Tuning of the ESS Drift Tube Linac using Machine Learning

Abstract: The European Spallation Source, currently under construction in Lund, Sweden, will be the world's brightest neutron source. It is driven by a linear accelerator designed to accelerate a beam of protons with 62.5 mA, 2.86 ms long pulses, working at 14 Hz. The final section of its normal-conducting front-end consists of a 39 m long drift tube linac divided into five tanks, designed to accelerate the proton beam from 3.6~MeV to 90 MeV. The high beam current and power impose challenges to the design and tuning of the machine. In order to keep the beam quality throughout the accelerator and beam losses at a minimum the radio frequency amplitude and phase within the accelerating components have to be set within 1% and 1 deg of the design values. One of the usual methods used to define the radio frequency set-point is called signature matching, which can be a challenging process, and new techniques to meet the growing complexity of accelerator facilities are highly desirable. Machine learning is a rapidly growing field which has found applications in a wide range of scenarios, accelerators being no exception, but the tuning of RF fields using machine learning has yet to be tried. This project explores the possibility of applying machine learning in this area of accelerator physics, comparing this novel technique with the established signature matching and introducing a new possibility for faster tuning using a different data structure. For this purpose, simulations of the first tank in the drift tube linac section of the ESS linac were used to produce large amounts of tuning data at varying setpoints for the machine. Data like this was then used with the signature matching and machine learning techniques to fit the necessary functions for the traditional technique and train artificial neural networks for the novel technique. Random machine errors were later introduced to test each method's generalized performance. This data was also restructured to allow for machine tuning in a single shot, while usually a parameter scan is necessary, and machine learning techniques were tried for tuning on this data. Machine learning has been found to perform well in comparison with the established method, with some select advantages inherent to machine learning. This faster RF tuning technique only possible with machine learning, is found to perform well, although not quite within the given 1% and 1 deg limitations. As the rest of the results of this project are all on simulated data for the ESS, a comparison with results of these techniques on real data collected from the Spallation Neutron Source in the USA was performed. Future improvements could include workarounds for faulty network inputs and further tuning of the networks used.