Quark vs Gluon Jet Discrimination With Machine Learning

September 25, 2026 · 5 min read

Quark-gluon discrimination is one of those problems that looks straightforward on a whiteboard and humbling in practice. The physics is clean: gluon jets are broader, carry more hadrons, and arise from a higher-color-charge parent than quark jets. Yet turning those differences into a reliable, simulation-independent discriminant is genuinely hard. If you are trying to build or evaluate a quark vs gluon jet classification model for the first time, this guide walks you through the key steps and the pitfalls worth anticipating.

Why the Problem Is Harder Than It Looks

The intuition is solid. Gluons couple more strongly to the QCD vacuum, so their jets are wider and produce higher particle multiplicities than quark jets at the same transverse momentum. That gives a classifier real signal to work with.

The difficulty is that "quark jet" and "gluon jet" are not perfectly defined quantum-mechanically — the assignment depends on the process, the parton shower, and the fragmentation model. Samples of pure quark or gluon jets are not available in data; you always get mixtures, and the mixture fractions depend on kinematics. Simulation-to-data agreement for multiplicity and substructure variables is also notoriously imperfect, which means a model trained on Monte Carlo can learn artifacts of the generator rather than genuine physics. These issues are not obstacles to doing the analysis — they are the analysis, in many ways.

Step 1: Choose Your Input Observables Carefully

Before touching any model, decide what your classifier will see. For quark gluon jet tagging ML, common choices fall into three levels of abstraction:

High-Level Hand-Crafted Features

Variables like charged-particle multiplicity, jet width, and fragmentation functions (such as the Les Houches angularity) are interpretable and tie directly to the color-charge physics. Starting here is good practice: it forces you to understand what the model is learning, and a simple boosted decision tree (BDT) on a handful of such observables already performs well.

Low-Level Particle Lists

Feeding the full set of particle four-momenta to a deep network removes the need to hand-engineer features, but adds a modeling burden. The network can learn whatever combination of observables discriminates best in simulation, including ones that are poorly modeled. Interpret gains over the high-level baseline with care.

Jet Images and Graph Representations

Representing a jet as a pixelated image in (η, φ) or as a particle-flow graph captures spatial correlations the particle list alone misses. These architectures are powerful and have become standard in the field. They are also more opaque, making it harder to diagnose where a performance difference is coming from.

Step 2: Build Your Training Sample Thoughtfully

Because pure quark or gluon samples do not exist in real data, training almost always happens on Monte Carlo, using generator-level parton labels to define the target classes. A few practical notes:

• Match kinematics between classes. Jet pT and η distributions differ between quark- and gluon-enriched processes. If you do not reweight or select carefully, the model can exploit kinematic differences rather than substructure. This is the quark-gluon discrimination analogue of learning the wrong observable.
• Test on a second generator. If your model's ranking of events changes substantially when you switch from, say, Pythia to Herwig, it has learned something generator-specific. Robustness across generators is a minimal sanity check before you trust the discriminant on data.
• Use sideband enrichment to access data. Dijet events at different rapidity intervals or different process selections give you samples with different quark/gluon fractions. Fitting the classifier output distribution in these sidebands is how you can partially validate performance without true labels.

Step 3: Evaluate Beyond the ROC Curve

The receiver operating characteristic — efficiency versus purity in physics language — is the standard figure of merit for quark gluon jet tagging ML. Report it, but do not stop there.

Check the working point behavior: does the discriminant sculpt the jet mass distribution? A classifier that creates a bump in an uninstrumented variable is a serious problem for searches. Decorrelation techniques exist and are worth applying.

Also compare your deep model to a well-tuned BDT on the high-level variables. If the gap is large, think about whether you understand why. If the gap is small, the simpler model may be preferable for systematics reasons.

Step 4: Quantify and Propagate Uncertainties

Every input variable carries a modeling uncertainty. For quark gluon discrimination, multiplicity and track-based variables are often the most sensitive and the least well-modeled. Estimate the impact of each by varying the generator tune or the fragmentation parameters within their uncertainties, propagating the effect through to your final discriminant output. This is more work than it sounds, but it is what turns a classifier into a measurement-ready tool.

If you want a structured introduction to building and validating classifiers like this from scratch, the full HEP ML course covers the complete workflow from feature engineering through uncertainty quantification. You can also get a taste with the free introductory module before committing.

For background on how deep learning fits into the broader LHC physics toolkit, the course blog has accessible overviews of jet tagging architectures and data-versus-simulation validation strategies that complement the steps above.

Quark vs gluon jet classification is a microcosm of everything that makes HEP machine learning interesting: the physics is real, the labels are imperfect, and simulation dependence is always lurking — get the validation right, and the discriminant becomes genuinely useful.

References

Guest, D., Cranmer, K., & Whiteson, D. (2018). Deep Learning and its Application to LHC Physics. Annual Review of Nuclear and Particle Science, 68, 161-181. arXiv:1806.11484.