Meta-learning-based optical vector beam high-fidelity communication under high scattering

Abstract

While spatial structured light based free space optical communication provides high-bandwidth communication with broad application prospect, severe signal distortion caused by optical scattering from ambient microparticles in the atmosphere can lead to data degradation. A deep-learning-based adaptive demodulator has been demonstrated to resolve the information encoded in the severely distorted channel, but the high generalization ability for different scattering always requires prohibitive costs on data preparation and reiterative training. Here, we demonstrate a meta-learning-based auto-encoder demodulator, which learns from prior theoretical knowledge, and then training with only three realistic samples per class can rectify and recognize transmission distortion. By employing such a demodulator to hybrid vector beams, high fidelity communication can be established, and data costs are reduced when faced with different scattering channels. In a proof-of-principle experiment, an image with 256 gray values is transmitted under severe scattering with an error ratio of less than 0.05%. Our work opens the door to high-fidelity optical communication in random media environments.