Efficient Gradient-Based Metasurface Optimization toward the Limits of Wavelength-Polarization Multiplexing

Polarization and wavelength multiplexing are the two widely employed techniques to improve capacity in metasurfaces. While previous studies have pushed the channel numbers of each technique to its individual limits, achieving simultaneous limits of both techniques still presents challenges. Furthermore, current multiplexing methods often suffer from computational inefficiencies, hindering their applicability in computationally intensive tasks. In this work, we introduce and experimentally validate a gradient-based optimization algorithm using deep neural network (DNN) to achieve the limits of polarization and wavelength multiplexing with high computational efficiency. By leveraging the computational efficiency of the DNN-based method, we further implement nine multiplexed channels (three wavelengths x three polarizations) for large-scale image recognition tasks with a total of 36 classes in the single-layer metasurface. The classification accuracy reaches 96% in simulations and 91.5% in experiments. Our work sets a new benchmark for high-capacity multiplexing with gradient-based inverse design for advanced optical elements.