Structural Color Design Method Based On Bidirectional Neural Network

Structural color is a kind of physical color formed by the interaction of subwavelength nanostructures with electromagnetic waves in the visible regime.Compared with traditional pigment colors,it has the advantages of vivid colors,stable chemical properties,easy recycling and less pollution.Structural colors can break the visible light diffraction limit with reported ultra-high resolution of 100,000 dpi,thus shows great application potential in related fields such as high-performance display and high-density optical data storage.Conventional design of structural color requires electromagnetic simulations combined with iterative optimizations.This process consumes significant amount computation time and resource,which may still not accurately fulfill the design needs.As a powerful fitting algorithm,artificial intelligence neural network tools can accurately explore and map the complex physical relationships between the nanophotonic structures and its optical response,thereby such tool is revolutionary in performing accurate,efficient prediction of optical properties and also direct inverse design of device structures.Based on the application prospects and design requirements of structural colors,this paper proposed to use the latest artificial intelligence deep learning algorithms to realize highly efficient structural color design,though the construction of a bidirectional neural network for both forward prediction and inverse design functions.The accurate design of structural color has been realized with a great decreasing in time complexity.With silicon nanopost as the basic unit,a periodic nanostructure that generates various structural colors is constructed.The structural parameters including period P,spacing G,diameter D,and height H were selected to simulate for their corresponding structural colors,and a total of 5214 training data sets of was constructed.The deep neural network training process determined the optimal hyperparameters and training results showed good convergence.Using 360 groups of data to test the effect of the network,the average absolute errors of the three color parameters x,y,Y in CIE 1931 standard color space are 0.0020,0.0031,0.0019 in forward prediction.If the value of less than0.01 is taken as the standard of accurate prediction,the accuracy rates of x,y,and Y have reached 99.2%,95.6%,and 99.2%,respectively.The inverse design is realized by using a tandem neural network to tackle the non-uniqueness problems.For the same 360 groups of test data,the average absolute errors of the three color parameters x,y,Y are 0.0019,0.0025,0.0022,and the accuracy rates are 97.8%,95.6%,and 97.5%,respectively.Additional examples are selected to demonstrate the entire design process of the network,and tests of reproducing actual paintings colors are made.These results show accurate and efficient designs of colors.Therefore,the developed tool can guide the structure color fabrication and application in the future.