Adapting Deep Learning Phase Detectors for Seismic Array Processing

Abstract Deep learning (DL) has considerably advanced single‐station seismic phase detection, but its integration into seismic array processing remains largely unexplored. We evaluate three strategies for adapting the previously developed TPhaseNet detector for regional phase detection to array data: (a) ensemble detection through stacking of single‐station detections, (b) beam detection using models trained directly on array beams, and (c) array detection in which a model with multi‐channel input learns array processing implicitly. Using more than 25 years of regional seismicity recorded at five arrays in Northern Europe and the European Arctic and an array in Asia, we assess performance on both event windows and fully labeled continuous data. All DL‐based array methods outperform single‐station detection and significantly improve S‐wave detection over standard beamforming and STA/LTA processing. Ensemble detection performs best for most arrays, while beam detection excels in challenging environments. The array detection model provides competitive performance at substantially lower computational cost. Our results demonstrate the feasibility and advantages of DL‐d