Speaker
Description
Real-time optimization of free-electron lasers (FELs) is challenged by high-dimensional parameter sensitivity and beam dynamics drift. We present a deep reinforcement learning (DRL) framework that models FEL tuning as a Markov Decision Process, with multi-diagnostic state representations and safety-regularized rewards. Our approach pre-trains SAC and MAPPO agents in a Genesis simulator with domain randomization, followed by deployment via EPICS control systems. Experiments demonstrate that DRL agents achieve 20-50% higher photon flux than baseline within fewer iterations, reducing optimization time from hours to minutes in dynamic scenarios and outperforming manual tuning and Bayesian optimization, while maintaining reduced pulse energy variance and operational safety. Ablation studies validate the importance of fused photon-beam diagnostics and hybrid sim-to-real training for robust performance. This work provides the first systematic validation of DRL for live FEL optimization, offering an open-source benchmark for autonomous accelerator control.
| In which format do you inted to submit your paper? | LaTeX |
|---|