Highly adaptive and energy efficient neuromorphic computation enabled by deep-spike heterostructure photonic neuro-transistors

Abstract

Recently, neuromorphic photonics using optical signal as a data domain are considered as a promising solution to realize the next generation neural network platform. Here, metal-chalcogenide/metal oxide semiconductor based photonic neuro-transistors with deep spike-like heterostructure are proposed as a highly adaptive and energy efficient neuromorphic device. In particular, the energy band structure of cadmium sulfide (CdS)/amorphous indium-gallium-zinc-oxide (a-IGZO) heterojunction is engineered via mediating the anion-to-cation ratio of CdS films. It is revealed that the S/Cd ratio is able to determine the work function of the film which consequently causes a variation in the degree of band-bending at the heterointerface. Using a CdS film with optimized S/Cd ratio (CdS1.2), deep spike-like heterostructure (DHS) can be constructed which enables efficient accumulation of photo-generated charge carriers and the emulation of biological synaptic functions including long-term potentiation (LTP) and depression (LTD) behaviors. Also, the a-IGZO/CdS1.2 DHS transistor exhibits low non-linearity value for LTP (1.1) and less energy consumption (45.04 pJ). Furthermore, 7 × 7 opteoelectronic neuromorphic arrays are successfully implemented to exhibit possibility of realization of hardware-based weight pixel training. In addition, the a-IGZO/CdS1.2 DHS transistor shows a high accuracy for image pattern recognition (85.96%) based on the artificial neural network simulation, proving the feasibility in the artificial intelligent systems. © 2022 Elsevier Ltd