Implementation of machine learning approaches to monitor pre-eruptive swarms at Piton de la Fournaise volcano

Eruptions of Piton de la Fournaise volcano (Reunion Island, France) are preceded by intense pre-eruptive seismicity swarms characterized by hundreds, or even thousands of micro earthquakes (magnitude < 2). These volcano-tectonic events are triggered by the upward migration of magma toward the surface and their location provides important information regarding the future eruption location. Yet, regarding the large number of earthquakes, it is difficult to locate them all during seismicity swarms. Hence, we have implemented an approach at the Piton de la Fournaise Volcano Observatory (OVPF-IPGP) based on machine learning to automatically detect and locate these events.. First, we use PhaseNet to pick P and S waves from 17 seismic stations installed on and around the volcano. Then, phase association and source location are done using a Graph Neural Network (GNN) approach called GENIE (Graph Earthquake Neural Interpretation Engine). To implement GENIE specifically at Piton de la Fournaise, we trained the code with seismic stations and velocity models used by OVPF-IPGP to monitor the volcano.. After phase association, we perform a final hypocenter localization using the probabilistic approach of NonLinLoc. To study the results quality, we compare origin time and source location to the OVPF manual catalog as well as a catalog resulting from template matching and double-difference relocation. In particular, we focus on pre- and syn-eruptive time-periods for multiple eruptions since 2014 in order to investigate the effect of elevated seismicity rate and eruptive tremor on the performance of the workflow. We also assess to what extent the quality of the resulting automatic locations are sufficient to provide an indication of the future eruption site without expert manual input.Applying this approach allows us to improve the monitoring of the seismicity at Piton de la Fournaise volcano. A work in progress is to implement the same approach over the entire island of La R&#233;union, which will enable the monitoring of other active areas in the region.