Spatially encoded broadband all-optical thermoelectric synapses for in-sensor computing

Abstract

In-sensor computing integrates the separated sensing, memory, and computing functions of conventional artificial intelligence (AI) systems, enabling low-power and efficient intelligent systems. In this study, we report a self-powered, broadband (from 405 to 1550 nm), spatially encoded all-optical artificial synapse based on screen-printed thermoelectric Ag₂Se for in-sensor computing. The excitatory and inhibitory behaviors of biological nervous systems can be emulated by a single Ag₂Se device through the generation of continuously tunable spatially dependent positive and negative photo-thermoelectric responses. Under consecutive optical pulses, the device can exhibit typical short-term potentiation (STP) and inhibitory facilitation, demonstrating dynamic synaptic plasticity, with synaptic weight updates regulating the enhancement and attenuation of neural signal transmission. Furthermore, the device enables logic operations including NOT and XOR gates. Using this artificial synapse, we further constructed a convolutional neural network (CNN) model for a multitask computing system, achieving smile detection, gender recognition, and image recognition, with accuracies of 92.2%, 98.6%, and 94.3%, respectively. This facile thermoelectric hardware platform enables integrated broadband photo-thermoelectric sensing and neuromorphic computing, offering a cost-effective solution for artificial intelligence systems.

Publication
Journal of Materials Science & Technology