PhyOAM-Net: Physical prior embedded data-driven demodulation for OAM shift keying under atmospheric turbulence

• Proposes PhyOAM-Net, a physics-embedded dual-stage model for 24-bit/symbol OAM-SK demodulation under atmospheric turbulence. • Integrates optical propagation laws into U-Net via physical consistency loss, reducing reliance on massive training data. • High bit accuracy (>98 %) is achieved under random turbulence ( C n 2 = 1 × 10 − 16 ∼ 1 × 10 − 12 m − 2 / 3 ) with small samples. • Real-time demodulation is ensured with 2.585 ms latency on a commercial GPU. • Provides a promising turbulence-resistant solution for long-distance, high-capacity free-space optical communication. Orbital angular momentum (OAM) serves as a novel multiplexing dimension that enables high-capacity transmission in free-space optical communication (FSO) systems. However, severe atmospheric turbulence (AT) inevitably degrades the demodulation performance of OAM shift keying (OAM-SK) systems. This paper proposes a physics-embedded data-driven model, named PhyOAM-Net, to demodulate 24-bit/symbol OAM-SK signals under AT interference. The model adopts a two-stage architecture: a physics-embedded U-Net (PE-UNet) first restores the optical intensity images distorted by AT, followed by a residual neural network (ResNet) for OAM mode recognition. Experimental results demonstrate that in a 1000-meter channel with AT strength ( C n 2 ) fluctuating randomly between 1 × 10 − 16 ∼ 1 × 10 − 12 m − 2 / 3 , PhyOAM-Net significantly outperforms existing models in recognition accuracy within sparse sample scenarios and maintains robustness even under strong AT conditions. It provides a promising solution for turbulence-resistant demodulation in long-distance, high-capacity FSO systems.