A Real-Time and Data-Driven Ground-Motion Prediction Framework for Earthquake Early Warning

Abstract

ABSTRACT The ShakeAlert earthquake early warning system in the western United States characterizes earthquake source locations and magnitudes in real time, issuing public alerts for areas where predicted ground-motion intensities exceed a threshold value. Although rapid source characterization methods have attracted significant scientific attention in recent years, the ground-motion models used by ShakeAlert have received notably less. This study develops a data-driven framework for earthquake early warning-specific ground-motion models by precomputing and incorporating site-specific corrections, while using a Bayesian approach to estimate event-specific corrections in real time. The study involves analyzing a quality-controlled set of more than 420,000 seismic recordings from 1389 M 3–7 events in the state of California, from 2011 to 2022. We first compare the observed ground motions to predictions from existing ground-motion models, namely the modified Boore and Atkinson (2008) and active crustal Next Generation Attenuation (NGA)-West2 ground-motion prediction equations, before implementing a new Bayesian model optimized for a real-time setting. Residual analysis of peak ground acceleration and peak ground velocity metrics across a host of earthquake rupture scenarios from the two ground-motion models show that the active crustal NGA-West2 model is better suited for ShakeAlert in California. In addition, the event-terms calculated using our Bayesian approach rapidly converge such that errors from earthquake magnitude estimation can be corrected for when forecasting shaking intensity in real time. Equipped with these improved ground-shaking predictions, we show that refined ShakeAlert warnings could be issued to the public within as soon as 5 s following ShakeAlert’s initial warning. This approach could be used both to reduce prediction uncertainties and thus improve ShakeAlert’s alerting decision.