U-Trans: a foundation model for seismic waveform representation and enhanced downstream earthquake tasks.

Area of Science:

• Geophysics
• Seismology
• Artificial Intelligence

Background:

• Earthquake monitoring systems (EQS) are vital for seismic hazard assessment and understanding fault mechanisms.
• Accurate earthquake catalogs are essential for public safety and infrastructure resilience.

Purpose of the Study:

• To develop a foundation model to improve the performance of earthquake monitoring systems.
• To enhance feature extraction from seismic waveforms for downstream earthquake analysis.

Main Methods:

• A two-stage framework utilizing a U-Trans architecture for self-supervised learning.
• Training on over 2 million three-component seismograms from diverse open-source datasets.
• Utilizing a U-Net encoder-decoder with a convolutional transformer for feature extraction in the latent space.

Main Results:

• The foundation model effectively extracts informative features from seismic waveforms.
• Incorporating latent features significantly improves downstream tasks like seismic phase picking, earthquake location, and magnitude estimation.
• Latent space analysis shows strong correlation with P- and S-wave arrival times.

Conclusions:

• The proposed foundation model offers a robust approach to enhance earthquake monitoring capabilities.
• The U-Trans architecture and self-supervised learning are effective for extracting crucial seismic features.
• This advancement supports more accurate seismic hazard assessment and improved public safety measures.