| Artificial electromagnetic metasurfaces have attracted more and more attention because of their unique electromagnetic characteristics.They are commonly used in antennas,radars and various electromagnetic devices.In the design of electromagnetic metasurfaces,electromagnetic simulation softwares are often used to carry out structure modeling,simulation calculation and parameter optimization.When the structure is complex,this is a very time-consuming work.How to design complex metasurface structures quickly has become an important topic in the research field.In recent years,with the rapid development of artificial intelligence technology,some experts and scholars try to construct neural network to design the structural parameters of the metasurfaces,which provides a different idea for the design method of artificial electromagnetic metasurfaces.Using the neural network to design the surface is a very convenient design method,we do not need to concern over the complex modeling process or to scan the various structure parameters.We just need to learn the corresponding relationship between structure parameters and electromagnetic properties,it has a great practical value for the design of metasurfaces.In this paper,the electromagnetic properties of the metasurface and the traditional methods are firstly introduced.Then,the electromagnetic properties of the metasurfaces in the terahertz frequency band are studied based on the deep learning method.Finally,the deep learning method is applied to the design of the metasurfaces constructed by our own.We use the method of software co-simulation to carry out structural modeling,random parameter setting and electromagnetic simulation on the terahertz metasurfaces,obtain the data set composed of the metasurface structural parameters and electromagnetic response.And then we build a deep neural network and train on the data set and learn the relationship between structural parameters and electromagnetic response.Finally,the trained neural networks are used for verification on the test set.When the verification effect is good,we believe that the network can accurately predict the electromagnetic characteristics only needs the structural parameters.Then,based on the forward network,the corresponding inverse design network is further trained.Based on the inverse networks,the metasurface structure parameters can be predicted according to the required electromagnetic characteristics.In order to verify the practicability of deep learning method in the field of metasurface design,several typical terahertz metasurfaces were studied.We build the network and verify the effect of the prediction network and the inverse design network.And then we compared the traditional metasurface research method and the new method in this paper.The results show that this new method has significant practical value. |
