19–24 May 2024
Music City Center
US/Central timezone

Injection optimization via reinforcement learning: from simulation to real-world application

TUPS11
21 May 2024, 16:00
2h
Blues (MCC Exhibit Hall A)

Blues

MCC Exhibit Hall A

Poster Presentation MC6.D13 Machine Learning Tuesday Poster Session

Speaker

Jan Hetzel (GSI Helmholtzzentrum für Schwerionenforschung GmbH)

Description

This research presents a framework for the application of Reinforcement Learning (RL) to optimize the injection process at particle accelerator facilities. By utilizing a tailored and enhanced RL agent, we demonstrate its capability to dynamically optimize the beam's cross-section to meet predefined targets effectively at the Cooler Synchrotron COSY facility in Jülich, Germany. The agent, trained exclusively in a simulated environment, successfully applied its learned strategies during live operations, achieving optimization accuracy comparable to that of a human operator but in a notable less time. An empirical analysis of the architecture components—dense layers, observation noise, history, and domain randomization—demonstrates their individual and collective importance in preparing the agent for real-world applications. The findings highlight the potential of RL to enhance the efficiency of operations in particle accelerators.

Region represented Europe
Paper preparation format LaTeX

Primary author

Awal Awal (GSI Helmholtzzentrum für Schwerionenforschung GmbH)

Co-authors

Jan Hetzel (GSI Helmholtzzentrum für Schwerionenforschung GmbH) Jörg Pretz (Rheinisch-Westfälische Technische Hochschule)

Presentation materials

There are no materials yet.