Machine learning application for dark matter - background classification in JUNO experiment

Abstract

Jiangmen underground neutrino observatory (JUNO) has a potential to indirectly detect dark matter (DM), observing neutrino events from annihilations of DM trapped by the gravitational force in the solar core. Weakly interacting massive particle (WIMP) DM candidate with mass > 4 GeV has significant solar capture rate. In this work, we simulate JUNO neutrino events from the most dominated WIMP annihilation channel, . Given the high-energy neutrinos from massive WIMPs, we extend the neutrino-nucleon interactions in the detector to include quasi-elastic (QE) and deep inelastic scatterings (DIS). The most challenging background events in the energy range above 100 MeV is the atmospheric neutrinos (atmo). The pulse shape discrimination (PSD) method is usually applied to distinguish between DM and atmo. In this work, we apply machine learning (ML) techniques to distinguish between the background (BG) atmo and DM neutrino events. We found that the random forest algorithm gives the best results. Using ML, the accuracy of DM-atmo events classification is 93.8% while similar f1-score could be achieved. On the other hand, only 78.1 % accuracy is obtained using conventional PSD with linear regression (LR). The upper limit for spin-dependent DM-proton cross-section with the applied ML is 5.84 × 10−40 cm2 for a 10 GeV WIMP DM which is an improvement over the sensitivity of the LR method of by a factor of ∼ 2.8x.