{"id":"8d095b72-30f0-4e9a-9c39-cacb46002763","arxiv_id":"2411.11001","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A CNN trained on AMPT-simulated Pb-Pb collisions reproduces elliptic flow v2 and impact parameter from eta-phi images with under 6% error, but only within the simulation's distribution.","lead":"A neural network was trained to look at images of particle directions from simulated lead-lead collisions and to output two numbers at once: the shape of the particle spray (elliptic flow) and how head-on the collision was. It works well on the simulation it was trained on; applying it to real LHC data would need extra validation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The in-simulation benchmark is plausible, but the direct-LHC claim rests entirely on generator-level AMPT images with no detector-response or cross-generator test, so it is not established.","rationale":"The reader's weakest_assumption identifies the same load-bearing gap: the model is trained and tested only on generator-level AMPT images, while the paper's headline practical claim is direct transfer to LHC data. The in-simulation benchmark is self-consistent and is not threatened by the absence of detector simulation, which is why I do not recommend a harsher verdict; the low MAE on held-out AMPT events supports the internal claim. However, the direct-implementation statement is exactly the step that requires validation on inputs whose pixel distribution differs from AMPT, and the paper itself notes in Section VI.5 that such deviations degrade performance. I considered whether the linear activations of all Conv and Dense layers are the more serious issue, but GroupNorm and dropout introduce nonlinearity, and the reported in-simulation performance is what it is, so this is secondary. I also considered label 'leakage', since v2 is computed from the same particles that form the image; this is the intended supervised setup rather than a flaw. The concrete detector-response test would settle whether the practical claim holds, and until that test is performed the conditional verdict is appropriate.","tokens_in":7836,"tokens_out":6871,"duration_ms":115366,"concrete_test":"Apply an ALICE-like detector response to held-out AMPT test events: keep only ~85% of tracks, smear pT with sigma(pT)/pT = 1.5%, and rebuild the pT+mass weighted (32,32) images; keep the original unsmeared v2 and b as labels. Evaluate the frozen best-performing CNN on these smeared images and compare MAE and centrality-binned v2 ratios to the quoted Table I values. If MAE degrades by more than a factor of two or the centrality dependence distorts beyond the quoted 6% level, the direct-LHC claim is not supported; if performance is essentially unchanged, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim has two parts: (i) a CNN can regress AMPT's v2 and b labels from weighted eta-phi images, and (ii) the model can be 'directly implemented' to LHC data. Part (i) is supported by held-out AMPT MAE values. Part (ii) is load-bearing for the stated practical value and is unsupported. Inputs are generator-level charged particles with 0.2 < pT < 5.0 GeV/c and |eta| < 0.8 (Section V), with no detector acceptance, tracking efficiency, pT resolution, or background/non-flow treatment. All training and test labels are computed from AMPT events with the reaction plane set to zero. Section VI.2 qualitatively overlays CNN v2 with ALICE data, but this cannot validate the regression because agreement depends on AMPT itself matching data; the ratio panels show CNN/AMPT agreement, not experimental agreement. Section VI.5 explicitly states that deviations in the pixel-value distribution will degrade performance and recommends training on experimental distributions, which contradicts the 'direct implementation' claim. Without a detector-simulation layer or a cross-generator test, the CNN's behavior on real LHC images is unknown.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":8105,"tokens_out":5433,"duration_ms":60297,"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":[{"comment":"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.","section":"VI.2 and V.A"},{"comment":"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.","section":"V.B"},{"comment":"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.","section":"Table I and VI.1"}],"minor_comments":[{"comment":"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.","section":"VI.1"},{"comment":"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.","section":"Table I caption"},{"comment":"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.","section":"III.A and Eq. (1)"},{"comment":"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.","section":"Fig. 3 caption"},{"comment":"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.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The gap between the supported in-simulation regression and the direct-LHC claim is the main concern. I would ask the editor to require a linear baseline and at least one detector-level or cross-generator test before reconsidering publication. The novelty claim of being the first CNN for simultaneous v2 and b prediction should also be checked carefully against Refs. [11-13] and related literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth your time as an incremental ML-in-heavy-ion paper, not as a new physics capability. The actual content: a single CNN with weighted (eta,phi) images predicts AMPT's v2 and b labels on held-out minimum-bias Pb-Pb events at 5.02 TeV, with mean absolute errors below 6%. Table I and Figures 3-4 show the network reproduces the centrality and pT dependence of v2. The architecture sweep — four input encodings times three output heads times two bin settings — is systematic, and the attention-map section is a nice touch. That part is honest and reproducible in principle. The soft spots are the ones the stress-test note names. First, the direct implementation claim in Sections V.A and VI.2 is not supported: inputs are generator-level charged particles with no detector acceptance, efficiency, pT resolution, or non-flow treatment. The ALICE overlay in Figure 3 cannot validate the regression; it only shows that AMPT roughly matches data and the CNN matches AMPT. Section VI.5 itself says deviations in pixel distributions will degrade performance and recommends training on experimental distributions, which undercuts the direct implementation claim. Second, all Conv and Dense layers use linear activations, so the network collapses to a linear transformation of the input image. No comparison against a linear baseline is given, so the CNN contribution beyond linear regression is unquantified. Third, no uncertainty estimates and no code or data are released, so the exact numbers cannot be checked. Fourth, the 'first time' framing is overstated: references [11-13] already train CNNs for v2 and b separately; the multi-task output and kinematic-weighted images are a routine extension, not a new observable. I agree with the reader's take. The in-simulation benchmark claim is plausible and supported. The burden is low because this is supervised fitting, but that is fine for a benchmark paper if framed as such. No circularity beyond self-citation, which is not a problem here. The paper deserves a serious referee — the benchmark is useful to the ML/heavy-ion community if claims are recalibrated and the detector-simulation question is addressed. I would not cite it until code and data appear and the linear-baseline comparison is done.","headline":"Solid in-simulation benchmark, but the direct-to-LHC claim is not established and the linear-activation network quietly limits the method.","tokens_in":8594,"tokens_out":2118,"would_cite":false,"duration_ms":23016,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["25.75.-q","25.75.Ld"],"model":"deepseek-v4-flash","headline":"A single CNN trained on simulated Pb-Pb collisions predicts both elliptic flow and impact parameter with less than six percent error.","keywords":["deep learning","convolutional neural network","elliptic flow","impact parameter","heavy-ion collisions","AMPT event generator","Pb-Pb collisions at 5.02 TeV","image-based regression"],"falsifier":"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.","tokens_in":7644,"feed_emoji":"⚛️","tokens_out":12975,"duration_ms":115752,"temperature":0.7,"pith_summary":"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.","feed_headline":"One image yields elliptic flow and impact parameter within 6%","feed_subtitle":"A single CNN trained on simulated Pb-Pb collisions outputs both at once and keeps their centrality and momentum trends.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Provides the string-melting AMPT Pb-Pb events at 5.02 TeV from which the CNN learns its $v_2$ and $b$ labels.","marker":"[14]"},{"why":"Supplies the experimental centrality dependence of $v_2$ used to judge whether the CNN predictions match ALICE data.","marker":"[10]"},{"why":"Establishes the Glauber relation between impact parameter and participating nucleons/binary collisions that underlies the $b$ label.","marker":"[4]"},{"why":"Gives the multiplicity-based estimate of centrality that replaces direct impact-parameter measurement in experiments.","marker":"[5]"},{"why":"Defines the event-plane method used to compute the $v_2$ target values from the simulated events.","marker":"[6]"}],"fun_headline_variants":["One CNN, two observables: v2 and impact parameter in one pass","Single event image feeds CNN that predicts both v2 and b","Joint v2 and impact parameter from one Pb-Pb image via CNN","CNN extracts elliptic flow and centrality from a single image","One image, two answers: v2 and impact parameter together"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["One CNN, two observables: v2 and impact parameter in one pass","Single event image feeds CNN that predicts both v2 and b","Joint v2 and impact parameter from one Pb-Pb image via CNN","CNN extracts elliptic flow and centrality from a single image","One image, two answers: v2 and impact parameter together"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001099,"raw_usage":{"total_tokens":4571,"prompt_tokens":917,"completion_tokens":3654,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":3565}},"tokens_in":533,"tokens_out":3654,"duration_ms":26610,"temperature":1.0,"reasoning_tokens":3565,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:01:31.128215+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Acharya et al., ALICE Collaboration, Eur","cited_arxiv_id":null,"evidence_quote":"Provides the string-melting AMPT Pb-Pb events at 5.02 TeV from which the CNN learns its $v_2$ and $b$ labels."},{"cited_title":"The best model for each, (16,16) and (32,32), bin is shown","cited_arxiv_id":null,"evidence_quote":"Supplies the experimental centrality dependence of $v_2$ used to judge whether the CNN predictions match ALICE data."},{"cited_title":"One such method is class activation mapping (CAM) [25], which 7 FIG","cited_arxiv_id":null,"evidence_quote":"Establishes the Glauber relation between impact parameter and participating nucleons/binary collisions that underlies the $b$ label."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the multiplicity-based estimate of centrality that replaces direct impact-parameter measurement in experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the event-plane method used to compute the $v_2$ target values from the simulated events."}],"review_version":1}