RATE: A Retrieval-Augmented Transformer for Regional Earthquake Early Warning

Abstract

Accurate and timely seismic intensity prediction is essential for effective regional earthquake early warning (EEW). This study presents a retrieval-augmented Transformer (RATE) model that leverages historical seismic events to enhance regional ground motion predictions. Upon receiving initial P-phase signals, RATE retrieves similar past events based on waveform similarity and integrates them into a Transformer-based prediction pipeline. This design allows the model to adapt the diverse seismic contexts and generalize across regions. The experiments on datasets from Japan and Taiwan demonstrate that the RATE consistently outperforms baseline models in terms of intensity estimation accuracy and alert precision. These results highlight the potential of a retrieval-augmented (RA) framework to enhance real-time EEW capabilities in diverse seismic regions.