Autonomous Accelerator

Modern particle accelerators provide exceptional beams for new discoveries in science. The required flexibility, number of operation modes, and better performance in simultaneously more compact and more energy-efficient accelerators demand advanced control methods. One major challenge is the start-up of such accelerators, which requires frequent manual intervention. Low repetition rates, often only one acceleration event per second, lead to slow optimization rates, thus demanding expert knowledge. Although a complete autonomous accelerator seems far from being reachable, this project takes the first steps by bringing reinforcement learning to accelerator operation. Reinforcement learning yields a policy for every initial state taking the impact of the current action on the future into account, eventually replacing the need for manual intervention. This project focuses on the longitudinal bunch profile of two accelerators, ARES and FLUTE, located at DESY and KIT, respectively. These two test accelerators are dedicated to research and thus offer the unique opportunity of sufficient beam time to develop such a reinforcement learning algorithm. Furthermore, these similar accelerators allow detailed research about the transferability of such algorithms and the resulting control agents. The interdisciplinary project team in this collaboration, including accelerator physics, control theory and computer scientist, opens new approaches to on the path toward autonomous accelerators. Read more about Autonomous Accelerator »

Publications

1. Learning to Do or Learning While Doing: Reinforcement Learning and Bayesian Optimisation for Online Continuous Tuning: Jan Kaiser et al., arXiv, doi: 10.48550/arXiv.2306.03739
2. Learning-based Optimisation of Particle Accelerators Under Partial Observability Without Real-World Training: Jan Kaiser et al., Proceedings of the 39th International Conference on Machine Learning, doi: https://proceedings.mlr.press/v162/kaiser22a.html
3. Accelerating Linear Beam Dynamics Simulations for Machine Learning Applications: Oliver Stein et al., Proceedings of the 13th International Particle Accelerator Conference, doi: 10.18429/JACoW-IPAC2022-WEPOMS036
4. First Steps Toward an Autonomous Accelerator, a Common Project Between DESY and KIT: Annika Eichler et al., Proceedings of the 12th International Particle Accelerator Conference, doi: 10.18429/JACoW-IPAC2021-TUPAB298
5. Machine Learning Based Spatial Light Modulator Control for the Photoinjector Laser at FLUTE: Chenran Xu et al., Proceedings of the 12th International Particle Accelerator Conference, doi: 10.18429/JACoW-IPAC2021-WEPAB289

Team Members

	Annika Eichler	Scientific areas: accelerator
	Jan Kaiser	Scientific areas: accelerator
	Oliver Stein	Scientific areas: accelerator

Project Partner