B0 -> K*0 tau+ tau- Decay: Using Machine Learning to Separate Signal from Background

TL;DR

This paper applies machine learning techniques, specifically Boosted Decision Trees and Neural Networks, to distinguish the rare B0 -> K*0 tau+ tau- decay signal from background in simulated data, aiding future particle physics analyses.

Contribution

It introduces the use of ML classifiers with feature engineering for rare decay signal-background separation in particle physics.

Findings

BDT achieved AUC of 0.912 and F1-score of 0.828

FCNN achieved AUC of 0.877 and F1-score of 0.799

Both models effectively separate signal from background in simulated data.

Abstract

This study investigates the rare decay B0 -> K*0 tau+ tau-, which is sensitive to potential violations of lepton flavor universality predicted by the Standard Model. A Monte Carlo simulated dataset containing both signal and the dominant background process B0 -> K*0 D+ D- was used to train and evaluate machine learning classifiers. After feature selection and parameter tuning, two supervised models -- Boosted Decision Trees (BDTs) and Fully Connected Neural Networks (FCNNs) -- were trained. Feature engineering was then applied to enhance classification performance. On the test set, the BDT achieved an AUC of 0.912 +/- 0.000 and an F1-score of 0.828 +/- 0.001, while the FCNN reached an AUC of 0.877 +/- 0.000 and an F1-score of 0.799 +/- 0.001. These results demonstrate that both models can robustly separate signal from background in rare decay searches, supporting their application in future LHCb analyses.