3D fault detection using vision transformer with a hybrid-attention mechanism

Fault detection is crucial for seismic interpretation, aiding reservoir characterization, and hydrocarbon exploration. Tra­ditional fault segmentation relies on seismic attributes, while deep learning (DL) methods (e.g., convolutional neural net­work (CNN)) offer enhanced performance. However, CNNs, limited by their receptive field, focus on local features, poten­tially compromising fault continuity and detection accuracy. To address this, we propose FaultVitNet, a vision-transformer­assisted CNN that leverages a hybrid attention mechanism to capture global fault distribution. Additionally, considering the unbalanced distribution of ones (fault) and zeros (non-fault) during prediction, we integrate binary cross-entropy and dice loss to mitigate class imbalance during training. Applications to both synthetic and field data demonstrate superior fault­tracking capabilities of FaultVitNet over conventional CNNs in terms of accuracy and continuity.