Preprint proposes DDPM framework for fault-conditioned seismic patch generation across three field datasets

The work trains U-DDPM and FC-DDPM variants on jointly standardized field data, using binary fault masks as conditioning input. Frequency-wavenumber analysis shows generated samples match the training distribution's spatiotemporal bandwidth and dip content without physics-based forward modeling. This addresses data scarcity for DL fault interpretation by enabling one-to-many realizations from a single structural template.

Standard geometric augmentations and noise injection fail to reproduce realistic seismic textures, while full-waveform modeling remains computationally prohibitive. The DDPM approach learns the high-dimensional distribution directly, offering controllable augmentation that generalizes across surveys with differing amplitude scales and processing histories. Cross-survey training demonstrates the method's robustness beyond single-basin applications.

Earthquake risk modeling and reservoir characterization both depend on accurate fault maps; improved synthetic training sets could accelerate DL deployment in regions with sparse labeled data. However, the preprint lacks quantitative downstream benchmarks showing improved fault detection F1 scores on held-out field data, leaving the practical gain unproven.

Next steps include integrating the generator into active learning loops for seismic interpretation and testing whether the synthetic variability reduces model overfitting on small labeled sets. Validation on a fourth independent survey would strengthen claims of cross-survey generalization.