Rapid magnitude estimation for Taiwan earthquake early warning via frequency-domain multi-station deep learning

Abstract Accurate and rapid earthquake magnitude estimation is essential for effective Earthquake Early Warning Systems (EEWS), particularly in seismically active regions such as Taiwan. In this study, we propose a frequency-domain multi-station deep learning approach that integrates a Fourier Neural Operator (FNO) and a Long Short-Term Memory (LSTM) network to estimate earthquake magnitudes using near real-time waveform data. The model is trained on 7,025 events recorded by Taiwan’s networks between 2012 and 2024. By jointly analyzing three-component waveforms from the ten nearest stations and incorporating hypocentral distance corrections, the proposed method achieves high accuracy within seconds after the fourth triggered P-wave arrival. Evaluation on a held-out test set demonstrates significant improvements over the current Pd-based single-station method currently used by the Central Weather Administration (CWA), with lower mean absolute error, root mean squared error, and prediction variance. In real EEWS scenarios from January to June 2025, the model consistently outperformed the operational method, even when using early-stage location estimates, and showed further gains when provided with precise hypocenters from the catalog. The proposed approach combines robustness, and computational efficiency, enabling practical deployment in EEWS of Taiwan.