Machine Learning Techniques for Charged Particle Tracking at the ATLAS Experiment

The Large Hadron Collider (LHC) uses proton-proton collisions to probe the fundamental building blocks of matter. Each collision produces thousands of particles scattering away from the detector center at nearly the speed of light. Reconstructing the trajectories of particles is a crucial task in most physics analysis. However, due to the rise in the number of simultaneous proton-proton interactions at the High Luminosity LHC (HL-LHC), the current tracking techniques will be the dominant component in CPU requirements. This thesis proposes the extension of existing as well as the design of novel Machine Learning (ML) approaches for the tracking of particles in the ATLAS experiment. We propose to describe and extend the similarity search problem in particle tracking through Approximate Nearest Neighbors (ANNs). In this context, the distance between data points is redefined with a tracking aware metric learning model termed TrackNet. Additionally, ANNs and metric learning models are evaluated on the TrackML dataset and on the ATLAS Inner Tracker Phase II dataset. We propose the Dynamic Tracking Linkage (DTL) clustering algorithm to process the output of the TrackNet model and to retrieve the final particle trajectories. This tracking inspired algorithm encapsulates physics constraints in its pairwise distance as well as a trained classifier that acts as an automatic stopping criteria