{"id":"1fad83fd-9a66-4065-b833-85dcbb94ca75","arxiv_id":"2507.12694","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A physics-informed TCN surrogate is trained on synthetic thermo-viscoelastic data with Mullins damage and reportedly generalizes to unseen stretch, rate, and temperature conditions, though validation stays within the same generative model.","lead":"This paper trains a Temporal Convolutional Network to reproduce stress softening and viscoelastic behavior of soft materials under different rates and temperatures. The authors claim the model generalizes to unseen conditions and can be used in finite element simulations, but the training data are synthetic, generated from a known constitutive model rather than experiments.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Synthetic-data circularity: all training, held-out 'unseen,' and FE-validation responses come from the same Yeoh+Ogden-Roxburgh+Maxwell generator, so the claimed generalization to real soft materials is unsupported; the abstract's 'high-fidelity experimental data' is contradicted by §4.1.","rationale":"I focused on the strongest claim: that a TCN trained on a small set of cyclic tests can extrapolate to unseen temperatures, rates, and stretches, and can be deployed in Abaqus for real soft materials. For that claim to hold, the training data must be representative of the target material class. The paper's own Section 4.1 shows the opposite: every training sample is generated from a closed-form Yeoh equilibrium model, an Ogden–Roxburgh damage law (Eq. 32), a four-branch Maxwell evolution, and a WLF shift. The word 'experimental' in the abstract and introduction is contradicted in the methods. Since the held-out 'unseen' conditions (10°C, 500%/min, 200% stretch) and the FE reference are also produced by this same generator, the reported R²>0.97 measures how well a TCN can approximate this specific parametric map, not its ability to generalize across real thermomechanical mechanisms. This is the single most load-bearing concern because every downstream claim—thermodynamic consistency in extrapolation, VUMAT usability, practical predictive value—depends on it. I also noted that the thermodynamic constraint is a penalty-based soft constraint (Section 3.5.1), not a hard guarantee, but that is secondary, since the synthetic targets already satisfy dissipation non-negativity and the penalty merely encourages the surrogate to preserve that property. The most direct way to settle the concern is to rerun the exact same training/held-out protocol on real experimental data; if performance persists, the surrogate is genuinely predictive, and if not, the generalization is an artifact. I see no independent support (e.g., reproducible code, data release) that would offset the missing experimental validation. Thus my read does not change the reader's REJECT verdict.","tokens_in":31339,"tokens_out":7824,"duration_ms":85616,"concrete_test":"Retrain the TCN on real multi-cycle uniaxial experimental data from a shape memory polymer (e.g., the material characterized in [48]) at the same training cases: 0°C/200%/min, 25°C/20%/min, 55°C/200%/min, first two cycles up to 150% strain. Then evaluate the held-out conditions claimed in Case 2: 10°C/500%/min, 200% strain, and the third cycle. If the held-out pointwise stress error or R² degrades materially relative to the synthetic-data results (e.g., R² falls below 0.97 or mean error doubles), the generalization claim was an artifact of generator-matching; if performance is preserved, the circularity objection is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—extrapolative prediction for real soft materials (unseen 10°C, 500%/min, 200% stretch) with thermodynamic consistency—depends entirely on the representativeness of the Section 4.1 data generator. Section 4.1 constructs the training set from a Yeoh isochoric energy, an Ogden–Roxburgh damage function (Eq. 32), a four-branch generalized Maxwell evolution (Eq. 33), and a WLF reduced-time shift. No experimental measurement appears in the paper; the abstract's phrase 'high-fidelity experimental data' is refuted by §4.1's explicit synthetic generation. Consequently, the 'ground truth' in Figures 5–10 is the same constitutive model the surrogate is designed to replace, and the Abaqus 'reference solution' in §4.3.4 is a VUMAT implementation of that same model. Case 2's R²>0.97 on 10°C/500%/min and 200% stretch therefore demonstrates only that a TCN can interpolate/extrapolate within this parametric family; it says nothing about physics outside that family, such as permanent set, damage saturation differences, non-WLF temperature dependence, or viscoplastic mechanisms. The thermodynamic admissibility is likewise a soft penalty term, not a hard constraint (Section 3.5.1), so it cannot certify out-of-distribution states not represented in training. Without experimental hold-out validation, the VUMAT deployment claim is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a physics-informed, data-driven constitutive model for strongly nonlinear thermo-viscoelastic soft materials with Mullins-type damage. The model decomposes stress into volumetric and unified isochoric contributions expressed as linear combinations of integrity basis tensors with scalar response functions, following Upadhyay et al. A Temporal Convolutional Network (TCN) learns these response functions (and the damage variable) from sequences of invariant inputs, with a loss function that adds a dissipation penalty and a damage-boundedness penalty. The authors report accurate fits on training cases, generalization to unseen temperatures, strain rates, and stretch levels, robustness to 20% Gaussian noise, and agreement with Abaqus VUMAT simulations of an open-hole specimen. The central claim is that the TCN surrogate captures real soft-material behavior and can be deployed in finite element workflows.","tokens_in":31708,"tokens_out":4147,"duration_ms":51414,"significance":"If the claims were supported, the framework would be a useful contribution to machine-learning constitutive modeling: it uses causal TCNs for history dependence, an invariant-based decomposition that enforces objectivity and isotropy by construction, and a workflow for VUMAT deployment. The synthetic benchmark shows that a TCN can fit and interpolate within a parametric viscoelastic-damage family, which is a legitimate proof-of-concept. However, the paper's significance as a statement about real soft materials is currently unsupported: no experimental data appear anywhere, and the validation loop is closed within the synthetic generator. The novelty relative to existing physics-augmented neural-network approaches is incremental, and the absence of released code/data (data only 'on request') limits reproducibility.","major_comments":[{"comment":"The abstract claims the TCN is 'trained on high fidelity experimental data across multiple temperatures, strain rates, and stretch levels', but §4.1 states that all training data are generated synthetically from a Yeoh hyperelastic energy, an Ogden–Roxburgh damage function (Eq. 32), a four-branch generalized Maxwell model (Eqs. 33–36), and a WLF shift factor. No experimental measurement is described anywhere in the paper. Consequently, the central claim of generalization to real soft materials—the headline of the paper and the Conclusions—is unsupported. The reported R² values in Figures 5–10 and the 'ground truth' curves demonstrate only that the TCN can approximate the specific generator used to create the data, not that it captures real material physics.","section":"§4.1 and Abstract"},{"comment":"The training targets for the surrogate—the response functions ω1, ω2 and the damage variable—are not measured or independently known; they are obtained by solving the linear system [b]=[A][x] (Eq. 18) using the same constitutive model that generated the raw stress–strain data (§3.2.1). The FEM 'reference solution' in §4.3.4 is produced by a VUMAT implementation of that same constitutive model. The validation is therefore circular: the TCN is trained to reproduce a known analytical model, and its 'unseen' test cases are extrapolations within that same model family. This does not test whether the learned representation transfers to a different material, different damage evolution law, or different temperature dependence.","section":"§3.2.1, Eq. (18), and §4.3.4"},{"comment":"The text repeatedly calls the Clausius–Duhem inequality a 'hard thermodynamic constraint' and claims 'strict enforcement' of thermodynamic admissibility. However, the implementation described in §3.5.1 and Figure 3 is a soft penalty: the loss is MSE plus a dissipation penalty term that is added when dissipation is negative, and violations are penalized rather than excluded from the feasible set. Therefore the model does not guarantee thermodynamic consistency for out-of-distribution states not seen in training. The claim of 'ensured' thermodynamic admissibility in the abstract is not justified by the actual loss construction.","section":"§3.5.1 and Figure 3"},{"comment":"The damage-variable convention is internally inconsistent. The text states 'ϑ = 1 signifying no damage and ϑ = 0 signifying complete mechanical damage', but Eq. (2) multiplies the intact isochoric energy by the integrity factor (1−ϑ), so ϑ = 0 is the undamaged state and ϑ = 1 is complete damage. The same inconsistency affects the damage evolution law in Eq. (32) and the interpretation of the damage contours in §4.3.4, where a saturation value of approximately 0.09 is reported without clarifying whether this is '9% damage' or '91% damage'. This needs to be fixed before the results can be interpreted.","section":"§2.1 and Eq. (2)"}],"minor_comments":[{"comment":"Equation numbering is inconsistent: Eq. (24) is used for the relative-error metric in §4.3, but Eq. (24) already denotes the residual-block formula in §3.5. Also, Eq. (31) is introduced without Eqs. (29) and (30), and several equation numbers appear out of sequence.","section":"§4.3"},{"comment":"The TCN architecture description is incomplete for reproducibility: no learning rate, dropout rate, number of residual blocks, sequence length, or loss-weighting coefficients for the dissipation/damage penalties are given. The claim that 'only three tests is sufficient' would also benefit from repeated-training statistics.","section":"§4.2"},{"comment":"The paper first describes a volumetric surrogate and a stress-decomposition procedure requiring volumetric stress data, then states that volumetric effects are 'intentionally excluded' and the material is assumed incompressible. Please clarify the role of the volumetric model in the actual experiments and how the decomposition is performed for incompressible data.","section":"§3.4 and §3.2.1"},{"comment":"The phrase 'high-fidelity experimental data' in the abstract and the 'trained directly on ground truth data' in the Conclusions should be replaced with 'synthetic data generated from a known constitutive model' throughout, to match the content of §4.1.","section":"Abstract and §5"},{"comment":"Reference [142] (about wall-shear stress from neural-network-enhanced fluid flow measurements) does not appear to support the sentence 'internal damage and microstructural reconfiguration tend to degrade the material’s shear resistance far more significantly than its volumetric response' in §2.1; please verify the citation.","section":"References"}],"recommendation":"reject","confidential_remarks":"The central problem is not the machine-learning methodology but the gap between the claims and the evidence: the abstract and conclusions assert experimental fidelity and real-material generalization, while the entire study is a self-consistent synthetic benchmark. This is a load-bearing overclaim that cannot be fixed by local edits; either the paper must be reframed as a synthetic-data proof-of-concept with all experimental-language removed, or it needs genuinely independent experimental or high-fidelity simulation data from outside the assumed Yeoh/Ogden–Roxburgh/Maxwell/WLF family. The damage-variable convention reversal is an additional correctness concern that would need to be addressed in any revision. I would not recommend acceptance in the current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe contribution here is the combination of a TCN with invariant-based response functions and time-temperature superposition to model stress softening, rate dependence, and damage in one surrogate. That is a sensible idea and the continuum formulation is careful: decomposition of stress into integrity bases, reduced time as input, bounded damage variable, and a dissipation penalty in the loss. The architecture is a reasonable choice for history-dependent material behavior, and the reported fits suggest the model can learn the synthetic family well.\n\nThe problem is that the validation is self-referential. All training and test data come from the same analytic generator described in §4.1: Yeoh elasticity, Ogden–Roxburgh damage, a four-branch Maxwell model, and WLF shift. The abstract says 'high-fidelity experimental data,' which is false. The 'unseen' temperatures, strain rates, and stretches are still within that parametric family, so the R²>0.97 results show interpolation/extrapolation within a fictitious world, not predictive power for real materials. The FEM validation in §4.3.4 compares the surrogate to a VUMAT of the same constitutive model; it's a consistency check, not independent validation. The thermodynamic constraint is a soft penalty, not a hard guarantee, despite the label in §3.5.1. The noise-robustness test trains on 20% Gaussian noise added to synthetic stress, which is a fitting exercise rather than a robustness certification.\n\nThe paper also omits the parameter values needed to reproduce the data (Yeoh coefficients, Ogden–Roxburgh parameters, Maxwell times, WLF constants), so as published it's not reproducible. That's a real weakness for a methods paper.\n\nWho is this for? People working on surrogate constitutive models for finite elements might find the architecture worth borrowing. But the paper currently overclaims: it presents a method demonstration on synthetic data as a validated tool. It deserves a serious referee because the framework has substance, but the expected outcome is major revision — add real experimental data or reframe the claims honestly. As it stands, I would not cite it as evidence of predictive capability.\n\nHope this helps.","headline":"A clever architecture for data-driven thermoviscoelastic surrogates with Mullins damage, but the validation is entirely self-referential: the 'high-fidelity experimental data' are generated by the same model the surrogate is supposed to replace.","tokens_in":32282,"tokens_out":3707,"would_cite":false,"duration_ms":40994,"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 Temporal Convolutional Network trained on three thermomechanical tests can act as a thermodynamically consistent surrogate for nonlinear viscoelastic soft materials with stress-softening damage, extrapolating to unseen temperatures…","keywords":["stress softening","Mullins effect","thermo-viscoelasticity","time-temperature superposition","Temporal Convolutional Network","physics-informed machine learning","constitutive modeling","finite element VUMAT"],"falsifier":"Train the identical surrogate on experimental two-cycle data at three thermomechanical states and compare its predictions with measured third-cycle responses at 10°C, 500%/min, and 200% elongation; any systematic divergence in peak stress, hysteresis area, or damage saturation would falsify the extrapolation claim.","tokens_in":31094,"feed_emoji":"🧪","tokens_out":11253,"duration_ms":115622,"temperature":0.7,"pith_summary":"The paper tries to establish that one physics-informed data-driven surrogate can capture rate dependence, temperature sensitivity, large-strain cyclic loading, and stress-softening damage in soft materials within a single model. Rather than learning raw stress-strain curves, the network learns scalar response functions in a stress decomposition built on deformation invariants, so objectivity, isotropy, and symmetry are satisfied by construction. The central demonstration is extrapolation: after training on only three two-cycle tests, the model predicts unseen temperatures, strain rates 2.5 times higher, elongations 50% larger, and a third loading cycle, while keeping dissipated energy nonnegative and the damage variable bounded. If correct, this gives a thermodynamically admissible surrogate that can be embedded in finite element analysis and used to simulate Mullins-type softening in elastomers and soft tissues.","feed_headline":"Neural net predicts stress softening at unseen temperatures and rates","feed_subtitle":"A neural net trained on three tests predicts new temperatures, rates, and stretches while obeying thermodynamics.","key_machinery":"The load-bearing object is a reduced-time, invariant-based stress decomposition: total isochoric stress is a linear combination of three integrity basis tensors, $\\mathbb{I}_1 = \\mathbf{C}^{-1}$, $\\mathbb{I}_2 = \\mathrm{Dev}(\\mathbf{I})$, and $\\mathbb{I}_3 = \\mathrm{Dev}(\\mathbf{C})$, with scalar response functions $\\omega_\\alpha(\\tau^*, I_1, I_2, I_{1,\\max})$ learned by a dilated causal Temporal Convolutional Network. Temperature enters through a shift-factor reduced time $\\tau^*$, while Mullins-type damage is tracked by the historical maximum of the first isochoric invariant $I_{1,\\max}$, which activates damage only when the current loading exceeds past maxima. The loss function appends a penalty whenever the Clausius-Duhem dissipation inequality is violated and constrains the damage variable to the interval $[0,1]$, so thermodynamic admissibility is imposed during training rather than checked afterward.","core_discovery":"The paper's central claim is that a Temporal Convolutional Network trained on an invariant-based stress decomposition can act as a thermodynamically consistent surrogate for strongly nonlinear thermo-viscoelastic soft materials with stress-softening damage. In the reported tests, the surrogate reproduces cyclic loading-unloading hysteresis and damage accumulation in training cases, then extrapolates to 10°C from training temperatures of 0, 25, and 55°C, to 500%/min from a 200%/min training maximum, and to 200% elongation from 150%, while maintaining nonnegative internal dissipation and damage in the physical range. The model also tracks the first and second viscoelastic branch contributions, tolerates 20% Gaussian noise in the input stress, and matches finite element reference results for Mullins damage growth around an open hole. This is presented as closing the gap left by quasi-static or isothermal Mullins-effect models and by unconstrained data-driven networks that may violate the second law.","pith_inferences":["The strongest unstated test is experimental: the paper validates extrapolation against a synthetic generator, so retraining on real multicycle data from an actual elastomer and repeating the holdout comparisons would tell whether the generalization claim survives contact with physical measurements.","Because the formulation is built on invariants and response functions rather than a fixed strain-energy form, the same architecture could in principle be transferred to multiaxial loading histories, provided training data with those paths are available.","The separation of time-temperature effects through a reduced-time shift factor suggests that a material's master curve behavior could be learned once and reused across temperatures, cutting the data needed for each new temperature.","The tolerance to 20% input noise hints that the dilated causal structure acts as a strong regularizer, which would be valuable when training on noisy lab data rather than clean simulations."],"forward_implications":["A surrogate trained on just three two-cycle tests reproduces unseen loading cycles with R-squared values above 0.97 in the reported holdout cases.","The model extrapolates to temperatures, strain rates, and stretch levels outside its training range, which suggests fewer physical tests may be needed to characterize a material across its service envelope.","Because dissipation is enforced nonnegative and damage is bounded, the surrogate can be used inside finite element solvers without producing energetically impossible stress cycles.","The noise-robustness result implies that modest experimental scatter in stress measurements need not destroy predictive accuracy.","Deployment through a user material subroutine allows the surrogate to replace an explicit constitutive law in structural-scale simulations of Mullins damage."],"supporting_citations":[{"why":"Supplies the pseudo-elastic damage evolution law used to generate the Mullins-type softening training data.","marker":"[146]"},{"why":"Provides the nonlinear continuum and generalized Maxwell framework from which the stress decomposition and branch evolution equations are taken.","marker":"[28]"},{"why":"Gives the thermodynamically consistent configurational free-energy form and driving force used to build the dissipation constraint.","marker":"[145]"},{"why":"Prior nonlinear thermo-visco-elastic constitutive model with Mullins damage that the surrogate framework extends.","marker":"[57]"},{"why":"Provides the shift-factor relation used to define reduced time for time-temperature superposition.","marker":"[143]"},{"why":"Establishes the invariant-to-response-function machine learning formulation that the surrogate's inputs and outputs are based on.","marker":"[144]"},{"why":"Supplies the dilated causal Temporal Convolutional Network architecture with residual blocks used as the sequence model.","marker":"[98]"},{"why":"A quasi-static symbolic-regression Mullins model that defines the rate-independent baseline this work aims to surpass.","marker":"[126]"},{"why":"A rate-independent, isothermal physics-augmented neural network for Mullins damage that this work extends to rate- and temperature-dependent response.","marker":"[108]"}],"fun_headline_variants":["Thermo-viscoelastic damage model learns Mullins effect","Data-driven model extrapolates stress softening across temps","Physics-informed neural net enforces second law in soft matter","Neural surrogate predicts damage in soft materials at new conditions","TCN captures Mullins hysteresis and extrapolates to hotter, faster, longer"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole extrapolation story is built on synthetic training data generated by a specific combination of mathematical material laws; if those laws do not faithfully represent real soft materials, the reported success shows only that the network learned the generator, not that it predicts physical behavior.","fun_headline_variants_meta":{"raw":{"variants":["Thermo-viscoelastic damage model learns Mullins effect","Data-driven model extrapolates stress softening across temps","Physics-informed neural net enforces second law in soft matter","Neural surrogate predicts damage in soft materials at new conditions","TCN captures Mullins hysteresis and extrapolates to hotter, faster, longer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000399,"raw_usage":{"total_tokens":2075,"prompt_tokens":922,"completion_tokens":1153,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":1068}},"tokens_in":538,"tokens_out":1153,"duration_ms":8863,"temperature":1.0,"reasoning_tokens":1068,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:41:43.396711+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the identical surrogate on experimental two-cycle data at three thermomechanical states and compare its predictions with measured third-cycle responses at 10°C, 500%/min, and 200% elongation; any systematic divergence in peak stress, hysteresis area, or damage saturation would falsify the extrapolation claim.","supporting_citations":[{"cited_title":"Rediscovering the Mullins effect with deep symbolic regression,","cited_arxiv_id":null,"evidence_quote":"A quasi-static symbolic-regression Mullins model that defines the rate-independent baseline this work aims to surpass."},{"cited_title":"Recovering Mullins damage hyperelastic behaviour with physics augmented neural networks,","cited_arxiv_id":null,"evidence_quote":"A rate-independent, isothermal physics-augmented neural network for Mullins damage that this work extends to rate- and temperature-dependent response."}],"review_version":1}