Frequency-domain filtering-based neural network for Shack-Hartmann super-resolution wavefront reconstruction

Area of Science:

• Astronomy and Astrophysics
• Optical Engineering
• Computer Science - Machine Learning

Background:

• Next-generation telescopes require higher-resolution wavefront sensing for advanced adaptive optics.
• The Shack-Hartmann wavefront sensor (SHWFS) is crucial for adaptive optics but limited by sub-aperture density.
• Existing methods struggle with aliased data from under-sampled wavefront sensors.

Purpose of the Study:

• To introduce a novel deep learning approach for super-resolution wavefront reconstruction (SRWR).
• To overcome the sub-aperture density limitations of traditional Shack-Hartmann wavefront sensors.
• To enable high-precision wavefront sensing for large-aperture telescopes.

Main Methods:

• Developed a frequency-domain filter-based neural network (FF-Net) for SRWR.
• Utilized learnable Gabor filters inspired by physical imaging principles and the convolution theorem.
• Employed physics-informed network design for feature extraction from aliased sub-aperture spots.

Main Results:

• FF-Net achieved state-of-the-art SRWR performance in numerical simulations.
• Outperformed conventional methods in accuracy with an under-sampled SHWFS, even compared to denser sensors.
• Successfully reconstructed higher-order aberration modes.
• Demonstrated GPU-accelerated inference time under 1 ms, meeting real-time adaptive optics requirements.

Conclusions:

• FF-Net offers a powerful strategy for improving wavefront sensing resolution in adaptive optics.
• Physics-informed deep learning provides accurate, robust, and interpretable solutions for optical systems.
• The method is suitable for real-time applications in astronomical adaptive optics.