Enhancing the Recognition Task Performance of MEMS Resonator-Based Reservoir Computing System via Nonlinearity Tuning

Area of Science:

• Neuromorphic computing
• Nonlinear dynamics
• Micro-electro-mechanical systems (MEMS)

Background:

• Reservoir computing (RC) offers low training cost and compatibility with nonlinear devices for AI realization.
• MEMS resonators, with their rich nonlinear dynamics, are promising for high-performance hardware RC.
• Previous work presented a non-delay-based RC using a single micromechanical resonator with hybrid nonlinear dynamics.

Purpose of the Study:

• To introduce a nonlinear tuning strategy for analyzing the computing properties of a micromechanical resonator-based RC.
• To investigate the influence of hybrid nonlinear dynamics on RC performance using an image classification task.
• To optimize the operating point of the RC for enhanced processing speed and recognition accuracy.

Main Methods:

• Numerical and experimental analysis of hybrid nonlinear dynamics in MEMS resonators.
• Study of transient nonlinear saturation by fitting quality factors under varying vacuum conditions.
• Identification of the optimal operating point (edge of chaos) via static bifurcation analysis and Duffing nonlinearity models.

Main Results:

• High classification accuracy of (93 ± 1)% achieved under optimal operating conditions.
• Processing speed several times faster than previous work on handwritten digits recognition.
• Performance gains attributed to high signal-to-noise ratios (quality factor) and utilized nonlinearity.

Conclusions:

• The proposed nonlinear tuning strategy effectively enhances the computing capabilities of MEMS resonator-based RC.
• Optimizing MEMS resonators at the edge of chaos significantly improves AI task performance.
• This approach demonstrates the potential of MEMS resonators for efficient and accurate hardware AI implementations.