Inverse Design Model for Meta-Atoms Based on Autoencoder

This paper proposes an inverse design model for meta-atoms based on the autoencoder. The meta-atom is non-parametrically modeled to obtain better electromagnetic (EM) performance. With the introduction of a loss function, the meta-atom structure is reduced in dimension to obtain a latent vector close to its corresponding EM response. To avoid overfitting, the EM responses are converted into low-dimensional vectors by the transfer function. After training, the expected EM response is input to obtain the structure of target meta-atom directly. The accuracy of the result is verified by CST simulations.