A Machine Learning approach to enhance the classification of dark matter signals

The Supersymmetric Model (SUSY) of particle physics explains the phenomena that the Standard Model (SM) fails to account for. SUSY, for example, puts forward Weakly Interacting Massive Particles (WIMPs) as the leading contender for the composition of dark matter. However, Time Projection Chamber (TPC) experiments fail to fully isolate WIMP signals due to background events—signals that mimic those emitted by the desired WIMP particles, especially in regards to nuclear-recoil signatures. Although conventional methods effectively reject the majority of background events, they fail to distinguish WIMPs in nuanced, specific cases. Therefore, this project proposes a new machine-learning pipeline - a calibrated soft-voting classifier that combines XGBoost for nonlinear discrimination, Random Forest for feature ranking, and a Support Vector Machine (SVM) for margin-based separation - to differentiate WIMP interactions from various types of background events such as cosmogenic neutrons (cosmic rays), surface events (radioactive contamination on detector boundaries), electronic recoils (residual gamma and beta radiation), and radiogenic neutrons (natural decay of detector materials). Resources from the XENON1T experiment were used to simulate 50,000 WIMP and 50,000 background events, where 70% were used for training and 30% for cross-validation and calibration diagnostic to test results. The ensemble model produced strong performance, with an accuracy of 0.847, F1 of 0.843, ROC-AUC of 0.918. These results serve as a basis for detector-systematic ranking, domain-shift quantification, probability-calibrated event scoring. Future extensions to the project can include testing on actual detector data.