A Deep Learning Framework for Pg/Sg/Pn/Sn Phase Picking and Its Nationwide Implementation in Chinese Mainland

Abstract Accurate seismic phase picking is critical for earthquake monitoring, tomography, and early warning systems. While prior deep learning models have achieved high performance on benchmark data sets, their applicability in real‐world, continuous waveform scenarios remains limited—particularly for phases beyond Pg and Sg. In this study, we develop a deep neural network trained on the Comprehensive Data set of the Chinese Seismic Network (CSNCD), the largest labeled seismic data set in the world, to simultaneously detect Pg, Sg, Pn, and Sn phases from continuous seismic waveform data. Our model employs an extended input window (102.40 s) and combines convolutional and recurrent architectures to capture long‐range temporal dependencies. It achieves F1‐scores of 0.832(Pg), 0.864(Sg), 0.814(Pn), and 0.491(Sn) on the CSNCD test set. When the analysis window fully contains the seismic signal, the and F1‐scores increase to 0.888 and 0.692, respectively. Operational deployment across Chinese mainland further demonstrates the model's robustness. In most provinces in China, our model achieves higher recall than PhaseNet, while significantly reducing false positives. This work presents the first nationwide real‐time application of a deep learning‐based multi‐phase seismic picker in China, offering a practical and scalable solution to support earthquake monitoring and subsurface imaging.