Collaborative Research: Energy Efficient Voltage Controlled Non-volatile Domain Wall Devices for Neural Networks

随着深度神经网络（dnn）越来越多地部署在低功耗嵌入式设备和物联网（IoT）应用中。他们需要能够实时学习，同时也要节能。这就需要使用比传统二进制“0”和“1”状态更多的多态存储器，它是非易失性的，这样当电源关闭时信息就会被保留，并且可以用很少的能量进行编程。这个项目的目标是研究和演示神经网络的突触元素，它可以存储在学习过程中使用电压控制磁畴壁（DW）设备更新的权重。信息被编码为DW在窄磁线中的位置。具体来说，该研究将侧重于利用对薄压电层施加小电压产生的应变，并将其转移到沉积在其上的磁丝上，以极其节能的方式控制DW的位置。这项研究可能会导致实现深度神经网络的密集，节能和强大的硬件范例。两名研究生，一名在弗吉尼亚联邦大学（VCU），另一名在麻省理工学院（MIT），将获得先进纳米制造、纳米表征和建模方面的多学科技能。VCU-PI和MIT- Co-PI将在他们所教授的课程中整合内存和计算领域墙技术。PI和Co-PI计划在他们的实验室招收从各自大学中代表性不足群体的外展项目中招募的研究实习生。学生将接受纳米磁体的纳米制造和其他磁性技术方面的培训。PI和Co-PI还计划合作为各自大学的高中生和教师举办纳米磁学讲习班。VCU和MIT之间的合作将研究和演示由磁致伸缩金属（如CoFe）组成的赛道的使用，其中DWs使用相邻Pt层的自旋轨道扭矩（SOT）移动，并使用压电层产生的电压产生的应变来确定地捕获，该压电层可以调制赛道不同区域的垂直磁各向异性（PMA）。研究小组进一步计划探索磁致伸缩稀土铁石榴石（REIG）的使用，该材料具有较低的饱和磁化强度和低阻尼，由于DW速度大，可以在较短的时间内实现较低的SOT，从而提高DW器件的能效。建议的工作将包括补充材料生长，表征，纳米制造，先进的磁性可视化，建模和仿真，包括：(i)金属铁磁体和绝缘铁磁体的生长（ii）研究磁致伸缩赛道中由SOT驱动的DW速度，以及用电压诱导应变阻止SOT驱动的DW运动的概念验证演示（iii）用SOT对畴壁运动进行微磁建模，并在存在缺口的情况下用电压诱导应变控制它。边缘效应和室温热噪声，并评估所提出的器件在实现深度神经网络中的整体性能优势。本项目的研究将推进局部电压诱导各向异性变化下的DW动力学知识，在表现出丰富的SOT物理和手性DW存在的异质结构中。它还将提供突触和神经元设备的概念验证演示，为dnn的节能硬件实现铺平道路。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

As Deep Neural Networks (DNNs) are increasingly deployed in low power embedded device and Internet of Things (IoT) applications. They need to be able to learn in real time while also being energy efficient. This necessitates the use of multi-state memory which is more than the conventional binary “0” and “1” states, is non-volatile such that information is retained when power is turned off, and can be programmed with very little energy. The goal of this project is to study and demonstrate synaptic elements of a neural network, which can store the weights updated during learning using voltage-controlled magnetic domain wall (DW) devices. Information is encoded as the position of a DW in a narrow magnetic wire. Specifically, the research will focus on using the strain generated by application of a small voltage to a thin piezoelectric layer and transferred to a magnetic wire deposited on it to control DW position in an extremely energy efficient manner. This research could lead to a dense, energy efficient and robust hardware paradigm for implementing DNNs. Two graduate students, one at Virginia Commonwealth University (VCU) and one at Massachusetts Institute of Technology (MIT), will gain multidisciplinary skills in advanced nanofabrication, nano-characterization and modeling. The VCU-PI and MIT- Co-PI will incorporate domain wall technology for memory and computing in the courses they teach. The PI and Co-PI plan to host research interns in their labs recruited from outreach programs for underrepresented groups in their respective universities. The students will be trained on nanofabrication of nanomagnets and other aspects of magnetic technology. The PI and Co-PI also plans to hold nanomagnetism workshops for high school students and teachers in their Universities collaboratively. This collaborative effort between VCU and MIT work will study and demonstrate the use of racetracks comprised of magnetostrictive metals such as CoFe, where DWs are moved using Spin Orbit Torque (SOT) from an adjoining Pt layer and arrested deterministically using voltage generated strain from a piezoelectric layer underneath that modulate perpendicular magnetic anisotropy (PMA) in different regions of a racetrack. The research team further plan to explore the use of magnetostrictive Rare Earth Iron Garnets (REIG) that have lower saturation magnetization and low damping, allowing for lower SOT applied for lesser time due to large DW velocities in order to improve the energy efficiency of DW devices. The proposed work will consist of complementary materials growth, characterization, nanofabrication, advanced magnetic visualization, modeling and simulation that includes: (i) Growth of metallic ferromagnetic and insulating ferrimagnets (ii) Study of SOT-driven DW velocity in magnetostrictive racetracks and proof-of-concept demonstration of arresting SOT-driven DW motion with a voltage induced strain (iii) Performing micromagnetic modeling of domain wall motion with SOT and its control with voltage-induced strain in the presence of notches, edge effects and room temperature thermal noise and evaluating the overall performance benefit of the proposed device in implementing DNNs. The research in this project will advance the knowledge of DW dynamics under local voltage- induced variations in anisotropy, in heterostructures that exhibit rich physics of SOT and the presence of chiral DWs. It will also provide a proof-of-concept demonstration of synaptic and neuron devices that could pave the way for energy-efficient hardware implementation of DNNs.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.