Signal and background discrimination for two-electron events in LDMX using a Boosted Decision Tree

Abstract: The Light Dark Matter Experiment (LDMX) is a fixed target experiment that will search for dark matter, but is still in the development phase. An important aspect for the experiment is the discrimination of signal and background events. Here this signal and background discrimination is inspected using a machine learning technique called a Boosted Decision Tree (BDT). This is done using Monte Carlo simulations of signal and background events, where two electrons hit the target. The signal sample used contains events where one electron undergoes a signal interaction in the target creating a dark matter mediator of a mass of 10 MeV, while the second electron can have any possible interaction (except for a signal interaction). The background sample contains events where one electron loses at least 1.5 GeV of energy by radiating a photon due to a bremsstrahlung interaction. This photon then has a photo-nuclear reaction in the Electromagnetic Calorimeter (ECal), while the second electron can again have any possible interaction. To train the BDT a number of features based on the information from the ECal are used. The distribution of these features is shown for the signal and background samples. The samples are divided into subsets of data, the training data, test data and independent test data. A BDT is trained on the training data and tuned using the performance on the test data. To get a unbiased measure of the performance of the BDT the independent test data is used. The BDT has an Area Under Curve (AUC), for a Receiver Operating Characteristic (ROC) curve, of 0.998981 on the test data and an AUC of 0.998720 for the independent test data. For a background rejection of 99% the signal efficiency is 97% with a BDT threshold of 0.7.