Learning Deep Earthquake Sources

TL;DR

• Symmetric Variational Autoencoders (SymVAE/SymAE) disentangle coherent source information from path effects in grouped waveforms.
• The model enables extraction of high‑resolution source time functions (STFs) and reveals directivity patterns that envelope averaging obscures.

Overview

This work presents methods and results demonstrating how SymVAE separates coherent earthquake source information from nuisance path effects, enabling conditional generation of virtual waveforms and improved directivity analysis.

Method (brief)

• Encoder/decoder architecture: coherent encoder infers source information shared across grouped seismograms; nuisance encoder infers path-specific effects.
• Training objective maximizes a variational lower bound and uses latent-space optimization to produce virtual waveforms that preserve coherent source information.

Key results (highlights)

• Most deep-focus earthquakes (Mw > 6.0) release seismic moment in fragmented bursts rather than continuous rupture. Individual rupture episodes have short durations of 5–10 seconds. Low-magnitude events (Mw 6.0–7.0) comprise 2–3 episodes; high-magnitude events (Mw > 7.0) consist of 4–10 episodes.
• Fragmented moment release supports the multi-mechanism hypothesis for deep earthquakes. Results suggest initiation by one mechanism (e.g., metastable olivine transformation) followed by propagation via another (e.g., thermal runaway or dehydration embrittlement). Consistent with cascading failure models of shear thermal instabilities.
• Minimal directivity in lower-magnitude earthquakes indicates closely-spaced source locations. Complex high-magnitude earthquakes show clear directivity patterns.