Direct Vertex Reconstruction of $\Lambda$ Baryons from Hits in CLAS12 using Graph Neural Networks

Anselm Vossen; Keegan Menkce; Matthew McEneaney

REVIEW 3 major objections 6 minor 12 references

Direct Vertex Reconstruction of $\Lambda$ Baryons from Hits in CLAS12 using Graph Neural Networks

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A graph neural network can reconstruct Lambda decay vertices directly from drift-chamber hits in CLAS12 simulation, and it does so more accurately than the standard track-based algorithm.

desk verdict A clean proof-of-concept for GNN vertex regression at CLAS12, but the Lambda purity gain is overstated because the evaluation uses MC truth for pair selection and an optimistic signal fraction. read the letter →

arxiv 2507.01868 v1 pith:OAIBRGX7 submitted 2025-07-02 hep-ex

Keegan Menkce , Matthew McEneaney , Anselm Vossen This is my paper

classification hep-ex

keywords graphneuralnetworksLambdabaryonvertexreconstructionCLAS12driftchambersdisplacedsemi-inclusivedeepinelasticscatteringparticleidentification

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that a graph neural network can read raw drift-chamber hits and directly predict the longitudinal position of the pion decay vertex from Lambda decays in CLAS12 simulation, and that this prediction is more accurate than the standard track-based reconstruction. This matters because displaced-vertex cuts are the main handle for separating Lambda decays from background, and the standard approach has to swim tracks back through a complicated magnetic field near the target. In simulation the GNN has lower mean absolute error and lower spread than the standard algorithm across true vertex positions, particularly in the 4-7 cm region that matters for Lambda identification, and cutting on its output raises the signal-to-background ratio from above unity to about 2 at 6 cm. The authors frame the result as a proof of concept that hit-to-vertex mapping can be learned directly, warranting larger training sets and tests on realistic data.

What carries the argument

The load-bearing object is a graph convolutional network with three convolutional layers, a dropout layer, and a final linear layer. Each input event is represented as a fully connected graph whose nodes are the drift-chamber hits of the proton and pion tracks; each node carries the six-dimensional hit information ($x,y,z,c_x,c_y,c_z$). The model also receives the standard algorithm's vertex estimate and the reconstructed momentum as baseline inputs, and Gaussian noise added to the input vertex serves as regularization during training. The network maps this variable-size hit set directly to the longitudinal ($z$) vertex coordinate, which is exactly the quantity used to require a minimum Lambda flight distance.

What would settle it

Recompute the signal-to-background ratio after the GNN vertex cut on a simulated sample whose Lambda fraction matches real data and whose multiple proton-pion pairs are paired without truth information; if the ratio no longer improves over the standard algorithm, the practical claim is refuted.

Watch

Extended reading notes

Core claim

The central discovery is that the full set of drift-chamber hits from the proton and pion can be turned into a vertex prediction without first fitting tracks. On simulated semi-inclusive deep inelastic scattering events with a Lambda in the final state, the GNN outputs the z-component of the pion vertex; compared with the default closest-approach reconstruction, its mean absolute error relative to the Monte Carlo truth is smaller at every true vertex value, its standard deviation is smaller in the 4-7 cm region, and its predictions are more strongly correlated with truth. Applying a minimum vertex displacement cut on the GNN output increases the Lambda signal-to-background ratio, reaching just over 2 around 6 cm. The paper also shows that performance improves with training-set size, so the reported accuracy is treated as a floor rather than a ceiling.

Load-bearing premise

The demonstrated purity gain relies on selecting the proton-pion pair with Monte Carlo truth; in real CLAS12 data, with multiple track pairs and no truth tag, the advantage may not survive.

Editorial extensions

If this is right

A minimum cut on the GNN-predicted pion vertex improves the Lambda signal-to-background ratio in simulation, reaching just over 2.0 at a 6 cm cut.
The GNN's mean absolute error and standard deviation are smaller than the standard algorithm's in the 4-7 cm region where Lambda identification is most sensitive.
Since the input is raw hits rather than fitted tracks, vertex reconstruction could avoid track swimming through the overlapping magnetic fields near the target.
Because the Kolmogorov-Smirnov statistic improves with dataset size, training on larger samples should push the accuracy beyond the values reported.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same hit-to-vertex architecture should transfer to other displaced-vertex decays, such as $K_S^0$ or other hyperons, because the input features are detector hits rather than particle-specific track parameters.
The paper leaves untested the realistic case of multiple candidate proton-pion pairs per event; a natural extension is to let the GNN score all pairings and use the highest-scoring pair for vertex selection.
The correlation plot's middle cluster suggests a resolution floor set by drift-chamber hit spacing; one testable prediction is that finer chamber resolution would compress this cluster and improve GNN accuracy further.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this paper does one clean, useful thing—shows that a GCN regressing the z-vertex of the pion from raw CLAS12 drift-chamber hits beats the old track-based algorithm in simulation on a held-out set. That regression result is believable and is the real contribution. The paper is much weaker as a demonstration that this improves Lambda selection, because the evaluation uses MC truth to pick the correct proton-pion pair and runs at a 50% signal fraction. So the s/b gain is measured after the hardest experimental problem—pair association—has already been solved by a magic wand.

What's new: previous work used RNNs/GNNs for vertex finding and the same group used a GNN for Lambda ID, but direct regression of the decay vertex from raw hits is a different input representation and target. That's a legitimate niche. The paper is honest about its own limits: it only tests against the older reconstruction, it notes the recent upgrade, and it explicitly says it avoided a low-signal dataset. The held-out evaluation with MC truth is a sound design for the regression question, and the error-vs-true-vertex plots are informative.

The soft spots are the ones the authors themselves half-concede. Section 2 says the pair is chosen by MC truth when multiple pairs exist. That removes the entire combinatorial background from the s/b plot in Fig. 2. In real data every proton-pi combination would feed the GNN, and wrong pairs would produce their own vertex predictions. The 50% signal fraction also makes the purity curve look better than any realistic SIDIS sample. And since the comparison is only against an outdated baseline, we don't know whether a simple cut on the old vertex at the same efficiency does as well. These are not a dealbreaker for the regression claim—the MAE comparison stands on its own—but they undermine the "we improve Lambda purity" sentence in the abstract. The authors should be asked to run the s/b comparison with all candidate pairs and at low signal fraction, and to compare against the current CLAS12 vertex algorithm.

Reproducibility is also weak: no code, no trained model, and the hyperparameters are only partially described. That's forgivable for a short proof-of-concept, but it means the only way to build on this is to reimplement from text.

Who should read it: researchers doing ML-based vertexing or working on CLAS12 Lambda analyses. It's a good reminder that a truthful regression result can overclaim practical impact when the surrounding selection is unrealistic. I'd take the regression result seriously and treat the purity improvement as unproven. A serious referee should engage with it and request the realistic evaluation, not desk reject it.

Referee Report

3 major / 6 minor

Summary. The paper reports a graph neural network (GNN) trained to reconstruct the longitudinal (z) component of the pion decay vertex from Lambda baryons in simulated CLAS12 events. The input is a fully connected graph whose nodes are drift-chamber hits (position and direction) of the proton and pion candidate tracks, augmented by the standard algorithm's vertex output and reconstructed momentum. The network is trained on approximately 8e4 simulated events with Monte Carlo truth vertex labels and evaluated on a held-out set of 20,000 events. The authors report that the GNN achieves lower mean absolute error and lower standard deviation of the error than the existing track-based reconstruction, shows a stronger correlation with the true vertex, and that applying a minimum-vertex cut based on the GNN prediction improves the signal-to-background ratio of the Lambda sample. The paper concludes that this approach is promising and warrants further study with larger training samples.

Significance. If the central claims are substantiated, the paper offers a credible proof-of-concept that secondary-vertex reconstruction can be performed directly from detector hits by a GNN, bypassing explicit track fitting in the complicated magnetic field of CLAS12. The supervised learning design is sound: the target (the MC truth vertex) is independent of the model, and evaluation is performed on a separate held-out sample, so the reported regression improvement is not a tautology. The paper is also commendably honest about its limitations, explicitly noting that only the older CLAS12 reconstruction algorithm was used as a baseline and that no low-signal-fraction test was performed. However, the practical claim that GNN-based vertex cuts improve Lambda sample purity is currently supported only under a Monte-Carlo-truth-assisted pair-selection procedure, which removes combinatorial background before the GNN is evaluated. The regression comparison is likely valid, but the selection-oriented claim needs additional evidence before it can be considered established.

major comments (3)

[Section 2, pair selection; Section 3, signal-to-background curve] The purity improvement shown in the right panel of Figure 2 is computed after selecting the proton-pion pair using Monte Carlo truth: Section 2 states that when multiple proton-pi pairs are present, 'we choose the pair that was produced by a Lambda in the Monte Carlo simulations.' This removes the combinatorial background that any realistic analysis must handle. In actual CLAS12 data the GNN input cannot be restricted to truth-matched pairs; every proton-pion combination is a candidate and each wrong pair would produce its own vertex prediction. The reported s/b curve therefore measures separation only after the hardest part of the problem, pair association, is solved by truth. I request either a test in which all possible pairs are passed through the GNN (with background defined by the full set of fake combinations), or a demonstration with a realistic pair-association step independent of MC truth. Without this, the practical Lambda-selection claim is not yet supported.
[Section 4, comparison baseline] The paper compares the GNN only against the older CLAS12 reconstruction algorithm and explicitly notes that 'CLAS12 has recently made an upgrade to the reconstruction algorithm that performs better, but we only tested our GNN against the older reconstruction algorithm.' Because the practical value of the GNN is its potential to replace or improve upon the current standard vertex reconstruction, the comparison against the outdated baseline is load-bearing for the claim that the GNN improves vertex resolution in a way that matters for ongoing analyses. I ask the authors to quantify the newer algorithm's performance on the same test sample, or at minimum to provide a detailed discussion of why the expected gain of the upgrade would not alter the qualitative conclusion.
[Section 4, low-signal evaluation] The test sample contains approximately 50% Lambda events, whereas real SIDIS datasets have substantially lower signal fractions. The paper states that the authors 'opted against' evaluating on a dataset with significantly fewer Lambda events because 'the statistics would be too low.' This decision weakens the central practical claim: the purity gain from a vertex cut depends on the background composition and on the distribution of vertex predictions for background tracks, which cannot be reliably inferred from the 50%-signal test set. I request a low-signal-fraction test (even with reduced statistics, or with a background-dominated sample generated by combining proton and pion tracks from different events) to demonstrate that the s/b improvement persists under realistic conditions. This is distinct from the pair-selection issue in my first comment because it concerns the overall sample composition rather than the per-event pair ambiguity.

minor comments (6)

[Section 3, Figure 2] The text refers twice to the 'right panel' of Figure 2: once for the mean absolute error comparison and once for the signal-to-background curve. From the caption, the left panel appears to show the absolute error and the right panel the s/b curve; please correct the figure references to avoid confusion.
[Section 3, s/b definition] The y-axis label is 'Signal over background ratio' but the text interchangeably calls it 'purity' and notes that the curve 'brings above unity' and reaches a maximum 'just over 2.0.' Please define explicitly whether this is S/B, S/(S+B), or another quantity, and state the corresponding signal fraction.
[Section 2, network architecture] The network is described as having 'three convolutional layers, a dropout layer and a final linear layer,' but no dimensions, hidden sizes, dropout probability, optimizer, learning rate, or training epochs per sub-sample are given beyond 200 epochs. For reproducibility, these details should be included.
[Section 3, typo] The sentence 'Although botherror for bothmethodsincrease with the true vertex' contains typographical errors; it should read something like 'Although both errors for both methods increase with the true vertex.'
[Section 4, uncertainty on s/b] The signal-to-background ratio plot in Figure 2 is shown without statistical uncertainties; adding error bars or a band would help assess the significance of the maximum around 6 cm.
[Figure 1] Figure 1 is a generic illustration of CNN versus GNN taken from the authors' previous work; it is not specific to this study. Consider moving it to the introduction or removing it to save space, unless a more problem-specific schematic is available.

Circularity Check

0 steps flagged · score 0.0 of 10

GNN vertex regression is evaluated against independent MC truth; the s/b claim is limited by truth-assisted pair selection and sample composition, but this is an applicability limitation, not circularity.

full rationale

The target of the network is the true longitudinal pion vertex from CLAS12 Monte Carlo, which is generated by the simulation independently of the network and of the input features. The GNN is trained on drift-chamber hits, reconstructed momentum, and the output of the existing vertex algorithm, then evaluated on a held-out set against the same external truth label. Because the label is neither an output of the network nor a deterministic function of the fitted input features, the reported regression improvement does not reduce by construction to the training labels; including the standard reconstruction as an input feature is a modeling choice, not a fitted-target renaming. The purity/signal-to-background curve is weakened by the paper's own admission that, when multiple proton-pion pairs exist, the pair is chosen using Monte Carlo truth ('If an event had multiple proton and pi pairs, we choose the pair that was produced by a Lambda in the Monte Carlo simulations'), and by the statement in Section 4 that the authors 'opted against' testing on a dataset with significantly fewer Lambda events. These choices mean the demonstrated purity gain may not hold under realistic combinatorial background and signal fraction, but they concern external validity rather than a circular derivation, since the GNN prediction itself is still scored against an independent truth. The only self-citation that appears, [11], is used for a generic GNN/CNN diagram and is not load-bearing for the vertex-reconstruction result. No circular step is exhibited.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The GNN introduces no new physical entities. The main free parameters are the trained network weights and an unreported noise hyperparameter. The load-bearing assumptions are simulation fidelity and the MC-truth-based pair selection, which limits the practical reach of the purity claim.

free parameters (2)

GNN network weights = Learned from ~8e4 simulated events
The central claim depends on this trained mapping from hits to vertex. These weights are fitted to MC truth labels, so performance is tied to the specific training set and architecture choices.
Gaussian noise level for input vertex regularization = not reported
The authors state that 'regularization by the addition of Gaussian noise on the input vertex significantly improved the performance' (Section 2) but do not specify the noise standard deviation, so it is a hand-chosen, unreported hyperparameter affecting results.

assumptions (3)

domain assumption CLAS12 Monte Carlo simulation accurately models drift chamber hit positions, directions, and the magnetic field geometry.
The GNN is trained and evaluated exclusively on simulation. If the simulation's hit responses do not match the real detector, the learned mapping will not transfer. Invoked in Section 2 where the training data is described.
ad hoc to paper The proton-pion pair selected using MC truth corresponds to the pair that would be associated in real data.
Section 2: 'we choose the pair that was produced by a Lambda in the Monte Carlo simulations.' This is a data-selection assumption that is not true in real events and is load-bearing for the claimed purity improvement.
domain assumption The standard vertex reconstruction algorithm used as baseline and as input feature is the same one the GNN is compared against.
Section 4 notes that only the older reconstruction algorithm was tested, so the comparison is valid only against that version, not the current standard.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Direct Vertex Reconstruction of $\Lambda$ Baryons from Hits in CLAS12 using Graph Neural Networks." pith.science (2026). https://pith.science/paper/OAIBRGX7

@misc{pith2026250701868,
  author       = {Pith},
  title        = {Pith review of: Direct Vertex Reconstruction of $\Lambda$ Baryons from Hits in CLAS12 using Graph Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OAIBRGX7}},
  note         = {Machine review of arXiv:2507.01868}
}

abstract

Machine learning techniques, including Graph Neural Networks (GNNs), have been used extensively for data analysis in high energy and nuclear physics. Here we report on the use of a GNN to reconstruct decay vertices of $\Lambda$ hyperons directly from hits in the tracking detector at the CLAS12 experiment at Jefferson Laboratory (JLab). We show that we can improve the vertex reconstruction in simulation compared to the standard, track based, algorithm. We believe this warrants further study. The current study is limited by available training resources but points to an interesting possibility to forgo vertex reconstruction by track fitting in a complicated magnetic field for a more direct approach where the hit to vertex mapping is encoded in a neural network.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

[1]

Collins,Foundations of Perturbative QCD, Cambridge Monographs on Particle Physics, Nuclear Physics and Cosmology, Cambridge University Press (2011)

J. Collins,Foundations of Perturbative QCD, Cambridge Monographs on Particle Physics, Nuclear Physics and Cosmology, Cambridge University Press (2011)

work page 2011
[2]

The STAR Collaborationcollaboration, Global Λhyperon polarization in nuclear collisions, Nature 548 (2017)

work page 2017
[3]

Particle Data Groupcollaboration, Review of particle physics, Phys. Rev. D110 (2024) 030001

work page 2024
[4]

Leemann, D.R

C.W. Leemann, D.R. Douglas and G.A. Krafft,The continuous electron beam accelerator facility: CEBAF at the Jefferson Laboratory, Annual Review of Nuclear and Particle Science51 (2001) 413 [https://doi.org/10.1146/annurev.nucl.51.101701.132327]

work page arXiv 2001
[5]

V.D. Burkert and et al.,The CLAS12 spectrometer at Jefferson Laboratory, Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment 959(2020) 163419

work page 2020
[6]

Mestayer, K

M. Mestayer, K. Adhikari, R. Bennett, S. Bueltmann, T. Chetry, S. Christo et al.,The CLAS12 drift chamber system, Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment959 (2020) 163518

work page 2020
[7]

Thais, P

S. Thais, P. Calafiura, G. Chachamis, G. DeZoort, J. Duarte, S. Ganguly et al.,Graph neural networks in particle physics: Implementations, innovations, and challenges, 2022

work page 2022
[8]

K. Goto, T. Suehara, T. Yoshioka, M. Kurata, H. Nagahara, Y. Nakashima et al.,Development of a vertex finding algorithm using recurrent neural network,Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment1047 (2023) 167836

work page 2023

Show all 12 references

[9]

Shlomi, S

J. Shlomi, S. Ganguly, E. Gross, K. Cranmer, Y. Lipman, H. Serviansky et al.,Secondary vertex finding in jets with neural networks,The European Physical Journal C81(2021)

2021
[10]

Battaglia, J.B

P.W. Battaglia, J.B. Hamrick, V. Bapst, A. Sanchez-Gonzalez, V. Zambaldi, M. Malinowski et al., Relational inductive biases, deep learning, and graph networks, 2018

2018
[11]

McEneaney and A

M. McEneaney and A. Vossen,Domain-adversarial graph neural networks forΛhyperon identification with CLAS12,Journal of Instrumentation18(2023) P06002

2023
[12]

Paszke, S

A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan et al.,Pytorch: An imperative style, high-performance deep learning library, inAdvances in Neural Information Processing Systems 32, pp. 8024–8035, Curran Associates, Inc. (2019), http://papers.neurips.cc/paper/90...

2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Collins,Foundations of Perturbative QCD, Cambridge Monographs on Particle Physics, Nuclear Physics and Cosmology, Cambridge University Press (2011)

J. Collins,Foundations of Perturbative QCD, Cambridge Monographs on Particle Physics, Nuclear Physics and Cosmology, Cambridge University Press (2011)

work page 2011

[2] [2]

The STAR Collaborationcollaboration, Global Λhyperon polarization in nuclear collisions, Nature 548 (2017)

work page 2017

[3] [3]

Particle Data Groupcollaboration, Review of particle physics, Phys. Rev. D110 (2024) 030001

work page 2024

[4] [4]

Leemann, D.R

C.W. Leemann, D.R. Douglas and G.A. Krafft,The continuous electron beam accelerator facility: CEBAF at the Jefferson Laboratory, Annual Review of Nuclear and Particle Science51 (2001) 413 [https://doi.org/10.1146/annurev.nucl.51.101701.132327]

work page arXiv 2001

[5] [5]

V.D. Burkert and et al.,The CLAS12 spectrometer at Jefferson Laboratory, Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment 959(2020) 163419

work page 2020

[6] [6]

Mestayer, K

M. Mestayer, K. Adhikari, R. Bennett, S. Bueltmann, T. Chetry, S. Christo et al.,The CLAS12 drift chamber system, Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment959 (2020) 163518

work page 2020

[7] [7]

Thais, P

S. Thais, P. Calafiura, G. Chachamis, G. DeZoort, J. Duarte, S. Ganguly et al.,Graph neural networks in particle physics: Implementations, innovations, and challenges, 2022

work page 2022

[8] [8]

K. Goto, T. Suehara, T. Yoshioka, M. Kurata, H. Nagahara, Y. Nakashima et al.,Development of a vertex finding algorithm using recurrent neural network,Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment1047 (2023) 167836

work page 2023