Improving deep learning for seismic phase picking with symmetries

Background : Earthquakes are among the most destructive natural hazards. While large earthquakes are rare, everyday thousands of tiny earthquakes occur. Mapping this myriad of small earthquakes is essential for our understanding of the buildup of large earthquakes. In recent years, deep learning methods have increased the number of earthquakes we can detect almost 10 fold. The most notable improvement here has happened for so-called seismic phase pickers : models detecting when a seismic wave has been recorded in a continuous waveform. While finding a lot more events, these models have also shown to be unstable : slight shifts in timing, rotations of the traces, or mirroring can completely destroy the predictions. These symmetry violations are unphysical and lead to inconsistent earthquake catalogs.

Objectives of the internship : The objective of this internship is to implement and test methods to enforce these symmetries in deep learning models for seismic phase picking. In particular, the intern will experiment with equivariant neural networks and data augmentation strategies. The work will be conducted based on the SeisBench framework that provides access to benchmark datasets, training codes and comparison models. If a successful new model is developed, this can be integrated into SeisBench, providing experience on contributing to open source development.

Prerequisites : The intern should pursue a degree in geophysics/data science/computer science or similar. Prior experience with programming in Python is required. Some knowledge of seismology or deep learning are helpful but not mandatory. A strong interest towards methodological developments in AI and a curiosity towards state-of-the-art research will be appreciated. The work environment is interdisciplinary, so communication and facilitation skills are required. The internship will be conducted in English, so fluency in English is required.

Environment : The intern will be based at ISTerre, the Earth science laboratory of Grenoble. At ISTerre, they will be supervised by Jannes Münchmeyer (expert in machine learning for seismic data processing and open source development) and Sophie Giffard-Roisin (expert in AI across solid earth geophysics).

Contact : Send an email as soon as possible to munchmej univ-grenoble-alpes.fr, with your CV and a brief explanation on why you want to apply and your interest in the topic.