Physical reservoir computing (PRC) is a neuromorphic computing framework that exploits the intrinsic nonlinear dynamics and fading memory of physical systems to process time-series information. Because only a simple output layer is trained digitally, PRC can reduce computational cost compared with conventional deep-learning approaches, making it attractive for edge applications where device volume, power consumption, and computing resources are limited. Ion-gating reservoirs (IGRs) are a promising class of PRC devices that use ion-gating-induced material property modulation in transistor structures.[1-5] Their large nonlinear transient responses enable efficient physical feature extraction from temporal signals.

Here, we report an ion-gel/graphene electric-double-layer transistor-based IGR for compact and computationally efficient processing of sensor data at the edge.[4,5] As distributed sensing systems generate continuous physical signals containing both fast and slow components, it is increasingly important to extract useful temporal features close to the sensor while reducing data-transfer volume, latency, power consumption, and digital computational load. The proposed IGR exploits the coexistence of ultrafast electric-double-layer charging at the ion-gel/graphene interface and slower surface-related memory processes on graphene. This combination provides nonlinear transformation and multi-timescale memory within a compact transistor structure, making the device suitable for broadband temporal signal processing.

In the fabricated six-channel ion-gel/graphene device, temporal input signals are applied to a common gate, and the resulting drain-current transients from multiple graphene channels are used as reservoir states. Owing to the ambipolar transport of graphene, channel-dependent geometries, and voltage-tunable ion dynamics, the device produces diverse transient responses over an ultrawide operating range from 1 MHz to 20 Hz, with a minimum relaxation time of 99 ns. These characteristics allow the IGR to respond to high-frequency signal components while retaining slower memory effects required for history-dependent information processing.

The reservoir performance was evaluated using benchmark time-series tasks, including nonlinear autoregressive moving-average tasks and chaotic time-series prediction based on the Mackey–Glass equation. The device achieved prediction accuracy comparable to deep-learning-based approaches, while reducing the computational load by approximately two orders of magnitude because the nonlinear feature expansion is performed physically by the device and only a simple readout layer is trained digitally. These results demonstrate that iontronic device dynamics can serve as a compact physical feature extractor for sensing data. The proposed IGR provides a promising platform for sensor-integrated edge AI hardware that combines ultrawideband operation, high temporal-processing performance, and lightweight computation for real-time processing of sensor signals.