{"id":"d63845a5-223b-48d5-9ea2-64ec3bfbac40","arxiv_id":"2412.14071","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A Bayesian optimization framework with an expanding search space inversely designs soft-interface parameters so that a nacre-inspired composite matches a target nonlinear stress-strain curve.","lead":"A computer optimization method finds the glue-like interface properties that make a layered, nacre-inspired composite follow a chosen stress-strain curve. It can return several different designs that behave the same macroscopically but fail in different ways.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Geometry-sensitivity study does not cover the optimized interface-parameter regime, so the single-geometry BO results may not transfer to a real composite.","rationale":"The reader's verdict is CONDITIONAL and identifies model fidelity and geometry sensitivity as the weakest assumption. My stress-test sharpens this into a concrete, testable gap: the sensitivity study in the SI covers only interface parameters within the initial design ranges, while the optimized designs and validation targets lie substantially outside those ranges, in regimes where damage localization is likely to be much more microstructure-sensitive. This does not undermine the paper's internal computational demonstration—the BO algorithm clearly recovers a known FEM-generated target and finds plausible non-unique designs—but it does weaken the broader claim that the recovered interface parameters are the material-level answer. The paper provides code and data, which is a real strength, and the validation against a known target is the strongest part of the evidence. However, the single-geometry pipeline means the load-bearing transferability claim rests on an extrapolation that the reported sensitivity study does not actually support. A simple multi-geometry re-evaluation of the final designs would settle the question, and the recommended condition is therefore unchanged: the paper should be accepted with the explicit condition that this geometry-robustness check is performed or the claims are re-scoped to a fixed microstructure.","tokens_in":12466,"tokens_out":5039,"duration_ms":49634,"concrete_test":"Evaluate the recovered designs from Table 1 (BO with expansion) and Table 2 (Design 1 and Design 2) in the other four Voronoi geometries from Fig. S1, plus at least five newly generated random Voronoi patterns at the same grain density, using the identical FEM setup. Compute the resulting stress-strain curves and the curve-difference metric Delta_TS relative to the target curves. If the inter-geometry spread in Delta_TS is comparable to the BO objective values (about 0.010) for all three designs, the geometry-independence assumption holds in the relevant regime and the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise for transferring the optimized interface parameters to a physical material is that the macroscopic stress-strain response is insensitive to grain arrangement at the chosen grain density. The Supporting Information tests this with five Voronoi patterns and three interface property sets (Table S1), but the BO solutions actually reported lie far outside those tested regions. The validation target has (sigma_n^o, sigma_s^o, delta_n^d, delta_s^d, delta_e) = (214.42, 159.16, 3.98, 12.38, 69.40) and the recovered expanded-space solution is (199.60, 167.17, 4.48, 10.14, 70.46); Design 2 in the non-unique case has sigma_s^o = 1.00 MPa, delta_s^d = 6.78 nm, delta_e = 80.02 nm (Table 2). The sensitivity study uses only strengths 125-200 MPa, separations 2-4 nm, and delta_e = 15 nm. Damage localization and crack-path sensitivity are known to depend strongly on interface strength and softening behavior; a microstructure that is insensitive for moderate interface properties can become highly sensitive when interfaces are very weak (sigma_s = 1 MPa) or very ductile (delta_e = 80 nm). Thus the statement that the composite's stress-strain response is insensitive to grain distribution at the current grain density is only established in a parameter regime that the optimization leaves. Since all BO iterations use geometry 4, the reported optimal designs and the non-unique failure-mode split could in principle be artifacts of one Voronoi realization. The Discussion acknowledges the 2D single-layer limitation, but it does not acknowledge this parameter-regime gap in the geometry-sensitivity claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper addresses the inverse design of interface parameters in a two-dimensional, Voronoi-tessellated nacre-inspired composite. The interface is modeled with a five-parameter bilinear traction-separation law (σ_n^o, σ_s^o, δ_n^d, δ_s^d, δ_e), and the forward response is computed by finite element analysis. The authors propose a Bayesian optimization (BO) framework with an expandable design space, using Expected Improvement as the acquisition function and a two-way pointwise curve-distance metric between the target and simulated stress-strain curves. They validate the method on a target generated from known parameters lying outside the initial design range, showing that the expansion recovers parameters close to ground truth, whereas a fixed design space pins at the boundary. They then demonstrate for a quadratic target curve that BO identifies two distinct interface designs with nearly identical stress-strain responses but different failure mechanisms (normal-dominated vs. shear-dominated fracture). The paper includes a geometry-sensitivity study in the Supporting Information and makes code and data available on GitHub.","tokens_in":12787,"tokens_out":7368,"duration_ms":60857,"significance":"If the claims hold, the paper makes a useful contribution by demonstrating that BO with a controlled, expandable design space can solve a class of inverse problems for nonlinear mechanical responses, and by showing that multiple physically distinct interface designs can produce nearly the same macroscopic response. The validation design (known ground truth outside the initial range) is clean, and the comparison with the no-expansion baseline clearly illustrates the benefit of expansion. The GitHub availability of code and data is a strength. However, as detailed in the major comments, the robustness of the central claims is currently limited by the narrow scope of the geometry-sensitivity study and the absence of repeated runs, and the non-uniqueness claim would be stronger if demonstrated for an attainable target.","major_comments":[{"comment":"The geometry-insensitivity claim, used in the main text to justify the use of a single Voronoi geometry for all BO runs, is established only for three moderate interface parameter sets (Table S1: strengths 125–200 MPa, separations 2–4 nm, δ_e = 15 nm). The optimized designs reported in Tables 1 and 2 lie far outside this regime: the validation solution has δ_e ≈ 70 nm and δ_s^d ≈ 10–12 nm, Design 1 has δ_e = 65.38 nm and δ_s^d = 0.99 nm, and Design 2 has σ_s^o = 1.00 MPa, δ_s^d = 6.78 nm, and δ_e = 80.02 nm. Because damage localization and crack paths are known to be sensitive to interface strength and softening behavior in this range, the statement that the response is insensitive to grain distribution at the current grain density has not been demonstrated for the parameter values that the optimization actually returns. Since every BO iteration uses geometry 4, the reported optimal parameters and the two-group failure-mode split could be artifacts of a single Voronoi realization. I request a geometry-sensitivity study that samples interface parameters from the optimized regions (including small σ_s^o and large δ_e), or an explicit statement that the optimized designs are only established for geometry 4.","section":"Model Material and Interface Law (main text) and Tables S1, 1, 2"},{"comment":"The central quantitative results are based on a single BO run per case, with one initial dataset of 50 curves and one geometry. The BO procedure is stochastic: the initial dataset is randomly generated, Gaussian-process hyperparameters are fitted, and the acquisition function is maximized numerically. The reported plateau of the objective after 45 iterations (Fig. 3C) and the appearance of two distinct groups among the top ten designs (Fig. 4F and Table S2) could depend on the particular run. Repeating the optimization with several random seeds or initial datasets would establish the robustness of these conclusions; without such repetitions, the claims of efficiency and of a natural split into two failure-mode groups are not fully supported.","section":"Results: Validation and Non-unique designs; Figs. 3 and 4"},{"comment":"The expansion scheme depends on two free parameters, the stall number S and the scaling parameter γ, and the refinement step in lines 11–13 contracts the search space to the bounding box of the previous space and the newly acquired designs. The paper adopts γ = N+1 and does not report the value of S or any sensitivity study with respect to these choices. The success of the expansion in the validation case may be sensitive to these parameters, particularly because an over-aggressive refinement could trap the search near the initial boundary, while an over-aggressive expansion could make the GP extrapolation unreliable. Please report the chosen S and include a brief sensitivity analysis (e.g., S ∈ {2, 5, 10} and a slower/faster γ) for the validation case.","section":"Methods: Algorithm 1"},{"comment":"The target curve in Case 2 is a quadratic function that is not known to be exactly representable by the five-parameter bilinear traction-separation model. The best objective values are 0.010–0.0176 (in the un-normalized units of Eq. (4)), and the paper does not establish whether the residual reflects optimization error or model mismatch. Consequently, the two distinct parameter groups may simply be two local minima for an unattainable target rather than genuinely non-unique solutions for an attainable target. To support the non-uniqueness claim, I recommend demonstrating with a target generated from a known parameter set (e.g., another ground-truth design) that BO, when run with different initialization or design-space expansion, recovers two or more distinct parameter sets that reproduce the target within the same tolerance. Alternatively, the paper should explicitly frame the result as approximate matching with non-unique local minima.","section":"Results: Non-unique Interface Designs; Fig. 4 and Table S2"}],"minor_comments":[{"comment":"Typo: 'feture' should be 'feature'.","section":"Results (text near Fig. 2F)"},{"comment":"Please specify the discretization counts N_T and N_S and the curve-interpolation procedure; the current definition leaves the metric's scale and normalization ambiguous.","section":"Methods, Eq. (4)"},{"comment":"Missing space in 'BO.(A)'; also, the caption does not identify which solution is Design 1 and which is Design 2.","section":"Fig. 4 caption"},{"comment":"The text selects 'geometry 4' but does not specify which of the five patterns in Fig. S1 corresponds to geometry 4; please label or reference it.","section":"Supporting Information, Fig. S1"},{"comment":"Reference [29] is incomplete: it lacks a journal or venue; please provide full bibliographic details.","section":"References"},{"comment":"The stall logic in Algorithm 1 is not immediately transparent: line 4 expands when t%S==1 and line 11 updates on t%S==0; a short clarification in the text or a pseudocode comment would help.","section":"Methods, Algorithm 1"},{"comment":"The paper does not report the number of FEM simulations needed to generate the 50-curve initial dataset or the per-simulation computational cost; stating these would help readers assess the practical budget for larger problems.","section":"Results: Computational budget"}],"recommendation":"major_revision","confidential_remarks":"The paper sits between machine-learning methodology and computational mechanics; it is within the journal's scope, but the contribution would be strengthened by a clearer separation of the novelty of the expandable-space BO algorithm from the specific nacre-inspired application. The geometry-sensitivity gap and the single-run nature of the results are the main obstacles to acceptance; both are addressable with additional experiments within the existing framework. I would not recommend rejection, as the core approach is sound and the presentation is generally clear."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent computational methodology paper with one genuinely nice result and one real gap. The nice result is the validation study: BO with an expandable design space recovers a known ground-truth parameter set that sits well outside the initial bounds, while the no-expansion baseline stalls at the boundary. That is a clean demonstration that the expansion mechanism does what it claims. The second case, where two distinct interface designs produce nearly the same stress-strain curve but fail by different mechanisms, is also a legitimate finding and is well supported by the damage and crack-path visualizations. Code and data are on GitHub, which is a plus.\n\nThe soft spots are in the transferability claims. The geometry-sensitivity study in the supplement tests five Voronoi patterns but only for three interface property sets whose parameters (strengths 125-200 MPa, separations 2-4 nm, delta_e=15 nm) are far from the optimized solutions reported in the main text. Design 2, for example, has sigma_s=1 MPa and delta_e=80 nm. So the claim that the stress-strain response is insensitive to grain arrangement at the current grain density is only established in a parameter regime the optimization leaves. Since all BO iterations run on geometry 4, the non-unique solution split could in principle be an artifact of one Voronoi realization. That is not fatal to the method, but it needs to be acknowledged and ideally fixed by running the sensitivity study at the optimized parameters.\n\nThere are also a few missing details for reproduction: the stall number S, the GP kernel and hyperparameters, and the number of runs or seeds are not reported. The paper would benefit from repeated BO runs with different initial datasets to show the non-unique result is stable.\n\nOverall, the central argument holds up: within the chosen 2D cohesive-interface model, BO with expandable space finds the target and returns multiple mechanisms. The paper is honest about the 2D/3D limitation and the need for experimental verification. I think this deserves a serious referee. My recommendation: send it to review, but ask the authors to close the geometry-sensitivity gap and report the optimizer settings.","headline":"Solid BO-for-inverse-design paper with a clean validation experiment; the geometry-sensitivity claim overreaches beyond the tested parameter regime but the method deserves peer review.","tokens_in":13338,"tokens_out":2373,"would_cite":true,"duration_ms":21778,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["74P10","74S05","62F15"],"pacs":[],"model":"deepseek-v4-flash","headline":"Bayesian optimization with an expandable design space recovers interface parameters that reproduce a target nonlinear stress-strain curve in nacre-inspired composites, and finds multiple distinct designs that achieve the same curve…","keywords":["Bayesian optimization","inverse design","nacre-inspired composite","interface engineering","traction-separation law","stress-strain curve","design space expansion","non-unique solutions"],"falsifier":"For Design 1 and Design 2 from Table 2, rerun the same finite-element simulation on several new Voronoi grain patterns (rather than the single geometry used for optimization) and measure the spread of the resulting stress-strain curves; if the spread is comparable to the ~0.01 curve-difference values that define a match, the geometry-insensitivity premise holds, whereas a spread an order of magnitude larger would mean the recovered parameters are artifacts of one grain arrangement.","tokens_in":12255,"feed_emoji":"🧱","tokens_out":10125,"duration_ms":75261,"temperature":0.7,"pith_summary":"Given a desired nonlinear stress-strain response, the paper asks whether the soft-interface parameters of a nacre-inspired composite can be recovered computationally. It answers yes: a Bayesian optimization loop, operating on a five-parameter bilinear traction-separation interface law and guided by a pointwise curve-difference metric, finds interface designs whose simulated tensile curves closely match the target. The framework deliberately expands the design space as the search proceeds, so targets that lie outside the initial parameter bounds remain reachable. In the second case study, two top solutions reproduce the same target curve through different mechanisms—one normal-dominated fracture, one shear-dominated fracture. The practical payoff is that designers can choose among several manufacturable interface recipes that meet the same mechanical specification.","feed_headline":"Bayesian search finds distinct designs for one stress-strain target","feed_subtitle":"The search returned separate interface recipes with nearly identical curves but different failure modes.","key_machinery":"The load-bearing machinery is the pairing of a Gaussian-process surrogate with an expected-improvement acquisition function inside a design-space expansion loop. The objective is a symmetric pointwise curve-difference metric $\\Delta_{\\mathrm{TS}}$, defined as the average minimum distance from points on the target curve to the simulated curve plus the average minimum distance in the reverse direction; this replaces the worst-case Hausdorff distance, which the paper argues can miss subtle variations. At regular 'stall' intervals the parameter bounds expand by a factor $\\sqrt[d]{\\gamma/N}$ with $\\gamma = N+1$, so early expansions are large and later ones taper, then contract to the minimal box enclosing the old space and the newly sampled designs. The interface itself is a bilinear traction-separation cohesive law with five independent parameters, simulated by cohesive elements in a two-dimensional Voronoi-grain finite-element model.","core_discovery":"The central claim is that a Bayesian optimization (BO) framework equipped with a controlled, expandable design space solves the inverse interface-design problem for a two-dimensional Voronoi nacre-mimetic composite: given a target stress-strain curve, it recovers the five parameters of the bilinear traction-separation law—the normal and shear interface strengths ($\\sigma_n^o$, $\\sigma_s^o$), the normal and shear damage-initiation separations ($\\delta_n^d$, $\\delta_s^d$), and the damage-evolution separation ($\\delta_e$)—that reproduce the curve in finite-element simulation. Validation against a ground-truth target outside the initial design space shows that space expansion recovers parameters close to the truth, whereas BO without expansion saturates at the boundary and misses the target. A second target, a purely quadratic curve not generated by FEM, yields top solutions that separate into two clusters differing by two orders of magnitude in shear strength; these designs give nearly indistinguishable stress-strain responses while failing by different mechanisms, one normal-dominated and one shear-dominated. The paper thereby claims that inverse design of nonlinear mechanics can be formulated as a tractable optimization problem and that non-uniqueness in interface design is not a nuisance but a resource.","pith_inferences":["A natural extension the paper does not explore is to mask the curve-difference metric so that only a portion of the stress-strain response (e.g., the post-peak softening tail) is targeted, which would let the same loop design interfaces for energy absorption without dictating the elastic regime.","The sharp drop in objective improvement after roughly 45 iterations suggests that a budget-allocation strategy—stopping early and spending the saved FEM evaluations on diverse candidates—could preserve solution diversity at a fraction of the computational cost; the paper notes the plateau but keeps the full 150-iteration budget.","The two design clusters imply a broader design rule for nacre-like composites: low-shear-strength and high-normal-strength interfaces spread damage broadly, while the converse concentrates damage into a few normal cracks; this trade-off could inform selections for damage tolerance versus predictable failure, though the paper does not make that application.","Because the framework treats the simulator as ground truth, its success on the FEM-generated target does not guarantee success on experimental data; a closed-loop version that replaces the simulator with physical measurements would be the decisive next step."],"forward_implications":["If the framework is correct, any specified nonlinear stress-strain curve that lies within the expressive range of the traction-separation law can be treated as a design target, not only extremal objectives like maximum strength or toughness.","Because the design space expands and contracts during the search, a user does not need to know the feasible bounds of interface parameters in advance.","The two recovered solution clusters show that multiple manufacturable interface designs can meet the same mechanical specification, so cost, processability, or preferred failure mode can be used as secondary selection criteria.","The same acquisition-and-expansion loop can be coupled to other forward solvers, such as discrete element models, and to other inverse design objectives beyond interface parameters."],"supporting_citations":[{"why":"Supplies the prior multi-objective Bayesian-optimization treatment of nacre-inspired composites that this work extends to inverse curve matching.","marker":"[20]"},{"why":"Supplies the Voronoi-tessellation geometry generation used to build the two-dimensional grain structures.","marker":"[30]"},{"why":"Supplies the bilinear traction-separation cohesive law that defines the interface design space.","marker":"[32]"},{"why":"Defines the Hausdorff distance that the paper's curve-difference metric improves upon.","marker":"[33]"},{"why":"Supplies the Gaussian-process regression surrogate model used in the optimization.","marker":"[34]"},{"why":"Supplies the expected-improvement acquisition function that guides the search.","marker":"[35]"},{"why":"Provides the weakly-specified-search-space Bayesian optimization scheme the expansion algorithm adapts.","marker":"[43]"},{"why":"Provides a second unbounded-Bayesian-optimization scheme used in the expansion approach.","marker":"[44]"}],"fun_headline_variants":["Same curve, different failure: Bayesian inverse design finds multiple solutions","Bayesian search finds many interface recipes for one stress-strain target","Inverse design with BO: distinct interfaces, identical stress-strain, varied failure","Bayesian optimization expands design space, reveals non-unique interface solutions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the two-dimensional Voronoi finite-element model with a bilinear traction-separation interface law faithfully represents the physical material; if the macroscopic response of the real composite is sensitive to grain arrangement or to interlayer shear that the 2D model cannot capture, the recovered interface parameters will not transfer to a fabricated specimen.","fun_headline_variants_meta":{"raw":{"variants":["Same curve, different failure: Bayesian inverse design finds multiple solutions","Bayesian search finds many interface recipes for one stress-strain target","Inverse design with BO: distinct interfaces, identical stress-strain, varied failure","Bayesian optimization expands design space, reveals non-unique interface solutions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001212,"raw_usage":{"total_tokens":5002,"prompt_tokens":968,"completion_tokens":4034,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":3957}},"tokens_in":584,"tokens_out":4034,"duration_ms":26438,"temperature":1.0,"reasoning_tokens":3957,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:31:01.048752+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For Design 1 and Design 2 from Table 2, rerun the same finite-element simulation on several new Voronoi grain patterns (rather than the single geometry used for optimization) and measure the spread of the resulting stress-strain curves; if the spread is comparable to the ~0.01 curve-difference values that define a match, the geometry-insensitivity premise holds, whereas a spread an order of magnitude larger would mean the recovered parameters are artifacts of one grain arrangement.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the prior multi-objective Bayesian-optimization treatment of nacre-inspired composites that this work extends to inverse curve matching."},{"cited_title":"Rousseau, et al., Sheet nacre growth mechanism: a Voronoi model","cited_arxiv_id":null,"evidence_quote":"Supplies the Voronoi-tessellation geometry generation used to build the two-dimensional grain structures."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the bilinear traction-separation cohesive law that defines the interface design space."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the expected-improvement acquisition function that guides the search."},{"cited_title":"Shahriari, A","cited_arxiv_id":null,"evidence_quote":"Provides a second unbounded-Bayesian-optimization scheme used in the expansion approach."}],"review_version":1}