The potential of prompt elasto-gravity signals and graph neural networks for tsunami early warning

Prompt Elasto-Gravity Signals (PEGS) are light-speed gravity perturbations that can be recorded by broadband seismometers before the arrival of P waves. This characteristics has raised interest for potential early warning applications but the emerging nature of PEGS and their extremely small amplitudes (nm/s²) have challenged their operational use. We developed a deep learning approach to rapidly estimate the magnitude and location of large earthquakes from PEGS. In order to optimize the performances, we designed a graph neural network (PEGSGraph) capturing the geometrical information of the seismic network. This approach is not subject to saturation and can reliably estimate the magnitude of M_w ≥ 7.6 earthquakes within 2 minutes from initiation in Alaska, making it a viable solution for tsunami warning. We are currently testing possible implementations of PEGSGraph into the tsunami early warning systems of Peru and Alaska and including GNSS version in the deep learning framework.