Machine learning applied to monochromatic diffraction: direct and inverse problems

We train a neural network using diffraction patterns to reproduce experimental diffraction patterns for a half-plane, a single slit, and a circular aperture. The training is done over theoretical data calculated under the Fresnel diffraction framework and applied to both calculated and experimental diffraction patterns. Experiments were designed to predict diffraction patterns from given obstacles and propagation distances and, conversely, to predict diffraction distances from known profiles. The second part was successfully carried out for theoretically calculated profiles and experimental data. The results show that the trained neural networks have outstanding performance as they reconstruct diffraction patterns with an error comparable to background noise of a standard experimental setup and recover propagation distances with millimeter precision from collected data.