Interface engineering of IGZO/TiO2 heterostructure for UV driven analog memristive dynamics in in-sensor reservoir computing

Abstract

Conventional machine vision systems face critical limitations, including substantial data latency and elevated energy consumption, due to the physical separation of memory, sensing, and processing units. In contrast, in-sensor computing, particularly using optoelectronic synapses, offers a promising paradigm shift by enabling direct sensing and processing of optical signals, thereby enhancing efficiency. Light-sensitive photodetectors exhibiting history-dependent dynamic responses further advance this field by facilitating temporal vision processing and time-series prediction without complex circuitry or excessive power demands. Recent approaches aim to emulate biologically inspired vision using light-modulated memristive behaviors. However, accurate realtime analog optical signal processing remains a significant challenge, necessitating the development of a photonic analog to electrically driven memristors. In this study, we introduce an optoelectronic synaptic device based on a TiO2/IGZO heterostructure, where ultraviolet (UV) photon flux triggers memristive behavior. We analyze photoexcited oxygen-vacancy (VO-) related defect states and activation energies to elucidate the correlation between photocarrier relaxation dynamics and optical synaptic behavior. The device demonstrates high recognition accuracy 98.01% (MNIST), 88.07% (FMNIST), and 96.10% (Weizmann dataset) within a 4-bit reservoir computing framework, highlighting its potential as an energy-efficient, cost-effective solution for real-time spatiotemporal optical signal processing in next-generation machine vision systems.