HZO/HSO Superlattice ReFET Array Integrating Optical Sensing for Neuromorphic Vision Computing.

Dang, Bingjie; Sun, Kaixuan; Su, Hanxin; Chen, Jiaxin; Yao, Pengyu; Zeng, Tao; Shi, Shu; Zhou, Guowei et al. · Adv Mater · 2025

basic_science · Level V

Where this comes from

Abstract

Neuromorphic vision systems require artificial synapses that integrate sensing, memory, and computation with high precision and stability. Conventional memristors face limitations including forming requirements, few multilevel states, low endurance, and poor integration density, while ferroelectric and flash-based transistors suffer trade-offs among endurance, retention, and switching ratio, and generally lack intrinsic photonic sensitivity, constraining in-sensor computing. Here, a photonic resistive-gate field-effect transistor (ReFET) array is presented that combines optical sensing, forming-free multilevel memory, and analog computation in a single device. The ReFET employs a Hf<sub>0</sub>.<sub>5</sub>Zr<sub>0</sub>.<sub>5</sub>O<sub>2</sub>/Hf<sub>0</sub>.<sub>95</sub>Sr<sub>0</sub>.<sub>05</sub>O<sub>2</sub> (HZO/HSO) superlattice gate and an amorphous InGaZnO (IGZO) channel, achieving 272 stable conductance states (>8-bit) in a 20 × 20 array, ON/OFF ratios >10⁶, endurance >10<sup>10</sup> cycles, and retention >10⁶ s. The array functions as an in-sensor optical convolutional layer, performing multiply-accumulate (MAC) operations with 94.45% accuracy on Fashion-MNIST using 8-bit quantized weights, while delivering high energy efficiency. This platform enables scalable, high-precision, energy-efficient photonic neuromorphic computing, integrating sensing, memory, and computation in one architecture.