Diffractive optical neural network for dual-wavelength vectorial vortex beams de-/multiplexing

Vectorial vortex beams (VVBs) exhibit unique polarization and mode characteristics, positioning them as promising candidates for advanced photonic applications. However, the efficient de-/multiplexing of VVBs, particularly across multiple wavelengths, remains a significant challenge. To address this, we propose a complex amplitude-modulation metasurface-based diffractive optical neural network (DNN) for dual-wavelength vector mode de-/multiplexing. Our approach enables full-dimensional control of vectorial optical fields by implementing a global optimization framework, achieving high mode conversion efficiency and low cross talk across up to 40 channels. Additionally, we investigate the impact of initial values and layer counts on optimization performance, offering valuable insights for enhancing DNN efficiency. This work provides a simple, yet efficient, platform for further applications of VVBs.