B-Tagging Algorithms Powered by Machine Learning

October 14, 2026 · 5 min read

B-tagging is one of those problems that makes you appreciate why particle physicists were among the first to embrace machine learning. You have a detector full of overlapping tracks, a handful of displaced vertices, and a handful of nanoseconds to decide whether a jet came from a bottom quark. Classical cut-based approaches worked, but they left a lot of discriminating power on the table. Modern b-tagging machine learning methods recover much of that power — and understanding how they work will help you apply the same ideas to your own jet-level analyses.

Why B-Jet Identification Is Hard

A bottom quark hadronizes before it decays, producing a B meson or baryon that travels a measurable distance — typically a few millimeters — before producing a secondary vertex. The signature is real, but it competes with tracking imperfections, pileup, and the combinatorial mess inside a busy jet.

Traditional approaches built discriminants from a small number of hand-engineered quantities: impact parameter significance, secondary vertex mass, number of tracks attached to the displaced vertex, and so on. These variables carry genuine physics, but combining them optimally by hand is difficult. That is exactly the kind of problem a learned classifier excels at.

From Cut-Based to MVA to Deep Networks

The evolution of b-jet identification ML mirrors the broader arc of machine learning in HEP:

Step 1 — Multivariate discriminants on engineered features

The first generation of modern b-taggers fed those hand-crafted variables (impact parameters, vertex properties, jet kinematics) into boosted decision trees or neural networks. This was already a substantial improvement over sequential cuts, for the same reason a multivariate likelihood beats a series of one-dimensional cuts: it exploits correlations. If you have used a BDT for event-level classification in your own analysis, you have already done something structurally similar.

Step 2 — Operating directly on track sequences

The real leap came when algorithms stopped summarizing the jet and started reading the raw track list directly. Instead of pre-computing a secondary vertex mass and handing it to a classifier, you feed every reconstructed track — its transverse and longitudinal impact parameter significances, its momentum, its quality flags — into a network that learns its own internal representation.

The natural architecture for ordered or variable-length sequences is a recurrent network (LSTM or GRU), but attention-based transformers have become increasingly common because they handle long-range dependencies between tracks more gracefully. The key idea is the same: let the network discover which combinations of track-level observables are discriminating, rather than hard-coding your physics intuitions.

Step 3 — Adding secondary vertex and jet-level context

A complete modern b-tagger typically runs multiple sub-networks in parallel:

• A track-level network that processes each track independently, then aggregates (by summing, averaging, or attention pooling).
• A vertex-level network that processes reconstructed secondary vertex candidates.
• A jet-level network that combines the outputs of the two above with global jet kinematics.

The outputs are then merged in a final dense network. This hierarchical structure mirrors the physics: tracks live inside jets, vertices are built from tracks, and the jet-level context (pT, η) sets the operating regime. If you are familiar with how different sub-detectors provide complementary information, you will recognize the same logic here.

What This Means for Your Working Vocabulary

When you read b-tagger documentation or papers, a few terms appear constantly:

Working point — a threshold on the tagger output score, chosen to fix the light-jet rejection at some operating value. This is exactly your efficiency-versus-purity trade-off, expressed as a ROC curve.

Charm rejection — modern taggers are often trained to separate b-jets from both light jets and c-jets simultaneously, because c-jets are the main irreducible background. This is a multiclass problem, and the output is a softmax over three categories.

Scale factors — because the tagger is trained on simulation, its efficiency in data differs from simulation by some factor that must be measured in control regions (typically tt̄ events) and applied as a weight. This is the familiar simulation-to-data correction you already know from cut-based analyses; the ML tagger does not eliminate the need for it.

Getting Practical in Your Own Analysis

If your analysis depends on b-tagging, you are almost certainly consuming a pre-trained tagger from your experiment's reconstruction framework rather than training your own. What you control is the working point and the associated scale factors. Choose the working point by reasoning about your signal-to-background needs — tighter working points give you cleaner samples but smaller yields — and propagate the scale-factor uncertainties through your analysis as systematic nuisance parameters.

If you want to go deeper — training your own jet classifiers, understanding attention pooling, or building hierarchical architectures for other jet substructure problems — the full HEP ML course covers graph networks and sequence models applied to jets in detail. You can also get started with the free Module 1 to see whether the material fits where you are right now. Related ideas around jet classification and multivariate analysis design appear throughout the course.

Deep learning b-tagging in HEP works because it respects the detector geometry, the track-level granularity, and the physics of heavy-flavor decay — and that same principle, match the architecture to the data structure, applies everywhere else in your analysis.

References

Radovic, A., Williams, M., Rousseau, D., Kagan, M., Bonacorsi, D., Himmel, A., Aurisano, A., Terao, K., & Wongjirad, T. (2018). Machine learning at the energy and intensity frontiers of particle physics. Nature, 560, 41-48.