⚛️ high-energy experiments

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

This paper demonstrates that employing state-of-the-art deep-learning jet flavor taggers combined with XGBoost classifiers at a future $e^+e^-$ Higgs factory operating at 240 GeV with 20 ab $^{-1}$ luminosity significantly improves the precision of measuring Higgs decays to $c\bar{c}$ and $gg$ while enabling the first quantitative sensitivity estimate for the challenging $H\to s\bar{s}$ channel.

Original authors: Xinzhu Wang, Yifan Zhu, Chunxiang Zhu, Jianfeng Jiang, Manqi Ruan, Kun Wang, Haijun Yang, Yongfeng Zhu

Published 2026-07-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xinzhu Wang, Yifan Zhu, Chunxiang Zhu, Jianfeng Jiang, Manqi Ruan, Kun Wang, Haijun Yang, Yongfeng Zhu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the universe as a giant, cosmic puzzle where every piece of matter is held together by invisible forces. For decades, scientists have been trying to figure out exactly how these pieces fit together, a theory known as the Standard Model. At the center of this puzzle is a special particle called the Higgs boson, often nicknamed the "God particle" (though physicists prefer just "Higgs"). Think of the Higgs as a cosmic glue that gives other particles their mass. Without it, atoms wouldn't form, and we wouldn't be here.

While scientists have already found the Higgs and measured how it interacts with heavy particles, there's a tricky part of the puzzle still missing: how it talks to the lighter, "second-generation" particles, like the strange quark. It's like trying to hear a whisper in a hurricane; the signal is so faint and the background noise so loud that it's incredibly hard to tell if the Higgs is actually talking to these light particles at all. To solve this, scientists are planning to build massive, super-clean particle colliders in the future—essentially giant microscopes that smash electrons and positrons together to create Higgs particles in a controlled environment. The goal is to catch the Higgs in the act of decaying into different types of particles and see if it follows the rules of our current theories or if it's hiding some new physics.

This paper is a practice run for that future experiment. The authors are using powerful computer simulations to test a new way of "listening" to the Higgs. They focus on a specific scenario where an electron and a positron collide to create a Higgs boson and a Z boson (a heavy cousin of the photon). The Z boson disappears into the void (decaying into invisible neutrinos), leaving the Higgs to decay into a pair of jets—sprays of particles that look like tiny fireballs. The challenge is that these fireballs can come from different sources: heavy bottom quarks, charm quarks, strange quarks, or even gluons (the glue holding quarks together). Distinguishing a "strange" fireball from a "gluon" fireball is like trying to tell the difference between two identical-looking twins wearing the same clothes, but one of them is actually a spy.

To solve this, the team developed a two-stage detective system powered by artificial intelligence. First, they use three different types of advanced deep-learning "jet taggers" (named ParticleNet, Particle Transformer, and More-Interaction Particle Transformer). Think of these as three different expert detectives, each looking at the microscopic details of the particle fireballs to guess what kind of particle started the spray. One detective might look at the shape of the spray, another at the specific types of particles inside, and a third at how they interact with each other.

Once these three experts have made their guesses, the second stage kicks in. A master classifier (using a tool called XGBoost) takes all those guesses and combines them with the big-picture view of the entire collision event. It's like a head detective who listens to the three experts, looks at the crime scene, and makes the final call on what happened. The paper simulates this process with a massive amount of data—equivalent to 20 ab⁻¹ of collisions, which would produce about four million Higgs bosons.

The results are promising. For the heavy particles (bottom and charm quarks), the new method is incredibly precise, improving the measurement accuracy by about 43% for charm quarks and 29% for gluons compared to previous estimates. The most exciting part, however, is the attempt to find the "strange" quark signal. In these simulations, the team managed to spot the strange quark signal with a statistical significance of about 1.5σ. In the world of particle physics, this isn't a confirmed discovery yet (you usually need 5σ for that), but it's a strong hint that the signal is there, like seeing a shadow that suggests a person is standing behind a curtain. The authors show that by using these smart AI tools, future colliders like the CEPC (Circular Electron-Positron Collider) could finally start to map out how the Higgs interacts with these elusive light particles, potentially revealing cracks in our current understanding of the universe.

The paper also checks to make sure their AI isn't just making things up. They found that the system relies on real physical clues, like the specific "flavor" of the particles and the energy missing from the collision, rather than random noise. They even compared their two-step detective method against a "holistic" approach where one giant AI looks at the whole event at once. They found both methods work about the same in terms of raw numbers, but their two-step method is better for understanding why the AI made a decision and easier to calibrate if the real experiment behaves differently than the simulation.

In short, this paper doesn't discover a new particle or prove a new theory. Instead, it builds a better pair of glasses for the future. It shows that with the right combination of deep-learning tools and clever analysis, the next generation of particle colliders will be able to see the Higgs boson in much sharper detail, finally allowing scientists to listen to the whispers of the strange quark and test the limits of the Standard Model.

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

Problem Statement
Precision measurements of Higgs boson decays into quarks and gluons are critical for probing Yukawa couplings and testing the flavor structure of the Standard Model (SM). While decays to third-generation fermions ( $H \to b\bar{b}, \tau^+\tau^-$ ) and gauge bosons are well-established, interactions with second-generation fermions, particularly $H \to c\bar{c}$ and $H \to s\bar{s}$ , remain challenging. The primary obstacle is the overwhelming Quantum Chromodynamics (QCD) background in hadronic environments, where jets from light quarks ( $u, d, s$ ) and gluons exhibit similar substructures, making flavor discrimination difficult. Although deep learning has improved $b$ - and $c$ -tagging at the LHC, separating light-flavor jets (especially $s$ vs. $g$ ) remains a significant hurdle. This paper addresses these challenges in the context of future $e^+e^-$ Higgs factories, specifically the Circular Electron-Positron Collider (CEPC), focusing on the $e^+e^- \to ZH$ process with $Z \to \nu\bar{\nu}$ and hadronic Higgs decays.

Methodology
The authors propose a two-stage analysis framework designed to separate four Higgs decay modes ( $H \to b\bar{b}, c\bar{c}, s\bar{s}, gg$ ) from dominant two-fermion ( $2f$ ) and four-fermion ( $4f$ ) SM backgrounds.

Simulation and Setup:
- The analysis assumes a CEPC operating at $\sqrt{s} = 240$ GeV with an integrated luminosity of $20 \text{ ab}^{-1}$ .
- Signal and background events are generated using MadGraph 3.6.3 and Pythia 8.313, with detector effects simulated via Delphes 3 using the official CEPC detector card.
- Events are reconstructed into exactly two jets using the $e^+e^-$ - $k_t$ algorithm.
Stage 1: Jet-Level Flavor Tagging:
- Three state-of-the-art particle-level deep neural networks are employed to identify the flavor of individual jets: ParticleNet (PN), Particle Transformer (ParT), and More-Interaction Particle Transformer (MIParT).
- Inputs include particle-level kinematics, particle identification (PID) flags, and track trajectory information (impact parameters).
- The models are trained to classify jets into 11 truth labels ( $b, \bar{b}, c, \bar{c}, s, \bar{s}, u, \bar{u}, d, \bar{d}, g$ ).
- Three independent training runs per architecture are performed to assess stability.
Stage 2: Event-Level Classification:
- The per-jet outputs from the three taggers are combined with global event observables (kinematics of the dijet system, missing energy variables, and correlations between jets and missing energy).
- An XGBoost classifier is trained to perform a six-class classification ( $H \to b\bar{b}, c\bar{c}, s\bar{s}, gg$ , and backgrounds $2f, 4f$ ).
- Four specific XGBoost configurations are tested: three using inputs from a single tagger (XGB_PN, XGB_ParT, XGB_MIParT) and one Combined classifier (XGB_Combined) utilizing outputs from all three taggers simultaneously.
Signal Extraction:
- A loose preselection is applied to enhance the $ZH(\nu\bar{\nu})$ topology.
- For each decay channel, the XGBoost score is treated as a discriminant. Optimal working points are chosen to maximize the statistical significance $Z$ using the profile-likelihood approximation.

Key Contributions and Results
The study provides projected sensitivities for hadronic Higgs decays at the CEPC, normalized to $20 \text{ ab}^{-1}$ :

Precision Improvements: Compared to the published CEPC Higgs benchmark study [40], the proposed method significantly improves precision:
- $H \to c\bar{c}$ : Precision improves from 1.85% to 1.06% (a ~43% relative improvement).
- $H \to gg$ : Precision improves from 0.70% to 0.50% (a ~29% relative improvement).
- $H \to b\bar{b}$ : Precision improves modestly from 0.20% to 0.17%.
Strange Quark Sensitivity ( $H \to s\bar{s}$ ):
- This channel was not included in the previous CEPC benchmark. The study provides a quantitative sensitivity estimate, predicting a statistical significance of approximately 1.5 $\sigma$ for the single $ZH \to \nu\bar{\nu}s\bar{s}$ channel.
- The projected relative precision is 68%, limited by low signal statistics and residual overlap with $H \to gg$ and $4f$ backgrounds.
Ensemble Benefits: The XGB_Combined classifier consistently outperforms single-tagger configurations, particularly for the difficult light-flavor channels ( $H \to gg$ and $H \to s\bar{s}$ ). The combination reduces the spread in results across independent training runs, indicating enhanced robustness against statistical fluctuations.
Interpretability: SHAP (SHapley Additive exPlanations) analysis confirms that the classifier relies on physically sensible features: heavy-flavor scores drive $b/c$ separation, while strange/gluon scores and global kinematics (missing energy, dijet mass) drive the separation of light flavors and backgrounds.

Significance and Claims
The paper claims that the developed two-stage framework demonstrates the potential of deep-learning-based jet flavor tagging for precision Higgs physics at future lepton colliders.

Methodological Advantage: The modular two-stage approach (jet-level tagging followed by event-level classification) is highlighted as offering better interpretability and flexibility for systematic uncertainty treatment compared to "holistic" event-level classifiers (where the entire event is fed directly into a single network). While the holistic approach yields similar statistical precision in simplified benchmarks, the two-stage method allows for independent calibration of jet taggers and easier integration of data-driven constraints.
Physics Impact: The results establish a concrete benchmark for probing the strange-quark Yukawa coupling, a previously unexplored regime. The study suggests that while $H \to s\bar{s}$ remains statistically limited in the baseline CEPC configuration, the sensitivity could be further enhanced with higher luminosity, additional $Z$ decay channels, or improved hadronization modeling.
Scope: The authors explicitly state that these are statistical-only projections. Systematic uncertainties related to jet energy scale, flavor-tagging calibration, and theoretical modeling are not included but are acknowledged as necessary for future realistic assessments. The analysis strategy is presented as broadly applicable to other $e^+e^-$ Higgs factories (e.g., FCC-ee, ILC) with appropriate detector re-interpretation.

More like this