{"id":"9a7d840e-1bc6-4309-9a89-bfff28adf90a","arxiv_id":"2607.06897","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"A LightGBM model trained on pion observables from one transport model predicts impact parameters in Au+Au collisions at 4 and 11 GeV with 0.2-0.4 fm error, generalizing to data from other models where polynomial fits fail.","lead":"This paper trains a machine-learning model on simulated heavy-ion collision data from three different transport models to predict the collision impact parameter, finding it generalizes across models with 0.2-0.4 fm error. A smart generalist might read it to see if ML can replace model-dependent calibration steps in nuclear physics experiments.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"Cross-model ML success with absolute-valued pion features is empirically demonstrated but mechanistically unexplained, leaving the robustness claim underdetermined.","rationale":"The reader correctly identified the load-bearing assumption: that the six pion features provide model-invariant physics rather than correlated model artifacts. I agree this is the weakest link. The paper demonstrates cross-model generalization empirically within a set of three transport models that share significant physics infrastructure, but does not investigate the mechanism or test whether the generalization extends beyond this narrow model family. The extrapolation to experimental data is explicitly hedged ('potential,' 'indicates'), which is appropriate. However, the CONDITIONAL verdict is already calibrated to this gap: the simulation-based results are solid and useful, while the broader claims await experimental validation. My concern does not move the verdict further toward REJECT because the within-model-space evidence is genuinely supportive and the paper does not overclaim beyond what the simulations show. The concern does suggest a specific, feasible test (feature perturbation + SHAP analysis) that would either strengthen or weaken the robustness claim, but its absence is consistent with a CONDITIONAL assessment rather than grounds for rejection. The lack of released code/data, noted by the reader, further supports keeping the verdict conditional rather than upgrading to ACCEPT.","tokens_in":14420,"tokens_out":2690,"duration_ms":141193,"concrete_test":"Perform a feature perturbation test: train LightGBM on UrQMD/CH, then test on AMPT data where each of the six features is independently rescaled by a factor drawn uniformly from [0.7, 1.3] (simulating model/detector systematic shifts beyond the tested trio). If MAE remains below ~0.5 fm, the ML is learning scale-invariant inter-feature patterns (supporting physics-learning). If MAE degrades substantially (e.g., >0.8 fm), the cross-model success depends on the specific scale relationships shared by UrQMD/AMPT/JAM, and the generalization claim weakens. Additionally, compute SHAP feature-importance values for the cross-model predictions to identify which features drive generalization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that LightGBM achieves MAE 0.2–0.4 fm even when trained on one transport model and tested on another. The six input features (N_π±, n_π±, p_t^π±) are all absolute quantities. Figure 1 shows these yields differ by up to ~60% between models (e.g., UrQMD/CH vs. AMPT at √sNN = 4 GeV). The polynomial-fit baseline fails across models precisely because of these scale differences (Figs. 4–5, dashed lines). Yet ML with the same absolute features succeeds (solid lines). The paper does not investigate why. This matters because the three models (UrQMD, AMPT, JAM) share substantial physics: similar nuclear initializations (Woods-Saxon or hard-sphere), cascade/mean-field propagation, and resonance-based particle production. If LightGBM is exploiting subtle inter-feature correlations that are shared by these three models but not guaranteed to hold in experimental data (where detector acceptance, efficiency, and backgrounds modify feature distributions), then the 'strong robustness' claim and the extrapolation to real data are unsupported. The paper itself acknowledges the features are 'very different from each other' across models (Sec. II.A) but does not perform feature-importance analysis, scale-invariance tests, or perturbation studies to distinguish physics-learning from shared-artifact exploitation. Without understanding the mechanism, the cross-model MAE could reflect the narrowness of the tested model space rather than genuine model independence.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"This manuscript applies supervised (LightGBM regression and classification) and unsupervised (K-means clustering) machine-learning methods to reconstruct the impact parameter in Au+Au collisions at sqrt(s_NN) = 4 and 11 GeV. The central result is that LightGBM achieves MAE of 0.2–0.4 fm even when trained on data from one transport model (UrQMD, AMPT, or JAM) and tested on another, whereas a conventional polynomial fit to charged-pion multiplicity fails in cross-model application. The K-means algorithm is also shown to autonomously partition events into six centrality classes. The methodology is clearly described, the cross-model validation is a well-designed test of robustness, and the comparison with a traditional baseline is appropriate. The paper represents a useful contribution to the growing literature on ML-based centrality determination at low-to-intermediate energies.","tokens_in":15251,"tokens_out":1225,"duration_ms":136782,"significance":"The cross-model validation across three mainstream transport models (UrQMD, AMPT, JAM) with different initialization and propagation modes is a genuine strength and goes beyond single-model ML studies. The finding that ML generalizes across models where polynomial fitting does not is the key novel result. The unsupervised K-means analysis provides an additional, model-independent cross-check. The six input features (charged-pion yields, mid-rapidity yields, and total transverse momenta) are experimentally accessible, enhancing practical relevance for STAR-FXT, NICA/MPD, and FAIR/CBM.","major_comments":[{"comment":"Section III.A, Figs. 4–5: The central claim of cross-model robustness rests on the observation that LightGBM achieves low MAE even when trained on one model and tested on another, despite the fact that the six input features are absolute quantities whose distributions differ by up to ~60% across models (Fig. 1, Sec. II.A). The polynomial-fit baseline fails precisely because of these scale differences, yet ML with the same absolute features succeeds. The paper does not investigate the mechanism behind this success. A feature-importance analysis (e.g., SHAP or gain-based importance from LightGBM) would help clarify whether the algorithm is exploiting inter-feature correlations that reflect genuine physics or shared model artifacts. Without this, it is difficult to assess whether the cross-model MAE reflects genuine model independence or the narrowness of the tested model space, which is a載","section":null},{"comment":"Section III.C, Fig. 10: The K-means clustering results are presented only for UrQMD/CH data. Since the paper's central theme is cross-model robustness, the absence of K-means results for AMPT and JAM is a gap. Showing that the unsupervised clustering produces comparable centrality partitions across at least one additional model would strengthen the claim that the method is model-independent. At minimum, the authors should discuss whether the cluster centroids and b-distributions are expected to be stable across models given the known yield differences shown in Fig. 1.","section":null}],"minor_comments":[{"comment":"Section II.A: The selection of six pion-related features is motivated, but the statement that observables like large-fragment yields and charged-particle multiplicity are 'highly model-dependent' and therefore excluded could be strengthened by briefly quantifying or citing the model dependence.","section":null},{"comment":"Section II.B: The statement that varying LightGBM hyperparameters 'did not significantly alter the results' should be supported with at least one quantitative example (e.g., MAE change for a specific parameter variation) in a footnote or supplementary table.","section":null},{"comment":"Section III.B, Figs. 6–9: The claim that deviations in centrality-class fractions are 'controlled at around 1%' is stated in the text but not tabulated. A small table comparing predicted vs. true fractions for each class and model pair would make this claim verifiable.","section":null},{"comment":"The abstract states the method 'has the potential to be generalized to handle real experimental data.' Given that no detector effects, acceptance cuts, or efficiency corrections are included in the present study, this claim should be softened to acknowledge these missing elements explicitly.","section":null},{"comment":"Section I: The phrase 'the classical picture is no longer valid' regarding the de Broglie wavelength at low energies is somewhat imprecise; a brief clarification of the energy scale at which this concern applies versus the energies studied here (4 and 11 GeV) would be appropriate.","section":null},{"comment":"Figures 4 and 5: The panel labels (a)–(l) are small and the distinction between solid (ML) and dashed (polynomial) lines is difficult to read in the peripheral region where statistics are low. Enlarging the legend or using thicker lines would improve readability.","section":null},{"comment":"Reference [21] is cited with a 2026 publication date and arXiv number 2601.20491; the editor should verify the publication status.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The reader's concern about whether the ML algorithm is learning physics versus shared model artifacts is legitimate and is the most important issue to address. However, I assess this as a minor revision rather than major because the cross-model validation itself is a meaningful test, and a feature-importance analysis or scale-invariance check would likely suffice to address the concern. The three models do share physics assumptions, but they also differ in initialization, mean-field treatment, and particle-production mechanisms, so the test is not vacuous. The authors should be asked to add mechanistic insight but the core result is publishable with this addition."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"The main thing to know: this paper shows that LightGBM trained on one transport model (UrQMD, AMPT, or JAM) predicts impact parameter in Au+Au at 4 and 11 GeV with MAE 0.2–0.4 fm even when tested on data from a different model. That cross-model generalization is the real result, and the polynomial-fit baseline failing where ML succeeds is a clean demonstration of the difference. The classification and K-means clustering results are secondary but consistent. This is a legitimate empirical contribution at a relevant energy regime (STAR-FXT, NICA, CBM). The method is not new—the authors cite their own prior work and studies going back to the 1990s—but the cross-model validation at these energies is new and useful. The experimental-style observables chosen (pion yields, mid-rapidity yields, total transverse momentum) are reasonable and practically measurable. The six-model setup with different initializations and propagation modes gives a fair spread of model space. The stress-test concern about absolute-valued features is worth raising but I think it is partially overblown. LightGBM is a tree-based method that splits on feature thresholds; it can learn relative patterns across features (e.g., the ratio between mid-rapidity and total yield) even when absolute scales differ between models. That is likely why it generalizes where a single-feature polynomial fit does not. The paper does not explain this mechanism, and it should—but the empirical result stands on its own as a finding worth reporting. The real soft spot is the extrapolation to experimental data. The three transport models share a lot of physics: similar initializations, resonance-based production, cascade or mean-field propagation. The cross-model MAE could reflect the narrowness of this model space rather than genuine model independence. No feature-importance analysis, no scale-invariance test, no perturbation study is performed to distinguish physics-learning from shared-artifact exploitation. The claim of ","headline":"Solid cross-model ML study for impact parameter, but the ","tokens_in":15233,"tokens_out":484,"would_cite":false,"duration_ms":78044,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["25.75.-q","25.75.Ag","25.75.Gz","24.10.Lx"],"model":"glm-5.2","headline":"ML pins collision geometry to 0.3 fm across rival models","keywords":["heavy-ion collisions","impact parameter","machine learning","LightGBM","transport models","centrality determination","UrQMD","AMPT"],"falsifier":"Train the ML algorithm on data from two of the three transport models and test on the third, then repeat with a fourth transport model (e.g., SMASH or GiBUU) that uses substantially different hadronic physics. If the MAE jumps well above 0.4 fm, the claimed robustness is an artifact of the tested model space rather than a model-independent physical mapping.","tokens_in":14713,"feed_emoji":"","tokens_out":1174,"duration_ms":98726,"temperature":0.7,"pith_summary":"The paper asks whether a machine-learning model trained on simulated heavy-ion collision data from one transport model can reliably predict the impact parameter (the transverse separation of the two colliding nuclei) when tested on data from a different transport model. This matters because the impact parameter is not directly measurable in experiment and is conventionally inferred through model-dependent fits that fail when the simulation model changes. The authors train a gradient-boosted-tree algorithm (LightGBM) on six pion-related observables — charged pion yields, mid-rapidity yields, and total transverse momenta — generated by three independent transport models (UrQMD, AMPT, and JAM) at collision energies of 4 and 11 GeV. They find that the mean absolute error in reconstructing the impact parameter stays between 0.2 and 0.4 femtometers even in cross-model tests, where the training model differs from the testing model. By contrast, the standard polynomial-fit method relating pion multiplicity to impact parameter collapses entirely in cross-model application. In a classification task, the ML method assigns events to six centrality classes with roughly one-percent-level accuracy across models. An unsupervised K-means clustering on the same observables, given no impact-parameter labels at all, autonomously separates events into six clusters that correspond to physically meaningful centrality ranges. The central claim is that the ML algorithm learns a model-independent mapping from pion observables to collision geometry, and that this mapping is robust enough to eventually be applied to real experimental data.","feed_headline":"ML pins collision geometry to 0.3 fm across rival models","feed_subtitle":"A tree-based algorithm trained on one nuclear transport model predicts impact parameters from another, beating conventional fits and hinting","key_machinery":"LightGBM (gradient-boosted decision trees) for supervised regression and classification; K-means for unsupervised clustering; six input features: charged pion yields, mid-rapidity charged pion yields, and total transverse momenta of charged pions; three transport models (UrQMD, AMPT, JAM) with different initialization and mean-field configurations; mean absolute error (MAE) and signed mean error as evaluation metrics.","core_discovery":"A gradient-boosted-tree ML algorithm trained on six pion-related observables from one transport model predicts the impact parameter of heavy-ion collisions at 4 and 11 GeV with 0.2–0.4 fm accuracy even when tested on data from different transport models, while the conventional polynomial-fit method fails in this cross-model setting. Unsupervised K-means clustering on the same observables recovers six centrality classes without any model-based binning.","pith_inferences":["The claim of model-independence rests on three transport models that share underlying physics assumptions; a truly model-independent mapping would need to be validated against models with more divergent physics, such as those incorporating different baryon transport or resonance dynamics.","The six pion observables may contain model-specific correlations that happen to be consistent across the tested models but could break when applied to experimental data where detector effects, acceptance cuts, and background processes introduce correlations absent from the simulations.","If the ML algorithm is learning a physical relationship rather than model noise, one would expect the learned feature importances to reflect known physics — for instance, that multiplicity and transverse momentum carry complementary geometric information — which could be tested by examining the learned decision-tree splits."],"forward_implications":["If the cross-model robustness holds, the method could replace Glauber-model-based centrality determination at intermediate energies where the eikonal approximation underlying the Glauber model is questionable.","The unsupervised K-means result suggests that centrality structure is encoded in pion observables in a way that does not require a model to define bin boundaries, which could make centrality calibration less model-dependent in future experiments at FAIR/CBM, NICA/MPD, and STAR-FXT.","The approach could be extended to additional observables (light charged particle yields, correlations) and additional transport models to further test whether the learned mapping is genuinely model-independent or an artifact of the limited model space tested.","If the method generalizes to real experimental data, it would reduce one of the dominant systematic uncertainties in extracting nuclear equation-of-state properties from heavy-ion collision measurements."],"fun_headline_variants":["ML predicts collision impact parameter within 0.4 fm across transport models","Tree-based ML reconstructs impact parameter from pion observables across rival models","Cross-model ML hits 0.2–0.4 fm accuracy for heavy-ion impact parameter","K-means autonomously recovers six centrality classes without model-based binning","ML trained on one transport model generalizes to impact-parameter prediction in others"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The paper assumes that six pion-related observables contain enough information for the ML algorithm to learn a genuinely model-independent mapping to the impact parameter, rather than learning model-specific patterns that happen to correlate with the impact parameter across the three models tested. The cross-model robustness is demonstrated only within a limited family of transport models that may share hidden commonalities.","fun_headline_variants_meta":{"raw":{"variants":["ML predicts collision impact parameter within 0.4 fm across transport models","Tree-based ML reconstructs impact parameter from pion observables across rival models","Cross-model ML hits 0.2–0.4 fm accuracy for heavy-ion impact parameter","K-means autonomously recovers six centrality classes without model-based binning","ML trained on one transport model generalizes to impact-parameter prediction in others"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":731,"prompt_tokens":628,"completion_tokens":103,"prompt_tokens_details":null},"tokens_in":628,"tokens_out":103,"duration_ms":28845,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T23:23:08.150803+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"Train the ML algorithm on data from two of the three transport models and test on the third, then repeat with a fourth transport model (e.g., SMASH or GiBUU) that uses substantially different hadronic physics. If the MAE jumps well above 0.4 fm, the claimed robustness is an artifact of the tested model space rather than a model-independent physical mapping.","supporting_citations":[],"review_version":1}