A Supervised U‐Net With Diffusion Prior Regularization for Efficient Seismic Diffraction Separation

Abstract

Seismic diffraction events contain high‐wavenumber information that is critical for imaging small‐scale discontinuities, yet their weak amplitudes are often obscured by strong reflection energy. Diffusion models offer a powerful framework for diffraction separation, but their iterative reverse‐diffusion sampling is computationally expensive. In contrast, supervised encoder–decoder separators enable efficient single‐pass inference but rely heavily on the quality and diversity of paired training data. To combine the advantages of both paradigms, we propose a diffusion‐prior‐regularized encoder–decoder separator that retains single‐pass inference while exploiting a frozen pretrained diffusion model as a statistical prior during training. The diffusion prior guides predicted diffraction wavefields towards the learned distribution of clean diffraction wavefields and is not required at inference time. Quantitative synthetic experiments demonstrate that the proposed method achieves higher separation accuracy than the supervised encoder–decoder baseline. Results on the Sigsbee2A benchmark further show reduced structural leakage, while the field‐data example provides a diagnostic comparison of component similarity. Because ground‐truth labels are unavailable for the field data, these results are interpreted as diagnostics rather than direct validation. On the synthetic test set, the conditional diffusion model achieves the highest accuracy but incurs substantially higher computational cost. In comparison, the proposed separator provides a more practical accuracy–efficiency trade‐off, maintaining competitive accuracy while significantly reducing inference time relative to the conditional diffusion model and approaching the computational efficiency of the supervised encoder–decoder separator.