Optimizing Temporal Data Forecasting for Stiffness-Modulated MEMS Reservoir Computing

This study details an advanced method of algorithm optimization, specifically developed for a novel reservoir computing (RC) architecture to forecast temporal information. Our physical RC is implemented using micro-electro-mechanical systems (MEMSs), ingeniously employing stiffness modulation, wherein the input is manifested as stiffness disturbances within a nonlinear resonator. We have eliminated the discretization of data that was typically required in conventional delay-based RC before it is injected into the reservoir. This allows our sensor to process natural signals directly, significantly reducing system complexity and power consumption. Theoretically, this idea is capable of integrating sensing and computing in a single device. Due to simplifications made at the hardware level, its capacity to handle forecasting tasks was initially compromised. To address this, we have applied software-level optimizations, coupled with nonlinearity tuning, to restore its capability for complex forecasting. We rigorously tested the method using data from three different chaotic systems, as well as with real sensor data for dynamic temperature compensation, all of which yielded impressive results. The proposed architecture demonstrates the concept of an intelligent sensor that can handle a variety of tasks and scenes.