Researchers tested a new, AI-based way to measure how often the Higgs boson decays to quarks and gluons. The method — called the holistic approach because it uses the full list of reconstructed particles from each collision — was applied in simulations of the Circular Electron–Positron Collider (CEPC) acting as a Higgs factory. For two clean channels, Z(μ+μ−)H and Z(νν)H, the team projects relative statistical uncertainties for the main hadronic decay modes that range roughly from 0.36% to 5.21% and 0.16% to 2.52%, respectively. They report a two- to fourfold improvement in precision compared with earlier CEPC estimates.

What the team did is a detailed simulation study. They assumed the CEPC running at 240 GeV center-of-mass energy and an integrated luminosity of 21.6 ab−1. Signal samples included Higgs decays to b b̄, c c̄, gluon pairs (gg), and four-quark final states from WW* and ZZ*. Standard model background processes such as e+e−→ZZ, WW and q q̄ were also simulated. Event generation used MadGraph5_aMC@NLO at leading order and Pythia8 for parton showering and hadronization. AURORA, a high-granularity detector concept with an approximate boson mass resolution of 2.7–3%, was modeled with a fast simulation tool (Delphes).

At a high level, the holistic approach treats each collision as a cloud of reconstructed particles and feeds many particle-level features into a deep neural network called ParticleNet. ParticleNet’s core operation, EdgeConv, gathers information from each particle and its nearby neighbors and learns patterns by message passing. Input features to the network included angular coordinates, transverse momentum and energy fractions, track impact parameters (useful to spot b and c decays), and a particle-type label (for example muon, electron, photon, charged pion, kaon, proton, neutral hadron). The network was trained in a multi-class setup so it directly outputs the probability that an event came from each Higgs decay or a background process. For the muon channel the reported training sample was 1,000,000 events, with separate validation and independent test sets; the neutrino channel used between 100,000 and 1,000,000 training events.