Transformer-based wavefront sensing for atmospheric turbulence aberration correction.

Area of Science:

• Optical Engineering
• Artificial Intelligence in Optics

Background:

• Traditional wavefront sensing methods face challenges in accurately estimating complex aberrations.
• Convolutional Neural Network (CNN)-based approaches have limitations in capturing global dependencies for wavefront sensing.

Purpose of the Study:

• To introduce a novel transformer-based wavefront sensing method utilizing cross-task pretraining.
• To enhance aberration estimation accuracy and model interpretability compared to existing methods.

Main Methods:

• Implementation of a transformer architecture for wavefront sensing.
• Application of a cross-task pretraining strategy to establish global dependencies.
• Analysis of attention maps for Zernike terms and their correlation with Point Spread Functions (PSFs).

Main Results:

• Significant improvement in aberration estimation accuracy, with a 70.5% reduction in test set loss and a 45.7% reduction in Root Mean Square (RMS) error.
• Attention maps demonstrated consistency with PSFs of individual aberrations, indicating effective decoupling of aberrations.
• The self-attention mechanism successfully captured complex physical relationships, enhancing model interpretability.

Conclusions:

• The proposed transformer-based method offers superior performance in wavefront sensing compared to CNN-based approaches.
• The self-attention mechanism facilitates the decoupling of individual aberrations and improves model interpretability.
• This novel method presents a unified approach with significant potential for advancing wavefront sensing technologies.