Speaker
Description
This work represents a preliminary step toward developing an AI-driven computational framework for accelerator science. The effort integrates accelerator physics, advanced computation, machine learning, and control to support future diagnostics, optimization, and operation. Initial activities focus on AI-enabled digital twins that combine physics-based models with machine learning, archived and live control data, and beam diagnostics to improve prediction and support accelerator tuning. In parallel, we are developing a unified Gymnasium-based interface connecting Bayesian optimization through Xopt with reinforcement learning through SOCT, enabling comparison of optimization and control methods. The program also explores agentic AI approaches through Osprey, with emphasis on translating operator goals into structured machine tasks while maintaining safety checks, human approval, and secure interaction with facility systems. Together, these efforts provide a foundation for scalable AI-assisted accelerator operation.
| Working group | WG5 |
|---|