Analysis of Alignment Errors Based on the Diffraction Pattern of Laser Radiation on Micro-Axicons Using a Convolutional Neural Network

This work presents the development of software for analyzing data obtained from simulations of laser radiation diffraction on micro-axicons, considering alignment errors such as displacement and tilt of the optical element relative to the incident beam. When modeling the axicon as a thin optical element, the tilt was simulated by introducing wavefront aberrations into the incident beam, specifically astigmatism and coma. A convolutional neural network was developed and trained to classify the resulting diffraction patterns, demonstrating high accuracy in identifying the type of aberration. The trained model achieved recognition accuracies of 100, 97.9, and 95.9% in three separate experiments. These experiments differed in the type of input data provided to the neural network: only intensity distributions, intensity combined with phase information, and intensity with defocused images, respectively. The results demonstrate that incorporating phase or additional spatial information significantly improves classification accuracy compared to intensity-only data. Even in the least accurate case (95.9%), the model shows strong generalization ability, making the proposed approach suitable for practical applications in optical diagnostics.