Performance of the neural network-based prediction model in closed-loop adaptive optics

Area of Science:

• Optical Engineering
• Atmospheric Physics
• Artificial Intelligence

Background:

• Adaptive optics (AO) systems compensate for atmospheric turbulence but are limited by inherent time delays.
• Deformable mirror (DM) phase compensation lags behind actual atmospheric distortions, reducing correction performance.
• Feed-forward prediction of atmospheric turbulence is crucial for offsetting time delays and enhancing AO system bandwidth.

Purpose of the Study:

• To evaluate the practical application and performance improvement of a deep learning-based spatiotemporal prediction model in an actual AO system.
• To compare the correction performance of an AO system integrated with a prediction model against traditional closed-loop control methods.

Main Methods:

• Development and implementation of a deep learning-based spatiotemporal prediction model for atmospheric turbulence.
• Integration of the prediction model into a 3 km laser atmospheric transport AO system.
• Experimental testing and comparison with traditional closed-loop control techniques.

Main Results:

• The study reports the first successful test of a deep learning-based spatiotemporal prediction model in an operational AO system.
• The AO system incorporating the prediction model demonstrated superior correction performance compared to traditional closed-loop methods.
• The feed-forward prediction effectively mitigated the limitations imposed by inherent time delays in the AO system.

Conclusions:

• Deep learning-based spatiotemporal prediction offers a viable solution to overcome the inherent time delay limitations in AO systems.
• The successful deployment in a real-world AO system validates the practical effectiveness of this predictive approach.
• This technology holds significant potential for advancing AO system performance in various applications, including astronomical observations and laser communications.