计算机科学
人工神经网络
深度学习
传输(电信)
算法
相(物质)
人工智能
架空(工程)
维数之咒
缩放比例
数学
物理
电信
几何学
量子力学
操作系统
作者
Zhenxiang Shi,Haiou Lu,Xinyu Yu,Kai Ni,Qian Zhou,Li Wang
摘要
In traditional metasurface structure design, it heavily relies on electromagnetic simulations to obtain transmission and phase spectral, followed by empirical adjustments. This iterative trial-and-error process, especially when dealing with multi-objective optimization tasks, demands intensive and time-consuming computations, which to a certain extent restricts the development of the metasurface research field. In this paper, a proposed method achieves rapid prediction of spectral responses corresponding to structural units by seeking analytical solutions within the constructed neural network model. The proposed deep learning-based method for predicting transmission and phase spectral of metasurface units consists of metasurface unit dataset construction and a ResNet-based network framework. In the dataset construction approach, an overhead view of the unit structure is extracted and transformed into a binary image, where scaling factors are coupled into the two-dimensional image to increase dimensionality. This enables the representation of different structures such as square pillars, elliptical cylinders, and varying sizes of metasurface units using the same data format, significantly enhancing network generalization. Within the network framework, ResNet is employed to predict the real and imaginary parts of the S21 parameter, which are then inverted to obtain transmission and phase information. The progressive training method employed in combination with this framework yields high prediction accuracy. The deep learning-based method for predicting transmission and phase spectral of dielectric metasurface units, as revealed in this paper, achieves a 7200-fold increase in prediction speed compared to traditional electromagnetic
科研通智能强力驱动
Strongly Powered by AbleSci AI