Matrix Diffractive Deep Neural Networks Merging Polarization into Meta-Devices

The all-optical diffractive deep neural networks (D²NNs) framework as a hardware platform is demonstrated to implement various advanced functional meta-devices with high parallelism and high processing speed. However, the design methodology merging trainable polarization modulation neurons into the D²NNs, which potentially possess higher integration and more task-loading capacity, is not yet fully explored. Here, the matrix diffractive deep neural networks (M-D²NNs) are proposed to deploy polarization-sensitive Jones matrix metasurfaces into the all-optical polarization multiplexing networks to perform sophisticated inference tasks as well as inverse designs for advanced functional meta-devices. Three polarization multiplexing meta-devices with advanced functionalities are implemented by the M-D²NNs, that is, high task-capacity integration classification, non-interleaved high-efficiency Jones matrix eight-channel regulation, and custom-polarization information cryptographic multiplexing. The M-D²NNs are demonstrated to provide a new strategy to merge polarization into electromagnetic and optical field modulators by Jones matrix metasurfaces, which may drive the evolution of all-optical networks toward multi-task integration and more advanced functional devices.