Abstract We present SeismicXM, a cross-task foundation model designed for single-station seismic waveform analysis. Unlike conventional approaches that target individual tasks, SeismicXM simultaneously performs seismic phase picking (Pg, Sg, Pn, and Sn), P-wave first-motion polarity classification, and event-type classification directly from raw three-component waveforms. The architecture integrates a convolutional encoder with a bidirectional Transformer backbone and multiple task-specific decoders, enabling shared representation learning while flexibly producing task-dependent outputs. SeismicXM is pretrained on the large-scale Comprehensive Dataset of Chinese Seismic Network dataset, which contains more than 45 million annotated phases, polarities, and event labels. Leveraging this diverse dataset, the model learns generalizable seismic representations that can be efficiently adapted to new regions with minimal labeled data. Benchmark experiments demonstrate strong performance: SeismicXM achieves F1 scores above 0.88 for Pg and Sg picking, and reaches F1 scores of 0.871 and 0.687 for Pn and Sn under full-waveform evaluation. The model also attains over 0.97 area under the curve in P-wave polarity classification and approximately 94% accuracy in event-type classification after fine-tuning on a small regional dataset. These results indicate that SeismicXM provides an effective and transferable solution for diverse waveform-processing tasks, highlighting the potential of pretrained multitask Transformers to streamline operational earthquake monitoring and analysis workflows.
Cai et al. (Tue,) studied this question.