Machine learning in online and offline reconstruction and identification with CMS

Machine learning (ML) plays an increasingly important role in both online and offline event reconstruction and identification at CMS experiment. A variety of ML techniques are used to improve the identification of physics objects. Dedicated algorithms enhance jet flavor tagging, including new approaches that strengthen sensitivity to Higgs boson decays to charm quarks. Tau identification has been significantly improved with ML-based methods, while in the electromagnetic calorimeter, ML-driven clustering techniques provide better energy reconstruction. Muon identification also benefits from multivariate approaches, leading to a higher signal efficiency and more background rejection. Looking at the future, ML will be central to the reconstruction strategy for the High-Granularity Calorimeter at high-luminosity LHC. New algorithms for the upgraded detectors are being developed to cope with extreme pileup conditions. All these advances ensure that CMS can fully exploit the physics potential of Run-3 and the HL-LHC, while also exploring novel ML strategies to maintain robust performance under evolving experimental conditions.

💡 Research Summary

The paper provides a comprehensive overview of how machine learning (ML) has become integral to both online and offline reconstruction and object identification in the CMS experiment, covering developments from Run 3 through the upcoming High‑Luminosity LHC (HL‑LHC). Four main areas are examined: jet flavor tagging, hadronic tau identification, electron/photon/muon identification, and Phase‑2 reconstruction for the High‑Granularity Calorimeter (HGCAL).

In jet flavor tagging, a lightweight ParticleNet (PNet) model has been deployed directly in the High‑Level Trigger (HLT) for AK4 and AK8 jets. Performance studies show that the per‑jet b‑tagging efficiency remains stable across Run C to Run I, while the mean transformed b‑tag score improves year‑by‑year due to retraining and working‑point adjustments. Building on the original ParticleTransformer, the Unified ParticleTransformer (UParT) adds adversarial training to increase robustness against simulation mismodelling. UParT delivers state‑of‑the‑art performance for b‑, c‑, and even s‑tagging, achieving roughly 10 % lower light‑jet mis‑identification for c‑jets and consistent gains in b‑jet tagging compared with the legacy DeepJet tagger.

For hadronic tau identification, the DeepTau algorithm has evolved from version 2.1 to 2.5. The newer version incorporates richer constituent‑level features, improved pile‑up mitigation, and a domain‑adaptation strategy that aligns simulation and data distributions. Across both low‑pT (< 100 GeV) and high‑pT (> 100 GeV) regimes, DeepTau v2.5 yields 3‑5 % higher signal efficiency and 7 % better background rejection relative to v2.1. While DeepTau remains the offline standard, ParticleNet‑based tau taggers will replace it in the HLT from 2025 onward, providing real‑time performance gains.

Electron and photon reconstruction have transitioned from the purely geometric Mustache clustering to the DeepSuperCluster approach, a deep neural network that ingests ECAL Rec‑hits and learns detailed shower shapes. This change improves energy resolution by about 15 % in low‑energy bins and brings the data‑to‑simulation ratio of the reconstructed energy close to unity across the full energy spectrum. Muon identification has similarly moved from cut‑based selections to a multivariate analysis (MVA) classifier that combines track quality, tracker‑muon matching, and other high‑level variables. The MVA achieves roughly 8 % higher background rejection at fixed efficiency and demonstrates stable performance throughout Run 3.

Phase‑2 reconstruction for the HGCAL addresses the extreme pile‑up conditions expected at the HL‑LHC (up to 200 simultaneous interactions). The Iterative Clustering (TICL) algorithm has been upgraded to version v5, which groups calorimeter hits into three‑dimensional “tracksters.” Although TICL retains a traditional clustering backbone, machine‑learning enhancements are applied at several stages. A graph neural network (GNN) classifier, built on a dynamic reduction network, discriminates electromagnetic from hadronic tracksters with a separation power improvement of about 20 %. This GNN‑enhanced reconstruction preserves both spatial and timing resolution, crucial for maintaining physics performance under high pile‑up.

Across all domains, the paper highlights the use of adversarial training and domain adaptation to mitigate simulation‑data mismatches, the importance of lightweight models for trigger latency constraints, and the systematic validation of ML‑driven algorithms over multiple data‑taking periods. The reported gains—typically 5‑20 % improvements in efficiency, resolution, or background rejection—translate directly into increased sensitivity for key physics analyses such as Higgs boson decays to heavy flavors, electroweak measurements, and searches for new phenomena. In summary, the integration of modern ML techniques positions CMS to fully exploit the physics potential of Run 3 and the HL‑LHC while ensuring robust, scalable performance in the face of evolving detector conditions.