Performance of a U-Net-based neural network for predictive adaptive optics

Area of Science:

• Astronomy
• Artificial Intelligence
• Optical Engineering

Background:

• Predictive adaptive optics (AO) is crucial for imaging fast-moving celestial objects.
• Traditional AO systems face limitations in tracking dynamic targets like low Earth orbit (LEO) satellites.
• Developing robust AO systems for space debris tracking presents significant challenges.

Purpose of the Study:

• To apply a U-Net-based convolutional neural network (NN) for predictive adaptive optics.
• To enhance the tracking and imaging capabilities for fast-moving targets in LEO.
• To evaluate the NN's performance against non-predictive methods in real-world astronomical conditions.

Main Methods:

• Utilized a U-Net-based convolutional neural network architecture.
• Trained the NN predominantly on simulated data.
• Tested the NN using 1 kHz Shack-Hartmann wavefront sensor data from open-loop observations.
• Collected data at the Advanced Electro-Optical System facility at Haleakala Observatory.

Main Results:

• Achieved an approximate 50% reduction in mean-squared wavefront error compared to non-predictive approaches.
• Demonstrated the NN's ability to predict wavefronts up to eight frames into the future.
• Successfully validated the NN's performance on real-sky data of LEO space debris.
• Reported the first successful on-sky test of a NN for predictive AO in space tracking.

Conclusions:

• The U-Net-based NN shows significant promise for predictive adaptive optics in astronomical applications.
• This AI-driven approach offers a substantial improvement for tracking and imaging challenging targets like LEO satellites.
• The study represents a pioneering step in applying neural networks to demanding space-based optical tracking scenarios.