more continuous, spatially coherent, and geologically plausible 3D fault system.

This work underscores the significant advantage of an AI-driven, integrated workflow over individual conventional attributes. It provides a robust, scalable template for multi-scale fracture characterization in complex reservoirs, effectively bridging the gap between geophysical data analysis and geological interpretation.