Modal decomposition of complex optical fields using convolutional neural networks.

Area of Science:

• Optical physics
• Machine learning applications
• Wavefront sensing

Background:

• Convolutional neural networks (CNNs) have been used for optical modal decomposition with intensity images.
• A key limitation is the inability to uniquely determine modal phases from intensity data alone.
• Modal phase information is vital for wavefront sensing, beam alignment, and mode matching.

Purpose of the Study:

• To develop a CNN-based modal decomposition technique that predicts complete modal phases.
• To utilize simulated heterodyne images, which contain complex amplitude and phase information.
• To overcome the phase ambiguity inherent in intensity-only modal decomposition methods.

Main Methods:

• Training a CNN using simulated heterodyne images of optical fields.
• Heterodyne imaging provides high-resolution transverse complex amplitude and phase profiles.
• The developed CNN processes this complex phase information for decomposition.

Main Results:

• The CNN successfully predicts complete modal phases, a novel capability for machine learning decomposition.
• The method demonstrates reduced sensitivity to beam centering compared to traditional algorithms.
• The network achieves higher average accuracy on simulated images than overlap integral and center-of-mass methods.

Conclusions:

• This work presents the first machine learning modal decomposition scheme leveraging complex phase information.
• The CNN approach offers improved accuracy and robustness for modal decomposition tasks.
• The technique has significant implications for advanced optical metrology and laser beam characterization.