研究论文 | 基于随机自旋轨道矩器件构建人工脉冲神经网络硬件

《中国科学：物理学 力学 天文学》英文版(SCIENCE CHINA Physics, Mechanics & Astronomy, SCPMA)出版南京大学刘荣华教授团队成果，文章题为“Stochastic spin-orbit-torque device as the STDP synapse for spiking neural networks”，于2023年第66卷第5期刊出。

目前，人工智能技术已渗透到人们生活生产的各个方面，然而随着其高速发展，传统冯·诺依曼架构计算机已经难以满足其庞大的计算量要求。近期，受人脑学习方式的启发，基于各类物理器件构建的神经形态计算硬件在人工智能领域受到了广泛关注。在众多的人工神经网络模型中， 脉冲神经网络展现出低能耗和高并行率的特性。拥有短时记忆效应、亚纳秒尺度非线性磁动力学和随机翻转行为的纳米自旋电子器件，可以用来构建循环神经网络和脉冲神经网络硬件。相比于传统的自旋转移矩，近期发现的自旋轨道矩可以使自旋电子器件实现更长的使用寿命和更快的写入速度，因此，自旋轨道矩器件受到学术界和工业界青睐。

脉冲神经网络中的神经元之间是基于众多突触传递离散脉冲序列进行通信的，突触权重可以根据与突触连接的前后神经元的激发状态进行调整，即突触具有可塑性。本文介绍了在一类低功耗自旋轨道矩器件中，通过研究电流诱导的自旋轨道矩效应和焦耳热效应，发现器件存储层磁矩的随机翻转概率（对应于电导率）随输入脉冲电流及其时间间隔成指数衰减关系，即时间间隔越长，随机翻转概率越低。为了演示该类自旋轨道矩器件能模拟人脑神经突触可塑性，本文进一步基于该自旋器件构建了两种人工脉冲神经网络，分别进行了手写数字识别的无监督学习（准确率高达80%）和逻辑运算的学习。该工作为如何运用新型低功耗自旋电子器件实现神经形态计算硬件提供了新的思路和具体方案。

创新要点：

本文利用自旋轨道矩器件中电流驱动存储层磁矩随机翻转和其温度特性实现了人工突触。由于存储层磁矩的随机翻转概率随输入脉冲电流及其时间间隔变化，其构建的人工突触的可塑性可以由存储层磁矩方向对应的霍尔电阻变化实现。基于该类自旋人工突触，研究人员构建了两种脉冲神经网络硬件原型，实现了非监督的手写数字识别和逻辑运算学习。

原文信息：

H. Li, L. Li, K. Zhou, C. Yan, Z. Gao, Z. Li, and R. Liu, Stochastic spin-orbit-torque device as the STDP synapse for spiking neural networks, Sci. China-Phys. Mech. Astron. 66, 257512 (2023), https://doi.org/10.1007/s11433-022-2081-5

Magnonic devices based on spin waves are considered as a new generation of energy-efficient and high-speed devices for storage and processing of information. Our new work entitled “Mode Structures and Damping of Quantized Spin Waves in Ferromagnetic Nanowires” is published in Chinese Physical Letter [Qingwei Fu¹, Yong Li², Lina Chen¹, Fusheng Ma², Haotian Li¹, Yongbing Xu^3,4, Bo Liu⁵, Ronghua Liu^1*, Chin. Phys. Lett. 37, 087503(2020)]. Congratulations to Mr. Qingwei Fu. In this work, we experimentally demonstrate that three distinct dominated magneto-dynamic modes are excited simultaneously and coexist in a transversely magnetized ferromagnetc wire by combining our home-made CPW broadband FMR spectroscopy and electron beam lithography technique. Besides the uniform FMR mode, the spin-wave well mode, the backward volume magnetostatic spin-wave mode, and the perpendicular standing spin-wave mode are experimentally observed and further confirmed with more detailed spatial profiles by micromagnetic simulation. Furthermore, our experimental approach can also access and reveal damping coefficients of these spin-wave modes. Our results make further insight into the complexity of the dynamical magnetization states in microscale confined FWs, which offers valuable information regarding the excitation of spin waves for the development of magnonic devices. This work is collaborated with Prof. Fusheng Ma’s group, School of Physics and Technology, Nanjing Normal University; Prof. Yongbing Xu in School of Electronic Science and Engineering, Nanjing University, and Dr. Bo Liu in Key Laboratory of Spintronics Materials, Devices and Systems of Zhejiang Province. We especially want to thank Prof. Fusheng Ma and Mr. Yong Li for providing their support in micromagnetic simulation.

In collaboration with Prof. Yongbing Xu’s group in School of Electronic Science and Engineering, Nanjing University, and Dr. Bo Liu, and Dr. Hao Meng in Key Laboratory of Spintronics Materials, Devices and Systems of Zhejiang Province, we recently published a paper entitled “Strong interface-induced spin-charge conversion in YIG/Cr heterostructures” in APL [Lijun Ni, Zhendong Chen, Xianyang Lu*, Yu Yan, Lichuan Jin, Jian Zhou, Wencheng Yue, Zhe Zhang, Longlong Zhang, Wenqiang Wang, Yong-Lei Wang, Xuezhong Ruan, Wenqing Liu, Liang He, Rong Zhang, Huaiwu Zhang, Bo Liu, Ronghua Liu*, Hao Meng*, and Yongbing Xu*, Appl. Phys. Lett. 117, 112402 (2020)]. Congratulations to Ms. Lijun Ni. In this work, we have investigated the spin pumping effect of Y3Fe5O12 (YIG)/Cu (t_Cu nm)/Cr heterostructures with the thickness of the Cu interlayer varying from 0.4nm to 5.0 nm by using our home-made CPW broadband FMR spectroscopy. A huge charge signal I_c = 0.239 uA is observed in a YIG/Cr bilayer with direct contact, whereas I_c drops dramatically by two orders of magnitude when thin Cu interlayers down to 0.4 nm are inserted between YIG and Cr. Meanwhile, the injected spin current J_s stays almost invariant for all the heterostructures. The effective spin Hall angle of the YIG/Cr interface is found to be three orders of magnitude larger than the spin Hall angle of the bulk Cr layer in YIG/Cu/Cr. The huge spin-charge conversion efficiency at the YIG/Cr interface is attributed to the inverse Rashba–Edelstein effect. Our experimental results demonstrate the dominant role of the interfacial effect in the spin-charge conversion process of the YIG/Cr heterostructures, suggesting that tailoring the interface of magnetic heterostructures provides an alternative strategy to achieve the higher spin-charge conversion efficiency for the development of low-power consumption spintronic devices. This paper was also selected as an Editor’s Pick.

Our new work entitled “Electrical generation and detection of spin waves in polycrystalline YIG/Pt grown on silicon wafers” is published in Materials Research Express [Rongxin Xiang, Lina Chen^*, Sheng Zhang, Haotian Li, J. Du, Y. W. Du, and R.H. Liu*, Materials Research Express 7, 046105 (2020)]. Congratulations to Mr. Xiang. In this work, we studied the magnetic properties of polycrystalline Y₃Fe₅O₁₂ (YIG) thin films (less than 100 nm) deposited on thermally oxidized silicon wafer by our home-made magnetron sputtering (<10^-9 Torr, DIY UHV sputtering guns) and followed by the post-annealing process. We found that sputtering at room temperature combined with the post-annealing treatment can be an efficient method to achieve large-area (inch scale) and highly uniform YIG thin films with a low damping constant α ~ 7 ×10^-3 on cheap oxidized Si wafer. The dynamical properties of the YIG thin films are characterized by our home-made CPW broadband FMR spectroscopy (10 MHz – 40 GHz, 4 K – 350 K, ± 2 T). Furthermore, the spin pumping experiments demonstrate that the polycrystalline YIG/Pt system has a good spin mixing conductance, where spin current can be effectively injected into the adjacent Pt layer from YIG through the interface. Then the electrical detection of magnetic properties (e.g., spin waves) of insulating YIG film can be achieved via the inverse spin Hall effect of Pt. The electrical detection of spin waves in the large-area polycrystalline YIG/Pt on silicon wafer may help to develop new spintronic devices (e.g., magnon-based devices) by utilizing the complementary metal-oxide-semiconductor (CMOS) technology. We thank Prof. Jun Du in our Department of Physics for his support in measurements and analysis of magnetic susceptibility.

Our new work of spin Hall nano-oscillator entitled “Dynamical mode coexistence and chaos in the nano-gap spin Hall nano-oscillator” is recently published in Phy. Rev. B [Lina Chen, Kaiyuan Zhou, S. Urazhdin, Wencong Jiang, Y. W. Du, and R. H. Liu*, Phys. Rev. B 100, 104436 (2019)]. Congratulations to Dr. Chen. In this paper, we utilize microwave spectroscopy to study the auto-oscillation modes in the nanogap spin Hall nano-oscillators based on Permalloy/Pt bilayers. We show that two distinct spin-wave modes appear in such oscillators, regardless of the magnetic film thicknesses or the size of the electrode gap. We identify the primary mode as a nonlinear self-localized bullet soliton localized at the center of the gap between the electrodes, which is excited over a broad range of currents, field magnitudes, and orientations. The secondary high-frequency mode appears at higher currents and coexists with the primary bullet mode. Micromagnetic modeling shows that this mode is stabilized by the dipolar field of the bullet, and its spatial profile exhibits two maxima offset from the center of the gap in two opposite directions collinear with the field. Simulations also suggest chaotic dynamics that emerges at large currents due to the incoherent coupling between the two modes. Our results demonstrate the possibility to induce and control complex nonlinear dynamical phenomena in spin Hall oscillators, which can be utilized in the emergent neuromorphic and reservoir computing applications. This work is collaborated with Prof. Sergei Urazhdin, Department of Physics, Emory Universty, USA.