An advanced feedback system based on reinforcement learning for autonomous accelerators

by Mr Luca Scomparin (KIT - Karlsruher Institut für Technologie)

Tuesday 16 May 2023, 11:00 → 13:00 Europe/Rome

M. Ceolin Meeting room

Modern particle accelerators aim to reach unprecedented luminosity or brilliance, which is possible only through stringent beam properties. This increases the complexity of the beam dynamics and accelerator operation, motivating the need of novel control methods.

The Karlsruhe institute of technology is world leading in the development of ultra-fast detector and feedback systems for accelerators, which allow a turn-by-turn and bunch-by-bunch analysis of the beam dynamics. These platforms, paired with Reinforcement Learning methods that can provide control policies by interacting with the accelerator environment, open the door to unparalleled on-line beam control.

Some of the control problems of interest, though, present latency and throughput requirements beyond the possibilities of currently available commercial platforms.

In this contribution, the KINGFISHER framework, based on the modern AMD-Xilinx Versal platform, will be presented. This new family of devices allows, through a heterogeneous computing architecture, to achieve unprecedented Machine Learning computation capabilities both from the latency and throughput point of view.

The integration of this platform with the advanced detector systems at the KARA storage ring will be described and the first closed loop feedback, which targets the control of horizontal betatron oscillations, will be discussed.