Deep-learning jet flavor tagging for precision hadronic Higgs measurements at future $e^+e^-$ Higgs factories

TL;DR

Problem: precise measurement of hadronic Higgs decays to $bb$, $cc$, $ss$, and $gg$ at future $e^+e^-$ colliders. Approach: a two-stage jet-plus-event framework combining three jet-level deep-learning taggers (ParticleNet, ParT, MIParT) with an event-level XGBoost classifier for the $e^+e^- \to ZH$ topology in a CEPC-like setting. Key results: projected relative precisions on $\sigma(ZH)\times\mathrm{Br}(H\to X)$ are $0.18\%$ (bb), $1.07\%$ (cc), $0.52\%$ (gg), and $78\%$ (ss); cc and gg reach improvements of about 42% and 26% over CEPC benchmarks, and $H\to s\bar{s}$ is estimated at roughly $1.3\sigma$ in a single channel. Significance: demonstrates that deep-learning-based jet flavor tagging can enable high-precision Higgs flavor studies at future lepton colliders and provides a modular, interpretable framework applicable to other facilities and $Z$-decay channels.

Abstract

Precise measurements of Higgs decays into quarks and gluons are essential for probing the Yukawa couplings of the Higgs boson and testing the flavor structure of the Standard Model. We investigate the process $e^+e^- \to ZH$ at $\sqrt{s}=240~\mathrm{GeV}$ at a future $e^+e^-$ Higgs factory, taking the CEPC design as a benchmark. The analysis focuses on events with $Z\toν\barν$ and hadronic Higgs decays $H\to b\bar b$, $c\bar c$, $s\bar s$ and $gg$. Jet flavor is identified using state-of-the-art particle-level deep neural network taggers (ParticleNet, Particle Transformer and More-Interaction Particle Transformer), whose per-jet outputs are combined with global event observables in a two-stage analysis employing XGBoost classifiers to separate the four Higgs decay modes from the dominant two- and four-fermion Standard Model backgrounds. Assuming an integrated luminosity of $20\,\mathrm{ab}^{-1}$, we obtain projected relative precision on $σ(ZH)\times\mathrm{Br}(H\to X)$ of 0.18% for $X=b\bar b$, 1.07% for $c\bar c$, 0.52% for $gg$ and 78% for $s\bar s$. Compared with the CEPC published results, the precisions for $H\to c\bar c$ and $H\to gg$ are improved by about $42%$ and $26%$, respectively. For $H\to s\bar s$ we present a quantitative sensitivity estimation corresponding to a statistical significance of about $1.3σ$. These results highlight the potential of deep-learning-based jet flavor tagging for precision studies of Higgs decays at future $e^+e^-$ Higgs factories.

Figures (10)

• Figure 1: Representative Feynman diagrams for the signal and background processes: (a) $e^+e^- \to ZH$ with $Z \to \nu\bar{\nu}$ and $H \to q\bar{q}$; (b) the two-fermion background $e^+e^- \to q\bar{q}$; (c) the representative four-fermion background $e^+e^- \to \nu\bar{\nu}q\bar{q}$.
• Figure 2: Schematic overview of the analysis strategy. Events are first processed by the three jet-level taggers (PN, ParT, MIParT). Their per-jet flavor scores, together with global event observables, are then used as inputs to a set of XGBoost classifiers: three using a single tagger each (XGB_PN, XGB_ParT, XGB_MIParT) and one combined classifier (XGB_Combined) that combines all three taggers simultaneously.
• Figure 3: Jet identification confusion matrices for $\nu\bar{\nu}H$, $H\rightarrow q\bar{q}$ events from ParticleNet (a), ParT (b) and MIParT (c). Each entry gives the fraction of jets of true flavor $i$ classified as flavor $j$.
• Figure 4: PCA projection of model-predicted $s$-jet probabilities for the $H \to s\bar{s}$ channel. Each point represents one independently trained model of ParticleNet (circles), ParT (squares) or MIParT (diamonds). Points of the same color correspond to the same model architecture, while the spread within a cluster reflects different training initializations. Models of the same architecture form compact clusters, indicating good training stability, while the separation between clusters reflects inter-architecture diversity that can be exploited in an ensemble.
• Figure 5: Performance of the XGBoost event classifier in the strange-enriched region. Left: confusion matrix for the XGB_Combined model, demonstrating clear separation for $H\to b\bar{b}$ and $H\to c\bar{c}$, while some confusion remains between $H\to s\bar{s}$ and $H\to gg$ due to similar jet substructure. Right: distribution of the XGB_Combined output score, taking $H\to s\bar{s}$ as signal, with all samples normalized to $20~\text{ab}^{-1}$. Higher scores correspond to higher signal purity, and $4f$ processes constitute the dominant background.