Jet Tagging for Boosted Objects With Machine Learning

September 7, 2026 · 5 min read

If you have spent time working with high-pT analyses, you have probably stared at an event display where a top quark or a W boson decayed into a messy blob of calorimeter deposits rather than cleanly separated jets. That blob is a fat jet, and deciding what produced it is one of the most practically important classification problems in modern collider physics. Jet tagging machine learning methods have become the standard toolkit for this, and getting started is more approachable than it might appear.

Why Boosted Objects Create a Distinct Problem

When a massive particle — a top quark, a Higgs, a W or Z — is produced with transverse momentum much larger than its mass, its decay products are collimated. The angular separation between daughters scales roughly as the mass divided by the transverse momentum, so at high enough boost the daughters land inside a single large-radius jet. This is the fat jet regime.

Boosted jet tagging is then the task of asking: given this single large jet, was it initiated by a top quark, a W boson, a Higgs, or just a QCD quark or gluon? Traditional cut-based taggers use variables like jet mass, N-subjettiness ratios, or energy correlation functions, and they work reasonably well. But they discard a great deal of structure that is present in the full radiation pattern of the jet. Machine learning can exploit that structure.

How Fat Jet Classification ML Is Typically Set Up

Before you build a model, you need to think carefully about your input representation and your training objective.

Choose your jet representation

Fat jets are rich objects. You can represent them as:

• Jet images — a pixelated grid of transverse momentum deposits in the η–φ plane, suitable for convolutional neural networks
• Particle-level lists — the four-momenta of all constituents, which graph neural networks or attention-based models handle naturally
• High-level engineered features — N-subjettiness, soft-drop mass, energy correlation functions, etc., which traditional boosted decision trees digest well

The particle-level approach has become popular because it makes fewer assumptions about which structure matters, but engineered features remain valuable, especially when your training sample is limited or when you want interpretability.

Define the training labels carefully

Your signal and background categories need to reflect what you actually want to separate in data. If you are building a top tagger, your signal is a fat jet genuinely from a hadronic top decay, and your background is QCD multi-jet production. Simulation from Pythia or Sherpa, plus a realistic detector model, provides the training sample. Be thoughtful about which generator settings you use and whether the physics you care about (underlying event, ISR, pile-up) is modeled adequately — the model will learn what you show it.

Watch for mass sculpting

This is the most practically dangerous failure mode. If your input features carry information about jet mass, a classifier trained to separate signal from background can learn to cut on mass implicitly, creating a bump or a dip in the mass distribution of the background. This makes your background estimate unreliable. Decorrelation techniques — adversarial training, distance correlation penalties, or designing mass-insensitive features — exist precisely for this. Think of it as keeping the tagger orthogonal to your signal region variable, the same instinct you would apply when choosing sidebands for a sideband fit.

Training and Evaluation

Use ROC curves the way you already think about purity versus efficiency

Fat jet classification ML performance is almost always reported as a ROC curve of signal efficiency versus background rejection (one minus false-positive rate). That is just your efficiency versus purity trade-off in a different notation. Pick an operating point that matches the needs of your analysis — if you are background limited, you want high rejection; if you are statistically limited, you can afford to loosen.

Validate on independent samples

Split your simulation into training, validation, and test sets and treat the test set as inviolable until you have fully committed to your architecture and hyperparameters. Comparing performance on training and validation sets is exactly the check you would do when looking at a chi-square fit on training data versus held-out data — same instinct, same logic.

Test on alternative generators

Once you have a working tagger, evaluate it on samples produced with a different generator or tune from the one you trained on. Large differences in performance signal that the model has latched onto generator artifacts rather than genuine physics. This is the single most informative robustness check available before you go anywhere near data.

Deploying in an Analysis

Keep the tagger as one discriminant among several. Combining a well-validated ML tagger with a jet mass window or a b-tag requirement is usually more defensible than relying on the tagger alone, because each requirement has a clear physics interpretation. Document what you trained on, what preprocessing you applied, and what your decorrelation strategy was — your collaboration review will thank you.

If you want to build these skills systematically, the complete HEP ML course covers architecture choice, training pipelines, and analysis-level validation in depth. You can also explore the free Module 1 to get a feel for the material before committing.

The one-line takeaway: jet tagging machine learning gives you access to the full radiation pattern of a fat jet, but your biggest job is making sure the model learns physics, not simulator artifacts or mass correlations.

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.