Research
Over the past few years, I have been involved in several exciting projects.
For example, the work of my former doctoral student Chenran Xu at KIT in collaboration with DESY led to the first detailed comparison between Bayesian optimisation and reinforcement learning algorithms in optimisation tasks, deployed in real particle accelerators.
This fruitful collaboration also created the first differentiable beam physics simulation code, Cheetah.
Cheetah has now been adopted by several international collaborators, and we hope to keep on expanding it to facilitate the use of machine learning tools in particle accelerators. This work has been hand picked by the Physical Review Accelerators and Beams editors for the 2024 highlights.
Another very interesting project I coordinated was the first deployment of online reinforcement learning in particle accelerators (not pre-trained at all!) running on hardware at very low latencies for the control of the microbunching instability. This was achieved thanks to a custom high-performance electronics platform develop by doctoral student Luca Scomparin.
This platform was benchmarked against a known accelerator control task and then applied to the challenging microbunching instability.
Finally, within the reinforcement learning for autonomous accelerators (RL4AA) international collaboration we work on zero shot/few shot reinforcement learning in particle accelerator control with advanced solutions like meta reinforcement learning and other model-based algorithms.