Reservoir computing (RC) is a machine learning model capable of predicting time-series data with low power consumption and fast learning. It consists of three layers: an input layer, a reservoir layer, and an output layer. The reservoir layer performs nonlinear transformations of the input signal. Physical reservoir computing (PRC) replaces this reservoir layer with a nonlinear physical phenomenon instead of software. These models are expected to be applied in edge AI devices that require real-time learning, such as autonomous driving systems. Although various materials and devices, such as optics [1] and magnetic spin systems [2], have been explored to enhance performance, an optimal device has yet to be discovered. This is because the relationship between the physical properties of the reservoir and the required characteristics of PRC, such as "nonlinearity, high dimensionality, and short-term memory", remains unclear, making it difficult to establish a clear strategy for improving performance.

In this study, we focused on ion-gated reservoirs using electric double-layer transistors (EDLTs) with solid-state Ionics [3] and developed a new device utilizing monolayer graphene as a material that can enhance nonlinearity [4]. Monolayer graphene is a semimetal with a Dirac cone-type electronic structure, exhibiting ambipolar transport characteristics for both n-type and p-type conduction. The evaluation of this EDLT revealed a characteristic V-shaped drain current-gate voltage (ID-VG) curve with hysteresis. Subsequently, the device’s performance was assessed using a benchmark task called the nonlinear autoregressive moving average (NARMA2) task. We set the VG application conditions by focusing on the influence of the V-shaped ID-VG. Condition (I) corresponds to a VG range that shows a monotonic ID-VG curve without including the Dirac point (Vmax = -0.8 V), while condition (II) corresponds to a VG range that includes the Dirac point, showing a V-shaped ID-VG curve (Vmax = 2.0 V). Compared to condition (I), a more complex ID response was observed under condition (II), suggesting that the V-shaped ID-VG complicates the real-time IDresponse.

We then conducted the NARMA2 task under these two conditions. By predicting the future values of time-series data and evaluating the reservoir performance using the normalized mean square error (NMSE), we found that the NMSE value for condition (II) (0.019) was approximately one order of magnitude lower than that for condition (I) (0.14), indicating a dramatic improvement in performance.

Furthermore, a detailed investigation of the relationship between the device's operating condition (Vmax) and the key RC characteristics (nonlinearity and short-term memory) through an information processing capacity (IPC) analysis revealed a trade-off between nonlinearity and short-term memory, leading to a peak in computational performance.

The approach adopted in this study is considered effective for elucidating the correspondence between the physical properties of EDLTs and the required characteristics of the reservoir. By further developing this approach, we aim to establish a foundation for enhancing the performance of PRC devices through optimal material selection and operational parameter control.