A new analysis of earthquake “families” reveals that hidden patterns in clustering, localization and strain release can occur before some large earthquakes.
Signs of a major earthquake, if they exist at all, are often buried among thousands of small, seemingly ordinary tremors. The problem for geoscientists is not only to find these signals, but to know whether they are meaningful before major disruption occurs.
Researchers at the GFZ Helmholtz Center for Geosciences, including Dr. Sadeg Kalimpuri and Professor Patricia Martínez Garzón, in collaboration with international partners, have built a data-driven method to detect changes in seismic activity before some major earthquakes. Instead of telling the computer which warning patterns to look for, we used unsupervised warning patterns. machine learninga type of artificial intelligence Find structures in your data without providing preset labels.
The method was tested on several large earthquake sequences whose history is already well-documented, including the Kahramanmaras earthquake (Turquier, 2023), the Iquique earthquake (Chile, 2014), and the L’Aquila earthquake (Italy, 2009). In those cases, the analysis detected distinct foreshock patterns that emerged weeks to months before the main shock.
When we applied the same method to earthquakes without known precursor signals, such as the Noto earthquake (Japan, 2024) and the Amatrice earthquake (Italy, 2016), we did not find the same pattern. The researchers claim this approach could help improve operational earthquake predictions. This research nature communications.
Challenge to earthquake prediction
Predicting when, where, and how strong earthquakes will occur remains one of the most difficult unsolved problems in Earth science. Some researchers question whether accurate predictions are possible. Instead, much of the field focuses on precursory phenomena, or changes that can occur before a large earthquake. These include foreshocks, where a small earthquake occurs before a large earthquake, and slow slip events, where a fault moves quietly without producing strong shaking.
The problem is that these signals are inconsistent. Their timing, size, and location depend on faults, plate boundaries, local geology, and stresses already stored in the Earth’s crust. Patterns that appear before one earthquake may not exist before another.
Pattern recognition using unsupervised machine learning
Machine learning is already helping geoscientists handle complex seismic interactions and search large earthquake catalogs for patterns that are difficult to find manually.
In this study, Dr. Kalimpuri and his colleagues changed their usual strategy. Rather than starting with a fixed idea of what a precursor should look like, they allowed the data itself to categorize seismic activity into meaningful patterns.
“Instead of searching for specific precursors, we are letting the structure of the data itself reveal itself and making use of so-called unsupervised learning, where the diagnostic criteria are not predefined,” said lead author Dr. Sadegh Kalimpuri, a scientist in Section 4.2 “Geomechanics and Scientific Drilling” at GFZ. Similar unsupervised methods have previously helped detect early changes before landslides and volcanic eruptions.
From individual earthquakes to “families” interacting with each other
The next challenge was how to represent earthquakes in a way that captured earthquakes and their relationships. Rather than treating each earthquake as a separate point in a catalog, Dr. Kalimpuri and his colleagues grouped related earthquakes into “families” based on their proximity in space, time, and magnitude.
This change is important because earthquakes can interact. Small breaks can change nearby stresses, sometimes making another break more or less likely.
“Earthquakes are not isolated events. They interact with each other, and the closer the rupture event is, the stronger this influence is,” explains co-author Professor Marko Bornhof, head of GFZ Section 4.2 “Geomechanics and Scientific Drilling”. “By analyzing their collective behavior, we can better understand how stress builds up in the Earth’s crust before major events.”
The researchers then described each earthquake cluster using a number of physical and statistical features. These include how tightly events are clustered, how localized they are in space and time, and other indicators related to stress within the Earth’s crust. An unsupervised algorithm then grouped those families into categories that reflected different stages of stress evolution.
Kalimpuri and his colleagues had already tested this approach in controlled laboratory seismic experiments. The new question was whether it would work in nature, where disorders are much more complex and available data are often incomplete.
Detection of transition to critical state
The researchers applied the method to several large earthquake sequences in different crustal environments where precursor events have already been reported. These include the 2023 Mw 7.8 Kahramanmaras (Turquie) earthquake along a major strike-slip plate boundary, the 2009 Mw 6.1 L’Aquila (Italy) earthquake that occurred on a fragmented normal fault, and the 2014 Mw 8.1 Iquique (Chile) earthquake that occurred in a subduction zone. In all these examples, the analysis identified different types of seismic activity before the mainshock.
These important patterns had three main characteristics. Stronger clustering and interaction between earthquakes, greater localization in space and time, and increased strain release due to earthquakes. Taken together, these features indicate that the faulty system is approaching instability.
“We are observing a transition from the relatively stable activity known from previous activity in the region to a more organized and critical state on the verge of collapse,” Dr Kalimpuri said. In the cases studied, changes appeared weeks to months before the main shock.
Not all earthquakes show warning signals
The results also demonstrated important boundaries for this method. In some earthquakes, there may be no detectable seismic preparation before collapse. When Karimpuri and his colleagues applied this method to the 2016 Amatrice earthquake in Italy, no clear critical categories emerged compared to previous activity. The 2024 Noto earthquake in Japan also lacked clear preparation signals despite long-term swarm activity in the region.
“This variation reflects the complexity of both the monitoring conditions and the seismic process,” says co-author Professor Patricia Martínez Garzón. “Some faults may fail without clear earthquake warning signs, which is a major challenge for prediction.” One of the main objectives of Professor Martínez Garzón’s ERC-initiated project QUAKEHUNTER, which supports this research, is to understand when earthquake preparedness begins and when monitoring systems can detect it.
Towards advanced earthquake prediction
To explore whether this method could be useful for practical predictions, the researchers went beyond just analyzing past earthquakes after the fact. Within the same earthquake sequence, they also tested a forward-looking approach. They first used early earthquakes in each region to define the normal pattern of seismic activity. They then updated the analysis as each new earthquake occurred, monitoring the moments when activity began to deviate from the established context.
In this setting, the sudden appearance of a new seismic category may suggest that the fault system is transitioning to a different, potentially more dangerous state.
“This does not mean that we can predict earthquakes in a deterministic way,” emphasizes Dr. Kalimpuri. “However, it is a powerful tool for recognizing when a faulty system is behaving in an unusual way.”
A new perspective on the evolution of major earthquakes
This study shows how earthquake physics and machine learning can be combined to uncover subtle patterns that may be missed by traditional methods. This approach provides an alternative way to observe the evolution of large-scale ruptures by focusing on how seismic events interact as a group.
“Our findings show that machine learning can help identify earthquake preparations when they are present and detectable with installed equipment,” concludes Professor Martínez Garzón. “The next step is to integrate such approaches into real-time monitoring to better understand why some earthquakes show a clear signal and others don’t.”
Reference: “Preparatory steps for major earthquakes revealed by unsupervised classification of earthquake catalog features” Sadegh Karimpouli, Patricia Martínez-Garzón, Sebastián Núñez-Jara, Matteo Picozzi, Daniele Spallarossa, Grzegorz Kwiatek, Georg Dresen, Marco Bohnhoff, Gregory C. Beroza, May 4, 2026. nature communications.
DOI: 10.1038/s41467-026-72279-x
Sadegh Karimpouli and Patricia Martínez-Garzón received funding from the European Research Council (ERC) for the project QUAKEHUNTER under the European Union’s Horizon 2020 research and innovation program 101076119.
Never miss a breakthrough: join the SciTechDaily newsletter.
Follow us on Google and Google News.
