{"id":"63177c12-4656-4383-b8f2-8f6b574d62ec","arxiv_id":"2505.23920","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A Gaussian Mixture Model decomposition of 28Si to 7alpha excitation spectra finds six peaks per dataset near the energies predicted for toroidal states, but the peaks are not tested against a null background hypothesis.","lead":"This paper applies a machine-learning curve-fitting method, a Gaussian Mixture Model, to excitation-energy spectra of silicon-28 nuclei that break into seven alpha particles, and reports peaks near energies predicted for exotic toroidal (donut-shaped) nuclear states. It also presents new model calculations of the collision that produces these fragments.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The GMM analysis is validated only on data already built from Gaussians; a smooth null spectrum with the same statistics may yield equally stable Gaussian components, so the toroidal interpretation is not yet established.","rationale":"The reader's weakest-assumption statement identifies the same load-bearing concern: the GMM decomposition is only meaningful if Gaussian components in a smooth histogram correspond to physical resonances, and the HαC validation is circular because that dataset was constructed from Gaussians. My read does not move the verdict; the paper remains conditionally acceptable pending a null-hypothesis test and honest uncertainty quantification. The time-split analysis is a genuine internal consistency check and the new cHαC calculations are useful, but neither addresses the false-positive rate of the method on a featureless spectrum. A Monte Carlo null test would settle whether the near-114 and near-138 MeV centroids are special or an artifact of the fitting procedure.","tokens_in":10664,"tokens_out":3623,"duration_ms":41987,"concrete_test":"Generate 1000 synthetic null spectra from a smooth model with no resonances, e.g., an exponential phase-space factor plus a low-order polynomial continuum, using the same excitation-energy range, 1.25 MeV binning, and total counts as the Hannaman data, with Poisson fluctuations per bin. Run the identical Section II pipeline on each synthetic spectrum: generate 10^6 points, initialize means via Eq. (2), choose NG by chi-square_nu close to 1, and apply the same stability check by splitting each realization into eight equal subsamples. Record how often a stable Gaussian centroid falls within 2 MeV or one HWHM of 114 and 138 MeV. If the false-positive rate is comparable to the rate in the real data, the claim that these peaks are toroidal resonances is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on treating GMM-decomposed Gaussians as physical resonances, but the only method validation in Section II uses the rotating-silicon HαC spectrum of Ref. [21], which is itself a sum of known Gaussians plus a smooth phase-space factor. That test shows the algorithm can recover injected Gaussian components, not that it discriminates real resonances from arbitrary smooth structure. In Eq. (3), the number of Gaussians is selected by requiring reduced chi-square near unity; this is necessary but not sufficient, since a flexible mixture can make chi-square near unity for a featureless continuum as well. The time-split stability check in Figs. 6 and 7 demonstrates reproducibility of the fitted decomposition within the Hannaman data, but reproducibility is not diagnosis: a smooth background can also produce stable Gaussian centroids when the same initialization and selection rules are used. The conclusion that the extracted peaks are 'confirmed underlying structure' therefore outruns the evidence. The arbitrary placement of out-of-range peaks in Fig. 8(b) further weakens the cross-dataset comparison, but the core issue is the missing null-hypothesis test.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a Gaussian Mixture Model (GMM) analysis of 7-alpha breakup excitation-energy spectra and applies it to the Cao and Hannaman experimental datasets, to the rotating-silicon HαC training data of Ref. [21], and to new collision HαC (cHαC) calculations performed by the authors. The central claim, stated in the abstract and conclusion, is that in all examined datasets the GMM decomposition reveals Gaussian components whose centroids lie close to the energies predicted for toroidal states of 28Si, notably near 114 MeV and 138 MeV, and that the time-ordered partitions of the Hannaman data show that these components are statistically stable. The paper also criticizes the 9th-order polynomial background-subtraction method used by Hannaman et al. and argues that the GMM approach resolves the dispute between Refs. [15] and [17] in favor of real resonant structure.","tokens_in":10868,"tokens_out":4105,"duration_ms":38562,"significance":"If the central claim were fully established, the paper would be a significant contribution to the toroidal-state controversy for 28Si and would offer a reusable analysis tool for sparse excitation-energy spectra. The new cHαC collision calculations are a useful theoretical addition, and the time-partition stability check in Figs. 6 and 7 is a commendable attempt to address statistical significance. However, the paper's diagnostic engine, the GMM decomposition, is validated only on data that are themselves sums of Gaussians, and no null-background test is provided. Because the interpretation of stable Gaussian components as physical resonances is the load-bearing assumption, the paper currently overclaims its conclusion.","major_comments":[{"comment":"The criterion that the reduced chi-square be close to unity is necessary but not sufficient to identify resonances; a flexible Gaussian mixture can also achieve chi-square near unity for a featureless continuum. The paper does not include a null-hypothesis test in which the same pipeline (binning, Np resampling, EM initialization, and N_G selection) is applied to a smooth background-only spectrum with similar statistics. Because the abstract and Section V conclude 'underlying structure' from this decomposition, the absence of such a test is load-bearing. Please add at least one null test, for example a phase-space or smoothed polynomial continuum with the same binning and errors, and report the resulting centroid stability and proximity to the x=y line.","section":"Sec. II, Eq. (3)"},{"comment":"The explicit statement that peaks outside the prediction limits are 'positioned arbitrarily close to the x=y line' means that part of the apparent agreement in panel (b) is by construction. This directly weakens the cross-dataset comparison with the toroidal predictions. Please report the unadjusted centroid values for all datasets, show the comparison without arbitrary repositioning, and state explicitly how many components fall outside the prediction limits and at what energies.","section":"Sec. IV, Fig. 8(b)"},{"comment":"The method validation on the rotating-silicon HαC data of Ref. [21] is not an external or blind test because that spectrum is itself generated as a sum of Gaussian components plus a phase-space factor. Recovering those injected Gaussians demonstrates that the algorithm can decompose a mixture of Gaussians, but it does not test the central diagnostic assumption that stable GMM components in an experimental histogram correspond to physical resonances rather than arbitrary smooth structure. Add at least one validation case with a known non-Gaussian background and, ideally, an independent experimental spectrum with confirmed resonances.","section":"Sec. II, Fig. 1"},{"comment":"The resampling step generates Np=10^6 points 'from a normal distribution around the center of each bin', but the standard deviation of these normal distributions is not specified. If the width is tied to the bin size, the procedure can artificially broaden or smooth the histogram and can introduce smooth Gaussian-like components into a featureless spectrum. Please specify the resampling width, justify the choice, and report the sensitivity of the extracted centroids and of N_G to Np and to the resampling width.","section":"Sec. II, Eq. (1)"}],"minor_comments":[{"comment":"There are several typographical and formatting errors, including 'Stasczak' for Staszczak, 'a only variation' for 'only a variation', 'un-binned' for 'unbinned', and inconsistent spacing in 'F AUST' and '7αcross section'.","section":"General"},{"comment":"The scaling of the Hannaman data by a factor of 2x10^-6 (mb/MeV) to the low-energy region of the Cao data is described only briefly; the units, the fitting procedure, and the uncertainty of this normalization should be stated, since the quantitative comparison of calculated and experimental cross sections depends on it.","section":"Sec. III and Fig. 3"},{"comment":"The phrase 'novel Artificial Intelligence (AI)-based machine learning method' overstates the degree of novelty; GMM with expectation-maximization is a standard unsupervised technique, and the paper would be clearer if it described the specific new elements beyond applying scikit-learn with a chosen initialization and chi-square selection.","section":"Sec. I"}],"recommendation":"major_revision","confidential_remarks":"The central claim is not yet supported because the principal diagnostic, the stability of GMM components, has not been tested against a null background. The authors' own statement about arbitrary placement in Fig. 8(b) is an additional concern. However, the new cHαC calculations and the time-partition analysis are useful contributions, and the missing tests appear to be feasible within the manuscript's scope. I therefore recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the cHαC collision model output and the GMM decomposition applied to it, along with the time-split stability test on the Hannaman data. The method section is readable, and the authors are candid about the drawbacks of polynomial fitting. The HαC training test in Fig. 1(c) shows the algorithm can recover known Gaussian constituents, and that is a useful sanity check.\n\nBut the central claim that these Gaussians are physical toroidal resonances outruns the evidence. The validation is circular in an important sense: the rotating-silicon HαC data used for training were themselves constructed as a sum of Gaussians plus a smooth phase-space factor. Recovering Gaussians from such data tests the curve-fitting bookkeeping, not the ability to discriminate resonances from arbitrary smooth structure. The chi-square selection criterion (Eq. 3) is necessary but far from sufficient; a flexible mixture can make chi-square near unity for a featureless continuum. The time-split stability in Figs. 6–7 demonstrates reproducibility, but reproducible is not the same as diagnosed.\n\nThe most serious soft spot is the explicit statement in Section IV and Fig. 8(b) that peaks outside the prediction limits are placed arbitrarily close to the x=y line. That panel does not show what it visually implies, and it weakens the cross-dataset comparison substantially. The scaling of the Hannaman data by an arbitrary factor of 2e-6 is also not justified beyond aligning tails. The uniform 5% error assumption and the 10^6-point resampling around bin centers add further smoothing that could manufacture Gaussian components.\n\nAll that said, the paper has real value. The cHαC collision spectra are new model output, and the GMM approach is a legitimate analytical tool worth exploring. The authors are honest about the limitations in low-statistics regions. The question of toroidal states in 28Si is genuinely contested, and this work contributes new information even if its interpretation is not yet established.\n\nThis deserves a serious referee. The right revision would add a null-hypothesis test—decompose smooth synthetic spectra with the same algorithm, selection rules, and statistics—plus honest plotting in Fig. 8 and proper uncertainty quantification. With those changes, the toroidal claim could be evaluated on firmer ground. As it stands, I would not cite the main conclusion, but I would take the cHαC spectra seriously as a benchmark.\n\nI'd send it to review with a clear request for those additions.","headline":"New cHαC collision results and a clear GMM write-up, but the toroidal claim rests on a method never tested against a smooth background, and Fig. 8(b) admits arbitrary peak placement.","tokens_in":11433,"tokens_out":1945,"would_cite":false,"duration_ms":21616,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["25.70.-z","21.60.-n","25.70.Pq"],"model":"deepseek-v4-flash","headline":"The paper claims that Gaussian Mixture Model analysis of the 7-alpha excitation spectra from the 28Si + 12C reaction at 35 MeV/A reveals reproducible components near 114 and 138 MeV in both published experimental datasets, indicating…","keywords":["28Si toroidal states","alpha clustering","Gaussian Mixture Model","7-alpha breakup","excitation energy spectra","heavy-ion reactions","machine learning in nuclear physics"],"falsifier":"Take the full statistics of the second experiment but replace the measured excitation energies with samples from a smooth phase-space or mixed-event distribution that matches the same total yield and errors, then run the same Gaussian Mixture Model pipeline on eight time-ordered partitions. If stable Gaussian components appear near 114 and 138 MeV in this background-only control, the claimed underlying structure is an artifact of the method rather than a resonance.","tokens_in":10394,"feed_emoji":"⚛️","tokens_out":10394,"duration_ms":85714,"temperature":0.7,"pith_summary":"Two published experiments on the peripheral 28Si + 12C reaction at 35 MeV/A disagreed about whether the 7-alpha breakup spectrum of 28Si* contains resonant peaks: one reported candidate toroidal resonances near 114, 126 and 138 MeV, while the other saw only statistical fluctuations over a smooth background. This paper argues that the disagreement is an artifact of how the background was subtracted. It develops a Gaussian Mixture Model analysis that decomposes the same spectra into Gaussian components without assuming a background shape, and applies it to both experimental datasets, to rotating-silicon model spectra, and to new collision calculations. The extracted components cluster near the toroidal-state predictions, and the centroids stay stable when the second experiment's events are split into eight time-ordered subsets. If the analysis is correct, it resolves the dispute in favor of real underlying structure close to toroidal excitations.","feed_headline":"AI fit finds the missing 28Si toroidal peaks","feed_subtitle":"A Gaussian mixture analysis of 7-alpha spectra finds peaks near 114 and 138 MeV in both disputed datasets.","key_machinery":"The central object is the Gaussian Mixture Model used as an unsupervised analytical tool: a histogram of counts or cross section versus excitation energy is normalized to a probability distribution, sampled with $10^6$ synthetic points drawn from normal distributions centered on each bin, and decomposed by expectation-maximization into a sum of Gaussians. The initial means are set by the paper's Eq. (2), $\\mu_{I_G}^{(0)} = (x_{\\max}-x_{\\min})\\,I_G(I_G+1)/(N_G(N_G-1)) + x_{\\min}$, and the number of Gaussians is chosen by requiring reduced chi-square $\\chi^2_\\nu$ close to 1. The same machinery applied to time-ordered partitions of the second experiment's data serves as the statistical-significance test. The HαC and cHαC models provide the theoretical spectra: a molecular-dynamics approach with $\\alpha$ particles as semiclassical degrees of freedom, in which 7-$\\alpha$ events are selected by two filters and excitation energies are computed from kinetic, Coulomb, and Q-value terms via Eq. (5).","core_discovery":"The central claim is that all examined datasets — the two experiments and the new cHαC collision model — contain excitation-energy structure compatible with toroidal states of 28Si* around 114 MeV and 138 MeV. The authors do not claim to reproduce every predicted resonance: the 126 MeV peak, for instance, is not recovered as an independent Gaussian but falls inside a wider component. The supporting demonstration is statistical: dividing the 186,097 unbinned events of the second experiment into eight time-ordered partitions and re-running the decomposition yields centroid shifts of only about 2.6 percent, which the paper presents as evidence that the extracted peaks are not random fluctuations. The paper further shows that a 9th-order polynomial background subtraction creates spurious maxima and can erase real peaks, and that the same GMM procedure recovers seven of ten known Gaussian components in the rotating-silicon model spectra where the polynomial-root method recovers only four of twelve.","pith_inferences":["The authors do not run the GMM on a purely smooth, resonance-free spectrum; that control would show how often the method finds stable Gaussians by construction, and it is the most direct test of the claim.","Because the rotating-silicon training data are themselves a known sum of Gaussians, recovering seven of ten components is a check on the fitting machinery rather than an external validation of physical content; the time-slice stability on real data carries more weight.","A natural extension is to apply the same pipeline to simulated background-only spectra with the same statistics as the experiments to map the false-positive rate as a function of counts per bin.","If the claim is right, higher-statistics future experiments should resolve the 126 MeV peak as an independent Gaussian rather than a shoulder inside a wider component, which is a concrete, testable prediction."],"forward_implications":["If the central claim is right, the null result of the second experiment reflects the polynomial background-subtraction procedure rather than the absence of resonances; the six Gaussian components found in its data align with the 114 and 138 MeV toroidal predictions.","The new cHαC collision calculations converge to the experimental data at high excitation energy, implying strong alpha clustering in the most energetic fragments and giving an upper-limit cross section because non-alpha-conjugate channels are absent from the model.","The about 2.6 percent centroid variation across time-ordered subsets offers a practical protocol for using GMM stability as a statistical-significance test for resonance claims in breakup spectra.","The widths and lifetimes of the extracted peaks being comparable to giant E0/E1/E2 resonances of 28Si places the claimed structure in a collective-excitation regime rather than a narrow-resonance regime."],"supporting_citations":[{"why":"Supplies the first experimental dataset and the reported resonance energies at 114, 126 and 138 MeV that the analysis aims to reproduce.","marker":"[15]"},{"why":"Supplies the second experimental dataset, the null-result analysis being challenged, and the 186,097 unbinned time-ordered events used for the stability test.","marker":"[17]"},{"why":"Supplies the rotating-silicon HαC spectra with known Gaussian components used to train the method, and the HαC model underlying the new collision calculations.","marker":"[21]"},{"why":"Provides the theoretical 28Si toroidal-state energies and angular momenta used as the comparison frame for extracted centroids.","marker":"[16]"},{"why":"Predicts toroidal-state energies in light nuclei, including the 143.18 MeV value used for comparison.","marker":"[14]"},{"why":"Provides the giant E0/E1/E2 resonance widths of 28Si used to argue that the extracted peaks have collective character.","marker":"[36]"},{"why":"Describes the Gaussian Mixture Model and expectation-maximization algorithm on which the analysis is based.","marker":"[30]"},{"why":"Provides the reduced chi-square criterion used to choose the number of Gaussian components.","marker":"[33]"}],"fun_headline_variants":["AI finds toroidal peaks in silicon-28 breakup","GMM spots 28Si toroid states in disputed data","Machine learning confirms exotic 28Si toroids","AI analysis of 7-alpha decay reveals toroidal resonances","Toroidal 28Si peaks emerge from AI mixture model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that decomposing a histogram into a few Gaussians with reduced chi-square close to one identifies physically meaningful resonances, rather than merely being a flexible way to approximate any smooth distribution.","fun_headline_variants_meta":{"raw":{"variants":["AI finds toroidal peaks in silicon-28 breakup","GMM spots 28Si toroid states in disputed data","Machine learning confirms exotic 28Si toroids","AI analysis of 7-alpha decay reveals toroidal resonances","Toroidal 28Si peaks emerge from AI mixture model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000552,"raw_usage":{"total_tokens":2594,"prompt_tokens":869,"completion_tokens":1725,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":1645}},"tokens_in":485,"tokens_out":1725,"duration_ms":12287,"temperature":1.0,"reasoning_tokens":1645,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:38:30.077531+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the full statistics of the second experiment but replace the measured excitation energies with samples from a smooth phase-space or mixed-event distribution that matches the same total yield and errors, then run the same Gaussian Mixture Model pipeline on eight time-ordered partitions. If stable Gaussian components appear near 114 and 138 MeV in this background-only control, the claimed underlying structure is an artifact of the method rather than a resonance.","supporting_citations":[{"cited_title":"Our observations of similarity between the results for all the available datasets, are also shown in Fig","cited_arxiv_id":null,"evidence_quote":"Supplies the first experimental dataset and the reported resonance energies at 114, 126 and 138 MeV that the analysis aims to reproduce."},{"cited_title":"Hannaman, B","cited_arxiv_id":null,"evidence_quote":"Supplies the second experimental dataset, the null-result analysis being challenged, and the 186,097 unbinned time-ordered events used for the stability test."},{"cited_title":"The Cao experimen- tal data obtained by Ref","cited_arxiv_id":null,"evidence_quote":"Supplies the rotating-silicon HαC spectra with known Gaussian components used to train the method, and the HαC model underlying the new collision calculations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the theoretical 28Si toroidal-state energies and angular momenta used as the comparison frame for extracted centroids."},{"cited_title":"Staszczak and C","cited_arxiv_id":null,"evidence_quote":"Predicts toroidal-state energies in light nuclei, including the 143.18 MeV value used for comparison."},{"cited_title":"Jothilakshmi and V","cited_arxiv_id":null,"evidence_quote":"Provides the reduced chi-square criterion used to choose the number of Gaussian components."}],"review_version":1}