Machine Learning in High-Energy Physics: Displaced Event Detection and Developments in ROOT/TMVA

Abstract: Many proposed extensions to the Standard Model of particle physics predict long-livedparticles, which can decay at a significant distance from the primary interaction point.Such events produce displaced vertices with distinct detector signatures when comparedto standard model processes. The Large Hadron Collider (LHC) operates at a collisionrate where it is not feasible to record all generated data—a problem that will be exac-erbated in the coming high-luminosity upgrade—necessitating an online trigger systemto decide which events to keep based on partial information. However, the trigger is notdirectly sensitive to signatures with displaced vertices from Long-lived particles (LLPs).Current LLP detection approaches require a computationally expensive reconstructionstep, or rely on auxiliary signatures such as energetic particles or missing energy. Animproved trigger sensitivity increases the reach of searches for extensions to the standardmodel.This thesis explores the possibility to apply machine learning methods directly on low-level tracking features, such as detector hits and hit-pairs to identify displaced high-massdecays while avoiding a full vertex and track reconstruction step.A dataset is developed where modelled displaced signatures from novel and knownphysics processes are mixed in a custom simulation environment, which models the in-ner detector of a general purpose particle detector. Two machine learning models areevaluated using the dataset: a multi-layer dense Artificial Neural Network (ANN), and aGraph Neural Network (GNN). Two case studies suggest that dense ANNs have difficultycapturing relational information in low-level data, while GNNs can feasibily discriminateheavy displaced decay signatures from a Standard Model background. Furthermore itwas found that GNNs can perform at a background rejection factor of 103and a signalefficiency of 20% in collision environments with moderate levels of pile-up interactions,i.e. low-energy particle collisions simultaneous with the primary hard scatter.Further work is required to integrate the approach into a trigger environment. Inparticular, detector material and measurement resolution effects should be included inthe simulation, which should be scaled to model the High-Luminosity Large HadronCollider (HL-LHC) with its more complicated geometry and its high levels of pile-up.In parallel, the machine learning landscape is quickly evolving and concentrating intolarge software frameworks with expanding scope, while the High-Energy Physics (HEP)community maintains its own set of tools and frameworks, one example being the Toolkitfor Multivariate Analysis (TMVA) which is part of the ROOT framework. This thesisdiscusses the long- and short-term evolution of these tools, both current trends and somerelations to parallel developments in Industry 4.0.