{"id":"601f61e4-ec80-40e2-bd58-592b15a967b3","arxiv_id":"2411.14902","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A machine-learning fit to lattice chromo field data yields a compact two-variable expression for E(d, xt) and reproduces flux tube string tension and width over existing separations.","lead":"Researchers trained two neural network types, MLP and KAN, on lattice QCD data for the electric field between a static quark and antiquark. Both reproduce the field shape, and the KAN fit yields a compact formula that depends on quark separation and transverse distance.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The analytic KAN expression Eq. (4) does not decay as x_t increases, so the string-tension and width integrals in Eqs. (1) and (5) are undefined over the transverse plane without an ad hoc cutoff; the paper never specifies such a cutoff, making the central quantitative comparison cutoff-dependent.","rationale":"The reader's verdict is CONDITIONAL, and I agree that the paper should not be accepted as is. My core concern is more specific and, in my view, more load-bearing than the ones listed in the reader's weakest_assumption: even if the lattice data were perfectly precise and the longitudinal-coordinate assumption were exact, Eq. (4) still cannot be used to compute string tension and width from Eqs. (1) and (5) without an explicit transverse cutoff, because the expression does not vanish as x_t grows. The paper presents the KAN-derived formula as a compact analytic surrogate and uses it for the quantitative comparison in Fig. (6), yet it never states the integration domain or shows that the result is insensitive to the outer radius. The internal warning in Sec. III A that KAN cannot predict unknown regions makes this omission particularly visible. The concern is concrete and checkable: one can numerically integrate Eq. (4) to increasing radii and see whether the derived observables stabilize; if they do not, the comparison in Fig. (6) is not a rigorous lattice comparison. The paper's MLP fitting exercise and the basic feasibility of using KAN for symbolic regression on lattice-like data are credible and are not the subject of my objection. The issue is that the central analytical expression, as written, is not a valid global field profile, so the derived physical quantities inherit a hidden cutoff dependence. This is an addressable flaw: adding a decaying envelope, restricting the use of Eq. (4) to the fitted x_t interval with an explicit cutoff, or constraining the symbolic regression to fields that vanish at infinity would resolve it. For that reason I recommend maintaining the CONDITIONAL verdict rather than rejecting the work outright, but the condition must include demonstrating that the string tension and width calculations are cutoff-independent or explicitly specifying and justifying the integration limit.","tokens_in":11361,"tokens_out":12536,"duration_ms":133517,"concrete_test":"Evaluate Eq. (4) on a fine grid and compute I(R) = (1/2) integral_{|x_t| < R} d^2x_t E^2 and W(R) = integral_{|x_t| < R} d^2x_t x_t^2 E for d = 0.37 fm and d = 1.19 fm with R = 1, 2, 5, 10 fm. If sqrt(sigma_T) or the width changes by more than a few percent as R grows, the quoted string tension and width are cutoff artifacts. Also evaluate Eq. (4) at the x_t values from Ref. [13] for the five smallest d; any predicted E < 0 where the lattice values are non-negative directly quantifies how closely the expression mirrors the data and exposes the need for a decaying or positivity-constrained parameterization.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central result, Eq. (4), is claimed to closely mirror lattice chromo field data and is used to compute string tension and width from Eqs. (1) and (5). But Eq. (4) is not square-integrable: as x_t grows, the factor exp[-3.5478 sin(1.68 x_t + 4.791)] oscillates with bounded amplitude rather than decaying, so E(d, x_t) does not approach zero at large transverse separation. Consequently, the integral in Eq. (1) for the string tension and the width integral in Eq. (5) do not converge over the full transverse plane. The paper does not state any integration cutoff, and no justification is given for truncating the domain. The problem is visible even inside the plotted range: for d = 0.37 fm and x_t = 0.85 fm, Eq. (4) gives E approx -0.010 GeV^2, a non-physical negative field where the lattice data are consistent with a non-negative tail. This matters because the KAN string-tension and width curves in Fig. (6) are the quantitative payoff of the paper. The paper itself warns in Sec. III A that KAN cannot reliably predict outside the training interval, yet Eq. (4) is used as if it were a global analytic surrogate. The 2D KAN training loss only reaches about 10^-2, much worse than the MLP loss, so the surrogate is not high-fidelity where it is used. The comparison in Fig. (6) therefore depends on an unspecified, physically unmotivated cutoff, and the claimed agreement with lattice results is not a well-defined statement as presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper trains MLP and KAN networks on lattice QCD chromoelectric field profiles for quark-antiquark separations d ≈ 0.37–1.19 fm, using data from Ref. [13]. It reports MLP training and test losses near 1e-5, obtains a compact bivariate KAN symbolic expression for E(d, x_t) in Eq. (4), and uses Eqs. (1) and (5) to compute flux-tube string tension and width, comparing them with lattice results in Fig. 6. The central claim is that the KAN expression closely mirrors the lattice field distributions and provides a continuous analytic surrogate over the studied separations, demonstrating the usefulness of machine learning for QCD flux-tube phenomenology.","tokens_in":11748,"tokens_out":9923,"duration_ms":111130,"significance":"If the KAN formula in Eq. (4) were reliable, it would be a genuinely useful compact interpolant for the flux-tube profile as a function of both transverse coordinate and quark separation, going beyond the per-separation Clem fits of Eq. (2). The paper also provides a concrete MLP-versus-KAN comparison on a small physics dataset, and the low MLP loss with a held-out test set is a useful sanity check that the lattice data can be interpolated by a neural network. However, the quantitative payoff of the paper—the KAN string-tension and width curves in Fig. 6—is not established by the analysis as presented, because Eq. (4) is not normalizable, the comparison is in-sample, and no uncertainties are propagated. These issues must be addressed before the central quantitative claims can be accepted.","major_comments":[{"comment":"The KAN expression in Eq. (4) does not approach zero as x_t increases: the factor exp[-3.5478 sin(1.68 x_t + 4.791)] is bounded and oscillatory, so the integrals defining sigma_T and sqrt(w^2) do not converge over the full transverse plane. Since the manuscript neither specifies an integration cutoff nor justifies truncating the domain, every KAN-based point in Fig. 6 depends on an unspecified regularization, and the claimed agreement with the lattice results is not a well-defined quantitative statement. The problem is visible even inside the training range: for d = 0.37 fm and x_t = 0.85 fm, Eq. (4) gives E approximately -0.010 GeV^2, where the lattice profiles are non-negative, so the surrogate is not locally faithful either. This is aggravated by the authors' own warning in Sec. III A that KAN cannot accurately predict the unknown region.","section":"Section III B, Eq. (4), Eqs. (1) and (5)"},{"comment":"The comparison in Fig. 6 is a self-consistency check rather than an independent validation, because Eq. (4) is fitted to the same lattice data from Ref. [13] that are then used as the 'Data' curves in that figure. The 20% held-out test set described in Sec. II B is selected from the same ten separations, so it cannot validate extrapolation or even interpolation in d. The predictive claim made in the abstract and in Sec. IV would require a separation-based holdout, such as training on nine values of d and testing on the tenth, or comparison with an independent lattice calculation; as it stands, the agreement only shows that the fitted formula reproduces its own training set.","section":"Section III B and Fig. 6"},{"comment":"The bivariate KAN training loss converges only to about 1e-2, three orders of magnitude worse than the MLP training and test losses of about 1e-5. Because Eq. (4) is the symbolic output of that same KAN training and the KAN curves in Fig. 6 are computed from Eq. (4), the KAN-based string tension and width carry an unknown but potentially large error. This is compounded by the explicit decision in Sec. III A not to consider uncertainties associated with the lattice simulation results or the parameterization; no error bars are propagated to Fig. 6, so the significance of the deviations between MLP, KAN, and lattice data cannot be assessed.","section":"Section III B and Fig. 5"},{"comment":"The network architectures are described inconsistently, which prevents reproduction of the central numerical results. Figure 1 and Sec. II B specify an MLP with two hidden layers of 128 ReLU neurons; Sec. III A states that the MLP input is a single neuron for the univariate feasibility test; and Sec. III B introduces a '(2,6,1)' configuration without clarifying whether it replaces the earlier 128-neuron description. For KAN, Sec. II B gives a (1,3,1,1) structure while Sec. III A gives (1,3,3,1). Because the MLP and KAN results in Fig. 6 depend on these architectural choices, the paper should provide a consistent architecture and hyperparameter table, or release the code, before the numerical results can be considered reproducible.","section":"Sections II B, III A, and III B"}],"minor_comments":[{"comment":"The y-axis label 'sqrt(w^2) [fm^2]' is dimensionally inconsistent: Eq. (5) defines sqrt(w^2) as a length, so the unit should be fm, or the label should be w^2.","section":"Fig. 6, second panel"},{"comment":"The sentence 'where the signal-to-noise ratio becomes excessively high' should presumably read 'low' or 'unfavorable', since lattice signals degrade at large separations.","section":"Sec. I"},{"comment":"The phrase 'with the calculation method outlined in the square root of Eq. (1)' is unclear; if Ref. [13] reports sqrt(sigma_T), the paper should explain how that quantity was obtained and how it is compared with the integral in Eq. (1).","section":"Sec. III B"},{"comment":"The output notation in Sec. II B says the output layer 'represents E(x_t)', although the bivariate model outputs E(d, x_t); please update the notation for consistency. In addition, Eq. (3) shares the non-decaying oscillatory behavior of Eq. (4); if KAN expressions are to be used as physical parameterizations in future work, a decaying or explicitly cutoff-regulated form should be imposed.","section":"Sec. II B and Eq. (3)"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The main new thing here is a two-variable symbolic fit for the chromo field, Eq. (4), which comes from applying KAN to Baker et al.'s lattice data. The paper is honest about what it does and doesn't do, but the central fit has a mathematical problem that undermines the payoff.\n\nWhat's new and good: this is the first KAN application to flux tube profiles as far as the citations go. The MLP fit to the 2D data looks solid (test loss about 1e-5), and the MLP-vs-KAN comparison is a legitimate use case. The paper is candid that KAN training only reaches about 1e-2 loss and is hyperparameter-sensitive, and it states up front that lattice uncertainties are not propagated. That honesty is worth something.\n\nNow the soft spots. The stress-test concern is real and load-bearing. Eq. (4) does not decay in x_t; the sine in the exponential just oscillates, so E(d,x_t) keeps oscillating with bounded amplitude as x_t grows. Both integrals in Eqs. (1) and (5) therefore diverge over the full transverse plane, and the paper never states an integration cutoff. The comparison in Fig. 6 is not a well-defined quantitative statement as written. This isn't a nitpick: the string tension and width are the paper's main quantitative results. The authors might reasonably integrate only over the lattice data range, but they need to say so and justify it. There is also the circularity issue the reader flagged: Eq. (4) is a fit to the same lattice data used for comparison, so Fig. 6 is a self-consistency check, not an independent prediction. The paper frames it as \"closely mirrors lattice results,\" which is fair if read as interpolation, but the conclusion overclaims when it suggests the approach can \"predict\" profiles for unknown d. The MLP architecture is also described inconsistently (128-neuron layers in Sec. II, then a (2,6,1) configuration in Sec. III B). Minor, but sloppy. No code or data are provided, which matters for reproducibility.\n\nWho this is for: researchers who want a quick look at whether KAN can produce compact fits to lattice flux tube data. As a tool-level contribution it is fine; as a physics result the current form doesn't hold up without a cutoff or a decaying functional form.\n\nRecommendation: send it to peer review with a clear request for major revision. The flaws are addressable: impose a physically motivated cutoff, switch to a parameterization that decays at large x_t, or restrict the string-tension and width claims to the data range. With that fixed, the paper would be a modest but useful contribution. Without it, the central comparison is meaningless.","headline":"Useful ML application to flux tube data, but the main KAN fit doesn't decay at large transverse distance, so the string tension and width integrals are undefined without a cutoff the paper never states.","tokens_in":12255,"tokens_out":4042,"would_cite":false,"duration_ms":34887,"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":"The paper claims that a Kolmogorov–Arnold network trained on lattice QCD data produces a compact two-variable formula for the quark–antiquark chromoelectric field that reproduces the flux-tube string tension and width.","keywords":["chromoelectric flux tube","Kolmogorov-Arnold network","multilayer perceptron","lattice QCD","string tension","quark confinement","QCD flux tube"],"falsifier":"A high-statistics lattice QCD calculation of the chromoelectric field at a separation not in the training set, such as $d = 0.8$ fm, with its transverse profile compared against Eq. (4), would falsify the claim if the disagreement exceeds the simulation's quoted uncertainties.","tokens_in":11171,"feed_emoji":"⚛️","tokens_out":9692,"duration_ms":82703,"temperature":0.7,"pith_summary":"The paper asks whether modern machine-learning architectures can turn sparse lattice QCD measurements of the color field between a static quark and antiquark into a compact, continuous description of the field. It trains two networks, a multilayer perceptron and a Kolmogorov–Arnold network, on the chromoelectric flux-tube profiles from ten quark separations between about 0.37 and 1.19 fm. The central claim is that the KAN delivers an explicit two-variable formula, $E(d,x_t)$, that closely tracks the lattice data and, when integrated through the standard flux-tube relations, reproduces the string tension and the root-mean-square flux-tube width. If this holds, it gives a parameter-free surrogate that can interpolate between simulated separations and offers a new way to connect machine learning with non-perturbative QCD phenomenology.","feed_headline":"One formula maps the quark-antiquark field","feed_subtitle":"Machine learning on lattice data yields the chromo field and string tension in one analytic expression.","key_machinery":"The central object is the Kolmogorov–Arnold network (KAN), a neural architecture in which learnable spline activation functions on the edges embody the Kolmogorov–Arnold superposition theorem, so a multivariate function is represented as a finite sum of univariate functions. The paper's KAN configuration (2,6,1) for the two-variable problem is pruned and refined until the surviving single path can be read out as the symbolic formula in Eq. (4). An MLP with two hidden layers of 128 ReLU neurons serves as the benchmark, and the Clem parameterization, $E(x_t) = \\frac{\\phi}{2\\pi}\\frac{\\mu^2}{\\alpha}\\frac{K_0[(\\mu^2 x_t^2 + \\alpha^2)^{1/2}]}{K_1[\\alpha]}$, is the conventional fitting form used for comparison.","core_discovery":"The paper's central discovery is a two-dimensional analytic expression for the chromoelectric field between a static quark–antiquark pair, obtained by training a Kolmogorov–Arnold network on the non-perturbative lattice data of Ref. [13]. Written as $E(d,x_t) = 0.0423\\,d - 0.0388 + 0.0103\\,\\exp[-3.5478\\,\\sin(1.68\\,x_t + 4.791)]$, it describes the transverse profile of the flux tube at ten separations and, according to the authors, closely mirrors lattice results. The expression is fed into $\\sigma_T \\simeq \\frac{1}{2}\\int d^2x_t\\, E^2(x_t)$ and the flux-tube width integral to compute the string tension and width, which the authors compare with lattice values; they report that the MLP surrogate remains more accurate for the string tension while the KAN supplies the interpretable symbolic form. The paper presents this as a proof of principle that a neural-network inverse-problem approach can yield continuous descriptions where traditional parameterizations require a separate fit at each separation.","pith_inferences":["The structure of Eq. (4) implies the transverse shape of the flux tube changes only mildly with separation, since $d$ enters linearly and $x_t$ enters only through the oscillatory exponential; this shape-similarity is a testable prediction for new lattice data.","Because the paper does not propagate lattice uncertainties into the KAN fit, the resulting formula has no error bars; a careful error analysis would determine whether the apparent agreement with the string tension is statistically meaningful.","A natural next step is to include the longitudinal coordinate or temperature as additional inputs, which would let the fit test the paper's working assumption of longitudinal independence instead of assuming it.","If Eq. (4) holds at intermediate separations not simulated, it could serve as an interpolation tool bridging the gap between small and large $d$ where lattice signal-to-noise degrades."],"forward_implications":["If Eq. (4) is correct, the chromo field can be interpolated continuously across quark separations from about 0.37 to 1.19 fm without refitting at each d.","The learned formula, used in Eqs. (1) and (5), yields string tension and flux-tube width that can be checked against lattice results, providing a fast forward model for these observables.","The KAN's symbolic output turns a black-box fit into an interpretable expression, making it straightforward to insert into analytic models of confinement and string breaking.","The same training approach can be applied to other lattice QCD observables, such as the chromomagnetic field or full-QCD data, as more simulations become available."],"supporting_citations":[{"why":"Supplies the ten non-perturbative chromo-field profiles at separations 0.37–1.19 fm that are the training data for both networks and the comparison baseline for string tension and width.","marker":"[13]"},{"why":"Introduces the Kolmogorov–Arnold network architecture with spline activation functions, which is the method used to produce the symbolic expression for the chromo field.","marker":"[53]"},{"why":"Provides the Clem vortex parameterization used as the conventional comparison fit in the feasibility test of the machine-learning models.","marker":"[20]"},{"why":"Supports the relation between string tension and the squared chromoelectric field integral used to compute σ_T from the learned field.","marker":"[21]"},{"why":"Establishes the dual-superconductor picture and the observation that longitudinal variations of the chromo field are negligible, supporting the paper's assumption of longitudinal independence.","marker":"[12]"}],"fun_headline_variants":["ML yields one formula for quark-antiquark field","Neural network reveals analytic QCD flux tube","One equation captures quark pair's chromo field","AI finds string tension from a single field map"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The fitted formula inherits any errors in the lattice data used for training and assumes the chromo field does not vary along the longitudinal coordinate, so if either fails, the derived string tension and width are unreliable.","fun_headline_variants_meta":{"raw":{"variants":["ML yields one formula for quark-antiquark field","Neural network reveals analytic QCD flux tube","One equation captures quark pair's chromo field","AI finds string tension from a single field map"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000227,"raw_usage":{"total_tokens":1454,"prompt_tokens":911,"completion_tokens":543,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":495}},"tokens_in":527,"tokens_out":543,"duration_ms":6058,"temperature":1.0,"reasoning_tokens":495,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:44:26.413986+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A high-statistics lattice QCD calculation of the chromoelectric field at a separation not in the training set, such as $d = 0.8$ fm, with its transverse profile compared against Eq. (4), would falsify the claim if the disagreement exceeds the simulation's quoted uncertainties.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Clem vortex parameterization used as the conventional comparison fit in the feasibility test of the machine-learning models."},{"cited_title":"Partial restoration of chiral symmetry in a confining string","cited_arxiv_id":"1404.7746","evidence_quote":"Supports the relation between string tension and the squared chromoelectric field integral used to compute σ_T from the learned field."}],"review_version":1}