A machine-learning program fed only routine earthquake catalogs flagged an unusual group of small quakes in the eight months before Turkey’s magnitude 7.8 earthquake in 2023. It found nothing before Japan’s magnitude 7.5 Noto Peninsula quake in 2024. Researchers at GFZ in Potsdam describe both outcomes in a new study.
Tech news in one quick read: Get the most interesting stories in startups, tech and programming, each with a short TLDR, in one daily email from TLDR.
We may earn a commission if you sign up.
The program was never told what a precursor looks like. It received the time, location and magnitude of each quake, grouped neighbors into families, and checked whether new kinds of families showed up before a large shock when the data were replayed in order.
GFZ’s program sorted 5,686 Turkish quakes
The study, in Nature Communications, comes from Sadegh Karimpouli, Patricia Martínez-Garzón and Marco Bohnhoff of the GFZ Helmholtz Centre for Geosciences, with colleagues in Germany, Italy and the United States. Earth.com’s account says each family was described with about 100 measurements and sorted into four or five groups. For Turkey the team used 5,686 quakes within about 50 kilometers of the epicenter. One group appeared mostly during the eight months before the magnitude 7.8 event, a period when a burst of small quakes had already been reported in a 2023 study. The program reached a similar signal without the authors telling it where to look. The design is unsupervised: with no labeled examples of precursors, a signal means only that the quakes before a shock differed from the usual patterns in that region. Replaying each dataset in order, as if the main quake had not yet happened, imitates a live feed, because the program sees only data up to each moment.
Hits in Turkey, Italy and Chile, misses in Italy and Japan
At L’Aquila, Italy, in 2009 (magnitude 6.1), two groups were active in the months before the quake, one of long-lasting families that appeared mainly beforehand. In northern Chile in 2014 (magnitude 8.1) the most unusual groups arrived about two weeks before the main shock, matching the burst of small quakes and slow fault slip that a 2014 Science paper reported in those weeks.
The three hits span magnitudes from 6.1 to 8.1, but the lead times differ widely: eight months for Turkey, months for L’Aquila, about two weeks for Chile. An alert whose warning time could fall anywhere in that range would be hard to act on. The program also did not detect the years of fault weakening described in other Chile studies, so even that hit captured only the final stretch. Two cases gave nothing. At Amatrice, Italy, in 2016 (magnitude 6.2), no warning group appeared in at least the six months beforehand. The Noto Peninsula case is the harder miss: a swarm began in late 2020, and 23,630 quakes were logged from 2021 onward, yet none of the groups showed the traits seen in the other three events.
The result matters for how the field frames prediction. Karimpouli and Martínez-Garzón present the method as a way to flag unusual fault behavior rather than to issue a forecast. A flagged anomaly says a fault system is acting differently; it does not give a date, a location or a magnitude for what follows. The Chile record shows the point: the 2014 Science paper had described a burst of small quakes and slow slip in the final weeks, so the two-week signal agrees with prior observation instead of contradicting it.
Limits the authors state themselves
Several records held few families, 96 for Turkey and 186 for L’Aquila, which may be too few for dependable groups. Choosing how many groups to use and labeling new ones also requires expert judgment, and the method can only work if clustered quakes come before the main shock.
Karimpouli said the approach offers a powerful way to recognize when a fault system is behaving differently than usual. He did not claim it predicts earthquakes in a deterministic way, and the team has not yet run it on live data. Martínez-Garzón named real-time monitoring as the next step, along with working out why some faults show signals and others do not, since some may fail without obvious seismic warning signs.
The coverage of the study quotes the authors on next steps rather than on any claim of imminent warning capability. The Amatrice miss is a clean negative, since the program saw no warning group in at least the six months before the quake, and the Noto miss is a noisy one, because a long swarm was present and still produced nothing that resembled the other cases. Both results are reported alongside the hits.
Five events were examined after the fact. Three showed a signal, with lead times from about two weeks to eight months, and two showed none, including a Noto Peninsula sequence of more than 23,000 quakes. Hindsight is the central caveat, since a live system would have to make the judgment calls about group counts and labels in real time. Whether it would raise a useful alert before the next large shock is unanswered.
This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.
Tech news in one quick read: Subscribe to TLDR →
We may earn a commission if you sign up.
More from Morning Overview
- Long use of a common prostate pill is tied to a higher chance of glaucoma
- Ford is recalling 223,472 F-150 pickups because the fuel tank can leak or detach
- Hybrids have 15% fewer problems than gas cars, while EVs and plug-in hybrids have about 80% more, Consumer Reports finds
- Regulators cleared the first U.S. small modular reactor, 4 months early
