Siamese Earthquake Transformer: A Pair‐Input Deep‐Learning Model for Earthquake Detection and Phase Picking on a Seismic Array

Earthquake detection and phase picking play a fundamental role in studying seismic hazards and the Earth’s interior. Many deep‐learning‐based methods, including the state‐of‐the‐art model called Earthquake Transformer (EqT), have made considerable progress. However, the processing of low signal‐to‐noise ratio (SNR) seismograms remains a challenge. Here, we present a pair‐input deep‐learning model called Siamese Earthquake Transformer (S‐EqT), which achieves good performance on low SNR seismograms using the latent information in the deep‐learning black box of the pre‐trained EqT model on a seismic array. We compare the EqT and S‐EqT models on 2 weeks of continuous seismograms recorded by stations around northern Los Angeles region in California. In addition to showing a good performance similar to the EqT model on high SNR seismograms, the S‐EqT model retrieves ∼40% more reliable picks from low SNR seismograms, resulting in better earthquake characterizations. Our method provides a novel perspective on earthquake monitoring by highlighting the importance of hidden responses inside a deep‐learning model and shows its great potential for seismology. Plain Language Summary Recently, deep‐learning‐based methods have made great progress in detecting earthquakes and picking seismic phases. It seems that previous deep‐learning models have reached their performance limits using large public seismic data sets and advanced network designs. However, their performance in processing low signal‐to‐noise ratio seismograms is still insufficient. In this study, we developed a deep‐learning model (Siamese Earthquake Transformer) to break the bottleneck by leveraging latent information in the deep‐learning black box and picking seismic phases on a seismic array. Our model performs better than previous deep‐learning models and provides more reliable picks in processing low SNR seismograms using the same training dataset. A significant amount of additional high‐quality seismic arrival times can contribute to characterizing earthquakes, which is essential for mapping the geometry of fault zones and illuminating the Earth’s structure. Hence, this study may significantly enhance our knowledge of the Earth’s interior. Key Points Siamese Earthquake Transformer (S‐EqT) is the first model to leverage the latent information in the deep‐learning black box to pick seismic phases The S‐EqT model processes seismograms on a seismic array rather than a single station The S‐EqT model performs