用于构建人工神经网络的纯电子型忆阻器突触阵列研究

基于人工神经网络的神经计算机，工作原理与人脑类似，除了速度快，更重要的是为人工智能的发展提供了可能的硬件条件。人工神经网络中，突触起着至关重要的作用。如果采用忆阻器作为人工突触，那么神经网络无论在超高集成度还是在超低能耗方面，都有望和人脑媲美。目前，找到性能稳定可靠的突触器件，构建高密度突触阵列，是该领域研究的重中之重。本项目旨在找到拥有自主知识产权、性能稳定、电学行为高度非线性的纯电子型忆阻器突触，其忆阻效应来源于电子在介质层或界面缺陷处的俘获/释放，不涉及到器件微结构变化，因此，与离子迁移型器件相比，有望表现出更加稳定、可靠的工作性能。在此基础上，构建高密度、低能耗、无串扰的交叉矩阵结构突触阵列。最后，搭建单层感知器神经网络，通过简单图像识别，测试突触阵列的有效性、稳定性和可靠性。研究单个突触器件电阻范围和非线性度等参数、电极尺寸和间距等因素对阵列集成度、工作稳定性和精确度的影响规律。

The artificial neural network-based neurocomputer, working like a human brain, provides the possibility of hardware implementation of artificial intelligence as well as a high speed operation capability. Synaptic devices are the most important components in artificial neural networks. By using memristors as artificial synapses, artificial neural networks will possess ultra-high density and ultra-low energy consumption comparable to that of the human brain. Therefore, it is extremely important to find out synaptic devices with stable and robust performance followed by constructing high density synaptic arrays. This project aims to find out purely electronic memristive synaptic devices with independent intellectual property rights, stable performance and high degree of nonlinearity, whose memristive behavior comes from an electron trapping/detrapping of the trap sites in films and interfaces. Compared to memristive synaptic devices based on ion migration, such purely electronic ones are expected to exhibit more stable and robust performance since no microstructural change occurs. Next, crosstalk-free crossbar structured synaptic arrays with high density and low energy consumption will be constructed. Finally, a simple artificial neural network, i.e. single-layer perceptron, will be built to test the performance of as-constructed synaptic arrays including effectiveness, stability and reliability by simple pattern recognition. In addition, the effects of the resistance and degree of nonlinearity of single synaptic devices and electrode size and space on the density, stability and degree of accuracy of arrays will be investigated.