In recent years, the rapid increase in energy consumption by AI systems has become a growing concern, driving heightened interest in neuromorphic computing systems—devices designed to perform efficient information processing inspired by biological neural networks. Among various approaches, physical reservoir computing (PRC), which actively leverages the complex spatiotemporal dynamics of material systems for information processing, has attracted significant attention. PRC offers a unique strategy to replace the roles of large-scale artificial neural networks composed of virtual neurons and synapses with a small number—or even a single—physical system that inherently exhibits nonlinearity, short-term memory, and high dimensionality. Although a wide variety of physical systems have been explored for PRC [1], achieving both high computational performance and practical integrability remains a major challenge. This challenge stems from the difficulty of appropriately evaluating the key characteristics of PRC in a given physical system and the resulting lack of design principles tailored to material-based reservoirs. Moreover, unlike simulation-based reservoir computing, where various hyperparameters that control the properties of the reservoir layer can be freely adjusted, the tunability of PRC remains limited, which further exacerbates the difficulty of establishing such design principles. Therefore, realizing high-performance PRC requires complementary efforts to rigorously evaluate the origins of information processing capability in developed systems and to feed these insights back into device development. Such efforts include not only the physical design of new devices but also the optimization of operating conditions for existing systems.

In this study, we report on the correlation between the fundamental characteristics and computational performance of an ion-gating reservoir (IGR), a PRC device we have developed based on ion-gating transistors [2-7]. The ion-gating reservoir provides a comprehensive framework for utilizing material property modulation induced by ion gating for information processing. To date, various types of nonlinear dynamics—such as electric double-layer effects, redox reactions, and modulation of molecular vibrational dynamics—have been demonstrated to enable PRC operations within this framework [2-7]. In particular, the ion-gating reservoir based on an ion-gel/monolayer graphene electric double-layer transistor (EDLT) achieves outstanding computational performance by leveraging the superior nonlinearity arising from the ambipolar behavior of graphene and the coexistence of complex multiple relaxation processes associated with molecular adsorption[7]. Benchmark tasks, including dynamical system prediction and chaotic time-series forecasting, have shown that this IGR achieves top-level performance compared to other PRC systems. In this presentation, we will provide a detailed analysis of the origins of the high computational performance of the IGR, focusing on the fundamental requirements for reservoirs—nonlinearity, short-term memory, and high dimensionality—and propose design principles for the development of high-performance physical reservoir computing systems.