Deep‐Learning Assisted Polarization Holograms

Abstract Multiplexing holography with metasurfaces using different degrees of freedom light has enabled recent applications in display and information processing. In terms polarization‐multiplexed holograms, the most general form is an arbitrary Jones matrix profile storing maximum amount information. It requires a relaxation to bianisotropic from conventional single‐layer implementation nanostructures, but it will also complicate both inverse design nanostructures hologram generation algorithm. Here, integrated neural network approach, being extended DeepCGH algorithm, developed obtain metasurface structural profiles directly independent holograms set polarizations another, maximally four co‐ cross‐polarization conversion channels. Such information‐driven approach enables designing complex polarization existing metamaterial library without detailed physical knowledge on constraints, can be other multiplexing further facilitate efficient usage stored metasurface.