Slow slip events at subduction zones using deep learning

6 mois, Janvier à Juin 2023

• Contexte du stage
This master project is part of the DEEP-trigger ERC project (PI A. Socquet, https://deeptrigger.osug.fr/) that aims at understanding the preparation mechanisms of large earthquakes. It could be continued and widened during a PhD for which a scholarship is already guaranteed through the project DEEP-trigger funded by the European Research Council. The student will interact with the group of students and researchers involved in the project, and with collaborators in France and abroad. The student will work within the “Seismic cycle and transient deformation” team in ISTerre laboratory, Grenoble. A solid background in Mathematics, Geophysics and Computation is required for this master project.

• Objectifs du stage
The discovery of slow slip events at plate boundaries in the 2000’s was made possible by the continuous monitoring of ground deformation by geodetic networks, and constituted a change of paradigm for the understanding of the earthquake cycle and of the mechanics of the faults interface. Twenty years later, the characterization of the full slip spectrum, and the understanding of the link between slow slip and associated seismological signals are hindered by our capacity of detecting, in a systematic manner, all the slow slip events, including the smallest ones.
The development of in-situ geophysical monitoring generates nowadays huge data sets, and machine learning techniques are now widely used to detect and characterize earthquakes.

In the frame of his PhD, Giuseppe Costantino has developped the first deep learning method that is successful in detecting slow slip events in raw geodetic time series. It is a supervised method that combines a convolutional and an attention-based neural networks, and that is trained on an ultra-realistic synthetic training set. Several group are trying to develop such methods and so far they fail because the noise in the geodetic time-series is huge and because such spatio-temporal data structures are tricky to handle with deep learning methods.
When applied to real data in Cascadia, our method allows to detect a total of 116 SSEs that compare well to existing benchmark, which is a strong indication that the method is robust.

The objective of this internship is to apply and test Giuseppe’s method to geodetic data on other subduction zones in the world (Japan for example) in order to construct unprecedented catalogues of Slow slip Events, to compare the results using independent data sets (e.g. other published catalogs of SSEs, catalogs of tremors LFEs or earthquakes) in order to assess the efficiency of the method and understand the relation between Slow slip events and seismological signals.
To achieve these objectives, these are the following steps to be followed during the internship :
– 1 - generate synthetic geodetic time series to train the model following our method of noise and deformation signal generation
– 2 - train the model and evaluate its performance on the synthetic test set
– 3 - apply the method on real data
– 4 - evaluate and interpret the results by comparing with independent indicators and data

• Points-forts
This internship deals with a ’hot’ subject in earth science, and uses novel machine learning techniques.

• Pour postuler
Send email with CV and cover letter to Anne Socquet.