An Optimization Approach Based on Separated Latent Space for Inverse Design of Metasurfaces

In this letter, we propose an inverse design framework of electromagnetic metasurfaces (EMMSs) realized by optimizing latent variables in a separated latent space. In the training phase, the framework sequentially connects an inverse network to generate the shapes of scatterers by the input of reflection coefficients and a forward network to recover the coefficients with the generated scatterers. The architecture of the generative network adopted the fundamental structure of variational autoencoder, but the latent space was split into four subspaces representing the predefined classes of metasurfaces by applying multivariate normal distribution. After model training, for the inverse design of EMMSs, the generator of the inverse model and the decoder of the forward model are linked to an optimizer. This network pipeline enabled to build more streamlined inverse design model, and on the other hand the optimization strategy in the separated latent spaces improved the prediction accuracy of EMMSs concerning desired electromagnetic (EM) properties reducing the misfit rate by 31.9% compared with the counterpart method having a single distribution latent space.