[论文解读] Focused Surface Acoustic Wave induced nano-oscillator based reservoir computing

Md. Fahim F. Chowdhury, Walid Al Misba+5 more|arXiv (Cornell University)|Dec 7, 2021

Neural Networks and Reservoir Computing参考文献 42被引用 9

一句话总结

本文提出了一种基于表面声波（SAW）驱动的纳米磁体阵列作为物理回声状态计算系统，其中4 GHz的聚焦SAW通过调制输入纳米磁体的磁化动力学，在耦合的输出纳米磁体阵列中诱导出非线性响应。采用包络状态读出与Moore-Penrose伪逆训练方法，该系统在分类100 MHz正弦波与方波输入时实现了100%的准确率，具备4.69比特的短期记忆（STM）能力与5.39比特的奇偶校验（PC）容量，同时其能耗比基于CMOS的回声状态网络低两个数量级。

ABSTRACT

We demonstrate using micromagnetic simulations that a nanomagnet array excited by Surface Acoustic Waves (SAWs) can work as a reservoir that can classify sine and square waves with high accuracy. To evaluate memory effect and computing capability, we study the Short-Term Memory (STM) and Parity Check (PC) capacities respectively. The simulated nanomagnet array has an input nanomagnet that is excited with focused SAW and coupled to several nanomagnets, seven of which serve as output nanomagnets. The SAW has a carrier frequency of 4 GHz whose amplitude is modulated to provide different inputs of sine and square waves of 100 MHz frequency. The responses of the selected output nanomagnets are processed by reading the envelope of their magnetization state, which is used to train the output weights using regression method (e.g. Moore-Penrose pseudoinverse operation). For classification, a random sequence of 100 square and sine wave samples are used, of which 80 % are used for training, and the rest of the samples used for testing. We achieve 100 % training accuracy and 100 % testing accuracy for different combination of nanomagnets as outputs. Further, the STM and PC is calculated to be ~ 5.5 bits and ~ 5.3 bits respectively, which is indicative of the proposed acoustically driven nanomagnet oscillator array being well suited for physical reservoir computing applications. Finally, the ability to use high frequency (4GHz, wavelength ~1 micron) SAW makes the device scalable to small dimensions, while the ability to modulate the envelope at lower frequency (100 MHz) adds flexibility to encode different signals beyond the sine and square waves demonstrated here.

研究动机与目标

• 开发一种基于表面声波（SAWs）激励纳米磁体的低能耗、可扩展的物理回声状态计算系统。
• 证明SAW诱导的纳米磁体阵列中的磁化动力学可作为具备强记忆能力与计算容量的非线性储水库。
• 通过微磁仿真评估系统在短期记忆（STM）与奇偶校验（PC）任务中的性能。
• 对比基于SAW的储水库与传统CMOS基回声状态网络的能量效率。

提出的方法

• 采用MuMax3进行微磁仿真，以模拟在SAW诱导的应力各向异性作用下纳米磁体的Landau-Lifshitz-Gilbert（LLG）动力学。
• 在LiNbO₃衬底上集成聚焦叉指换能器（FIDT），产生4 GHz的SAW，通过周期性应力驱动输入纳米磁体的铁磁共振。
• 输入信号通过100 MHz的SAW包络调制进行编码，形成正弦波与方波的随机序列。
• 通过七个输出纳米磁体的磁化动力学包络读取储水库状态，每信号周期采样N=20个虚拟节点。
• 采用Moore-Penrose伪逆的线性回归方法训练输出权重，实现分类与容量评估。
• 基于SAW功率、电压与器件尺寸计算能量耗散，并与CMOS基回声状态网络进行比较。

实验结果

研究问题

• RQ1 SAW在纳米磁体阵列中诱导的应力是否能产生足够强的非线性动力学，以作为计算用的物理储水库？
• RQ2 SAW驱动的纳米磁体储水库的短期记忆（STM）与奇偶校验（PC）容量分别是多少？
• RQ3 该SAW基储水库的能量消耗与CMOS基回声状态网络相比如何？
• RQ4 该系统能否在正弦波与方波等时间序列信号上实现高分类准确率？

主要发现

• SAW驱动的纳米磁体阵列在分类100 MHz正弦波与方波输入时，实现了100%的训练准确率与100%的测试准确率。
• 系统展现出约4.69比特的短期记忆（STM）容量，表明其具备强大的时间记忆能力。
• 奇偶校验（PC）容量测量值约为5.39比特，反映出其高度的非线性特性与计算能力。
• SAW基储水库的能量耗散约为每输入0.87 × 10⁻¹² J，比CMOS基回声状态网络低两个数量级。
• 4 GHz SAW的使用使系统具备向纳米尺度扩展的潜力，而100 MHz包络调制则支持灵活的信号编码，超越正弦波与方波的限制。