Prediction in a driven-dissipative system displaying a continuous phase transition using machine learning

Area of Science:

• Complex Systems Science
• Geophysics
• Machine Learning

Background:

• Complex systems near criticality are notoriously difficult to predict.
• Earthquake size distributions often follow a power-law (Gutenberg-Richter) relationship.
• The predictability of earthquakes is a key question in seismology.

Purpose of the Study:

• To investigate the predictability of event sizes in a critical system.
• To determine if earthquakes, following Gutenberg-Richter scaling, are inherently unpredictable.
• To analyze how proximity to criticality affects prediction accuracy.

Main Methods:

• Utilized the Olami-Feder-Christensen model to simulate earthquake events.
• Employed a convolutional neural network (CNN) for event size prediction.
• Examined model behavior at varying degrees of proximity to criticality.

Main Results:

• Event sizes in the model follow a power-law with a cutoff for large events.
• Predictability of event sizes decreases significantly as the system approaches criticality.
• Accurate prediction was only achieved for large, nonscaling events.

Conclusions:

• Earthquake faults exhibiting Gutenberg-Richter scaling are inherently difficult to forecast.
• The findings suggest limitations in predicting typical earthquake events due to system criticality.
• Machine learning can identify predictability in specific event types within complex systems.