Machine Learning Applications in Accelerator Physics: Multi-Objective Bayesian Optimization of the Korea-4GSR Injector Linac and Orbit Correction in the RAON LEBT

Prof. Chong Shik Park (Korea University, South Korea)

This talk presents two complementary studies on the application of machine learning to accelerator design and beam control. The first study focuses on the multi-objective design optimization of the Korea-4GSR injector linac using Bayesian optimization, where multiple competing objectives—such as beam emittance, energy spread, transmission efficiency, and beam stability—are simultaneously considered. This study directly couples Bayesian optimization with high-fidelity beam dynamics simulations to efficiently explore the design space and construct Pareto-optimal solutions, providing quantitative insight into trade-offs among key performance metrics. The second study addresses beam orbit correction in the RAON Low Energy Beam Transport (LEBT) section, where machine learning techniques are employed to improve control performance under realistic operating conditions. In this work, reinforcement learning-based strategies are developed to determine optimal steering magnet settings for minimizing beam centroid deviations along the beamline. Together, these studies demonstrate how machine learning can be effectively applied across both design and operation phases of accelerator systems, highlighting a pathway toward systematic multi-objective optimization and autonomous beam tuning in next-generation accelerator facilities.