Nobody can yet tell when the next large earthquake will strike, but a team in Potsdam believes it has found a better way to notice when a fault starts acting strangely. Geoscientists headed by Sadegh Karimpouli at the GFZ Helmholtz Centre for Geosciences, describe the approach in Nature Communications.
What sets the work apart is that the computer was never told what a warning sign should look like. Using unsupervised machine learning, the researchers gave it only the bare entries found in any seismic catalogue: time, location and magnitude of small tremors. The software then lumped closely related events into clusters the team calls “families” and characterised each one with around a hundred measurements, from how tightly packed it was to how much strain it released.
The researchers replayed the records of five major earthquakes. Ahead of the magnitude 7.8 Kahramanmaraş disaster in Türkiye in 2023, the 2009 L’Aquila quake in Italy and the 2014 Iquique event in Chile, a distinct type of family emerged weeks or months early, more clustered, more localised and more energetic than the background. Before Amatrice in 2016 and Japan’s Noto Peninsula in 2024, by contrast, the method picked up nothing at all.
Karimpouli is careful about what that means. He stresses that firm forecasts remain out of reach; what the method offers is an alert that a particular fault has departed from its normal rhythm. Co-author Marco Bohnhoff stressed that quakes influence one another rather than occurring in isolation.
The authors acknowledge limits: some sequences offered only a few dozen to under 200 families, and choices such as the number of groups still need expert judgement. Their next step is testing the tool on live monitoring data.
