面向仿脑计算的自旋电子学理论研究

In recently years, great progress has been achieved in the software implementation of artificial intelligence, where "deep learning" is a representative example. The hardware implementation of the brain-inspired computing, on the other hand, is just an emergent field. The magnetic nanostructures that are studied in spintronics, such as magnetic tunnel junctions, magnetic domain walls, have the required physical properties of the elements in brain-inspired computing. So they are naturally suitable to be used for the corresponding hardware implementation of artificial neural networks. The research on spintronics for brain-inspired computing will significantly change and enrich the present focuses of the spintronics community. For example, the random behaviors in magnetization dynamics due to thermal fluctuations that are unwanted in previous studies of magnetic storage devices, will play a key role and hence become very important in spintronics researches. This proposal is to study theoretically the physical implementation of the brain-inspired computing scheme based on spintronic nanostructures. We combine first-principles spin transport calculation and micromagnetic simulation to demonstrate the principles of the designed brain-inspired computing scheme without the help of other complex techniques. Specifically, we construct a unique scale-free neural network based on spintronic nanostructures. Using theoretical calculation and design, we will show how the spintronic neural network can encode different rhythms after the machine learning process. This project will result in a breakthrough on the time-resolved brain-inspired computing, which enables processing dynamical information, beyond the focus of the present researches of machine learning–the recognition and classification of static objects.

近年来以“深度学习”为代表的人工智能的软件实现方面取得了巨大的进步，而相应的“仿脑计算”硬件研发尚处于起步阶段。自旋电子学研究的磁性隧道结、磁畴壁等纳米结构具备“仿脑计算”硬件元件所需的物理特性，因此在其物理实现上存在天然的优势。面向“仿脑计算”的研究将很大程度上改变并丰富目前自旋电子学的研究内容和关注点，例如过去希望抑制的热扰动导致的磁动力学随机行为将发挥关键作用。本项目着重研究基于自旋电子纳米结构的“仿脑计算”的物理实现方案，结合第一原理自旋输运计算与微磁学模拟，在不借助其它工程学技术的条件下，从理论上验证并展示“仿脑计算”的设计方案。具体来说，我们构建一种新颖的自旋电子无标度神经网络，借助理论模拟设计自旋神经元和突触特性，使该系统能够通过机器学习，获得编码不同时长的节律信息的能力。这个问题的研究将突破目前机器学习普遍研究的空间感知功能，将“仿脑计算”拓展到时间维度，能够处理动态信息。

我们基于磁性隧道结、磁畴壁等自旋电子学器件，设计了递归神经网络，连续吸引子神经网络和脉冲前馈神经网络等自旋电子学类脑计算方案，实现了通过机器学习编码时空信息，对动态目标进行预测追踪，多模态感知信息的高效整合等功能，提出通过群体编码有效对抗器件的不均一性和系统噪音等不利因素，并结合微磁学模拟和网络动力学仿真验证了上述方案的可行性。在研究过程中，发展了一套独特的多尺度计算仿真程序，自主开发了基于Python语言的网络动力学仿真程序，并且与PyTorch等开源软件对接，结合微磁学模拟软件实现全硬件动态仿真，最终通过物理仿真环境演示，为“仿脑芯片”的实验研究提供理论基础和工具。我们在算法方面的探索也可以应用于阻变、相变、铁电等新材料和新器件的类脑计算方案。