Deep Neural Networks for Efficient Simulation of Metasurface

This paper proposes a neural network-based method for the efficient simulations of the metasurfaces. A convolutional neural network (CNN) is used to extract features of the metasurfaces, and a long short-term memory network (LSTM) is utilized to predict the electromagnetic fields. To ensure the accuracy, a coarse mesh finite-difference time-domain (FDTD) simulation data is employed as an additional input of the LSTM. By integration the CNN and LSTM with a coarse grid compensation data, the accuracy of optical response prediction is enhanced. The network achieves a MSE of 0.0142 and an accuracy of 98%. This framework can output spectral predictions for given metasurface geometric structure image within only 5 seconds, while the time for the software is 5min. Experimental results demonstrate that the proposed method can predict the electromagnetic fields efficiently with satisfactory accuracy for the metasurfaces.