Equivalent Circuit Theory-Assisted Deep Learning for Accelerated Generative Design of Metasurfaces

In this article, we propose an equivalent circuit theory-assisted deep learning approach to accelerate the design of metasurfaces. By combining the filter equivalent circuit theory and a sophisticated deep learning model, designers can achieve efficient metasurface designs. Compared with most existing metasurface generative design methods that rely on arbitrarily generated training dataset (TDS), the proposed method can adaptively produce highly relevant and low-noise training samples under the guidance of filter equivalent circuit theory, resulting in a significantly narrowed target solution space and improved model training efficiency. Furthermore, we select the variational autoencoder (VAE) as a generative model, which can compress the raw training samples into a lower-dimensional latent space where optimization methods, such as genetic algorithm, can be more efficiently executed to find the optimal design than a brute-force search. To verify the effectiveness of the proposed method, we apply it in the creation of three examples of frequency selective surfaces (FSSs), presenting wide-band, dual-band, and band-stop responses. Experimental results show that the proposed method can realize much faster and more stable convergence than deep learning design methods without domain knowledge.

Automated summary

In this article, we propose an equivalent circuit theory-assisted deep learning approach to accelerate the design of metasurfaces. By combining filter equivalent circuit theory and a sophisticated deep learning model, efficient metasurface designs can be achieved. The proposed method can adaptively produce relevant and low-noise training samples under the guidance of the filter equivalent circuit theory, resulting in a significantly narrowed solution space and improved model training efficiency. The use of a variational autoencoder as a generative model further enhances the optimization process.