(PD)²: physics-driven phase diversity neural network for wavefront sensing

Phase-diversity wavefront sensing reconstructs aberrations from focused and defocused images without dedicated sensors, but it remains a high-dimensional nonlinear inverse problem. Conventional iterative methods are sensitive to initialization, noise, and model mismatch, whereas purely data-driven methods depend heavily on labeled data and often suffer from simulation-to-real gaps. This work focuses on point-source phase-diversity wavefront sensing and proposes a physics-driven self-supervised neural-network-based solution that estimates wavefront parameters from focused and defocused observations using image consistency under the forward physical model, without requiring large amounts of real labels. Simulations and real-world experiments show improved accuracy and robustness over L-BFGS and a purely data-driven baseline under the evaluated settings; under the ideal clean simulation setting, the proposed method reduces the wavefront residual RMS to 0.0058λ and achieves a success rate of 98.40%.