How AI is shaking up the study of earthquakes (www.understandingai.org)

Over the past seven years machine learning has transformed earthquake detection and phase picking from a labor-intensive human task into a largely automated process. Models like the Earthquake Transformer (Stanford, ~2020) and simpler systems such as PhaseNet now find many times more small earthquakes than traditional workflows or computationally expensive template matching (e.g., a Caltech study found 1.6M small quakes using 200 P100 GPUs). These AI tools detect quakes in noisy environments (cities), run on modest hardware, generalize across regions, and produce richer catalogs that improve imaging of Earth structure and hazard assessment—effectively “putting on glasses” for seismologists. Technically, success stems from combining large labeled datasets (STEAD’s ~1.2M segments) with standard neural building blocks: one‑dimensional convolutions over time, deconvolution to localize events, and attention layers to relate P- and S-wave patterns across time. Models output per-timestep probabilities for an event, P-wave arrival, and S-wave arrival; typical architectures are compact (~350k parameters) yet highly accurate. Limitations remain: AI has largely replaced older detection/phase-picking methods, but it hasn’t delivered reliable short-term earthquake forecasting. The field’s leap owes more to data scale and pragmatic architectures than to novel math—suggesting further gains will come from richer labels, broader deployment, and integrating ML outputs into physical forecasting workflows.