Reinforcement Learning Approaches for Particle Accelerators: The PIAVE Case Study at the INFN National Laboratories of Legnaro

Particle accelerators are employed worldwide for fundamental physics research, medical diagnostics, and industrial applications. These systems are highly intricate, with control mechanisms consisting of thousands of sensors and actuators that generate vast amounts of data. The more parameters are available, the harder it is to converge to a working solution, with no guarantee of reaching the optimal one. Recently at the INFN (National Institute of Nuclear Physics) laboratories in Legnaro, Bayesian optimization approaches have been employed to automate the process. This has led to significant improvements in accelerator performance and a notable reduction in the time required for proper setup. Having already significantly improved accelerator performance, the results suggest that there is still potential for further enhancement. The objective of this thesis is to propose a machine learning approach to simulate the accelerator and then optimize it using reinforcement learning techniques to speed up the setup and improve the performance of the beam.