Research On The Application Of Neural Networks In Metasurface Device Design

Metasurfaces are special metamaterial structures composed of artificially designed two-dimensional wavelength scale elements.By ingenious design of metasurface structures,researchers achieve complete manipulation of the physical properties of incident electromagnetic waves,such as phase,polarization and dispersion.However,the design of metasurface structures frequently relies on massive unit structure optimization processes,which are time-consuming and heavily dependent on research experience in related fields.In recent years,with the rise of artificial intelligence algorithms,neural network methods based on deep learning have been widely used in cutting-edge scientific and engineering fields and have shown excellent performance.At present,neural networks also have injected new vitality into the field of metasurfaces design.Through the training of a large number of data sets,the neural network can explore the complex implicit relationship between the metasurface structure and its electromagnetic spectrum,so as to simplify the design difficulty,shorten the design time and improve the design performance.Therefore,the neural network can well deal with traditional metasurfaces facing challenges in the process of design,and has important research significance and application value in the application research of metasurface device design.In this paper,a neural network model composed of forward modeling and reverse design networks is proposed.We obtained the data set through the MatlabCST co-simulation method and completed the training of the model.The data set contains 20,000 data pairs composed of metasurface structure and corresponding electromagnetic response.From our training results,it can be seen that the trained network has high robustness,which can help to discover the complex,non-intuitive relationship between the unidirectional transmission metasurface structure and its corresponding optical response,thus avoiding the time-consuming and forward deductive numerical simulation process in the past metasurface design.In addition,the neural network model can retrieve a reliable design solution from a given requirement in a few seconds.Therefore,the method can be used as a powerful tool to study the complex interactions between light and matter in the real world,thus accelerating the on-demand design of metasurface device structures.In order to further verify the reliability and practicability of the proposed method,we also use the reverse design network to invert the unidirectional transmission structure parameters of a target electromagnetic response,and prepare the corresponding metasurface structure samples experimentally.The sample was measured by terahertz time-domain spectroscopy system.The experimental results show that the asymmetric transmission response based on the inversion structure is basically consistent with the target electromagnetic response.The above results confirm our view and provide a new method for the design and application of functional metasurface.