Reservoir Computing with Graphene-Based Lithium-Ion Operated Electric Double Layer Transistors

Abstract

Listen

Translate article

In recent years, the rapid development of AI technology has faced serious problems such as increasing power consumption and computational costs, while the demand for AI technology has increased. One of the possible solutions to these problems is physical reservoir computing (PRC), a form of physical implementation of machine learning that is suitable for efficiently processing time series data. It utilizes nonlinear dynamics of physical phenomena in materials or devices for information processing. There is a potential for low power consumption and miniaturization. Therefore, PRC is promising for applications such as edge AI, which requires in-situ learning without transmitting to the cloud, in environments with limited computational resources. In this context, PRC has been studied using various physical phenomena such as optical response and magnetic spin, but its performance needs to be improved. Previously, our team developed ion-gating reservoirs (IGRs) using solid-state ionic materials, such as electric double layer (EDL) and redox-based transistors (PRC devices).[1,2] In particular, a hydrogen-terminated diamond-based IGR operating in an EDL mechanism has achieved high PRC performance by the nonlinear drain (I D) current response, although the performance needs to be further improved for practical use.[1] Since the nonlinear response of IGR originates largely from the I D -V G (gate voltage) characteristic of the transistor, choosing the channel semiconductor with unique electronic transport properties can give IGRs with different characteristics. In this viewpoint, we considered using semiconductors with ambipolar electron transport (n-type and p-type) in the transistor channel to improve the performance. Therefore, in the present study, EDL ambipolar transistors were fabricated using monolayer graphene as the channel semiconductor, which exhibits ambipolar conduction due to its electronic structure (Dirac cone). We investigated the fundamental electrical characteristics of the transistor and the PRC performance. This study used monolayer graphene deposited by chemical vapor deposition as the channel semiconductor with Au/Cr source and drain electrodes. Then, an electric double layer transistor (EDLT) was fabricated by depositing a Li+- conducting solid electrolyte (amorphous Li-Nb-O) and a gate electrode (LiCoO2) by pulsed laser deposition, as shown in Fig.(a). Applying various V G to the gate electrode causes Li+ transport inside the electrolyte. The formation of EDL at the graphene channel/electrolyte interface changes the electron carrier density in graphene. The I D through the channel can thus be controlled. We evaluated the electrical characteristics of this transistor and its performance as a PRC device through a benchmark task. The I D-V G characteristics measured by sweeping the V G from -1.5 V to +1.5 V are shown in Fig. (b). When a negative V G is increased (e.g., V G < -0.2 V), Li+ moves toward the gate electrode, and holes are injected into the channel surface, forming an electric double layer (p-type conduction). On the other hand, increasing a positive V G (e.g., V G>-0.2 V) injects electrons into the channel surface (n-type conduction). Therefore, V-shaped ambipolar transistor characteristics were clearly observed. The performance of this EDLT was evaluated by the 2nd-order nonlinear autoregressive moving average (NARMA2) task, which is a time-series data prediction task requiring nonlinearity and short-term memory and is thus widely used as a typical benchmark task of PRC. The performance is evaluated by the value of the normalized mean square error (NMSE) with the nonlinear transform output of the PRC and the target (i.e., a low NMSE means high performance). The predicted waveform by the EDLT reproduced the characteristics of the target waveform well, indicating that it predicted the time-series data with high accuracy. The NMSE (the test phase) was 0.015, which was notably lowered by 40% compared to the high-performance diamond-EDLT (NMSE: 0.020 in the test phase) [1] The high performance is attributed to the unique I D-V G and transient characteristic of the ambipolar EDLT. This work was supported by JST PRESTO (grant number, JPMJPR23H4).[1] D. Nishioka, et al., Sci. Adv. 8, eade1156(2022) [2] T. Wada et al., Adv. Intell. Syst. 5, 2300123(2023). Figure 1