Reservoir computing with MEMS resonators

2021 Fiscal Year Annual Research Report

• Project/Area Number: 21F20799
• Research Institution: Kyoto University
• Principal Investigator: 土屋 智由 京都大学, 工学研究科, 教授 (60378792)
• Co-Investigator(Kenkyū-buntansha): MEFFAN ROBERT CLAUDE 京都大学, 工学(系)研究科(研究院), 外国人特別研究員
• Project Period (FY): 2021-04-28 – 2023-03-31
• Keywords: Reservoir Computing / Surface Acoustic Wave / Phase modulation / noise resiliency / neuromorphic computing

Outline of Annual Research Achievements

The goal of this research is to create a “single domain” piezo-electric reservoir computer. It is hoped that this approach will allow for power efficient reservoir computing algorithms to be created and applied. To achieve this goal, so far this research has focused on two areas. The design and fabrication of a non-linear Surface Acoustic Wave (SAW) resonator, and the simulation of a delay-based reservoir computer using wave-based encoding.
For surface acoustic waves to be used in a reservoir computer, it is important to understand how wave-based properties effect the implementation. In particular, the frequency and phase of the wave. To study this, a simulation program has been written which can simulate a delay-loop reservoir computer with an arbitrary non-linear processing node, and user definable encoding. This program has allowed a phase modulated reservoir computer to be studied and compared to amplitude modulated systems from literature. This investigation has revealed several interesting sources of computational power, which differ from that of the equivalent amplitude modulated system. The nature of this difference is still under investigation and will directly inform future designs for SAW resonators.
In addition, a non-linear SAW resonator has been successfully designed and fabricated. Key properties for reservoir computing applications, such as non-linearity and fading memory, have been established through the specific design of the device.

Current Status of Research Progress

1: Research has progressed more than it was originally planned.

Reason

The non-linear SAW resonator design has been successfully fabricated. Based on the simulation results, the devices measured properties appear to be suitable for reservoir computing. Presently, work is being conducted to experimentally verify this. In particular, to verify that the non-linear action of the resonator is sufficient for a reservoir computer when it is implemented into a delay loop. This requires the design and fabrication of custom electronics, as well as a custom computer program to manage the real-time processing of the resonator response.
An area of concern for the current device is the magnitude of the non-linear response. So far, the device has a small non-linear response when compared to previously reported MEMs dynamical system reservoir computers. The simulation results, as well as discussions with collaborating students and professors, indicate that the SAW resonators level of non-linearity may still facilitate processing. However, increasing the non-linearity will allow processing to be carried out at lower power levels. As such, pursuing a heightened level on non-linearity is still desirable. In parallel to the experimental verification of the current resonator design, alternative non-linear designs are also being explored. These designs are inspired by the literature of non-linear acoustic convolvers, which use beam width compression to boost the wave power density. As these devices are an established non-linear processing system, they represented a good benchmark for non-linear operation.

Strategy for Future Research Activity

Currently, the experimental apparatus for reservoir computer testbed is being designed and tested. Once this apparatus is established, the simulation results can be experimentally verified. The non-linear node may then be assessed, and the true suitability for reservoir computing confirmed. In the future, the non-linearity of the resonator is sought to be increased to reduce the power required by the system. The research and design work required for this is being carried out in parallel to the experimental verification.
The final goal of this research is to integrate a delay line onto the device structure, thereby realizing a “single domain” piezoelectric reservoir computer. While the final device structure will render the “delay” aspect of the testbed irrelevant, the design is such that it can be reused for post-processing. In addition, the testbed can be used to confirm that physical and virtual delay lines behave in functionally identical manners with the same
There may be an additional apparatus required to measure the spectral response of these devices. The requirements of this apparatus are being established presently.