Improving Photonuclear and Kaon Background Veto Efficiency Using Boosted Decision Trees in LDMX

The Light Dark Matter Experiment (LDMX) is an electron-based, fixed-target, missing momentum experiment aimed at detecting sub-GeV dark matter. Currently, we are focusing on the rare but challenging photonuclear background, especially the Kaon subset. To address this, a machine learning algorithm, the boosted decision tree (BDT), was developed to improve veto efficiency for both photonuclear backgrounds and the kaon subset while maintaining high signal efficiency. The current BDT model incorporates segmented electromagnetic calorimeter data with MIP tracking information, as well as radius of containment variables that characterize the shower development. By using the BDT and hadronic calorimeter veto, we successfully reject most photonuclear background while preserving a high signal efficiency. For the kaon sample, additional cuts on MIP tracking variables further reduce this subset of background. A comparison with the previous BDT model, which lacked segmentation and MIP tracking, demonstrates the improvement in rejecting both photonuclear and kaon backgrounds with the updated BDT. The BDT plays a critical role in separating signal from background, and future enhancements could further improve its performance, allowing LDMX to better explore low-mass dark matter scenarios.