Machine learning applications for Hadron Colliders: LHC lifetime optimization

The Large Hadron Collider is a indescribably complicated system with numerous intertwined systems, each impacts in it’s own way the dynamics and stability of the protons. As such, building a model of the particle losses occurring withing the LHC is an extremely daunting task, but it would offer valuable insight into the inner workings of the machine, and could potentially be used to further optimize the working points of the systems. This project aims to characterize the beam losses during the ramp of energy and use Machine Learning to build models of the LHC capable of predicting the intensity lifetimes of the beams at injection energy given a set of input parameters, such as tunes, chromaticities, intensities, octupole currents and a number of other features. The main goal is to determine the optimal settings of these parameters which would help operators in the decision making process, when striving to improve the performance and physics of such colliders. Two input feature sets are defined, one encompasses as many lifetime relevant measurements as possible and allows us to train a model as information complete as possible. The other, more practical feature set, only covers parameters which the operators can control, the knobs of the machine. The models are trained on experimental data using a Gradient Boosted Decision Tree supervised learning algorithm. From the models and covariance calculations we are able to extract the most relevant features in terms of contribution to the beam lifetimes. These results would be useful in understanding the machine on a fill by fill basis and also shed light on the physics mechanisms behind the lifetime variations. To go further, a similar method will be used on simulation data. In parallel a numerical model has been developed to physically explain the results gleaned from the machine learning analysis. A similar method could be applied to train on a dataset produced by simulations. The aim is to then use this model instead of the extremely computationally expensive SixTrack simulations, thus allowing the gained CPU time to be spent elsewhere, leading to more thorough studies.