Exploring Boosted Decision Trees for an ATLAS Search for Dark Mesons

This project explores the usage of the machine learning method of boosted decision trees (BDTs) for the ATLAS search for dark mesons, which are a dark matter candidate. As this search tries to extract a tiny signal from a huge and very similar background using a set of sensitive kinematic variables, the question whether a multivariate method could improve the sensitivity has been studied. For this purpose a BDT model has been trained and tested on simulated signal and background samples. For evaluation receiver operating characteristic (ROC) curves and precision recall curves (PRC) have been studied. As an estimation of whether or not the sensitivity could be improved, selections on the BDT discriminant have been applied and evaluated using the resulting significance. This method has successfully been applied and the significances using one common selection for the whole parameter space mostly exceed the ones that are reached without the usage of BDTs. However, one common selection for the whole parameter space is not the ideal choice, as the signal distributions depend heavily on the free parameters of the theory.