Physics-inspired autoencoder for inverse design of achromatic polarization-multiplexed varifocal metalenses

Abstract In this study, we propose an innovative artificial intelligence-driven approach for designing metasurface structures with light field modulation capabilities. By directly encoding the Rayleigh-Sommerfeld diffraction equations into the autoencoder architecture, our method establishes a physics-grounded mapping from structural parameters to optical responses. Leveraging this physics-embedded AI framework, we successfully developed a achromatic polarization-multiplexed varifocal metalenses that achieves 67.3% average focusing efficiency at 10 mm focal length under x-polarized incidence and 57.5% at 5 mm under y-polarized incidence across 0.8-0.95 THz. Comparative analysis with particle swarm optimization-designed achromatic polarization-multiplexed varifocal metalenses demonstrates the superior focusing performance and focal control characteristics of our methodology. Full-wave electromagnetic simulations validate the reliability of this physics-inspired autoencoder.