Performance of the ATLAS tau trigger in Run 3 and future prospects
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: High Energy Physics: Experiment. The day in high-energy experiment: detectors, searches and limits.
Kai: We're looking at the paper: "Performance of the ATLAS tau trigger in Run 3 and future prospects".
Mira: Here's the gist of the paper. The paper describes a detailed study of the algorithms and selections employed by ATLAS at the trigger level to select events containing hadronically decaying tau leptons in proton–proton (p p) collisions at sqrt s = 13.6 TeV during Run 3 of the Large Hadron Collider (LHC).
Kai: The selection process utilizes a two-level trigger system: a custom, hardware-based first-level trigger (L1) feeding a Central Trigger Processor (CTP), and software-based algorithms implemented in the High Level Trigger (HLT), which runs AthenaMT.
Mira: Performance studies utilizing data from 2022 and 2023 show that the efficiency curve for a given single tau trigger varies between 90–98% in the plateau region.
Kai: Data-to-simulation correction factors derived from these datasets, enriched with hadronically decaying tau lepton final states, range from 0.95 to 1.05, with an associated uncertainty of 2–3% over a wide range of the transverse momentum (p T) of hadronic tau candidates.
Paper summary: Kai: So, we're looking at this paper right now, "Performance of the ATLAS tau trigger in Run three and future prospects <ref:2611.16454#pg1,Performance of the ATLAS tau trigger in Run 3 and future prospects>." It’s a deep dive into how ATLAS selected those tau leptons that decay into hadrons during the LHC’s third run at fifteen point six TeV.
Mira: The main thing this paper is claiming is they are detailing exactly how they built and tuned the two-level trigger system—the hardware part and the software part—to find these specific signals. It sets out to show what actually worked in that high pile-up environment, which is pretty tough for any detector setup.
Lev: From my side, I'm interested in the engineering aspect here. They’re talking about how this works on real hardware, and they’re looking at things like the Phase-I triggers versus the older ones to see if we actually got better performance there, which is where the real overhead lands.
Kai: Right, so it’s not just a theory piece; it's about what was built and what was measured in those Run three data sets from twenty twenty two and twenty twenty three <ref:2611.16454#pg1>. The core thesis is about the efficiency of these tau triggers in that plateau region, which they say varies between ninety percent and ninety eight percent.
Mira: And the numbers they give for the data-to-simulation correction factors are pretty tight, ranging from zero point nine five to one point zero five with a two to three percent uncertainty across all those different transverse momentum values of the tau candidates. That suggests a very controlled measurement process.
Lev: That uncertainty range, two to three percent over that wide range of transverse momentum, that’s something I think we have to keep in mind when we try to scale this up for real hardware deployment; you can't just assume perfect scaling across the board.
Kai: The paper mentions they used data enriched with those hadronically decaying tau final states, which is a smart way to calibrate their efficiency curves, so they’re not just looking at simulations in a vacuum. It grounds the performance in real physics outcomes.
Paper summary: Mira: They also break down how the L1 trigger works by looking at both non-isolated and isolated tau triggers, and then the High Level Trigger selection is based on a mix of calorimeter information, track information, and identification methods. That gives you a good view of their layered approach.
Lev: When you talk about that L1 non-isolated tau trigger achieving ninety five percent efficiency for one prong or three prong tau leptons with a transverse momentum greater than thirty five or forty GeV, that’s a concrete metric we can actually test against our own simulation models for the hardware.
Kai: They also compared their Phase-I triggers against legacy systems and found that the new system maintained at least the same efficiency as the older ones, but it showed a sharper turn-on curve and about ten percent improvement in the plateau region for those three prong tau candidates.
Mira: That ten percent improvement in the plateau region is significant because it means they managed to squeeze more signal out of the trigger system without drastically increasing their background rejection requirements. It shows iterative refinement actually had a tangible effect on the performance metrics.
Lev: From an error-correction research standpoint, that kind of measurable improvement in efficiency at that level means less data needs to be stored and processed downstream, which is where real hardware savings start to show up.
Kai: They also mentioned further refinements made in Run three like commissioning the Phase-I L1Calo system and using more complex identification methods like a Boosted Decision Tree for physics object identification <ref:2611.16454#pg1>. It shows they weren't just relying on the initial design but actively tuning things as they went.
Mira: And then they brought in Graph Neural Networks at the High Level Trigger stage to further enhance the identification of those tau candidates, even with the tougher environment of Run three pile-up conditions <ref:2611.16454#pg1>. That’s a big step in using machine learning for particle identification right there on the trigger path.
Lev: Using Graph Neural Networks at HLT is interesting from an overhead perspective; it means more computation per event, but if it significantly boosts the signal yield against the background, that trade-off is often worth analyzing when you build real systems.
Paper summary: Kai: So, to summarize what we’ve covered in this first part of the paper about "Performance of the ATLAS tau trigger in Run three and future prospects," it’s a detailed look at how they achieved ninety to ninety eight percent efficiency for single tau triggers in the plateau region using data from twenty twenty two and twenty twenty three <ref:2611.16454#pg1,Performance of the ATLAS tau trigger in Run 3 and future prospects>.
Mira: They establish those correction factors between zero point nine five and one point zero five with a two to three percent uncertainty over the relevant transverse momentum range, showing a very robust measurement technique for this signal.
Lev: The paper is also focused on looking ahead, discussing possible future improvements to how they optimize computing resources and how they can make the system more robust against higher track density, which is where real hardware challenges will definitely come back into play.
Kai: Moving into the conclusion of this paper, we look at the authors and what this all means for us in terms of understanding new physics or just refining our Standard Model measurements. The paper itself is a big piece of work because it documents the actual operational performance under high-luminosity conditions.
Mira: In simple terms, it tells us that their selection algorithms are performing quite well, giving them a solid foundation for what they expect to see in future runs at the LHC. It validates their methodology for finding these tau signals.
Lev: For someone who is building real error correction hardware, this paper shows them the baseline performance of the trigger system they’ll be trying to integrate with, including where those computational demands are landing.
Kai: The implication for us is that this level of tuning and performance validation is crucial before we can even plan out what the High-Luminosity LHC will actually look like in terms of data handling. It sets a benchmark for what’s achievable right now.
Mira: It means they have a clear roadmap based on these results, showing exactly which areas need further development to maintain that high efficiency and low uncertainty we saw here.
Lev: Ultimately, it provides the necessary empirical data to inform the design of future trigger algorithms, making sure they don't just rely on theory but are backed by actual performance numbers from running experiments.
Conclusion: Kai: So, we've been looking at how ATLAS selected those tau leptons in Run three during their fifteen point six TeV runs, and this paper is about their performance of that selection system.
Mira: It boils down to how they tuned the hardware and software triggers to find these specific hadronic tau signals under those high pile-up conditions.
Lev: From my side, I'm looking at what this means for building real error correction hardware; it tells us what kind of signal strength we’re dealing with right now.
Kai: They show that the efficiency for a single tau trigger in the plateau region is between ninety percent and ninety eight percent using data from twenty twenty two and twenty twenty three.
Mira: And they provide those data-to-simulation correction factors, which are between zero point nine five and one point zero five with a two to three percent uncertainty over the relevant transverse momentum range.
Lev: That uncertainty tells us we can’t just assume perfect scaling when we move this to real hardware; that two to three percent spread is what you have to account for when you design your system.
Kai: They also detail how they improved their Phase-I triggers compared to older systems, showing they actually got a sharper efficiency turn-on curve and a ten percent improvement in the plateau region for three prong tau candidates.
Mira: That ten percent boost in the plateau region is significant because it means they managed to squeeze more signal out without making the background rejection requirements drastically worse.
Lev: If you can improve efficiency like that, it directly translates to less data processing overhead, which is exactly what we need when designing systems for high-luminosity environments.
Kai: The authors also describe using new tools like Graph Neural Networks at the High Level Trigger stage to help identify those tau candidates even in the tougher Run three environment.
Mira: That machine learning approach is interesting because it shows they are actively refining identification methods on top of their traditional calorimeter and track based selections.
Lev: Using Graph Neural Networks adds computational load, but if it helps overcome those pile-up issues, that trade-off between computation and signal gain is a key design point.
Kai: So, the authors conclude by proposing specific ways to improve these tau trigger algorithms for what they call the high luminosity conditions of the high luminosity LHC.
Mira: It's basically saying this is a solid baseline performance study, providing an empirical roadmap based on what they actually built and measured in those runs.
Lev: For us, it means we have concrete numbers to work with before we start designing the actual trigger algorithms for the next stage of accelerator operation.
The ATLAS Collaboration
EUROPEAN ORGANISATION FOR NUCLEAR RESEARCH (CERN)
hep-ex
Submitted: 2026-11-12
Updated: 2026-11-12
Comments: 61 pages in total, author list starting in page 44, 25 figures, 8 tables, submitted in JINST. All figures including auxiliary figures are available at https://atlas.web.cern.ch/Atlas/GROUPS/PHYSICS/PAPERS/TRIG-2024-01/
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 90/100
The gist: The paper describes a detailed study of the algorithms and selections employed by ATLAS at the trigger level to select events containing hadronically decaying tau leptons in proton–proton (p p)…
Key concepts
- Two-Level Trigger System
- This is a two-stage selection process: a fast, hardware-based first level (L1) trigger and a more complex software-based second level (HLT). L1 filters data very quickly, while HLT uses advanced algorithms to make finer selections on the remaining events.
- Efficiency Curve
- This measures how well the trigger system selects tau leptons across different energy levels. The study found that for a single tau trigger in the plateau region, efficiency stays consistently between 90% and 98%, indicating reliable performance under specific conditions.
- Data-to-Simulation Correction Factors
- These factors adjust simulated data to match real detector performance. The values range from 0.95 to 1.05, with a small uncertainty of 2–3%. These are essential for accurately converting observed event counts into physical tau lepton yields.
- Graph Neural Networks (GNN)
- These are advanced machine learning algorithms used in the HLT stage to identify hadronic tau candidates. They help improve the selection process, especially in the challenging high-pile-up environment of Run 3, by better recognizing complex event structures.
Terminology
Summary
The paper describes a detailed study of the algorithms and selections employed by ATLAS at the trigger level to select events containing hadronically decaying tau leptons in proton–proton (p p) collisions at sqrt s = 13. p. 2 Keeping the Run 3
(2022–2026) tau trigger rate under control is even more challenging than in Run 1 (2009–2012) or Run 2
(2015–2018) of the LHC, because of the higher production rate of jets from the increase in collision energy
and the higher number of collisions per bunch crossing (pile-up). This paper describes the tau trigger used by ATLAS to select events with hadronically decaying tau leptons
in proton–proton (𝑝 𝑝) collisions at a centre-of-mass energy of √
𝑠 = 13. The performance of the hadronic
tau trigger is presented using proton–proton collision data recorded by the ATLAS experiment
in Run 3 at √
𝑠 = 13. 6 TeV at the Large Hadron Collider. is composed of a custom, hardware-based first-level trigger (L1) followed by software-based algorithms
implemented in the High Level Trigger (HLT). The L1 trigger is a synchronous, pipelined system consisting of a Central Trigger Processor (CTP) fed by
signals coming primarily from dedicated trigger hardware in the calorimeter (L1Calo) and muon (L1Muon)
detector systems.Where the paper says it
Where the paper says it
Where the paper says it
Performance studies utilizing data from 2022 and 2023 show that the efficiency curve for a given single tau trigger varies between 90–98% in the plateau region. p. 0 In the plateau region of the efficiency
curve for a given single tau trigger, efficiency varies between 90–98%. Correction factors
for the simulated tau trigger efficiency in Monte Carlo samples were derived using 2022 and
2023 data enriched with hadronically decaying tau lepton final states, and these range from
0. In the plateau region of the efficiency
curve for a given single tau trigger, efficiency varies between 90–98%. Correction factors
for the simulated tau trigger efficiency in Monte Carlo samples were derived using 2022 and
2023 data enriched with hadronically decaying tau lepton final states, and these range from
0. Correction factors
for the simulated tau trigger efficiency in Monte Carlo samples were derived using 2022 and
2023 data enriched with hadronically decaying tau lepton final states, and these range from
0. 05 with an uncertainty of 2–3% over a wide range of the transverse momentum of the
hadronic tau candidates.Where the paper says it
Where the paper says it
Where the paper says it
The L1 trigger incorporates both non-isolated and isolated tau triggers, and the HLT selections are based on calorimeter-based, track-based, and identification methods. p. 1 2 Isolated tau triggers at L1 10
6 HLT tau trigger 11
6. 1 Calorimeter-based selection 11
6. 1, enriched in 𝑍 → 𝜏𝑒/𝜇𝜏had events where one tau lepton decays into an electron or
muon and the other decays hadronically, with associated neutrinos. The chosen tag-and-probe approach
consists of selecting events triggered by the presence of an electron or muon (tag) that also contain a 𝜏had
candidate (probe) in the final state. The efficiencies of the isolated and non-isolated tau triggers at L1 are evaluated using 𝛾
∗ → 𝜏𝜏 simulated
MC events. The denominator, in the efficiency fraction, includes all particle-level 𝜏had-vis candidates
with 𝑝
vis
T
> 20 GeV and |𝜂| < 2. The expected rates of the single
and di-tau triggers were required to match with that of the legacy triggers, while maximising the selection
efficiency of events with 𝜏had-vis signatures. As described below, the Phase-I system shows an improved
efficiency relative to the Legacy system in the plateau region for the same rate.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Further refinements in Run 3 involved implementing updates to selection algorithms, including the commissioning of the Phase-I L1Calo system and incorporating more sophisticated algorithms such as a Boosted Decision Tree (BDT) for physics object identification. p. 9 2 Isolated tau triggers at L1
To suppress background events and have manageable trigger rates, L1 hadronic tau triggers with low
thresholds (𝐸T ≲ 40 GeV) use a combination of two additional quality criteria: a boosted-decision-tree
(BDT) based algorithm and an energy-based isolation ring algorithm. The BDT-based algorithm for eTAUs, implemented in the eFEX FPGAs, is designed to reject background
events based on the shape of the energy deposits in the SCs that constitute eTAU RoIs. Based on the minimum 𝐸T requirement of each isolated L1 tau
trigger, a selection on the BDT score is applied only on RoIs with 12 GeV < 𝐸T < 50 GeV. For RoIs with
𝐸T > 50 GeV, the L1 trigger rates are low enough that the BDT isolation isolation criteria is not applied. A multivariate 𝑏-tagging
discriminant value is calculated for each jet. The reconstruction and identification procedure for 𝜏had-vis candidates is described in Refs.Where the paper says it
Where the paper says it
Where the paper says it
The final performance metrics confirm that for a given single tau trigger in the plateau region, efficiency is between 90–98%, with data-to-simulation correction factors ranging from 0. p. 0 In the plateau region of the efficiency
curve for a given single tau trigger, efficiency varies between 90–98%. Correction factors
for the simulated tau trigger efficiency in Monte Carlo samples were derived using 2022 and
2023 data enriched with hadronically decaying tau lepton final states, and these range from
0. Correction factors
for the simulated tau trigger efficiency in Monte Carlo samples were derived using 2022 and
2023 data enriched with hadronically decaying tau lepton final states, and these range from
0. 05 with an uncertainty of 2–3% over a wide range of the transverse momentum of the
hadronic tau candidates. The paper is
organised as follows: After a brief overview of the ATLAS detector and LHC operating conditions for
Run 3 in Section 2 and the ATLAS trigger system in Section 3, the details of the data and simulated samples
along with the object reconstruction are described in Section 4. The ATLAS hardware and software tau
trigger algorithms and selection criteria are described in Sections 5 and 6, respectively, while Section 7
details the tau triggers used during Run 3 along with the topological selection requirement and observed
online rates.Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improve event reconstruction by implementing a transformer-based model adapted from GN2 to estimate primary vertex positions before track reconstruction, using information from pixels and strips to predict z position with an associated uncertainty of less than 10 mm. This allows the system to reduce the z width of the first tracking stage RoI based on whether this predicted uncertainty is below a threshold sigma thr.
-
Enhance tau identification robustness by incorporating a prong-inclusive approach for training Graph Neural Networks, which is trained inclusively on the overall sample without any preselection based on track counting. This reduces the number of models to be trained and maintained compared to splitting the sample into zero-prong, one-prong, and multi-prong classes.
-
Improve trigger performance by deploying test chains that modify the tau reconstruction workflow so that the new algorithm is always run on the first stage tracking RoI if its uncertainty value is below a chosen threshold sigma thr. This leads to a 68% decrease in the execution time for track seeding tasks compared to GNTau Reference.
-
Develop an improved performance model by using the extended GNTau model, which has a larger number of inputs (up to 256 attention heads and 4 transformer layers) and is trained on the same samples as the reference model. This extended model shows up to a 5% increase in rejection against light jets compared to the reference GNTau prong-inclusive approach.
-
Optimize computing resources by removing the track-based candidate selection step, incorporating the number of tracks directly into the training of the GNTau algorithm. This moves towards a modern end-to-end approach to tau identification and reduces dependence on track counting which can deteriorate in high pile-up environments at the HL-LHC.
Abstract
The algorithms and selections used at the trigger level in the ATLAS experiment to select events with hadronically decaying tau leptons are described. The performance of the hadronic tau trigger is presented using proton-proton collision data recorded by the ATLAS experiment in Run 3 at sqrt s = 13.6 TeV at the Large Hadron Collider. In the plateau region of the efficiency curve for a given single tau trigger, efficiency varies between 90-98%. Correction factors for the simulated tau trigger efficiency in Monte Carlo samples were derived using 2022 and 2023 data enriched with hadronically decaying tau lepton final states, and these range from 0.95-1.05 with an uncertainty of 2-3% over a wide range of the transverse momentum of the hadronic tau candidates. In addition, potential improvements and their performance for the identification of hadronic tau candidates at trigger level are discussed using data recorded in 2026, having been implemented in selected tau trigger test chains.
Sources
- Evidence for the Higgs-boson Yukawa coupling to tau leptons with the ATLAS detector
- Differential cross-section measurements of Higgs boson production in the $H\to\tau^+\tau^-$ decay channel in $pp$ collisions at $\sqrt{s}=13$ TeV with the ATLAS detector
- Search for charged Higgs bosons produced in top-quark decays or in association with top quarks and decaying via $H^{\pm} \to \tau^{\pm}\nu_{\tau}$ in 13 TeV $pp$ collisions with the ATLAS detector
- Search for direct stau production in events with two hadronic $\tau$-leptons in $\sqrt{s} = 13$ TeV $pp$ collisions with the ATLAS detector
- Search for heavy Higgs bosons decaying into two tau leptons with the ATLAS detector using $pp$ collisions at $\sqrt{s}=13$ TeV
- The ATLAS Trigger System for LHC Run 3 and Trigger performance in 2022
- The ATLAS Experiment at the CERN Large Hadron Collider: A Description of the Detector Configuration for Run 3
- Software and computing for Run 3 of the ATLAS experiment at the LHC
- The ATLAS Inner Detector Trigger performance in pp collisions at 13 TeV during LHC Run 2
- An Introduction to PYTHIA 8.2
- Parton distributions with LHC data
- Emulating the impact of additional proton-proton interactions in the ATLAS simulation by pre-sampling sets of inelastic Monte Carlo events
- Parton Ladder Splitting and the Rapidity Dependence of Transverse Momentum Spectra in Deuteron-Gold Collisions at RHIC
- A comprehensive guide to the physics and usage of PYTHIA 8.3
- EPOS LHC : test of collective hadronization with LHC data
- A New Method for Combining NLO QCD with Shower Monte Carlo Algorithms
- Matching NLO QCD computations with Parton Shower simulations: the POWHEG method
- A general framework for implementing NLO calculations in shower Monte Carlo programs: the POWHEG BOX
- QCD matrix elements and truncated showers
- New parton distributions for collider physics
Related papers
- DrSai: an agentic AI system toward autonomous physics analysis workflows at BESIII
- CANTON- mu Proposal: A Next-Generation Muon g-2 Measurement at Sub-0.1 ppm Precision
- Search for dark photon dark matter near 1,meV using a superconducting mixer
- ANUBIS detector and its sensitivity to neutral long-lived particles
- How to scale your HEP ML models: A recipe for robust architecture comparisons at scale
- Search for the 16 O(ppp) to 13 C pi+ pi+ e+ Decay Mode in Super-Kamiokande Using Machine Learning Techniques