Development of an AI device using ion gel and graphene that dramatically streamlines machine learning computations

A joint research team from NIMS, Tokyo University of Science, and Kobe University has developed a new artificial intelligence (AI) device that exploits ion behavior to perform information processing. The team succeeded in reducing the computational load to about 1/100 of that required for conventional deep learning. The technology is expected to contribute to enhancing the information processing performance of "edge AI" operating directly on terminal equipment (an edge device). This research was published in ACS Nano on October 14, 2025.

Background

In recent years, power consumption by machine learning technologies, represented by deep learning and generative artificial intelligence (AI), has increased exponentially, creating a serious social challenge. To address this problem, demand is growing for AI devices with low power consumption and high computational performance. "Physical reservoirs"—AI devices that perform efficient brain-inspired information processing called reservoir computing—have attracted attention due to their low computational load (the required number of multiply-accumulate operations) and low power consumption, but their lower computational performance compared to software processing has been a drawback.

Key Findings

A research team from NIMS, Tokyo University of Science, and Kobe University developed a physical reservoir device utilizing ions that achieved high computational performance comparable to that of deep learning while reducing the computational load by orders of magnitude. By combining graphene, which has high electron mobility and ambipolar behavior, and an ion gel, various responses with different speeds (ions and electrons moving in various manners) develop through complex interactions, enabling the device to respond to input signals with time constants (rates of change) that vary over an extremely wide range. The device exhibited the highest-level computational performance among conventional physical reservoirs, comparable to that of deep learning performed using software, while succeeding in reducing the computational load to about 1/100 (Figure).

Future Outlook

Going forward, the research team aims to develop an ultra-low-power edge AI device capable of high-performance and efficient information processing by mounting the device obtained in this research.

Other Information

• This research was conducted by a research team led by Daiki Nishioka (Research Fellow, International Center for Young Scientists, NIMS) and Takashi Tsuchiya (Group Leader, Neuromorphic Devices Group, Research Center for Materials Nanoarchitectonics (MANA), NIMS; Visiting Professor, Tokyo University of Science). Other team members include Hina Kitano (NIMS Junior Researcher, Neuromorphic Devices Group, MANA, NIMS; a first-year doctoral student, Tsuchiya Laboratory, Joint Graduate School System, Tokyo University of Science), Wataru Namiki (Researcher, Ionic Devices Group, MANA, NIMS), Kazuya Terabe (Group Leader, Ionic Devices Group, MANA, NIMS), and Satofumi Souma (Associate Professor, Department of Electrical and Electronic Engineering, Graduate School of Engineering, Kobe University). The research was conducted as part of research project "Creation of ultrafast iontronics" (JPMJPR23H4) under JST PRESTO "Nano Materials for New Principle Devices (Research Supervisor: Yoshihiro Iwasa)."
• This research was published in the online version of ACS Nano on October 14, 2025.