Improving Lepton Identification in CLAS12 Using Machine Learning

In this talk, a new algorithm based on machine learning techniques is presented. It aims to improve the lepton identification of the CLAS12 analysis. The primary focus of this work is minimizing pion contamination in the identified lepton sample. This background has been observed in both experimental and simulated CLAS12 datasets. Two machine learning models, Boosted Decision Trees (BDT) and Multilayer Perceptron (MLP), were developed, evaluated, and validated. Each model was trained using two sets of variables, and rigorous validation was conducted with simulated and experimental data to ensure reliability. The results demonstrated the effectiveness of these models in mitigating background from misidentified charged pions. While retaining 90% of the true leptons, the models achieved a ten-fold reduction of the pion contamination. This approach has been successfully applied to the study of the J/ψ photoproduction near threshold, showing significant improvements in signal extraction.