Using machine learning to predict extreme events in complex systems

Area of Science:

• Complex Systems Science
• Turbulence and Dynamical Systems
• Machine Learning Applications

Background:

• Extreme events and anomalous statistics are common in natural systems, posing prediction challenges.
• Deep learning shows promise for analyzing complex dynamical systems beyond image processing.
• Predicting rare events in turbulent systems requires advanced modeling techniques.

Purpose of the Study:

• To investigate deep learning strategies for predicting extreme events in turbulent dynamical systems.
• To develop and apply a novel neural network model for capturing extreme events.
• To assess the model's performance across various statistical regimes, including skewed distributions.

Main Methods:

• Utilized a densely connected mixed-scale network for extreme event prediction.
• Trained the neural network on data from the truncated Korteweg-de Vries (tKdV) model, focusing on near-Gaussian regimes.
• Employed a relative entropy loss function and empirical partition functions to measure prediction accuracy.

Main Results:

• The neural network successfully captured extreme events within the tKdV statistical framework.
• The model demonstrated high skill in predicting solutions across diverse statistical regimes, including highly skewed extreme events.
• Training data was sourced from near-Gaussian regimes, yet the network generalized to predict extreme values.

Conclusions:

• Deep learning, specifically the proposed network architecture, is highly effective for predicting extreme events in complex turbulent systems.
• The method shows robustness even when trained on data lacking extreme values.
• This approach holds significant potential for application to other high-dimensional and complex systems.