基于外延铪基薄膜畴壁调控忆阻器仿神经网络计算与物理机制研究

Brain-like artificial intelligence may become an important engine of the new industrial revolution. Memristor is the key device to realize artificial neural network. However, most of the memristor conductance control is nonlinear, which impacts the accuracy of pattern recognition. This project is intended to develop a linear continuous conduction modulation behavior and multi-value memory memristor by modulating two-dimensional ferroelectric domain wall of the epitaxial Hf0.5Zr0.5O2 films. In order to clarify the intrinsic physical mechanism, we wolud first to solve the difficult problem that it is hard to prepare the epitaxial Hf0.5Zr0.5O2 film successfully, because the ferroelectric material Hf0.5Zr0.5O2 reported is polycrystalline structure. We would explore the process conditions on the epitaxial Hf0.5Zr0.5O2 film growth, develop the ferroelectric domain wall prototype memory, achieve multi-level data storage and linear conduction change through tunning the generation and removal kinetics of the domain wall. Meanwhile, we study the effects of stress and interface on the ferroelectric polarization-domain wall modulation and synaptic plasticity. we also intend to simulate images of static and dynamic learning function by using random noise learning methods though building a neural network based on a memristor crossover array. The LIF spiking neuron circuit model is used to verify the feasibility of the neuron circuit with memristor. The implementation of this project can achieve high performance domain wall memristor devices and pave the scientific and technical way for the construction of large scale neural networks.

类脑人工智能可能将成为新一轮工业革命的重要引擎。忆阻器是实现人工智能神经网络计算的关键器件，但是目前多数忆阻器电导调控是非线性的，影响了图像识别的准确率。本项目拟通过施加电压控制在外延Hf0.5Zr0.5O2二维铁电畴壁上的电子导电特性，制备线性连续电导可调行为和多值存储的忆阻器。为了阐释清楚内在物理机制，首先解决铁电材料Hf0.5Zr0.5O2主要是多晶结构，不易制备外延薄膜的难题，探索该薄膜的外延生长工艺条件，研制出铁电畴壁原型存储器，并通过调控畴壁产生和去除动力学过程实现多级数据存储和线性电导变化，研究应力和界面对铁电极化-畴壁调控-突触可塑性能的影响。构建基于忆阻器交叉阵列的神经网络，通过采用随机噪声学习方法，模拟图片静态和动态学习功能。采用LIF脉冲神经元电路模型，验证神经元电路的可行性。本项目的实施能够完成高性能畴壁型忆阻器件的研发，为构建大规模神经网络奠定器件科学和技术基础。

类脑人工智能可能将成为新一轮工业革命的重要引擎。忆阻器是实现人工智能神经网络计算的关键器件，但是目前多数忆阻器电导调控是非线性的，影响了图像识别的准确率。类脑人工智能可能将成为新一轮工业革命的重要引擎。亿阻器是实现人工智能神经网络计算的关键器件，但是目前多数忆阻器电导调控是非线性的，影响了图像识到的准确率。本项目优化了更宽范围的电导线性调控，进一步推进与CMOS兼容的外延铪基铁电薄膜材料用作忆阻器原型器件在人工神经网络类脑智能技术方面的应用，并且多方位、深层次的研究了给基薄膜亿阻器的铁电畴壁的调控基础研究，阐释了微结构-畴壁调控-学习与记忆行为(类脑性能)之间的影响规律。进一步优化忆阻器性能，借助微结构分析对器件畴壁动力学分析以及氧空位在失效机制中扮演的角色进入了深入的探讨，制备的铪基铁电薄膜忆阻器成功实现了低功耗的尖峰神经网络，深入研究通过衬底应变和电压控制的铪基薄膜在神经形态计算中的阻变开关性能模拟类似于大脑的学习行为。铪基铁电薄膜忆阻器可以实现高精度的人脸识别，通过不同的脉冲调制方法在人脑的学习和遗忘行为中显示出卓越的优势。通过理论计算和硬件联想学习电路测试，实现了基于铪基铁电忆阻器的生物联想学习行为，并且采用LIF脉冲神经元电路模型，验证神经元电路的可行性。该项目有助于在高精度和强大的AI芯片中开发铪基铁电忆阻器。