Inverse Color Design Of Silicon Metasurface Based On Tandem Network

Local optical resonance in silicon nanostructures is increasingly used in color printing,which can achieve ultra-high resolution of more than 100,000 pixels per square inch.And compared with metal plasmonic nanostructures such as gold and silver,by changing the geometric parameters of silicon nanostructures to obtain different structural colors,a larger color coverage and higher color saturation can be obtained in the CIE color space.However,the design of specific colors in traditional methods involves iterative optimization of geometric parameters,which is a time-consuming process.So getting millions of different colors in a color space is a very hard job.In recent years,the use of neural networks to complete the design of nanophotonic structures has become a hot research object of many researchers at home and abroad.Through the fitting of large-scale data,neural network can greatly improve the design efficiency of nanostructures.In this paper,the following research work is carried out based on neural network:(1)A forward prediction network model based on a cross-attention mechanism is proposed for accurate prediction of the structural color produced by silicon nano truncated cone.Compared with the traditional fully connected network,the cross-attention network can obtain the nonlinear combination of different input features,and assign different weights to the combination items to achieve the purpose of distinguishing the importance of features.The effectiveness of the cross-attention mechanism is verified by comparative experiments.(2)In the inverse design,the loss function of the tandem network is improved to avoid outputting invalid values.Two improved methods are used in this paper.One is to dynamically add structural loss based on the original color loss of the tandem network.The second is to use the discriminator in the generative adversarial network as a structural loss added to the tandem loss.Both methods can effectively resolve the conflict between the tandem network and the one-to-many data in the dataset.(3)Design and implement a nano truncated cone structural color demonstration system.Based on the first two research points,this paper realizes the forward prediction and reverse design network of the nano truncated cone color,and uses the B/S architecture to build a visual demonstration system.Predict the results and return the results to the frontend interface.