{"id":"b4741a46-0e5d-44dd-8a96-cf006d2e3531","arxiv_id":"2507.02991","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A physics-augmented neural network learns composition-dependent, rate-dependent stress-strain behavior of PolyJet digital materials, interpolating well to held-out blends but extrapolating poorly.","lead":"This paper measures how mixtures of soft and rigid 3D-printed photopolymers stretch and twist, then trains a physics-informed neural network to predict the material's behavior from its composition. It shows the model accurately interpolates to a never-seen intermediate blend, but fails when extrapolating beyond the tested range.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The learned constitutive law is only as valid as the exact-incompressibility/isotropy kinematics; ν≥0.47 at λ<1.05 does not justify J≡1 and isotropy at rupture-scale stretches where the model is trained and validated.","rationale":"The reader’s weakest assumption identifies the same load-bearing point: the model treats the printed digital materials as perfectly incompressible and isotropic over the full deformation range, although the supporting Poisson’s-ratio measurement is limited to small stretches. I agree that this is the most fundamental gap because it affects the kinematic transformation from measured forces/torques to all quantities used in training and evaluation. I considered the torsion data-range inconsistency (φ≤360° in §2.2 versus φ≤720° in Figure 6) and the overstated polyconvexity claim; both are real presentation issues, but they are less load-bearing than the kinematic assumption. The suggested retraining test directly settles whether the incompressibility assumption changes the conclusion. It does not make me reject the paper: the empirical interpolation result is plausible, the staged training is described in detail, and the paper honestly reports extrapolation failure for DM-70. The concern only reinforces the conditional verdict: the paper should either provide large-strain transverse-strain evidence or retrain with the measured kinematics before claiming discovery of the true constitutive law.","tokens_in":19525,"tokens_out":11662,"duration_ms":256992,"concrete_test":"Retrain the identical pipeline of §4.2 with the actual transverse stretch histories recorded by the video extensometer in place of the incompressible λ^{-1/2} in Eq. 17, and compare held-out DM-40 R²/sMAPE and the resulting sparse expression Eq. 26. If the held-out metrics change materially (for example, DM-40 R² drops below 0.95 or the active terms in Eq. 26 change by more than 10%), the exact-incompressibility assumption is load-bearing and the constitutive claim needs to be qualified. If the metrics are essentially unchanged, the assumption is benign for the empirical interpolation claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the PANN discovers the composition-dependent constitutive law, not merely that it interpolates stress-stretch curves. Every training target and every reported R² depends on converting displacement/force measurements into stresses and invariants through assumed deformation gradients in §3.3 and §3.4. In uniaxial tension, Eq. 17 sets the transverse stretch to λ^{-1/2}, so the invariants used to train the pICNN are computed under exact incompressibility. In torsion, Eq. 19 enforces J=1 by construction and neglects any radial expansion or axial force (Poynting) response. The supporting measurement is only a Poisson’s ratio ν≥0.47, taken at λ<1.05 (§2.1.1, §4.1), while the model is trained on data all the way to rupture, where microcavitation and damage can change volume. The paper itself flags the limit of the Poisson measurement, but no large-strain transverse strain data are shown. If J deviates from 1 in the training range, the isochoric invariants, the pressure reconstruction in uniaxial stress, and the torsion shear stress in Eq. 21 are all computed from a mis-specified kinematic target. In that case the high DM-40 test accuracy is a fit to the wrong invariants, and the discovered Ψ(I1,I2,c) is not the true material response.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper combines an experimental campaign on five PolyJet digital materials (A, DM-40, DM-50, DM-60, DM-70) under uniaxial tension and torsion at several rates with a physics-augmented neural network model. The model uses a partially input convex neural network (pICNN) to represent a composition-dependent hyperelastic strain energy and a quasi-linear viscoelastic (QLV) formulation with a composition-dependent relaxation coefficient to capture rate dependence. Training is performed on A, DM-50, and DM-60, with DM-40 held out for interpolation and DM-70 for extrapolation. The authors report high R^2 and low sMAPE on training data and on the interpolated DM-40, and they honestly document the failure to extrapolate to DM-70. L0 sparsification reduces the model to closed-form expressions for the energy and relaxation parameter.","tokens_in":19842,"tokens_out":9630,"duration_ms":116360,"significance":"If the physical assumptions behind the kinematic reductions are valid, this is a useful contribution: it demonstrates a scalable, composition-aware framework for constitutive model discovery, provides experimental data for an under-characterized class of printed materials, and includes a rare, clearly reported extrapolation failure that delimits the method's validity. The sparse closed-form expressions in Eqs. (26) and (27) are a concrete interpretability outcome, and the staged training and honest reporting of held-out performance are methodological strengths. The main reservations concern the strength of the 'polyconvexity' guarantee and the kinematic assumptions used to convert raw force/torque measurements into the invariants and stresses that define the training targets.","major_comments":[{"comment":"The paper states that convexity of the learned strain energy with respect to (I1, I2) 'ensures the polyconvexity conditions required for hyperelastic constitutive modeling.' This is not sufficient. Polyconvexity of an isotropic incompressible energy requires additional structure beyond convexity in the two invariants, typically monotonicity conditions on the derivatives ∂Ψ/∂I1 and ∂Ψ/∂I2 (or an explicit representation in terms of F and cof F), and convexity in (I1, I2) alone does not imply rank-one convexity or ellipticity. Since the paper does not verify ellipticity, monotonicity, or a suitable invariant-based polyconvexity theorem, the guarantee is overstated. Please either prove the required conditions for the specific architecture or soften the claim to 'convex in I1 and I2' and discuss what stability property this actually provides.","section":"§3.1–3.1.2"},{"comment":"The incompressibility and kinematic reductions are load-bearing because every training target is computed from them. Eq. (17) imposes transverse stretch λ^{-1/2} for the entire tensile deformation range up to rupture, justified only by ν ≥ 0.47 measured at λ < 1.05. The video extensometer recorded transverse strain; please report large-strain transverse stretch data, either to confirm J ≈ 1 over the training range or to use measured transverse stretches in training. Similarly, Eq. (19) assumes pure simple torsion with no radial expansion and no axial stretch. For an incompressible isotropic cylinder with a free lateral surface, this deformation is not generally an equilibrium solution at the twists used here; Poynting-type normal stresses and radial deformation can be significant at shear strains of order 0.5. If the actual deformation in the torsion tests was not verified, the invariants and shear stresses in Eqs. (20)–(21) may be systematically wrong, and the high reported R^2 values would be fits to a mis-specified kinematic target.","section":"§3.3–3.4, §2.1.1"},{"comment":"The torsion data range is internally inconsistent. Section 2.2 states that torsion results are shown for φ ≤ 360° because at higher angles micro-cracks evolve and the rod buckles and deforms out-of-plane, but Section 4.1 and Figure 6 plot T L/Jp up to φ ≤ 720°, and Figure 2 shows a specimen at φ = 720°. Please clarify which angular range is used for training and how post-buckling or damaged data are excluded. This matters because the apparently linear torque response at large angles may include structural instability rather than intrinsic material response.","section":"§2.2 vs §4.1/Fig. 6"},{"comment":"The QLV convolution is written as σ(t) = σe(t) + ∫_0^t D′(t−s) σe(t) ds, with σe(t) inside the integral. As written this is not a history-dependent convolution: σe(t) is constant with respect to the integration variable, so the integral reduces to a time-dependent scalar factor multiplying the current stress. The standard QLV form should have σe(s) under the integral. If the implementation uses the correct form, please correct the equation; if not, the model is not the QLV model claimed in the text. This is a central equation and needs to be fixed.","section":"Eq. (14)"}],"minor_comments":[{"comment":"The sentence 'Out of 24 training and test datasets for these compositions, 21 have an R2 above 0.98, and 22 have a sMAPE below 8%' is not supported by the rounded values in Table 3. For example, DM-60 has rows with R2 of 0.977, 0.956, and 0.925, DM-50 has rows of 0.980, and several sMAPE values exceed 8%. Please recompute these counts from unrounded metrics or adjust the sentence.","section":"Table 3 and §5"},{"comment":"It is unclear whether Poisson's ratio was measured in the present study or taken from reference [28]. Section 2.1.1 describes a calculation from the uniaxial tests, while Section 4.1 attributes the values to Levin and Cohen [28]. Please state this explicitly and report the measured ν values for each composition if they are new data.","section":"§2.1.1 and §4.1"},{"comment":"The mapping from % Digital ABS to the composition parameter c is not a simple linear scaling of the digital ABS percentages (e.g., 25.2% maps to 0.1755 while 27.2% maps to 1.0). Please provide the exact scaling formula or clarify how c is defined, since Eq. (26) and Eq. (27) depend on this input definition.","section":"Table 2"},{"comment":"The first line of Eq. (11), 's = s(ζ − γ) + γ1', appears to contain a typo or undefined notation. Please correct or define the intermediate variable s̄ (or similar) used in the hard-sigmoid approximation.","section":"Eq. (11)"}],"recommendation":"major_revision","confidential_remarks":"The core empirical contribution is solid and the authors report negative extrapolation results honestly, which is rare and valuable. The main risk is not the curve fitting but whether the reported R^2 values are computed against the correct physical targets. I would support a major revision rather than rejection because the authors have access to transverse strain measurements and can either validate the incompressibility assumption over the full strain range or retrain with measured kinematics; the torsion inconsistency and the equation typo are also fixable. I would not accept the paper before these points are resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this paper earns its conditional verdict. The empirical core is convincing and the authors are honest about the model's limits. Held-out composition DM-40 is predicted well (R2 ≥ 0.96 in tension, 0.99 in torsion), and the extrapolation to DM-70 fails—they report that failure instead of hiding it. The L0 sparsification yields compact closed-form expressions for the energy and relaxation coefficient, which is a real plus for interpretability.\n\nWhat's actually new: a composition-conditioned pICNN for the hyperelastic energy combined with an MLP for the QLV relaxation coefficient, applied to a new tension/torsion dataset for five Agilus/Digital ABS blends. The building blocks are established, but the combination and the held-out composition benchmark are new.\n\nSoft spots, in proportion. First, the polyconvexity claim is overstated. Convexity in (I1, I2) is not sufficient for polyconvexity; the paper equates the two in Section 3.1. This is a framing error, not a flaw in the fitted model, but it should be fixed. Second, the incompressibility assumption is the bigger issue. They measure ν ≥ 0.47 only at λ < 1.05, yet impose J=1 all the way to rupture and in torsion. If the material dilates at large stretch or in twist, every training target is computed from the wrong kinematics, and the learned Ψ is not the true material response. The paper flags the small-strain measurement but gives no large-strain transverse strain data; the stress-test concern lands. Third, a minor internal inconsistency: Section 2.2 says torsion results are shown for φ ≤ 360°, while Figure 6 plots to 720°. The training range needs clarification. Also, no data/code or error bars are provided; the explicit final equations help, but a data deposit would strengthen reproducibility.\n\nWho this is for: people in computational mechanics and additive manufacturing who want composition-aware constitutive models for simulation. It deserves a serious referee. I would send it to review with requests to correct the polyconvexity language, justify or test the large-strain incompressibility assumption with direct kinematic data, and clear up the torsion angle range.","headline":"Composition-aware pICNN+QLV model with a real held-out interpolation test; the main caveats are the overclaimed polyconvexity guarantee and the unverified large-strain incompressibility assumption.","tokens_in":20388,"tokens_out":3909,"would_cite":true,"duration_ms":44067,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A physics-augmented neural network learns composition-dependent stress and torque laws for 3D printed blends, predicting a held-out mixture accurately but not blends beyond the trained composition range","keywords":["Multi-material 3D printing","Physics-augmented neural network","Partially input convex neural network","Constitutive modeling","Hyperelasticity","Viscoelasticity","Digital materials","L0 sparsification"],"falsifier":"Measure volume change directly during uniaxial tension to large stretch ($\\lambda > 1.2$) and during torsion on a printed DM-50 specimen; if the material dilates or compresses by more than a few percent, the incompressibility assumption ($J=1$) that defines all strain invariants and stresses in Sections 3.3 and 3.4 fails, and the learned energy function does not describe the true material.","tokens_in":19348,"feed_emoji":"🖨️","tokens_out":18273,"duration_ms":167522,"temperature":0.7,"pith_summary":"The paper sets out to show that a single composition-aware physics-augmented neural network can learn the mechanical response of 3D printed digital materials, a continuum of blends between a soft rubber-like resin and a stiff glassy resin, without fitting a separate constitutive model for each blend. The network learns a hyperelastic strain-energy function that is convex in the strain invariants and depends non-convexly on composition, then feeds that energy into a quasi-linear viscoelastic model whose relaxation coefficient also depends on composition. On experimental data from five blends tested in tension and torsion at multiple rates, the model reproduces the nonlinear, rate-dependent curves and, for a held-out intermediate blend (DM-40), predicts the response with $R^2 \\geq 0.96$ in tension and $R^2 = 0.990$ in torsion. The broader goal is a scalable, automated route to constitutive-model discovery for multi-material printing, where the space of possible compositions is effectively continuous.","feed_headline":"Held-out 3D-printed blend predicted at R² ≥ 0.96","feed_subtitle":"Single model trained on five blends reproduces tension and torsion for an unseen mixture; extrapolation fails.","key_machinery":"The load-bearing mechanism is the partially input convex neural network (pICNN) coupled to a quasi-linear viscoelastic (QLV) model. The pICNN enforces polyconvexity of the strain-energy function with respect to strain invariants while allowing arbitrary non-convex dependence on composition, ensuring thermodynamic admissibility; the QLV layer with relaxation time $\\tau = 10$ s fixed and composition-dependent coefficient $\\gamma$ (output by an MLP) captures rate dependence; $L_0$ sparsification prunes the trained network to 17+4+11 active parameters, yielding the closed-form strain energy of Eq. 26 and $\\gamma(c)$ of Eq. 27.","core_discovery":"The central claim is that a partially input convex neural network (pICNN) coupled to a quasi-linear viscoelastic (QLV) kernel can serve as a unified, composition-aware constitutive law for Agilus/Digital-ABS digital materials. Strain invariants $\\bar{I}_1$ and $\\bar{I}_2$ enter the convex branch, the resin mixing ratio $c$ enters the non-convex branch, and a separate multilayer perceptron maps $c$ to the QLV relaxation coefficient $\\gamma$; the resulting model is trained jointly on uniaxial tension at three stretch rates and torsion at 360 deg/min. After $L_0$ sparsification the network shrinks from 1,357+75 weights to 17+4+11 active parameters and yields closed-form expressions for the strain energy (Eq. 26) and $\\gamma(c)$ (Eq. 27). The paper establishes that this model achieves high accuracy on training compositions ($R^2 > 0.98$ in 21 of 24 datasets), interpolates well to the held-out DM-40 composition ($R^2 \\geq 0.964$ in tension, $R^2 = 0.990$ in torsion), and degrades sharply when extrapolating to the stiffer held-out DM-70 ($R^2$ as low as 0.251 in tension, negative in torsion).","pith_inferences":["A practical consequence the authors leave implicit: to apply this method to a new material family, the training set should bracket the intended composition range with endpoint formulations, because interpolation is reliable but extrapolation is not.","Because the torsion tests in this study stayed in the quasi-static regime, the relaxation coefficient $\\gamma$ is constrained mainly by tension data; adding higher-rate or cyclic torsion would likely produce a different $\\gamma$ and sharpen the viscoelastic branch.","The near-linearity of torque vs twist angle alongside strongly nonlinear tension suggests the second strain invariant $\\bar{I}_2$ of the learned energy is weakly identified by the current experiments; biaxial or shear tests would pin down the $\\bar{I}_2$ dependence.","The architecture is not specific to resin jetting; the same pICNN-QLV combination should transfer to any continuously tunable material family, provided the composition input is encoded monotonically with the property of interest."],"forward_implications":["A single trained model replaces per-composition constitutive fitting: once trained on a handful of blends, it supplies stress and torque predictions for any intermediate composition in the training range.","The $L_0$-sparsified result is a closed-form strain-energy function and $\\gamma(c)$ mapping with only 17 convex, 4 non-convex, and 11 connection parameters, so the discovered law can be written down, inspected, and embedded in simulations.","The framework covers both uniaxial tension and torsion within one model, so multiaxial behavior of an interpolated composition can be predicted without new experiments.","Extrapolation in composition is not reliable: for held-out DM-70 the model's $R^2$ drops to 0.251 at the slowest tension rate and becomes negative in torsion, so the model's valid range is bounded by the trained composition interval."],"supporting_citations":[{"why":"Supplies the input-convex neural network construction that guarantees convexity of the learned energy with respect to strain invariants.","marker":"[51]"},{"why":"Introduces partial input convexity, allowing composition to enter through a non-convex branch; the paper modifies this architecture.","marker":"[52]"},{"why":"Provides the smoothed L0 gating mechanism that prunes network weights to a compact, interpretable set.","marker":"[50]"},{"why":"Supplies the stress-normalization procedure that subtracts the reference-state energy so stress vanishes at the undeformed configuration.","marker":"[53]"},{"why":"Demonstrates physics-augmented neural networks with extreme sparsification for interpretable constitutive discovery, the methodological template extended here.","marker":"[40]"},{"why":"Supplies the quasi-linear viscoelastic formulation and kernel used for the time-dependent stress response.","marker":"[32]"},{"why":"Provides the finite-torsion torque integral used to convert the learned strain-energy into torque predictions.","marker":"[55]"},{"why":"Supplies the measured Poisson's ratio (about 0.47) that supports the incompressibility assumption behind the deformation gradients.","marker":"[28]"}],"fun_headline_variants":["AI learns 3D-printed blend laws, predicts unseen mix","Physics-augmented net: exact for seen, off for stiffer","Sparse neural net yields closed-form rules for printed blends","pICNN+QLV predicts new blend, then hits extrapolation wall","Neural net with convex strain finds unseen blend mechanics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The printed digital materials are assumed to be isotropic and perfectly incompressible ($J=1$) over the full deformation range, based on a Poisson's ratio of about 0.47 measured only at small stretches ($\\lambda < 1.05$).","fun_headline_variants_meta":{"raw":{"variants":["AI learns 3D-printed blend laws, predicts unseen mix","Physics-augmented net: exact for seen, off for stiffer","Sparse neural net yields closed-form rules for printed blends","pICNN+QLV predicts new blend, then hits extrapolation wall","Neural net with convex strain finds unseen blend mechanics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1524,"prompt_tokens":1096,"completion_tokens":428,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":712,"completion_tokens_details":{"reasoning_tokens":339}},"tokens_in":712,"tokens_out":428,"duration_ms":6478,"temperature":1.0,"reasoning_tokens":339,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:00:17.468704+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure volume change directly during uniaxial tension to large stretch ($\\lambda > 1.2$) and during torsion on a printed DM-50 specimen; if the material dilates or compresses by more than a few percent, the incompressibility assumption ($J=1$) that defines all strain invariants and stresses in Sections 3.3 and 3.4 fails, and the learned energy function does not describe the true material.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the input-convex neural network construction that guarantees convexity of the learned energy with respect to strain invariants."},{"cited_title":"Inverse design of anisotropic microstructures using physics-augmented neural networks","cited_arxiv_id":"2412.13370","evidence_quote":"Introduces partial input convexity, allowing composition to enter through a non-convex branch; the paper modifies this architecture."},{"cited_title":"K., Kalina, K","cited_arxiv_id":null,"evidence_quote":"Supplies the stress-normalization procedure that subtracts the reference-state energy so stress vanishes at the undeformed configuration."},{"cited_title":"N., Jones, R","cited_arxiv_id":null,"evidence_quote":"Demonstrates physics-augmented neural networks with extreme sparsification for interpretable constitutive discovery, the methodological template extended here."},{"cited_title":"(2009) Nonlinear Viscoelastic Solids - A Review.Mathematics and Mechanics of Solids 14, 300–366","cited_arxiv_id":null,"evidence_quote":"Supplies the quasi-linear viscoelastic formulation and kernel used for the time-dependent stress response."},{"cited_title":"O., and Saccomandi, G","cited_arxiv_id":null,"evidence_quote":"Provides the finite-torsion torque integral used to convert the learned strain-energy into torque predictions."},{"cited_title":"(2023) Swelling under Constraints: Exploiting 3D-Printing to Optimize the Performance of Gel-Based Devices.Advanced Materials Technologies 8, 2202136","cited_arxiv_id":null,"evidence_quote":"Supplies the measured Poisson's ratio (about 0.47) that supports the incompressibility assumption behind the deformation gradients."}],"review_version":1}