| Metamaterials are two-dimensional equivalent metamaterials composed of subwavelength elements on ultra-thin surfaces.The metasurface can completely control the amplitude,phase,dispersion,momentum and polarization of the incident electromagnetic wave.Metasurface design mainly includes two ideas:forward prediction result and reverse design structure.Forward prediction results refer to the output of the total optical response according to the given geometric parameters.Reverse design of the structure refers to the output of geometric parameters according to the desired optical response.In general,advanced iterative calculation methods,combined with finite element modeling or finite-difference time-domain(FDTD),are used to predict the optical properties of metasurface materials.This traditional design process is affected by human guided error and time-consuming.Common reverse design methods include genetic algorithm and adjoint method.For multi-parameter problems,genetic algorithm needs complex computing power and a lot of time to solve,and with the increase of the number of parameters,the computing time will increase exponentially.The establishment of adjoint method often requires deep photonic knowledge,and is limited by experimental conditions or existing theoretical basis,which makes it impossible to adjust and re-evaluate a few parameters for many times to approach the target.Compared with traditional optimization algorithms,machine learning algorithm can predict unknown problems by learning the complex relationship between model variables from a large data set.This data-driven scheme significantly reduces the calculation time of metasurface design,and can provide a more comprehensive and systematic optimization of metasurface characteristics.In this paper,an artificial neural network based method for forward spectral prediction and structural reverse design of a monolayer metal-film metasurface periodic unit under linear polarization incident is presented.Main work includes:firstly,using the finite difference time domain method to produce super surface spectrum training data set,by putting a cycle unit coding for 8×8 binary grid to generate arbitrary shape of the surface structure,and its coding for 64×64 pixel grayscale,input to the network training,to generate the output of 64×64 pixels,which can generate arbitrary shape with expectations,optical properties of super surface design patterns.Secondly,a convolutional neural network is trained to achieve forward spectrum prediction.Trained convolutional neural network model can be implemented within the 5 s rapid spectral prediction,and the model output prediction and FDTD simulation spectroscopy,the average accuracy of 0.982,model has good fitting effect,can replace the traditional simulation software to calculate,reduce the computing time,a lack of experience in using software services.Finally,a conditional deep convolution generation adversation network model is built to realize the reverse structure design.Use generative models to explore alternatives to large design Spaces.After training,the generated model can realize the user’s on-demand design.For each test expected spectrum,the generated model will produce reasonable hypersurface geometric structure images.The predicted spectrum of the generated geometry is basically consistent with the expected spectrum,and the average accuracy of the spectrum is about 0.923.Compared with the traditional generation model,the proposed model improves the accuracy by 0.056 and generates clearer images.It can improve the accuracy of metasurface design and provide some guidance for parameter design optimization. |
