{"id":"afe53c19-54d6-46cb-89f9-cea1b5eccf31","arxiv_id":"2506.04628","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A fully-connected neural network classifies which of five partial waves contains a resonance from K-matrix-generated differential cross sections, reaching about 69% accuracy on synthetic data.","lead":"This paper trains a small neural network to identify which angular momentum wave, from S through G, contains a resonance in synthetic scattering data. It reports about 69% accuracy, showing that machine learning can pick out resonance-bearing partial waves, though real-data validation is still missing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported 69% accuracy may ride on a scale confound: K-matrix parameters are sampled per partial wave, so overall |dσ/dΩ| magnitude can correlate with the label; no normalization ablation is reported.","rationale":"The paper is an honest proof-of-concept with a modest 69% accuracy claim, and the reader's conditional verdict is reasonable. My concern is more specific and more directly tied to the central claim than the transfer-to-real-data issue: even within the synthetic setup, the network may not be learning what the paper says it learns. In §3 the parameters are sampled per partial wave from 'literature-based ranges', which can create a class-dependent normalization. Since the FCNN receives raw dσ/dΩ values and no energy, the simplest decision rule is the overall magnitude. The absence of any ablation or baseline means the 69% figure is uninterpretable as evidence of resonance-feature extraction. The proposed normalization test settles this. If accuracy holds after normalization, the concern is resolved and the paper's claim is supported for the synthetic in-distribution case; if not, the central claim needs to be reframed. Either way, the reader's CONDITIONAL verdict should stand, with the added condition of the ablation. I disagree with the reader's choice of weakest assumption because the in-distribution confound is more load-bearing than external transfer for a proof-of-concept.","tokens_in":3656,"tokens_out":10309,"duration_ms":129332,"concrete_test":"Retrain the exact FCNN on a normalized version of the dataset where each 15-point curve is divided by its empirical integral (or maximum |dσ/dΩ|) before training, and report validation accuracy; if accuracy drops from ~69% toward the 20% chance level, the original classifier relied on the overall cross-section scale rather than the resonance's angular signature. Additionally, run a logistic-regression baseline on the per-sample mean and standard deviation of dσ/dΩ; if it approaches 69%, the FCNN's 'feature extraction' claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (abstract, §4) is that the FCNN extracts resonance-location information from K-matrix differential cross sections. But the generative protocol in §3 activates the resonant term in a single partial wave while sampling 'physical parameters m_α, γ_αi, Γ0_α, a_ij and b_ij' within per-partial-wave literature ranges for each T_l. This makes the overall scale of dσ/dΩ (Eq. 4) a potential class-dependent confound: if, e.g., l=4 resonances have larger couplings or different non-resonant a_ij,b_ij ranges than l=0, the network can classify by mean/peak magnitude without using the angular shape that carries the resonance information. The input to the FCNN is the raw 30-dimensional vector of 15 cosθ values and 15 dσ/dΩ values; no normalization is mentioned, and no baseline (e.g., classifying from the sample mean alone) is reported. The confusion-matrix analysis in Fig. 2c is consistent with scale-driven decisions: class 4 (largest l) is easiest, while classes with overlapping magnitudes are confused. Without an ablation that removes the overall normalization, the 69% accuracy does not establish that the network is extracting the resonant partial wave; it may only be exploiting a sampling artifact of the generative model.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper describes a machine-learning classification study for partial wave analysis. Synthetic differential cross sections are generated from a K-matrix amplitude in which exactly one partial wave (l=0..4) contains a resonant term and the remaining partial waves contribute non-resonant terms, with parameter ranges inspired by pion-nucleon scattering. A fully-connected neural network with 30 inputs (15 cosθ values and 15 dσ/dΩ values) is trained to classify which partial wave contains the resonance. The authors report about 69% validation accuracy and a confusion matrix, and they conclude that the network can distinguish the resonant partial wave from purely non-resonant partial waves.","tokens_in":3984,"tokens_out":5220,"duration_ms":52418,"significance":"The paper is a proof-of-concept with a clear application: using a fully-connected neural network to identify which partial wave carries a resonance from K-matrix-parametrized differential cross sections. Its strengths are the physically motivated K-matrix parametrization with unitarity and Blatt-Weisskopf barrier factors, the explicit generation of synthetic datasets, and the transparent reporting of a confusion matrix. If the reported performance survives normalization ablations and an independent test-set evaluation, it would support the viability of ML-assisted partial wave analysis for broad or overlapping structures. As it stands, the central claim is plausible but not yet sufficiently established, because a scale artifact has not been ruled out and no baselines are provided.","major_comments":[{"comment":"The FCNN input is the raw 30-dimensional vector of 15 cosθ values and 15 dσ/dΩ values with no normalization described, while Section 3 says that the parameters m_α, γ_αi, Γ0_α, a_ij, b_ij are sampled within πN literature-based ranges for each T_l(m^2). If the ranges differ by partial wave, the overall magnitude of dσ/dΩ can be class-discriminative on its own, and the confusion-matrix pattern (class 4 easiest, class 0 frequently misclassified as class 4) is consistent with scale-driven decisions. To support the abstract's claim that the network extracts resonance-location information from the angular dependence, the authors should add an ablation that normalizes each sample (e.g., dividing by its angle-averaged value or maximum) and a baseline trained only on the mean or total cross section. Without such controls, the 69% accuracy does not establish that the network is using the angular shape of the differential cross section.","section":"Section 3 and Section 4, Eq. (4), Fig. 2c"},{"comment":"The paper reports training and validation accuracy/loss curves and a confusion matrix, but the number of samples per class, the train/validation split, and whether Fig. 2c is computed on the validation set are not stated. No independent held-out test set and no repeated training runs with different random seeds are reported, so the 'around 69%' figure has no uncertainty estimate. The authors should evaluate on a held-out test set and report the mean plus or minus one standard deviation over several initializations.","section":"Section 4, Fig. 2"},{"comment":"The generative protocol is under-specified: the number of samples per label, the exact probability distributions for m_α, γ_αi, Γ0_α, a_ij, b_ij, the energy sampling step, and the choice of the 15 cosθ values are not given. The phrase 'systematically varied within πN scattering experiment- or literature-based ranges for each T-matrix partial wave amplitude' is ambiguous: if the parameter ranges are l-dependent, the scale confound in my first comment is direct; if they are common, the authors should state that explicitly. This detail is essential for judging representativeness and for reproducibility.","section":"Section 3, dataset generation"}],"minor_comments":[{"comment":"The sentence beginning 'The term 'class' refers to either the true label or the predicted label, with 'label' to denote...' is confusing; 'label' should simply be defined as the index l of the resonant partial wave.","section":"Section 4, Fig. 2"},{"comment":"Please verify the phase-space factor convention in the denominator (√ρ_iρ_j versus ρ_iρ_j) against reference [2].","section":"Section 2, Eq. (1)"},{"comment":"'Smoothening' should be 'smoothing', and 'the model is learning reasonably' would read better as 'the model is learning reasonably well'.","section":"Throughout"},{"comment":"The phrase 'the first three labels' is vague; specify that (a), (b), and (c) correspond to resonant partial waves l=0, 1, and 2, respectively.","section":"Fig. 1 caption"}],"recommendation":"major_revision","confidential_remarks":"This is a short proceedings-style paper. The technical concern is not that supervised learning on synthetic K-matrix data is circular, but that the reported accuracy may be inflated by an uncontrolled scale variable. The paper should be evaluated after the authors provide the normalization ablation, a baseline comparison, and a proper held-out test-set evaluation. I see no issue with scope or novelty for a hadron-spectroscopy conference."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: it's a small, honest proof-of-concept that a plain fully-connected net can pick out which of five partial waves carries a resonance from synthetic K-matrix differential cross sections. The components are standard—K-matrix generation plus an FCNN—and the result is a new empirical demonstration at 69% accuracy. What the paper does well: the generative model is physically motivated (unitary K-matrix, Blatt-Weisskopf factors, piN-inspired parameter ranges), the setup is clearly described, and the authors don't oversell generalization. They explicitly note the small loss gap as a possible generalization issue.\n\nThe soft spot that matters is a likely scale confound. The K-matrix parameters are sampled separately per partial wave, so the overall magnitude of dσ/dΩ can correlate with the label. The network gets the raw dσ/dΩ values as input, with no normalization. The paper reports no ablation that removes the overall scale (e.g., normalizing each curve to unit integral or classifying from the mean alone). The confusion matrix pattern—class 4 easiest by a wide margin, class 0 leaking into class 4—is consistent with a magnitude-driven decision. So the headline claim that the network 'distinguishes' the resonant partial wave is only weakly supported as shape-based feature extraction. This is not a manufactured flaw; it falls directly out of the text in Section 3.\n\nOther gaps are in proportion for a proceedings: no held-out test set with error bars, no baseline classifier, no code/data, and only three references (no prior ML-PWA work to calibrate novelty). But these are second-order compared to the normalization issue. The reader's CONDITIONAL verdict is fair; I'd go a bit stronger—the 69% should be treated as an upper bound until the ablation is done.\n\nWho is this for: people working in ML-assisted hadron spectroscopy or PWA who want a baseline demonstration. It's not a conceptual advance, but it's a reasonable starting point. I'd send it to peer review, but with the clear expectation that the authors add a normalization ablation and a baseline. If they do, the claim becomes defensible. Reading group? Maybe, mostly as a case study in how synthetic-data confounds can sneak into ML physics.","headline":"Modest proof-of-concept with a real scale confound: the 69% classification accuracy likely rides on per-partial-wave magnitude differences, not shape.","tokens_in":4450,"tokens_out":3117,"would_cite":false,"duration_ms":37618,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A fully connected neural network trained on K-matrix differential cross sections can label the resonant partial wave at 69% accuracy in a synthetic pion-nucleon-like scattering model.","keywords":["partial wave analysis","K-matrix","deep neural network","fully-connected network","resonance classification","differential cross section","pion-nucleon scattering","hadron spectroscopy"],"falsifier":"Generate a validation set with two resonant partial waves active simultaneously using the same K-matrix parametrization and parameter ranges, and require the network to identify all resonant waves; if accuracy falls to near chance, the reported 69% is an artifact of the one-resonance-per-sample generative model.","tokens_in":3462,"feed_emoji":"🧠","tokens_out":6207,"duration_ms":59478,"temperature":0.7,"pith_summary":"Partial wave analysis is the standard tool for deciding whether a bump in an invariant-mass spectrum is a resonance, but the task is hard when peaks are broad, overlapping, or sit on non-resonant backgrounds. This paper proposes a hybrid step: generate synthetic differential cross sections from a K-matrix parametrization with pion-nucleon-inspired parameter ranges, and train a fully connected neural network to label which partial wave contains the resonance. The model reaches about 69% validation accuracy, with the clearest separation for the highest tested partial wave and predictable confusion between classes whose angular shapes overlap. The significance, if the result holds, is a proof of concept that a neural network can extract resonance-location information directly from the shape of a differential cross section, potentially assisting conventional fits.","feed_headline":"Neural net spots resonant partial waves at 69 percent accuracy","feed_subtitle":"In synthetic K-matrix cross sections, a fully connected network picks the partial wave carrying the resonance","key_machinery":"The load-bearing object is the K-matrix parametrization of the scattering amplitude, $K_{ij}=\\sum_\\alpha g_{\\alpha i}(m)g_{\\alpha j}(m)/[(m_\\alpha^2-m^2)\\sqrt{\\rho_i\\rho_j}]+a_{ij}+b_{ij}m^2$, with residue functions built from Blatt-Weisskopf centrifugal barrier factors. That K-matrix is converted to the transition amplitude by $T=(I-iK\\rho)^{-1}K$, expanded in Legendre polynomials $T(\\theta,m^2)=\\sum_l(2l+1)T_l(m^2)P_l(\\cos\\theta)$, and turned into the differential cross section $d\\sigma/d\\Omega = 4\\rho_f/(m^2\\rho_i)|T|^2$. The neural network then maps 15 $\\cos\\theta$ values and 15 $d\\sigma/d\\Omega$ values to one of five labels, identifying the partial wave whose resonant K-matrix term is active.","core_discovery":"The paper's central claim is that the angular shape of a K-matrix-parametrized differential cross section encodes enough information to identify which partial wave carries the resonance, even when every partial wave also has active non-resonant terms. In the synthetic dataset, exactly one of the five partial waves $l=0,\\ldots,4$ has its resonant K-matrix term turned on per sample, and the fully connected network classifies that wave at about 69% accuracy. The confusion matrix shows the network performs best on class 4 and most often confuses classes whose differential cross sections share edge and base features, which the authors interpret as expected from the overlapping plotted shapes. The claim is specifically about this one-resonance-per-sample synthetic problem, not yet about arbitrary experimental data.","pith_inferences":["My inference: the 69% figure is likely specific to the one-resonance-per-sample generative model; a natural test is to run the same architecture on data with two simultaneously active resonant partial waves and see whether classification degrades.","My inference: the strong class-4 performance and class-0-to-class-4 confusion suggest the network is keying on high-angular-momentum shape features; an ablation that removes the highest $l$ contributions would test this.","My inference: feeding the full energy-dependent differential cross section, rather than 15 sampled points per curve, could push accuracy beyond the reported 69%."],"forward_implications":["A trained classifier can flag which of the five partial waves most likely carries a resonance in a new K-matrix-generated differential cross section, shrinking the search space for conventional PWA fits.","The confusion pattern (classes 1 and 2 spilling into 3 and 4, and class 0 into class 4) identifies the overlapping shape features that limit the current architecture, giving a concrete target for better feature engineering.","Because training and validation accuracies stay close (68–70%), the model is not memorizing the training set; the gap signals a mild generalization limit rather than overfitting.","Extending the input to include energy distributions or total cross-section information, as the paper outlines as its next step, is a direct route to improving on the 69% baseline."],"supporting_citations":[{"why":"Supplies the standard motivation for partial wave analysis as the tool for interpreting resonance structures, framing the classification problem the DNN is meant to assist.","marker":"[1]"},{"why":"Provides the specific K-matrix parametrization form used in Eq. (1), the central object from which the synthetic differential cross sections are generated.","marker":"[2]"},{"why":"Supplies the Blatt-Weisskopf centrifugal barrier factors used in the residue function, which shape the energy dependence of the couplings in the generated data.","marker":"[3]"}],"fun_headline_variants":["Neural network singles out resonant partial wave at 69% accuracy","Deep learning finds which partial wave holds the resonance (69% accuracy)","Resonant partial wave identified by neural net in synthetic scattering data","FCNN distinguishes resonant partial wave with 69% accuracy in K-matrix data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the synthetic dataset, built with only one resonant partial wave per sample and parameter ranges borrowed from pion-nucleon scattering, captures the structure of real partial-wave-analysis classification problems.","fun_headline_variants_meta":{"raw":{"variants":["Neural network singles out resonant partial wave at 69% accuracy","Deep learning finds which partial wave holds the resonance (69% accuracy)","Resonant partial wave identified by neural net in synthetic scattering data","FCNN distinguishes resonant partial wave with 69% accuracy in K-matrix data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000663,"raw_usage":{"total_tokens":3070,"prompt_tokens":1030,"completion_tokens":2040,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":646,"completion_tokens_details":{"reasoning_tokens":1962}},"tokens_in":646,"tokens_out":2040,"duration_ms":15025,"temperature":1.0,"reasoning_tokens":1962,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:37:20.662869+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a validation set with two resonant partial waves active simultaneously using the same K-matrix parametrization and parameter ranges, and require the network to identify all resonant waves; if accuracy falls to near chance, the reported 69% is an artifact of the one-resonance-per-sample generative model.","supporting_citations":[{"cited_title":"Peters, https://doi.org/10.1142/S0217751X06034811 Int","cited_arxiv_id":null,"evidence_quote":"Supplies the standard motivation for partial wave analysis as the tool for interpreting resonance structures, framing the classification problem the DNN is meant to assist."},{"cited_title":"Chung et","cited_arxiv_id":null,"evidence_quote":"Provides the specific K-matrix parametrization form used in Eq. (1), the central object from which the synthetic differential cross sections are generated."}],"review_version":1}