Novel method of performance-optimized metastructure design for electromagnetic wave absorption in specific band using deep learning

In this paper, we propose a new method that utilizes deep learning techniques to design a metastructure for an electromagnetic absorber. This method enables the effective design of a metastructure with the desired performance (spectrum of S11<−10 dB) in the frequency band specified by the designer, within a wideband range from 2 to 40 GHz. The proposed absorber consists of two dielectric layers with varied conductive patterns and a back reflector. Critical to the absorber's microwave performance is the binary pattern configuration, organized in a 20-pixel square, along with the sheet resistance and layer thickness of each layer, contributing to a significant design freedom exceeding ${10}^{37}$ degrees of freedom. Our model for performance-optimized design involves three steps: Initially, with limited data from 26,000 sets, a Variational Autoencoder (VAE) was trained to map S11 spectra and arrange a latent space linked to metastructure. Subsequently, we developed a spectrum prediction network to correlate patterns with S11 spectra, leveraging a pre-trained decoder from the auxiliary VAE in the first step. The final step trains a network for designing a metastructure with broadband absorption. To verify the performance of a metastructure designed by the developed method, we compared their performances with those obtained through Finite Difference Time Domain (FDTD) simulation and the developed network. And also to further validate our approach experimentally, the designed metastructures were fabricated by silkscreen printing using carbon paste ink, and some bands (1–18 GHz, 26.5–40 GHz) were measured to compare with the performance predicted by the VAE network.