Semiconductor Materials - 04

Abstract

Physical reservoir computing, in which physical phenomena exhibiting nonlinear responses are used as reservoirs, is expected to be applied to next-generation edge AI devices because of its low learning cost and highly efficient processing. Ion gating reservoirs using high-density carrier injection with solid electrolytes have been realized in various material systems[1-4]. The surface-enhanced Raman scattering enables physical reservoir computing using Raman scattering signals from a single molecule as reservoir states. Figure (a) shows a schematic diagram of the measurement system used for molecular reservoir computing. Figure (b) shows a TEM image of WO_x nanorod with a unique crystal structure having a two-dimensional conductive plane used for Raman enhancement[5]. Silver nanoparticles were deposited on WO_x nanorod for additional enhancement of Raman scattering. It has been demonstrated that Raman signals of adsorbed molecules can be used as reservoir states to perform nonlinear waveform transformation tasks and second-order nonlinear dynamical equation solving tasks with high precision. Furthermore, it has been demonstrated that molecular reservoirs enable high- accuracy prediction of blood glucose level changes [4].

We fabricated an electric double-layer transistor using MoS₂ as a channel material on a solid electrolyte substrate and realized a MoS₂-based Raman-ion-gating reservoir (MoS₂-RIGR) that uses both Raman scattering signals and current responses in the channel region as computational resources. Figure (c) shows a schematic diagram of the fabricated MoS₂-RIGR device. Multilayer MoS₂ fabricated by the exfoliation method was used as the channel material. Here we use an excitation laser with a wavelength of 632.8 nm, which provides resonant Raman scattering effects for multilayer MoS₂. The nonlinear waveform transformation task was performed using Raman scattering signal, drain current, and gate current as reservoir states. Higher accuracy was obtained in all types of waveform transformation when both Raman and current responses were used as reservoir states. Furthermore, it was found that including the Raman signal in the reservoir state reduces the normalized mean square error in solving the second-order nonlinear dynamical equation by 37%.