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

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Wavefront Sensing

Background:

• Convolutional neural networks (CNNs) are used for Shack-Hartmann wavefront sensing, but their black-box nature limits interpretability and robustness.
• Challenges include low sampling rates and strong turbulence, which can affect model performance.

Purpose of the Study:

• To develop an interpretable and robust deep learning framework for Shack-Hartmann wavefront sensing.
• To explicitly map CNN-extracted features to physically meaningful optical parameters.
• To improve model generalization across various conditions.

Main Methods:

• Proposed an optics-prior-driven (OPD) framework integrating physical spot descriptors (centroid shifts, shape distortions) with CNN features.
• Developed a novel network architecture, OPD-Net, incorporating optical priors.
• Validated feature correspondence through spatial-response and correlation analyses.

Main Results:

• Demonstrated that latent CNN features directly correspond to physical parameters like centroid shifts and spot distortions.
• OPD-Net achieved comparable reconstruction accuracy to conventional CNNs.
• The proposed model showed superior generalization across varying turbulence strengths, microlens-array configurations, and sampling conditions.

Conclusions:

• Provided the first physics-grounded interpretation of deep learning in Shack-Hartmann wavefront sensing.
• Established OPD-Net as a robust and interpretable alternative to end-to-end CNN models.
• Highlighted the potential of physics-informed deep learning for optical sensing applications.