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

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning Applications

Background:

• Phase-diversity wavefront sensing is crucial for aberration correction but is a complex nonlinear inverse problem.
• Existing iterative methods struggle with noise and initialization, while data-driven approaches require extensive labeled datasets and face simulation-to-real gaps.

Purpose of the Study:

• To develop a physics-driven, self-supervised neural network for point-source phase-diversity wavefront sensing.
• To overcome limitations of conventional and purely data-driven methods by utilizing image consistency under a forward physical model.

Main Methods:

• A novel physics-driven, self-supervised neural network architecture was developed.
• The method leverages image consistency between focused and defocused observations based on the forward physical model.
• No large amounts of real-world labeled data are required for training.

Main Results:

• The proposed method demonstrates improved accuracy and robustness compared to L-BFGS and a purely data-driven baseline.
• In ideal clean simulations, the wavefront residual RMS was reduced to 0.0058λ.
• A high success rate of 98.40% was achieved under evaluated settings.

Conclusions:

• The physics-driven, self-supervised neural network offers a robust and accurate solution for phase-diversity wavefront sensing.
• This approach effectively addresses the challenges of high dimensionality, noise sensitivity, and data requirements in wavefront reconstruction.