Abstract Fault detection represents a crucial task in seismic interpretation, with significant implications for hydrocarbon exploration. Advancements in artificial intelligence have driven widespread adoption of machine learning techniques to tackle complex scientific and engineering problems. CNNs, Transformers, and their adaptive variants have been extensively applied to seismic fault detection, demonstrating promising efficacy across geological exploration applications. However, the large parameter counts and high computational complexity intrinsic to these models severely limit their applicability to resource‐constrained edge devices. With the progressive maturation of horizontal well drilling, data processing demands for the mobile Logging‐While‐Drilling (LWD) systems during drilling operations have become increasingly urgent. To address these requirements, we present LightGEUnet: a lightweight Unet architecture enhanced with two innovative modules–the Grouped Hadamard Product Attention for Multi‐axis (GHPA‐M) for cross‐dimensional feature refinement and Group Aggregation Fusion (GAF) for hierarchical feature integration. The GHPA‐M architecture employs grouped feature decomposition with triaxial Hadamard Product Attention (HPA) to extract 3D fault features through cross‐orientation interactions among geological axes (inline, crossline, time), while the GAF module integrates low‐level and high‐level features at each processing stage, effectively fusing multi‐scale information. The model was trained using a synthetic seismic data set, with multiple data augmentation techniques applied to enhance its generalization capacity and robustness under geological uncertainty. Our research conducted comprehensive validation and application of the model on the Netherlands North Sea F3 offshore seismic data set. In this study, the model is simultaneously implemented and validated on F3 offshore seismic data from the Netherlands, and LightGEUnet still shows great accuracy and continuity while being lightweight and efficient compared with traditional popular models.
Tang et al. (Sat,) studied this question.