How AI is helping seismologists hear smaller earthquakes

Earthquake detection begins with a difficult listening problem. A seismometer records ground motion, but the trace is crowded with ocean swell, traffic, weather, quarry blasts and instrument noise. A small earthquake may be visible only as a faint pattern that arrives at several stations with slightly different timing. That is why artificial intelligence is interesting to seismologists: not because it predicts the next disaster, but because it can help find weak signals already hidden in the data.

A recent research highlight in the Earth and space sciences describes work by Köhler and colleagues using decades of readings from seismic arrays, including data operated by the Norwegian research foundation NORSAR. Instead of treating one instrument as the whole story, the researchers trained models to combine signals from many sensors spread across a local area. The geography of the array is part of the method. A real seismic wave should appear across stations in a physically consistent order, while random noise usually does not.


The machine-learning task is pattern recognition with safeguards. Models are trained on archived waveforms where analysts already know which wiggles correspond to earthquakes or other seismic events. They then scan new or withheld data for similar arrivals, sometimes detecting events that are too small, distant or messy for older automatic methods. Better detection can fill gaps in earthquake catalogues, especially for small events that reveal how faults behave between larger shocks.

This matters geographically because earthquake risk is built from place-specific evidence. A richer catalogue can show where clusters of small events occur, how stress migrates through a fault zone, or whether a volcanic, mining or reservoir region is changing. In some settings, faster detection can also support alerts after shaking has already begun, giving automated systems a few seconds to slow trains, stop equipment or warn nearby communities. Those seconds are useful only when the detection is reliable.


This does not make seismology less local. In fact, the best systems need local knowledge: which stations sit on hard rock, which lie near roads or coasts, how winter storms appear in the traces, and which mines or industrial sites produce repeating vibrations. AI can rank candidates quickly, but regional experts still know which signals make sense for a mountain belt, a rift valley, a volcano or a stable shield.

The limits are just as important as the promise. AI does not remove the need for dense instruments, good maintenance or expert review. A model trained in one region can make mistakes in another if geology, station spacing or noise sources differ. False positives can waste attention, while false negatives can hide important activity. Researchers therefore test models against independent data and compare them with established seismological methods before relying on them operationally.

The useful optimism is modest and practical. More sensitive catalogues make the underground map less blank. They help scientists see small fractures, aftershock sequences and human-induced vibrations with better resolution. For the public, that does not mean earthquakes become predictable like tomorrow’s rain. It means the ground’s memory can be read more carefully, and better reading is one of the quiet foundations of safer buildings, wiser planning and faster response.