Deep learning control model for adaptive optics systems.

Area of Science:

• Optics
• Control Systems Engineering
• Artificial Intelligence

Background:

• Adaptive optics (AO) systems commonly use proportional-integral control, which relies on the deformable mirror's response matrix.
• Alignment errors between wavefront sensors and deformable mirrors in AO systems limit conventional control accuracy.
• Recalibrating the response matrix can mitigate, but not eliminate, the impact of alignment errors.

Purpose of the Study:

• To propose a novel deep learning control model (DLCM) for adaptive optics systems.
• To eliminate the dependence on the deformable mirror response matrix for wavefront aberration correction.
• To achieve online identification and self-adaptive control of AO systems.

Main Methods:

• Developed a deep learning control model (DLCM) utilizing wavefront slope data.
• Defined cost functions for model and actor networks, optimized using gradient algorithms.
• Implemented a parameter-sharing mechanism between networks for system gain control.

Main Results:

• The DLCM demonstrated strong adaptability and stability in simulations.
• The model network ensured system stability and convergence speed.
• The actor network enhanced control accuracy through online identification and self-adaptive capabilities.

Conclusions:

• The DLCM effectively compensates for wavefront aberrations, surpassing conventional methods.
• The proposed DLCM offers improved convergence accuracy, iteration efficiency, and adjustment tolerance.
• Deep learning provides a robust approach for advanced adaptive optics control.