Semiconductor Materials - 11

Abstract

To realize edge computing playing a role of artificial intelligence on terminal device, low power consumption and real-time information processing are required. Reservoir computing using physical device attracts attentions as promising from the perspectives of low learning cost high-speed processing, and physical implementation. Spintronic reservoirs possess particularly the potential of excellent miniaturization and high computational performance [1,2], but significant issues remains: increased power consumption and increased device complexity due to necessity of external magnetic fields. To address such issues, we developed an ion-gated reservoir (IGR) utilizing a transient response of ion in electrolyte and electron carrier in channel in transistor structure [3-5] and newly fabricated an IGR utilizing in situ manipulation of magnetization vector. Thus, this IGR does not need external magnetic fields. [6,7] In this study, by utilizing the complex transient response of the magnetization vector-controlled IGR [8], we performed a prediction task of blood glucose level with slow dynamics (300 seconds/step) where an operating speed of the IGR could not be synchronized so far.[7]

The IGR in this study is an all-solid-state oxidation-reduction device (Figure (a)) using room-temperature ferrimagnetic Fe₃O₄ thin film and Zr-Li₄SiO₄ (LSZO) thin film as the Li electrolyte. The magnetization vector can be manipulated through electronic structure in the thin film by applying a gate voltage (V_G) (Figure (b)) [6]. An in-plane Hall voltage V_XY, expressed as the nonlinear function sin2φ using the magnetization direction φ, reflects φ information in response to pulsed V_G input (Fig. (c)). The relaxation time of the IGR is relatively slow with a distribution of 30–70 seconds depending on the past input. Such complex and gradual state transition enabled the reservoir to predict precisely the blood glucose level in patient with type1 diabetes (Figure (d)). Despite synchronizing the blood glucose sampling time with the input time, the error for a prediction 3 steps ahead (i.e., 15 minutes ahead) was 29.88 mg/dl, achieving the high prediction accuracy among physical reservoirs [7].