Next-generation reservoir computing for dynamical inference.

Area of Science:

• Computational science
• Nonlinear dynamics
• Machine learning

Background:

• Dynamical systems modeling is crucial for scientific understanding.
• Existing reservoir computing methods have limitations in flexibility and scalability.
• Time-series data analysis requires robust modeling techniques.

Purpose of the Study:

• To introduce a simple, scalable implementation of next-generation reservoir computing (NGRC).
• To model dynamical systems from time-series data using a novel nonlinear projection method.
• To demonstrate the framework's effectiveness on benchmark tasks and its suitability for real-world applications.

Main Methods:

• Utilized a pseudorandom nonlinear projection of time-delay embedded inputs.
• Implemented NGRC with feature-space dimension independent of observation size.
• Applied the method to attractor reconstruction and bifurcation diagram estimation using noisy, partial measurements.

Main Results:

• The NGRC models demonstrated stability over long prediction rollouts.
• Models generalized effectively beyond training data, showing robust performance.
• Small amounts of training noise acted as a regularizer, enhancing autonomous stability.
• Achieved accurate results on benchmark tasks like attractor reconstruction and bifurcation analysis.

Conclusions:

• The proposed NGRC framework offers a flexible and scalable alternative to polynomial-based methods.
• The method provides explicit control over system state during prediction.
• NGRC is well-suited for surrogate modeling and digital-twin applications due to its stability and generalization capabilities.