Physics-enhanced neural network for phase retrieval from two diffraction patterns.

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning Applications

Background:

• Phase retrieval is crucial for characterizing complex wavefronts.
• Traditional methods often require multiple measurements or prior information.
• Diffraction pattern analysis offers a path for non-invasive wavefront reconstruction.

Purpose of the Study:

• To introduce a novel physics-enhanced Y-neural network for efficient phase retrieval.
• To enable accurate reconstruction of complex wavefronts from limited diffraction data.
• To demonstrate the network's capability without requiring constraints or a priori knowledge.

Main Methods:

• A two-input, one-output Y-neural network architecture was employed.
• A hybrid loss function combining root-mean-square error and Pearson correlation coefficient was used for optimization.
• Self-supervised training was facilitated by an angular spectrum method network.
• Wavefronts were reconstructed from two diffraction patterns.

Main Results:

• The Y-neural network successfully retrieved amplitudes and phases of complex wavefronts.
• Reconstructions were achieved for a USAF-1951 resolution target, a phase grating, and a biological sample (skeletal muscle cell).
• The method demonstrated fast reconstruction capabilities within 100 learning iterations.
• Performance was validated using metrics like root-mean-square error and normalized Pearson correlated coefficient.

Conclusions:

• The proposed physics-enhanced Y-net offers a powerful and efficient solution for complex wavefront phase retrieval.
• The method's ability to perform fast, unconstrained reconstructions from two diffraction patterns is a significant advancement.
• This approach holds promise for various applications in optics, microscopy, and material science where precise wavefront characterization is needed.