Image-based wavefront correction using model-free reinforcement learning.

Area of Science:

• Astronomy and Astrophysics
• Optical Engineering
• Machine Learning

Background:

• Optical aberrations limit telescope resolution, preventing them from achieving their theoretical diffraction limit.
• Current methods for estimating and correcting these aberrations rely on physical models, which can be limited by inaccuracies.
• Accurate aberration correction is crucial for enhancing the performance of astronomical instruments.

Purpose of the Study:

• To develop a novel data-driven method for automatic estimation and correction of optical aberrations.
• To utilize model-free reinforcement learning for aberration compensation, bypassing limitations of physical models.
• To improve the precision and reliability of wavefront sensing and correction in optical systems.

Main Methods:

• Formulated aberration correction as a reinforcement learning problem.
• Trained a reinforcement learning agent using simulated phase diversity images.
• Employed model-free reinforcement learning for direct estimation and correction of optical aberrations.
• Utilized deformable mirrors in a closed-loop system for aberration compensation.

Main Results:

• The reinforcement learning method successfully learned an efficient control strategy for aberration correction.
• The approach demonstrated reliable performance under various realistic observational conditions.
• The method showed robustness to a wide range of noise levels in the acquired images.
• Achieved accurate estimation and correction of optical aberrations directly from focal plane images.

Conclusions:

• Data-driven reinforcement learning offers a powerful alternative to model-based approaches for optical aberration correction.
• The developed method enhances telescope performance by overcoming limitations of traditional wavefront sensing.
• This approach holds significant potential for improving astronomical imaging and other optical applications.