Study on the Zr doping effect on the performance of BaTiO3-based artificial synapses

Abstract

Memristors are considered as key components of brain-like hardware to meet the demand for energy-efficient computing in the era of big data. The realization of synaptic and neuronal functions based on memristors is a prerequisite for building artificial neural networks. In this study, we fabricated artificial synaptic devices based on Zr-doped BaTiO3 (BT-xZr) (BTO) thin films grown by the pulsed laser deposition method. Compared with the BT-0.04Zr film, the BT-0.15Zr film effectively inhibits the oxygen vacancy generation, which greatly optimizes the problem of large leakage current in the device. By applying pulse sequences to the devices, the BT-0.15Zr devices show better synaptic properties than the BT-0.04Zr devices from several aspects, such as pair-pulse facilitation, spike time-dependent plasticity, and long-term potentiation/depression. We further simulated the image recognition performance of artificial neural networks based on the two types of devices. When cycle-to-cycle variation of the devices was not considered, the recognition accuracy of BT-0.15Zr and BT-0.04Zr neural networks is 92% and 86%, respectively. With cycle-to-cycle variation included, the recognition accuracy of BT-0.15Zr and BT-0.04Zr neural networks becomes 90.5% and 31%, respectively. Our work reveals the great impact of Zr-doping on the performance of BTO-based memristors for neuromorphic computing.