AI is changing our understanding of earthquakes – Knowable Magazine (knowablemagazine.org)

Machine learning is rapidly reshaping earthquake science by automating labor-intensive signal analysis and revealing vast numbers of previously undetected tiny quakes. Algorithms now routinely do phase picking — identifying P and S wave arrival times — far faster than humans and with comparable accuracy, and they can trawl terabytes of seismometer, fiber‑optic, and smartphone accelerometer data to pull out events hidden in noise. Landmark work includes a 2019 study that recovered 1.5 million small Southern California quakes and recent AI catalogs (e.g., Taiwan’s 2024 M7.3 study) that are roughly five times more complete than human‑made lists, produced within a day. Expanded catalogs let researchers image fault structure with much finer detail and study the small‑event behavior that may presage large ruptures — in one case showing that 80% of larger quakes could be anticipated from preceding small ones. Despite these gains, ML has not yet beaten traditional statistical methods for probabilistic forecasting of future quakes; current AI models match but don’t outperform classic approaches. Where ML excels today is speed and scale: faster aftershock forecasts, richer catalogs for hazard modeling, and potential improvements in early warning where systems exist. Experts caution about quality control and interpretation, but overall the field sees ML as opening a “new window” on rupture physics and seismic risk, with practical benefits for mapping hidden faults and accelerating post‑quake decision support.