Large-Scale and High-Accuracy OAM Mode Recovery, Recognition, and Demodulation Based on Hybrid Transfer Networks Under Atmospheric Turbulence

Vortex beams transmitted through free space are highly susceptible to atmospheric turbulence, which significantly impair their performance in optical communication systems. To address these challenges, a deep-learning-based scheme for the recovery and recognition of large amount of orbital angular momentum (OAM) modes is proposed and experimentally demonstrated. Sixteen networks based on UNet structure, residual block and dense connection block are designed and evaluated on nearly one million OAM modes collected from 400-meter real-world turbulence experiments, demonstrating excellent performance in both OAM recovery and recognition. The results show that the hybrid network can achieve mode recovery and recognition accuracy exceeding 98% while minimizing resource consumption. Furthermore, by combining transfer learning with mode selection, the accuracy can be increased to nearly 100%. In OAM shift-keying free space optical communication systems, the pre-trained model was employed, achieving high-fidelity information transmission, thereby demonstrating excellent demodulation performance. The proposed CNN-based OAM recovery method paves a new way toward future long-distance and low-crosstalk wireless optical communication through atmospheric turbulence channels.