An End-to-End CNN Framework for Electromagnetic Response Prediction of Metasurface Design

To solve the problems of traditional full-wave electromagnetic (EM) simulations, such as heavy reliance on expert knowledge, inefficient parametric modeling, and high computational costs, an end-to-end convolutional neural network (CNN) framework is proposed to expedite the metasurface design. The proposed CNN consists of three convolution-pooling layers, five fully- connected layers, and five batch-normalization layers. An automated model-data generation framework is used to build metsurface structure, solve EM simulation, and generate dataset. The images of metasurfaces are input to CNN to quickly predict the electromagnetic response. To validate the design concept, an electromagnetically induced transparency (EIT) metasurface is designed to predict the transmission spectra. The results show that the CNN model has a high accuracy of above 94%. The prediction period can be reduced by four orders in comparison with the traditional EM simulations. This method offers an efficient solution for metasurface design and optimization.