Deep-learning-enabled robust OAM mode demodulation for optical vortices in atmospheric turbulence

Vortex beams carrying orbital angular momentum (OAM) offer a promising approach to increase channel capacity in freespace optical (FSO) communication by enabling multiplexing through topological charge (TC) states. However, the presence of atmospheric turbulence introduces significant distortions, posing challenges to the accurate recognition of OAM. In this paper, we proposed ARNet, a deep-learning-based network combining channel attention-enhanced residual blocks with dynamic L2 regularization, to enhance the robustness of OAM identification under turbulence. Experiments demonstrate that ARNet achieves 95.52% recognition accuracy for OAM states ranging from −5 to +5 under strong turbulence (C_n² = 5×10⁻¹³ m^−2/3) conditions. The proposed method significantly outperforms traditional interference-based techniques, demonstrating enhanced resilience and potential for practical FSO system deployment in turbulent environments.