Physics-informed deep learning framework for wavefront sensing via optical beam pattern analysis

We propose PIRNet, a lightweight, physics-informed deep learning framework for non-interferometric wavefront sensing in optical beam expansion systems. The network simultaneously estimates spherical aberration, coma, and astigmatism from single-shot beam intensity patterns. A large-scale dataset is generated by simulating vortex beam propagation through combined aberrations using the ABCD transfer matrix method. To ensure physical plausibility, we introduce a physics-consistency loss that reconstructs the beam pattern from the predicted coefficients via an optical propagation model and compares it with the input, dynamically balanced with the regression loss through a learnable uncertainty weighting mechanism. A staged training strategy stabilizes convergence by first focusing on regression before introducing the physical constraint. Comparative experiments demonstrate that PIRNet outperforms ResNet18, ResNet34, ResNet50, and Xception across multiple metrics under varying noise levels and cropping ratios. The integration of physical priors enhances both accuracy and generalization, positioning PIRNet as a promising approach for model-driven wavefront characterization in adaptive optics and free-space optical communication.