{"id":"25c0e6ed-0bb1-4e83-8500-24bb219e2e0f","arxiv_id":"2507.01706","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A modified Levenberg-Marquardt optimizer with an autocorrelation-based phase objective reduces the number of simulations needed to identify elastic parameters of isotropic polymer waveguides.","lead":"This paper proposes a faster way to estimate the elastic properties of polymers from ultrasonic measurements, using a modified optimization algorithm and a new way to compare measured and simulated signals. If the method works on real specimens, it could speed up industrial quality control and materials testing.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed speed advantage rests on the autocorrelation-phase objective being effectively convex over the initial-condition region, but the paper tests only one visualized surface and unreported starting points, and explicitly concedes it has no proof.","rationale":"The reader's weakest assumption identifies the same load-bearing gap: the new objective's convexification is essential for the local optimizer to find the correct minimizer, yet it is explicitly unproven. I agree with the conditional verdict and would keep it. The derivation of the step-size rule is internally consistent and the synthetic benchmark is a reasonable first test of optimizer iteration counts, but the validation does not establish the claimed global or regional convexity. The paper reports one illustrative objective surface, does not report the distance of the initial estimates to the true parameters, and provides no code or data for independent replication. A broader multistart study and a Hessian positive-definiteness check would settle whether the speed advantage is robust across the declared search space or only valid for favorable starting points. The conclusion's caveat about transverse isotropy and damping is an acknowledged limitation, but it does not address the isotropic case, where the convexification claim is used. No theatrical language is needed: the concern is precise and testable, and the conditional verdict already captures the required additional validation.","tokens_in":21425,"tokens_out":10667,"duration_ms":132720,"concrete_test":"Run a multistart experiment over the full declared isotropic search space, for example 1000 Latin-hypercube initial points spanning at least three standard deviations of the fitted gamma distributions for PEEK, PA6, and PP. For every start, run both the modified Levenberg-Marquardt method and BFGS with Hager-Zhang line search to the same tolerance, recording success rates and total forward-model evaluations. In parallel, evaluate the Hessian of the Eq. (59) objective on a dense grid in E and nu and check for negative eigenvalues. If any start outside a small neighborhood of the true parameters fails to reach the 1e-6 error threshold, or if the Hessian is not positive semidefinite on the grid, then the convexification assumption is falsified and the speed claim does not extend to the plotted search space.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. 3.4 claims that the residual in Eq. (59) \"completely convexifies the objective function within the search space,\" but the same paragraph states: \"we hypothesize that this mixture ... relaxes and untangles the phases, although we do not have a concrete proof.\" The central comparison in Sec. 4.2 counts forward-model evaluations until the relative parameter error drops below 1e-6, using initial estimates generated \"using the methodology presented in [83].\" The distance of those starting points from the true parameters is not reported, and Fig. 8 shows the objective surface for only a single reference. If the starts already lie in one locally convex basin, the experiment measures only local convergence speed and does not substantiate the abstract's practical claim that the method shortens parameter identification over the intended search region. If Eq. (59) has additional local minima or flat regions elsewhere, the modified Levenberg-Marquardt method, which has no line search or global safeguard, can converge to the wrong parameters and the evaluation-count advantage disappears. The unproven convexification is therefore the load-bearing link between the new residual and the headline speed claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two methodological innovations for estimating elastic material parameters of polymer waveguides from ultrasonic transmission signals: (i) an automatic step-size rule for the Levenberg-Marquardt (LM) method, derived from a geometric interpretation of nonlinear least squares, and (ii) a new objective function based on the phase of the autocorrelation of the signal's frequency content. The method is tested on virtual measurements for three isotropic polymers (PEEK, PA6, PP), where the SBFEM forward model is used both to generate synthetic reference signals and to evaluate the objective during inversion. The central performance claim is that the proposed LM variant with the autocorrelation-phase residual reaches a relative parameter error below 1e-6 in fewer forward-model evaluations than BFGS with Hager-Zhang line search for all three materials considered.","tokens_in":21732,"tokens_out":6048,"duration_ms":70473,"significance":"If the claimed speedup and convexification hold beyond the presented test cases, the paper makes a useful contribution to ultrasonic material characterization by reducing the number of expensive wave-propagation solves in gradient-based inversion. The geometric derivation of the LM damping parameter is elegant and parameter-free, and the careful construction of marginal gamma distributions for PEEK, PA6, and PP from aggregated datasheet values is a reproducible and practical contribution. The comparison metric is fair in counting line-search evaluations of BFGS. However, the significance is tempered by two factors the manuscript itself acknowledges: the validation is entirely synthetic (references come from the same SBFEM forward model used in inversion), and the convexification property of the new residual is stated as a hypothesis, not a proven fact. As a methods paper with empirical support on a synthetic benchmark, the work is worthwhile, but the abstract and conclusions currently claim more than the evidence substantiates.","major_comments":[{"comment":"The claim that the autocorrelation-phase residual 'completely convexifies the objective function within the search space' is load-bearing for the speedup claim, yet the same section states: 'we hypothesize that this mixture ... relaxes and untangles the phases, although we do not have a concrete proof.' The visual evidence in Fig. 8 shows only one reference (caption: 'for a particular reference') and only a two-standard-deviation box. Since the proposed LM method has no line search or global safeguard (Sec. 3.3), any unobserved local minima or flat regions outside this box could invalidate the convergence and the evaluation-count advantage. The authors should either provide a proof of unimodality on the search space (even for a simplified or reduced-order problem) or, failing that, run a systematic numerical scan over many randomly drawn references and initial points, reporting the basin of attraction and failure rates. Without this, the abstract's statement that the method 'shortens the time to identify the material parameters' is not fully supported.","section":"Sec. 3.4, Fig. 8"},{"comment":"The performance comparison is self-referential and the initial conditions are not characterized. All 60 virtual measurements are produced by the same SBFEM forward model that is used in the inversion (the abstract acknowledges that the study 'primarily relies on simulation data'), so the test measures the optimizer's behavior on synthetic data with perfect model correspondence, not on real measured signals. More importantly, the initial estimates are 'generated using the methodology presented in [83]' and the distance of those starting points from the true parameters is never reported. If these starts lie inside the single visually convex basin of Fig. 8, the experiment measures local convergence speed only; it does not substantiate a claim about global robustness over the intended search region. Please report the distribution of initial parameter errors, success/failure counts as a function of initial distance, and ideally repeat the comparison with deliberately perturbed starting points spanning the full search space defined by the gamma distributions in Table 1.","section":"Sec. 4.2"},{"comment":"The proposed step-size rule λ_k squared equals (Δx*_k)ᵀG⁻¹Δx*_k / (Δx*_k)ᵀGΔx*_k, which requires the metric G = JᵀJ to be positive definite and the gradient descent step Δx*_k to be nonzero. Near a stationary point, in a flat direction, or when the Jacobian is rank-deficient (a common situation in parameter-identification problems with correlated parameters), λ_k and hence η = η_k λ_k⁻¹ may be undefined or unbounded. The manuscript states that η_k may exceed one for uphill steps and that 'This is not problematic in and of itself', but no safeguard or fallback is described. Since no line search is used, the authors should state explicitly under what conditions Eq. (38) yields a descent direction and describe what the implementation does when G is singular or when λ_k cannot be computed.","section":"Sec. 3.3, Eq. (34)"},{"comment":"The new residual introduces a new free hyperparameter, the damping coefficient C ∈ [1,10] in the stability weights γ_k. This is in tension with the paper's claim to avoid hyperparameters (Sec. 1), and the objective surface shown in Fig. 8 is computed with C=1 only. No sensitivity study with respect to C is presented, so it is unknown whether the claimed convexification and the subsequent speedup are robust to the choice of C. A small scan (e.g., C = 1, 2, 5, 10) for at least one material should be added, and the role of C as a user-set parameter should be acknowledged in the summary of the method.","section":"Sec. 3.4, Eq. (57)"}],"minor_comments":[{"comment":"The text states 'We choose ¯f = 1 GHz' for the center frequency. This appears inconsistent with the time axis of Fig. 2 (in microseconds) and with typical ultrasonic polymer testing; if this is a typo for 1 MHz, please correct it in the text and in any related discussion.","section":"Sec. 2, Eq. (5)"},{"comment":"Several figure captions contain unresolved placeholder text '( ??)', e.g., Fig. 1, Fig. 3, and Fig. 8. These internal cross-references need to be completed before the manuscript is sent to production.","section":"Figure captions"},{"comment":"The phrase 'has to equally fullfull' contains a typo and should read 'has to equally fulfill'.","section":"Sec. 3.2, p. 7"},{"comment":"The procedure for fitting gamma distributions is described, but the actual data ranges from the cited sources are not reported (due to licensing). For reproducibility, the authors should at least provide the aggregated parameter intervals or quartiles that were used, so that readers can reproduce the marginal distributions without access to the original datasheets.","section":"Sec. 4.1"},{"comment":"The generation of initial estimates via 'the methodology presented in [83]' is not described in this paper. Since the initial-point distance is critical to the interpretation of the comparison, a one-sentence description of that methodology (or a citation to an accessible source) is needed for reproducibility.","section":"Sec. 4.2, reference [83]"}],"recommendation":"major_revision","confidential_remarks":"The paper is in scope for a computational/engineering journal, and the geometric derivation is interesting. My main concern is the gap between the headline claim of practical time savings and the purely synthetic validation with unreported initial conditions. The authors are transparent about the hypothesis status of the convexification, which is to their credit, but the abstract should be toned down until either a proof or a systematic robustness study is provided. The introduction of the new C hyperparameter also weakens the 'no hyperparameters' narrative. I would look favorably on a revision that adds a basin-of-attraction study and a sensitivity analysis with respect to C."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version: this is a genuinely useful methods paper, not a breakthrough, and it deserves a serious referee. The two contributions are real: the automatic step-size adaptation for Levenberg-Marquardt (Eq. 34 and 38) and the autocorrelation-phase residual objective (Eq. 59). Both are new relative to the cited envelope-based work, and the derivations hold up. The geometric story—Gauss-Newton and gradient descent as dual projections—is clearly told, and Eq. (34) follows from the metric-distance argument without fitting. The autocorrelation derivation (Hilbert transform, convolution, Eq. (51)) is correct. I also give the authors credit for counting BFGS line-search evaluations in the comparison, not just iterations, which is fairer than most papers in this space. The material parameter distributions for PEEK, PA6, and PP are a practical reproducibility artifact.\n\nThe soft spots are in the validation, not the derivation. The load-bearing claim is that Eq. (59) 'completely convexifies' the objective, but the same paragraph concedes 'we hypothesize... we do not have a concrete proof.' That is an explicit gap, and the paper only shows one visualized surface for one reference. The convergence experiments use starting points generated by [83] without reporting their distance from the true parameters, so we cannot tell whether the speed comparison reflects global basin behavior or just local convergence rate. If Eq. (59) has other minima outside the visualized region, the no-line-search LM method can converge to the wrong parameters and the evaluation-count advantage disappears.\n\nSecond, the validation is entirely virtual: the references are simulated with the same SBFEM forward model used for inversion. The authors are upfront about this, which I respect, but it means the practical claim is not yet supported by evidence that includes model error or measurement noise. And there is no code or data, so reproducing the comparison requires reimplementation.\n\nWho is this for? Anyone working on inverse problems with expensive forward models and oscillatory signals, and anyone interested in geometric interpretations of least squares. I would send it to peer review. The referee should require concrete progress on the convexification question—a proof, a counterexample, or at least a systematic numerical survey over the parameter range—and ideally a real-measurement or perturbed-model case before the abstract's broad speed claim is accepted.","headline":"A clean derivation of a step-size-adapted Levenberg-Marquardt method and an autocorrelation-phase objective, but the headline speed claim rests on unproven convexification and same-model virtual measurements.","tokens_in":22190,"tokens_out":3328,"would_cite":true,"duration_ms":37972,"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 geometry-derived step-size rule and an autocorrelation-phase residual identify polymer-waveguide elastic parameters with fewer forward-model evaluations than a quasi-Newton baseline.","keywords":["material parameter estimation","inverse problem","Levenberg-Marquardt method","autocorrelation residual","model manifold","ultrasonic waveguides","polymers","least squares"],"falsifier":"Sample the Hessian of the autocorrelation-phase least-squares objective on a fine grid of $(E,\\nu)$ pairs inside the two-standard-deviation boxes of the three materials; if any sampled Hessian has a negative eigenvalue, the claimed convexification inside the search space is false. Equivalently, run the optimizer from many starting points on real measurement signals of PEEK, PA6 and PP and count failures to reach a $10^{-6}$ relative parameter error.","tokens_in":1950,"feed_emoji":"📉","tokens_out":2477,"duration_ms":119750,"temperature":0.7,"pith_summary":"The paper tries to establish that two coupled changes make inverse material-parameter identification in ultrasonic waveguides faster and more reliable. The first is a parameter-free automatic step-size rule for the Levenberg-Marquardt method, derived by viewing the forward model's outputs as points on a curved model manifold and treating Gauss-Newton and gradient descent as dual projections. The second is a new objective function that compares the unwrapped phases of the autocorrelation of the squared signal envelopes in the frequency domain, instead of comparing raw time signals. The paper argues that this autocorrelation-phase transformation convexifies the objective over the relevant parameter ranges for isotropic polymers, so a local optimizer can find the true parameters from realistic starting points. On virtual measurements for PEEK, PA6 and polypropylene, the method reaches a relative parameter error below $10^{-6}$ with fewer total forward-model evaluations than the reference quasi-Newton method, and since the forward model is the computational bottleneck, fewer evaluations mean shorter identification times.","feed_headline":"Fewer simulations to pin down polymer elastic parameters","feed_subtitle":"A phase-based residual and a self-scaling optimizer cut the forward solves needed for PEEK, PA6 and PP.","key_machinery":"The organizing object is the model manifold $\\mathcal{M} = \\{f(x) \\in \\mathbb{R}^n \\mid x \\in \\mathbb{R}^m\\}$, the curved surface swept out in signal space as the parameters vary. Gauss-Newton is the orthogonal projection of the residual $r = \\hat{y} - f(x)$ onto the tangent space $T_{f(x)}\\mathcal{M}$, while gradient descent projects onto the dual basis. The new machinery is the scalar factor $\\lambda_k$ that rescales the pure descent direction to the metric length of the Gauss-Newton step, combined with the Levenberg-Marquardt damping $\\eta = \\bar{\\eta}_k \\lambda_k^{-1}$, and the objective residual $r = \\mathrm{arg}_{\\mathrm{stable}}(\\hat{a}) - \\mathrm{arg}_{\\mathrm{stable}}(a)$ computed from the positive-frequency autocorrelation coefficients $E_k = \\sum_{i=k}^{n_+-1} U_{i+1} \\bar{U}_{i-k+1}$ of the squared envelope, stabilized by a damping factor $\\gamma_k$. The autocorrelation-phase step is intended to \"relax and untangle\" the phases, producing a locally convex objective within the search space for isotropic symmetry, which is the property that makes a local optimizer reliable.","core_discovery":"The central claim is that the inverse problem of estimating frequency-dependent elastic parameters from ultrasonic waveguide responses can be made both more robust and cheaper by replacing the standard least-squares comparison of time signals with a phase residual between autocorrelations of the squared envelopes, and by equipping the Levenberg-Marquardt method with a self-contained geometric step-size selection. The step size is derived from the metric tensor $G=J^TJ$ of the model manifold: the proposed factor $\\lambda_k = \\sqrt{(\\Delta x_k^*)^T G^{-1} \\Delta x_k^* \\big/ (\\Delta x_k^*)^T G \\Delta x_k^*}$ rescales the gradient-descent direction to match the scale of the Gauss-Newton step, and the Levenberg-Marquardt damping is set to $\\bar{\\eta}_k \\lambda_k^{-1}$ with $\\bar{\\eta}_k = \\|\\Delta y_k\\| / \\|\\Delta y_0\\|$, requiring no hand-tuned hyperparameters. The paper demonstrates on virtual measurements for three isotropic polymers that this combination reaches relative parameter errors below $10^{-6}$ in fewer forward-model evaluations than the baseline quasi-Newton method, counting the line-search evaluations the baseline needs.","pith_inferences":["If the convexification property extends to transverse-isotropic symmetry and damping, which the authors flag as an open uncertainty, the same residual could stabilize a harder inversion problem that currently defeats first-order methods.","The phase-untangling mechanism is a hypothesis rather than a proof; a Hessian audit of the objective inside the search box would settle whether \"convexify\" is literal or merely empirical.","The geometric step-size derivation is independent of the autocorrelation objective, so either component could be swapped into other inverse problems; for instance, the same Levenberg-Marquardt scaling could accelerate any differentiable least-squares fit with expensive forward solves.","The gamma-distributed material priors give a ready-made test set for experimental validation: if real measurements on PEEK, PA6 or PP do not reproduce the virtual convergence statistics, the mismatch would point to model error rather than optimizer error."],"forward_implications":["For the isotropic case, the method reaches the $10^{-6}$ parameter-error cutoff in fewer total forward-model evaluations than the reference quasi-Newton method for PEEK, PA6 and PP; because the forward solve dominates compute time, this translates into shorter identification time.","The autocorrelation-phase objective makes the least-squares landscape convex over the tested parameter ranges, so first-order optimizers starting inside the material ranges should converge to the true parameters rather than to local minima.","The step-size rule is parameter-free, so it transfers to other least-squares inverse problems once Jacobians are available.","Counting line-search evaluations matters: the comparison shows that the baseline's implicit line search adds a significant number of forward solves, so fair speed comparisons should include them."],"supporting_citations":[{"why":"Defines the transmission measurement setup and transducer modeling that the inverse problem is based on.","marker":"[3]"},{"why":"Original Levenberg method that the proposed step-size adaptation modifies.","marker":"[12]"},{"why":"Original Marquardt damping interpolation used as the baseline for the Levenberg-Marquardt framework.","marker":"[13]"},{"why":"Shows that envelope-based residuals improve optimization for defect reconstruction, motivating the new autocorrelation-phase objective.","marker":"[14]"},{"why":"Provides the information-geometric model-manifold view used to derive the step-size rule.","marker":"[26]"},{"why":"Supplies the semi-analytical SBFEM forward solver whose evaluations are the computational bottleneck.","marker":"[31]"},{"why":"Provides the inverse-problem metric tensor and dual-step interpretation used in the geometric derivation.","marker":"[32]"},{"why":"Demonstrates differentiation of the SBFEM, enabling the analytic Jacobian required for the optimizer.","marker":"[34]"},{"why":"Supplies the reference quasi-Newton implementation with line search used as the performance baseline.","marker":"[82]"}],"fun_headline_variants":["Self-scaling LM with phase residual cuts forward solves","Autocorrelated residual and tuned LM speed polymer fits","Geometric LM step size reduces simulations for elastic params","New LM method and autocorrelation mean fewer model calls","Modified LM plus envelope autocorrelation lowers evaluations"],"cache_read_input_tokens":24320,"weakest_assumption_plain":"The decisive premise is that the autocorrelation-phase residual makes the objective function convex over the whole relevant parameter space for isotropic materials, a property the paper explicitly states as a hypothesis rather than a proof; if the objective retains hidden local minima, the optimizer's speed and reliability claims would not generalize beyond the tested instances.","fun_headline_variants_meta":{"raw":{"variants":["Self-scaling LM with phase residual cuts forward solves","Autocorrelated residual and tuned LM speed polymer fits","Geometric LM step size reduces simulations for elastic params","New LM method and autocorrelation mean fewer model calls","Modified LM plus envelope autocorrelation lowers evaluations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000165,"raw_usage":{"total_tokens":1286,"prompt_tokens":1016,"completion_tokens":270,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":632,"completion_tokens_details":{"reasoning_tokens":195}},"tokens_in":632,"tokens_out":270,"duration_ms":3963,"temperature":1.0,"reasoning_tokens":195,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:45:13.009509+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Sample the Hessian of the autocorrelation-phase least-squares objective on a fine grid of $(E,\\nu)$ pairs inside the two-standard-deviation boxes of the three materials; if any sampled Hessian has a negative eigenvalue, the claimed convexification inside the search space is false. Equivalently, run the optimizer from many starting points on real measurement signals of PEEK, PA6 and PP and count failures to reach a $10^{-6}$ relative parameter error.","supporting_citations":[{"cited_title":"Ultrasonic transmission measurements in the characterization of viscoelasticity utilizing polymeric waveguides","cited_arxiv_id":null,"evidence_quote":"Defines the transmission measurement setup and transducer modeling that the inverse problem is based on."},{"cited_title":"An Algorithm for Least-Squares Es- timation of Nonlinear Parameters","cited_arxiv_id":null,"evidence_quote":"Original Marquardt damping interpolation used as the baseline for the Levenberg-Marquardt framework."},{"cited_title":"Defect reconstruction in a two-dimensional semi-analytical waveguide model via derivative-based optimiza- tion","cited_arxiv_id":null,"evidence_quote":"Shows that envelope-based residuals improve optimization for defect reconstruction, motivating the new autocorrelation-phase objective."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the information-geometric model-manifold view used to derive the step-size rule."},{"cited_title":"Society for Industrial and Ap- plied Mathematics, 2005.doi:10.1137/1.9780898717921(cit","cited_arxiv_id":null,"evidence_quote":"Provides the inverse-problem metric tensor and dual-step interpretation used in the geometric derivation."},{"cited_title":"Shape sensitivity analysis of stress intensity factors by the scaled boundary finite element method","cited_arxiv_id":null,"evidence_quote":"Demonstrates differentiation of the SBFEM, enabling the analytic Jacobian required for the optimizer."}],"review_version":1}