Modeling nonlinear oscillator networks using physics-informed hybrid reservoir computing.

Area of Science:

• Complex Systems
• Non-linear Dynamics
• Computational Science

Background:

• Surrogate modeling of non-linear oscillator networks is difficult due to the gap between analytical models and real-world complexity.
• Existing methods struggle to capture intricate dynamics accurately.

Purpose of the Study:

• To investigate hybrid reservoir computing (RC) by integrating RC with expert analytical models for improved surrogate modeling.
• To assess the performance and robustness of these hybrid models under simulated model inaccuracies and in residual physics tasks.
• To evaluate their utility for short-term forecasting and control applications in diverse dynamical regimes.

Main Methods:

• Developed hybrid reservoir computers combining standard RC with analytical models.
• Tested surrogate models with parameter errors in the expert model.
• Assessed performance in a residual physics task where the expert model lacked key non-linear coupling terms.
• Focused on short-term forecasting across various dynamical regimes.

Main Results:

• Hybrid reservoir computers generally outperformed standard RCs.
• Hybrid models demonstrated greater robustness to parameter tuning.
• Performance advantage was less pronounced in the residual physics task.
• Hybrid models showed good performance in dynamical regimes inaccessible to the expert model, highlighting the reservoir's contribution.

Conclusions:

• Hybrid reservoir computing offers a promising approach to surrogate modeling of complex non-linear systems.
• These models provide enhanced robustness and forecasting capabilities compared to standard RCs.
• The integration of analytical models with RC effectively bridges the gap between theoretical simplification and real-world complexity.