Two-Stage Machine Learning Strategy for Scalar and Vector Leptoquark Discrimination at the LHC

Abstract

We present a machine learning framework for the characterization of leptoquark (LQ) signals at the Large Hadron Collider, focusing on the discrimination between scalar (SLQ) and vector (VLQ) hypotheses. The method is based on a two-stage inference pipeline that combines a classifier trained to separate Standard Model backgrounds from a mixed LQ signal with a second classifier designed to distinguish between SLQ and VLQ scenarios, using the signal yield inferred from the first-stage classifier to guide the corresponding scalar and vector mass hypotheses. A test statistic is constructed from the classifier outputs and interpreted using reference probability density functions. The approach is applied to realistic LHC final states with hadronically decaying tau leptons, multiple jets, and missing transverse momentum, and its performance is assessed using simulated pseudo-experiments. We show that the proposed strategy provides a robust and statistically consistent procedure to discriminate between SLQ and VLQ signals across a wide range of masses and signal stregths. We find that the spin identification power closely follows the discovery potential, demonstrating that determining the spin nature of a newly discovered LQ does not require substantially larger datasets than those needed for discovery itself.