Scientists have developed a machine learning algorithm that independently identifies anomalous changes in fault behavior before powerful tremors, infohub.kz reports.
Researchers at the German Research Centre for Geosciences (GFZ) in Potsdam, working with colleagues from Italy and the United States, tested the program on archival seismic records from five destructive disasters of the past. The developers did not build pre-set templates for warning signals into the system. Instead, the algorithm was fed standard catalogs containing the precise timing, coordinates and magnitude of thousands of small tremors that seismologists continuously record around epicenters.
The program grouped small tremors into so-called "families" based on similarities in time, location and magnitude, then clustered them according to a hundred physical parameters. In three out of five cases — before the earthquakes in Turkey in 2023, in Italy's L'Aquila in 2009 and in northern Chile in 2014 — the algorithm detected rare combinations of tremors. These anomalies appeared weeks and months before the main shock: sources became more tightly compressed in space and released unusually large amounts of mechanical stress.
However, before the tremors in Italy's Amatrice in 2016 and on Japan's Noto Peninsula in 2024, the program found no warning signs. The authors stress that the model does not predict the exact date of a future disaster, but it could become an effective tool in operational monitoring systems, tracking zones where a geological fault begins to behave atypically.