Highly transparent reconfigurable non-volatile multilevel optoelectronic memory for integrated self-powered brain-inspired perception

Highlights: • Two-terminal, highly transparent (>65%), non-volatile, programmable, self-powered ultraviolet photodetector is developed. • Self-powered photoresponse was customized at various levels by fine-tuning an electric pulse. • Photodetector mimics optical-electrical-coupled versatile features of a bio-synapse such as manifold memory capability. • Photoconductive AFM revealed tunable and scalable photocurrent, providing high-density integration of ~716 GB/in². • Array was developed and integrated with the well-developed camera, demonstrating self-adaptive human-like visual perception. Photonic image sensors with programmable non-volatile manifold memory can offer an essential breakthrough for the advancement of optoelectronic memory, smart machine vision, and optical neuromorphic computing. Here, we developed a two-terminal, nickel oxide and titanium dioxide-based, highly transparent (> 65%), non-volatile, programmable, self-powered ultraviolet photodetector. The self-powered photoresponse was customized at various levels by fine-tuning an electric pulse, even without changing illumination intensity. Moreover, the photodetector mimics the optical-electrical-coupled versatile features of a bio-synapse, such as manifold memory capability, paired-pulse facilitation, and excitation or depression. The observed results are quantitatively explained by the dynamics of oxygen vacancy migration-induced junction width modulation. Furthermore, photoconductive atomic force microscopy revealed a tunable and scalable (over the desired area) photocurrent even at the nanoscale (~30 nm), providing high-density integration, with a pixel density of ~716 GB/in². Moreover, an array was developed and integrated with the well-developed camera, which was trained dynamically to memorize and classify the desired input optical patterns, demonstrating self-adaptive human-like visual perception. Our programmable photodetector represents a unique possibility to develop a trainable photoresponse that collects information for the desired shape, offering profound implications for building a complex, trainable, and energy‐efficient neuromorphic imaging system.