Intelligent Control Systems and Machine Learning Approaches for Particle Accelerators

Particle accelerators are used all around the world for fundamental physics research, medical diagnosis and industrial applications. These can be extremely complex machines, with control systems composed of thousands of sensors and actuators producing an enormous amount of data. By exploiting this data, it's possible to reach new levels of performance, improve the uptime of the accelerator and reduce the effort required to setup, control and maintain it. This thesis focuses on the application of Machine Learning and Deep Learning models to the field of particle accelerator control systems. Anomaly Detection is applied to the task of fault prediction, both using classical Machine Learning algorithms and Deep Learning time-series forecasting models. By discovering anomalies in the trends of the process variables we can predict the insurgence of fault conditions, or the breakage of a critical component, thus allowing to intervene in time to avoid it. Reinforcement Learning is applied to the task of beam emittance optimization, with the aim of training a model which is able to automatically tune the beam transport parameters online to reach the optimal beam dynamics. These methods are validated on the particle accelerators at the INFN National Laboratories of Legnaro, Italy, but can be extended to other facilities with similar challenges. Finally, we explore data-driven Dynamic Sampling strategies to optimize metrology plans for quality control in industrial manufacturing processes.