Fast Prestack Wavefield Separation via 3D U-Net Knowledge Distillation

We present a hybrid approach for automated prestack seismic data processing that combines expert-configured High-Resolution Wavefield Separation (HRWS) algorithms with 3D U-Net neural networks through knowledge distillation. The methodology uses HRWS with expert-tuned parameters to generate high-quality training targets, enabling neural networks to learn expert processing decisions. The model features a 51-channel input structure (50 amplitude offsets plus offset-density mask) and incorporates advanced attention mechanisms (CBAM, ASPP) for residual prediction. For industrial deployment, models are exported to ONNX format and deployed in distributed C++ environments with GPU acceleration. Experimental results demonstrate 15–20× speed improvement over classical methods while maintaining expert-level quality on diverse datasets without manual parameter tuning.