Machine learning applied to monochromatic diffraction: direct and inverse problems.

Area of Science:

• Optics and Photonics
• Computational Physics
• Machine Learning Applications

Background:

• Diffraction patterns are crucial for understanding light-matter interactions.
• Predicting diffraction patterns and propagation distances is essential in optical metrology and imaging.
• Current methods may lack precision or require extensive computational resources.

Purpose of the Study:

• To train neural networks for reconstructing diffraction patterns from various obstacles.
• To develop a method for predicting propagation distances using diffraction data.
• To evaluate the performance of neural networks on both theoretical and experimental diffraction data.

Main Methods:

• Training a neural network on theoretical Fresnel diffraction data.
• Applying the trained network to reproduce experimental diffraction patterns for a half-plane, single slit, and circular aperture.
• Utilizing the network to predict propagation distances from diffraction patterns.

Main Results:

• Neural networks achieved high accuracy in reconstructing diffraction patterns, with errors near experimental background noise levels.
• Propagation distances were recovered with millimeter precision from both calculated and experimental data.
• The trained models demonstrated robust performance across different obstacle geometries.

Conclusions:

• Neural networks offer a powerful and precise tool for analyzing diffraction phenomena.
• This approach enables accurate prediction of diffraction patterns and propagation distances in optical systems.
• The method shows significant potential for applications in optical sensing and characterization.