Single-shot wavefront sensing with deep neural networks for free-space optical communications

Area of Science:

• Optics and Photonics
• Machine Learning
• Wavefront Sensing

Background:

• Deep neural networks enable real-time wavefront sensing.
• Phase diversity methods typically require two images, limiting speed.
• Single-image methods often sacrifice accuracy.

Purpose of the Study:

• To demonstrate accurate wavefront retrieval using a single pupil plane intensity image.
• To train and evaluate deep learning models for wavefront sensing.
• To apply the method to free-space optical communications (FSOC).

Main Methods:

• Generated a dataset with low-order aberrations for training.
• Trained EfficientNet to regress Zernike polynomial coefficients from intensity patterns.
• Compared EfficientNet against ResNet-50 and Inception-V3.
• Validated models using experimental adaptive optics data.

Main Results:

• EfficientNet significantly outperformed ResNet-50 and Inception-V3.
• High accuracy wavefront retrieval was achieved using only one image.
• The method proved effective with experimental data.

Conclusions:

• Single-image wavefront sensing is feasible with deep learning.
• EfficientNet offers a superior approach for real-time wavefront reconstruction.
• This technique has potential for improving FSOC systems.