Machine learning based searches for new physics at the ATLAS experiment

A primary goal of the ATLAS experiment at the LHC is to discover new physics. In recent years, however, such discoveries have been scarce, creating a need for more sophisticated analysis techniques to probe for new physics in more intricate ways. This thesis focuses on two searches for new physics using machine learning techniques. Firstly, a search for electroweak supersymmetry at ATLAS, using the full Run-2 dataset of 139 fb¯¹ at √s = 13 TeV, is presented. This investigated a particularly challenging region of parameter space where there is a low mass-splitting between the supersymmetric particles considered – the lightest chargino and the lightest neutralino. Using a boosted decision tree to perform multiclass classification, separate regions in phase space can be defined that are enriched in either signal or a certain background. These regions are used to search for the supersymmetric signals and for improving the background modelling, respectively. Exclusion limits are set on the masses of the charginos and neutralinos, which cover an important gap in sensitivity between previous searches. Secondly, a novel method for performing model-independent searches for parity-violating new physics is presented. This analysis method uses convolutional neural networks which are parity-odd by construction. Asymmetries in the parity-odd output of the convolutional neural network indicates that there is parity-violation in the dataset. The efficacy of this method is demonstrated using a simplified model of parity-violating physics. Future searches using this method on data at the LHC allows for the investigation of previously unexplored forms of parity-violating physics at the energy scales of the LHC.