Programmable memtransistor array with temporal dynamics modulation for efficient time-series data processing.

Abstract

Hardware-based reservoir computing systems promise energy-efficient time-series data processing. However, their non-adjustable temporal dynamics at the hardware level limit multiscale feature extraction and computing capacity. Here, we present a programmable dynamic memtransistor featuring a dual-functional gate stack that enables hardware-level control of dynamic behavior without constant biasing or input preprocessing. The memtransistor integrates a charge storage layer for volatile temporal processing and a charge trap layer for non-volatile modulation of the band structure, enabling a 5-fold tunability in relaxation time constants. By configuring memtransistors with distinct dynamics in parallel, we realize a wide reservoir computing system capable of processing signals across multiple timescales. This architecture delivers a 40-fold reduction in the error for the multiple superimposed oscillator prediction task compared to a single reservoir baseline, and achieves software-comparable accuracy in forecasting the chaotic Lorenz attractor. Our results establish a compact and energy-efficient hardware platform for scalable wide reservoir computing implementation.