arxiv: 2604.11648 · v1 · submitted 2026-04-13 · ✦ hep-ex

Recognition: unknown

Filtering hits for speeding up online track reconstruction at hadron colliders

Andrea Coccaro , Carlo Schiavi , Alessandro Zaio

Authors on Pith no claims yet

Pith reviewed 2026-05-10 15:41 UTC · model grok-4.3

classification ✦ hep-ex

keywords track reconstructionhit filteringconvolutional neural networkpile-upLHC triggerHigh-Luminosity LHCaccelerator cardsonline processing

0 comments

The pith

A convolutional neural network filters detector hits to speed up track reconstruction in high-pileup LHC events.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents a method to assist track reconstruction by using a convolutional neural network to discard irrelevant hits from the innermost detector layers before the main reconstruction step begins. Current combinatorial approaches slow down dramatically as the number of simultaneous collisions per bunch crossing rises, and this problem will intensify after the High-Luminosity LHC upgrade. The CNN is built to run on accelerator hardware, removing unnecessary data early while aiming to keep all hits needed for correct trajectories. If the filter works as described, trigger systems could maintain acceptable processing times without discarding potentially valuable physics events. A sympathetic reader would care because the technique directly targets the scaling bottleneck that threatens real-time selection at future collider luminosities.

Core claim

The authors introduce and characterize a convolutional-neural-network architecture that filters unnecessary detector hits to assist track reconstruction. The network is designed for straightforward deployment on accelerator cards, and the work assesses its effect on processing speed under the elevated hit occupancies expected at the High-Luminosity LHC.

What carries the argument

Convolutional neural network architecture that identifies and removes irrelevant hits from detector layers to reduce the combinatorial load on subsequent track reconstruction.

If this is right

Track reconstruction processing time grows more slowly with increasing pile-up than with current methods.
The filtering step can be executed in real time on accelerator cards inside the trigger system.
Overall computational cost for online tracking decreases while preserving the ability to reconstruct charged-particle trajectories.
The approach supports continued operation of trigger strategies after the luminosity upgrade without major redesign of downstream algorithms.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same early-filtering idea could be tested on other detector subsystems that also face high occupancy.
Combining the CNN filter with existing seeding or pattern-recognition steps might produce further speed gains.
Performance measured in simulation will ultimately need confirmation on recorded collision data to establish real-world reliability.
If the filter proves robust, it could influence the design of future trigger architectures that rely on aggressive early data reduction.

Load-bearing premise

The neural network can remove large numbers of hits without eliminating any that are required to reconstruct tracks from interesting physics processes.

What would settle it

A direct comparison of track-finding efficiency and fake-track rate on the same set of high-pile-up simulated events when the reconstruction algorithm receives only the CNN-filtered hits versus the full unfiltered hit collection.

Figures

Figures reproduced from arXiv: 2604.11648 by Alessandro Zaio, Andrea Coccaro, Carlo Schiavi.

**Figure 2.** Figure 2: Display of the hitmap obtained from the two-dimensional projection of a toy [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Schematic representation of the architecture being developed for the filtering [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: The distribution of the raw Dout output of the filtering algorithm using the model Kstart = 128, Kend = 4 and the input dataset with a binning along the φ direction corresponding to ∆φ = 0.0002 rad is displayed on the left. The corresponding distribution of Dlog, the scaling of the Dout, is displayed on the right. how to filter hits. The chosen figure of merit for comparing the different trained models is … view at source ↗

**Figure 5.** Figure 5: Background rejection factors when selecting a cut on the output score [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

**Figure 6.** Figure 6: Efficiency as a function of pT of the signal track in the event, computed for the 90%, 95% and 98% WPs, defined as in the text. 4.1. Pile-up While the training process was performed on events generated by overlaying an average of 25 pile-up interactions on top of the additional random noise, the performance of the network is now evaluated on samples in which the average number of secondary collisions is se… view at source ↗

**Figure 7.** Figure 7: Displays of the hitmaps as 2D projections in the ( [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗

**Figure 8.** Figure 8: ROC curves when the best model is evaluated against events with a higher [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗

**Figure 9.** Figure 9: ROC curves for the best model evaluated on samples where the Gaussian [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗

**Figure 10.** Figure 10: ROC curves for the best model evaluated on samples where the per-hit signal [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗

read the original abstract

Collider experiments are equipped with trigger systems that rapidly inspect the physics content emerging from collisions to decide whether the resulting products are worth saving for later analysis. One crucial aspect for analyzing the final states originating from the collisions is to process the information produced by charged particles in the innermost detectors to reconstruct the corresponding trajectories. This task is a challenge for the experiments running at the Large Hadron Collider (LHC) at CERN because of the large number of secondary collisions per bunch crossing, the so-called pile-up vertices, giving rise to extremely high hit occupancies in the detector layers close to the beam line. Reconstructing tracks is a combinatorial problem and its processing time strongly depends on the average pile-up per event. The future accelerator-complex upgrade to the High-Luminosity LHC, implying even higher detector occupancies, will result in a considerable growth of the computational cost of the current trigger strategies. To face this issue, a new technique for assisting track reconstruction by filtering out unnecessary detector information is presented and characterized in this work. The algorithm is based on a convolutional-neural-network architecture which can be easily deployed on accelerator cards. The impact of this approach is assessed and future prospects are also discussed.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

CNN hit filtering for faster LHC tracking is a sensible idea, but the efficiency for low-pT tracks needs stronger proof before the speed-up claim holds.

read the letter

The key point is that this paper applies a CNN to filter hits and reduce the load on track reconstruction for HL-LHC pile-up conditions. It positions the network for easy deployment on accelerator hardware. What stands out positively is the focus on a practical constraint: the filter has to run fast enough for the trigger. They evaluate its effect in simulation and outline next steps. This is a direct response to the growing computational demands from higher luminosity. The main weakness is in the validation of track efficiency. The central assumption is that the CNN removes mostly noise hits without losing reconstructible tracks, especially those with low transverse momentum or displaced origins. In dense environments these tracks produce fewer hits, so any bias in the filter could reduce efficiency. The assessment mentioned in the abstract does not appear to include the full kinematic breakdown needed to confirm the speed-up survives. If efficiency drops by more than a couple of percent in key categories, the claimed benefit shrinks. This paper is for physicists and engineers working on online tracking algorithms. A reader interested in ML applications for trigger systems would get some value from the architecture and deployment angle. It deserves peer review because the problem is timely and the method is concrete, even though the results section needs strengthening on efficiency preservation.

Referee Report

1 major / 1 minor

Summary. The manuscript proposes a convolutional neural network (CNN) architecture to filter unnecessary hits from detector data, thereby assisting online track reconstruction by reducing combinatorial complexity in high pile-up environments at hadron colliders such as the HL-LHC. The method is presented as deployable on accelerator cards, with the work including characterization of the algorithm, assessment of its impact on processing time, and discussion of future prospects.

Significance. If validated, the approach could meaningfully address the computational scaling challenges of track finding under increased luminosity by reducing hit multiplicity while preserving essential physics information. The explicit design for hardware acceleration is a practical strength that aligns with real-time trigger requirements. However, without quantitative benchmarks the significance remains prospective rather than demonstrated.

major comments (1)

The central claim that hit filtering speeds up track reconstruction without compromising quality rests on the untested assumption that the CNN retains essentially all hits from reconstructible tracks, including low-pT and displaced ones. No efficiency-vs-pile-up scan broken down by track kinematics or origin is provided, leaving the speed-up claim unsupported at the level required for trigger deployment.

minor comments (1)

The abstract states that the impact is assessed, but the manuscript provides no tables, figures, or numerical results quantifying speed-up, efficiency retention, or comparison to baseline track finders.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading of the manuscript and the constructive feedback. We appreciate the acknowledgment of the practical relevance of a hardware-deployable CNN hit filter for addressing combinatorial challenges in high-pile-up track reconstruction. We respond to the major comment below and will revise the manuscript to strengthen the supporting evidence.

read point-by-point responses

Referee: The central claim that hit filtering speeds up track reconstruction without compromising quality rests on the untested assumption that the CNN retains essentially all hits from reconstructible tracks, including low-pT and displaced ones. No efficiency-vs-pile-up scan broken down by track kinematics or origin is provided, leaving the speed-up claim unsupported at the level required for trigger deployment.

Authors: We agree that the manuscript would benefit from a more detailed breakdown of hit-retention efficiency to substantiate the central claim. The current version reports overall efficiency and processing-time gains but does not include efficiency-versus-pile-up curves separated by track pT (including the low-pT regime) or by track origin (prompt versus displaced). In the revised manuscript we will add these scans, using the same simulated samples already employed for the aggregate results, to demonstrate that the filter preserves hits from reconstructible tracks across the relevant kinematic and topological phase space. revision: yes

Circularity Check

0 steps flagged

No circularity: new CNN hit-filtering algorithm validated empirically on simulation

full rationale

The paper introduces a convolutional neural network to filter detector hits and thereby accelerate combinatorial track reconstruction under high pile-up. The central contribution is an engineering proposal whose performance (speed-up factor and track-finding efficiency) is measured directly from training and inference on simulated events rather than derived from any self-referential equation, fitted parameter renamed as prediction, or load-bearing self-citation. No step in the described chain reduces by construction to its own inputs; the method is tested against independent external metrics (reconstruction quality and timing) and remains self-contained.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Limited information from abstract; the method relies on standard ML assumptions and the effectiveness of CNNs in pattern recognition for detector data.

free parameters (1)

CNN hyperparameters
Architecture details, training parameters not specified in abstract but typically fitted.

axioms (1)

domain assumption Convolutional neural networks can accurately identify relevant hits for track reconstruction.
Core assumption enabling the filtering approach.

pith-pipeline@v0.9.0 · 5508 in / 1052 out tokens · 35593 ms · 2026-05-10T15:41:53.243349+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

27 extracted references · 2 canonical work pages · 1 internal anchor

[1]

Introduction Experimental particle physics studies the fundamental particles and forces that make up the universe by using high-energy accelerators, such as the LHC [1], and detectors, such as ATLAS [2] and CMS [3]. At the LHC, interesting events from the proton-proton collisions are rare and the ATLAS and CMS experiments are equipped with sophisticated t...

work page internal anchor Pith review Pith/arXiv arXiv 2026
[2]

Synthetic dataset Synthetic data is generated with a custom event generator simulating a cylindrical detector consisting of eight concentric layers. Its geometry is based on the barrel part of the inner detector of the ATLAS experiment, and with the four layers matching the Filtering hits for speeding up online track reconstruction at hadron colliders3 pi...
[3]

The network architecture is inspired by the denoising autoencoders in Ref

The filtering algorithm The filtering algorithm is designed to receive the data in the image-like form described in the previous section as input. The network architecture is inspired by the denoising autoencoders in Ref. [19, 20, 21], which are designed to reconstruct a clean version of noisy input data by learning to classify between signal and noise at...
[4]

The following sub-sections examine these three aspects individually

Robustness Tests The robustness of the trained network is assessed by evaluating its classification performance on dedicated samples in which increased pile-up, enhanced hit smearing, and reduced signal-hit collection efficiency are simulated. The following sub-sections examine these three aspects individually. This procedure is intended to quantify how t...
[5]

pile-up tracks originating from secondary collisions and detector noise

Conclusions and outlook This work presents a novel machine learning approach based on a convolutional neural network to classify hits in a simulated tracking detector, distinguishing those produced by charged-particles from the hard-scattering primary vertex from those produced by Filtering hits for speeding up online track reconstruction at hadron collid...
[6]

Filtering hits for speeding up online track reconstruction at hadron colliders13

Acknowledgments This work was partially supported by ICSC – Centro Nazionale di Ricerca in High Performance Computing, Big Data and Quantum Computing, funded by European Union – NextGenerationEU. Filtering hits for speeding up online track reconstruction at hadron colliders13
[7]

Evans and P

L. Evans and P. Bryant, LHC Machine, JINST3(2008) S08001

2008
[8]

ATLAS Collaboration, The ATLAS Experiment at the CERN Large Hadron Collider, JINST3 2008, S08003

2008
[9]

CMS Collaboration, The CMS experiment at the CERN LHC,JINST3(2008), S08004

2008
[10]

ATLAS Collaboration, Operation of the ATLAS trigger system in Run 2, JINST15(2020) 10, P10004

2020
[11]

The ATLAS Trigger System for LHC Run 3 and Trigger performance in 2022, JINST192024, P06029

ATLAS Collaboration. The ATLAS Trigger System for LHC Run 3 and Trigger performance in 2022, JINST192024, P06029

2022
[12]

CMS Collaboration, The CMS trigger system, JINST122017, P01020
[13]

ATLAS Collaboration, The ATLAS inner detector trigger performance in pp collisions at 13 TeV during LHC Run 2, Eur. Phys. J. C82(2022) no.3, 206

2022
[14]

Aberle,et al., High-Luminosity Large Hadron Collider (HL-LHC): Technical design report, CERN-2020-010 (2020)

O. Aberle,et al., High-Luminosity Large Hadron Collider (HL-LHC): Technical design report, CERN-2020-010 (2020)

2020
[15]

Duarte,et al., Graph Neural Networks for Particle Tracking and Reconstruction, arXiv:2012.01249

J. Duarte,et al., Graph Neural Networks for Particle Tracking and Reconstruction, arXiv:2012.01249

work page arXiv 2012
[16]

DeZoort,et al., Charged Particle Tracking via Edge-Classifying Interaction Networks, Comput

G. DeZoort,et al., Charged Particle Tracking via Edge-Classifying Interaction Networks, Comput. Softw. Big Sci.5(2021), 26

2021
[17]

Bocci, ,et al., Heterogeneous Reconstruction of Tracks and Primary Vertices With the CMS Pixel Tracker, Front

A. Bocci, ,et al., Heterogeneous Reconstruction of Tracks and Primary Vertices With the CMS Pixel Tracker, Front. Big Data3(2020), 601728

2020
[18]

ATLAS Collaboration, ATLAS Software and Computing HL-LHC Roadmap, CERN-LHCC-2022- 005

2022
[19]

CMS Collaboration, CMS Phase-2 Computing Model: Update Document, CERN-CMS-NOTE- 2022-008

2022
[20]

Coccaro,et al., Fast neural network inference on FPGAs for triggering on long-lived particles at colliders, Mach

A. Coccaro,et al., Fast neural network inference on FPGAs for triggering on long-lived particles at colliders, Mach. Learn. Sci. Tech.4(2023) no.4, 045040

2023
[21]

Soybelman,et al., Accelerating graph-based tracking tasks with symbolic regression, Mach

N. Soybelman,et al., Accelerating graph-based tracking tasks with symbolic regression, Mach. Learn. Sci. Tech.5(2024) no.4, 045042

2024
[22]

ATLAS Collaboration, Technical Design Report for the ATLAS Inner Tracker Pixel Detector, Technical design report, CERN-2017-21 (2017)

2017
[23]

ATLAS Collaboration, Expected tracking performance of the ATLAS Inner Tracker at the High- Luminosity LHC, JINST20(2025) 02, P02018

2025
[24]

ATLAS Collaboration, Charged-particle distributions in √s= 13 TeV pp interactions measured with the ATLAS detector at the LHC, Phys. Lett. B758(2016), 67-88

2016
[25]

G. E. Hinton,et al., Reducing the Dimensionality of Data with Neural Networks, Science, 313 (5786): 504-507 (2006)

2006
[26]

Pascal,et al., Extracting and composing robust features with denoising autoencoders, ICML, ACM International Conference Proceeding Series, 1096-1103 (2008)

V. Pascal,et al., Extracting and composing robust features with denoising autoencoders, ICML, ACM International Conference Proceeding Series, 1096-1103 (2008)

2008
[27]

Long,et al., Fully convolutional networks for semantic segmentation, 2015 IEEE Conference on Computer Vision and Pattern Recognition, 3431-3440 (2015)

J. Long,et al., Fully convolutional networks for semantic segmentation, 2015 IEEE Conference on Computer Vision and Pattern Recognition, 3431-3440 (2015)

2015