Deep learning wavefront sensing for fine phasing of segmented mirrors.

Area of Science:

• Optical Engineering
• Astronomy
• Artificial Intelligence

Background:

• Extra-large space telescopes utilize segmented primary mirrors, but their imaging quality is sensitive to phasing errors between segments.
• Current deep learning techniques are primarily used for coarse phasing and estimating piston errors, with limitations in efficiency and problem-solving.

Purpose of the Study:

• To introduce a novel deep Bidirectional Gated Recurrent Unit (Bi-GRU) neural network for high-precision fine phasing of segmented mirrors.
• To address the limitations of existing deep learning methods in achieving fine phasing for advanced optical systems.

Main Methods:

• Development and application of a deep Bi-GRU neural network for segmented mirror fine phasing.
• Incorporation of phasing errors (piston and tip-tilt), low-order aberrations, and practical considerations into the neural network model.
• Validation through simulations and real-world experiments.

Main Results:

• The Bi-GRU network demonstrates a simpler structure compared to Convolutional Neural Networks (CNNs) or Long Short-Term Memory (LSTM) networks.
• Effectively solves the gradient vanishing problem inherent in training recurrent neural networks with long-term dependencies.
• Achieves accurate and effective fine phasing of segmented mirrors, as confirmed by simulations and experiments.

Conclusions:

• Deep Bi-GRU neural networks offer a powerful and efficient solution for the fine phasing of segmented mirrors in large space telescopes.
• The proposed method enhances imaging quality by precisely correcting phasing errors, including piston and tip-tilt.
• This advancement has significant implications for the development of next-generation astronomical observatories.