ROSERS—A Deep Learning Framework for Earthquake Early Warning and Its Interpretation

Earthquake early warning (EEW) systems are swiftly evolving from standalone physics-inferred methods (requiring computationally expensive inversions) to data-driven strategies for efficient earthquake hazard mitigation in real time. Besides being speedy in prediction, data-driven approaches such as artificial neural networks usually require minimal assumptions in the training and execution processes. This study discusses and attempts to interpret the data-driven EEW framework: ROSERS (Real-Time On-Site Estimation of Response Spectra) proposed by Fayaz and Galasso (2022). ROSERS aims to utilize the early non-damaging p-waves and the recording-site characteristics to predict the acceleration response spectrum (S_a(T) ) of the anticipated on-site ground motion waveform. The framework's efficacy is analyzed using an extensive database of ground motions, and it is observed that ROSERS leads to exceptional prediction power when implemented in a real-time backdrop. To provide a better interpretation of the framework, this study utilizes the concepts of explainable artificial intelligence (i.e., Shapley additive explanation, SHAP) to obtain insights into the decision-making process of the trained neural networks. Particularly, the cause-effect relationship of the computed latent variables and S_a(T) is explored. The analyses showcase that the two latent variables of the framework complement each other in capturing stiff short-period and flexible long-period S_a(T) thereby leading to excellent reconstruction power.