From controlled earthquakes to the monitoring of unstable slopes

Earthquakes are among the greatest natural hazards. However, it is almost impossible to observe in nature what happens immediately before a rock fracture, because no one knows in advance where and when an earthquake will strike. ETH Zurich’s BedrettoLab turns this situation on its head: in a tunnel in Ticino, with around one and a half kilometres of rock above it, researchers inject water into the rock, thereby deliberately triggering minute vibrations, whilst a dense network of sensors records every movement.

This generates volumes of data that are difficult to analyse using standard seismological tools. This is where the project came in, carried out in 2024 and 2025 by Anne Obermann of the Swiss Seismological Service (SED) in collaboration with Léonard Seydoux of the Institut de physique du globe de Paris (IPGP): A method of unsupervised machine learning developed in France was designed to identify patterns in the recordings that no one had specifically been looking for. The collaboration was supported by the Germaine de Staël funding programme.

Ms Obermann, at BedrettoLab you deliberately trigger tiny earthquakes deep within the rock. What can you observe there that is not visible either in a conventional laboratory or on the Earth’s surface?

At the BedrettoLab, in a tunnel running beneath the Alps, we can study earthquakes directly where they occur: deep within the rock and under natural pressure and stress conditions. By injecting water in a controlled manner, we alter the pressure within small rock fractures and trigger tiny earthquakes. High-precision sensors measure how fractures form, propagate and release stress. This enables us to gain new insights into how earthquakes originate and propagate. The scale of the BedrettoLab is unique in its kind. In a conventional laboratory, the samples are very small and it is difficult to realistically replicate the natural stress conditions of the Earth’s crust. At the Earth’s surface, on the other hand, the distance to the earthquake’s focus is too great to observe these processes directly and in detail.

Your sensors record data continuously and at very high resolution. Where do traditional seismological analysis methods reach their limits when dealing with such large volumes of data?

Traditional seismological methods usually search specifically for earthquake signals that exhibit certain characteristics. At BedrettoLab, however, vast amounts of data are generated, within which even very weak and superimposed signals are hidden. Many of these remain undetected because they are drowned out by noise or do not correspond to the expected patterns. Furthermore, manual analysis is time-consuming and can hardly be applied to continuous measurements on this scale. New, automated methods therefore help us to detect even the smallest events, identify patterns in the data and gain a better understanding of the processes taking place within the rock.

In the project with the IPGP, you searched the data using an unsupervised machine learning method – in other words, without telling the algorithm what to look for. What did it find that had previously been overlooked? And where did the difficulties lie?

The method identified significantly more seismic events than in the previous reference catalogue: in the experiment studied, there were around nine times as many. Furthermore, the method separated earthquake signals, electronic noise and other interference, as well as signals from active seismic sources, much more quickly and reliably than conventional methods. A particular challenge was processing the large volumes of data at extremely high sampling rates in the megahertz range, which required the calculations to be carried out in parallel. Further investigation is needed to determine whether all the patterns found can be unambiguously interpreted in physical terms.

The project title refers to precursor signals of earthquakes. How close is the research to being able to predict earthquakes? And what do your findings mean for Switzerland, for example in relation to deep geothermal energy?

We are still a long way from being able to reliably predict earthquakes. However, our results show that unsupervised machine learning can identify additional seismic events and certain signal groups, including tremor-like signals prior to the activation of a fault zone. Whether such signals can actually be classified as precursors requires further investigation. For Switzerland, this approach could be particularly relevant in the context of deep geothermal energy. A better understanding of fracture processes and the associated seismic activity could, in the long term, help to better assess the risks involved in rock stimulation. However, the method does not currently allow for the reliable prediction or prevention of earthquakes in deep geothermal energy.

How close is research to predicting earthquakes? And what potential does unsupervised machine learning have in seismology in general?

We are still a long way from being able to reliably predict earthquakes. However, our results show that unsupervised machine learning can identify additional seismic events and signal groups that have received little attention to date. This is not only of interest for research into fracture processes and deep geothermal energy. The same fundamental idea can also be applied to other geophysical processes: instead of searching exclusively for known signals, we can systematically scan continuous measurement data for recurring patterns and changes. This could open up new possibilities, for example for monitoring unstable slopes. We still need to investigate how reliably this works and under what conditions.

The method originates from France, whilst the data comes from Ticino. What did the visits to Paris and Grenoble, as well as the return visit to Switzerland, achieve that would not have been possible via video conference? And what did the young researchers in your team take away from these experiences?

The face-to-face meetings in Paris, Grenoble and Switzerland facilitated a direct scientific exchange that goes beyond what is possible via video conferencing. Spontaneous ideas and new approaches, in particular, often emerge more easily in face-to-face discussions when people discuss data, methods and open questions together. This allowed the collaboration between the French and Swiss teams to be deepened and the joint research to be further developed. For the young researchers, the visits were also an opportunity to forge contacts and expand their scientific network beyond their own institution – an important foundation for future collaborations.

The funding programme covers only travel and accommodation costs. What impact does such a small amount actually have, and what advice would you give to researchers wishing to apply?

The budget is, of course, not comparable to the funding for large-scale research projects. Rather, it aims to facilitate collaboration between individual researchers. This is precisely where its strength lies: researchers can spontaneously exchange ideas outside large research consortia, develop them further together and try out new approaches without first having to set up a major project. In our case, the idea of collaborating with France arose spontaneously at a conference; we had not originally set aside any funds for this. The Germaine de Staël grant made this collaboration possible in the first place. I therefore advise researchers to make the most of such opportunities, to actively seek out suitable partners and to be open to shared ideas – it is precisely from small encounters that long-term collaborations can emerge.

I find it particularly exciting that the impact now extends beyond the original project. We initially applied the method to seismic data from controlled experiments in Bedretto. During a reciprocal visit, the idea arose to try out the approach on data from the Blatten ice and rockfall. The initial results were very promising. Thus, a collaboration to research earthquakes has given rise to a new research question concerning the monitoring of natural hazards.

To me, this highlights the particular value of such funding programmes: they not only finance travel, but also create the conditions for personal exchange, new ways of thinking and collaborations, the potential of which often only becomes apparent through working together.

The project has now been completed. What does the future hold for the collaboration and the method? What question would you like to tackle next?

When Léonard Seydoux recently visited us in Zurich, I was in the process of analysing seismic data from the Blatten ice and rockfall in May 2025. That’s when the idea arose to apply our method to this data as well. No sooner said than done: just one day later, we had our first, very promising results. They suggest that, using unsupervised machine learning and data from a single station, we can accurately track activity on a slope. This is not yet a fully operational early-warning system, but it offers an exciting prospect for the future monitoring of natural hazards. We would now like to develop this approach further. Perhaps there will be an opportunity to continue our collaboration with another Germaine-de-Staël grant!