REWFERS: A Regional Early Warning Framework for Estimating Response Spectra

Abstract

Earthquake early warning (EEW) systems aim to provide critical information (i.e., magnitude, ground-shaking intensity, and/or potential consequences) about the characteristics of incoming earthquake waves at a target site to mitigate the hazard in real-time/near-real-time rapidly. In general, the EEW systems are classified into two groups: 1) on-site systems which use the early features of the incoming waves at the target site itself to enable preventive action and mitigate seismic risk; 2) regional systems which use the seismic waves recorded at sensors located closer to the epicenter to estimate ground shaking intensity at the target site and implement appropriate precautionary measures. Conventional EEW methods have often been based on the inferred physics of the fault ruptures combined with simplified empirical models. Based on the recent boost in computational resources, data-driven methods are widely accepted as effective alternatives for EEW. This study introduces a hybrid deep learning- and Bayesian statistics-based regional EEW framework called REWFERS (Regional Early Warning Framework for Estimating Response Spectra). The proposed framework is based on variational autoencoders (VAE), deep neural networks (DNN), gaussian process regressions (GPR), and spatial correlation models trained to estimate the acceleration response spectrum () at a target site in real-time using the early non-damage-causing p-waves recorded at the stations closer to the epicenter. The framework is trained and thoroughly assessed using a carefully selected extensive database of ground motions. Due to the well-known correlation of with structuresﾒ seismic response and resulting damage/losses, rapid and accurate knowledge of expected is highly beneficial to various end-users to make well-informed decisions