Advancing LEIR Automation: From Commissioning Tools to Reinforcement Learning-Based Continuous Control

by Borja Rodriguez Mateos (Universitat Politecnica Catalunya (ES))

Friday 4 Sept 2026, 16:00 → 17:00 Europe/Zurich

774/R-013 (CERN)

Abstract:

The Low Energy Ion Ring (LEIR) turns long lead-ion pulses from Linac3 into the dense bunches required by the LHC heavy-ion program, accumulating successive multi-turn injections under electron cooling. The accumulated intensity depends sensitively on tens of coupled machine settings whose optimum drifts continuously, as the stripper foil ages, and abruptly, whenever the injectors are recommissioned. This talk follows a multi-year program using machine learning to assist operations with this tuning burden. High-dimensional Bayesian optimization automates commissioning across some forty parameters and outperforms the intensities reached in routine operation. A reinforcement learning agent, deployed on the operational machine, then holds the machine at its optimum, acting on the nine settings that steer the accumulation and compensating foil aging cycle by cycle: it observes the machine through a learned compression of Schottky spectrograms, pretrains on data-driven digital twins, re-locks to a new campaign in one to two hours of beam time, and explores under a learned vacuum-risk constraint. I will close with first-principles simulations of the accumulation process and how they could complement the data-driven pipeline.

 

BE_seminar_LEIR.pdf