[Paper Review] Jet tagging in the Lund plane with graph networks

Frédéric A. Dreyer, H. Qu|arXiv (Cornell University)|Dec 15, 2020

Particle physics theoretical and experimental studies60 references4 citations

TL;DR

This paper introduces LundNet, a graph neural network-based jet tagging method that represents jets via their Lund plane clustering history, using kinematic features from pairwise emissions to improve top and W boson tagging performance. It achieves state-of-the-art results—especially in top tagging—while reducing training time by an order of magnitude and enhancing robustness to non-perturbative effects through kinematic cuts in the Lund plane.

ABSTRACT

The identification of boosted heavy particles such as top quarks or vector bosons is one of the key problems arising in experimental studies at the Large Hadron Collider. In this article, we introduce LundNet, a novel jet tagging method which relies on graph neural networks and an efficient description of the radiation patterns within a jet to optimally disentangle signatures of boosted objects from background events. We apply this framework to a number of different benchmarks, showing significantly improved performance for top tagging compared to existing state-of-the-art algorithms. We study the robustness of the LundNet taggers to non-perturbative and detector effects, and show how kinematic cuts in the Lund plane can mitigate overfitting of the neural network to model-dependent contributions. Finally, we consider the computational complexity of this method and its scaling as a function of kinematic Lund plane cuts, showing an order of magnitude improvement in speed over previous graph-based taggers.

Motivation & Objective

• To develop a robust, high-performance jet tagging method for identifying boosted heavy particles like top quarks and W bosons at the LHC.
• To address the challenge of overfitting in deep learning-based taggers by mitigating reliance on model-dependent Monte Carlo effects in training data.
• To improve computational efficiency by leveraging the hierarchical structure of jet clustering in the Lund plane for faster training and inference.
• To enable better generalization to real experimental data by incorporating kinematic cuts in the Lund plane to suppress non-perturbative contributions.
• To provide a scalable, interpretable framework for jet substructure analysis using graph neural networks with minimal kinematic input.

Proposed method

• Represents each jet as a graph where nodes correspond to Lund declusterings from the CA clustering tree, with kinematic features (Δ and kt) as node features.
• Uses graph convolutional networks (GCNs) to aggregate information across the hierarchical jet clustering structure, enabling effective substructure learning.
• Applies a kt cut in the Lund plane to remove low-kt emissions, reducing graph size and improving robustness to non-perturbative effects and detector smearing.
• Trains two variants: LundNet-3 (3D input: Δ, kt, and jet mass) and LundNet-5 (5D input: adds z and θ variables), enabling comparison of performance vs. complexity.
• Employs a message-passing architecture that leverages the tree-like structure of the Lund decomposition to efficiently propagate information across nodes.
• Reduces computational cost by avoiding nearest-neighbor searches and limiting edges per node, achieving ~10× faster training and inference than prior graph-based taggers.

Experimental results

Research questions

• RQ1 Can graph neural networks applied to the Lund plane representation of jets achieve superior performance in top and W boson tagging compared to existing state-of-the-art methods?
• RQ2 How does the inclusion of kinematic cuts in the Lund plane affect the robustness of graph-based jet taggers to non-perturbative and detector effects?
• RQ3 To what extent can a simplified kinematic input (e.g., LundNet-3) maintain high performance while reducing sensitivity to model-dependent features in Monte Carlo training data?
• RQ4 What is the scaling behavior of training and inference time with respect to the number of Lund declusterings, and can it be optimized via kinematic filtering?
• RQ5 Can the graph-based framework be made significantly faster than previous graph-based taggers without sacrificing tagging performance?

Key findings

• LundNet-5 achieves an AUC of 0.987 for top tagging at pt > 500 GeV, outperforming ParticleNet (AUC 0.983) and significantly improving background rejection at high signal efficiency.
• At εS = 0.5, LundNet-5 achieves a background rejection of 609.8 for top tagging, compared to 500.0 for LundNet-3 and 480.8 for ParticleNet.
• For W tagging at pt > 2 TeV, LundNet-5 achieves a background rejection of 12,500 at εS = 0.5, outperforming ParticleNet’s 8,333.3.
• The inference time per jet scales linearly with the number of Lund declusterings, and is reduced by up to an order of magnitude through kt cuts, with a 4.3 ms pre-processing cost per jet in the current implementation.
• LundNet-3 achieves comparable performance to ParticleNet (AUC 0.904) in quark/gluon discrimination with a 3D input, demonstrating high efficiency with minimal kinematic input.
• Applying kt cuts in the Lund plane reduces overfitting to non-perturbative effects and improves generalization, confirming that kinematic filtering enhances robustness without sacrificing performance.