Boosting Sensitivity to $HH o bar{b} γγ$ with Graph Neural Networks and XGBoost

In this paper, we explore the use of advanced machine learning (ML) techniques to enhance the sensitivity of double Higgs boson searches in the ( HH \to b\bar{b}γγ) decay channel at $\sqrt{s} = $ 13.6 TeV. Two ML models are implemented and compared: a tree-based classifier using XGBoost, and a geometrical-based graph neural network classifier (GNN). We show that the geometrical model outperform the traditional XGBoost classifier improving the expected 95% CL upper limit on the double Higgs boson production cross-section by 28%. Our results are compared to the latest ATLAS experiment results, showing significant improvement of both upper limit and Higgs boson self-coupling ($κ_λ$) constraints.

💡 Research Summary

This paper investigates the potential of state‑of‑the‑art machine‑learning techniques to improve the sensitivity of double‑Higgs boson searches in the HH→b b̄ γγ decay channel at the LHC operating at √s = 13.6 TeV with an integrated luminosity of 168 fb⁻¹ (representative of the ATLAS Run‑3 data set). The authors compare two classifiers: a conventional gradient‑boosted decision‑tree model (XGBoost) that uses a large set of high‑level kinematic variables, and a graph‑neural‑network (GNN) that explicitly encodes the geometric and topological relationships among the four reconstructed physics objects (two photons and two b‑tagged jets).

The theoretical background outlines the dominant gluon‑gluon‑fusion (ggF) and sub‑dominant vector‑boson‑fusion (VBF) production mechanisms for Higgs‑pair production, emphasizing the dependence of the ggF cross‑section on the trilinear Higgs self‑coupling λ₃, parametrised as κ_λ = λ₃^BSM/λ₃^SM. Quadratic parameterisations (Eqs. 2 and 3) are provided for both ggF and VBF as functions of κ_λ.

Monte‑Carlo event generation is performed with Powheg‑Box (NLO) for ggF HH, single‑Higgs processes, and tt̄H, while MadGraph5_aMC@NLO (LO) is used for VBF HH and the γγ+jets continuum background. Parton showering, hadronisation, and underlying‑event modelling are handled by Pythia 8.186, and a fast detector simulation is carried out with Delphes using an ATLAS Run‑3 configuration (updated tracking, calorimeter resolutions, b‑tagging efficiencies, etc.).

Object reconstruction follows ATLAS Run‑3 recommendations: photons with p_T > 20 GeV and |η| < 2.37 (excluding the barrel‑endcap transition), jets built with anti‑k_T (R = 0.4) and p_T > 25 GeV, b‑tagging at 85 % efficiency (c‑jet mistag 0.17, light‑jet mistag 0.01), and lepton vetoes to suppress tt̄H contamination. Events are required to pass a di‑photon trigger (E_T > 35 GeV and > 25 GeV) and to satisfy invariant‑mass windows 105 GeV < m_γγ < 160 GeV and a loose m_bb selection. Up to six central jets are allowed, and VBF‑tagged forward jets are identified when present, though no dedicated VBF category is defined.

For the machine‑learning stage, the XGBoost classifier ingests ~30 engineered features (p_T, η, ΔR, invariant masses, angular variables, etc.) and is tuned via cross‑validation of learning rate, tree depth, and number of estimators. The GNN treats each photon and b‑jet as a node; edges are constructed based on ΔR, and a three‑layer message‑passing network with global pooling produces an event‑level embedding that is fed to a binary classifier. Training uses the Adam optimizer, binary cross‑entropy loss, and early stopping.

Performance is evaluated with ROC curves, AUC, and signal‑efficiency versus background‑rejection plots. The GNN achieves an AUC of 0.89 compared to 0.82 for XGBoost, corresponding to roughly a 28 % improvement in the expected significance at a fixed signal efficiency (~30 %). The GNN also shows greater robustness against limited background statistics and systematic variations (b‑tag efficiency, energy scale, etc.).

Statistical inference is performed with the CLs method (RooStats). The expected 95 % confidence‑level upper limit on the HH production cross‑section improves from 0.72 pb (XGBoost) to 0.52 pb (GNN), a 28 % reduction. Constraints on the self‑coupling parameter tighten accordingly, e.g. from –2.0 < κ_λ < 5.0 (ATLAS Run‑2) to roughly –0.8 < κ_λ < 2.5 in the GNN‑based analysis.

The authors conclude that graph‑based neural networks, by exploiting the full spatial and relational information of final‑state objects, provide a substantial gain over traditional high‑level feature classifiers in the HH→b b̄ γγ channel. They suggest that extending the approach to a dedicated VBF category, incorporating larger Run‑4/5 data sets, and applying similar graph techniques to other multi‑object searches (e.g., tt̄H, SUSY cascades) could further enhance the LHC’s sensitivity to Higgs self‑interactions and beyond‑Standard‑Model physics.