{"id":"62e3575e-7611-44c3-9310-d63b702cdff0","arxiv_id":"2412.13370","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A partially input convex neural network learns anisotropic hyperelastic response from homogenized stress-strain data, and an evolution strategy inverts it to recover microstructure design parameters and preferred directions.","lead":"This paper trains physics-aware neural networks to learn how composite microstructures stretch, then inverts the trained model to recover design choices such as fiber orientation and volume fraction from a desired stress-strain response. The work targets engineers who want to design materials with a specific mechanical behavior without running expensive simulations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The homogenized-data pipeline filters training deformations under an assumed symmetry class; if the true RVE symmetry is lower (cubic/tetragonal), the surrogate and the inverted parameters are biased, and the paper's checks do not rule this out.","rationale":"The central claim is that the framework can learn anisotropy class/orientation and invert design parameters from homogenized stress-strain data. The reader correctly identified the assumed symmetry class in the data-generation filtering as the weakest point. I agree and add two observations that make the concern concrete. First, the isotropy check in Appendix D is limited: pure-shear comparisons at four extreme parameter sets cannot exclude cubic anisotropy if the shear directions are aligned with cubic axes; a standard discriminator is the Zener ratio or normal loading along different crystal directions. Second, the fiber RVE is treated as transversely isotropic with no symmetry verification, even though a periodic fiber array generally has at most tetragonal symmetry; the Section 6.3 result that two distinct fiber orientations give the same objective value is a red flag that the surrogate may be blind to in-plane orientation. If the filtering assumption fails, the omitted modes are not representable by the surrogate, so the inverse problem can return parameters that are correct only within the assumed isotropic/transversely isotropic subspace. This does not refute the synthetic macroscale results, which are unaffected, and the paper's own caveats show awareness; hence the verdict remains CONDITIONAL rather than REJECT, with the requested verification being a concrete, feasible computation.","tokens_in":23185,"tokens_out":8944,"duration_ms":82576,"concrete_test":"Re-run the RVE homogenization at the four extreme parameter sets of Appendix D without invariant-space filtering, and compare responses to deformation states that distinguish cubic/tetragonal from isotropic/transversely isotropic symmetry: e.g., for the inclusion RVE, extract the small-strain tangent and check the Zener ratio C11-C12-2C44; for the fiber RVE, compare in-plane shear moduli for loadings at 0° and 45° to the fiber lattice. If the Zener ratio deviates from 1 or the 0°/45° shear responses differ by more than a few percent, retrain the pICNN on an unfiltered dataset and re-solve the Section 6.3 orientation inversion; if the two previously distinct converged orientations now give different objective values, the filtering bias is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is the symmetry class used to filter homogenized training states. In Section 5.2 (footnote 2), deformation gradients are sorted and pruned in invariant space assuming the spherical-inclusion RVE is isotropic and the fiber RVE is transversely isotropic. The pICNN is then trained on this filtered dataset and the inverse problem is solved on targets generated from the same filtered distribution. If the true effective symmetry is lower (cubic for a single inclusion in a cubic periodic cell; tetragonal for a periodic square array of fibers), the omitted deformation modes are exactly the ones that would reveal the symmetry breaking. The paper's own check (Appendix D, Section 6.2.1) tests only four extreme parameter combinations and only pure-shear responses, and cites [83] noting that single-inclusion RVEs 'often show effective cubic properties in the small strain context.' No symmetry verification is reported for the fiber RVE. A concrete symptom appears in Section 6.3: five FE inverse runs converge to two different fiber orientations with nearly identical maximum von Mises stress. The authors attribute this to design-space symmetry, but it is equally consistent with a transversely-isotropically filtered surrogate that is insensitive to in-plane lattice orientation. If the symmetry assumption is wrong, the homogenized-data demonstrations (Figures 13, 14) and the FE inversion (Figure 16) do not validate the abstract's claim about recovering correct microstructure parameters from homogenized data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a physics-augmented neural-network framework for the inverse design of anisotropic hyperelastic microstructures. A partially input-convex neural network (pICNN) maps deformation invariants and design parameters to a polyconvex strain-energy potential, from which stresses and tangents are obtained by automatic differentiation. The anisotropy class and preferred directions are learned during training through regularized coefficients α1, α2 and a rotation-parameterized structural tensor. After training on stress-strain data from synthetic constitutive laws and from FE homogenization of inclusion and fiber RVEs, the design parameters and preferred directions are recovered with CMA-ES. The trained fiber-RVE surrogate is also embedded in an FE solver to invert fiber orientation from a target maximum von Mises stress. Results are presented for isotropic, transversely isotropic, and orthotropic cases, including extrapolation to unseen parameters and preferred directions.","tokens_in":23452,"tokens_out":8773,"duration_ms":86298,"significance":"If the claims hold, the framework is a valuable step toward a complete forward-inverse pipeline for anisotropic microstructure design: it combines automatic anisotropy-class/orientation detection, a polyconvex surrogate construction, and parameter/orientation inversion, and it demonstrates extrapolation to preferred directions not seen in training. The architectural choices—invariant-based inputs, monotone convex activations, stress normalization, and derivative-free inverse optimization—are well motivated and build sensibly on prior work. The central limitation is that the evidence is largely qualitative: the paper reports no quantitative error metrics, and the homogenized-data demonstrations rest on a symmetry-assumption check that is incomplete and partly outside the sampled parameter range. These issues are addressable, and with quantitative reporting and symmetry verification the contribution would be significant for the computational mechanics community.","major_comments":[{"comment":"The central inverse-design claims are supported only by visual inspection of stress-strain overlays and parameter-trajectory plots; no quantitative error metric (e.g., relative error of recovered design parameters, stress residuals, or an accepted tolerance for \"correct\" recovery) is reported anywhere. The reader cannot judge whether the recovered parameters in Figures 6c, 8e, 10e, 13d, and 14e are within, say, 1% or 20% of the truth. Please report, for every inverse problem, the relative L2 error in the recovered design parameters and the forward stress error on the target dataset, and state the convergence criterion used for the CMA-ES runs.","section":"Section 6, Figures 6-14"},{"comment":"The homogenized datasets are pruned in invariant space under an assumed symmetry class (isotropic for the single-inclusion RVE, transversely isotropic for the fiber RVE), but the symmetry verification is incomplete and partly outside the sampled parameter range. The isotropy check in Appendix D uses R=0.1 and R=0.5, which for the unit-cell RVE correspond to volume fractions of about 0.004 and 0.52, outside the training range φ∈[0.15,0.35]; no symmetry check is reported for the fiber RVE. Section 6.2.1 itself cites [83] noting that single-inclusion RVEs often show effective cubic properties in the small-strain context, so the possibility that the true effective symmetry is lower than assumed cannot be dismissed. Because the invariant-space filtering removes the deformation modes that would reveal lower symmetry, the surrogate and any inverse-design conclusion drawn from it would be biased. Please verify the effective symmetry of both RVEs over the full sampled parameter range using deformation modes that discriminate between the assumed and lower symmetry classes, and show that the filtering does not discard symmetry-breaking information.","section":"Section 5.2 and Appendix D"},{"comment":"The five FE inverse runs converge to two different fiber orientations with almost identical maximum von Mises stress, and the authors attribute this to symmetry in the design space. The same observation would result from a surrogate that is insensitive to in-plane lattice orientation because the fiber-RVE training data were filtered under the transverse-isotropy assumption; the current experiment cannot distinguish these explanations. Figure 16a shows only the stress evolution, not the orientation evolution, so the claimed convergence to two distinct orientations is not actually visible. Please compare the surrogate's predicted response for orientations that are not equivalent under the assumed symmetry group (e.g., an in-plane rotation of the fiber lattice) against direct FE homogenization, and report whether the surrogate distinguishes them.","section":"Section 6.3, Figure 16"}],"minor_comments":[{"comment":"The displayed set for \\bar{I}_{iso} contains an extra closing brace; this appears to be a typo.","section":"Equation (27a)"},{"comment":"The text contains several typos that should be corrected: \"V on Mises\" should be \"von Mises\", \"Covariant Matrix Adaptation\" should be \"Covariance Matrix Adaptation\", and \"modulii\" should be \"moduli\"; in Appendix C, \"nice-dimensional\" should be \"nine-dimensional\".","section":"Sections 4 and 6.3"},{"comment":"The convex, monotone activation function Θc is never specified concretely (e.g., softplus); specifying the exact activation and initialization would improve reproducibility.","section":"Section 3"},{"comment":"Figure 22 does not label its axes or clarify which stress component and loading path are shown; please add axis labels and a description of the deformation mode, and align the tested R and μ1/μ2 values with the sampled parameter ranges.","section":"Section 6.2.1 and Appendix D"},{"comment":"The notation \"s.t. arg min_θ ...\" is unusual because the forward problem is solved before the inverse problem rather than being a constraint in the optimization; please clarify that this is a sequential two-stage procedure.","section":"Section 4, Eq. (32)"},{"comment":"The statement that code will be made available only after acceptance limits reproducibility of the reported experiments; please provide at least the datasets and a reference implementation in a public repository or in the supplementary material.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of a computational mechanics journal and the core methodology is sound in principle. The main risk is that the current version's claims outpace the quantitative evidence, particularly for the homogenized-data demonstrations where the symmetry assumption is load-bearing. I would be willing to re-review a version that adds error metrics, strengthens the symmetry verification, and addresses the FE-inversion ambiguity."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real method paper, not a toy demo. The new contribution is the closed loop: train a pICNN surrogate with an Lp-sparsity penalty that automatically detects anisotropy class and preferred directions, then invert design parameters and orientation with CMA-ES, including inside a finite element solver. The synthetic macroscale tests hold together: class and orientation recovery are convincing, and the inverse problems find parameters outside the training set. The FE beam example is a reasonable integration.\n\nThe soft spots are in the homogenized-data validation. The biggest is the symmetry assumption in data generation. The paper filters deformation states assuming the single-inclusion RVE is isotropic and the fiber RVE is transversely isotropic. Isotropy for the inclusion RVE is only checked at four extreme parameter combinations under pure shear, and the authors themselves note that single-inclusion RVEs often show effective cubic behavior in the small-strain limit. If the true effective symmetry is lower than assumed, the omitted deformation modes are exactly the ones that would reveal it, and both the surrogate and any inverted parameters would be biased. The fiber RVE gets no symmetry verification at all. The stress-test note is right to make this the central concern.\n\nThe second issue is that validation is entirely qualitative. Fits are eyeballed; there are no error norms, no tolerances, no convergence measures for the inverse optimizations. Code is withheld, and tuning constants like epsilon and gamma are unreported. That makes it hard to separate the method's merits from careful hand-tuning.\n\nThe paper deserves a serious referee. The forward framework is a reasonable assembly of prior building blocks, the inverse formulation is a genuine extension, and the authors are transparent about polyconvexity loss and the alternate formulations they tried. What a revision needs is (a) a proper symmetry check for both RVEs, or a demonstration that the results are insensitive to the assumption; (b) quantitative errors and hyperparameters; and (c) available code. This is for people working on data-driven constitutive models and microstructure design. I would send it to peer review with the expectation of substantial revision before acceptance.","headline":"A solid forward-inverse pipeline for anisotropic microstructure design, with the main caveat being an unverified symmetry assumption in the homogenized-data validation and no quantitative error reporting.","tokens_in":23994,"tokens_out":3182,"would_cite":true,"duration_ms":28853,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["74B20","74Q05","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A physics-augmented neural network trained on macroscopic stress-strain data learns the anisotropy class and preferred directions of a composite, and then solves the inverse design problem of recovering the microstructure parameters that…","keywords":["inverse design","anisotropic hyperelasticity","polyconvex neural networks","partially input convex neural networks","computational homogenization","microstructure design","CMA-ES","invariant-based constitutive modeling"],"falsifier":"Take the single-spherical-inclusion RVE at parameter pairs inside the sampled range rather than only the four corners, apply pure shear along several material directions at small strain, and compute the full anisotropic tangent; if the directional shear responses separate measurably, the isotropy-filtered training set has dropped relevant deformation modes and the inverse predictions from the surrogate would be systematically wrong.","tokens_in":22999,"feed_emoji":"🎯","tokens_out":7074,"duration_ms":60910,"temperature":0.7,"pith_summary":"This paper tries to establish that inverse design of anisotropic hyperelastic microstructures can be solved in two linked steps using one physics-augmented neural representation. The forward step trains a partially input convex neural network to represent the strain energy as a polyconvex function of invariants of the right Cauchy-Green tensor, with trainable coefficients that reveal whether the material is isotropic, transversely isotropic, or orthotropic and what the preferred directions are. The inverse step then uses an evolution strategy to find the material and geometric design parameters whose predicted stress-strain response matches a target, including targets whose preferred direction was never seen during training. If the claim holds, a designer could replace costly multiscale simulation-based searches with a surrogate that simultaneously supplies the constitutive law, the symmetry classification, and the parameter inversion.","feed_headline":"Neural network designs anisotropic microstructures from stress data","feed_subtitle":"A physics-augmented surrogate finds material symmetry and the design parameters for a target response.","key_machinery":"The central object is the partially Input Convex Neural Network (pICNN) strain-energy representation $\\Psi(\\mathbf{I},\\mathbf{D})$, convex and monotonically non-decreasing in the invariants $\\mathbf{I}$ of $\\mathbf{C}=\\mathbf{F}^T\\mathbf{F}$ and arbitrary in the design parameters $\\mathbf{D}$. Anisotropy enters through structure tensors $\\mathbf{N}_i(R)=\\mathbf{n}_i\\otimes\\mathbf{n}_i$ built from trainable rotations, and through sigmoid-gated coefficients $\\alpha_1,\\alpha_2$ that an $L^p$ penalty drives to zero unless the data require them. Polyconvexity is enforced by the convex monotone network; volumetric growth and normalization terms ensure coercivity, zero energy, and zero stress at the undeformed state. Stresses are obtained by differentiating the energy with respect to $\\mathbf{C}$, and the inverse problem is solved with CMA-ES over the design variables plus the orientation parameters. This machinery is what lets a single surrogate perform symmetry classification, forward prediction, and inverse design.","core_discovery":"The paper's central claim is that an invariant-based partially input convex neural network can serve as a complete forward-inverse pipeline for anisotropic hyperelastic microstructures: trained on homogenized stress-strain data, it learns a polyconvex strain energy that is convex in the deformation invariants and arbitrary in the design parameters, identifies the anisotropy class through sparse regularization of two anisotropy coefficients, recovers the preferred direction(s) through a rotation parametrization, and then—via CMA-ES—inverts both design parameters and orientation for target stress-strain data, including data with a preferred direction different from the training set. The framework is demonstrated on synthetic isotropic, transversely isotropic, and orthotropic datasets, on two homogenized microstructures (a single spherical inclusion and aligned fibers), and inside a finite element beam optimization that finds the fiber orientation minimizing peak von Mises stress. For homogenized data where polyconvexity may be lost, the paper proposes an alternate formulation that keeps the tangent modulus positive definite while relaxing polyconvexity through the stress-normalization term.","pith_inferences":["If orientation inversion works as reported, the surrogate implicitly learns a rotation-equivariant material map; a natural stress test is to interpolate between two orientations and check that inverted parameters vary smoothly.","The invariant-space sorting used to build the homogenized training sets is a general data-reduction idea that could benefit other surrogate constitutive models, not only pICNNs.","The same forward-inverse split might extend to inelastic behavior or higher-order anisotropy, but the structure tensors and invariant set would need generalizing beyond the transverse-isotropy and orthotropy cases treated here.","The ability to invert orientation from stress-strain data in a finite element context suggests a practical design loop for additively manufactured composites: specify a target peak-stress response and read off the fiber angle directly."],"forward_implications":["A surrogate trained on one set of preferred directions can invert target stress-strain data with a different orientation, recovering both the design parameters and the new orientation.","The anisotropy class is discovered rather than assumed: the anisotropy coefficients are driven to zero unless the stress data require them.","The learned polyconvex energy supplies stresses and a positive semi-definite tangent modulus, making it usable inside a finite element solver for structural inverse problems.","For homogenized data that break polyconvexity, the framework offers an alternate stress-normalization formulation that preserves positive definiteness of the tangent modulus.","The inverse procedure recovers material parameters outside the training range and solves a beam-level fiber-orientation optimization for minimum maximum von Mises stress."],"supporting_citations":[{"why":"Supplies the tensor-basis invariant formulation and the idea of learning anisotropy class and preferred directions from data.","marker":"[48]"},{"why":"Introduces parametrized polyconvex hyperelasticity with physics-augmented neural networks, the pICNN construction that the paper adapts for design parameters.","marker":"[50]"},{"why":"Shows how polyconvex anisotropic hyperelasticity is enforced with input convex neural networks, the constraint scheme used here.","marker":"[47]"},{"why":"Provides the stress-normalization technique that the paper adapts to keep the free energy stress-free at the undeformed state.","marker":"[46]"},{"why":"Defines partially Input Convex Neural Networks, the architecture class the framework is built on.","marker":"[61]"},{"why":"Is the CMA-ES evolutionary strategy used to solve the inverse design problem.","marker":"[70]"},{"why":"Supplies the invariant-space sampling of deformation states and a neural-network-based multiscale framework that the data generation follows.","marker":"[77]"}],"fun_headline_variants":["Physics-augmented neural net inverts anisotropic microstructures","AI learns anisotropy and inverts design from stress data","Neural network solves inverse design of anisotropic composites","Convex neural net recovers microstructure from stress-strain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The homogenized RVE data are generated under an assumed symmetry class—isotropic for the spherical-inclusion RVE and transversely isotropic for the fiber RVE—so if a real microstructure's effective symmetry is lower than assumed, the omitted deformation modes would bias both the surrogate and the inverse design.","fun_headline_variants_meta":{"raw":{"variants":["Physics-augmented neural net inverts anisotropic microstructures","AI learns anisotropy and inverts design from stress data","Neural network solves inverse design of anisotropic composites","Convex neural net recovers microstructure from stress-strain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00017,"raw_usage":{"total_tokens":1319,"prompt_tokens":1044,"completion_tokens":275,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":660,"completion_tokens_details":{"reasoning_tokens":211}},"tokens_in":660,"tokens_out":275,"duration_ms":3501,"temperature":1.0,"reasoning_tokens":211,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:12:23.672548+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the single-spherical-inclusion RVE at parameter pairs inside the sampled range rather than only the four corners, apply pure shear along several material directions at small strain, and compute the full anisotropic tangent; if the directional shear responses separate measurably, the isotropy-filtered training set has dropped relevant deformation modes and the inverse predictions from the surrogate would be systematically wrong.","supporting_citations":[{"cited_title":"Learning hyperelastic anisotropy from data via a tensor basis neural network","cited_arxiv_id":null,"evidence_quote":"Supplies the tensor-basis invariant formulation and the idea of learning anisotropy class and preferred directions from data."},{"cited_title":"Klein, Fabian J","cited_arxiv_id":null,"evidence_quote":"Introduces parametrized polyconvex hyperelasticity with physics-augmented neural networks, the pICNN construction that the paper adapts for design parameters."},{"cited_title":"Neural networks meet hyperelasticity: A guide to enforcing physics","cited_arxiv_id":null,"evidence_quote":"Provides the stress-normalization technique that the paper adapts to keep the free energy stress-free at the undeformed state."},{"cited_title":"M ¨uller, and Petros Koumoutsakos","cited_arxiv_id":null,"evidence_quote":"Is the CMA-ES evolutionary strategy used to solve the inverse design problem."},{"cited_title":"Neural network-based multiscale modeling of finite strain magneto-elasticity with relaxed convexity criteria","cited_arxiv_id":null,"evidence_quote":"Supplies the invariant-space sampling of deformation states and a neural-network-based multiscale framework that the data generation follows."}],"review_version":1}