Data-informed reservoir computing for efficient time-series prediction.

Area of Science:

• Computational Science
• Applied Mathematics
• Machine Learning

Background:

• Dynamical system forecasting is crucial for understanding complex phenomena.
• Traditional methods often require extensive hyper-parameter optimization or rely on unavailable knowledge-based models.
• Reservoir computing (RC) offers a powerful machine learning framework but can be computationally intensive and sensitive to parameter choices.

Purpose of the Study:

• To propose and evaluate a novel data-informed reservoir computing (DI-RC) approach for dynamical system forecasting.
• To enhance prediction accuracy and reduce computational cost compared to existing methods.
• To mitigate the need for tedious hyper-parameter optimization in reservoir computing.

Main Methods:

• Developed a hybrid approach combining a data-driven model discovery technique with a delay-based reservoir computer (RC).
• Utilized sparse identification of nonlinear dynamical systems (SINDy) for the data-driven component.
• Tested the DI-RC approach on the Lorenz and Kuramoto-Sivashinsky systems.

Main Results:

• DI-RC demonstrated improved time-series forecasting accuracy compared to individual component approaches.
• The method significantly reduced computational cost.
• Performance gains were most pronounced when reservoir parameters were not optimized, highlighting reduced hyper-parameter sensitivity.

Conclusions:

• Data-informed reservoir computing (DI-RC) offers a computationally efficient and accurate alternative for dynamical system forecasting.
• This approach is particularly valuable when knowledge-based models are unavailable.
• DI-RC successfully integrates data-driven model discovery with machine learning for robust predictions.