Rapid inverse design of multi-objective transparent absorbing metasurfaces based on deep neural networks

Transparent absorbing metasurfaces with tailored spectra are promising for radar shielding, electromagnetic radiation protection, and emerging communication fields. However, traditional methods remain inefficient, whereas current intelligent networks are constrained by producing fabrication-infeasible designs and by poor generalization to ideal objectives. Here, a deep-learning-based inverse design framework is proposed for multi-objective transparent absorbing metasurfaces, enabling target-driven microstructure design by mapping spectra to structures that determine absorption. The network integrates automated generation, instantaneous forward prediction, and adaptive optimization into a closed-loop framework, where real-time spectral prediction feedback iteratively refines the structure generation, achieving over 8000-fold efficiency improvement and high design agreement (>0.95) with the target. The continuous generative space of the variational autoencoder (VAE) and the global search of the covariance matrix adaptation evolution strategy (CMA-ES) synergistically resolve the one-to-many mapping, expanding the feasible design space and simplifying fabrication. Through the collaborative mechanism, metasurfaces are designed to meet multiple ideal targets, achieving over 90% absorptance in the single band and above 85% across broadband (>10 GHz), with tunable number, frequency, and efficiency of multiple absorption peaks. The fabricated absorbers exhibit >90% transmittance while achieving diverse absorption characteristics. Additionally, the network provides a general framework for designing metasurfaces based on targeted reflection or transmission spectral characteristics, with promising applicability to absorption tuning from the terahertz to infrared range.