Berlin 2024 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
AKBP: Arbeitskreis Beschleunigerphysik
AKBP 3: Instrumentation and Beam Diagnostics I
AKBP 3.7: Talk
Monday, March 18, 2024, 16:30–16:45, E 020
Reinforcement Learning Techniques for Injection Control at the Cooler Synchrotron COSY — •Awal Awal1,2, Jan Hetzel2, and Jörg Pretz1,3 — 1RWTH Aachen University — 2GSI Helmholtzzentrum für Schwerionenforschung — 3Forschungszentrum Jülich
Machine learning, particularly Reinforcement Learning (RL), holds significant promise in enhancing operations and optimisation within particle accelerator facilities. This study explores the application of RL for optimising particle accelerators, with a focus on the injection process at the COSY facility in Jülich, Germany. We propose a general formulation for RL problem and utilise it to optimise the injection into the synchrotron by manipulating four quadrupoles and seven steerers in the last section of the Injection Beam Line IBL.
Our methodology employs a soft actor-critic agent with dense neural networks, adapted for continuous action spaces, and training it with domain randomization to handle a variety of complex environmental dynamics. This results in a robust policy capable of generalizing to new, unseen environments. The integration of modernized viewer and control systems enabled direct analysis and automated adjustments of the beam cross section based on the RL agent's decisions. We extend this study with an in-depth analysis of the different components of the proposed RL framework and their significance. The successful implementation of this technique demonstrates a proof of concept in automating and optimizing accelerator operations, presenting a leap towards more efficient and consistent particle accelerator performance.
Keywords: machine learning; reinforcement learning; optimization; cosy