(Invited) Flexible Resistive Random-Access Memories Based on Carbon Nanotubes and Their Application for Reservoir Computing

Abstract

Reservoir computing (RC) is one of the machine learning methods derived from a recurrent neural network (RNN) and has attracted attention because of the small amount of computation required for learning. In recent years, various forms of physical RC using nonlinear physical systems such as memristors, optical amplifiers, and soft octopus robots have been proposed for low power consumption and high-speed processing. [1] The physical reservoir system needs non-linearity, high dimensionality, and short-term memory. It has also been shown that transient response currents due to electrochemical reactions and ion migration in solution have the ability of reservoirs [2]. In this study, we report carbon nanotube (CNT)-based resistive random-access memories (ReRAMs) and their application for reservoir computing.The network-like CNT film is used as electrodes sandwiching a thin (10 nm) HfO2 oxide film to form CNT/HfO2/CNT ReRAMs. The CNT film was synthesized using the floating-catalyst chemical vapor deposition method. Then, the CNT film was transferred onto a plastic substrate. HfO2 was deposited by the atomic layer deposition method onto CNT film. Finally, another layer of CNT film was transferred to form CNT/HfO2/CNT stack. The ReRAM exhibited clear memory behavior with SET and REST voltages around +6 V and –6 V, respectively. The ReRAM can sustain the current level of each state for more than 1 × 104 s and can be predicted to store them for ten years. The ReRAMs also exhibited stable operation under bending stress with a 0.5 cm bending radius.We used the ReRAMs as a reservoir for reservoir computing. A time-series prediction task called NARMA10, which was often used as a benchmark of reservoir computing, was performed with a normalized mean square error (NMSE) of 0.38. The chaotic motion of a double pendulum was also successfully demonstrated by using the CNT ReRAM-based reservoir.[1] G. Tanaka et al., Neural Netw. 115, 100 (2019), [2] S. Kan et al., Adv. Sci. 9, 2104076 (2022)