Answers are grounded in this paper and cite their sources.
Accurate seismic fault detection is essential for structural interpretation, reservoir characterization, and geological decision-making. However, automatic fault delineation remains challenging because fault responses are often weak, sparse, discontinuous, and embedded in complex seismic backgrounds. In this paper, we propose FaultUKAN, a Kolmogorov–Arnold Network (KAN)-enhanced U-Net architecture for seismic fault segmentation. While retaining the robust multi-scale encoder–decoder framework of U-Net, FaultUKAN enhances nonlinear feature representation by embedding KAN-based refinement blocks into the bottleneck and decoding pathways. Furthermore, attention-gated skip connections are integrated to selectively suppress fault-irrelevant responses and improve weak structural reconstruction. Comprehensive evaluations on the FaultSeg3D synthetic dataset demonstrate that FaultUKAN outperforms representative baselines, including U-Net, Attention U-Net, Fault-Seg-Net, and TransUNet, in both segmentation accuracy and fault-structure continuity. Controlled ablation experiments, including an MLP-only variant with a comparable parameter count, further confirm that the performance improvement cannot be attributed solely to increased model capacity. Field-data applications on the F3 and Kerry3D volumes also demonstrate the capability of FaultUKAN to generate visually continuous and geologically plausible fault responses under different structural settings.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: