ISTerre seminar

---

Using Lab Earthquakes, Data Science and Machine Learning to Advance Understanding of Earthquake Physics and Forecasting

Thursday 16 October 2025 - 11h00

Chris Marone - Dipartimento di Scienze della Terra La Sapienza Università di Roma, Italia
--- 
I summarize recent works on earthquake physics that include lab results on the full spectrum of slip modes, from fast to slow, and applications of machine learning to seismic data that probe the evolution of fault properties during the seismic cycle. For the lab works we now have a high quality understanding of the full range of slip modes from slow, quasi-dynamic failure to fast elastodynamic slip events. These data include repetitive events under controlled conditions and provide an opportunity to apply machine learning (ML) to build predictive models of the underlying physics. Such works show that ML models can use microearthquakes (AE) that originate in the lab fault zone to predict the timing of labquakes, their magnitude, and the fault zone stress state during the complete lab seismic cycle. Moreover, we have developed ML models that use changes in fault zone elastic properties during the lab seismic cycle to predict lab earthquakes. Our work shows that labquakes are preceded by a cascade of AE events within the fault zone and that there are systematic changes in fault zone elastic wave speed that foretell catastrophic failure. The methods include traditional ML techniques based on regression, deep learning (DL) prediction of failure times, and also DL methods to autoregressively forecast fault zone shear stress. The lab data suggest sensible connections to tectonic faulting and we have found, using data from the 2016 seismic sequence of central Italy, that DL models can successfully distinguish seismic waves pre/post mainshock. DL models using seismic waves that pass through the hypocentral region of the 2016 M6.5 Norcia earthquake successfully distinguish between foreshocks, aftershocks and time-to-failure (TTF). Binary and N-class models defined by TTF correctly identify seismograms in test with > 90% accuracy. These data are in accord with lab and theoretical expectations of progressive changes in crack density prior to abrupt changes at failure and gradual postseismic recovery. DL model performance is lower for band-pass filtered seismograms (below 10 Hz) suggesting that DL models learn from the evolution of subtle changes in elastic wave attenuation. If such models are generalizable they could dramatically improve earthquake early warning and seismic hazard analysis in areas of fluid-injection induced seismicity.

---

Organizing team : Grands Séminaires ISTerre

Amphithéâtre Killian, Maison des Géosciences, 38400 Saint Martin d'Hères

Informations de visio :

https://univ-grenoble-alpes-fr.zoom.us/j/93103679388?pwd=h0D9jyzkjKYcn3AyqvcMExlYGYaNlC.1