第175期 2021年7月刊
 
 
 
发行人:黄建璋所长  编辑委员:曾雪峰教授  主编:林筱文  发行日期:2021.07.30
 
 

Surface Plasmon Resonance-Induced Diffusion-Limited Aggregation in the Formation of Ag/AgOx Nano-Networks as Broad-Spectrum Transparent Conductors

Professor C. C. Yang

Graduate Institute of Photonics and Optoelectronics, National Taiwan University

台湾大学光电所 杨志忠教授

Based on the generation and transportation of hot electrons and the diffusion-limited aggregation of Ag+ ions induced by the surface plasmon resonance of surface Ag nanoparticles (NPs) with ambient moisture, we develop a new mechanism for fabricating a surface Ag nano-network (NNW) structure on a slightly conductive template through Ag NP reorganization and AgOx formation. Such an extended NNW structure consists of distributed Ag NPs covered and connected by AgOx to form a highly transparent conductive network. Its sheet resistance is as low as ~140 W/square while its transmittance lies between 80 and 90 %. The process of diffusion-limited aggregation is regulated by electrostatic induction between diffuse Ag+ ions in a thin surface water layer condensed from moisture and the sharp tips of an NNW structure during its formation. Oxygen atoms can be dissociated from AgOx when a post thermal treatment at a temperature higher than 412 oC is applied, leaving behind an Ag thin layer, which still connects the remaining Ag NPs to form a different conductive network of ~240 W/square in sheet resistance. Such an NNW structure can stand an elevated temperature up to ~400 oC and is thermally stable. It can extend the transparent spectrum into ultraviolet and near-infrared ranges and can be applied to touch-panel displays, ultraviolet light-emitting diodes, and flexible optoelectronics devices. Figure 1(a) below shows the SEM image of the initial surface Ag NPs before illumination. Figures 1(b)-1(d) show the SEM images of the formed NNW structures with different magnifications after step-5 illumination. In Fig. 1(d), we show the SEM image in the clear area of the NNM structure, in which one can see the remaining low-density, small Ag NPs. Figure 2 shows the transmission spectrum of the sample after each illumination step. We can see that after step-5 illumination, the transmittance is always higher than 83% in the whole visible range.

 Fig. 1 SEM images of an NNW structure.

Fig. 2 Transmission spectrum after each illumination step.

 

Quantitative spectroscopic comparison of the optical properties of mouse cochlea microstructures using optical coherence tomography

Professor Hsiang-Chieh Lee

Graduate Institute of Photonics and Optoelectronics, National Taiwan University

台湾大学光电所 李翔杰教授

Currently, the cochlear implantation procedure mainly relies on using a hand lens or surgical microscope, where the success rate and surgery time strongly depend on the surgeon’s experience. Therefore, a real-time image guidance tool may facilitate the implantation procedure. In this study, we performed a systematic and quantitative analysis on the optical characterization of ex vivo mouse cochlear samples using two swept-source optical coherence tomography (OCT) systems operating at the 1.06-µm and 1.3-µm wavelengths. The analysis results demonstrated that the 1.06-µm OCT imaging system performed better than the 1.3-µm OCT imaging system in terms of the image contrast between the cochlear conduits and the neighboring cochlear bony wall structure. However, the 1.3-µm OCT imaging system allowed for greater imaging depth of the cochlear samples because of decreased tissue scattering. In addition, we have investigated the feasibility of identifying the electrode of the cochlear implant within the ex vivo cochlear sample with the 1.06-µm OCT imaging. The study results demonstrated the potential of developing an image guidance tool for the cochlea implantation procedure as well as other otorhinolaryngology applications.

Fig. (a) Example cross-sectional OCT image of the mouse cochlea sample treated with ethylenediaminetetraacetic acid (EDTA) for 7 days, acquired with the 1.06 µm OCT system. RM: Reissner’s membrane; SV: scala vestibuli; SM: scala media; ST: scala tympani; OC: organ of Corti. Scale bar: 250 µm. (b) Schematic diagram showing the anatomic features of the mouse cochlea sample. (c) The volumetric rendering of three-dimensional OCT images of the cochlear sample. Red box: bounding box of the rendering.

 

Reference:

Ting-Yen Tsai, Ting-Hao Chen, Hsin-Chien Chen, Chuan-Bor Chueh, Yin-Peng Huang, Yi-Ping Hung, Meng-Tsan Tsai, Bernhard Baumann, Chih-Hung Wang, and Hsiang-Chieh Lee, "Quantitative spectroscopic comparison of the optical properties of mouse cochlea microstructures using optical coherence tomography at 1.06 µm and 1.3 µm wavelengths," Biomed. Opt. Express 12, 2339-2352 (2021)

 

     
 
 

— 资料提供:影像显示科技知识平台 (DTKP, Display Technology Knowledge Platform) —

— 整理:林晃岩教授、卓真禾 —

竞争式光子神经网络

神经网络(NNs)是受神经启发的人工智能概念,它基于非线性元素(神经元)的集体响应来实现计算。透过这种方式,神经网络模仿了生物大脑中计算的最基本面向之一。特定的讯息处理任务通常是在统计优化的基础上,透过调整网络拓扑来进行“编程”(programming),亦即是学习。神经网络可以区分狗和猫,或者进行更严肃和有用的杰作,并且代表了计算器领域激动人心的最新进展之一。

但是,每个神经元的状态都需要透过网络的连接进行计算,且其相关的计算量大大地击败传统处理器。与其用数字计算器来仿真神经网络,不如将大量的精力投入到硬件上,其硬件的物理定律模仿特定的神经网络的概念[1]。一个主要的重点是在物理上实现神经网络的连接,而光子处理的平行化则有望实现更快,更高效的神经网络处理器[2]。

在《自然光子学》中,周(Zhou)等人表明,即使使用非专业和现成的组件实现的光子神经网络,也可以胜过专为神经网络应用量身定制的一流图形处理单元(GPU)[3]。他们所报导的准确性、速度和能源效率,可与竞争式数字神经网络基准测试模型以及现今的GPU媲美,展示了光子解决方案与未来高性能计算的关联性。

在周(Zhou)等人的工作中,当固态激光照射,单个神经元的状态是由数字微镜装置(DMD)内部的微镜的反射来达成。液晶空间光调制器(SLM)上显示的相位屏蔽,对DMD信号的绕射会产生平行和可重新配置的网络连接,透过快速摄影进行光学检测的模块平方,会为每个神经网络增加非线性。这种串联的非线性运算会大大地放大数据表示的维数,这最终使神经网络能够挖掘出隐藏的特征,然后可以利用它们来进行具有挑战性的计算。它们执行的互连受到严格限制,但是它们通过串联多层实现了具有竞争式性能。他们计算出他们的绕射处理单元(DPU)达到240.1 TOP s-1,能量效率为1.578 TOP J-1。预算中包括了与运作系统有关的所有组件和设备。值得注意的是,在这两个指标上,作者均击败了辉达(Nvidia)的顶级Tesla V100 Tensor Core GPU。作者所达到的准确性是进一步的实质性进步。让我们以MNIST数字识别和DPU透过分时多任务的三层深神经网络为例。透过物理模型进行的初始仿真可达到97.6%的测试准确度。但是,当将预优化的网络配置转移到物理DPU时,该值大大降低了。作者继续通过迭代实验更新来优化SLM的相位屏蔽,直到他们在15个训练时期后透过实验获得96.2%的测试准确度。更复杂的DPU分时多任务方案(D-NIN-1(++))(图1a)实现了卷积神经网络,在MNIST中具有99%的测试准确度胜过LeNet-4架构的98.9%的准确度(图1b)。在进一步测试中,例如时尚的MNIST(图1c)以及人类动作识别,证实了具有竞争式的表现。

图1、光学神经网络计算。(a)周(Zhou)等人创建了一个主要单元,该单元包括透过绕射处理单元(DPU)并联连接的光子神经元,以非线性创建特征空间。空间光调制器使连接可重新配置以进行训练,并允许一个这样的原理单元来计算网络的特征图和层的配置,这是深层的网络中的网络(D-NIN-1)拓扑。(b,c)不同的光学神经网络拓扑以性能为基准,例如在标准手写数字(MNIST)和时尚物品(Fashion-MNIST)数据集之中,周(Zhou)等人的报告推理准确性可与经典计算机上仿真的先前突破性神经网络体系结构(LeNet 1-5)相提并论。

周(Zhou)等人的研究中具有许多含义。DPU利用光子学的平行化可以物理式实现网络连接。最近在具有固定拓扑结构的随机重复[4]和深层线性网络[5]中显示,光子方法具有竞争式,并且可以实现GPU优越的缩放性。周(Zhou)等人在一系列通用的拓扑结构中量化此优势,包括学习并确认光子神经网络可以与GPU上运行的类似神经网络模型竞争,因为它们在当今的一些基准测试数据集之中具有较高的推理精度。此外,该概念是可扩展的:神经元在二维平面中实现,可连接利用沿第三维的光学传播,而第三维在增加神经元数量时使物理尺度大小线性变化[6]。

我们希望当前对机器学习在光学上的浓厚兴趣仅仅是个开始[7]–[9]。最重要的是,实现光学神经网络中的非线性可以使用电子方式,经由电子到光学到电子的转换来完成。未来的挑战是以光学方式实现非线性,避免转换带来的瓶颈。这有望大大提高性能[10][11]。光领域的原位学习是另一个需要解决的开放性问题,需要在硬件中加以解决和证明。

 

参考资料:

Daniel Brunner & Demetri Psaltis, “Competitive photonic neural networks,” Nature Photonics volume 15, pages323–324(2021)

https://www.nature.com/articles/s41566-021-00803-0

DOI: s41566-021-00803-0

参考文献:

[1] Marković, D., Mizrahi, A., Querlioz, D. & Grollier, J. Nat.Rev. Phys. 2, 499–510 (2020).

[2] Psaltis, D., Brady, D., Gu, X.-G. & Lin, S. Nature 343, 325–330 (1990).

[3] Zhou, T. et al. Nat. Photon. https://doi.org/10.1038/s41566-021-00796-w (2021).

[4] Rafayelyan, M., Dong, J., Tan, Y., Krzakala, F. & Gigan, S. Phys.Rev. X 10, 41037 (2020).

[5] Lin, X. et al. Science 361, 1004–1008 (2018).

[6] Dinc, N. U., Psaltis, D. & Brunner, D. Photoniques 114, 34–38 (2020).

[7] Wetzstein, G. et al. Nature 588, 39–47 (2020).

[8] Xu, X. et al. Nature 589, 44–51 (2021).

[9] Feldmann, J. et al. Nature 589, 52–58 (2021).

[10] Teğin, U., Yıldırım, M., Oğuz, İ., Moser, C. & Psaltis, D. Preprint at https://arxiv.org/abs/2012.12404 (2020). [11] Porte, X. et al. Preprint at https://arxiv.org/abs/2012.11153 (2020).

 
       
       
 
 
版权所有   国立台湾大学电机信息学院光电工程学研究所   http://gipo.ntu.edu.tw/
欢迎转载   但请注明出处   http://gipo.ntu.edu.tw/monthly.htm/