Dual-plane wavefront sensing using a vision transformer.

Area of Science:

• Optics and Photonics
• Machine Learning
• Image Processing

Background:

• Deep learning enables direct estimation of Zernike coefficients from intensity measurements in wavefront sensing.
• Convolutional Neural Networks (CNNs) are the predominant deep learning models used for this task.
• Limitations exist in CNN performance, particularly with downsampled image data.

Purpose of the Study:

• To introduce and evaluate a dual-plane wavefront sensor utilizing a vision transformer (ViT) model.
• To compare the performance of the ViT-based sensor against a CNN-based approach.
• To assess the efficacy of ViT in handling downsampled image data for Zernike coefficient estimation.

Main Methods:

• Development of a dual-plane wavefront sensor architecture.
• Training a vision transformer model for wavefront sensing.
• Comparative analysis of ViT and CNN performance using experimental and simulation data.
• Evaluation of prediction accuracy for high-order Zernike coefficients.

Main Results:

• The vision transformer model demonstrated superior performance compared to the CNN.
• Outperformance was particularly evident when dealing with significantly downsampled image data.
• The ViT model showed enhanced accuracy in predicting high-order Zernike coefficients.

Conclusions:

• Vision transformers offer a powerful alternative to CNNs for image-based wavefront sensing.
• ViT-based wavefront sensing is particularly advantageous for applications with limited image resolution.
• This approach advances the field of optical metrology and adaptive optics.