Optics-prior-driven network for wavefront sensing with downsampled Shack-Hartmann sensors

Convolutional neural networks (CNNs) have demonstrated significant potential in Shack-Hartmann wavefront sensing by directly recovering phase information from intensity spot patterns. However, the black-box nature of CNNs obscures the physical interpretability of the learned features, thereby limiting model robustness under conditions such as low sampling rates or strong turbulence. To address this, we propose an optics-prior-driven (OPD) framework that explicitly maps CNN-extracted features to physically meaningful spot descriptors, including centroid shifts and spot-shape distortions. Through spatial-response and correlation analyses, we verify that the latent features learned by CNNs correspond directly to these physical parameters. Building on this insight, we develop OPD-Net, which is believed to be a novel network architecture that integrates optical priors into the learning process. The proposed model achieves reconstruction accuracy comparable to conventional CNNs while exhibiting generalization across varying turbulence strengths, microlens-array configurations, and sampling conditions. This study provides the first physics-grounded interpretation of deep learning in Shack-Hartmann wavefront sensing and establishes a robust, interpretable alternative to conventional end-to-end CNN models.