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

The real-time identification and selection of b-jets play a crucial role in the CMS experiment, particularly in searches involving heavy-flavor jets. The High-Level Trigger (HLT) is designed to efficiently select events of interest while maintaining a manageable output rate of a few kilohertz. This report presents the commissioning and performance evaluation of b-jet triggers in the CMS HLT system using proton-proton collision data collected during Run-3 (2022-2024). Key aspects include algorithm optimization, efficiency studies, and comparisons with offline reconstruction. The results provide valuable insights into the current b-jet selection strategy and highlight potential refinements for future data-taking campaigns.

💡 Research Summary

The paper presents a comprehensive study of the commissioning, optimization, and performance of b‑jet triggers in the CMS High‑Level Trigger (HLT) during LHC Run 3 (2022‑2024). The authors begin by motivating the need for efficient real‑time heavy‑flavor tagging, which is essential for a broad physics program that includes measurements of tt̄H, vector‑boson‑fusion Higgs production, Higgs‑pair production (HH → bb̄bb̄), and searches for heavy resonances decaying to b‑jets. In Run 2, the online b‑tagging relied on the DeepJet deep‑neural‑network classifier, but aging tracker conditions and the higher instantaneous luminosity of Run 3 forced the use of higher pT thresholds, limiting physics reach.

To overcome these limitations, the CMS collaboration introduced ParticleNet@HLT, a graph‑neural‑network (GNN) specifically optimized for the trigger environment. ParticleNet treats each jet as a set of particle‑flow (PF) candidates and secondary vertices (SVs) forming a dynamic graph. Edge‑convolution layers learn both local and global correlations among constituents, allowing the network to exploit displaced tracks, flight‑distance observables, SV kinematics, and jet‑composition variables. The model is trained on simulated AK4 jets (pT > 30 GeV, |η| < 2.5) matched to generator‑level jets, and predicts per‑jet probabilities for five classes: b, c, light‑flavor (uds), gluon, and hadronic τ.

Three working points (WPs) – Loose, Medium, Tight – are defined to correspond to mis‑identification rates of 10 %, 1 %, and 0.1 % for c‑ and light‑flavor jets, respectively. In simulation, ParticleNet@HLT delivers a 10‑15 % increase in b‑tag efficiency at fixed mistag rates compared with DeepJet, and an even larger gain over the older DeepCSV tagger.

The algorithm was deployed online in 2022, retrained with updated heavy‑flavor enriched samples in 2023, and became the standard b‑tagger for the HLT from 2023 onward. Performance is evaluated using two complementary approaches. First, per‑jet efficiency is measured in a tt̄‑enriched data set requiring at least two offline jets (pT > 35 GeV, |η| < 2.5) matched to online jets (pT > 30 GeV). The efficiency, defined as the fraction of offline jets matched to online jets that pass the ParticleNet WP, remains stable at ≈ 84 % across the 2024 data set (109 fb⁻¹) and shows less than 2 % variation across the individual Run periods (RunC‑RunI). The three WPs exhibit consistent efficiency curves as a function of the transformed offline ParticleNet score (BvsAll), confirming the robustness of the online calibration.

Second, trigger‑level efficiency is probed in a dilepton (e‑μ) tt̄+jets control region. Here the trigger efficiency is defined as the ratio of events passing the HLT e μ + 2b path (with ParticleNet) to events satisfying the offline e μ + 2b selection (offline ParticleNet WP ≈ 84 % efficiency, ≈ 1 % mistag). The efficiency as a function of the mean offline ParticleNet score of the two leading b‑jets shows a modest year‑to‑year increase, with 2024 achieving roughly a 6 % gain over 2023 for the same physics path. This improvement is attributed to a slightly looser online selection and the 2023‑2024 retraining of the GNN.

The impact on physics‑driven triggers is illustrated with the 4j2b path (≥ 4 jets, ≥ 2 b‑tagged at HLT) and the HH → 4b trigger. Simulated HH → 4b events reveal that the 2024 HLT menu yields an efficiency increase of about 6 % relative to 2023, at the cost of an additional ≈ 50 Hz in output rate. The gain is most pronounced for high‑mass HH resonances (MHH > 1 TeV), where the enhanced b‑tagging directly translates into higher signal acceptance.

Technical upgrades in Run 3 also include an improved track‑reconstruction algorithm, which reduces HLT latency and enables the deployment of computationally intensive GNN inference without exceeding the overall budget. This synergy between faster tracking and more powerful machine‑learning models is a key factor in achieving the observed performance gains.

In the conclusions, the authors emphasize that ParticleNet@HLT has demonstrably outperformed the previous DeepJet and DeepCSV taggers in both efficiency and stability, while maintaining acceptable trigger rates. Regular retraining with up‑to‑date simulated samples and careful tuning of WP thresholds are highlighted as essential practices for sustaining performance. Looking ahead, the paper outlines plans to explore transformer‑based architectures for the HLT, motivated by their ability to capture global context efficiently and to scale to the higher data rates expected in the HL‑LHC Phase‑2 upgrade. The authors anticipate that such next‑generation models, combined with continued improvements in tracking and computing resources, will further extend the physics reach of CMS in heavy‑flavor and Higgs‑pair searches. Overall, the study provides a clear roadmap for maintaining and enhancing b‑jet trigger capabilities throughout Run 3 and into the high‑luminosity era.