Pith. sign in

REVIEW 3 major objections 5 minor 40 references

Simultaneous Estimation of Elliptic Flow Coefficient and Impact Parameter in Heavy-Ion Collisions using CNN

T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A single CNN trained on simulated Pb-Pb collisions predicts both elliptic flow and impact parameter with less than six percent error.

desk verdict Solid in-simulation benchmark, but the direct-to-LHC claim is not established and the linear-activation network quietly limits the method. read the letter →

arxiv 2411.11001 v1 pith:ZJNBMUJ7 submitted 2024-11-17 hep-ph nucl-exnucl-th

Praveen Murali , Sadhana Dash , Basanta Kumar Nandi This is my paper · ORCID
classification hep-phnucl-exnucl-th PACS 25.75.-q25.75.Ld
keywords deeplearningconvolutionalneuralnetworkellipticflowimpactparameterheavy-ioncollisionsAMPTeventgeneratorPb-Pbat5.02TeVimage-basedregression
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

Deep learning can recover two key heavy-ion collision properties—the elliptic flow coefficient $v_2$ and the impact parameter $b$—from a single two-dimensional image of charged particles. The paper trains a convolutional neural network on 50,000 minimum-bias Pb-Pb events at 5.02 TeV simulated with the AMPT event generator, using $\eta$–$\phi$ histograms weighted by $p_T$, mass, transverse mass, or both. Across twelve input/output combinations and two bin settings, all twenty-four trained models reach mean absolute errors below six percent, and the best model (both $p_T$- and mass-weighted input at $32 \times 32$ bins) reproduces the centrality and $p_T$ dependence of $v_2$ seen in simulation and in experimental data. This matters because the impact parameter is not directly measurable, and a single network that returns both $v_2$ and $b$ in one forward pass offers a fast, simultaneous estimator for LHC-energy heavy-ion collisions.

What carries the argument

The central object is the weighted $(\eta,\phi)$ image: a binned histogram of all charged particles with $0.2<p_T<5.0$ GeV/$c$ and $|\eta|<0.8$, with pixel values weighted by $p_T$, mass, transverse mass, or both, treated as a multichannel image. This image is the sole input to a CNN regressor whose four convolutional layers (16, 16, 32, and 32 filters, linear activation) extract spatial correlations, followed by flattening, dense layers of 64 and 32 neurons, and one or two output neurons; group normalization, dropout, and kernel regularizers control overfitting. The machinery works by mapping the spatial distribution of produced particles directly to the target labels, with the reaction-plane-zero AMPT setup supplying exact $v_2$ labels.

What would settle it

Train the same CNN on real ALICE Pb-Pb events at 5.02 TeV, using standard experimental estimates of centrality and $v_2$ as labels, and compare per-event predictions; if the mean absolute error on genuine data exceeds the roughly 6% level seen on AMPT, or if the predicted $v_2(p_T)$ curves deviate from measured ones beyond the stated ratios, the claimed direct transfer to LHC data fails.

Watch

Extended reading notes

Core claim

The paper claims that the azimuthal structure of a single minimum-bias event, encoded as a weighted $(\eta,\phi)$ pixel image, carries enough information for one CNN to regress $v_2$ and $b$ jointly. With the AMPT reaction plane set to zero, $v_2 = \langle \cos(2\phi)\rangle$ serves as the exact training label, and $b$ is normalized to $[0,1]$. The architecture stacks four convolutional layers with linear activations, group normalization, dropout, and L1L2 regularization, then flattens into dense layers and one or two output neurons. The result: mean absolute errors below 6% for all models, relative errors as low as 0.5% for $b$ in favorable centrality classes, preservation of the centrality and $p_T$ dependence of $v_2$, and attention maps showing the network concentrates on sparse, low-occupancy image regions. The authors present this as the first CNN that predicts both observables simultaneously.

Load-bearing premise

The load-bearing premise is that AMPT string-melting simulations of Pb-Pb collisions at 5.02 TeV produce $\eta$–$\phi$ particle distributions similar enough to real LHC data that a CNN trained only on simulated images will return trustworthy $v_2$ and $b$ values on experimental events.

Editorial extensions

If this is right

  • A single forward pass through the network returns both $v_2$ and $b$ for a given event, replacing separate estimation pipelines with one image-based regressor.
  • The centrality dependence and $p_T$ dependence of $v_2$ are preserved across all centrality classes, so the model can produce differential flow curves directly from weighted $\eta$–$\phi$ images.
  • Minimum-bias training lets one model cover central, mid-central, and peripheral collisions without retuning; the largest deviations occur in low-statistics regions such as 0–10% centrality, where $b$ still deviates by only about 0.25 fm.
  • All twenty-four input/output/bin configurations stay under 6% mean absolute error, indicating the approach is not narrowly tied to one input weighting scheme.

Reading between the lines

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

  • A testable follow-up the paper leaves implicit: because every Conv and Dense layer uses a linear activation, the whole network is formally a linear map from image pixels to outputs; comparing against a nonlinear-activation twin would show whether genuine feature learning accounts for the accuracy or a simple linear correlation in the binned images is sufficient.
  • The attention-map results suggest the network learns from sparse, low-occupancy image regions; a practical consequence is that detector acceptance or efficiency losses altering the pixel sparsity pattern could shift predictions more than changes in overall particle count.
  • The same architecture could be extended to higher harmonics ($v_3$, $v_4$) or to the symmetry-plane angle $\psi_n$, since the input image already contains the full azimuthal structure; the paper does not test these outputs.
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.

Referee Report

3 major / 5 minor

Summary. The manuscript develops a CNN-based regression to estimate the elliptic flow coefficient v2 and the impact parameter b simultaneously from weighted (eta, phi) images of charged particles in Pb-Pb collisions at sqrt(s_NN) = 5.02 TeV. The authors generate 50k minimum-bias AMPT string-melting events with the reaction plane set to zero, and build 12 models covering four input weightings (pT, mass, mT, and pT+mass) and three output configurations (v2 only, b only, both) under two bin settings. They report mean absolute errors below about 6% on a held-out AMPT test set, show the centrality and pT dependence of v2, and compare the centrality-averaged predictions with ALICE data. The paper claims that the model can be directly implemented to LHC data.

Significance. The work provides a systematic comparison of input representations and output configurations for a supervised flow and impact-parameter regression, and the attention-map analysis is a useful diagnostic. As an in-simulation benchmark, the result is plausible and the recipe is clearly described. However, the practical claim that the network can be applied directly to experimental data is not established: there is no detector simulation, the external comparison is qualitative, and the fully linear activation scheme means the specific role of the CNN architecture is not demonstrated. If a linear baseline and detector-level validation were added, the paper would be a solid benchmark study; in its current form, the significance is more modest and the headline claim overreaches.

major comments (3)
  1. [VI.2 and V.A] The claim in Section VI.2 that the model is "directly implemented" to LHC data is not supported by the presented benchmark. Training inputs are generator-level charged particles with 0.2 < pT < 5 GeV/c and |eta| < 0.8, with no detector acceptance, tracking efficiency, momentum resolution, or non-flow treatment. The comparison with ALICE data in Fig. 3 is qualitative, and the ratio panels compare CNN outputs with AMPT, not with data, so they only confirm that the CNN reproduces its training generator. Moreover, Section VI.5 explicitly states that deviations from the training pixel-value distribution will degrade performance, which contradicts the direct-implementation claim. The authors should either remove this claim or add a detector-response layer and a quantitative comparison on detector-simulated or experimental input images.
  2. [V.B] Since every Conv and Dense layer uses a linear activation and no pooling or nonlinear activation is present in the described architecture, the entire network is mathematically a linear function of the input image: linear convolutions followed by linear dense layers (GroupNorm and dropout do not restore nonlinearity). The reported MAEs could therefore likely be reproduced by a simple linear regression on the binned pixel values. Without a comparison against a linear baseline, the paper cannot support the claim that the CNN architecture is responsible for the performance or that it captures nonlinear correlations between flow and impact parameter. Please add a linear baseline or introduce nonlinear activations and compare the results.
  3. [Table I and VI.1] The central performance claim rests on the MAE values in Table I, but the normalization of these percentages is not stated. Since b is normalized to [0,1] while v2 is naturally in [0,1], a "4%" error could mean 0.04 in normalized units or 4% of the physical range, and the two are not equivalent. The paper should specify the normalization, report MAE in physical units where relevant, and provide uncertainties or run-to-run spreads for the reported values. Without this information, the claim that all models perform within 6% is ambiguous.
minor comments (5)
  1. [VI.1] The text says "all 24 models," but 4 inputs x 3 outputs gives 12 models; presumably the factor of two comes from the two bin settings. Please clarify this in the text.
  2. [Table I caption] The input labels "mass," "mT," and "both" are defined only in Section V.A; the caption should explicitly restate these definitions so the table is self-contained.
  3. [III.A and Eq. (1)] The text uses both "reaction plane angle" and "symmetry plane angle" (psi_n) without defining their relationship. In the AMPT setup the reaction plane is set to zero, so the distinction should be stated explicitly.
  4. [Fig. 3 caption] The caption should state that the ratio panels compare CNN predictions with AMPT labels, not with ALICE data. The current wording could mislead readers into thinking the overlay validates the CNN against experiment.
  5. [References] Reference [25] appears to have an inconsistent year (listed as 2013 in the bibliography but cited as 2023 in the text); please verify and correct.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the CNN benchmark is a standard supervised regression internally validated on held-out AMPT events; the unsupported LHC extrapolation is a domain-shift limitation, not a circular step.

full rationale

The paper's central result is a supervised regression benchmark: a CNN maps pT/mass/mT-weighted (eta,phi) histograms from AMPT Pb-Pb events to AMPT-computed v2 and b labels. The training/test split is explicit (Sec. V: '80% of the total event sample was used for training and validation while 20% was used for testing and evaluation'), so the reported MAEs are genuine held-out generalization numbers. The v2 label is defined from the same azimuthal distribution that forms the input (Eq. 1 with psi_n = 0, so v2 = <cos 2phi>), but that is the intended target function, not a derivation of the label from the model; the CNN must learn the mapping from the pixel images, and it is not given the label formula. No fitted parameter is renamed as a prediction, and no load-bearing conclusion is imported from a self-citation; reference [11] is contextual only. The claim that the model 'can be directly implemented' to LHC data is not supported by detector simulation or a cross-generator test, and Sec. VI.5 itself warns that deviations in the pixel-value distribution will degrade performance. That is an external-validity (domain-shift) limitation, not circularity. Hence no circular step is identified.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests entirely on supervised labels from AMPT and on hand-chosen CNN hyperparameters; the ledger lists the fitted components and the domain assumptions required for the experimental comparison.

free parameters (4)
  • CNN trainable parameters (kernels, dense weights, biases) = not released
    These are the actual fitted values; the central estimation result depends entirely on them, trained on 40K AMPT events with MSE loss.
  • Model hyperparameters = filters 16/32, dense 64/32, dropout 0.1, GroupNorm groups 4, L1L2 regularizer, batch 256, epochs 70
    Chosen by hand without a systematic sweep; the paper reports convergence but no ablation showing these choices are optimal.
  • Input bin settings = (16,16) and (32,32)
    Two grid resolutions chosen as powers of two; the paper states 16 was a lower bound and 32 an upper bound based on feature sparsity.
  • Impact parameter normalization = b scaled to [0,1]
    Ad hoc normalization to make training easier; v2 was left unchanged.
assumptions (4)
  • domain assumption AMPT string melting mode at sqrt(s_NN)=5.02 TeV gives eta-phi distributions and v2 close enough to Pb-Pb data for the method to transfer
    Used in Sections II and V as ground truth. If AMPT is inaccurate, predictions on real data inherit the bias.
  • domain assumption The reaction plane angle in AMPT is zero, so v2 = <cos(2 phi)> is the true elliptic flow label
    Section V.A states psi_n is set to zero; this removes event-plane resolution and non-flow effects present in experimental data.
  • domain assumption All charged particles with 0.2 < pT < 5 GeV/c and |eta| < 0.8 capture sufficient information
    Section V imposes these cuts; no study of acceptance dependence is provided.
  • ad hoc to paper A stack of linear-activation Conv and Dense layers can represent the mapping from images to v2 and b
    Section V.B says every Conv and Dense layer uses linear activation. Composing linear layers yields a linear map, so the model is effectively a linear regression on binned inputs; the paper does not acknowledge this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Simultaneous Estimation of Elliptic Flow Coefficient and Impact Parameter in Heavy-Ion Collisions using CNN." pith.science (2026). https://pith.science/paper/ZJNBMUJ7

@misc{pith2026241111001,
  author       = {Pith},
  title        = {Pith review of: Simultaneous Estimation of Elliptic Flow Coefficient and Impact Parameter in Heavy-Ion Collisions using CNN},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZJNBMUJ7}},
  note         = {Machine review of arXiv:2411.11001}
}
abstract

A deep learning based method with Convolutional Neural Network (CNN) algorithm is developed for simultaneous determination of the Elliptic Flow coefficient ($v_{2}$) and the Impact Parameter in Heavy-Ion Collisions at relativistic energies. The proposed CNN is trained on Pb$-$Pb collisions at $\sqrt{s_{NN}}$ = 5.02 TeV with minimum biased events simulated with the AMPT event generator. A total of twelve models were built on different input and output combinations and their performances were evaluated. The predictions of the CNN models were compared to the estimations of the simulated and experimental data. The deep learning model seems to preserve the centrality and $p_{T}$ dependence of $v_{2}$ at the LHC energy together with predicting successfully the impact parameter with low margins of error. This is the first time a CNN is built to predict both $v_{2}$ and the impact parameter simultaneously in heavy-ion system.

Figures

Figures reproduced from arXiv: 2411.11001 by the authors.

Figure 1
Figure 1. FIG. 1: The three layers of the weighted [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2: The variation of MAE as a function of the target observables for all 24 models. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3: Centrality dependence of the target observables for the model where the input has both [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: FIG. 4: Variation of [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5: Attention Maps for two different input space distributions. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6: Distribution of pixel values for [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 31 canonical work pages

  1. [1]

    It is these kernels that get updated during training

    Convolution operation - A kernel-matrix is slid across the input map to produce a lower (or same) dimensional representation of the features. It is these kernels that get updated during training. Since the same kernel con- volves all the pixels of the image, the CNN picks up spatial correlations inherent in the input image

    work page

  2. [2]

    The input image is divided into2 × 2 windows, where the largest (or average) value is chosen and taken to the next step

    Max/Average Pooling - The main purpose of using a Pooling layer is to reduce the dimen- sion of the input image. The input image is divided into2 × 2 windows, where the largest (or average) value is chosen and taken to the next step. In general, aN × N input image becomes N 2 × N 2 after one pooling step

    work page

  3. [3]

    Flattening - All the pixels of the input fea- tures are flattened to form one long vector as it is easier to manipulate to get the desired output shape. V. ANAL YSIS METHOD The AMPT event generator was used to gener- ate 50K minimum bias Pb−Pb events at√sN N= 5.02 TeV. The string melting version of AMPT ISOFT mode was used. All charged particles with tran...

    work page

  4. [4]

    One such method is class activation mapping (CAM) [25], which 7 FIG

    Activation Maps In the case of convolutional neural networks (CNNs), commonly used in visual recognition tasks, various methods have been developed to visualize their behavior. One such method is class activation mapping (CAM) [25], which 7 FIG. 3: Centrality dependence of the target observables for the model where the input has bothpT and mass-weighted l...

    work page

  5. [5]

    Each Conv layer was followed by a Group Normalization technique with the number of groups set to four

    work page

  6. [6]

    Conv layer 2 and 4 were followed by a dropout layer with a dropout rate of 0.1

    work page

  7. [7]

    The model was compiled with a Mean-Squared Error loss function and the Adam algorithm was used to update the kernel weights

    Every Conv and Dense layer had an L1L2 kernel regularizer. The model was compiled with a Mean-Squared Error loss function and the Adam algorithm was used to update the kernel weights. The Mean- Absolute Error (MAE) was also tracked while training to evaluate the learning. The model trained with a batch size of 256 for 70 epochs showed a good convergence w...

    work page

  8. [8]

    The MAE col- umn shows the error obtained while testing the models

    General Performance Table I shows the performances of all the 24 models that were built and trained for the 2 bin settings, (16 × 16) and (32× 32). The MAE col- umn shows the error obtained while testing the models. The percentage indicates the magnitude of the error but not the relative error. It can be observed that all 24 models performed well with max...

    work page

Show all 40 references
  1. [9]

    1: The three layers of the weighted(η, ϕ) space with (32, 32) bins for a single Pb−Pb collision at√sN N= 5.02 TeV

    Mean Absolute Error Figure 2 shows the variation of MAE for different values of target observables and therefore compares the performance of all the 12 models 5 FIG. 1: The three layers of the weighted(η, ϕ) space with (32, 32) bins for a single Pb−Pb collision at√sN N= 5.02 T...

  2. [10]

    The best model for each, (16,16) and (32,32), bin is shown

    pT dependence of v2 Figure 4 shows the variation ofv2 with pT for different centrality classes. The best model for each, (16,16) and (32,32), bin is shown. Usually, fewer particles are detected for highpT, suggesting that the input regions for highpT would be very sparse. Desp...

  3. [11]

    Mallick et al. , Phys. Rev. D105, 114022 (2022)

    2022

  4. [12]

    A sophisticated network like the one built in this work, performs well in regions of fewer statistics like for v2 > 0.2 and highly central ( or peripheral) collisions

    Quality over Quantity Neural nets usually rely on a large dataset to effectively extract mappings between input and output neurons. A sophisticated network like the one built in this work, performs well in regions of fewer statistics like for v2 > 0.2 and highly central ( or p...

  5. [13]

    Pei Xiang, Yuan-Sheng Zhao, Xu-Guang Huang, arXiv:2112.03824 (2021)

    2021 arXiv

  6. [14]

    Acharya et al., ALICE Collaboration, Eur

    S. Acharya et al., ALICE Collaboration, Eur. Phys. J. C 84 (2024) 813

    2024

  7. [15]

    G. F. Chapline, M. H. Johnson, E. Teller, and M. S. Weiss. Phys. Rev. D8:4302–4308 (1973)

    1973

  8. [16]

    T. D. Lee and G. C. Wick. Phys. Rev. D9:2291, 1974

    1974

  9. [17]

    Miller, Klaus Reygers, Stephen J

    Michael L. Miller, Klaus Reygers, Stephen J. Sanders, Peter Steinberg, Ann. Rev. Nucl. Part. Sci. 57, 205-243 (2007)

    2007

  10. [18]

    Aamodt et al.[The ALICE Collaboration], Phys

    K. Aamodt et al.[The ALICE Collaboration], Phys. Rev. Lett. 105, 252302 (2010)

    2010

  11. [19]

    A. M. Poskanzer and S.A. Voloshin, Phys. Rev. C58, 1671 (1998)

    1998

  12. [20]

    Aamodt et al.[ALICE Collaboration], Phys

    K. Aamodt et al.[ALICE Collaboration], Phys. Rev. Lett., 105, 252302 (2010)

    2010

  13. [21]

    Aamodt et al.,ALICE Collaboration, Phys

    K. Aamodt et al.,ALICE Collaboration, Phys. Rev. Lett., 107,032301 (2011)

    2011

  14. [22]

    Schmidhuber, Neural Networks, 61, 85–117 (2015)

    J. Schmidhuber, Neural Networks, 61, 85–117 (2015)

    2015

  15. [23]

    Adam et al., ALICE Collaboration, Phys

    J. Adam et al., ALICE Collaboration, Phys. Rev. Lett. 116, 132302(2016)

    2016

  16. [24]

    Martín Abadi et al., arXiv:1605.08695, (2016)

    2016 arXiv

  17. [25]

    X.Zhang et al.,Phys. Rev. C105, 034611 (2022)

    2022

  18. [26]

    Selvaraju, Ramprasaath R. et al. , International Journal of Computer Vision, 128, 2, 336–359 (2019). 11

    2019

  19. [27]

    In fact, to test the robustness of the models built in this work, one can generate data from different 10 event generators and analyze the performance of the predictions

    can also be used for training but for those datasets, the quality matters over quantity. In fact, to test the robustness of the models built in this work, one can generate data from different 10 event generators and analyze the performance of the predictions. The (η − ϕ) distr...

    2022

  20. [28]

    Zi-Wei Lin, Che Ming Ko, Bao-An Li, Bin Zhang, Subrata Pal, Phys.Rev.C 72,064901 (2005)

    2005

  21. [29]

    Gyulassy and X

    M. Gyulassy and X. N. Wang, Comput. Phys. Commun. 83, 307 (1994)

    1994

  22. [30]

    Zhang, Comput

    B. Zhang, Comput. Phys. Commun. 109, 193 (1998)

    1998

  23. [31]

    B. Li, A. T. Sustich, B. Zhang, and C. M. Ko, Int. J. Mod. Phys. E 10, 267 (2001)

    2001

  24. [32]

    All the twelve models were constructed with the same architecture to ensure uniformity while comparing the performance of these models

    bins weighted by bothpT and mass. All the twelve models were constructed with the same architecture to ensure uniformity while comparing the performance of these models. The first and the second convolutional(Conv) layer had 16 filters. The third and the fourth convolutional l...

  25. [33]

    Borghini, P

    N. Borghini, P. M. Dinhand J. Y. Ollitrault, Phys. Rev. C63, 054906 (2001)

    2001

  26. [34]

    R. S. Bhalerao, J. Y. Ollitrault, S.Pal and D. Teaney, Phys. Rev. Lett. 114, 152301 (2015)

    2015

  27. [35]

    Coadou, arXiv:2206.09645 (2022)

    Y. Coadou, arXiv:2206.09645 (2022)

    2022 arXiv

  28. [36]

    Goodfellow et al., arXiv:1406.2661 (2014)

    Ian J. Goodfellow et al., arXiv:1406.2661 (2014)

    2014 arXiv

  29. [37]

    Dalya Baron, arXiv:1904.07248 (2019)

    2019 arXiv

  30. [38]

    Jeff Heaton, arXiv:2009.05673,(2020)

    2020 arXiv

  31. [39]

    Pham Thi Minh Anh, arXiv:2309.14304 (2013)

    2013 arXiv

  32. [40]

    Lokhtin, L.V

    I.P. Lokhtin, L.V. Malinina, S.V. Petrushanko, A.M. Snigirev, I. Arsene, K. Tywoniuk, Comput. Phys. Commun. 180,779-799 (2009)

    2009

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.