Machine learning for improved image-based wavefront sensing

Area of Science:

• Optical engineering
• Machine learning applications
• Wavefront sensing

Background:

• Gradient-based optimization methods in image-based wavefront sensing struggle with convergence when initial wavefront estimates are poor.
• Large wavefront errors pose a significant challenge for traditional optimization techniques.

Purpose of the Study:

• To develop a machine learning approach for generating accurate initial wavefront estimates.
• To improve the convergence and effectiveness of wavefront sensing algorithms.

Main Methods:

• Utilized a convolutional neural network (CNN) trained on point-spread function (PSF) data.
• Simulated detector noise to test the robustness of the CNN model.

Main Results:

• The trained CNN successfully provided good initial wavefront estimates.
• The machine learning approach demonstrated superior performance compared to using multiple random starting guesses for large wavefront errors.

Conclusions:

• Machine learning, specifically CNNs operating on PSFs, offers a robust solution for initial wavefront estimation.
• This method enhances the reliability of wavefront sensing, particularly in scenarios with significant wavefront aberrations and noise.