Deep Learning-Based Inverse Design of Stochastic-Topology Metamaterials for Radar Cross Section Reduction.

Area of Science:

• Electromagnetic Metamaterials
• Deep Learning in Electromagnetics
• Computational Electromagnetics

Background:

• Electromagnetic (EM) metamaterials offer unique properties but traditional designs are limited by specific topologies.
• Stochastic topologies provide diverse EM properties but require extensive computational resources and designer expertise for optimization.
• Current design processes are time-consuming, relying heavily on full-wave simulations.

Purpose of the Study:

• To develop a deep learning framework for efficient electromagnetic metamaterial design.
• To establish a rapid mapping between metamaterial structure and EM response, replacing slow simulations.
• To enable inverse design for applications like 1-bit coding metamaterials.

Main Methods:

• Utilized a deep learning agent model combining a Convolutional Block Attention Module-enhanced Variational Autoencoder (CBAM-VAE) with a Transformer-based predictor.
• CBAM-VAE enhances feature extraction and reconstruction of metamaterial structures.
• Transformer predictor uses an encoder-only configuration for efficient EM response prediction from latent variables.

Main Results:

• The trained model rapidly generated two cells for 1-bit coding metamaterials.
• Designed coding metamaterials achieved over 10 dB reduction in radar cross-section (RCS) from 6 to 18 GHz compared to metallic plates.
• The framework requires a small data fraction for training, significantly reducing computational cost.

Conclusions:

• The deep learning approach offers a novel and efficient perspective for designing EM metamaterials.
• Simulation and experimental results validate the reliability and effectiveness of the proposed design methodology.
• This framework accelerates the discovery of metamaterials with tailored electromagnetic responses.