| A metasurface is a two-dimensional artificial composite material composed of a series of small metasurface units.It has attracted widespread attention from researchers because it can achieve precise modulation of light waves by accurately controlling each element.Currently,metasurfaces have been applied in fields such as artificial optics,optoelectronic devices,and optical sensors,and there are vast application prospects in the future.Traditional metasurface design requires structural modeling,full-wave numerical simulation,and parameter optimization using simulation software,which is time-consuming and requires high professional skills from designers.With the rapid development of artificial intelligence,more and more researchers have begun to apply deep learning to metasurface design.This method does not require complex modeling or time-consuming simulation and can assist researchers in metasurface design with high precision.It has significant research significance and application value.The main focus of this paper is on the deep learning-based metasurface modeling method,and the main work includes:1.Proposing a convolutional neural network model based on feature fusion for fast prediction of the optical response of metasurface elements.The model extracts features from the metasurface 2D images through convolution and fuses them with the one-dimensional structural parameters of the metasurface to increase the dimensionality of the network input.Meanwhile,genetic algorithm is used to optimize the number of layers and neurons of the network model to obtain the best network model configuration.Finally,transfer learning is introduced to train other metasurface structures,significantly reducing training time.Joint simulation is used to automatically obtain the metasurface dataset,and data preprocessing such as cropping,binarization,and normalization is performed to improve data quality.Experiments have shown that the model can achieve spectral prediction of metasurface elements with different structures in a very short time,and the predicted results are basically consistent with the full-wave simulation results of CST software,with an average accuracy rate of 98%.Comparing with deep neural networks and convolutional neural networks,the performance of the proposed network model is verified to be superior to other models,with mean square error of the real part and imaginary part is reduced to 3.148×10-4 and 4.015×10-4.2.Building a network model composed of a deep convolutional generative adversarial network and a deep neural network,aiming to achieve rapid reverse design of metasurface structures.The discriminator and generator of this model use convolution and deconvolution operations,respectively,to generate metasurface structural images.The deep neural network processes the optical response to obtain its one-dimensional structural parameters.Finally,in view of the low accuracy of the image generated by the deep convolution generation confrontation network,a method based on loss function optimization is proposed to further optimize the generator,so that the accuracy of the generator can be further improved.Through simulation experiments,this model has high accuracy,with the MAE of the generator reduced to 5.19×10-3 and the MSE of the deep neural network reduced to 7.27×10-4.3.A metalens with a numerical aperture of 0.55 and a focusing efficiency of 68.86%under the transmission of electromagnetic waves at a wavelength of 405nm was produced based on the transfer phase principle using the trained network model.The design process using deep learning was ten times faster than traditional design methods. |
PDF Full Text Request