Fault detection in seismic data using a true 3D global attention convolutional network with self-supervised denoising...

Accurate fault detection in 3D seismic data requires modeling extended geological structures across multiple dimensions. Recent CNN-transformer hybrids employ windowed attention (Swin Transformer) or factorized attention, fragmenting volumetric fault geometries into constrained local interactions that cannot fully capture continuous fault planes. Additionally, models frequently misclassify seismic noise as faults due to limited labeled data and insufficient noise exposure. We present the first application of unfactorized 3D global attention-where every voxel attends to all other voxels-integrated with volumetric convolution for seismic fault segmentation. Building on Mahzad et al. (2026), our hybrid U-Net combines 3D convolution for local features with true global self-attention for complete spatial context. Training employs two-stage learning: self-supervised denoising pretext training across multiple real-world 3D surveys, then discriminative transfer learning on fault-labeled data using layer-wise learning rate decay and Unified Focal Loss. Validation on Thebe survey data achieved Dice = 0.853, IoU = 0.744, precision = 0.842, MCC = 0.841-statistically significant improvements (p < 0.001, Cohen's d = 0.54-1.02) of 4.5-8.1% over state-of-the-art 3D CNN+Swin architecture. On unseen test data, improvements widened to 9.9-16.7% with superior generalization (3.6% vs. 8.3% degradation). Results establish that unfactorized global attention with noise-aware pretext training significantly improves fault detection accuracy and generalization.