Model-assisted deep learning of rare extreme events from partial observations

Area of Science:

• Complex Systems
• Computational Science
• Machine Learning

Background:

• Predicting rare extreme events is difficult due to the 'small data problem' in observational datasets.
• Deep neural networks (DNNs) offer potential for extreme event prediction but require sufficient training data.

Purpose of the Study:

• To investigate a model-assisted framework for predicting rare extreme events using DNNs trained on simulated data.
• To assess the feasibility of this framework using observable quantities for practical applicability.

Main Methods:

• Utilized numerical simulations to generate adequate samples of extreme events for training DNNs.
• Applied a subset of observable quantities from dynamical systems (Rössler attractor, FitzHugh-Nagumo model, turbulent flow) for network training.
• Evaluated three DNN architectures: feedforward, long short-term memory (LSTM), and reservoir computing.

Main Results:

• Long short-term memory (LSTM) networks demonstrated the highest robustness to noise and prediction accuracy.
• LSTM networks required minimal hyperparameter tuning for effective performance.
• The study analyzed prediction accuracy, noise robustness, reproducibility, and input data sensitivity across systems and architectures.

Conclusions:

• The model-assisted framework, using simulated data and observable quantities, is feasible for predicting rare extreme events.
• LSTM networks are a promising DNN architecture for robust and accurate extreme event prediction in data-scarce scenarios.