Abstract

The two-dimensional optical metasurface can realize the free regulation of light waves through the free design of structure, which is highly appreciated by researchers. As there are high requirements for computer hardware, long time for simulation calculations, and data waste in the process of using the time-domain finite-difference method to solve the optical properties of the metasurface, the deep neural network (DNN) is proposed to predict the spectral response of an optical metasurface. The structural parameters of the metasurface are taken as inputs and the metasurface transmission spectrum is used as the output. To achieve better prediction results, different gradient descent algorithms were selected and the parameters of the DNN model were optimized. After 5 × 10⁴ times of epoch training, the loss function mean squared error (MSE) reaches 2.665 × 10⁻³, the sum error of 98% test data is less than 3.23, and the relative error is less than 2%. The results show that the DNN model has an excellent prediction effect. Compared with the traditional simulation method, the efficiency of this model is improved by 10⁴ times, which can improve the efficiency of optical micro-nano structure design.

© 2022 Optica Publishing Group

Convolutional neural networks used for random structure SPP gratings spectral response prediction

Tianle Qu, Liping Zhu, and Zhenghua An
Opt. Lett. 48(2) 448-451 (2023)

Inverse design paradigm for fast and accurate prediction of a functional metasurface via deep convolutional neural networks

Xudong Du, Chengan Zhou, Hongbai Bai, and Xingxing Liu
Opt. Mater. Express 12(10) 4104-4116 (2022)

Laser-based selective BTEX sensing using deep neural networks

Mhanna Mhanna, Mohamed Sy, Ayman Arfaj, Jose Llamas, and Aamir Farooq
Opt. Lett. 47(13) 3247-3250 (2022)

Long short-term memory neural network for directly inverse design of nanofin metasurface

Wenqiang Deng, Zhengji Xu, Jinhao Wang, and Jinwen Lv
Opt. Lett. 47(13) 3239-3242 (2022)

Reconfigurable origami hologram based on deep neural networks

Kang Wang, DaShuang Liao, and Haogang Wang
Opt. Lett. 49(8) 2041-2044 (2024)