Intelligent wavefront correction via self-supervised predictive Zernike phase inversion network

Wavefront correction under dynamic atmospheric turbulence is essential for achieving high-speed and high-fidelity optical communication and imaging. However, the rapidly varying phase distortions caused by turbulence pose significant challenges to conventional adaptive optics systems and supervised learning-based compensation methods. In this paper, we propose a self-supervised predictive Zernike phase inversion network (SS-PZPIN) that integrates dual-intensity Fresnel-consistent inversion and temporal prediction within a physics-informed learning framework. The network reconstructs Zernike coefficients directly from paired intensity distributions and predicts their temporal evolution without relying on labeled phase data. This self-supervised learning paradigm effectively bridges optical propagation physics with data-driven prediction. Simulation results demonstrate that SS-PZPIN achieves over 25% improvement in phase correction accuracy under strong turbulence and maintains more than 100% higher fidelity than the traditional Gerchberg-Saxton algorithm under 2.5-5 ms control delays, with an inference time of 2.67 ms. Furthermore, the model exhibits strong generalization to different turbulence conditions and temporal variations. These results confirm that SS-PZPIN provides a scalable and interpretable approach for real-time adaptive optics and turbulence-resilient free-space optical communication systems.