Efficient Reservoir Computing by Nonlinearly Coupled Piezoelectric MEMS Resonators

In this work, we propose reservoir computing (RC), a framework derived from recurrent neural network theory and a promising technology for edge AI computing, using piezoelectric MEMS resonators. Since the output voltage characteristics of the resonator depend heavily on the load resistance, the output of one resonator was inputted to the gate of a MOSFET to modulate the channel resistance, which was then connected to another resonator. With this nonlinear node-to-node connection and virtual nodes based on the transient response, the proposed piezoelectric MEMS RC system can perform computational tasks related to time-series data processing without using delayed feedback.