Domain Knowledge Assisted Training Dataset Generation for Metasurface Designs

Domain knowledge assisted training dataset generation for deep learning aided metasurface designs is investigated. By combining the domain knowledge of metasurface from the designer and the advanced machine learning (ML) technique, an efficient training dataset generation approach has been successfully achieved. Unlike most existing metasurface generative designs that allow for arbitrary target pattern generation, which results in a time-consuming or even nonconvergence model training process, the proposed method takes the full advantages of the prior knowledge from designer to reduce the target solution space greatly, leading to reduced design cycles and higher explainabilities of the designs. The proposed ML model combines generative adversarial network (GAN) and variational autoencoder (VAE) as an encoder to transfer the original data into the latent space, which greatly improves the design efficiency as demonstrated in the validation results.