{"id":"24286558-17ce-47bc-a861-5e0b36aa9fbd","arxiv_id":"2412.15474","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A KAN-based surrogate model is trained on simulated scattering data and then used to fit experimental SANS curves, extracting lamellar defect parameters and colloidal interaction parameters.","lead":"This paper trains a neural network called a KAN to act as a stand-in for scattering formulas, then uses it to fit experimental small-angle scattering data and extract structural parameters of soft materials. It demonstrates the approach on two systems, defective lamellar phases and charged colloids, but the fits show notable low-angle discrepancies and the framework is not as model-independent as claimed.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The KAN is a surrogate for the authors' own forward models; the claimed model-independence and direct structural extraction are not validated because the paper never tests parameter recovery on independently generated scattering data.","rationale":"The reader's weakest assumption (the three-parameter uniqueness of the lamellar descriptor) is part of a broader issue: the training data for both case studies come from the authors' prior forward models, so the KAN is a surrogate for those specific libraries rather than a model-independent inversion tool. The paper does provide plausible internal evidence, such as the held-out MD colloid test, and the KAN architecture is a reasonable approach to continuous scattering functions. However, the central claim of direct, model-independent structural extraction goes beyond what is demonstrated: the experimental lamellar fit has acknowledged low-Q discrepancies, no uncertainties are reported, and no test shows that the fitted parameters correspond to true real-space structure. I therefore agree with the conditional verdict rather than rejecting the paper, since the missing validation is concrete and obtainable, and the core methodology is sound.","tokens_in":13343,"tokens_out":4296,"duration_ms":39711,"concrete_test":"Generate synthetic SANS curves from an independent forward model not used in training, e.g., a real-space Monte Carlo simulation of defective lamellae with known sigma_k, Gamma, alpha, or a Nallet-type lamellar model with controlled disorder, across the experimental Q-range. Run Algorithm 1 with the trained lamellar KAN on these curves, from at least 20 random initializations, and compare the recovered parameters to the known true values. If recovery is biased, fails, or has overlapping uncertainties that cannot distinguish distinct true parameter sets, the model-independent inversion claim is refuted; if recovery is accurate with tight uncertainties, the claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that the training libraries constitute a complete and faithful generative description of the experimental systems, because the KAN only reproduces those libraries. Section III.A trains on two-point correlation functions from the wave-field representation of Ref. 10, and Section III.B trains on Yukawa/MD structure factors from the authors' earlier work. Consequently, the fitted parameters are meaningful only insofar as these forward models, not the KAN itself, correctly describe AOT lamellae and silica colloids. The paper explicitly asserts in Section III.A that defective lamellar phases are 'uniquely characterized by three parameters' (sigma_k, Gamma, alpha), but this premise is imported from Ref. 10 and is not re-derived or tested here. The paper's own low-Q fit residuals for AOT (Q*tilde_d < 1) show that the training model can miss real experimental features, and no independent ground truth or uncertainty is provided for the extracted (sigma_k, Gamma, alpha) or V_R(r). Thus the abstract's 'model-independent' claim and the conclusion's 'transformative' claim are not supported unless parameter recovery is validated on data generated outside the training family.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a Kolmogorov-Arnold Network (KAN) framework for small-angle scattering inversion: a continuous, differentiable surrogate I(Q;\\lambda) that can be used in least-squares fitting of experimental SANS data at arbitrary Q points. Two applications are presented: defective lamellar phases, where the KAN maps (\\sigma_k, \\Gamma, \\alpha) to the scattering intensity, and charged colloidal suspensions, where the KAN maps (\\phi, 1/\\kappa D, \\ln A, QD) to the direct correlation function relative to a hard-sphere reference. The authors fit experimental AOT SANS data and silica colloidal SANS data, and benchmark the colloidal S(Q) against an MD test set.","tokens_in":13634,"tokens_out":4452,"duration_ms":39731,"significance":"If validated, the KAN approach would be a useful addition to the SAS analysis toolbox because it provides continuous, differentiable scattering functions that avoid fixed-Q grids and can support gradient-based minimization. The colloid case is the stronger part: the KAN reproduces an independent MD test set across \\phi from 0.045 to 0.405, and the hard-sphere-referenced formulation is a sensible way to reduce the learning target. The lamellar case is weaker: the experimental fits show acknowledged low-Q discrepancies and no reported parameter values or uncertainties. The main gap is that the central claim of a 'model-independent, data-driven approach' is not supported, because the training data are generated from the authors' own forward models and no test on independently generated scattering data is reported. The paper also does not report quantitative goodness-of-fit statistics for the experimental inversions, which limits the ability to judge the accuracy of the extracted structural parameters.","major_comments":[{"comment":"The central claim of a 'model-independent, data-driven approach' is not supported by the presented evidence. The KAN is trained exclusively on simulation libraries produced by the authors' own wave-field model (Ref. 10) for lamellae and by the authors' MD pipeline (Ref. 3) for colloids. Consequently, the experimental inversion returns parameters of those same forward models, and the fitted parameters have real-space meaning only to the extent that those models are correct. Held-out tests within the same pipeline validate the KAN as a surrogate for the forward model, but they do not establish model independence. The authors should either temper the 'model-independent' claim in the abstract and conclusions or demonstrate parameter recovery on scattering data generated by an independent model outside the training family (e.g., a different simulation approach or a different closure relation).","section":"Abstract; Section III.A; Section III.B"},{"comment":"The lamellar experimental fits are not quantitatively documented. The text acknowledges significant discrepancies for Q\\tilde{d} < 1, yet the paper reports no goodness-of-fit statistics, no fitted values of (\\sigma_k, \\Gamma, \\alpha) for the 30%, 40%, and 50% AOT samples, and no uncertainty estimates from Algorithm 1. Without these numbers, the three-dimensional structural renderings in Fig. 5 and the claim that the KAN 'enabled the resolution of structural distortions' cannot be assessed. Please report the fitted parameters, their uncertainties, and residual statistics separately for Q\\tilde{d} < 1 and Q\\tilde{d} > 1.","section":"Section III.A; Fig. 4"},{"comment":"The experimental colloid inversion is not sufficiently validated or documented. The particle diameter D=115 nm and 6% polydispersity are obtained from a fit to the 1 wt% data, but no confidence intervals for D are given. The fitted potential parameters (A, \\kappa, \\phi) for each weight fraction are also not reported. The VR(r/D) curves in Fig. 8(b) are shown without error bars, and no comparison with an independent inversion method (e.g., HNC, RMSA, or an alternate closure) is provided. As a result, the conclusion that the KAN achieves 'excellent agreement with both molecular dynamics simulations and experimental data' is only demonstrated for S(Q) on the MD test set, not for the experimental potential inversion. Please provide the fitted values, uncertainties, and an independent cross-check of the inverted potential.","section":"Section III.B; Fig. 8"}],"minor_comments":[{"comment":"The sentence 'This report contends that achieving a perfect closure for solving the inverse scattering problem is unattainable' uses 'report' in an unusual way; 'we contend' or 'the present work contends' would be clearer.","section":"Section III.B"},{"comment":"The notation 'D 2O' should be rendered as 'D2O' with proper subscript formatting, and the tilde placement in 'Q \\tilde{d}' should be defined explicitly to avoid ambiguity in plain text.","section":"Section III.A; Section IV"},{"comment":"The training set details are not reported: no number of training samples, no parameter ranges for (\\sigma_k, \\Gamma, \\alpha) or (\\phi, 1/\\kappa D, \\ln A), and no network hyperparameters (depth, width, spline knots, learning rate). Providing these details would aid reproducibility.","section":"Section III.A; Section III.B"},{"comment":"The convergence threshold L_c and the specific minimization algorithm used in each case study are not stated; please specify these choices so that the fitting procedure is reproducible.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The paper's main novelty claim rests on the KAN framework being 'model-independent', but the training data are generated from the authors' own forward models. A revision that either adds an independent-model validation or substantially softens this claim, together with proper reporting of the fitted parameters and uncertainties, would address the most serious concerns. Given that the manuscript is a methods paper in a machine-learning-for-physics area, I also encourage the editor to request code and trained KAN weights for reproducibility; the current data availability statement ('available from the corresponding author upon reasonable request') is unlikely to be sufficient for the journal's standards."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuine methodological step for the SAS community — a KAN-based continuous surrogate for I(Q) that slots into least-squares fitting and sidesteps fixed Q grids. The colloid case is the stronger half; the lamellar case is a sketchier feasibility demo. The abstract's 'model-independent' and 'transformative' language is not supported.\n\nWhat's actually new: first application of KAN to scattering inversion, as far as I know. The reference-subtraction architecture (HS reference for colloids, linear-plus-quadratic standardization for lamellae) is a real design choice, and it makes sense. The colloid test on an MD test set shows quantitative agreement across a wide phi range, and the experimental silica fit gives a plausible D=115 nm and VR(r) curves that trend the right way with concentration. The authors are also honest about the lamellar low-Q discrepancy and offer plausible physical speculations (grain orientation, P(k) ansatz) rather than sweeping it under the rug.\n\nThe soft spots are the ones the stress-test flagged. The KAN is a surrogate for the authors' own forward models (wave-field representation for lamellae, MD/Yukawa for colloids). Testing on a held-out set from the same pipeline is internal consistency, not model-independence. There is no test on synthetic data generated from a different forward model, so 'model-independent' is simply wrong. The lamellar fit has no uncertainties, no goodness-of-fit statistic, and the three-parameter uniqueness premise is imported from Ref. 10. D=115 nm is fixed from the 1 wt% fit and propagated without error. No code or data are released.\n\nBut these are fixable. The core method is sound and likely useful. I'd want the authors to add uncertainty quantification (e.g., bootstrap or MCMC around the KAN surrogate), test recovery on out-of-family synthetic data, report the fitted parameters with errors, and soften the claims. That is a revision path, not a rejection path.\n\nI'd send this to a referee who knows both SAS analysis and ML surrogates. It's not a desk reject. For my own reading group, maybe — the colloid half is a nice example of surrogate-based inversion, and the discussion of the lamellar limitations is instructive.","headline":"A useful KAN surrogate for SAS fitting, with a strong colloid demo and a weaker lamellar half; the 'model-independent' claim overreaches.","tokens_in":14185,"tokens_out":2521,"would_cite":false,"duration_ms":21026,"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 Kolmogorov-Arnold network trained on computed scattering curves can invert measured small-angle scattering data into real-space structural parameters without closed-form analytical models.","keywords":["small-angle neutron scattering","Kolmogorov-Arnold network","structural inversion","lamellar phases","colloidal suspensions","machine learning","structure factor","effective interaction potential"],"falsifier":"Take a lamellar sample whose real-space defect structure has been determined independently by electron microscopy, compute the KAN-inferred $(\\sigma_k, \\Gamma, \\alpha)$ from its SANS curve, and check whether the reconstructed three-dimensional structure reproduces the observed layer connectivity and defect distribution; if it does not, the three-parameter representation is insufficient for structural inversion.","tokens_in":13122,"feed_emoji":"🔬","tokens_out":14902,"duration_ms":109820,"temperature":0.7,"pith_summary":"The paper sets out to show that a Kolmogorov-Arnold Network (KAN) can serve as a continuous, differentiable surrogate scattering function for small-angle scattering data, allowing the inverse problem—recovering real-space structure from reciprocal-space intensity—to be solved by ordinary least-squares fitting without a closed-form analytical model. The authors argue this matters because many soft-matter systems, such as defective lamellar phases, have no tractable analytical scattering function, and because earlier convolutional-network approaches require evenly spaced $Q$ grids that real instruments rarely supply. They demonstrate the idea on two systems: lyotropic lamellar phases, where the KAN recovers three structural parameters that render the layer topology, and charged colloidal suspensions, where the KAN maps the measured structure factor to an effective screened-Coulomb interaction potential. If the demonstrations hold, a trained KAN becomes a reusable model library for scattering analysis, and the same recipe should transfer to other soft-matter systems.","feed_headline":"Neural network inverts scattering data without analytical models","feed_subtitle":"Using a Kolmogorov-Arnold network, the fitted function recovers lamellar and colloidal structures from scattering data.","key_machinery":"The load-bearing object is the KAN itself, built on the Kolmogorov-Arnold representation theorem, which writes a multivariate function as sums of univariate spline functions. Two stacked KAN blocks separate physical parameters from $Q$: the first maps parameters to latent variables, the second combines those latent variables with $Q$ to emit a continuous $I(Q)$, so no fixed $Q$ grid is needed. A second device is reference-system standardization: the lamellar network learns the ratio $I(Q)/I_{\\mathrm{ref}}(Q)$ against a linear baseline, and the colloidal network learns deviations from the analytic hard-sphere structure factor, which concentrates the network's capacity on small corrections and improves numerical stability.","core_discovery":"The authors' central claim is that a two-stage KAN can learn the mapping from structural parameters and wave vector $Q$ to scattering intensity $I(Q)$ well enough that the trained network, inserted into a least-squares fitting loop, returns physically meaningful parameters from experimental data. For defective lamellar phases, the network takes $(\\sigma_k, \\Gamma, \\alpha)$—the wave-vector spread, symmetry ordering, and amphiphile-to-water volume ratio—and outputs $\\ln I(Q\\tilde d)$; the fitted values reproduce the measured SANS curves and, when fed back into the wave-field representation, show a concentration-driven evolution from isolated anisotropic plates to connected, more isotropic structures. For charged colloids, the network takes $(\\varphi, 1/(\\kappa D), \\ln A, QD)$ and outputs corrections to a hard-sphere reference structure factor; the resulting $S(Q)$ matches molecular-dynamics test sets, and fitting experimental data from silica dispersions yields effective interaction potentials whose repulsive strength grows with concentration.","pith_inferences":["The method's independence from closed-form scattering expressions does not make it independent of the physical assumptions built into the training library; the inversion inherits those assumptions through the simulated or modeled training data.","The fitted parameters carry no built-in uncertainty estimate, so a natural extension is to attach one, for example by training an ensemble of KANs or adding synthetic noise.","The continuous-$Q$ property suggests the framework can be applied to time-resolved or scanning scattering experiments where the $Q$ grid changes between frames, without the data preprocessing that grid-based networks require.","One could test the scheme's limits by generating scattering curves from structures outside the training library and checking whether fitting recovers a plausible parameter set or silently returns an in-distribution compromise."],"forward_implications":["A trained KAN can fit SANS data recorded at arbitrary, unevenly spaced $Q$ values without rebinning or interpolation, so one network can analyze data from different instruments and configurations.","Because the KAN is differentiable in its input parameters, standard gradient-based least-squares minimizers can be used for fitting, which is not practical for activation-heavy convolutional networks.","For defective lamellar phases, the extracted $(\\sigma_k, \\Gamma, \\alpha)$ parameters yield three-dimensional structures, so scattering alone can visualize the transition from ordered lamellae toward sponge-like, interconnected phases.","For charged colloids, the KAN replaces approximate closure relations used in integral-equation inversion, extending feasible analysis to concentrated or strongly coupled regimes where those closures fail to converge.","The same surrogate-plus-least-squares recipe should apply to other scattering problems, as long as a reference system and a sufficiently representative training library can be constructed."],"supporting_citations":[{"why":"Introduces the KAN architecture with spline-based univariate functions that the paper adapts to generate continuous scattering intensities.","marker":"[11]"},{"why":"States the Kolmogorov-Arnold representation theorem that motivates the network's decomposition of multivariate scattering functions into univariate components.","marker":"[12,13]"},{"why":"Provides the wave-field representation, the three-parameter characterization of defective lamellar phases, and the training library of two-point correlation functions used in the lamellar inversion.","marker":"[10]"},{"why":"Supplies the earlier machine-learning inversion framework and the training set of structure factors and potential parameters for charged colloidal suspensions.","marker":"[3]"},{"why":"Gives the analytic hard-sphere structure factor used as the reference system from which the colloidal KAN learns corrections.","marker":"[77]"},{"why":"Defines the screened-Coulomb interaction potential whose parameters the colloidal inversion recovers.","marker":"[58]"},{"why":"Identifies the convergence limits of a standard closure approximation that motivate replacing integral-equation closures with the KAN regression.","marker":"[70]"}],"fun_headline_variants":["KAN turns scattering spectra into real-space structure","Model-free KAN decodes soft-matter structure from scattering","No analytical models needed: KAN inverts scattering data","KAN maps scattering to structure, skipping analytical models","KAN extracts real-space structure from scattering without closed forms"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The lamellar half of the paper stands on the assumption, taken from earlier work, that every defective lamellar structure encountered in the experiments is uniquely captured by the three parameters $(\\sigma_k, \\Gamma, \\alpha)$ used to build the training library; if that wave-field representation is wrong or incomplete, the fitted numbers will have no direct real-space meaning.","fun_headline_variants_meta":{"raw":{"variants":["KAN turns scattering spectra into real-space structure","Model-free KAN decodes soft-matter structure from scattering","No analytical models needed: KAN inverts scattering data","KAN maps scattering to structure, skipping analytical models","KAN extracts real-space structure from scattering without closed forms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001581,"raw_usage":{"total_tokens":6283,"prompt_tokens":898,"completion_tokens":5385,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":5306}},"tokens_in":514,"tokens_out":5385,"duration_ms":29099,"temperature":1.0,"reasoning_tokens":5306,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:23:56.791340+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a lamellar sample whose real-space defect structure has been determined independently by electron microscopy, compute the KAN-inferred $(\\sigma_k, \\Gamma, \\alpha)$ from its SANS curve, and check whether the reconstructed three-dimensional structure reproduces the observed layer connectivity and defect distribution; if it does not, the three-parameter representation is insufficient for structural inversion.","supporting_citations":[{"cited_title":"\\ Tung , author Y.-J","cited_arxiv_id":null,"evidence_quote":"Introduces the KAN architecture with spline-based univariate functions that the paper adapts to generate continuous scattering intensities."},{"cited_title":"\\ Tung , author M.-Z","cited_arxiv_id":null,"evidence_quote":"Provides the wave-field representation, the three-parameter characterization of defective lamellar phases, and the training library of two-point correlation functions used in the lamellar inversion."},{"cited_title":"Pedersen ,\\ in\\ @noop booktitle Neutron, X -Ray and Light: Scattering Methods Applied to Soft Condensed Matter \\ ( publisher edited by Th","cited_arxiv_id":null,"evidence_quote":"Supplies the earlier machine-learning inversion framework and the training set of structure factors and potential parameters for charged colloidal suspensions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the analytic hard-sphere structure factor used as the reference system from which the colloidal KAN learns corrections."},{"cited_title":"Kékicheff , author B","cited_arxiv_id":null,"evidence_quote":"Defines the screened-Coulomb interaction potential whose parameters the colloidal inversion recovers."},{"cited_title":"Klein \\ and\\ author B","cited_arxiv_id":null,"evidence_quote":"Identifies the convergence limits of a standard closure approximation that motivate replacing integral-equation closures with the KAN regression."}],"review_version":1}