Speakers
Description
Radio-frequency (RF) bunch-merging gymnastics is used in the RHIC heavy-ion program to combine individual source pulses into single bunches with suitable intensity. To preserve both intensity and emittance during these gymnastics, the voltages and phases of RF cavities at several harmonic numbers must be carefully coordinated, which is labor-intensive and fragile. Recent work using a physics-based simulator of the Brookhaven Alternating Gradient Synchrotron (AGS) has shown that reinforcement learning (RL) can learn effective merge configurations. However, RL is highly data-intensive and requires many training interactions with the environment. Recent advances in large language models (LLMs) have demonstrated their capability of extracting patterns from large, noisy data. In addition, LLM is able to integrate domain knowledge in the control loop to improve sample efficiency and improve robustness. Therefore, it is an attractive solution for tuning complex accelerator systems. However, domain adaptation (i.e., prompt engineering, finetuning, etc.) is always required for deploying LLM in the target domain and has not been investigated in particle accelerators. To fill this gap, we propose an active supervision framework in which the LLM-based teacher first transfers general control principles from human operators to the student agent. Then, the student agent further finetunes the control policy by interacting with the simulator/experiments with improved sample efficiency.
Funding Agency
DE-SC0024617
| In which format do you inted to submit your paper? | LaTeX |
|---|