{"id":"54ad46b1-4a32-45ca-a56e-d0a463a5318c","arxiv_id":"2607.16865","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Optimized specimen topologies, chosen by a Bayesian D-optimality criterion, increase expected information gain and improve parameter identification for hyperelastic materials in simulated uniaxial tests.","lead":"This paper uses computer optimization to design test-specimen shapes that produce the most informative deformation fields for identifying material properties. It couples Bayesian experimental design with topology optimization and shows in simulations that optimized shapes recover hyperelastic material parameters more accurately than unoptimized shapes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Design objective (18) relies on an unvalidated Laplace/Gauss-Newton proxy for EIG; if posterior is non-Gaussian or Hessian varies from P0, optimized specimens may not maximize true information gain.","rationale":"Agreement: The reader's weakest assumption (missing validation of the Laplace/Gauss-Newton EIG approximation) is the same load-bearing point. It is the most fundamental because it directly affects whether optimizing (18) is a valid way to maximize true information gain. The paper's own results largely evaluate the design with the same proxy (Table 2), so the 'better expected information gain' part of the central claim is circular absent an external check. The independent synthetic recovery tests (Table 3) are a step in the right direction, but they are based on a handful of single noise realizations and do not recover the full posterior/information gain; some anisotropic parameters (e.g., θ1) are not improved in the shown cases. The Laplace assumption is especially questionable here because the fiber-reinforced model (31) has multiple parameters on very different scales and orientations θ1, θ2 enter nonlinearly through invariants; the posterior may be far from Gaussian. A nested MC check would settle whether the optimized design actually outperforms arbitrary designs in expected KL. If it does, the conditional concerns are resolved; if not, the central claim fails. Secondary limitations—the failure of the image-based formulation for anisotropic materials (Section 4.2) and absence of code/data—further support conditionality but do not change the verdict.","tokens_in":18211,"tokens_out":8078,"duration_ms":86176,"concrete_test":"For the anisotropic example, case 7 (grip-constrained optimal design) at constant noise, compare true EIG against the OEIG values in Table 2 for designs B1–B5 using a nested Monte Carlo estimator: draw N=200 parameter samples from the prior (32), for each draw M=50 synthetic datasets with noise as in (30), solve the forward problem, approximate the posterior via MCMC or a Laplace approximation at each sample's MAP, and estimate the average KL divergence (6). If the optimized design is not top-ranked under this true-EIG estimate, the design objective (18) is not a reliable proxy and the central claim weakens.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2 defines true EIG in (6) and approximates it by (8) as 1/2 log det(I+C0 H(PMAP)) for nonlinear G, with H the Hessian of the negative log-likelihood. In the design step, (15)–(16) evaluate H at the prior mean P0, and the design problem (19) maximizes this local proxy. There is no validation that this proxy tracks the true EIG for the strongly nonlinear finite-strain hyperelastic models in Section 4. If the posterior is multimodal or the Hessian changes materially between P0 and the actual MAP, the optimized topology may be suboptimal for real information gain. Moreover, Table 2 compares the same approximate OEIG for optimized and arbitrary designs, so the 'better expected information gain' claim is circular: the design is chosen to maximize that quantity, and it is then evaluated with that same quantity. Table 3 provides independent synthetic recovery, but it reports single realizations without error bars and does not quantify information gain. The paper's own Section 4.2 limitation—that the image-based formulation failed to converge for anisotropic materials—further narrows the demonstrated scope. Therefore, the central claim that the framework 'leads to better expected information gain' is not established without a check of the approximation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for designing specimen geometries that maximize information gain for identifying hyperelastic material parameters, coupling Bayesian optimal experimental design with density-based topology optimization. The design objective is the local Laplace/Gauss-Newton D-optimal criterion of Eq. (18), with the Hessian of the negative log-likelihood evaluated at the prior mean (Eqs. 15–16). Adjoint equations (A1–A2) are derived for the design gradient, and the method is demonstrated on isotropic and fiber-reinforced hyperelastic models using synthetic data. The authors report that optimized specimens yield higher expected information gain and more accurate parameter recovery than a set of arbitrary designs (Tables 2–3).","tokens_in":18535,"tokens_out":4508,"duration_ms":47184,"significance":"If the claims are supported, the framework is a worthwhile contribution to material parameter identification: it replaces intuition-based specimen selection with a systematic, PDE-constrained optimization over a high-dimensional design space, and the adjoint-based sensitivity derivation is a concrete technical contribution. The synthetic recovery tests, using reference parameters different from the prior, are an appropriate first validation step. However, the central claim rests on a Laplace approximation to the expected information gain that is not independently verified, and the statistical evidence is based on single noise realizations. These gaps need to be addressed before the stated conclusions can be fully accepted.","major_comments":[{"comment":"The design objective replaces the true EIG with a local Laplace/Gauss-Newton approximation in which the Hessian H is evaluated at the prior mean P0 (Eqs. 15–16). Table 2 then compares optimized and arbitrary designs using this same OEIG, so the statement that the optimized specimen has 'better expected information gain' is, at this stage, a statement about the proxy, not about the true information gain. For a nonlinear, possibly multimodal posterior, the proxy may not track the true EIG. I request an independent check for at least the fiber-reinforced case: estimate the true EIG (e.g., by nested Monte Carlo or by MCMC posterior sampling using a small number of map evaluations) for the optimized design and for one or two arbitrary designs, and report the comparison. This is load-bearing for the central claim.","section":"§2.2, Eqs. (8), (15)–(16); §3.3, Eq. (18)"},{"comment":"The synthetic recovery experiments report percentage errors for a single noise realization at each setting. Because the data are noisy, a single draw may not be representative, and it is unclear whether the improvements over arbitrary designs (e.g., C1 error 4.4% vs. 4.4% for A2 in the last block) are statistically meaningful. Please repeat the experiment with multiple independent noise realizations (e.g., 10–20 seeds) and report means and standard deviations, or a paired comparison, for the optimized and the non-optimized designs. The claim that optimized designs 'lead to more accurate parameter inference' requires this type of evidence.","section":"§4.2, Table 3"},{"comment":"The text states that the image-based formulation 'often led to non-convergence' for anisotropic materials and that the anisotropic results are therefore confined to the displacement-based formulation. The image-based recovery results in Table 3 use displacement-optimized designs, not image-optimized ones. Consequently, the paper does not demonstrate the image-based branch of the framework for the main anisotropic target. Please either include a stabilized image-based design example or clearly delimit the conclusions to the displacement-based formulation. This is important because the introduction discusses image-based methods as a motivation, and the conclusion claims that 'direct image-based inference is valuable.'","section":"§4.2, image-based formulation"}],"minor_comments":[{"comment":"The fiber direction vectors are written as a=(cos θ1, sin θ2, 0) and b=(cos θ2, sin θ2, 0). The second component of a should presumably be sin θ1; please correct.","section":"Eq. (31)"},{"comment":"The heading 'V arious priors' contains a typo; it should read 'Various priors.'","section":"Table 1 caption"},{"comment":"The stopping condition '∥OEIG,k+1 − OEIG,k1∥' appears to have a typo: the subscript 'k1' should likely be 'k'; also use consistent norm notation.","section":"Algorithm 1"},{"comment":"The sentence 'These are also shown in Figure 2' seems to refer to the anisotropic results of Figures 3–4 or Table 2, not Figure 2 (isotropic). Please update the cross-reference.","section":"§4.2, text near Table 2"},{"comment":"The displacement-based Φd omits the image weighting ρ^{2p} present in Eq. (13); this may be intentional (displacements are defined on the body) but should be stated explicitly to avoid confusion.","section":"§3.2, Eq. (14)"},{"comment":"In Eq. (8), the notation H(P_MAP) is used, but in the design phase (Eqs. 15–16) H is evaluated at P0. The distinction between the two should be clarified in the text, since the approximation error is one of the main concerns.","section":"§2.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within a reasonable scope for this journal, and the general approach has merit. My major concerns are not about the feasibility of the method but about the strength of the evidence for the headline claims. The paper would be considerably stronger if the authors add a verification of the Laplace approximation against an independent EIG estimator and repeat the synthetic recovery experiments over several noise realizations. I do not see any citation or novelty concerns; the related work is cited appropriately."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper deserves a serious referee. It is a real contribution: the integration of Bayesian D-optimal experimental design with density-based topology optimization for designing specimens in anisotropic hyperelastic parameter identification is new, and the authors work through it carefully. The adjoint sensitivity derivation is sound, the implementation is described in enough detail to follow, and the synthetic recovery tests in Table 3 give independent support that the optimized specimen improves parameter inference compared to arbitrary designs. The paper also deserves credit for honestly reporting that the image-based formulation often failed to converge for the anisotropic case and then pivoting to a displacement-based formulation. That limitation is real but admitted, and the displacement-based route is defensible.\n\nThe soft spots are in proportion. The main one is the unvalidated Laplace/Gauss-Newton proxy for expected information gain. The design objective replaces true EIG with 1/2 log det(I + C0 H) evaluated at the prior mean, and the authors never check this approximation against a nested Monte Carlo estimate or any other full-EIG computation. For strongly nonlinear finite-strain models, posterior multimodality or a MAP far from the prior mean could make the optimized topology suboptimal for actual information gain. This is addressable but currently open. Second, the headline comparisons in Table 2 are partly circular: the same proxy that was optimized is used to evaluate the designs. The independent evidence is Table 3, which shows lower recovery errors for optimal designs, but it reports single realizations without error bars. Third, no code or data are released, so the central numerical results cannot be reproduced. None of these are fatal. The displacement-based central claim is credible, and the synthetic recovery tests are a genuine attempt to ground it.\n\nThe audience for this paper is experimental mechanicians and researchers in Bayesian OED who care about constitutive model identification. A reader interested in anisotropic soft tissues or full-field testing will find the framework and the resulting multi-directional specimen designs useful. It should go to peer review, not desk rejection, but the referees should push for a validation of the Laplace approximation, error bars on the recovery results, and ideally code or data release. If those are added, this could become a standard reference in the area.","headline":"A genuinely new coupling of Bayesian D-optimal design with topology optimization for hyperelastic specimen design; the displacement-based results hold up, but the EIG approximation is unvalidated and the image-based route is dropped for the anisotropic case.","tokens_in":645,"tokens_out":1155,"would_cite":true,"duration_ms":23745,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["74P10","62K05"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that specimen geometry can be designed by maximizing the expected information gain of a mechanical test, and that such optimized specimens yield more accurate material parameter identification than conventional or arbitrary","keywords":["Bayesian optimal experimental design","topology optimization","material parameter identification","hyperelasticity","expected information gain","D-optimality","adjoint sensitivity","anisotropy"],"falsifier":"Run the same OED for a hyperelastic model with a bimodal posterior (or with a prior mean far from the true parameters), computing the true EIG by nested Monte Carlo sampling for both the topology-optimized design and several conventional designs; if the optimized design's true EIG is not higher than the alternatives, the local D-optimal approximation is inadequate.","tokens_in":18074,"feed_emoji":"🧪","tokens_out":4964,"duration_ms":50466,"temperature":0.7,"pith_summary":"In material testing, constitutive parameters are inferred from experiments, but conventional specimen shapes (dogbones, notches) sample few deformation states, and identifying anisotropic models requires many tests. This paper proposes designing the specimen geometry itself as a topology optimization problem, using the Bayesian expected information gain as the objective: the design that maximizes the predicted information from a uniaxial test is found by gradient-based optimization over a material density field. The method is demonstrated on isotropic and fiber-reinforced hyperelastic models, where the optimized specimens produce heterogeneous deformation fields that excite multiple fiber orientations, leading to higher expected information gain and lower parameter inference errors than arbitrary designs. If it works in practice, a single well-designed specimen could replace several conventional experiments.","feed_headline":"Optimized specimens sharpen material parameter identification","feed_subtitle":"Bayesian experimental design shapes test samples so one uniaxial test excites the deformation states that identify anisotropic models.","key_machinery":"The expected information gain (EIG), approximated by the Laplace/Gauss-Newton D-optimal criterion 1/2 log det(I + C0 H(ρ)), with H the parameter-to-observable sensitivity (Hessian of the data misfit) evaluated at the prior mean. This criterion is optimized over a density field ρ via SIMP interpolation and adjoint-based gradient computation, coupling the design to the mechanical equilibrium equations.","core_discovery":"The paper's central claim is that the locally D-optimal Bayesian design criterion—maximizing 1/2 log det(I + C0 H(ρ)), where H is the Gauss-Newton Hessian of the data misfit at the prior mean—can be used as the objective of a density-based topology optimization to generate specimen geometries that are maximally informative for constitutive parameter identification. Optimizing this criterion over a SIMP density field, with sensitivities computed by adjoint equations, produces designs with multiple load-bearing struts oriented in different directions for anisotropic materials, which generate heterogeneous deformation states in a single uniaxial test. The paper reports that these optimized desi","pith_inferences":["Since the D-optimal approximation is exact only for linear maps, its accuracy for strongly nonlinear hyperelastic responses with possible multimodal posteriors remains untested; a nested Monte Carlo validation of the true EIG for the optimized designs would be a useful check.","Because the Hessian is evaluated at the prior mean, the method implicitly trusts the prior; in a truly unknown regime, sequential design (updating the prior with each experiment) would be needed to correct a poor initial guess.","The observed design robustness across priors may break down for parameters with weakly coupled sensitivities; a parameter-wise information budget could reveal which parameters actually drive the optimized geometry.","The computational cost of several hours per design suggests that surrogate models or reduced-order methods would be needed to extend the approach to 3D specimens or history-dependent materials."],"forward_implications":["Optimized specimens yield higher expected information gain and lower inference errors than conventional or arbitrary designs (Tables 2–3).","A single uniaxial test on an optimized specimen can recover anisotropic fiber angles and moduli with errors reduced relative to initial guesses or arbitrary designs.","Optimizing under the D-optimal criterion also improves the posterior mean-square error (A-optimal criterion) for the tested cases.","Designs are similar across ~15% prior variations, suggesting one-shot design is robust for these models.","The displacement-based OED route is more stable than image-based, pointing to a practical workflow for experimental implementation."],"fun_headline_variants":["Topology optimization designs specimens for maximum material info","Bayesian design shapes test specimens for better material models","Optimized specimens extract more from a single test","Specimen geometry tuned to reveal material parameters","Designing test samples to maximize what experiments tell you"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The framework's objective replaces the true expected information gain with a local Gaussian/Laplace approximation evaluated at the prior mean; if the posterior is multimodal or the true parameters lie far from the prior mean, the optimized design may not be the most informative.","fun_headline_variants_meta":{"raw":{"variants":["Topology optimization designs specimens for maximum material info","Bayesian design shapes test specimens for better material models","Optimized specimens extract more from a single test","Specimen geometry tuned to reveal material parameters","Designing test samples to maximize what experiments tell you"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000309,"raw_usage":{"total_tokens":1537,"prompt_tokens":613,"completion_tokens":924,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":357,"completion_tokens_details":{"reasoning_tokens":862}},"tokens_in":357,"tokens_out":924,"duration_ms":6575,"temperature":1.0,"reasoning_tokens":862,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T19:42:28.258313+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same OED for a hyperelastic model with a bimodal posterior (or with a prior mean far from the true parameters), computing the true EIG by nested Monte Carlo sampling for both the topology-optimized design and several conventional designs; if the optimized design's true EIG is not higher than the alternatives, the local D-optimal approximation is inadequate.","supporting_citations":[],"review_version":1}