⚛️ high-energy experiments

Measurement of the $b$ -jet identification efficiency in dileptonic $t\bar{t}$ events using proton-proton collision data at $\sqrt{s}=13.6$ TeV collected with the ATLAS detector

This paper presents the first measurement of the $b$ -jet identification efficiency for the transformer-based GN2 algorithm using 56 fb $^{-1}$ of 13.6 TeV proton-proton collision data collected by the ATLAS detector, demonstrating significant performance improvements over its predecessor and providing essential correction factors for physics analyses.

Original authors: The ATLAS Collaboration

Published 2026-07-07

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: The ATLAS Collaboration

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the Large Hadron Collider (LHC) as the world's most powerful particle smasher. When it smashes protons together, it creates a chaotic explosion of debris, much like shattering a complex clock and watching thousands of gears, springs, and screws fly out in every direction.

The ATLAS experiment is the giant camera and sensor array trying to catch these flying pieces. One of the most important pieces they are looking for is the "bottom quark" (or b-quark). These particles are special because they often hide inside a jet of other particles (called a "jet") and decay very quickly. Finding them is crucial for understanding the universe, but it's like trying to find a specific type of red screw in a pile of thousands of red, blue, and green screws that all look almost identical.

The Problem: The "Needle in a Haystack" Challenge
In this paper, the ATLAS team is testing a new, super-smart computer program called GN2. Its job is to act as a highly trained detective. When a jet of particles flies through the detector, GN2 has to decide: "Is this a jet containing a bottom quark (a 'b-jet'), or is it just a regular, boring jet made of lighter stuff?"

Previously, the detectives used older methods (like DL1d). They were good, but the new GN2 detective is based on a "Transformer" architecture (the same kind of AI technology behind modern chatbots). This allows it to look at the complex relationships between the particles inside the jet, rather than just checking them one by one.

The Test: A Controlled Crime Scene
To see how good this new detective is, the scientists didn't just look at random debris. They set up a specific "crime scene" where they knew bottom quarks had to be present.

They looked for a very specific event: a collision that produces a pair of "top quarks." Top quarks are the heaviest particles in the Standard Model, and they almost always decay into a bottom quark and a W-boson. If the W-boson then turns into an electron and a muon (two different types of "leptons," or light particles), the scientists know for a fact that two bottom quarks must be hiding in the jets of that event.

It's like finding a crime scene where two specific suspects (the bottom quarks) are guaranteed to be there. The scientists then asked: "Can our new AI detective correctly identify the jets containing these two suspects?"

The Results: A Major Upgrade
The paper reports on data collected in 2022 and 2023 at a record-breaking energy level (13.6 TeV). Here is what they found:

Superior Performance: The new GN2 detective is significantly better than the old one.
- The Analogy: Imagine the old detective could correctly spot the red screw 80% of the time, but occasionally mistook a blue screw for a red one. The new GN2 detective can still spot the red screw 80% of the time, but it is twice as good at ignoring the blue screws and three times as good at ignoring the green screws. It makes far fewer mistakes.
Calibration (The "Scale Factor"): Even though the computer simulation predicted how well GN2 should work, real-world data is messy. The scientists measured the actual efficiency in the real data and compared it to the simulation. They found that the simulation was slightly off, so they created "correction factors" (numbers between 0.9 and 1.3) to adjust the simulation so it matches reality perfectly.
Precision: For jets with high energy (above 60 GeV), the measurement is incredibly precise, with an uncertainty of only about 1%. This is like measuring the length of a football field and being off by only the width of a human hair.

Why This Matters (According to the Paper)
The paper states that this measurement is the "calibration" needed for all future physics analyses at ATLAS. Because GN2 is so much better at telling the difference between a bottom-quark jet and a regular jet, scientists can now:

Find the Higgs boson decaying into bottom quarks more easily.
Search for new, rare particles that might also decay into bottom quarks.
Study the top quark with greater accuracy.

In short, the ATLAS team has successfully upgraded their "bottom-quark detector" with a smarter AI. They have proven it works better than the previous generation, measured exactly how well it works in the real world, and provided the necessary tools for other scientists to use this improved detector in their own research.

Technical Summary: Measurement of the $b$ -jet identification efficiency in dileptonic $t\bar{t}$ events using proton–proton collision data at $\sqrt{s}= 13.6$ TeV collected with the ATLAS detector

Problem and Context
Jet flavour tagging is a critical component of the ATLAS physics programme at the Large Hadron Collider (LHC), essential for identifying jets containing $b$ -hadrons ( $b$ -jets) and $c$ -hadrons ( $c$ -jets) while rejecting the substantial background of light-flavour jets. This capability underpins precision measurements of the Standard Model (SM), searches for Higgs boson decays, top-quark physics, and beyond-SM searches. While previous generations of tagging algorithms utilized Deep Neural Networks (DNN) and Recurrent Neural Networks (RNN), the ATLAS Collaboration has adopted GN2, a transformer-based machine-learning algorithm, for Run 3 data analysis. GN2 offers significant performance improvements over its predecessor, DL1d, by predicting jet flavour probabilities ( $p_b, p_c, p_\tau, p_u$ ) and defining a discriminant $D_{GN2}$ . However, to utilize GN2 in physics analyses, its performance must be calibrated against real data to derive correction factors (Scale Factors, SF) that account for discrepancies between data and Monte Carlo (MC) simulation. This paper addresses the measurement of the GN2 $b$ -jet tagging efficiency using data collected in 2022 and 2023 at a centre-of-mass energy of $\sqrt{s}= 13.6$ TeV.

Methodology
The measurement utilizes a dataset corresponding to an integrated luminosity of $56.0 \pm 1.1 \text{ fb}^{-1}$ . The analysis selects a sample enriched in dileptonic top-quark pair ( $t\bar{t}$ ) events, characterized by the decay chain $t \to Wb$ followed by $W \to \ell\nu$ (where $\ell$ is an electron or muon). The event selection requires exactly two isolated leptons of different flavours and opposite charges, and exactly two hadronic jets.

The core of the methodology is a simultaneous extraction of the $b$ -jet tagging efficiency and the jet flavour composition using a maximum likelihood approach. The selected events are categorized into:

Kinematic Bins: Defined by the transverse momentum ( $p_T$ ) of the leading ( $T_m$ ) and subleading ( $T_n$ ) jets, with boundaries at [20, 40, 60, 140, 400] GeV.
Pseudo-Continuous $b$ -jet Tagging (PCBT) Bins: Exclusive ranges of the $D_{GN2}$ discriminant corresponding to cumulative efficiency intervals: [100%, 90%], [90%, 85%], [85%, 77%], [77%, 70%], [70%, 65%], and [65%, 0%].
Analysis Regions: Events are further divided into a Signal Region (SR) enriched in $b\bar{b}$ events (requiring jet-lepton invariant masses $m_{j\ell} < 175$ GeV) and three Control Regions (CRs) enriched in $b\ell$ , $\ell b$ , and $\ell\ell$ events (where $\ell$ denotes non- $b$ jets).

A likelihood function is constructed to fit the observed event yields in these regions. The fit simultaneously determines the $b$ -jet tagging efficiencies ( $P_b$ ) in each PCBT and $p_T$ bin, as well as jet flavour composition correction factors ( $c_{bb}, c_{bl}, c_{lb}, c_{ll}$ ) that adjust the MC flavour composition to match data. Systematic uncertainties are evaluated by varying detector response parameters (lepton selection, jet energy scale/resolution, pile-up modelling) and physics modelling parameters (parton distribution functions, QCD radiation, parton shower/hadronisation models, and $t\bar{t}$ vs. single-top interference schemes).

Key Contributions

First Data-Driven Calibration of GN2: This work presents the first measurement of the GN2 algorithm's performance in ATLAS Run 3 data, establishing the necessary Scale Factors for physics analyses.
Simultaneous Extraction: The analysis employs a robust simultaneous fit to extract both tagging efficiencies and flavour compositions, reducing sensitivity to mis-modelling of the jet flavour composition in simulation.
Comprehensive Uncertainty Budget: The study provides a detailed breakdown of uncertainties, distinguishing between statistical limitations (data and MC) and systematic sources (detector and physics modelling), specifically for the new transformer-based architecture.

Results

Performance Improvement: The GN2 algorithm demonstrates significant performance gains over DL1d, achieving up to a factor of two higher rejection of light-flavour jets and a factor of three higher rejection of charm-flavour jets at the same $b$ -tagging efficiency.
Scale Factors: The derived simulation-to-data Scale Factors range from 0.9 to 1.3 across the measured $p_T$ and PCBT intervals.
Precision: For jets with transverse momentum $p_T > 60$ GeV in the [65%, 0%] cumulative efficiency interval, the total uncertainty on the measured efficiency is approximately 1%. In the lowest $p_T$ bin (20–40 GeV) and the highest efficiency bin ([100%, 90%]), uncertainties are larger (up to ~25%), dominated by jet energy scale/resolution uncertainties and physics modelling.
Uncertainty Sources: In the medium $p_T$ region (60–140 GeV), the total uncertainty is dominated by systematic effects, particularly $t\bar{t}$ modelling and detector response. At high $p_T$ ( $>140$ GeV), the data statistical uncertainty becomes the dominant contribution.
Flavour Composition: Post-fit correction factors for jet flavour compositions are reported, ranging from 0.7 to 1.3, indicating the adjustments required to align MC predictions with the observed data.

Significance
The paper validates the outstanding $b$ -jet tagging performance of the GN2 algorithm during Run 3. By providing the central values of the Scale Factors, their uncertainties, and the correlation scheme across $p_T$ and PCBT bins, this work enables ATLAS physics analyses to correctly interpret $b$ -tagged jets in diverse search and measurement channels. The results confirm that the transformer-based GN2 algorithm maintains high efficiency and rejection power in the high-luminosity, high-pileup environment of the LHC, supporting the precision physics goals of the ATLAS experiment.

More like this