Broadband high-purity mode converter based on diffractive neural networks

Mode-division multiplexing (MDM) technology has supported breakthroughs in ultra-high-capacity optical interconnects and optical communication systems. In wavelength and mode-hybrid multiplexing systems, broadband mode converters are crucial for optical chips, fibers, and free-space optical communication (FSOC) systems. Mode converters based on multi-plane cascaded diffractive neural networks (DNN) have attracted widespread attention due to their high degrees of freedom. However, they typically only operate within narrow bandwidths, and even when multi-wavelength optimization is employed, it is challenging to ensure the purity and consistency of the output modes across the entire bandwidth. Here, we propose an optimization scheme for achieving broadband, high-purity mode conversion. Compared to existing approaches, our method demonstrates improved consistency across the designed bandwidth. Using this scheme, we realize mode conversion between seven perfect vortex beams and Gaussian beams within a 30 nm bandwidth centered at 1550 nm. The average fidelity of the output light field on the right and left sides is 0.987 and 0.971, respectively, with minimum fidelities of 0.958 and 0.980.