Inverse design of electromagnetically induced transparency (EIT) metamaterials based on autoencoder with reconstruction error

In the paper, we design a new deep learning (DL) network to analysis and design the electromagnetically induced transparency (EIT) metamaterials. The network is divided into two parts: dimensionality reduction reconstruction network and inverse design. The dimensionality reduction reconstruction network mainly deals with high-dimensional input data. In the inverse design, we combine the convolutional neural network (CNN) and long short-term memory (LSTM) to predict the EIT metamaterial structure. The inverse design has reduced the mean square errors (MSE) on the validation set to 0.0032 and predict the structure parameters within the error of 0.26 µm, which are better than other reports. Besides, we avoid the overfitting problem and one-to-many problem to some extent. By comparing the spectra of real and predicted parameters on the validation set, we are confident that the network will pave a new path for designing EIT metamaterials.