Earthquake magnitude and location estimation from real time seismic waveforms with a transformer network

SUMMARY Precise real time estimates of earthquake magnitude and location are essential for early warning rapid response. While recently multiple deep learning approaches fast assessment earthquakes have been proposed, they usually rely on either seismic records from a single station or fixed set stations. Here we introduce new model real-time estimation using the attention based transformer networks. Our approach incorporates waveforms dynamically varying stations outperforms baselines in both performance. Furthermore, it classical algorithm considerably shows promising performance comparison to localization algorithm. is applicable prediction provides realistic uncertainty probabilistic inference. In this work, furthermore conduct comprehensive study requirements training data, procedures typical failure modes. Using three diverse large scale data sets, targeted experiments qualitative error analysis. analysis gives several key insights. First, can precisely pinpoint effect data; example, four times larger reduces average errors by more than half, required factor four. Secondly, basic systematically underestimates events. This issue be mitigated, some cases completely resolved, incorporating events other regions into through transfer learning. Thirdly, highly precise areas with sufficient but strongly degraded outside distribution, sometimes producing massive outliers. suggests that these characteristics not only present our model, most models published so far. They result black box modeling their mitigation will likely require imposing physics derived constraints neural network. These need taken consideration practical applications.