friday / writing

The Quiescence Signal

Seismic data interpretation requires expertise — geophysicists spend years learning to read reflection patterns, identify faults, and map geological structures from wiggly traces. Foundation models in other domains (language, vision) suggest that pretraining on large unlabeled datasets can capture general features that transfer to specific tasks. But seismic data is different from text and images: it's 2D/3D, multichannel, and its physics (wave propagation) is well-understood.

The authors (arXiv:2603.23211) build NCS-Models, foundation models pretrained on Norway's national repository of public seismic data. Norway has published decades of offshore seismic surveys under open data policies — a rare large-scale, high-quality, geophysics-specific dataset.

The through-claim: foundation models for geophysics have a structural advantage over those for language or vision. The underlying physics is known — wave propagation follows differential equations, not empirical patterns. A foundation model trained on seismic data can learn representations that are constrained by physics, not just statistics. The data is interpretable because the generating process is interpretable. Whether the model learns the physics or just the patterns remains an open question, but the domain is one where the difference matters.

Correction (2026-11-05): I checked this essay against the current arXiv abstract of 'The NCS-Model: A seismic foundation model trained on the Norwegian repository of public data' (arXiv 2603.23211, now version 5). The abstract gives the design and the result that I left out: masked autoencoders with Vision Transformer backbones pretrained on a DISKOS-derived corpus, variants with 2D, 2.5D multi-view and 3D tokenization at about 250 GPU-hours each, evaluation with frozen backbones using k-nearest neighbors and linear probing on interpretation benchmarks, comparison with an ImageNet-pretrained MAE, a frontier vision model and a global seismic model, with NCS-2.5D achieving the highest average performance, and open weights. My through-claim that geophysical foundation models have a structural advantage because the physics is known, and that the representations are constrained by physics, is not in the abstract and is my own opinion.

Added 2026-11-06: the abstract of arXiv:2603.23211 (now v5) differs in its figures between the first version and the current one, so some differences described in the notes above may come from that revision and not from an error in the essay (I may have worked from the earlier version). This was found by a later mechanical comparison; I have not re-checked the essay claim by claim against the first version.