Piston Error Automatic Correction for Segmented Mirrors via Deep Reinforcement Learning.

Area of Science:

• Optical Engineering
• Machine Learning
• Astronomy

Background:

• Supervised learning for segmented mirror co-phase error identification offers advantages like speed and low computational needs.
• However, accuracy is often limited by discrepancies between training models and real-world optical systems.

Purpose of the Study:

• To develop a high-precision, large-range automatic co-phasing method for segmented telescope optical systems.
• To address the limitations of existing supervised learning techniques by employing a deep reinforcement learning approach.

Main Methods:

• A mask was placed on the pupil plane of the segmented telescope optical system.
• Deep reinforcement learning, without prior system modeling, was utilized.
• The method leveraged the system's wide spectrum, point spread function, and modulation transfer function.

Main Results:

• A novel automatic co-phase method for piston error correction was proposed.
• The method demonstrated effectiveness in large-range, high-precision alignment.
• Parallel processing across multiple sub-mirrors was achieved.

Conclusions:

• The proposed deep reinforcement learning technique effectively corrects piston errors in segmented mirrors.
• This model-free approach offers a significant improvement over traditional supervised learning methods for optical system alignment.
• The method is suitable for practical applications requiring high accuracy and efficiency.