Earthquake monitoring systems (EQS) play a critical role in seismic hazard assessment and tectonic studies. A complete and accurate earthquake catalog improves our understanding of fault mechanisms and supports the development of strategies to enhance public safety and the resilience of infrastructure. We propose a foundation model to improve EQS performance through a two-stage framework. Stage 1 involves training the foundation model in a self-supervised manner using a U-Trans architecture, which combines a U-Net encoder-decoder structure with a compact convolutional transformer in the bottleneck layer. At this stage, the model learns to reconstruct corrupted seismic waveforms in both the time and frequency domains, enabling it to extract well-structured and informative features in the latent space. Stage 2 focuses on downstream earthquake tasks. The latent features extracted by the encoder are flattened and concatenated as an additional input channel to the downstream models, effectively guiding them toward better performance. The U-Trans network is trained on more than 2 million three-component seismograms collected from three open-source datasets, ensuring diverse coverage and robust feature extraction. Extensive testing demonstrates that incorporating the encoder's latent features significantly improves the performance of several key downstream tasks, including seismic phase picking, earthquake location, magnitude estimation, and P-wave polarity classification. Analysis of the latent space reveals that the extracted features strongly correspond to P- and S-wave arrival times, which are crucial for many earthquake monitoring applications.
© 2026. The Author(s).