We place particular emphasis on connecting observed phenomena to the underlying physical processes driving the sequences. From our findings, we propose a new way of thinking about how earthquakes happen: not as the result of one single process, but as the outcome of several overlapping physical processes, unfolding across different times and places, that together increase stress in the Earth's crust until a major quake is triggered (Martínez-Garzón and Poli, 2024).
Artificial intelligence (AI) is rapidly transforming how we study earthquakes. One major breakthrough in recent years has been the use of AI techniques to process massive amounts of seismological data. This has opened up exciting new possibilities—especially in understanding and potentially forecasting large earthquakes, including those created in controlled laboratory experiments that simulate fault behavior. As part of our ongoing research in the team, another key development has been an unsupervised AI-based method that analyzes patterns in earthquake catalog data. This approach does not rely on pre-labeled datasets, but instead detects subtle changes in seismic activity on its own (i.e. as an unsupervised machine learning method). It identifies distinct seismicity patterns, which we interpret as signals related to the changing stress along a fault—essentially, how close a fault is to slipping.
We first developed and tested this method using laboratory earthquake data, where we can precisely measure stress changes using a machine that applies controlled pressure to rock samples (Karimpouli et al., 2024). The AI method closely followed changes in fault stress throughout these experiments, suggesting it could be a powerful new tool for identifying when a fault is approaching failure. Encouragingly, we’re now beginning to apply this technique to real-world data.
Currently, the first applications of this method to two earthquakes (2023 MW 7.8 Kahramanmaras, Türkiye and 2009 MW 6.1 L’Aquila) are promising (Karimpouli et al., submitted). Preliminary results from two major earthquakes—the 2023 magnitude 7.8 Kahramanmaraş earthquake in Türkiye and the 2009 magnitude 6.1 L’Aquila earthquake in Italy—are promising (Karimpouli et al., submitted). These applications show potential for using AI to track fault stress in the field, which is an essential step toward real-world earthquake forecasting.
The 2023 Kahramanmaraş earthquake caused catastrophic loss of life and destruction. While attention understandably focused on this tragic event, our research continues to highlight another highly vulnerable region: the eastern Sea of Marmara, near the densely populated city of Istanbul. This area is considered one of the highest-risk seismic zones in Europe (Crowley et al., 2020). In the next phase of our project, we will focus on better understanding the conditions that might allow us to detect seismic or aseismic precursors—early warning signs—of major earthquakes in this region. We also aim to refine our AI methods to identify when a fault is transitioning from a stable to an unstable state, a critical clue that could indicate an impending rupture.Our ultimate goal is to move toward near-real-time testing, applying these techniques to ongoing seismic activity as it happens. While we're still far from being able to predict exactly when and where an earthquake will occur, our work is paving the way toward earlier detection and, hopefully, better preparation and resilience for communities at risk.
