Metasurface-Based Wavelength-Multiplexed Diffractive Neural Networks for Multilabel Intelligent Vision

Diffractive neural networks (DNNs) offer an energy-efficient platform for optical artificial intelligence by exploiting the inherent parallelism of light propagation. However, most existing DNN architectures are limited to single-label recognition, restricting their applicability to complex real-world visual perception tasks. Here, we propose a metasurface-based wavelength-multiplexed diffractive neural network (WMDNN) capable of multilabel object recognition by simultaneously identifying the object category and color in a single optical forward pass. The system is implemented using cascaded titanium dioxide (TiO₂) metasurfaces based on the Pancharatnam-Berry phase principle. By spatially interleaving TiO₂ nanoblocks with distinct geometries within each optical neuron, we achieve independent and parallel phase modulation across the RGB wavelengths for intelligent recognition of trichromatic objects. The proposed architecture is validated on a joint category-color Fashion-MNIST data set, achieving numerical and experimental accuracies of 91.7% and 85.8%, respectively. This work paves the way for high-dimensional intelligent perception systems through wavelength multiplexing.