Deep learning for robust orbit control of Elettra 2.0 storage ring

Recent advancements in light source accelerators have significantly enhanced beam quality. These accelerators are continually upgraded to meet scientific demands. A key element of their design is the precise control of particle orbits and the correction of errors to maintain a stable, high-quality beam. To facilitate the automatic electron beam position correction, we propose a computational framework to control the beam based on the Elettra 2.0 storage ring lattice, a process commonly known as orbit control. Our method exclusively relies on beam position monitors to predict the correctors' magnetic field strength using a deep convolutional neural network. We aimed to train our system and controller deep learning models sequentially, adopting a transfer learning strategy to fine-tune the controller and bring the final orbit closer to reference orbit. Since Elettra 2.0 is still in the design and construction stage, the proposed method is evaluated entirely in simulation. Results demonstrated the effectiveness and efficiency of our approach in orbit control, showcasing robust generalizability to real-world applications. Notably, when applied to the Elettra 2.0 storage ring, our method achieved 7 additional percentage-points of suppression stability relative to the uncorrected orbit over the singular value decomposition method. The proposed computational framework can also be applied to other synchrotrons, thereby facilitating the use of machine learning in accelerators.