Abstract

Deep learning has rapidly transformed seismology, and more generally geophysics, by shifting the focus from task-specific algorithms to learning general representations directly from data. Early successes came from supervised applications such as earthquake, as well as polyphonic seismo-volcanic signals detection upon introducing the “scattering network”, and phase picking but a key advance occurred with unsupervised (deep clustering) and self-supervised methods which we developed and that can organize seismic waveforms without reliance on labeled catalogs. These approaches uncover latent structure across multiple time scales and have enabled the discovery (and separation) of previously undetected events and subtle seismic phenomena. Building on these results, we introduced SeisLM, a large-scale pretrained (transformer-based) “foundation model”, trained on continuous global datasets to provide transferable representations that can be adapted across regions and tasks through fine tuning with minimal additional data, such as tremor detection associated with slow slip events. We show how self attention mechanisms naturally support forecasting, and introduce HARPA , a high-rate phase association framework that lifts arrival sequences associated with (unknown) microseismic events from arrays to probability distributions and compares them using an optimal transport metric; a generative travel time neural field is used to estimate the wave speed (and event locations) simultaneously with association. Leveraging an architecture reminiscent of cross attention encoding geometry, we further demonstrate that ray transforms – and implicitly the underlying wave speeds – can be learned directly from travel-time data. Finally, we present designs of foundation models, including Flowers, combining data-driven with physical principles, aimed at enabling scientific discovery. We show examples of approximating wave propagation and scattering, and fluid dynamics through data-driven surrogates.