Artificial synaptic properties of zirconium-doped barium titanate film for neuromorphic computing

Abstract

The traditional Von Neumann Architecture relies on the separation of memory and CPU, leading to bottlenecks like slow data transfer and high energy consumption. This is especially challenging when addressing the demands of large-scale parallel computing tasks such as artificial intelligence. Devices capable of simulating synaptic properties have gotten significant attention. Neuromorphic computing combines those devices with methods that work with neural networks. They are considered a viable approach to overcoming the Von Neumann Architecture bottleneck. This paper presents a Zr-doped BaTiO3 ferroelectric thin film prepared by a sol-gel process, which mimics the synaptic properties of the nerve: Short-Term Synaptic Plasticity (STP), paired-pulse facilitated (PPD), Long-Term Synaptic Plasticity (LTP) and Spike-Timing-Dependent Plasticity (STDP). An enhanced stochastic adaptive method is used to a built-in Convolutional Neural Network (CNN) to evaluate the device's neuromorphic computing capabilities. According to the results, the device obtains recognition accuracies of 96.5 % and 75.1 % for the MNIST and Fashion-MNIST datasets, respectively. This significantly contributes to the advancement of ferroelectric materials in neuromorphic computing applications.