The warning signs before a major earthquake, if they exist at all, are often buried in thousands of smaller tremors that seem ordinary at first glance. The problem for geoscientists is not only finding those signals, but knowing whether they are meaningful before the main rupture arrives.
Researchers from the GFZ Helmholtz Center for Geosciences, including Dr. Sadegh Karimpouli and Prof. Dr. Patricia Martínez-Garzón, have now built a data-driven method that spots changes in seismic activity before some large earthquakes. Instead of telling a computer what warning pattern to search for, they used unsupervised machine learning — a type of AI that looks for structure in data without being given preset labels.
The method was tested on several major earthquake sequences whose histories are already well documented: Kahramanmaraş (Türkiye, 2023), Iquique (Chile, 2014) and L'Aquila (Italy, 2009). In every case, the analysis detected distinct foreshock patterns appearing weeks to months before the mainshock.
Crucially, the researchers changed the usual strategy by grouping related events into families based on their closeness in space, time, and magnitude — capturing how earthquakes influence one another as stress builds in the crust. The unsupervised algorithm identified three hallmark traits of an impending quake: stronger clustering and interaction among events, greater localization in space and time, and increased release of seismic strain.
We observe a transition from relatively stable activities to a more organized, critical state shortly before rupture, said Dr. Karimpouli.
Not all earthquakes show these signals. When applied to the 2016 Amatrice earthquake in Italy and the 2024 Noto earthquake in Japan, the method found no clear preparatory patterns — a reminder that some faults may fail without detectable warning.
The study was published in Nature Communications and supported by the European Research Council's QUAKEHUNTER project. While not a deterministic prediction system, the approach offers a powerful new tool for operational earthquake forecasting by recognizing when a fault system is behaving differently than usual.
