Deep Learning–Based Surrogate Models and RL Control for Extreme Vacuum Environments in Fusion Research

Progress in nuclear fusion is tightly bound to the research of materials that can endure extreme irradiation environments. To evaluate these materials, the IFMIF-DONES facility is being developed as a high-power particle accelerator, with the MuVacAS prototype reproducing the vacuum conditions of its final beam-line segment. Precise regulation of argon pressure inside the ultra-high vacuum chamber is crucial in this setup. This work presents a fully data-driven approach for autonomous pressure control. A Deep Learning Surrogate Model, trained on operational data, emulates the dynamics of the argon injection system and replaces the prototype with a safe simulator. Within this environment, a Deep Reinforcement Learning agent learns a control policy that maintains pressure within strict limits despite disturbances.Then the validation is donde by the deployment of the agent in real prototype. The implementation relies exclusively on open-source tools, even the control system, ensuring transparency, reproducibility, and accessibility. The results demonstrate that combining surrogate modeling and reinforcement learning enables intelligent control strategies for next-generation fusion energy facilities.