Deep learning of aftershock patterns following large earthquakes

Area of Science:

• Seismology
• Geophysics
• Machine Learning

Background:

• Aftershocks are triggered by stress changes from large earthquakes.
• Empirical laws describe aftershock magnitude and decay, but spatial forecasting remains challenging.
• Coulomb failure stress change is a common but debated criterion for aftershock distribution.

Purpose of the Study:

• To develop a deep-learning approach for forecasting aftershock locations.
• To identify a static-stress-based criterion for aftershock prediction without fault orientation assumptions.
• To improve the accuracy of aftershock forecasting compared to existing methods.

Main Methods:

• Utilized a neural network trained on over 131,000 mainshock-aftershock pairs.
• Applied the model to an independent test dataset of over 30,000 mainshock-aftershock pairs.
• Compared deep-learning predictions against classic Coulomb failure stress change.

Main Results:

• The deep-learning model achieved a higher accuracy (AUC of 0.849) than Coulomb failure stress change (AUC of 0.583).
• Learned patterns are physically interpretable, with shear stress and stress tensor invariants explaining prediction variance.
• Identified key physical quantities controlling earthquake triggering.

Conclusions:

• Deep learning offers a more accurate method for forecasting aftershock locations.
• The study reveals physically meaningful stress-based criteria for earthquake triggering.
• This approach enhances understanding of the seismic cycle and earthquake forecasting capabilities.