A Prompt Elastogravity Signal–Based Rapid Source Characterization Module for Tsunami Warning Support in Alaska

Abstract

State-of-the-art earthquake early warning systems relying on seismic waves tend to underestimate the magnitude of large earthquakes, which typically results in dramatic underestimation of tsunami forecasts. Systems relying on other types of data (W phase, Global Navigation Satellite System) may provide more reliable magnitude estimates for very large seismic events but at the cost of slower warning. Prompt elastogravity signals (PEGS) have shown potential to provide unsaturated magnitude estimates more rapidly than systems based on other data. In this study, we present the first fully automated pseudo-real-time implementation of a PEGS-based approach on the Alaska seismic network. We designed a continuous pipeline that combines an automatic detection module, a real-time waveform processing protocol, and a PEGS-based deep-learning model for rapid source characterization. To evaluate operational resilience, we conducted a 13-month continuous playback, alongside targeted testing of 42 moderate regional earthquakes and 37 major teleseismic events recorded between 2018 and 2025. By implementing a multiparametric alert criterion—requiring a stable Mw ≥7.8 inference, network consensus, and regional location—the system achieved a strict zero false-alarm rate. Crucially, the pipeline accurately tracked the magnitude and focal mechanism of three major regional earthquakes (the 2018 Mw 7.9 Kodiak, 2020 Mw 7.8 Shumagin, and 2021 Mw 8.2 Chignik events) within 90–120 s of their onset. Although highly robust to typical automated picking errors (± 2 s), comparative tests reveal that the system’s sensitivity limit is strictly bound to network density. We are presently running a real-time test system in partnership with the Alaska Earthquake Center. In addition to its validation against simulated real-time data presented in this study, the regular occurrence of Mw ≥7.8 earthquakes in the region offers a concrete possibility to test the system against real-time-acquired data in the next 2–4 yr.