Pith. sign in

REVIEW 3 major objections 5 minor 9 references

Analysis of Atomic Charge State and Atomic Number for VAMOS++ Magnetic Spectrometer using Deep Neural Networks and Fractionally Labelled Events

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A deep network can sort the full spectrum after seeing only the lowest, cleanest charge states and atomic numbers labeled by hand.

desk verdict A practical time-saver for VAMOS++ analysis, but the claimed accuracy on unlabeled events is not independently established because the network's own rounded predictions are used as training targets. read the letter →

arxiv 2507.07109 v2 pith:ELQKKF4X submitted 2025-06-20 physics.ins-det cs.LGnucl-exphysics.atom-phphysics.data-an

M. Rejmund , A. Lemasson This is my paper · ORCID
classification physics.ins-detcs.LGnucl-exphysics.atom-phphysics.data-an
keywords deepneuralnetworksatomicchargestatenumbermagneticspectrometerfractionallabelingpseudo-labeltrainingionizationchamberisotopicidentification
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

This paper claims that a deep regression network can take over the two hardest calibration steps in magnetic-spectrometer analysis, finding an ion's atomic charge state $q$ and atomic number $Z$, using only a small slice of precisely labeled events. On a fission-fragment dataset, the network was trained with integer labels for only the lowest, best-resolved charge states and atomic numbers, while all higher values were labeled by rounding the network's own prediction in each training step. The resulting spectra reach FWHM($q_{NN}$)/$q_{NN}$ between 1.0% and 1.2% and FWHM($Z_{NN}$)/$Z_{NN}$ between 1.1% and 1.2%, a 9% tighter resolution than the conventional step-by-step analysis, and the whole analysis runs in hours rather than months. If correct, this would make high-quality isotopic identification routine and reproducible for this class of spectrometers.

What carries the argument

The working engine is a dense feed-forward regression network, eight layers of 32 units, trained on $5\times10^7$ experimental events split 80/20 into training and validation sets. In the labeled region the target is a fixed integer; in the unlabeled region the target is the network's own output rounded to the nearest integer at each training step. The input mix is deliberately low-level: the ten raw ionization-chamber row energies, positions on the entrance window and focal-plane detectors, and the reciprocal mass-over-charge and Lorentz-factor terms, so the network has to infer the calibration-related correlations itself rather than receiving them as formulas.

What would settle it

Take a held-out sample from the unlabeled ranges, determine the true charge state and atomic number independently, for instance through coincident characteristic gamma rays or a complete manual analysis of a few thousand events, and compare with the network's predictions. A systematic offset in the unlabeled region would show that the self-labeling loop is not being checked against the ground truth.

Watch

Extended reading notes

Core claim

At its center, the paper reports a training scheme that supplies true integer labels only for the low-lying, well-resolved states and lets the network generate its own labels for the rest: for every unlabeled event the target is updated to the rounded value of the network's current prediction, so the model effectively pulls the entire ridge of charge-state and element bands into integer steps. With raw ionization-chamber energies and position variables as inputs, the trained network removes the window-deformation and position dependencies visible in the conventional spectra and yields charge-state and atomic-number distributions with 9% better FWHM resolution than the complete conventional analysis. The authors present this as a complete replacement for the months-long iterative correction procedure, producing the same or better isotopic identification from the same experimental data.

Load-bearing premise

The load-bearing premise is that the network's self-generated rounded labels for the unlabeled events are correct enough to serve as training targets, because the reported resolutions and the 9% improvement are measured on those self-labeled events rather than on independently confirmed identities.

Editorial extensions

If this is right

  • Routine VAMOS++ charge-state and $Z$ analysis would take hours instead of months, with no manual iteration.
  • The 9% FWHM improvement over the conventional method propagates directly into cleaner mass-number identification when $q_{NN}$ is combined with mass-over-charge.
  • The procedure makes the analysis reproducible: replaying the same trained network on the same raw data removes the human-bias component.
  • Because accurate absolute energy calibration of the ionization chamber is not needed, the method can be applied when calibration sources are unavailable.

Reading between the lines

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

  • If the self-labeling scheme generalizes, it becomes a template for any detector where a smooth response function distorts an intrinsically integer-valued observable: anchor a few resolved integers and let the network fill in the ladder.
  • An independent check is still missing: identities for events in the unlabeled ranges, established by gamma-ray coincidences or a manual reanalysis of a small subset, would settle whether the self-generated labels match true charge states and elements.
  • Run-to-run transfer may require retraining, since the window deformation and electronics response can change between experiments; the paper demonstrates one dataset rather than a universal trained model.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper presents a deep neural network approach to determine the atomic charge state q and atomic number Z of heavy ions detected by the VAMOS++ magnetic spectrometer. The network is trained on an experimental dataset where only 18% of events (q in [25,34]) and 13% of events (Z in [30,40]) are labeled with integer values; for the remaining unlabeled events (q in [35,50] and Z in [41,63]) the training target is the network's own current prediction rounded to the nearest integer. The authors report FWHM/q_NN from 1.0% to 1.2%, FWHM/Z_NN from 1.1% to 1.2%, and a 9% improvement in FWHM over conventional analysis, with analysis time reduced from months to hours.

Significance. If the claimed accuracy on the unlabeled region were independently verified, the method would be a practical and time-saving contribution to heavy-ion spectrometer data analysis, and the use of a minimal labeled fraction is an interesting and potentially valuable idea. The paper also explicitly uses raw detector energies and positions without providing calibration relationships, which is a useful goal for reducing human bias. However, the central quantitative claims rest on self-generated labels, and the current manuscript does not provide any external check on the correctness of the unlabeled-region predictions. The reported FWHM values and the 9% improvement measure self-consistency of the network rather than physical accuracy.

major comments (3)
  1. [Sections 4.1 and 4.2] The training targets for the unlabeled majority are the network's own rounded predictions: for q in [35,50] (82% of events) the target is q_t = floor(q_p + 0.5), and for Z in [41,63] (87% of events) the target is Z_t = floor(Z_p + 0.5). The validation split is subject to the same pseudo-labeling, and early stopping uses the validation loss computed on these self-generated targets. Therefore the validation RMSD values (0.177 for q, 0.227 for Z), the FWHM values in Sections 4.1, 4.2 and 5, and the 9% improvement over conventional analysis are not independent measurements of accuracy. A network can produce arbitrarily narrow distributions on the unlabeled region by learning to output values close to the nearest integer, even if the absolute q or Z assignments are systematically wrong. The central claim that the network 'autonomously and accurately classifies the remaining events' (Abstract) is therefore not established. I request an external validation on the labeled region only (e.g., reporting FWHM and peak positions for q in [25,34] and Z in [30,40] where ground truth is known), or a physics-based check such as gamma-ray coincidences as mentioned in Section 3, or a comparison against a fully independently analyzed subset of the unlabeled region.
  2. [Sections 5 and Figures 5(d), 6(d)] The comparison between the neural-network results and the 'conventional analysis' results (q_conv, Z_conv) is not quantified in enough detail to support the claimed 9% FWHM improvement. It is not stated how q_conv and Z_conv are obtained (the paper says only that they come from 'a complete process of the conventional analysis method'), what peak-fitting procedure is used, how many events are included, and whether the improvement is dominated by the self-labeled region. If the same unlabeled events are used for both q_NN and q_conv, the comparison is partly circular because the q_NN distribution is generated under the pseudo-label objective. Please provide the exact definition of FWHM used, the uncertainty on the quoted values, and a breakdown of the improvement for the labeled and unlabeled ranges separately.
  3. [Section 4] The paper provides no evidence that the pseudo-labeled training converges to the true physical q and Z rather than to a self-consistent but inaccurate solution. Because only 18% (q) and 13% (Z) of events carry externally supplied labels, the network's behavior on the unlabeled range is constrained only by the smoothness of the input-output mapping and by the rounding rule, which can reinforce initial systematic errors. I suggest the authors perform a stability analysis by retraining from several random initializations and reporting the variance of the predicted q and Z for the unlabeled events; if the predictions are stable across seeds, this still does not prove correctness, but it would at least demonstrate that the result is not a single-run artifact. Additionally, the paper mentions that coincident characteristic gamma rays can verify isotopic identification, but this verification is not performed or reported; applying it to a subset of the data would provide the missing external reference.
minor comments (5)
  1. [Section 4.1] The input list states 'the reciprocal of the mass-over-charge, (A/q), and of the Lorenz term, (gamma-1)' which is confusing: (A/q) and (gamma-1) are not reciprocals. Please clarify whether the inputs are q/A and 1/(gamma-1) or A/q and (gamma-1), and correct the typo 'Lorenz' to 'Lorentz'.
  2. [Equation (3.4)] Equation (3.4) as written is dimensionally inconsistent with the definition of u as 931.494 MeV/c^2: the denominator 1 u * (gamma-1) * (A/q) has units of MeV/c^2, not MeV. Please either define u as an energy (931.494 MeV) or include the factor c^2 explicitly, and state the assumed units of E_tot.
  3. [Figure 4] The caption of Figure 4 does not indicate which curves correspond to q and which to Z, although the text refers to both. Please add labels or a legend to the figure.
  4. [Section 4.2] The sentence beginning 'Figure demonstrates a sharp definition...' is missing a figure number. It should refer to Figure 6(a) and (b).
  5. [Section 3] The derivation of q in equation (3.4) uses the total energy E_tot, but the text says E_tot = E_um + Delta E + E_res, while the input variables to the network for the q analysis include raw energies and the reciprocal of A/q; it would be helpful to state explicitly whether E_tot is used in the network inputs or only in the reference q_conv.

Circularity Check

3 steps flagged · score 6.0 of 10

Accuracy metrics on unlabeled events are circular: training targets for 82% (q) and 87% (Z) of events are the network's own prediction rounded to the nearest integer, so FWHM and 9% gains measure self-consistency, not external correctness.

  1. self definitional [Section 4.1, training-target definition for q_u = [35, 50]]
    "For the labeled events in the range q_l = [25, 34], the training target was the corresponding q_t = q_l, while for the unlabeled events in the range q_u = [35, 50], it was the nearest integer value to the predicted in each learning step q_p, q_t = ⌊q_p + 0.5⌋."

    For the 82% of events in the unlabeled range q_u = [35, 50], the training target is defined as the network's own current prediction rounded to the nearest integer. Minimizing the regression loss against this target pushes outputs toward the nearest integer regardless of the true atomic charge state, artificially narrowing the q_NN peaks. The reported FWHM(q_NN)/q_NN values and the 9% improvement over q_conv are computed on the full spectrum that includes this self-labeled population, so they are a measure of self-consistency, not an independent test of correctness.

  2. self definitional [Section 4.2, training-target definition for Z_u = [41, 63]]
    "Conversely, for the unlabeled events within the range Z_u = [41, 63], the target was the nearest integer value to the predicted value in each learning step Z_p, i.e., Z_t = ⌊Z_p + 0.5⌋."

    The same construction is applied to atomic number: for 87% of events (Z_u = [41, 63]), the target equals the network's own prediction rounded. The reported FWHM(Z_NN)/Z_NN of 1.1% to 1.2% and the 9% improvement over Z_conv are therefore obtained on data whose labels were generated by the network itself. The apparent sharpness of the Z_NN bands can be produced even if the absolute Z assignment is systematically wrong, because the loss explicitly rewards integer-valued outputs without any external ground truth in this range.

1 more flagged steps
  1. fitted input called prediction [Section 4, dataset partitioning and validation, before Section 4.1]
    "The labeled and unlabeled events were provided to the networks simultaneously, following the experimental repetition frequency. The dataset used to train the neural networks consisted of 5 × 10^7 events. The dataset was randomly partitioned into the training set (80%) and the validation set (20%)."

    The validation set is a random 20% split of the same dataset, and for unlabeled events in that split the targets are also floor(q_p + 0.5) or floor(Z_p + 0.5), i.e., generated from the network's own predictions. The reported training/validation RMSD convergence therefore does not provide external evidence of accuracy on the unlabeled range; it only shows that the network reproduces its own rounded outputs. The few precisely labeled low-state events do anchor the model, but the central claim that the network 'autonomously and accurately' classifies the vast majority of events is not independently demonstrated.

full rationale

The paper's headline claim is that a network trained with a small fraction of precisely labeled events autonomously and accurately classifies the remaining events, with FWHM(q_NN)/q_NN = 1.0-1.2%, FWHM(Z_NN)/Z_NN = 1.1-1.2%, and a 9% improvement over conventional analysis. The central circularity is explicit in Sections 4.1 and 4.2: for the unlabeled majority (q_u = [35,50], covering 82% of events; Z_u = [41,63], covering 87% of events), the training target is the network's own prediction rounded to the nearest integer. Thus the loss, the validation RMSD, and the reported FWHM and 9% gains all measure how well the network agrees with itself on the unlabeled region, not how well it agrees with true charge states or atomic numbers. This is a self-consistency measure, not an external accuracy measure. The labeled low-state subset does provide physical anchoring, and the inputs (raw energies, positions, A/q) plus the physics of the spectrometer constrain the problem, so the method may well work. But the paper does not exhibit any independent verification for the unlabeled region, and therefore the strongest accuracy claims reduce by construction to the network's own rounded outputs. This warrants a partial circularity score of 6. There is no evidence of load-bearing self-citation or renamed known results; the issue is specifically the self-generated labels being used as both training target and validation truth.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical entities. The free parameters are the network size, the labeled ranges, and training hyperparameters, all chosen by hand. The central assumption is that the small labeled subset is correct and representative, and that the self-generated labels for the remaining events are valid training targets.

free parameters (5)
  • Network architecture size (8 layers x 32 units)
    Chosen by hand, 'similarly as described in Ref. [5]', not optimized or justified for this task.
  • Labeled charge state range q_l = [25, 34] = 25-34
    Corresponds to ~18% of events; selected because these are the best-resolved charge states. This free choice determines the amount of supervision.
  • Labeled atomic number range Z_l = [30, 40] = 30-40
    Corresponds to ~13% of events; selected because these are the best-resolved atomic numbers. Free choice affecting the method's reliance on self-labels.
  • Early stopping tolerance and patience = 0.001, 60 epochs
    Chosen without sensitivity analysis; no evidence that other values would not change results.
  • Pseudo-label rounding rule floor(x + 0.5)
    The rounding rule determines the self-generated targets for unlabeled events; no alternative rules (e.g., confidence-based thresholds) are explored.
assumptions (5)
  • domain assumption The manually labeled events in q_l = [25,34] and Z_l = [30,40] have correct integer labels.
    The paper relies on these labels as ground truth, saying identification can be verified by coincident gamma rays (Section 3), but presents no such verification in this work.
  • domain assumption The input variables (raw energies, positions, A/q, gamma-1) contain sufficient information to determine q and Z.
    No proof or information-theoretic analysis is given; the network is expected to learn this mapping from the data.
  • domain assumption A dense feed-forward network with 8 layers of 32 units can represent the required mapping.
    The architecture is taken from prior work (Ref. [5]) without testing alternative architectures.
  • domain assumption The approximate online calibration coefficients C_i used to define the labeled gates are accurate enough for the low-q and low-Z regions.
    The labels are selected using the q vs A/q correlation that depends on these approximate coefficients; if the coefficients are wrong, labels could be incorrect.
  • ad hoc to paper For the unlabeled events, the network's own prediction rounded to the nearest integer is a valid training target (pseudo-label assumption).
    This is the core assumption of the fractional labeling method: the network can bootstrap itself on the unlabeled region. If the network makes systematic errors there, the training reinforces them.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Analysis of Atomic Charge State and Atomic Number for VAMOS++ Magnetic Spectrometer using Deep Neural Networks and Fractionally Labelled Events." pith.science (2026). https://pith.science/paper/ELQKKF4X

@misc{pith2026250707109,
  author       = {Pith},
  title        = {Pith review of: Analysis of Atomic Charge State and Atomic Number for VAMOS++ Magnetic Spectrometer using Deep Neural Networks and Fractionally Labelled Events},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ELQKKF4X}},
  note         = {Machine review of arXiv:2507.07109}
}
read the original abstract

The VAMOS++ magnetic spectrometer is a multi-parametric system that integrates ion optical magnetic elements with a multi-detector stack. The magnetic elements, along with the tracking and timing detectors and the trajectory reconstruction method, provide the analysis of the magnetic rigidity, the trajectory length between the beam interaction point and the focal plane of the spectrometer, and the related velocity and mass-over-charge ratio. The segmented ionization chamber provides the energy measurements necessary to analyze the atomic charge state and atomic number. However, this analysis critically suffers from inherent limitations due to the variable thickness and non-uniformity of the entrance window of the ionization chamber and other detector imperfections. Conventionally, this meticulous, detailed analysis is exceptionally tedious, often requiring several months to complete. We present a novel method utilizing deep neural networks, trained on an experimental dataset with only a small fraction of precisely labeled events for the lowest and best-resolved atomic charge states or numbers. This innovative approach enables the networks to autonomously and accurately classify the remaining events. This method drastically accelerates the acquisition of high-resolution atomic charge state and atomic number spectra, reducing analysis time from months to mere hours. Crucially, by discarding human bias, this approach ensures standardized, optimal, and reproducible results with unprecedented efficiency.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

9 extracted references · 9 canonical work pages

  1. [1]

    Pullanhiotan, M

    S. Pullanhiotan, M. Rejmund, A. Navin, W. Mittig and S. Bhattacharyya,Performance of VAMOS for reactions near the Coulomb barrier, Nuclear Instruments and Methods in Physics Research Section A 593(2008) 343

    work page 2008

  2. [2]

    Rejmund, B

    M. Rejmund, B. Lecornu, A. Navin, C. Schmitt, S. Damoy, O. Delaune et al.,Performance of the improved larger acceptance spectrometer: VAMOS++, Nuclear Instruments and Methods in Physics Research Section A 646 (2011) 184

    work page 2011

  3. [3]

    Montagnoli, A.M

    G. Montagnoli, A.M. Stefanini, M. Trotta, S. Beghini, M. Bettini, F. Scarlassara et al.,The large-area micro-channel plate entrance detector of the heavy-ion magnetic spectrometer PRISMA, Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment 547(2005) 455

    work page 2005

  4. [4]

    Lemasson and M

    A. Lemasson and M. Rejmund,Fast trajectory reconstruction techniques for the large acceptance magnetic spectrometer VAMOS++, Nuclear Instruments and Methods in Physics Research Section A 1054 (2023) 168407

    work page 2023

  5. [5]

    Rejmund and A

    M. Rejmund and A. Lemasson,Seven-dimensional trajectory reconstruction for VAMOS++, Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment 1076 (2025) 170445

    work page 2025

  6. [6]

    10.26143/GANIL-2022-E826_21

    E826-21 Collaboration,E826-21 GANIL dataset, 2022. 10.26143/GANIL-2022-E826_21

    work page doi:10.26143/ganil-2022-e826_21 2022

  7. [7]

    Vandebrouck, A

    M. Vandebrouck, A. Lemasson, M. Rejmund, G. Fremont, J. Pancin, A. Navin et al.,Dual Position Sensitive MWPC for tracking reaction products at VAMOS++, Nuclear Instruments and Methods in Physics Research Section A 812(2016) 112

    work page 2016

  8. [8]

    Kinsley and D

    H. Kinsley and D. Kukieła,Neural Networks from Scratch in Python, Harrison Kinsley (2020)

    work page 2020

Show all 9 references
  1. [9]

    Subasi,Practical Machine Learning for Data Analysis Using Python , Academic Press (2020), 10.1016/C2019-0-03019-1

    A. Subasi,Practical Machine Learning for Data Analysis Using Python , Academic Press (2020), 10.1016/C2019-0-03019-1. – 11 –

    2020 doi

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.