Summary

When and where the next large earthquake will strike remains one of the most difficult questions in geosciences. Researchers from the GFZ Helmholtz Centre for Geosciences around Dr Sadegh Karimpouli and Prof. Dr Patricia Martínez-Garzón have now – together with international partners – developed a new data-driven approach that can identify characteristic changes in seismic activity before some major earthquakes occur. The team used unsupervised machine learning to detect previously hidden patterns in earthquake catalogues, without relying on predefined assumptions. They applied their method to several well-documented major earthquakes, including the Kahramanmaraş (Türkiye, 2023), Iquique (Chile, 2014) and L’Aquila (Italy, 2009) events, and were able to identify distinct patterns in the foreshock activity in these cases, which occurred weeks to months before the mainshock. For other earthquakes for which no precursor phenomena were known – the Noto (Japan, 2024) and Amatrice (Italy, 2016) events – the method found no patterns. The researchers therefore believe that their new approach has potential for the further development of operational earthquake forecasting approaches. The study has been published in the journal Nature Communications.

Background: The challenge of earthquake prediction

Predicting the timing, location and magnitude of future earthquakes represents a long-standing and still unsolved – if not impossible – challenge. Geoscientific research in this area focuses on attempting to identify precursor phenomena, i.e., specific patterns in the processes that may occur prior to some major earthquakes. Such preparatory processes may include, for example, foreshocks and slow-slip events in the vicinity of the future epicentre. Precursor signals observed to date vary considerably in their spatial and temporal extent and intensity, depending on the type of fault and plate boundary, as well as geological conditions and stress.

New Approach I: Pattern recognition using unsupervised machine learning

Machine learning methods have for some time been successfully applied to unravel the complexity of earthquake interactions and to identify distinct patterns in existing earthquake data catalogues. In their current study, the researchers are breaking new ground in several respect: 

“Instead of searching for specific precursors, we let the data reveal its own structure and make use of so-called unsupervised learning in which diagnostic criteria are not predefined”, says lead author Dr Sadegh Karimpouli, scientist in Section 4.2 “Geomechanics and Scientific Drilling” at GFZ. This approach has already been used successfully to detect the early stages of landslides and volcanic eruptions.

New approach II: From individual earthquakes to interacting “families”

A key innovation of the study is the shift from analysing single earthquakes to studying “families” of events, groups of earthquakes that are closely linked in space, time, and magnitude. These families reflect how earthquakes interact and evolve collectively.

“Earthquakes are not isolated events, they influence each other, and the closer the rupture event gets, the stronger this influence becomes”, explains co-author Prof. Marco Bohnhoff, Head of GFZ Section 4.2 “Geomechanics and Scientific Drilling”. “By analysing their collective behaviour, we can better capture how stress builds up in the Earth’s crust before large events.”

The researchers extracted a large number of physical and statistical features describing seismicity, such as clustering, spatial localization, and stress-related indicators. Using an unsupervised machine learning algorithm, these families were automatically grouped into distinct categories representing different states of stress evolution.

They had previously used this approach successfully in well-controlled laboratory earthquake experiments. The question now was whether it would also prove effective in the far more complex events that occur in nature.

Detecting the transition to a critical state

The research team applied the method to several well-documented earthquake sequences in different tectonic settings for which precursor phenomena were already known, including the 2023 Mw 7.8 Kahramanmaraş (Türkiye) event at a major strike-slip plate boundary, the 2009 Mw 6.1 L’Aquila (Italy) event on a set of fragmented normal faults, and the 2014 Mw 8.1 Iquique (Chile) event in a subduction zone. In each case, the team identified a distinct category of seismicity emerging before the mainshock.

These “critical” patterns are characterized by three groups of features: 1. increased clustering and interaction between earthquakes, 2. stronger spatial and temporal localization, and 3. enhanced release of seismic strain. Together, these features indicate that the fault system is approaching instability. “We observe a transition from relatively stable activities – known from previous activities in the region – I to a more organized, critical state shortly before rupture,” says Dr Karimpouli. Depending on the case, these changes appeared weeks to months before the main earthquake.

Not all earthquakes show warning signals

However, the study also highlights an important limitation: not all earthquakes are preceded by detectable seismic preparation. For example, when the method was applied to the 2016 Amatrice earthquake in Italy, no clear “critical” category, relative to the previous activities, emerged before the event. Similarly, in the case of the 2024 Noto earthquake in Japan, long-lasting swarm activity did not evolve into a clear seismic preparatory signal.

“This variability reflects the complexity of both monitoring conditions and earthquake processes,” says co-author Prof. Patricia Martínez-Garzón. “Some faults may fail without obvious seismic warning signs, which is a major challenge for forecasting.” Investigating the conditions under which earthquake preparatory processes emerge and are more detectable is one of the major goals of Prof. Martínez-Garzón’s ERC Starting Project QUAKEHUNTER, which funds the research developed here.

Towards improved earthquake forecasting

To assess the suitability of their method for practical use in earthquake prediction, the researchers tested their approach not only on past earthquakes but also – within the context of the same earthquake sequences mentioned above – prospectively: they based their analysis on previous earthquakes in the region. By continuously updating the analysis of further new earthquakes, the method can then detect when seismic activity begins to deviate from previously observed patterns. In this way, a sudden emergence of a new category of seismic behaviour may indicate that the system is entering a different and potentially more critical state.

“This does not mean we can predict earthquakes in a deterministic way,” emphasizes Dr Karimpouli. “But it provides a powerful tool to recognize when a fault system is behaving differently than usual.”

Resume: A new perspective on the developments of large earthquake 

The study demonstrates how combining physics-based features with machine learning can reveal subtle processes that are difficult to detect with conventional methods. By focusing on interactions within seismicity, the approach offers a new window into how large earthquakes develop.

“Our findings show that machine learning can help identify earthquake preparatory phases, when they exist and are detectable with installed instrumentation,” resumes Prof. Martínez-Garzón. “The next step is to integrate such approaches into real-time monitoring and to better understand why some earthquakes show clear signals while others do not.”

Funding:

Sadegh Karimpouli and Patricia Martínez-Garzón have received funding from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme 101076119 for project QUAKEHUNTER.

Original publication:

Karimpouli, S., Martínez-Garzón, P., Núñez-Jara, S. et al. Preparatory phase of large earthquakes illuminated by unsupervised categorization of earthquake catalog features. Nat Commun 17, 4024 (2026). 
DOI: 10.1038/s41467-026-72279-x