Application of artificial synapse based on Al-doped SrTiO3 thin film in neuromorphic computing.

To alleviate the von Neumann architecture bottleneck, scholars have proposed neuromorphic computing consisting of artificial synapses and neural network algorithms, in which the ability to well simulate biological synaptic functions is considered a key step in building neuromorphic computing. In this work, the artificial synapse based on Al-doped SrTiO 3 (STAO) thin film is fabricated by a low-cost sol-gel method. It effectively emulated a diverse array of significant synaptic functions, including spike-time-dependent plasticity and short/long plasticity. We constructed a convolutional neural network (CNN) to perform the training/recognition task of handwritten digit images and demonstrated the applicability of the memristor device in neuromorphic computing. We compared STAO with SrTiO 3 (STO) and found that the Al-doped devices had better stability, higher linearity, and higher image recognition accuracy from 84.5% to 96.2%. We have studied the main influencing factors for these improving performances, which may be due to the generation of more oxygen vacancies by Al doping. The x-ray photoelectron spectroscopy (XPS) test showed an increase in the oxygen vacancy content of the sample from 23.59% to 37.91%, which supports our hypothesis. This has positive ramifications for building artificial neural networks capable of processing vast quantities of data. [Display omitted] • The device implements important synaptic learning and memory functions like spike-time-dependent plasticity. • CNN architecture was built and the device achieves a preferable handwritten digit recognition accuracy of 96.2%. • The introduction of Al increases oxygen vacancies from 23.59% to 37.91%, promoting device performance. [ABSTRACT FROM AUTHOR]