Dark Matter signals at the Large Hadron Collider with Deep Learning

Abstract: While holding a firm position in popular culture and science fiction, Dark Matter (DM) is nonetheless a highly relevant topic at the forefront of modern particle physics. We study the applicability of characterizing DM particle candidates SUSY neutralino and sneutrino using Deep Learning (DL) methods. We focus on the monojet and mono-Z signatures and the emergence of missing transverse energy as the result of the undetectable DM candidates. Based on kinematic distributions of outgoing particles as input, a DM candidate classifier is built for each signature, along with a DM mass regressor. The DM candidate classifier obtained near perfect accuracy of 0.995 for the monojet, and 0.978 for mono-Z signature. The monojet and mono-Z mass regressors achieved a Mean Absolute Percentage Error (MAPE) of 17.9 % and 8.5 % respectively. Furthermore, we discuss both the shortcomings and simplification that our choice of model implied, as well as an interpretation of the results. Finally, we debate the prospects of DL in the discovery of new physics and it's use in experiments.