基于生成对抗网络模型的高自由度超表面原子逆向设计  

Reverse design of high-degree-of-freedom metasurface based on generative adversarial network model

作　　者：王军凯 林森 刘港成 伍滨和 王春瑞[1] 周健[2] 孙浩[2] WANG Junkai;LIN Sen;LIU Gangcheng;WU Binhe;WANG Chunrui;ZHOU Jian;SUN Hao(College of Science,Donghua University,Shanghai China;Shanghai Institute of Microsystem and Information Technology,Chinese Academy of Sciences,Shanghai,China)

机构地区：[1]东华大学理学院,上海 [2]中国科学院上海微系统与信息技术研究所,上海

出　　处：《东华大学学报（自然科学版）》2024年第5期61-68,共8页Journal of Donghua University(Natural Science)

基　　金：国家自然科学基金(61975029)。

摘　　要：为解决传统方法在设计电磁响应差异显著的超表面时需要消耗大量时间以及计算资源的问题,打破超表面在波前调控领域的应用瓶颈,将生成对抗神经网络与预测神经网络组合的模型引入设计流程中,实现了对高自由度超原子的快速、精确逆向设计。研究表明,使用卷积网络模型作为光谱预测器代替复杂的电磁数值模拟时,可实现对超原子电磁响应光谱的快速、准确分析。生成对抗网络模型则取代了传统的迭代优化试错法,能够根据设计要求快速产生多个满足条件的候选超原子结构。该模型结合工艺容差分析等手段,为全息显示、光学功能器件等新颖应用提供可靠的设计方法。To solve the problem that the traditional method of designing metasurfaces with different electromagnetic responses requires a lot of time and computing resources.It is necessary to break the bottleneck of the application of metasurfaces in the field of wavefront regulation.the research This paper introduces a model that combines generative adversarial neural networks and predictive neural networks into the design process to achieve rapid and accurate reverse design of high-degree-of-freedom meta-atoms.The research demonstrates that using the convolutional network model as a spectral predictor,instead of complex electromagnetic numerical simulations,can achieve fast and accurate analysis of the electromagnetic response.The generative adversarial network replaces the traditional iterative optimization trial-and-error method,and achieves the goal of quickly generating multiple candidate meta-atomic structures meeting design requirements.This model combines process tolerance analysis and other methods to provide reliable design methods for novel applications such as holographic displays and optical functional devices.

关 键 词：超表面 逆向设计 预测网络 生成对抗网络 高自由度 

分 类 号：O43[机械工程—光学工程]