⚛️ high-energy experiments

Development and performance of a transformer-based lepton isolation algorithm with the ATLAS experiment

This paper presents the development, performance evaluation, and calibration of the Prompt Lepton Isolation Tagger (PLIT), a transformer-based algorithm for the ATLAS experiment that significantly improves the identification of prompt electrons and muons in 13.6 TeV proton-proton collisions compared to traditional isolation methods.

Original author: ATLAS Collaboration

Published 2026-11-06

📖 4 min read🧠 Deep dive

CC BY 4.0

Original author: 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 author. For technical accuracy, refer to the original paper. Read full disclaimer

In the high-energy collisions inside the Large Hadron Collider, particles smash together with such force that they create a shower of new matter. Among the debris, physicists look for specific particles called leptons, such as electrons and muons, which are often produced when heavy, unstable particles decay. These "prompt" leptons are valuable because they act as clean signatures of rare, fundamental processes, like the creation of the Higgs boson or the decay of top quarks. However, the collision environment is chaotic. Many other particles, particularly those born from the decay of heavy-flavor hadrons, can mimic these clean leptons. These "non-prompt" leptons are usually born slightly displaced from the main collision point and are surrounded by a cloud of other debris, making them look messy compared to their prompt counterparts. Distinguishing the clean signal from this noisy background is essential for measuring the laws of physics with precision, yet it remains one of the most difficult tasks in the experiment.

A researcher at the ATLAS experiment has developed a new tool to solve this problem, moving beyond traditional methods to use a type of artificial intelligence known as a transformer. This new system, called the Prompt Lepton Isolation Tagger, or PLIT, is designed to tell the difference between a lepton that came directly from a primary collision and one that is a byproduct of a heavier particle's decay. Instead of relying on simple rules that count the amount of energy near a particle, PLIT looks at the entire neighborhood of the lepton. It examines the paths of up to forty charged particles moving near the candidate lepton, analyzing their trajectories, speeds, and how they relate to one another. By processing this complex web of information simultaneously, the algorithm learns to recognize the subtle patterns that indicate whether a lepton is truly isolated or if it is hiding within a cluster of other particles.

The researcher trained this neural network using vast amounts of simulated data from top-quark pair production events, a process that naturally generates both the clean prompt leptons and the messy non-prompt ones the algorithm needs to learn from. They tested the system on data recorded during the 2022 to 2024 run of the collider, where protons collided at a center-of-mass energy of 13.6 tera-electronvolts. The results show that the new algorithm is significantly better at its job than the previous generation of tools. For muons, the new system can identify prompt particles with ten percent higher efficiency while keeping the same level of rejection for background noise. For electrons, the improvement is even more dramatic; in the barrel region of the detector, efficiency gains reach fifty percent, and in the endcap regions, the system doubles the efficiency for the same background rejection. This means physicists can now see more of the rare events they are searching for without being overwhelmed by false signals.

To ensure these results are reliable for real-world physics measurements, the researcher did not stop at simulation. They calibrated the algorithm using real data from collisions where a Z boson decays into two leptons, a process that provides a known sample of prompt particles. By comparing how the algorithm performed on this real data against the simulations, they derived correction factors that allow physicists to adjust their models accurately. The study confirms that the new tagger remains robust even as the number of simultaneous collisions increases, a condition known as pile-up, which can confuse simpler detectors. The work represents a shift in how the ATLAS collaboration handles particle identification, moving from static rules to a dynamic, learning system that understands the context of every particle it encounters. This advancement promises to sharpen the precision of future measurements, allowing scientists to probe the fundamental structure of the universe with greater clarity than ever before.

Technical Summary: Development and performance of a transformer-based lepton isolation algorithm with the ATLAS experiment

Problem Statement
In high-energy physics measurements at the Large Hadron Collider (LHC), particularly those targeting multilepton final states (e.g., $t\bar{t}H$ , $t\bar{t}W$ , four-top production), a critical challenge is distinguishing prompt leptons from non-prompt and misidentified leptons. Prompt leptons originate directly from the decay of electroweak gauge bosons ( $W$ , $Z$ ) or subsequent leptonic decays of $\tau$ -leptons at the primary interaction vertex. In contrast, non-prompt leptons arise primarily from heavy-flavour ( $b$ and $c$ ) hadron decays at displaced secondary vertices, often accompanied by nearby charged-particle tracks. Misidentified leptons and photon conversions further contribute to the background.

Conventional isolation algorithms rely on scalar sums of transverse momentum ( $p_T$ ) from tracks and calorimeter energy deposits within a cone around the lepton candidate. While effective, these methods often struggle at low lepton $p_T$ , where non-prompt backgrounds are abundant. Previous multivariate approaches, such as the Prompt Lepton Improved Veto (PLIV) based on Boosted Decision Trees (BDT), improved discrimination by incorporating jet flavour-tagging information but operated at low prompt lepton efficiencies (below 50%) in low- $p_T$ regions, limiting signal acceptance in precision measurements.

Methodology
The ATLAS Collaboration developed the Prompt Lepton Isolation Tagger (PLIT), a transformer-based neural network designed to overcome the limitations of previous algorithms. PLIT adapts the architecture of the GN2 jet flavour-tagging algorithm, specifically utilizing a transformer encoder to process relationships between input features rather than treating them independently.

Input Features: For each lepton candidate, PLIT processes up to 40 charged-particle tracks reconstructed within a cone of $\Delta R < 0.4$ $Δ R < 0.4$ . The input vector includes:
- Lepton-level variables: Kinematic properties ( $p_T$ , $\eta$ , $\phi$ ) and conventional isolation observables (track-based and calorimeter-based isolation variables).
- Track-level variables: Kinematic properties of associated tracks relative to the lepton, impact parameters ( $d_0$ , $z_0$ ), and detailed silicon detector hit information (pixel and SCT hits, shared hits, split hits, and holes).
Architecture: The network employs a transformer encoder with three stacked layers and eight attention heads. It processes combined lepton and track embeddings to capture correlations between tracks associated with the same lepton candidate.
Training Strategy: The model is trained on simulated $t\bar{t}$ $t \overset{ˉ}{t}$ events, which provide a rich sample of both prompt and non-prompt leptons. The training utilizes a multi-task learning approach:
1. Primary Objective: Classification of the lepton as prompt or non-prompt.
2. Auxiliary Objectives: Classification of individual track origins (e.g., primary vertex, $b$ -hadron decay, $c$ -hadron decay, pile-up) and a vertexing task to predict if pairs of tracks originate from a common vertex. These auxiliary tasks improve the network's representation learning capabilities without being used in the final discriminant.
Implementation: Separate models were trained for muons, barrel electrons ( $|\eta| < 1.37$ ), and endcap electrons ( $1.52 < |\eta| < 2.47$ ). The training samples for Run 3 ( $\sqrt{s} = 13.6$ TeV) are approximately ten times larger than those used for the previous PLIV algorithm.

Key Contributions

Algorithm Development: Introduction of a transformer-based architecture for lepton isolation that exploits detailed track-level information and correlations between tracks, eliminating the dependency on jet reconstruction or jet flavour-tagging tools.
Operating Points (OPs): Definition of new PLIT operating points (Tight and VeryTight) for both electrons and muons. These are defined to either match the efficiency of the previous PLIV algorithm (yielding improved rejection) or match the rejection of PLIV (yielding improved efficiency).
Calibration: Derivation of data-to-simulation correction factors (scale factors) using the standard tag-and-probe method on $Z \to \ell^+\ell^-$ events from Run 3 data (2022–2024).

Results
The performance of PLIT was evaluated using simulated $t\bar{t}$ events and compared against the PLIV algorithm and conventional isolation criteria.

Discrimination Power: PLIT demonstrates significantly improved discrimination power across the entire $p_T$ $p_{T}$ spectrum.
- For muons at low $p_T$ , PLIT achieves a non-prompt lepton rejection 1.5 to 2 times higher than PLIV for a given prompt efficiency.
- For electrons, the rejection improvement is a factor of 2 or more in both barrel and endcap regions, reaching up to a factor of 5 for endcap electrons at low efficiency.
Operating Point Performance:
- When matching the rejection of PLIV, PLIT improves prompt lepton efficiency by approximately 10% for muons, 50% for barrel electrons, and 100% for endcap electrons at low $p_T$ .
- The algorithm shows robustness against pile-up, with prompt efficiency decreasing only mildly and rejection remaining constant as the average number of interactions per bunch crossing increases.
- Performance is stable across different pseudorapidity regions, with PLIT maintaining or improving rejection at higher $|\eta|$ where PLIV performance degrades.
Calibration: Scale factors derived from data are close to unity across the measured $p_T$ range (10–150 GeV). The largest deviations occur at low $p_T$ , where uncertainties increase. Systematic uncertainties are dominated by the modelling of hadronic activity in the simulation, decreasing from ~10% at 10 GeV to ~2% above 40 GeV.

Significance
The paper claims that the PLIT algorithm represents a significant advancement in lepton isolation for the ATLAS experiment. By leveraging transformer architectures and auxiliary learning tasks, PLIT achieves a substantial increase in prompt lepton efficiency while maintaining high rejection of non-prompt backgrounds. This improvement is particularly critical for multilepton measurements where signal acceptance is often limited by isolation requirements at low transverse momentum. The author states that these performance gains will directly benefit physics analyses relying on multilepton signatures, such as measurements of $t\bar{t}H$ , $t\bar{t}W$ , and four-top-quark production, by reducing systematic uncertainties associated with non-prompt lepton backgrounds and increasing the statistical power of the signal samples. The algorithm is now calibrated and available for use in Run 3 physics analyses.

More like this