Piezoelectric MEMS-based physical reservoir computing system without time-delayed feedback

Abstract

In this study, a physical reservoir computing system, a hardware-implemented neural network, was demonstrated using a piezoelectric MEMS resonator. The transient response of the resonator was used to incorporate short-term memory characteristics into the system, eliminating commonly used time-delayed feedback. In addition, the short-term memory characteristics were improved by introducing a delayed signal using a capacitance-resistor series circuit. A Pb(Zr,Ti)O3-based piezoelectric MEMS resonator with a resonance frequency of 193.2 Hz was employed as an actual node, and computational performance was evaluated using a virtual node method. Benchmark tests using random binary data indicated that the system exhibited short-term memory characteristics for two previous data and nonlinearity. To obtain this level of performance, the data bit period must be longer than the time constant of the transient response of the resonator. These outcomes suggest the feasibility of MEMS sensors with machine-learning capability.