The Seismic Foundation

Deep learning for seismic data processing has a training data problem. Models are typically trained on narrow datasets — one geological setting, one noise profile, one acquisition geometry. When deployed on different data, they fail. The models learn the specific dataset, not the physics.

Gong, Fomel, and Chen build SWAN — a large-scale, open-source benchmark combining diverse synthetic and real seismic waveforms across geological structures and noise conditions. Then they train a single diffusion model on this breadth. The model, conditioned on the specific processing task (missing-trace reconstruction), generalizes across geological scenarios it wasn’t explicitly trained on.

The through-claim: the generalizability bottleneck in seismic ML was the dataset, not the model. Narrow training data produced narrow models. Broad training data produces broad models. The diffusion architecture matters less than the distribution it trains on.

The SWAN-trained model outperforms both existing deep learning methods and physics-based approaches on heterogeneous test cases. The physics-based methods have the correct theory but limited flexibility. The deep learning methods have flexibility but limited training. SWAN gives the flexibility access to breadth.

This is a pattern beyond seismology: when domain-specific ML fails to generalize, the first diagnosis should be the training distribution, not the model architecture. The foundation model paradigm — train broadly, deploy specifically — works when someone builds the foundation dataset. Most domains haven’t.