Optimizing b-Jet Performance in the CMS High-Level Trigger with Run-3 Data

This paper evaluates the commissioning and performance of b-jet triggers within the CMS High-Level Trigger system using Run-3 data to optimize real-time heavy-flavor jet selection.

The Gist

Imagine you are a security guard at the world’s busiest airport. Every second, thousands of people (data particles) rush through the gates. Your job is to let the "important travelers" (interesting physics events) through to the VIP lounge (the permanent storage) while quickly turning away the "regular commuters" (background noise) so the lounge doesn't get overcrowded.

This paper describes how scientists at the CMS experiment (part of the Large Hadron Collider) upgraded their "security scanners" to better identify a very specific, high-value traveler: the b-jet.

The Challenge: The "Identity Crisis"

In the world of subatomic particles, a b-jet is like a VIP traveler who is wearing a very convincing disguise. They look a lot like "light jets" (regular travelers), but they have one tiny giveaway: they tend to take a slightly different path or "stumble" a little bit because they are heavier and live just a fraction of a second longer before decaying.

In previous years (Run 2), the CMS team used an older scanning system called DeepJet. It was good, but as the LHC got more powerful and the "crowds" at the airport got bigger, the old scanners couldn't keep up. They were either too slow or they were accidentally letting too many regular commuters into the VIP lounge, which would clog up the system.

The Solution: The "Super-Scanner" (ParticleNet@HLT)

To fix this, the team introduced a new, high-tech AI scanner called ParticleNet@HLT.

Think of the old scanner like a person looking at a photo of a traveler and trying to guess who they are based on a checklist. ParticleNet is more like a sophisticated AI that looks at the entire movement of the crowd. It doesn't just look at the person; it looks at how they walk, how they interact with the people around them, and the subtle "wobble" in their step.

Because it uses a "Graph Neural Network," it treats the particles inside a jet like a social network, looking at the connections and relationships between them to spot the "VIP" signature.

The Results: Faster, Smarter, Better

The paper reports three big wins from using this new AI:

1. Better Eyesight (Efficiency): The new scanner is about 10–15% better at spotting the VIPs (b-jets) than the old one. It’s like upgrading from a blurry CCTV camera to 4K high-definition.
2. Fewer Mistakes (Mistagging): It is much better at not being fooled by regular travelers. It keeps the "VIP lounge" exclusive to the right people.
3. Stability: Even as the LHC's "traffic" changed over the years (from 2022 to 2024), the scanner remained incredibly consistent. It didn't get confused by the changing crowds.

Why does this matter?

By being able to spot these b-jets instantly and accurately, scientists can now study rare and mysterious phenomena—like the Higgs Boson (the particle that gives everything mass) decaying into b-jets—much more effectively.

In short: The CMS team just gave their "security system" a massive brain upgrade, ensuring that when the next big discovery happens, they won't miss it in the crowd.

Technical Summary

Technical Summary: Optimizing b-Jet Performance in the CMS High-Level Trigger with Run-3 Data

1. Problem Statement

In the CMS experiment at the Large Hadron Collider (LHC), identifying jets originating from heavy quarks (b- and c-jets) is essential for studying Standard Model processes (like $t\bar{t}H$) and searching for new physics (such as Higgs boson pair production, $HH \to b\bar{b}b\bar{b}$).

The challenge lies in the High-Level Trigger (HLT), the software-based system responsible for reducing the collision rate from ~30 MHz to ~1–3 kHz. During Run-3, the aging conditions of the tracker made previous multivariate classifiers (like DeepJet) difficult to operate at sustainable rates without significantly raising energy thresholds. To maintain high physics sensitivity, CMS needed a more efficient algorithm capable of distinguishing b-jets from light-flavor and c-jets with higher precision and lower computational latency.

2. Methodology

The paper details the transition from traditional feed-forward Deep Neural Networks (DNNs) to a more advanced architecture: ParticleNet@HLT.

• Algorithm Architecture: ParticleNet is a dynamic graph convolution neural network (GCNN). Unlike previous methods, it treats a jet as an unordered set of Particle-Flow (PF) candidates and secondary vertices (SV). It uses successive edge-convolution layers to learn both local and global correlations between jet constituents in a permutation-invariant manner.
• Input Features: The model utilizes track impact parameters, flight-distance observables, SV kinematics, and jet composition variables.
• Training and Implementation: The model was trained on simulated AK4 jets ($p_T > 30$ GeV, $|\eta| < 2.5$). It classifies jets into five categories: b, c, light (uds), gluon, and hadronic $\tau$.
• Working Points (WPs): Three operational thresholds were established to control mistag rates:
  • Loose: 10% mistag rate.
  • Medium: 1% mistag rate.
  • Tight: 0.1% mistag rate.

3. Key Contributions

• Deployment of ParticleNet@HLT: The successful integration of a GCNN into the real-time HLT environment, replacing the older DeepJet and DeepCSV algorithms.
• Multi-Flavor Tagging: For the first time, the HLT can support dedicated triggers for c-jets and $\tau$-jets, significantly expanding the searchable physics landscape.
• Continuous Optimization: The implementation of a regular retraining strategy using updated simulated samples to ensure the algorithm remains robust against changing detector conditions across different data-taking eras (RunC through RunI).

4. Results

• Algorithmic Superiority: ParticleNet@HLT provides a 10–15% gain in b-tagging efficiency at fixed misidentification rates compared to the DeepJet algorithm.
• Stability in Data: Performance studies using 2024 data (109 $\text{fb}^{-1}$) show that the per-jet efficiency remains stable across different data-taking eras, with ratios relative to the reference era (RunC) remaining near unity.
• Trigger Efficiency Improvements:
  • In the $t\bar{t}$ control region, the trigger efficiency shows high correlation with offline ParticleNet scores.
  • For the $HH \to 4b$ physics trigger ($4j2b$ topology), the 2024 HLT menu achieved an approximate 6% efficiency improvement over 2023, with only a marginal increase in the trigger rate (~50 Hz).

5. Significance

This work represents a major upgrade in the CMS real-time selection capabilities. By moving to graph-based neural networks, CMS has unlocked the ability to efficiently trigger on fully hadronic final states that were previously difficult to capture. This optimization is vital for maximizing the physics reach of Run-3. Furthermore, the success of ParticleNet paves the way for future transitions to transformer-based architectures, which will be necessary to handle the even higher data rates and complexities expected during the High-Luminosity LHC (HL-LHC) Phase-2 upgrade.