Deep learning based approach for multi-functional fabrication-friendly metasurfaces

Metasurfaces, comprising arrays of ultrathin and planar nanostructures (termed "meta-atoms"), hold immense potential for high-performance optical devices, enabling the precise control of electromagnetic waves with subwavelength spatial accuracy. However, the design of meta-atom structures that satisfy multiple functional criteria and workability presents a formidable challenge that significantly increases the design complexity. To address this challenge, we developed an expedited process for constructing a versatile, fabrication-friendly meta-atom library. This process utilizes deep neural networks in conjunction with a meta-atom selector, which considers the practical fabrication limitations. To corroborate the effectiveness of our method, we successfully employed it to empirically validate a dual-band metasurface collimator utilizing intricate free-form meta-atoms.