Bogotá - Ciencias - Maestría en Ciencias - Física · 2026

Tagging Particles with the Lund Jet Plane and Machine Learning at the LHC

This thesis studies boosted-jet tagging using the JetClass public dataset as a high-statistics, fully labeled benchmark. Jets with 500 < pT < 1000 GeV are reclustered with the Cambridge–Aachen algorithm and declustered to build the primary Lund jet plane. Each splitting is encoded by physically motivated variables such as (∆R, kt ≃ pT,soft∆R, z) and used as graph input. A LundNet graph neural network is trained for binary signal-vs-QCD tasks with 106 signal and 106 background jets per task, without a jet-mass cut. The resulting Lund-plane distributions show the expected jet substructure. For hadronic W/Z jets, the classifier achieves QCD background rejection of O(102–103) at ϵsig ∼ 0.3–0.5, with mild pT dependence. Top-jet discrimination depends on the decay mode, with hadronic tops outperforming semileptonic tops at intermediate efficiencies. Higgs tagging shows a consistent hierarchy across decay channels, with multi-prong H → WW modes being the most distinctive. A four-class multiclass LundNet (trained with 2 × 106 jets per class) achieves correct classification rates of ∼ 83% (H → b¯b), 93% (top), 84% (W), and 71% (QCD). From per-class probabilities of the model, a H → b ¯b-optimized discriminant is constructed while preserving stable performance across the boosted interval. Overall, the Lund jet plane provides a compact, interpretable representation that graph networks exploit for boosted-object tagging.