Reservoir computing for forecasting non-autonomous dynamics with hidden regime variations

Area of Science:

• Nonlinear dynamics
• Complex systems modeling
• Machine learning for scientific discovery

Background:

• Predicting nonlinear dynamical systems is difficult due to hidden, time-varying factors.
• Standard reservoir computing (RC) methods struggle with accuracy in non-autonomous settings.

Purpose of the Study:

• To develop an improved RC framework for enhanced prediction accuracy in non-autonomous systems.
• To enable regime identification and adaptation to unseen dynamics.

Main Methods:

• Propose a multi-regime RC framework with multiple trained readouts.
• Combine readouts using a trajectory-dependent linear readout within a short observation window.
• Evaluate the method on Duffing and Rössler systems with time-varying and chaotic forcing.

Main Results:

• The multi-regime RC framework significantly improved prediction accuracy.
• The proposed method outperformed both regime-specific and single global models.
• Demonstrated effective adaptation to unseen and intermediate dynamics.

Conclusions:

• The multi-regime RC framework offers a robust solution for predicting complex nonlinear dynamics with hidden factors.
• This approach enhances the adaptability and accuracy of reservoir computing in non-autonomous settings.
• The method shows promise for applications requiring precise modeling of evolving systems.