Combining deep learning and nano-optics as a new enabling technology for nano-scale characterization and information processing

将深度学习和纳米光学相结合，作为纳米级表征和信息处理的新使能技术

• 批准号: 415025779
• 负责人: Dr. Peter Wiecha
• 金额: --
• 依托单位: Department of Physics and Astronomy
• 依托单位国家: 德国
• 项目类别: Research Fellowships
• 财政年份: 2018
• 资助国家: 德国
• 起止时间: 2017-12-31 至 2019-12-31
• 项目状态: 已结题
• 来源: https://gepris.dfg.de/gepris/projekt/415025779?language=en
• 关键词: Combining; deep; learning; nano; optics

Despite their ability to solve complicated mathematical problems, classical computational techniques are pretty bad at other tasks, which humans usually solve without any difficulty.Such problems include for instance image or speech recognition.During the last decade, great progress has been made in the field of artificial neural networks (ANNs) - computational models inspired by how the human brain works. ANNs can be trained to categorize such problems and to eventually solve them very efficiently.The goal of this DFG research project is to apply deep artificial neural networks to the field of nano-optics.In a first step, ANNs will be used for the prediction of optical properties of photonic nano-structures and meta-surfaces.In preliminary studies that I have done in preparation of this proposal, I have demonstrated the capability of neural networks for the ultra-rapid prediction of the optical scattering of complex photonic nanostructures.By training ANNs with experimental datasets, I will obtain fully phenomenological models for the prediction of optical effects.In a second work package, deep learning techniques will be applied to the design of nano-optical devices by solving "inverse" problems -- the prediction of nanostructure geometries which offer a desired optical response.Next to the conception of individual nano-optical components, I will explore the use of deep learning methods in complex, multi-modal systems such as speckle-based compressive sensing or all-optical reconfigurable photonic routing and for optical information encoding.

尽管经典计算技术能够解决复杂的数学问题，但在其他任务上却相当糟糕，而人类通常不会有任何困难地解决这些任务。这类问题包括图像或语音识别。在过去的十年里，人工神经网络(ANN)领域取得了巨大的进步，人工神经网络是受人脑工作原理启发的计算模型。神经网络可以被训练来分类并最终非常有效地解决这些问题。这个DFG研究项目的目标是将深度人工神经网络应用于纳米光学领域。在第一步，神经网络将被用于预测光子纳米结构和亚表面的光学性质。在我为准备本提案所做的初步研究中，我已经证明了神经网络用于超快速预测复杂的光子纳米结构的光学散射的能力。通过用实验数据训练神经网络，我将获得用于预测光学效应的完全唯象模型。在第二个工作包中，我已经证明了神经网络能够超快地预测复杂的光子纳米结构的光散射。通过用实验数据集训练神经网络，我将获得用于预测光学效应的完全唯象模型。在第二个工作包中，深度学习技术将通过解决提供期望的光学响应的纳米结构几何形状的预测问题来应用于纳米光学器件的设计。接下来，我将探索深度学习方法在复杂的多模式系统中的使用，例如基于散斑的压缩传感或全光可重构的光子路径，以及用于光学信息编码。

项目成果

Design of plasmonic directional antennas via evolutionary optimization.

• DOI: 10.1364/oe.27.029069
• 发表时间: 2019-06
• 期刊: Optics express
• 影响因子: 3.8
• 作者: P. Wiecha;Cl'ement Majorel;C. Girard;A. Cuche;V. Paillard;O. Muskens;A. Arbouet