Monte Carlo simulations of X-ray sources by machine learning approaches

Abstract

This master thesis introduces an approach to using Generative Adversarial Networks for the generation of phase space to replace the generated phase space instead of large phase space datasets. The original approach was produced by Monte Carlo method of ImagingRing system at MedAustron. This is intended to create the generated particles that can be used in research areas, while creating the conventional sampling of phase space is time consuming and challenging. To evaluate the outcome of GAN, some methods are proposed to validate the generated particles. The efficiency of the generated particles produced by GAN has been checked and satisfactory results have been gained for this research. As the main result, the particles in the energy range of 20 keV-60 keV were generated with the maximum statistical and theoretical significance. In addition, the superimposition of the original phase space and generated one can be obtained in this given range of the energy. This study shows that no particles were generated in the energy above 70 keV for X-Y parameters and the particles in the energy range between 60 keV-70 keV were generated with lower superimposition of the generated particle and original one due to lack of the reference particles. The generated particles by GAN requires only around 10 MB storage compared to the phase space produced by ImagingRingTM System which contains tens of Gigabyte data. Besides, the process of the particle generation is fast and it is efficient to use. Moreover, a novel research pathway in the statistical techniques for validation of the generated phase space has been opened so that further research can be developed.