Source-linked AI summary
Jet Flavour Classification Using DeepJet
Emil Bols, Jan Kieseler, Mauro Verzetti, Markus Stoye, Anna Stakia
TL;DR
Jet flavour classification is important for exploiting flavour-dependent heavy-hadron production and decay signatures, but prior models use small, strongly selected input subsets. DeepJet processes full low-level information from jet constituents and related variables through a dedicated architecture, achieving improved heavy-flavour performance and quark-gluon discrimination. The study reports an almost 20% efficiency increase at 10^-3 misidentification probability for jets with pT > 90 GeV.
Problem
Prior jet-flavour classifiers use only small subsets of charged constituents selected with stringent quality criteria, risking information loss and performance degradation.
Method
DeepJet is a multiclass architecture that processes charged and neutral particles, secondary vertices, global variables, and approximately 650 low-level input variables without constituent selection.
Results
Almost 20% efficiency increase at 10^-3 misidentification probability is reported for jets with pT > 90 GeV versus DeepCSV, with gains also observed for c-jet and quark-gluon identification.
Takeaways & Limitations
DeepJet extends heavy-flavour tagging to quark-gluon discrimination while exploiting the full information in a jet.
Abstract
from arXiv · showhide
Jet flavour classification is of paramount importance for a broad range of applications in modern-day high-energy-physics experiments, particularly at the LHC. In this paper we propose a novel architecture for this task that exploits modern deep learning techniques. This new model, called DeepJet, overcomes the limitations in input size that affected previous approaches. As a result, the heavy flavour classification performance improves, and the model is extended to also perform quark-gluon tagging.
1 Introduction
Jet flavour classification matters because heavy-flavour jets carry distinctive fragmentation and lifetime information, while existing classifiers use limited, highly selected inputs. DeepJet addresses this by processing all constituent types without constituent selection.
- Motivation: Heavy-flavour jet classification is important because bottom and charm production and decay create flavour asymmetries, displaced tracks, and secondary vertices.Bottom and charm hadrons have sizeable lifetimes, with cτ values of approximately 0.5 mm and 0.3 mm, respectively.
- Existing approaches: Combined information from jet tracks and reconstructed secondary vertices provides stronger discrimination than isolated features.Machine-learning classifiers exploit collective track behaviour together with secondary-vertex information.
- Existing approaches: ATLAS RNN and CMS DeepCSV improved jet classification, but they process restricted inputs rather than the full constituent set.DeepCSV combines selected tracks and secondary vertices through fully connected layers without sequence processing.
- Limitations: Stringent constituent selection loses information and can degrade performance, while relaxing track selection has produced significant gains.Deep Sets transforms individual track features into latent representations before combining them and achieves similar performance to an RNN with reduced training and inference time for the same inputs.
- DeepJet proposal: DeepJet removes constituent selection and simultaneously processes charged particles, neutral particles, secondary vertices, and global event variables.The design targets previous limitations in input purity and input count while using the full jet information.
2 Setup
The study trains and evaluates DeepJet on simulated CMS-reconstructed jets with flavour labels assigned by ghost association. Its inputs comprise approximately 650 variables spanning global, charged-candidate, neutral-candidate, and secondary-vertex information.
- Samples and simulation: DeepJet is trained and tested on simulated anti-kT jets with R = 0.4 from QCD multijet and fully hadronic top-pair samples.PYTHIA8 and POWHEGv2 generate the samples, while PYTHIA8 performs hadronization and showering.
- Samples and simulation: Approximately 130 million jets are split into training, validation, and testing samples in a 0.765:0.135:0.1 ratio.CMS Phase 1 detector response is simulated with GEANT4, and constituents are reconstructed using the Particle Flow algorithm.
- Labelling: Ghost association labels jets as b when they contain a b hadron, as c when they contain c but no b hadron, and otherwise as light-flavour jets.The last generated b and c hadrons before decay are included as low-momentum directional ghosts for clustering.
- Input features: DeepJet uses approximately 650 input variables divided among global variables, charged candidates, neutral candidates, and secondary-vertex features.Global inputs include jet kinematics, track and secondary-vertex counts, and the number of reconstructed primary vertices to represent pileup effects.
- Preprocessing: Preprocessing matches each flavour class in pT and η to the b-jet shape and bounds several variables within physically defined ranges.Unavailable or infinite object features are replaced with an appropriately chosen value.
3 DeepJet
DeepJet processes large sets of jet constituents through separate convolutional and recurrent branches before combining the resulting information with fully connected layers. The architecture uses 1x1 convolutions for per-constituent feature engineering and LSTM layers for constituent sequences.
- 3.1 Architecture: DeepJet applies 1x1 convolutional branches to perform automatic feature engineering for each jet constituent.The filter size applies the same feature transformation to candidates without using information from other constituents.
- 3.2 Training procedure: Batch normalization is applied throughout the network, and dropout with rate 0.1 provides regularisation.Batch normalization occurs at the network beginning and between layers.
- 3.2 Training procedure: The model is trained with Adam for 65 epochs using categorical cross entropy, halving the learning rate after validation loss stagnates for more than 10 epochs.Over-training is monitored with validation loss and ROC curves across transverse-momentum ranges.
- 3.1 Architecture: DeepJet’s architecture is closely related to the Deep Sets concept.The paper identifies this relationship in a note accompanying the architecture discussion.
- 3.1 Architecture: Three branches separately process charged candidates, neutral candidates, and secondary vertices.The branch structure preserves distinct constituent categories before their information is combined.
- 3.1 Architecture: Three RNN (LSTM) layers combine information within the charged-candidate, neutral-candidate, and secondary-vertex sequences.The recurrent layers follow the category-specific convolutional processing.
- 3.1 Architecture: Fully connected layers combine the resulting branch information into full-jet information.This final combination follows the separate constituent-category processing and sequence aggregation.
4 Performance
DeepJet outperforms DeepCSV across heavy-flavour and quark/gluon classification tasks, with gains especially evident at higher jet pT. Comparisons also show that exploiting larger, less-selected inputs requires an architecture designed to process them efficiently, while pileup degradation remains similar.
- Comparison with other b-tagging algorithms: Almost 20% efficiency increase at 10^-3 misidentification probability is observed for jets with pT > 90 GeV when comparing DeepJet with DeepCSV.This comparison uses fully hadronic tt̄ events.
- Comparison with other b-tagging algorithms: DeepJet shows an increasing performance gain over DeepCSV at higher jet pT in QCD multijet events.The comparison is made at fixed light-jet misidentification probabilities and is attributed presumably to information from all constituents.
- Comparison with other b-tagging algorithms: DeepJet outperforms DeepCSV for light-versus-c jet classification in fully hadronic tt̄ events.Both models use the binary discriminator P(c) / [P(c)+P(uds)+P(g)].
- Comparison with other b-tagging algorithms: DeepJet improves quark/gluon discrimination over the CMS quark-gluon likelihood on pure light-quark and gluon jets.The paper also expects a larger gain when light-quark and gluon jets must be separated from heavy-flavour jets.
- Qualitative assessment of the performance gain source: Similar performance degradation with increasing pileup is observed for DeepJet and DeepCSV despite DeepJet using more low-purity input.Pileup is assessed through the number of reconstructed primary vertices at fixed light-jet mistag rates.
- Qualitative assessment of the performance gain source: Increasing inputs alone provides little benefit without an architecture that efficiently processes larger, less-pure track sets.Changing the architecture alone while retaining selected inputs also has limited gain; comparable preprocessing effects were observed for a Deep Sets tagger.
5 Conclusion
DeepJet is a multiclass flavour-tagging algorithm that exploits full jet information and extends beyond heavy-flavour tagging to quark–gluon discrimination. It improves performance over fully connected models using smaller engineered-feature sets, with gains sometimes exceeding two-fold at the same misidentification rate.
- DeepJet is a multiclass flavour-tagging algorithm designed to exploit the full information in a jet.
- The model uses low-level variables, loose input selection, and an architecture that processes inputs efficiently.
- DeepJet improves performance across all tested flavour-tagging topologies compared with fully connected models using fewer engineered features.
- In some cases, DeepJet exceeds a two-fold efficiency gain at the same misidentification rate.
- The approach also performs quark–gluon discrimination and has been extended to tagging algorithms for exotic long-lived particles.
A.1 List of global variables
The global-variable list includes counts describing jet constituents, secondary vertices, and primary vertices in the event.
- A.1 List of global variables: Global variables include the number of charged particle-flow candidates in the jet.
- A.1 List of global variables: Global variables include the number of neutral particle-flow candidates in the jet.
- A.1 List of global variables: Global variables include the number of secondary vertices in the jet.
- A.1 List of global variables: Global variables include the number of primary vertices in the event.
A.2 List of charged candidate variables
The charged-candidate variables describe momentum alignment and fraction, impact parameters, track–jet geometry, pileup weighting, vertex association, and track-fit quality.
- A.2 List of charged candidate variables: Charged-candidate variables include dot products with jet momentum, including a version divided by jet-momentum magnitude.
- A.2 List of charged candidate variables: The list includes 2D and 3D track impact-parameter values and significances.
- A.2 List of charged candidate variables: Charged-candidate variables include the track distance to the jet axis and the fraction of jet momentum carried by the track.
- A.2 List of charged candidate variables: The inputs include the charged-candidate PUPPI weight and a flag indicating whether the track entered the primary-vertex fit.
- A.2 List of charged candidate variables: Track-fit information includes χ2 and an integer quality flag based on detector hits and overall χ2.
A.3 List of neutral candidate variables
The neutral-candidate variables describe momentum fraction, particle identity, calorimeter deposition, proximity to secondary vertices, and pileup weighting.
- A.3 List of neutral candidate variables: Neutral-candidate variables include the fraction of jet momentum carried by the neutral candidate.
- A.3 List of neutral candidate variables: The list includes an integer flag indicating whether the neutral candidate is a photon.
- A.3 List of neutral candidate variables: Neutral-candidate inputs include the fraction of energy deposited in the hadronic calorimeter.
- A.3 List of neutral candidate variables: The variables include ΔR between the neutral candidate and the closest secondary vertex.
- A.3 List of neutral candidate variables: The neutral-candidate PUPPI weight is included as an input.
A.4 List of secondary vertex variables
The secondary vertex variables include track multiplicity, impact-parameter measurements and significances, flight-direction alignment, and vertex-to-jet energy ratio.
- Secondary vertex track multiplicity is included as an input variable.
- Secondary vertex displacement is characterized by 2D and 3D impact-parameter values and significances.
- The variable set also includes flight-direction alignment and the ratio of secondary vertex energy to jet energy.