Real-time inference of vortex beam modes in time-varying atmospheric turbulence via diffractive optical neural networks

Abstract Orbital angular momentum (OAM) multiplexing offers potentially infinite information capacity in optical wireless communication, yet its performance is significantly degraded by time-varying atmospheric turbulence (AT), which induces inter-mode crosstalk. Conventional electronic-based mitigation strategies are inherently constrained by latency bottlenecks for real-time operation. Here, we propose a diffractive optical neural network tailored to the transmission properties of vortex beams under a time-varying AT propagation model, enabling real-time recognition of OAM modes. Experimentally, the network successfully identifies ten distinct OAM modes under time-varying AT conditions with 90.10% accuracy at a turbulence intensity of D / r 0 ≈ 0.83. Crucially, the phase-only trained network demonstrates strong generalization across varying turbulence intensities, sustaining accuracy above 80% for untrained turbulence intensities up to D / r 0 ≈ 1.11 without retraining. These results establish a light-speed intelligent optical system capable of tracking dynamically evolving AT, thereby overcoming the inherent latency-accuracy trade-off and enabling rapid, accurate, and energy-efficient all-optical demodulation of OAM modes in real-world fluctuating environments.