Subwavelength Grating Design Based On Deep Learning

As a new optical device in recent years,sub-wavelength grating has been paid close attention by researchers.It has been widely used in many kinds of optical devices,such as filters,solar absorption films,polarizers and so on.In the structural design of sub-wavelength gratings,it is often necessary to solve the optimal solution by setting the geometric parameters,setting up the optimization algorithm,and consuming a lot of time and computing resources.So how to design subwavelength grating quickly and accurately to meet the needs of reality is an urgent problem to be solved in grating design.The sub-wavelength grating is designed by the method of deep learning,and the subwavelength grating is reverse designed by learning the structure and function of optical device through constructing neural network,it provides a new idea for the design and research of optical field.The main findings and findings are as follows:(1)Construction of sub-wavelength grating data set: modeling,parameter setting and simulation of sub-wavelength grating based on RCWA software,the sub-wavelength grating structure parameters and spectral response curves are obtained,and the required data set is established.There are 46200 sets of one-dimensional sub-wavelength raster data set,including80% training set and 20% test set.Two-dimensional subwavelength raster data set 46080 groups,including 80% training set,20% test set.(2)Establishment of grating model and neural network model: based on the traditional spectral response theory and physical structure of sub-wavelength grating,combining the deep learning theory and neural network framework;The filtering characteristics of subwavelength gratings with different structures are studied,and the grating and the depth neural network(DNN)are constructed.For one-dimensional subwavelength raster networks,there are input layer,hidden layer and four hidden layers,each layer node of hidden layer is(200,200,500,200).The nodes of the hidden layer of the two-dimensional subwavelength grating network layer are(128,512,512,128).(3)Realization of forward simulation and reverse design of sub-wavelength grating: through deep study of the relationship between grating structure parameters and spectral response,the structure parameters of neural network are optimized,and the forward simulation calculation is realized,the error of one-dimensional subwavelength grating is less than 0.014,the error of twodimensional subwavelength grating is less than 0.023,and the time consumed by the traditional method is 817 times that of this method,which is obviously better than the traditional method.The inverse design of the grating is realized,the required spectral response curve is input,the geometric structure parameters of the grating are calculated,the response time is 1.35 s,and the correlation degree of the theoretical spectrum is more than 0.65,which belongs to strong correlation.(4)Experimental verification of forward simulation and reverse design of one-dimensional and two-dimensional filter gratings: using the trained network,input the grating structure made by published literatures into DNN,and carry out forward simulation,the spectral response of DNN network and references are correlated,and the correlation is more than 0.5,which belongs to strong correlation.For the inverse design,one-dimensional and two-dimensional subwavelength gratings are designed and analyzed.By inputting the target spectral curve,the DNN network can give the correct grating structure,and the correlation degree between the corresponding spectral curve and target spectral curve is more than 0.6,which belongs to strong correlationIn this paper,a depth neural network(DNN)is constructed based on the structure and characteristics of the grating,and the prior knowledge of the data set is obtained,finally,the performance of the neural network is tested by the pre-set test set,and the experimental verification of the neural network is carried out by the published references,the spectral response curve of the grating can be obtained at 1.2 s by simply inputting the grating geometry parameters.More importantly,the required grating spectral response curve can be input to get the correct grating geometry parameters.