Phase-diversity wavefront sensing enhanced by a Fourier-based neural network.

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning Applications

Background:

• Phase diversity wavefront sensing (PDWS) quantifies aberrations using intensity measurements and nonlinear optimization.
• Non-convexity in PDWS inverse problems can lead to local minima, hindering accurate wavefront retrieval.
• Effective initialization is crucial for avoiding local minima and improving accuracy in PDWS.

Purpose of the Study:

• To develop a novel neural network for improved wavefront aberration estimation in PDWS.
• To leverage low-frequency Fourier domain coefficients for accurate aberration prediction.
• To reduce the amount of simulation data required for training PDWS models.

Main Methods:

• A neural network was designed utilizing low-frequency coefficients from the Fourier domain.
• The network was trained on a significantly reduced dataset compared to existing methods.
• Performance was evaluated against established wavefront retrieval techniques.

Main Results:

• The proposed neural network method achieved superior accuracy in wavefront aberration retrieval.
• Training time was drastically reduced to 1.4 minutes.
• Root mean square (RMS) residual errors were consistently low, with 95% below 0.05λ.

Conclusions:

• The neural network approach offers a more robust and efficient solution for PDWS.
• This method significantly lowers the data requirements for training wavefront sensing models.
• The technique demonstrates high accuracy and reduced computational cost for aberration correction.