北京邮电大学学报

  • EI核心期刊

北京邮电大学学报 ›› 2022, Vol. 45 ›› Issue (3): 112-116.doi: 10.13190/j.jbupt.2021-281

• 研究报告 • 上一篇    

基于神经网络的纳米光子结构逆设计

田亮, 潘鹏, 刘玉敏   

  1. 北京邮电大学 电子信息工程学院, 北京 100876
  • 收稿日期:2021-12-03 出版日期:2022-06-28 发布日期:2022-06-01
  • 作者简介:田亮(1987—),男,博士生,邮箱:tianliang667@sina.com;刘玉敏(1976—),男,教授,博士生导师。

The Inverse Design of Nanophotonic Structure Based on Neural Network

TIAN Liang, PAN Peng, LIU Yumin   

  1. School of Electronic Information Engineering, Beijing University of Posts and Telecommunications, Beijing 100876, China
  • Received:2021-12-03 Online:2022-06-28 Published:2022-06-01

摘要: 同扫描参数优化的方法相比,神经网络模型可以在设计误差很小的情况下大大提高结构逆设计的效率。在模拟器与发生器串联的复合模型基础上,提出了2种有效的神经网络复合模型:基于生成对抗思想引入评价模块的复合模型和基于多任务学习增加模拟器模块的复合模型。在传输矩阵法模拟的数据集上进行有效训练后,将2个模型应用于多层薄膜吸收光谱的逆设计,并用具有特殊形状的目标光谱验证了2个模型的逆设计结果。最后,利用多任务串联网络模型设计了多层太阳能吸收器结构参数,在300~1500nm范围内,平均吸收率可达96%。

关键词: 纳米光子结构, 对抗生成, 多任务学习, 复合模型

Abstract: Compared with the method based on scanning parameters and optimization,neural network model can greatly improve the efficiency of structural inverse design with low design error. Two kinds of efficient neural network composite models are proposed: the composite model that introduces critic module based on generation adversarial idea and the composite model that adds simulator module based on multi-task learning. The two models to the inverse design problem of multilayer thin film absorption spectra after effective training on the data set simulated by the transfer matrix method, and the results of the inverse design of the two models are verified by the target spectra with special shapes. Finally, multi-task tandem network model is used to design a set of solar absorber structural parameters, which achieves an average absorption rate of 96% between 300nm and 1500nm.

Key words: nanophotonic structure, adversarial generation, multi-task learning, composite model

中图分类号: