Machine learning predicts meter-scale laboratory earthquakes

Area of Science:

• Geophysics
• Earthquake Science
• Machine Learning Applications

Background:

• Growing interest in machine learning (ML) for predicting laboratory quakes (shear-slip failures) in rock friction experiments.
• Uncertainty exists regarding ML applicability to larger-scale laboratory quakes and natural earthquakes due to vast timescale variations.

Purpose of the Study:

• To apply an advanced ML approach to meter-scale laboratory quake data.
• To assess ML's capability in predicting time-to-failure for larger-scale seismic events.

Main Methods:

• Utilized an advanced ML approach on meter-scale laboratory quake data.
• Employed a network representation of the event catalog for ML model training.
• Compared ML predictions with a dynamic model of shear failures.

Main Results:

• Accurately predicted time-to-failure for meter-scale mainshocks, from seconds to milliseconds prior.
• Demonstrated ML's ability to forecast events across timescales relevant to natural earthquakes (decades to weeks).
• Identified tracking shear stress evolution on creeping faults as key to ML prediction.

Conclusions:

• ML can effectively predict laboratory quakes by analyzing acoustic emission events.
• Findings suggest ML can indirectly track fault stress, crucial for earthquake prediction.
• Provides critical insights for short-term forecasting of natural earthquakes.