TL;DR
This paper opens the "black box" of ParticleNet, a state-of-the-art graph neural network (GNN) used for jet tagging at the CERN Large Hadron Collider (LHC). By adapting Layerwise Relevance Propagation (LRP), the authors demonstrate that the network implicitly learns traditional physics observables—specifically, the multi-prong substructure of top quarks—without being explicitly trained on them.
What Problem Does This Paper Address?
In high-energy physics, collision events produce dense sprays of collimated particles called jets. Identifying the origin of these jets (e.g., distinguishing a top quark from lighter quarks or gluons, a task known as jet tagging) is crucial for discovering new physics.
While deep learning models, particularly GNNs like ParticleNet, have achieved state-of-the-art performance by treating jets as particle point clouds, they are notoriously difficult to interpret. This "black box" nature leaves physicists asking a critical question: Is the machine learning model actually learning underlying physical principles, or is it merely exploiting unphysical artifacts in the simulation data? This paper addresses this gap by visualizing and quantifying the internal decision-making process of ParticleNet.
Methods
The authors leverage ParticleNet, which utilizes EdgeConv blocks to dynamically construct graphs using $k$-nearest neighbors ($k\text{NN}$) and learn edge weights between particle nodes.
To interpret the model, they adapt Layerwise Relevance Propagation (LRP), an explainable AI (XAI) technique. LRP works backward from the model's final output prediction, systematically redistributing the prediction score layer-by-layer to assign a "relevance score" ($R$) to individual neurons, nodes, and crucially, the learned edges connecting the particles.
To verify if these relevant edges correspond to physical structures, the authors use a traditional, non-ML physics clustering method called the Cambridge–Aachen (CA) algorithm to independently group the jet particles into three distinct subjets (clusters). They then analyze whether the GNN's most "relevant" edges are bridging these independent subjets.
Datasets
The model was evaluated using the standard benchmark top quark jet tagging dataset:
- Signal: Hadronically decaying top quarks (which characteristically decay into three distinct energetic clusters, or "prongs").
- Background: Quantum chromodynamic (QCD) dijets from light quarks and gluons (typically single-pronged).
- Simulation: Generated at 14 TeV using
PYTHIA 8and fast-detector simulation viaDELPHES, selecting jets with transverse momentum ($p_T$) between 550 and 650 GeV.
Key Insights and Ideas
- The Emergence of Physical Distance: In an untrained, randomly initialized model, the distribution of edge relevance is identical for both top quark and QCD jets. However, as ParticleNet trains, it learns to favor edges spanning larger angular distances ($\Delta R$) specifically for top quarks, reflecting the physical separation of the top quark's decay products.
- Implicit Substructure Learning: By analyzing the final
EdgeConvblock, the authors discovered that the model actively prioritizes edges that connect different subjets in top quark events. This means the GNN is relying on the multi-prong nature of the jet to make its classification, effectively reinventing traditional jet substructure techniques.
Key Results
- High Classification Performance: The trained ParticleNet model achieved an Area Under the Curve (AUC) of 98.23%, converging after 27 epochs.
- Distinct Angular Distributions: For trained top quark jets, the angular distribution ($\Delta R$) of the 5 most relevant edges shows a prominent peak around 0.5, corresponding to the expected separation of the three prongs. In contrast, QCD background jets show no such peak, with relevance concentrated at very low angular separation.
Limitations and Future Work
A primary limitation is that the interpretation relies on the choice of the subjet-finding algorithm (the CA algorithm); a different clustering metric might yield slightly different subjet boundaries. Additionally, the authors note an inherent mathematical trade-off between attribution-based XAI robustness and explanation accuracy.
Future work will focus on mapping these highly relevant edges directly to specific, well-known jet substructure variables, such as $N$-subjettiness, energy correlation functions, and energy flow polynomials, to establish a more rigorous dictionary between GNN representations and analytical physics.
Summary updated