Compact 32-channel metasurface hologram enabled by the UNet model

With unique advantages in parallel information manipulation, multi-channel metasurface holography exhibits extensive application prospects in optical information encryption, high-density data storage, and integrated photonic chips. However, most existing multi-channel metasurface holograms usually require a large footprint with more than 500 × 500 meta-atoms, accompanied by significant image quality degradation. To tackle these limitations, we propose a neural network-based strategy to achieve compact multi-channel metasurface holographic imaging. The metasurface (∼184 λ × 184 λ ), composed of 256 × 256 TiO 2 meta-atoms, is partitioned into 4 × 4 independently designed sub-arrays, enabling 16-channel holographic image storage within an ultracompact footprint. Benefiting from the polarization multiplexing capability of TiO 2 nanopillars, each sub-array encodes two independent image channels. To compensate for information degradation caused by the ultrasmall sub-array size (64 × 64-pixel) and fabrication imperfections, a UNet model is introduced for high-fidelity image reconstruction. Experimentally, both 16-channel spatially multiplexed and 32-channel polarization-spatially multiplexed meta-holograms are demonstrated. The reconstructed images achieve PSNRs exceeding 23.1 dB and fidelity above 96.7 % across the 550–700 nm spectral range, while maintaining ultralow crosstalk below −13 dB. This work provides a feasible route toward compact, high-capacity, and high-fidelity metasurface holographic systems.