Magnitude estimation of large earthquakes based on high-rate GNSS and end-to-end deep learning

Traditional Earthquake Early Warning (EEW) systems provide initial prediction from empirical relationship between P-wave and final magnitude, which often saturate during large events (MW > 7). High-rate GNSS data address this limitation by directly capturing co-seismic displacement, but struggle with unilateral ruptures or sparse earthquake records. In contrast, deep learning presents an end-to-end solution for this challenging problem, with its powerful nonlinear fitting capability. In this study, we develop a real-time magnitude estimation approach for large earthquakes, which couples high-rate GNSS data with Temporal Fusion Transformer (TFT). Based on the Japan Trench structure, we simulate over 50000 earthquakes (MW 7.0 ~ 9.5) and generate displacement waveforms across multiple GNSS stations. The end-to-end model uses raw three-component waveforms as input and outputs real-time magnitude series, while quantifying the uncertainty by quantile loss function. In simulated earthquakes of testing dataset, the deep learning model achieves over 93% accuracy within 60 s after P-wave arrival. Even with limited station availability, the model maintains ~ 80% accuracy, significantly outperforming the Peak Ground Displacement (PGD) scaling law, which delivers only ~ 70%. In real earthquake cases, the deep learning model estimates the final magnitude within 90 s, whereas traditional method shows a notable delay, requiring at least 120 s for a reliable alert during the MW 9.1 event. The results demonstrate that deep learning can effectively extract critical information from raw waveforms, which can offer substantial improvement to current EEW systems.