Design Of Phase-modulating Metasurface Device Based On Deep Learning

Metasurface,an artificial two-dimensional composite structure or composite material,due to its unconventional controllability of electromagnetic waves,has received extensive concerns from researchers in the field of subwavelength electromagnetics.The derived metasurface devices,such as radar,antennas,sensors,and other electromagnetic waves-controlled devices,have the advantages of easy integration and flexible control of electromagnetic waves and show the tremendous potential to replace traditional optical components.The classical optimization design approach of metasurface devices features the disadvantages of slow convergent speed and ease to fall into the local minimum,which restricts the development of the metasurface.As an artificial intelligence algorithm with spanking computing speed,a variety of methods based on deep learning have been proposed for the design of phase-modulating metasurface devices.In this paper,according to the existing basic design framework of phase-modulating metasurface devices.A simple and fast design scheme of high-efficiency phase-modulating metasurface devices based on transfer learning is proposed.The validity of this method is demonstrated by the actual device design.Our works include:1.The existing methods often use a large dataset or complex network structure to achieve higher prediction accuracy of metasurface spectrum,which needs a lot of time for establishing dataset or adjusting and training the neural network.Here,we propose a "prescreening" method,which is to screen the dataset according to the existing experience and by referencing the metasurface structures or parameters in other researches.As a result,the neural network can also achieve hochpraezis spectral prediction ability after training on the dataset consisting of about 10000 instances.2.Because of the existing research methods,to enrich the basic structure of metaatoms,a large amount of additional electromagnetic simulation data and repeated neural network adjusting and training are required.We propose to use transfer learning technology to migrate the knowledge between similar tasks.It not only reduces the time for network tuning and training by 10 times but also the MSE by 9.2%.3.Finally,we propose an expeditious and accurate method for designing highefficiency phase-modulating metasurface devices,which is based on transfer learning technology and evolution algorithm.The deflection efficiency of the designed metasurface deflector is up to 92%,and the performance of the designed metalens approximates the preset goal.The successful design of these devices shows that the proposed method shows great prospects in the domain of fast metasurface devices designing.