What can neural networks tell us about earthquake rupture predictability? Using machine learning for real-time magnitude estimation
Geb 06.42 - Raum 001 (Seminarraum) / Online
Jannes Münchmeyer is a PhD student at GFZ Potsdam and Humboldt-University Berlin under the supervision of Frederik Tilmann and Ulf Leser. He is a mathematician by training and has previously worked in bioinformatics and biomedical text mining. In his research, he develops machine learning approaches for seismic data, focusing on applications to earthquake early warning and rupture predictability. He is a developer of the SeisBench framework (http://seisbench.org) for machine learning in seismology.
The ruptures underlying very large earthquakes can last from several seconds up to minutes. To which extent the final magnitude of an earthquake can be determined while its rupture is still ongoing is an open question. Evidence has been brought forward for both early predictability and full stochasticity. The answer to this question defines the fundamental limitations of earthquake early warning and is thereby integral for mitigating seismic hazard.
One appealing option to study rupture predictability is building real-time magnitude assessment systems. However, until recently such systems lacked either accuracy or timeliness to use them for studying rupture predictability. In this talk, we present a novel deep learning model for real-time magnitude estimation. We conduct an intensive study of this model to highlight its advantages and limitations. Finally, using this model, we study the question of rupture predictability of very large events.