Skip to main content
European Commission logo
español español
CORDIS - Resultados de investigaciones de la UE
CORDIS
CORDIS Web 30th anniversary CORDIS Web 30th anniversary

Detection of Early seismic signal using ARtificiaL Intelligence

Periodic Reporting for period 2 - EARLI (Detection of Early seismic signal using ARtificiaL Intelligence)

Período documentado: 2022-07-01 hasta 2023-12-31

Earthquakes caused nearly one million fatalities in the last two decades and billions of euros of economic losses. The hazardous nature of earthquakes is largely due to their unpredictability. We do not know how to predict earthquakes as no consistent warning signs have ever been detected. Early warning systems have been developed but they are based on data recorded after the initiation of the seismic rupture and therefore only provide, at best, a few seconds of warning. These systems also suffer from systematic underestimation of the magnitude of the largest earthquakes : they typically cannot make the difference between a Magnitude 9 and a Magnitude 8 earthquake. This is a major issue for tsunami warning, as the tsunami expected in the first case is about 30 times as large as in the second one.
EARLI aims at designing Artificial Intelligence (AI) algorithms to improve both the rapidity and accuracy of earthquake and tsunami early warning systems by extracting the information contained in recently identified low-amplitude signals (named PEGS for Prompt Elasto-Gravity Signals), which propagate at the speed of light. The second, more exploratory, objective of EARLI is to design AI algorithms to search for earthquake precursors in seismic and geodetic data.
We designed a prototype AI algorithm capable to estimate almost instantaneously the magnitude of large earthquakes from PEGS (and PEGS exclusively) recorded by seismometers in Japan (Licciardi et al., Nature, 2022). The algorithm outperforms existing early warning systems for Magnitude ≥ 8.3 earthquakes, which can be already transformative for tsunami warning. We applied this algorithm to Chile and concluded that the present seismic network there is to sparse for PEGS-based early warning applications (Arias et al., JGR Solid Earth, 2023).
In parallel, we designed a complementary early warning AI algorithm for Magnitude ≤ 8 earthquakes. This approach uses 3 seconds of "classical" seismic signal recorded by a single seismic station to rapidly estimate the earthquake location and magnitude (Lara et al., in review at JGR Solid Earth). We implemented this algorithm in the early warning system of Peru, currently in construction. We are now complementing the Peruvian system with the PEGS-based algorithm.
We stacked all the GPS data ever recorded before large earthquakes after projecting the measurements in the direction of expected motion assuming precursory slip on the fault. We found evidence that earthquakes likely start, on average, with a precursory phase of slow slip in the ~2 hours preceding large events (Bletery and Nocquet, Science, 2023).
The development of the PEGS-based AI algorithm and the evidence of precursory slow slip preceding large earthquakes represent significant progresses beyond the state of the art. The former opens the way to the field of light-speed seismology with considerable potential for improvements in early warning systems. The latter indicates that earthquake prediction is not physically impossible and, if confirmed, will surely opens the way to a quest for earthquake precursors.
In the second part of the project, we hope to lower the magnitude at which the PEGS-based algorithm starts working, we plan to incorporate it to the early warning system of Peru (as a complement to the other algorithm we already installed) and we aim to design a global version of this algorithm that could offer a global early warning cover.
We also plan to apply AI approaches to lower the noise in GPS data with the aim of identifying precursory signals at the scale of individual events.
Illustration of our AI algorithm estimating earthquake magnitude from light-speed gravity signals