Seismology: innovation also goes through artificial intelligence

How is artificial intelligence used in earthquake analysis?
DT: There are several applications. Specifically, during my academic career I worked on improving the performance of existing models (physical, statistical, mathematical, AI-based), such as in the project dedicated to seismic denoising. To understand what this means, it’s important to know that seismic records can also contain signals produced by volcanic eruptions, ocean waves, road traffic, and other types of noise: all elements that contribute to ground motion recorded by seismic stations. To better analyze earthquakes, we need to extract this noise using AI. Our team drew inspiration from a paper published by Mitsubishi Electric Research Laboratories on audio denoising. Think of the cocktail party paradox: you are in a room with multiple people talking at the same time, and your goal is to listen to just one conversation. So you need to separate the different tracks—the noises we mentioned earlier (eruptions, waves, explosions, etc.)—and to do this you need a dataset containing all these signals so that the AI can reconstruct the single conversation you want to hear, the only signal you are interested in. The task our research team assigned to AI was to understand what is an earthquake and what is not, filtering out all other noise. In our project, we used the Stanford Earthquake Dataset with 1.2 million seismic traces recorded between 1984 and 2018, along with a catalog of noise identified by experts.

LL: One of the first projects I worked on was for my Master’s thesis in Data Science, where I combined AI with laboratory data to predict a specific element called “stress”, that is, the force acting on rocks in the Earth’s crust which, when very high and exceeding the rocks’ resistance, accumulates along faults and causes an earthquake. This value cannot be measured directly in reality because faults are located too deep underground; however, it can be measured in the laboratory.

Currently, colleagues at Sapienza University are developing machines (BRAVA2 and BIGBIAX) used to study fault slip, with the aim of obtaining useful information to analyze the seismic process based on selected parameters. Another project I worked on during my PhD focused on the 2016 Amatrice–Norcia–Visso seismic sequence. Usually, after a strong earthquake, smaller aftershocks occur, but it is rare for these to happen before the main event. The Norcia earthquake was a rare case, because the main and strongest shock was preceded by many smaller ones. In this project, we tried to distinguish seismic wave data (seismograms) recorded before and after the main shock—information that is difficult for the human eye to detect. To do this, we used AI. This project was particularly interesting because it showed how artificial intelligence can support humans in analyzing the evolution of multiple seismic events, something that traditional seismological studies cannot easily capture.

Laura, during your PhD you worked in Los Alamos, New Mexico (USA): what did you focus on?
I worked on studying a hypothetical AI model capable of performing multiple tasks. To better understand, think of ChatGPT: a single model capable of handling multiple tasks—answering questions, summarizing texts, completing sentences, and so on. These are different tasks, but they can be managed by a single structure. In earthquake science, however, we still rely on separate models (one focuses on predicting a specific value, another on a different one, and so on). What we aim to develop is a model capable of doing multiple tasks—just like ChatGPT. At Los Alamos, we started from a specific audio model developed by Facebook (if you think about it, seismograms and audio are similar, as they are both signals with oscillations). We created a prototype and trained it with seismological data with the goal of using it for additional tasks. The results were better than expected: we found that an audio model trained in seismology performs better than one designed only for seismology. After completing this work, we thought: “This is our starting point.” It’s a positive and optimistic beginning, though it still does not always work perfectly. What I am currently working on at ETH Zurich is developing the Los Alamos project further, improving the prototype and configuring it so that it can be used for specific tasks of interest.

Daniele, can the information obtained through AI be useful for post-earthquake emergency response or for prediction?
In the future, if research progresses well, yes. At the moment, using AI-derived information to assist emergency response units immediately after an earthquake is still in the design phase. However, some prototypes are currently being tested, although many steps and validations are still required before such a model can be deployed. That said, research in this area is active. I worked on a project in collaboration with the National Institute of Geophysics and Volcanology (INGV) on this very topic. In Italy, there are about 600 seismic stations, and when an earthquake occurs, these stations detect two types of waves: the P-wave, which is the first to arrive and provides information about the onset of the earthquake, and the S-wave, which arrives later and is important for determining its magnitude. However, the signal does not reach all stations at the same time: those closest to the epicenter receive it first, while more distant stations receive it later and in a weaker form. Once the signal reaches the nearest station, researchers work to predict—using AI—key values to be shared with other stations that would otherwise receive them too late. Among these values, for example, is the level of ground shaking. This information is crucial for assessing the impact on buildings, identifying the most affected areas, estimating damage, and deciding where to intervene urgently.