Long Short-Term Memory Recurrent Neural Network for the Fully-Automatic Collimator Beam-Based Alignment in the Large Hadron Collider (LHC)

The Large Hadron Collider (LHC) at the European Organization for Nuclear Research (CERN) is the world’s largest particle accelerator. Due to the high energy and high luminosity that the LHC can reach, a complex beam collimation system, comprising some 100 collimators is required to clean beam losses before they damage sensitive equipment. Collimators consist of two movable jaws made of robust materials whose purpose is to dispose of the halo around the proton or ion beams, to prevent quenching of superconducting magnets and protect the machine itself against damage. These jaws are positioned around the beam following well established automated beam-based alignment (BBA) techniques. The latter, use time series of beam losses detected by Beam Loss Monitors (BLM) to classify whether a jaw is aligned or not. The aim of this thesis is to analyze all aspect underlying the collimators automatic alignment based on machine learning, analyze the data gathered during previous alignment campaigns, determine the possibility of speeding up the automatic collimators alignment and propose a new machine learning model that can accomplish this goal. Since simple neural networks are not suitable for solving sequence problems, a Long Short-Term Memory Recurrent Neural Network has been designed and tested, obtaining an excellent precision on proton data. The new machine learning model developed has shown to be more efficient than the previous one whilst maintaining the high accuracy needed for such sophisticated collimation system as that of the LHC. The work described in this dissertation is planned to be tested in accelerator operating conditions in Run 3 starting in 2022. In addition, the currently used supervised learning models have been cross-checked with heavy ions beams, which have different loss patterns, to study the possibility to apply machine learning also to these beams.