Deep Sets and event-level maximum-likelihood estimation for fast pile-up jet rejection at the ATLAS High Level Trigger
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: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "Deep Sets and event-level maximum-likelihood estimation for fast pile-up jet rejection at the ATLAS High Level Trigger".
Mira: Here's the gist of the paper. Deep Sets and event-level maximum-likelihood estimation for fast pile-up jet rejection in ATLAS introduces a novel uncertainty-aware jet regression model called DIPz and an event-level discriminant called MLPL to efficiently reject pile-up jets in multijet final states at the ATLAS High-Level Trigger Multiple proton–proton collisions occur at every bunch crossing at the LHC, with the mean number of interactions reaching about 64 during Run 3 and expected to rise to around 200 at the High-Luminosity LHC.
Kai: Events with multijet signatures will appear at increasingly high rates, making it crucial to efficiently group jets according to their origin along the beamline, particularly at the trigger level.
Mira: The paper presents a novel approach for fast and efficient pile-up jet rejection in ATLAS.
Kai: DIPz Architecture The DIPz algorithm follows the Deep Sets framework, which is well suited for variablelength unordered inputs such as tracks within a jet.
Paper summary: Kai: So to wrap up this paper, "Deep Sets and event-level maximum-likelihood estimation for fast pile-up jet rejection at the ATLAS High Level Trigger," they put together a Deep Sets based regression model called DIPz and an event level discriminant called MLPL.
Mira: The core claim is that this combination provides a robust way to predict jet vertex positions with uncertainty, and then uses those predictions to build an event score that efficiently rejects pile-up jets in multijet final states.
Lev: For me, it’s about making sure that the overhead from this pre-selection stage doesn't become a bottleneck when you try to run it on actual hardware at the High Level Trigger <ref:2512.10819#pg2>.
Kai: It seems like they’ve managed to fit this into the existing ATLAS Athena framework, and it runs with negligible additional CPU cost compared to what's already there.
Mira: The implication is that this DIPz plus MLPL algorithm is ready for use in future HL-LHC operations, offering a scalable and low latency solution for targeting multijet signatures <ref:2512.10819#pg1>.
Lev: It’s good to hear it’s already being used in collision data taking right now, which shows the practical feasibility of this approach for real experiments.
Kai: So, we've seen how they combine deep learning structures with statistical likelihood methods to solve a very specific, high-throughput problem in particle physics.
Conclusion: Kai: So we've been looking at how ATLAS used this new method, DIPz and MLPL, to filter out pile-up jets in those high-energy multijet collisions.
Mira: Exactly. This paper focuses on using a deep sets approach for jet vertex prediction with uncertainty, and then building an event level score called MLPL to reject the pile-up background at the trigger stage.
Lev: And what's interesting is how they combine those two ideas: the neural network part for predicting where the jets are, and then this maximum likelihood estimation discriminant that looks at all those predictions together.
Kai: It seems like they’ve managed to make this whole system fit into the existing ATLAS Athena framework, which means it could actually be implemented pretty quickly in real hardware.
Mira: That's a big deal because if you can get something like this running with low overhead, it opens the door for much faster selections on the High-Luminosity LHC data.
Lev: Low overhead is key. If this adds too much processing time to the trigger chain, it just slows everything down, so they’re showing that it doesn't take up significant new CPU cycles compared to what we already have in place for tracking.
Kai: The authors of this work are using simulated t bar events from the Powheg and Pythia eight generators at fifteen point six tera electron volts to train their models.
Mira: That simulation setup is standard, but what they're doing with the Deep Sets structure is unique because it handles those variable length inputs—like the tracks inside a jet—in a way that respects how they are unordered.
Lev: From my side, I’m thinking about how this uncertainty aware modeling helps. Predicting not just the position of a jet vertex, but also its standard deviation, that gives us a much better handle on how reliable that prediction actually is.
Kai: So what does this mean for the bigger picture? It shows a way to build these complex rejection systems using deep sets and likelihood methods without needing massive, slow training datasets or huge computational resources upfront.
Mira: It’s about making the selection process itself smarter, event by event, rather than relying on simple cuts that might miss real signal jets or let too much background through.
Lev: It's a practical tool for the future. For us building actual error correction systems, this is an example of how you can integrate machine learning and statistical estimation into a fast pipeline without blowing the budget on experimental time to test every possible configuration.
Kai: So, it boils down to having a robust method that can distinguish between real jets originating from the hard physics process and those just being background interactions from the pile-up.
Mira: Right. This work demonstrates how combining deep learning for geometry with event level likelihood estimation is a viable path for fast and reliable jet rejection.
Lev: And it sets up a blueprint for how we might approach building future real hardware trigger systems where you have to make decisions in milliseconds based on complex inputs.
ATLAS Collaboration · Southern Methodist University
hep-ex, cs.LG
Submitted: 2025-12-11
Updated: 2026-11-04
Comments: 6 pages, 2 figures. Contribution to the 43rd International Conference on High Energy Physics (ICHEP 2026), on behalf of the ATLAS Collaboration
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
Importance score: 78/100
The gist: Deep Sets and event-level maximum-likelihood estimation for fast pile-up jet rejection in ATLAS introduces a novel uncertainty-aware jet regression model called DIPz and an event-level discriminant…
Key concepts
- Deep Sets
- Deep Sets is a machine learning framework designed to handle variable-length, unordered inputs, such as the tracks within a jet. It processes these inputs through shared layers to create representations that are summed up. This ensures the model's output remains consistent regardless of the order in which the input elements are presented.
- DIPz
- DIPz is an uncertainty-aware regression model built on Deep Sets. It takes jet features and track information, processes them to predict both the mean ($μ_z$) and standard deviation ($σ_z$) of a Gaussian distribution describing the jet's position along the beamline. This uncertainty estimation is key for robust pile-up rejection.
- MLPL Discriminant
- The Maximum Log Product of Likelihoods (MLPL) is an event-level discriminant that combines predictions from DIPz across multiple jets. It maximizes the product of likelihoods based on the predicted jet positions, allowing it to effectively separate signal events from pile-up backgrounds during trigger selection.
- Pile-up Jet Rejection
- Pile-up refers to extra proton–proton interactions occurring simultaneously with the main collision. These interactions create additional jets that can confuse analyses. The MLPL system is specifically designed to identify and reject these unwanted pile-up jets quickly, which is essential for efficiently selecting true physics events.
Terminology
Summary
Deep Sets and event-level maximum-likelihood estimation for fast pile-up jet rejection in ATLAS introduces a novel uncertainty-aware jet regression model called DIPz and an event-level discriminant called MLPL to efficiently reject pile-up jets in multijet final states at the ATLAS High-Level Trigger
Motivation for the Research
Multiple proton–proton collisions occur at every bunch crossing at the LHC, with the mean number of interactions reaching about 64 during Run 3 and expected to rise to around 200 at the High-Luminosity LHC. Events with multijet signatures will appear at increasingly high rates, making it crucial to efficiently group jets according to their origin along the beamline, particularly at the trigger level. p. 0 As a direct consequence, events with multijet signatures will appear at
increasingly high rates. To cope with the higher luminosity, efficiently grouping jets according to
their origin along the beamline is crucial, particularly at the trigger level. Used primarily for fast 𝑏-tagging [2], this super-ROI tracking information can also be exploited
to identify and reject pile-up jets early in the trigger selection, improving the efficiency of true
multi-jet signal event reconstruction. This work presents a novel approach for fast and efficient pile-up jet rejection in ATLAS.Where the paper says it
Where the paper says it
DIPz Architecture
The DIPz algorithm follows the Deep Sets framework, which is well suited for variablelength unordered inputs such as tracks within a jet. Each track is processed through shared feedforward layers, producing latent representations that are summed to ensure permutation invariance. p. 1 DIPz for Uncertainty-Aware Regression
The DIPz algorithm follows the Deep Sets framework [3], which is well suited for variablelength unordered inputs such as tracks within a jet. Each track is processed through shared feedforward layers, producing latent representations that are summed to ensure permutation invariance. Each track is processed through shared feedforward layers, producing latent representations that are summed to ensure permutation invariance. These are merged with a small set of jet-level features (jet 𝑝T and 𝜂), and the combined representation
is passed to a final network that predicts the mean (𝜇𝑧) and standard deviation (𝜎𝑧) of a Gaussian
describing the jet-vertex position along the beamline. The training target corresponds to the
true 𝑧-position of the collision vertex from which the jet originated. The model is trained using simulated 𝑡𝑡¯events at √
𝑠 = 13. The model is trained using simulated 𝑡𝑡¯events at √
𝑠 = 13. 6TeV, generated with Powheg+Pythia 8
[4] [5], with pile-up overlaid to emulate Run-3 conditions [6].Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
MLPL Discriminant Construction
The event-level discriminant, the Maximum Log Product of Likelihoods (MLPL), is constructed by combining the DIPz per-jet predictions. For a given n-jet final state selection, a cut is optimized and applied to this discriminant to reject pile-up-dominated events efficiently. p. 3 MLPL: The Fast Rejection Algorithm
This approach relies on constructing an event-level discriminant based on the per-jet vertex
predictions provided by DIPz. For a given 𝑛-jet final state selection, a cut is optimized and applied
to this discriminant to reject pile-up-dominated events efficiently, enabling fast decision-making
within tight latency constraints faced at the ATLAS trigger system. When evaluating 𝐿(𝑧ˆ; 𝐽
(𝑚)
𝑛 ) for each group of jets
𝐽
(𝑚)
𝑛 , the jet group corresponding to the group of n-jets coming from the primary vertex (collision
of interest) is expected to have the largest value compared to the other n-jet groups. Following this
logic, the Maximum Log Product of Likelihoods (MLPL) discriminant is introduced as:
MLPL(𝑛, 𝑚) = max
𝐽
(𝑚)
𝑛
max
𝑧
log 𝐿(𝑧; 𝐽
(𝑚)
𝑛 )
, (4)
where:
• m is the number of the highest-𝑝𝑇 jets considered for combinations in the event (i.Where the paper says it
Where the paper says it
Performance Evaluation
The performance of the MLPL discriminant is evaluated using a validation sample of simulated t¯events. Receiver Operating Characteristic (ROC) curves are constructed by scanning thresholds on the discriminant and computing the corresponding signal efficiency and background rejection. p. 4 Running the algorithm over the entire signal and background datasets (∼ 25𝑘 events each),
MLPL distributions are obtained for multiple MLPL(𝑛, 𝑚) configurations. Then, Receiver Operating Characteristic (ROC) curves are constructed by scanning thresholds on the discriminant and
computing the corresponding signal efficiency and background rejection (Figure 3). The comparison between configurations—varying the number of jets entering the likelihood (𝑛 = 2 − 4) and
the number of highest-𝑝T jets considered (𝑚 = 5, 6, all)—shows that larger 𝑛 values impose stricter
topological requirements, yielding stronger signal–background separation, while reducing 𝑚 limits
the combinatorial complexity with only a modest impact on performance. Results show that for a
four-jet trigger selection, the pre-selection rejection provided for a signal efficiency of 80% allows
to reduce the input rate to the later CPU-intensive selection steps by a factor of four, reducing CPU
needs.Where the paper says it
Where the paper says it
Conclusion and Implementation
The presented approach combines an uncertainty-aware Deep Sets-based regression model, DIPz, for jet-vertex prediction with an event-level likelihood discriminant, MLPL(n,m), to enable fast and robust pileup jet rejection. Both components are fully implemented in the ATLAS Athena framework and integrated into the High-Level Trigger, where they run with negligible additional CPU cost compared to existing tracking stages. p. 4 The presented
approach combines an uncertainty-aware Deep Sets-based regression model, DIPz, for jet-vertex
prediction with an event-level likelihood discriminant, MLPL(n,m), to enable fast and robust pileup jet rejection. Both components are fully implemented in the ATLAS Athena framework [11]
and integrated into the High-Level Trigger, where they run with negligible additional CPU cost
compared to existing tracking stages, effectively providing a near cost-free enhancement. Both components are fully implemented in the ATLAS Athena framework [11]
and integrated into the High-Level Trigger, where they run with negligible additional CPU cost
compared to existing tracking stages, effectively providing a near cost-free enhancement. Deployed
online at Point 1, the DIPz + MLPL algorithm is already used in ATLAS collision-data taking,
optimizing trigger chains targeting multijet signatures in 2025 and offering a scalable, low-latency
solution for future HL-LHC operations.Where the paper says it
Where the paper says it
References
[1] ATLAS Collaboration. The ATLAS Experiment at the CERN Large Hadron Collider. p. 5 DIPz and MLPL for Fast Pile-Up Jet Rejection in ATLAS Mohammed Aboelela
References
[1] ATLAS Collaboration. The ATLAS Experiment at the CERN Large Hadron Collider.Where the paper says it
[2] ATLAS Collaboration and G. Aad et al. Fast b-tagging at the high-level trigger of the ATLAS experiment in LHC Run 3. p. 5 [2] ATLAS Collaboration and G. Fast b-tagging at the high-level trigger of the ATLAS
experiment in LHC Run 3.Where the paper says it
[3] Patrick T. Komiske, Eric M. p. 5 [3] Patrick T. Komiske, Eric M. Komiske, Eric M. Metodiev, and Jesse Thaler. Metodiev, and Jesse Thaler. Energy flow networks: deep sets for
particle jets.Where the paper says it
Where the paper says it
Where the paper says it
[4] Stefano Frixione, Paolo Nason, and Carlo Oleari. Matching NLO QCD computations with parton shower simulations: The POWHEG method. p. 5 [4] Stefano Frixione, Paolo Nason, and Carlo Oleari. Matching NLO QCD computations with
parton shower simulations: The POWHEG method.Where the paper says it
[5] Torbjörn Sjöstrand, Stephen Mrenna, and Peter Skands. A brief introduction to PYTHIA 8. p. 5 [5] Torbjörn Sjöstrand, Stephen Mrenna, and Peter Skands. A brief introduction to PYTHIA 8.Where the paper says it
[6] ATLAS Collaboration. Emulating the impact of additional proton–proton interactions in the ATLAS simulation by presampling sets of inelastic Monte Carlo events. p. 5 [6] ATLAS Collaboration. Emulating the impact of additional proton–proton interactions in the
ATLAS simulation by presampling sets of inelastic Monte Carlo events.Where the paper says it
[7] François Chollet et al. Keras. https://keras. p. 5 [7] François Chollet et al. https://keras.Where the paper says it
[8] Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean et al. p. 5 [8] Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro,
Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz
Kaiser, Manjunath Kudlur, Josh Levenberg, Dandelion Mané, Rajat Monga, Sherry Moore,
Derek Murray, Chris Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever,
Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol
Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng.Where the paper says it
[9] Diederik P. Kingma and Jimmy Ba. p. 5 [9] Diederik P. Kingma and Jimmy Ba.Where the paper says it
[10] ATLAS Collaboration. Deep sets and Event-level Maximum (Jet Trigger Public Results). p. 5 [10] ATLAS Collaboration. Deep sets and Event-level Maximum (Jet Trigger
Public Results). Deep sets and Event-level Maximum (Jet Trigger
Public Results). https://twiki.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
-
Deep Sets based Jet Regression for Pile-Up Mitigation: The system can predict
the longitudinal production vertex z-position of each jet from the charged-particle tracks reconstructed within the corresponding jet super-ROI.
This allows for an event level discriminant, MLPL, to be constructed toefficiently reject events dominated by pile-up jets.
-
Uncertainty Aware Jet Regression: The DIPz model is trained using a loss function that learns both
the central value and its associated uncertainty,
providingreliable per instance uncertainty estimates.
This capability is critical for high throughput environments where both predictive accuracy and uncertainty quantification are needed. -
Fast Rejection Algorithm for Trigger Optimization: The MLPL discriminant, defined as
MLPL(n, m) = max J(m)n max z log L(z; J(m)n),
enables the system to identify events whereat least n jets originate from the same primary collision vertex.
This allows for a pre-selection rejection that canreduce the input rate to the later CPU-intensive selection steps by a factor of four.
-
Scalable and Low Latency Deployment: The DIPz and MLPL approach is fully implemented in the ATLAS Athena framework, running with
negligible additional CPU cost compared to existing tracking stages.
This makes it suitable forreal-time event selections at the ATLAS High-Level Trigger
and offers ascalable, low-latency solution for future HL-LHC operations.
Abstract
Multiple proton-proton interactions occur in the same bunch crossing at the LHC ("pile-up"), with the mean number of interactions per bunch crossing reaching up to about 60 in Run 3 and up to about 200 at the High-Luminosity LHC. As a direct consequence, hadronic activity within each bunch crossing increases, making it more difficult to identify collisions of interest pertaining to multijet final states. This work introduces DIPz, an uncertainty-aware Deep Sets-based machine-learning model that estimates each jet's origin along the beam axis as a Gaussian likelihood using the set of charged-particle tracks matched to the jet. Next, the DIPz predictions for each jet are combined into a likelihood-based event-level discriminant, which is optimised to select events compatible with a single collision producing a given jet multiplicity. This combined approach provides a robust and computationally efficient method for pile-up rejection in multijet final states, applicable to real-time event selections at the ATLAS High Level Trigger (HLT). Deployed online during Run 3 data taking, this approach reduces background rates by more than a factor of two, and decreases CPU consumption by approximately one third relative to the baseline data-taking configuration. It provides an early, flavour-agnostic pile-up jet rejection layer at the HLT preselection stage, introducing a previously unavailable handle for high-multiplicity jet signatures targeted by ATLAS.
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