Speaker
Description
This study investigates the enhancement of boosted top quark tagging at the ATLAS detector by integrating jet substructure features, captured through the Lund Jet Plane (LJP) and the LundNet Graph Neural Network, with b-tagging information from the GN3X transformer model. The study demonstrates that combining these orthogonal data sources improves background rejection compared to using individual taggers, revealing that the inclusion of b-tagging information (secondary vertices) is a dominant factor in improving discrimination power. Despite the challenges of mass sculpting, where the networks learn the top quark mass as a primary discriminator, the results provide a preliminary validation for using multi-tagger workflows to optimize signal purity for future analyses, such as charm quark direction reconstruction in $W \to cq$ decays