ZeMS-Net: Physics Constrained Computational Wavefront Sensing via Self-Supervised Zernike-Modulated Swin-Unet

Adaptive optics (AO) systems are utilized to mitigate atmospheric turbulence in free-space optical links; however, conventional implementations rely on hardware wavefront sensors (WFS) that introduce considerable optomechanical complexity. While computational wavefront sensing offers a sensor-less alternative, traditional iterative algorithms are often constrained by high latency and a tendency to stagnate at local minima. Deep learning provides high-speed inference, yet existing supervised methods typically require ground-truth data from the hardware sensors they aim to substitute. To address these challenges, we propose the Zernike-Modulated Swin-Unet (ZeMS-Net), a physics-constrained computational framework that decouples macroscopic low-order geometric aberrations from high-frequency scintillation via Feature-wise Linear Modulation (FiLM). By embedding a fully differentiable Angular Spectrum Method propagation model, ZeMS-Net establishes an in-situ self-supervised learning loop, utilizing diffraction principles to recover the latent phase without external labels. Simulations indicate that ZeMS-Net achieves a reconstruction fidelity comparable to a simulated hardware WFS, improving single-mode fiber coupling efficiency from 16.44% to 47.9% under stronger turbulence conditions (𝐶2 𝑛 = 5 × 10−14 m−2/3). Bench-top experiments further support its practical applicability, recovering the focal profile and increasing the Strehl Ratio from 0.38 to 0.70. This approach provides a low-latency, sensor-less solution with the potential for continuous in-situ self-calibration.