Physical reservoir computing using all-solid-state LaZrO/InO electric double-layer thin-film transistor

Machine-learning-based AI prediction drives rising electricity use and processing demands, raising sustainability concerns. Reservoir computing reduces training by optimizing only output weights, and physical reservoir computing (PRC) further accelerates processing by exploiting physical phenomena. Yet, many hysteresis-based PRCs depend on costly or unstable materials and fabrication methods unsuited to scale, limiting practical deployment. This study evaluated the potential of all-solid-state electric double-layer (EDL) thin-film transistors, incorporating a lanthanum zirconium oxide solid electrolyte and an indium oxide semiconductor, for PRC applications. The drain current (ID) and gate current (IG) as functions of gate voltage (VG) displayed pronounced hysteresis, with the extent of hysteresis dependent on both the drain voltage (VD) and the VG scanning rate. The ID response to pulsed VG inputs demonstrated a clear dependence on the preceding input pulse VG, thereby validating the device's short-term memory properties. In the second-order nonlinear autoregressive moving average task, the minimum normalized mean square error (NMSE) was 7.71 × 10−3 and 7.56 × 10−3 when utilizing the IG response as virtual nodes comprising 40 and 80 points, respectively. Furthermore, the NMSE demonstrated a strong correlation with memory capacity, particularly with the coefficient of determination associated with the correlation for up to two previous data points. The all-solid-state EDL-TFT-based physical reservoir, with its miniaturization, air stability, and broad pulse operating range, offers a scalable and robust platform that could enable practical integration of PRC into next-generation AI devices, advancing energy-efficient and durable hardware solutions.