{"id":"e6318202-431b-47bc-ba75-b9c3d5527dce","arxiv_id":"2607.15049","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"PANO and CANO learn a single-pass map from full-field displacement and force data to the strain-energy density of an incompressible hyperelastic material.","lead":"Two neural operators, PANO and CANO, take measured displacement fields and reaction forces from a stretched specimen and return the material's strain-energy function in a single forward pass. If the approach extends beyond the cubic material family used in training, it could replace slow iterative calibration in routine material characterization.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Trained and tested only on the six-parameter separable cubic family (Eqs. 23–24); since CANO's trunk is exactly that basis, out-of-family constitutive functions are never demonstrated, so the 'discovery' claim reduces to coefficient identification within a pre-chosen model family.","rationale":"The reader identified essentially the same load-bearing weakness: the method is trained and tested only within the six-parameter separable cubic family, with the CANO trunk exactly matching that family's features. This is the central gap relative to the 'discovery' framing, and no amount of within-family testing or identifiability analysis restricted to that class can close it. The paper is otherwise careful and the within-family results appear credible, but the headline claim of mapping into an infinite-dimensional space and discovering constitutive functions is not yet supported. A single out-of-family experiment of the type described would directly test whether the operator generalizes beyond the training parameterization. Since the reader's conditional verdict already reflects this concern, no verdict change is needed.","tokens_in":30490,"tokens_out":5099,"duration_ms":59642,"concrete_test":"Generate an out-of-family test set by simulating the same boundary value problem (Appendix A) for a non-separable polyconvex strain-energy function, e.g., W = C10 I1* + C01 I2* + C11 I1* I2*, where the coupling term is present in the Taylor expansion (Eq. C.2) but omitted from Eq. (24). Feed the resulting displacement and reaction-force data to the already-trained PANO and CANO and compare the predicted strain-energy density against ground truth over the invariant sampling domain of Section 2.5. If the operators fail to reproduce the coupling term—CANO structurally cannot—the 'constitutive discovery' claim is not supported. If PANO nonetheless approximates the out-of-family function well, the claim is strengthened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that the operators map displacement and reaction-force data to the infinite-dimensional space W of admissible strain-energy functions and thereby enable constitutive-model discovery (Abstract; Sections 2.3, 3.3–3.4)—is only tested on functions of the six-parameter separable cubic form of Eq. (24). Training data are generated exclusively from this family (Section 2.5), and the CANO trunk features in Eq. (23) are exactly the six monomials of that family. Thus CANO cannot represent any strain-energy function outside this parametric family by construction, and PANO, while architecturally more flexible, is trained only on labels from this family and is never evaluated on out-of-family data. The empirical results are therefore consistent with a weaker claim: the operators identify six coefficients of a known separable cubic from data. This matters because the paper's novelty over prior parameter-calibration work rests on reducing a priori assumptions about the functional form of the strain-energy density (Section 1). The identifiability analysis in Section 4.2 is explicitly restricted to the six-parameter class, assumes plane stress and exact displacement data, and does not address the general W. The manuscript itself lists enriching the constitutive model space as future work (Section 5), confirming that out-of-family generalization is untested. Without such evidence, the 'discovery' claim is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two neural operator architectures, PANO and CANO, that map full-field displacement measurements and net reaction forces directly to the hyperelastic strain-energy density function, eliminating the need to solve iterative inverse calibration problems. The operators combine a DeepONet-style branch net, which encodes displacement fields via Laplacian eigenfunctions and force data, with a physics-constrained trunk net based on PANN or CANN architectures. Training data are generated from 3000 finite element simulations of a thin plate with a central hole, using the six-parameter separable cubic strain-energy family of Eq. (24). The authors test on unseen samples, noisy displacement data, missing spatial data, different discretizations, and scaled geometries, and they analyze identifiability within the six-parameter CANO class via singular values of the weak-form equilibrium system. The central claim is that, after training, a single forward pass yields a physically admissible strain-energy density from a new experimental dataset.","tokens_in":30858,"tokens_out":6017,"duration_ms":62531,"significance":"If the claimed generalization holds, the framework is a useful step toward amortized inverse material characterization: it avoids per-material retraining and iterative optimization, builds physical admissibility (polyconvexity, objectivity, stress-free reference state) directly into the output, and introduces a discretization-robust encoding via Laplacian eigenfunctions. The simulated-data study is carefully designed within its family, and the SVD identifiability analysis in Section 4.2 is a valuable diagnostic. However, the demonstrated scope is narrower than the title and abstract suggest: all training and test materials lie in a six-parameter polynomial family, and CANO's trunk uses exactly that basis. The evidence therefore supports coefficient identification within a preselected model family rather than discovery of general constitutive functions in the infinite-dimensional space W of Eq. (2).","major_comments":[{"comment":"All training and test labels are generated from the six-parameter separable cubic family in Eq. (24), and the CANO trunk in Eq. (23) is exactly the monomial basis of that family. The learned operator therefore cannot represent any strain-energy function outside this family, and the empirical results demonstrate six-coefficient identification within a known basis rather than discovery of a general function in the infinite-dimensional space W of Eq. (2). This undercuts the abstract's \"infinite-dimensional output space\" claim and the Introduction's statement that the framework \"reduces a priori assumptions about material behavior\" and \"can predict a variety of functional forms\" (Section 1). PANO is architecturally more flexible but is trained and tested only on this same family, so it provides no out-of-family evidence. Either add out-of-family experiments or rephrase the claims to in-famil","section":"§2.3.2, §2.5, Eq. (24); Abstract"},{"comment":"The quantitative evaluation is limited to histograms and selected best/median/worst curves; no numerical MSE values, quantiles, or confidence intervals are reported anywhere. This makes it impossible to assess the frequency of outliers and the magnitude of the \"excellent agreement\" claimed in Sections 3.1–3.2. For instance, the PANO worst-case MSE is dismissed as an \"exceptional outlier\" without reporting its rank or frequency. Please report numeric error statistics (e.g., median, 95th percentile, max MSE, or relative L2 errors) over the full test set, and if feasible confidence intervals across repeated training runs.","section":"§3.1–3.3"},{"comment":"The noise-robustness study perturbs only the displacement fields; the scalar reaction force R_data, which enters the branch net and provides the absolute scale through Eq. (21), is left unchanged. Since the abstract claims robustness to \"noisy data\" generally and the load-cell signal is an equally important measurement, the current experiment does not fully support the claim. Please add tests with reaction-force noise (and combined noise), or explicitly restrict the robustness claim to displacement noise.","section":"§3.3; Eq. (21)"},{"comment":"The identifiability analysis is an honest and useful check, but its scope is narrow: it assumes plane stress, exact displacement data, and that the true model lies in CANO's six-parameter feature space. The reported σ_min/σ_max ≳ 0.15 is only for \"representative datasets,\" and the tolerance τ is not specified. Because the conclusion \"no non-identifiable directions\" is used to argue that the framework regularizes the inverse problem, the analysis should be documented across all 3000 simulations and for noisy or missing-data cases; otherwise the conclusion should be stated more cautiously.","section":"§4.2, Eqs. (32)–(38)"}],"minor_comments":[{"comment":"In the parameter-scaling sentence, \"C01 = ¯C10/5\" appears to be a typo; it should be \"C01 = ¯C01/5\".","section":"§2.5"},{"comment":"The scaling by ∥R_data∥ is used both in normalizing the branch-net input and in rescaling the final strain energy. Please clarify this two-step use in the notation to avoid confusion.","section":"Eq. (21)"},{"comment":"The histograms in Figs. 5a and 6a lack axis labels in the captions; the x-axis presumably is MSE, but this should be stated explicitly.","section":"Figures 5–6"},{"comment":"The geometry-scaling test is performed for a single representative material model. This demonstration is suggestive, but a statement about the robustness across the full test set would strengthen the claim of geometry invariance.","section":"§3.4"},{"comment":"The determinant expression 1/det(I+∇u_t(X)) is written with a scalar denominator; the notation would be clearer if the inverse of the 2×2 block and the scalar determinant were distinguished explicitly.","section":"§4.2, Eq. (27)"}],"recommendation":"major_revision","confidential_remarks":"The paper's main weakness is the mismatch between the 'discovery' framing and the in-family evidence. I would support publication after either (a) adding out-of-family experiments with broader constitutive model classes or (b) reframing the contribution as amortized in-family parameter identification. The SVD analysis is a strength but does not resolve this gap. The manuscript is otherwise careful and reproducible in its simulated-data design."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is that a trained DeepONet-style operator maps full-field displacement and reaction-force data directly to a strain-energy function, so a new material can be characterized in one forward pass instead of an inverse optimization. That is a real workflow change for full-field identification, and the paper implements it carefully. The Laplacian-eigenfunction encoding gives discretization independence and noise robustness, the PANN/CANN-inspired trunks enforce physical admissibility, and the tests on unseen data, noisy data, missing data, and rescaled geometries are mutually consistent. Within its chosen model family, the within-class result holds up: PANO and CANO recover the six coefficients of the separable cubic model accurately, and CANO is the more stable of the two.\n\nWhere the paper is soft is the gap between what it claims and what it tests. The training distribution is exclusively the six-parameter separable cubic of Eq. (24), and the CANO trunk features in Eq. (23) are exactly the six monomials of that family. CANO cannot represent any strain-energy function outside that parametric family by construction, and PANO, while architecturally more flexible, is trained only on labels from this family and is never evaluated on out-of-family data. So the demonstrated result is coefficient identification within a pre-chosen model family, not discovery of general constitutive functions. The abstract's language about mapping to the infinite-dimensional space W is not supported by the evidence presented. The SVD identifiability analysis in Section 4.2 is explicitly restricted to the six-parameter class, plane stress, and exact displacement data; it is a useful check but does not address the general problem. There are also no quantitative MSE values in the text, only histograms and representative curves, and the code and data are promised but not yet released, so the exact numerical results cannot be independently verified.\n\nNone of this is fatal. The paper is honest about several limitations and lists richer constitutive spaces as future work. The within-class engineering is solid, the architectures are principled, and the robustness experiments are a real strength. What it needs before the discovery claim can be taken at face value is out-of-family validation: test on a model with coupling terms, or a model outside the separable cubic class, or an experimentally measured dataset. It also needs quantitative error numbers and the promised code and data.\n\nThis paper is for people working on neural operators for material identification, and it deserves a serious referee. I would send it to peer review, with the clear request that the authors either narrow the claim to parametric identification or add out-of-family evidence. I would also cite it if I were working in this area, because it is a clean baseline for the single-forward-pass approach.","headline":"A useful, mostly credible operator-learning framework for parametric hyperelastic identification, but the 'discovery' claim is only demonstrated within the six-parameter separable cubic family that the CANO trunk encodes by construction.","tokens_in":31303,"tokens_out":1805,"would_cite":true,"duration_ms":23473,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A pair of neural operators, PANO and CANO, learns to map full-field displacement data and reaction forces directly to a hyperelastic material's strain-energy density function, returning the model in a single forward pass without solving an","keywords":["neural operators","constitutive model discovery","hyperelasticity","strain-energy density","inverse problems","physics-augmented neural networks","Laplacian eigenfunction encoding","material characterization"],"falsifier":"Generate synthetic displacement and reaction data from a strain-energy function outside the six-parameter separable cubic family—e.g., one with a coupling term C11(¯I1−3)(¯I2−3^(3/2))—and feed it to the trained operator. If the predicted energy deviates substantially from the truth over the sampled invariant domain, the claim that the operator maps into the infinite-dimensional space is refuted.","tokens_in":30397,"feed_emoji":"⚡","tokens_out":6169,"duration_ms":67547,"temperature":0.7,"pith_summary":"The paper proposes two neural operator architectures, PANO and CANO, that learn a map from measurable quantities—surface displacement fields and net reaction forces—to the infinite-dimensional space of admissible hyperelastic strain-energy density functions. In contrast to standard material characterization, which solves an optimization problem each time, a trained operator returns the material model in one forward pass. The output is constrained to be physically admissible (objective, stress-free, polyconvex, etc.) by construction. The authors demonstrate robustness to noise, missing data, different discretizations, and specimen size, and present a numerical identifiability analysis showing that within the six-parameter cubic family of strain-energy functions the standardized experiment yields no non-identifiable parameter directions. The load-bearing claim is that the operator captures the physics-to-data map accurately enough for near-instantaneous material characterization without retraining.","feed_headline":"Neural operators map strain data to a material law in one forward pass","feed_subtitle":"Skip the optimization loop: trained operators recover hyperelastic strain-energy densities from noisy, partial, or resized specimens.","key_machinery":"The load-bearing construction is the branch-and-trunk operator architecture. The branch net takes the flattened coefficients of the displacement field's Laplacian-eigenfunction expansion together with the normalized reaction-force history, and outputs a non-negative latent vector. The trunk net maps the isochoric invariants to a latent feature vector: in PANO these features are learned convex monotone functions (input-convex networks); in CANO they are the six fixed monomials of a cubic generalized Mooney–Rivlin expansion. The strain-energy density is the inner product of the two latent vectors scaled by the reaction-force norm. The Laplacian-eigenfunction projection in the branch's first la","core_discovery":"The central claim is that the inverse problem of constitutive model discovery can be reformulated as supervised operator learning: a neural operator approximates the conditional inverse operator that maps an observation tuple—surface displacements and net reaction force—to the physically admissible strain-energy density function. The operator is trained on thousands of simulated experiments for strain-energy functions sampled from a six-parameter cubic family. Its branch net encodes the displacement field through Laplacian eigenfunctions, which makes the prediction independent of the measurement grid and robust to noise, while its trunk net outputs latent features designed to enforce objecti","pith_inferences":["Editorial inference: The current demonstration is essentially coefficient identification within a known six-parameter family, so the more ambitious 'discovery' claim needs a test on out-of-family functions; if the operator were trained on a mixture of model forms, it might interpolate across families in ways the fixed-feature CANO cannot.","Editorial inference: The Laplacian-encoding trick generalizes beyond this setting: representing input fields in the eigenbasis of the problem domain regularizes any inverse operator learning task, potentially reducing the data needed for other ill-posed identification problems.","Editorial inference: Conditioning the operator on the loading protocol or geometry descriptor (rather than a fixed standardized experiment) would be a natural next step and is implicitly supported by the scaling-invariance results, since the operator must currently be retrained for fundamentally different boundary conditions.","Editorial inference: The identifiability matrix could be used as a cheap a priori experiment-design tool: maximizing its smallest singular value over candidate loading protocols would give an optimization criterion for designing tests that most reliably separate material parameters."],"forward_implications":["Trained operators characterize a new material without any optimization loop, so material identification becomes near-instantaneous and could be run in real time during experiments.","Because the branch input is a spectral encoding rather than raw grid values, the same operator works for different measurement resolutions, partial fields, and specimens of different side lengths and thicknesses, subject to the stated scaling relations.","The predicted models are admissible by construction, so they can be plugged directly into finite element simulations without additional regularization or stability checks.","The SVD-based identifiability analysis provides a quantitative way to judge whether a given experiment, material class, and set of measured quantities can uniquely determine material parameters, which can guide experimental design.","The framework reframes simulation databases as reusable assets: data generated for one-off calibration can instead train operators that serve many future inverse problems."],"fun_headline_variants":["Neural operators get material laws in one forward pass","Single-pass neural operators map strain to constitutive models","Skip optimization: neural operators learn material laws directly","Physics-constrained neural operators infer hyperelastic laws from data","From noisy strain fields to material model with one neural pass"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The operator is trained and tested only on strain-energy functions of the six-parameter separable cubic form, and the CANO trunk features are exactly those six monomials; so the demonstrated accuracy establishes recovery of six coefficients within a pre-chosen model family, not discovery of a general constitutive function.","fun_headline_variants_meta":{"raw":{"variants":["Neural operators get material laws in one forward pass","Single-pass neural operators map strain to constitutive models","Skip optimization: neural operators learn material laws directly","Physics-constrained neural operators infer hyperelastic laws from data","From noisy strain fields to material model with one neural pass"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000511,"raw_usage":{"total_tokens":2317,"prompt_tokens":733,"completion_tokens":1584,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":477,"completion_tokens_details":{"reasoning_tokens":1508}},"tokens_in":477,"tokens_out":1584,"duration_ms":14007,"temperature":1.0,"reasoning_tokens":1508,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T00:17:41.353995+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate synthetic displacement and reaction data from a strain-energy function outside the six-parameter separable cubic family—e.g., one with a coupling term C11(¯I1−3)(¯I2−3^(3/2))—and feed it to the trained operator. If the predicted energy deviates substantially from the truth over the sampled invariant domain, the claim that the operator maps into the infinite-dimensional space is refuted.","supporting_citations":[],"review_version":1}