Accurate broadband wavefront sensing for space telescopes via a compact neural network

Area of Science:

• Optical Engineering
• Machine Learning
• Astronomy

Background:

• Broadband imaging is crucial for space telescopes observing faint celestial targets.
• Chromatic dispersion and spectral incoherence in broadband light challenge accurate wavefront sensing.
• Traditional phase retrieval algorithms struggle with accuracy under broadband illumination.

Purpose of the Study:

• To develop and evaluate a deep learning-based solution for accurate broadband wavefront sensing.
• To assess the performance of a compact convolutional neural network, U-EFFNet, for this task.
• To compare U-EFFNet against classical and other deep learning methods.

Main Methods:

• Utilized a compact convolutional neural network, U-EFFNet, integrating U-Net and EFFNet architectures.
• Conducted extensive simulations across diverse aberration and noise conditions.
• Performed experimental validation using a broadband spectral range of 400-750 nm.

Main Results:

• U-EFFNet demonstrated superior performance compared to Misell Gerchberg-Saxton (MGS) and monochromatic methods.
• The network outperformed other deep learning models in comprehensive performance metrics.
• Experimental validation confirmed U-EFFNet's high reconstruction accuracy and robustness.

Conclusions:

• U-EFFNet offers an effective balance of local/global feature representation and a lightweight structure.
• The network shows significant promise for onboard wavefront sensing in spaceborne optical systems.
• Deep learning provides a robust solution for challenging broadband wavefront sensing applications.