Jet flavor tagging with Particle Transformer for Higgs factories

This paper demonstrates that the Particle Transformer (ParT) significantly outperforms traditional BDT-based taggers for jet flavor identification at Higgs factories, achieving a 5–10-fold improvement in $b$/$c$ tagging efficiency while also enabling effective strange tagging and quark/antiquark separation through the integration of multivariate hadron identification data.

The Gist

Imagine you are a detective trying to solve a crime scene inside a giant, high-tech factory. This factory is a Higgs Factory, a machine designed to smash particles together to create the Higgs boson (the particle that gives other particles mass).

When these particles smash, they don't just disappear; they explode into sprays of smaller particles called jets. Your job is to look at these jets and figure out: "What kind of particle started this explosion?"

Was it a heavy, slow-moving "b-quark"? A slightly lighter "c-quark"? Or was it a common, light "u-quark" or a "gluon"?

This is called Flavor Tagging. It's like trying to identify a suspect by looking at the footprints they left behind.

The Old Way: The Manual Detective

Traditionally, physicists used a method called LCFIPlus. Think of this as a detective who looks at the crime scene, manually finds the "secondary footprints" (displaced tracks from heavy particles), measures them with a ruler, and then feeds those numbers into a standard calculator (a BDT) to make a guess. It works, but it's slow and relies on the detective finding the right clues first.

The New Way: The AI Super-Scanner

This paper introduces a new detective: the Particle Transformer (ParT).

Instead of looking for specific footprints one by one, ParT is like a super-powered AI scanner that looks at the entire spray of particles at once. It doesn't just see "a particle here" and "a particle there"; it understands how every single particle in the jet is talking to every other particle.

• The Analogy: Imagine a crowded room.
  • The Old Detective asks, "Who is standing near the door? Who is wearing a red hat?" and builds a profile based on isolated facts.
  • The ParT AI looks at the whole room and instantly understands the vibe. It sees how people are clustering, how they are moving in relation to each other, and instantly knows, "This group is definitely a b-quark family," or "That group is a light-flavor tourist."

What Did They Test?

The researchers tested this AI on data from the ILD detector (a very detailed camera system for the Higgs Factory). They trained the AI in three different ways:

1. The Simple Test (3 Categories): Can you tell the difference between a heavy "b", a medium "c", and everything else ("light")?
2. The Harder Test (6 Categories): Can you distinguish between b, c, strange (s), up (u), down (d), and gluons (g)?
3. The Expert Test (11 Categories): Can you not only identify the flavor but also tell if it's a particle or its anti-particle (like a quark vs. an anti-quark)?

They also gave the AI a special "superpower": Particle Identification (PID).

• The Analogy: Imagine the particles are wearing ID badges. The detector can read these badges (using ionization and time-of-flight data). The AI uses these badges to know exactly what kind of particle it is looking at, which is crucial for spotting "strange" particles that are otherwise very hard to catch.

The Results: A Massive Leap Forward

The results were impressive.

• For Heavy Particles (b and c tagging): The new AI was 5 to 10 times better than the old method.
  • The Metaphor: If the old method was like finding a needle in a haystack with a magnet, the new method is like using a metal detector that beeps only when it's right on the needle. It can spot a "b-jet" while ignoring 99.9% of the fake "c-jet" noise.
• For Strange Particles: This was a new frontier. By using the "ID badge" data, the AI could actually identify strange jets reasonably well, something previous methods struggled with.
• The Data Scale: They trained the AI on 1 million jets, then 10 million. The more data they fed it, the smarter it got, suggesting that with even more training, it could become even sharper.

The "Quark vs. Anti-Quark" Trick

In the most advanced test (11 categories), the AI managed to tell the difference between a particle and its "evil twin" (antiparticle).

• The Analogy: It's like looking at a shadow and knowing if the person casting it is walking left or right. For heavy particles (like charm), the AI was surprisingly good at this. For light particles, it was still a bit like guessing, but for the heavy stuff, it was a breakthrough.

Why Does This Matter?

Future Higgs factories (like the one planned in Japan) will produce millions of these events. To find rare, new physics, scientists need to be able to sort through this massive pile of data with extreme precision.

This paper proves that modern AI (Transformers) is the perfect tool for the job. It's faster, more accurate, and can handle the complex, messy details of particle collisions better than the old, manual methods. It's the difference between a human sorting mail by hand and a high-speed robotic sorter that never gets tired and never makes a mistake.

In short: They built a super-smart AI that looks at particle explosions and instantly knows exactly what caused them, making the search for new physics at the Higgs Factory much easier and more precise.

Technical Summary

1. Problem Statement

Future $e^+e^-$ colliders (Higgs factories) require precise event reconstruction to identify the flavor of jets initiated by specific partons (quarks and gluons). This "flavor tagging" is critical for measuring Higgs boson decay properties (e.g., $H \to b\bar{b}, c\bar{c}, s\bar{s}$).

• Challenges: As detector granularity increases (e.g., the ILD concept), pattern recognition becomes complex. Traditional methods (like LCFIPlus) rely on explicit secondary vertex reconstruction and multivariate classifiers (BDTs), which may not fully exploit the fine-grained, low-level information available in modern detectors.
• Specific Needs: There is a need for improved discrimination between heavy flavors ($b, c$) and light flavors ($u, d, s, g$), particularly for strange ($s$) tagging, which relies heavily on hadron-level particle identification (PID) rather than just vertex displacement. Additionally, separating quarks from antiquarks is a novel challenge for precision physics.

2. Methodology

The authors employ the Particle Transformer (ParT), a deep learning architecture based on the Transformer model, adapted for jet physics.

• Data Samples:

  • Detector: Simulated using the International Large Detector (ILD) concept, featuring a silicon vertex tracker, TPC (for $dE/dx$), and timing layers for Time-of-Flight (ToF).
  • Physics Process: $e^+e^- \to ZH \to \nu\bar{\nu} q\bar{q}$ at $\sqrt{s} = 250$ GeV.
  • Samples:
    • Full Simulation: $\sim 2 \times 10^6$ jets per category (ILD central production).
    • Fast Simulation (SGV): Large statistics samples ($10^7$ jets per category) used to study scaling behavior.
  • Inputs: Per-particle features including kinematic variables, impact parameters, track uncertainties, and Comprehensive PID (CPID) outputs (probabilities for $\pi, K, p$) derived from $dE/dx$ and ToF.

• Model Architecture:

  • ParT: Utilizes self-attention mechanisms with a pairwise "interaction" term (derived from kinematics like $k_t, z, \Delta R, m$) injected as a bias into attention weights. This captures correlations between jet constituents.
  • Input Embedding: Separate projections for charged and neutral particles to handle feature availability differences.

• Training Scenarios:

  1. 3-Category: $b, c, d$ (light).
  2. 6-Category: $b, c, s, u, d, g$.
  3. 11-Category: Distinguishes quarks from antiquarks ($b, \bar{b}, c, \bar{c}, s, \bar{s}, u, \bar{u}, d, \bar{d}, g$). Labels are assigned via MC truth matching based on color-singlet systems and angular proximity.

3. Key Contributions

• Application of ParT to Higgs Factories: First comprehensive study applying the Particle Transformer to ILD full simulation and large-scale fast simulation samples for Higgs factory conditions.
• Integration of PID Information: Successfully incorporated multivariate hadron PID (from $dE/dx$ and ToF) as per-particle inputs, specifically targeting the difficult task of strange jet identification.
• Quark-Antiquark Separation: Demonstrated the feasibility of distinguishing quarks from antiquarks using deep learning, a task previously unaddressed in standard tagging workflows.
• Statistical Scaling Analysis: Quantified the performance gains achieved by increasing training statistics from 1M to 10M jets in fast simulation.

4. Results

• Heavy Flavor ($b/c$) Tagging:

  • Performance Gain: ParT-based taggers show a factor of 5–10 improvement in background rejection compared to traditional LCFIPlus BDTs at fixed signal efficiency.
  • Metrics: At 80% $b$-tag efficiency, $c$-jet background acceptance drops to $O(10^{-3})$ and light-flavor ($d$) background to $O(10^{-3} \text{--} 10^{-4})$.
  • Scaling: Increasing training data from 1M to 10M jets in fast simulation further reduced background acceptance (e.g., $c$-background for $b$-tagging dropped from $1.48 \times 10^{-3}$ to $4.64 \times 10^{-4}$).
  • Comparison: 3-category and 11-category trainings yielded comparable $b$-tagging performance, though 11-category training faced challenges in rejecting gluons at high purity due to $g \to b\bar{b}$ splittings.

• Strange ($s$) Tagging:

  • Performance: Achieved reasonable performance, heavily dependent on PID inputs.
  • Limitations: Separation between light flavors ($u, d, s$) and gluons remains probabilistic due to fragmentation (non-strange jets containing strange hadrons). However, gluon jets were distinguishable from quarks based on kinematic/radiation patterns.

• Quark-Antiquark Separation (11-Category):

  • The model successfully separated quarks from antiquarks for heavy flavors ($b, c, s$).
  • Charm ($c$): Showed the strongest separation accuracy, likely due to the larger electric charge of charm hadrons providing distinct kinematic signatures.
  • Light Flavors: Separation of $u/\bar{u}$ and $d/\bar{d}$ pairs remained difficult, with performance consistent with random guessing for charge-conjugate pairs.

5. Significance and Future Outlook

• Physics Impact: The substantial improvement in $b/c$ tagging and the new capability for $s$-tagging and quark/antiquark separation will significantly enhance the precision of Higgs coupling measurements and rare decay searches at future colliders.
• Scalability: The results confirm that deep learning models like ParT benefit significantly from larger training datasets, motivating the generation of even larger simulation samples.
• Deployment: The taggers are being integrated into the ILD analysis workflow via an ONNX-based inference pipeline.
• Future Work: Focus will shift to refining input representations, optimizing training strategies, and systematically validating the differences between full and fast simulations to ensure robustness in physics measurements.