Abstract
The reconstruction and identification of hadronic objects has a wide variety of use-cases inside the physics program of the LHC, ranging from Standard Model measurements and Higgs searches up to the more exotic searches of boosted resonances. The CMS collaboration devoted a constant effort in improving such tools in order to extend the analyses reach. A review of the current status of the identification algorithms for hadronic taus, heavy flavour jets, quark/gluon discrimination, and boosted resonances is presented together with their performance, either measured on data or expected from simulation. Special attention is devoted to novel techniques recently introduced, and their impact on the final performance on the classification task.
Introduction
In the high-energy environment of the Large Hadron Collider (LHC), most interesting physics processes manifest as "jets"—collimated sprays of hadrons resulting from the fragmentation of quarks and gluons. The fundamental challenge for the CMS experiment is "tagging": accurately determining the flavor of the particle that initiated the spray. Whether it is a light quark, a gluon, a heavy $b$-quark, or a tau lepton, precise identification is crucial for nearly every physics analysis.
As the LHC pushes into higher energy regimes, particles are increasingly produced with high transverse momentum ("boosted"), causing their decay products to merge into a single, wide-radius jet. Distinguishing these signals from standard QCD backgrounds requires sophisticated algorithms. CMS is currently transitioning from traditional cut-based methods to advanced Deep Learning architectures to meet these challenges.
Reconstruction and Identification Methods
Jet Clustering and Grooming
CMS utilizes the anti-$k_T$ algorithm as the standard for jet clustering, typically using a radius parameter ($R$) of 0.4 for resolved jets and 0.8 for "fat" boosted jets. To handle the "noise" from soft, wide-angle radiation that can obscure the hard decay structure of boosted objects, techniques like SoftDrop and jet pruning are employed to "groom" the jets. Furthermore, the DDT (Designed Decorrelated Taggers) method is used to ensure that tagger performance remains independent of the jet's mass or momentum, which is vital for reliable background estimation.
Hadronic Tau Identification
The identification of hadronic taus ($\tau_h$) relies on the Hadron Plus Strip (HPS) algorithm. For Run II, this algorithm was enhanced with dynamic "strips" in the calorimeter that adjust their size based on transverse momentum ($p_T$) to better capture decay products. The current state-of-the-art approach uses a Boosted Decision Tree (BDT) to combine information regarding isolation, jet shape, and lifetime.
As shown in the performance plots, the MVA-based tau identification achieves a significantly lower mis-identification rate—nearly an order of magnitude better in some regions—compared to traditional cut-based isolation methods.
Heavy Flavor Tagging
The evolution of heavy flavor tagging at CMS represents a major leap in machine learning application:
- CSVv2: A shallow Neural Network.
- DeepCSV: A Deep Neural Network (DNN) with four output nodes, enabling simultaneous classification of jets as $b$, $bb$, $c$, or light flavor.
- DeepFlavour: The latest iteration, utilizing Convolutional Neural Networks (CNNs) to process 674 input features, a massive increase from the 68 features used in DeepCSV.
By training models like DeepCSV to recognize multiple categories simultaneously, CMS has gained highly effective charm-taggers as a secondary benefit, outperforming previous dedicated algorithms.
Performance and Validation
The algorithms are developed using simulated $t\bar{t}$ events for $b$-tagging and $H \to \tau\tau$ samples for tau identification. These simulations are validated against real LHC data from 2016 and 2017 using $Z$+jets and dijet events. "Scale Factors" are calculated to ensure that the performance gains observed in simulation are accurately reflected in real-world data, with most factors remaining near 1.0.
DeepFlavour, in particular, demonstrates dominance at high $p_T$, catching significantly more true $b$-jets for a fixed mis-identification rate compared to its predecessors.
Conclusion and Future Outlook
While the shift to complex models like DeepFlavour offers superior precision, it introduces challenges in computational speed and optimization. Future efforts will focus on reducing systematic uncertainties and exploring even more advanced techniques, such as Graph Neural Networks or Transformers, to handle the increasingly "boosted" environment of high-luminosity LHC runs.
Summary updated