{"id":"1f77d4a2-84b1-4c38-b7f2-319218cb54dd","arxiv_id":"2411.13185","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A compact physics-augmented neural network trained only on three plane-stress test curves reproduces Mullins damage hyperelasticity in full 3D finite element simulations.","lead":"Researchers trained a small neural network on simulated rubber test data to reproduce the Mullins effect, the way rubber softens after being stretched the first time. The network then predicted stresses, energy, and damage in 3D Abaqus simulations of a rolling disc and a twisted cylinder, matching the reference model within about one percent.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim that three plane-stress training curves determine the full 2-D energy and damage model is not verified for the actual invariant states reached in the 3-D examples; until those trajectories are shown to lie inside the trained envelope, the 'full 3-D recovery' conclusion rests on an…","rationale":"The reader's weakest assumption is that the three training curves determine the full 2-D strain energy and damage evolution over the region reached in the 3-D examples, and that the 3-D states are not verified to lie inside the trained envelope. This is exactly the load-bearing weak point. The paper is otherwise well executed: the shared-weight architecture is a genuine contribution, the validation against the Ogden generator is thorough, and the Abaqus implementation is useful. The empirical error map in Fig. 13 provides partial support, but it is not connected to the actual invariant trajectories of the 3-D boundary value problems, so the strongest claim remains conditional. I therefore do not change the reader's verdict; the same verification step should be requested before the full 3-D recovery claim is accepted.","tokens_in":22419,"tokens_out":15313,"duration_ms":164731,"concrete_test":"Instrument the Abaqus UHYPER subroutine to log, at every integration point and converged increment in the rolling-disc and diabolo simulations, the current invariants (I1,I2) and the stored maxima (I1,max,I2,max). Plot these trajectories on the (I1,I2) plane together with the three training curves from Sec. 2.1 and the random-sampling domain of Fig. 12. Then evaluate the NN's relative error in psi0 and in the damaged energy (1-zeta)psi0 against the Ogden reference (Eqs. 4, 9, 10) at those exact states. If all 3-D states lie inside the training-region envelope and the maximum pointwise energy error is at or below the ~1% level of Fig. 13, the coverage concern is resolved; if any states fall outside or errors exceed that level, the 3-D examples do not demonstrate recovery outside the training envelope.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that three one-parameter plane-stress tests (uniaxial, equibiaxial, planar tension, Sec. 2.1) determine the full 2-D undamaged energy psi0(I1,I2) and the Mullins damage evolution, so that the trained network acts as a 3-D constitutive model. This requires an identifiability/coverage step that the paper does not establish. The training data are 1-D curves in the (I1,I2) plane; the network is low-capacity, but no argument is given that these curves pin down psi0 and the damage split over the 2-D region sampled in the rolling-disc (Sec. 4.2) and diabolo (Sec. 4.3) examples. Fig. 13 validates the energy on a random invariant domain, but the paper does not report where the 3-D FE integration points actually lie: if those states are inside the region traced by the training curves, the result is interpolation; if outside, it is extrapolation. The paper gives no such classification. Furthermore, the trainable-beta subnetwork in Sec. 4.1.1 (and Remark 2) already demonstrates a stress-equivalent solution that fits the data but predicts wrong energy and damage (Fig. 14), so recovery from stress data alone is not automatic. Without an invariant-trajectory check or an identifiability argument, the 'correctly captures the full 3-D behaviour' conclusion is not supported for states away from the training paths.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a physics-augmented feed-forward neural network for incompressible hyperelasticity with isotropic Mullins-type damage. The network outputs strain energy as a function of the invariants I1, I2 and of a maximum-history undamaged energy, with stress obtained by automatic differentiation; a damage branch multiplies the undamaged energy by (1−ζ), using either a specialised exponential saturation law or a trainable tanh subnetwork. Training data are synthetic uniaxial, equibiaxial, and planar tension stress-stretch curves with loading/unloading cycles, generated from a three-term Ogden model with damage (ζ∞=0.8, ι=1). Six constrained and unconstrained architecture variants are compared; the unconstrained network (case 4) gives the lowest loss and the best agreement in subsequent tests. The model is implemented in an Abaqus UHYPER subroutine and tested on homogeneous verification tests, a rolling solid rubber disc, and a diabolo under tension-torsion, with reported median errors mostly below 1.5% for stresses and reactions, and good qualitative agreement for damage evolution.","tokens_in":22798,"tokens_out":4816,"duration_ms":47734,"significance":"The practical aim — a compact, few-parameter neural-network constitutive model for incompressible rubber with Mullins damage, trained only on three plane-stress tests and usable in a commercial finite-element package — is attractive, and the paper provides a useful engineering recipe. Strengths include the complete UHYPER implementation details (Algorithms 1 and 2), a systematic comparison of six architecture/constraint cases, and quantitative error reporting on stresses, energies, reaction forces, and damage fields. The claim of 'full 3D recovery' is plausible but not fully established: the paper does not show that the 3D deformation states lie inside the trained invariant envelope, and the specialised damage branch partly assumes the generator's functional form. The paper's own Fig. 14 example demonstrates that stress data alone can admit models with wrong energy and damage, so the identifiability gap is real and should be addressed before the strongest conclusions are drawn.","major_comments":[{"comment":"The central claim that three one-parameter training curves determine the full two-dimensional energy and damage evolution is not verified for the 3D examples. Training data are uniaxial, equibiaxial, and planar tension curves (Sec. 2.1); Fig. 13 validates the energy on a random invariant domain with I1≤50 and I2≤625, but the paper never reports the (I1,I2) states actually visited at integration points in the rolling-disc (Sec. 4.2) and diabolo (Sec. 4.3) simulations. If those states fall within the trained envelope, the examples are interpolation; if they fall outside, the conclusion that the network 'correctly captures the full 3D behaviour' rests on extrapolation. Please add invariant-trajectory plots for the 3D examples overlaid on Fig. 12, and report the maximum energy error along those trajectories.","section":"Sec. 2.1, Sec. 4.2, Sec. 4.3"},{"comment":"The specialised damage branch uses the same exponential saturation law as the data generator (Eq. (10)), with ζmax and ι_NN trained. Consequently the agreement of ζ and energy in Figs. 11c, 22, and 27 confirms parameter identification within a known functional form rather than discovery of the damage law from data. Report the trained values of ζmax and ι_NN for the case-4 network and compare them with the generator values (0.8 and 1.0); this will clarify the epistemological status of the 'recovery' claimed in the abstract.","section":"Sec. 3.1, Eq. (10)"},{"comment":"The abstract states that the architecture fulfils polyconvexity and non-negativity of the energy, but the recommended model is case 4 (Sec. 3.2.1), for which no weights are constrained and neither property is guaranteed. Sec. 3.2 explicitly relaxes the polyconvexity criteria, and the conclusion says overconstraining the network to be polyconvex is too restrictive. The abstract should be rephrased to distinguish the constraints enforced a priori (objectivity, thermodynamic consistency, normalisation) from properties that are only numerically observed (non-negativity) or not enforced (polyconvexity).","section":"Abstract, Sec. 3.2.1, Sec. 5"},{"comment":"The trainable-β subnetwork experiment is a load-bearing counterexample that deserves more discussion than a brief remark. It fits the stress curves with 4.3% error but mispredicts energy and damage for λ>5, with β=0.86 exceeding the generator's ζ∞=0.8. This demonstrates non-uniqueness of energy and damage given stress-only training data. The paper should either supply an identifiability argument for the specialised damage branch (e.g., that the exponential form plus shared weights removes the degeneracy), or add information (energy measurements, dissipation data, or additional loading paths) to justify the claimed recovery of energy and damage.","section":"Sec. 4.1.1, Fig. 14"}],"minor_comments":[{"comment":"The paragraph begins with the typo 'In tIn this section'; please correct it.","section":"Sec. 4.3.1"},{"comment":"The word 'refferred' should be 'referred'.","section":"Sec. 1"},{"comment":"The word 'futlfilled' should be 'fulfilled'.","section":"Sec. 3.2.1"},{"comment":"The notation C_i is introduced in the brace in Eq. (A.3) but is not used elsewhere; the derivative expression is correct, but the brace annotation is unnecessary and could be removed.","section":"Eq. (A.3)"},{"comment":"The color map for relative error and the number of random samples used to create the invariant domain are not specified; please add a color bar and state the sample count.","section":"Fig. 13"},{"comment":"Algorithm 2 is referenced in the text but appears without a formal caption; label it consistently with Algorithm 1.","section":"Algorithms 1 and 2"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the empirical demonstrations are valuable. The main risk is that the 'full 3D recovery' language overstates identifiability; the requested trajectory check and a more careful statement of what is learned versus assumed would resolve this. The manuscript appears already published in JMPS per the copyright notice, so the revision concerns the arXiv version and the clarity of its claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a solid methods paper on learning Mullins-type damage hyperelasticity with physics-augmented NNs. The core idea: a compact feed-forward energy network with an exponential activation, a shared-weight branch that computes the historical maximum undamaged energy, and a multiplicative damage factor. Trained on stress from three plane-stress tests (uniaxial, equibiaxial, planar tension), the network reproduces the generating Ogden-plus-Mullins model in 3D FEA benchmarks (rolling disc, diabolo) with median errors around 0.3–1.5%. That is a genuinely useful result for data-driven constitutive modeling, and the Abaqus implementation details are practical.\n\nWhat is new: the shared-weight max-energy branch is a neat trick to let a network that only sees damaged stress still recover the undamaged backbone and damage evolution; the tanh subnetwork alternative and the comparison of six constraint cases are also useful. The paper is honest about a real identifiability issue: a trainable beta subnetwork can fit the stress data but get the energy and damage wrong (Fig. 14, Remarks 1–2). That is an important caveat, and they document it clearly.\n\nThe soft spots are real but mostly fixable. The abstract says the architecture fulfills polyconvexity and non-negativity, but the recommended case 4 enforces neither; the constrained cases underperform. That mismatch should be corrected in the abstract. More importantly, training and validation share the same synthetic Ogden generator, so there is no independent experimental test, though that is normal for a methods paper. The paper does not release code or trained weights, so the exact model is not reproducible from the text alone. Finally, the strongest claim—that three 1D plane-stress curves determine the full 2D energy and damage over the region reached in the 3D examples—is supported empirically but never checked directly: no invariant trajectory plots for the 3D integration points, so we do not know if those states fall inside the trained envelope or outside it. The random invariant-domain test (Fig. 13) helps, and the small 3D errors suggest the network is generalizing rather than extrapolating, but showing the actual invariant paths would close the gap.\n\nWho is this for? Researchers working on neural constitutive models in computational mechanics, especially anyone wanting an FE-ready NN for rubber damage. It deserves serious peer review; a competent referee should ask for code/weights, the invariant-trajectory check, and a toned-down abstract. I would not desk-reject it.","headline":"A capable, honest NN constitutive model for Mullins damage that trains on three stress curves and validates in 3D FEA at roughly one percent error; the strongest claim needs an invariant-coverage check, but the paper deserves serious peer review.","tokens_in":23326,"tokens_out":3585,"would_cite":true,"duration_ms":33886,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["74B20","74A45","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a small feed-forward neural network trained only on uniaxial, equibiaxial, and planar tension data can reproduce the full three-dimensional hyperelastic response with Mullins-type damage.","keywords":["Mullins effect","hyperelasticity","physics-augmented neural networks","isotropic damage","incompressible rubber","invariant-based constitutive modelling","finite element implementation"],"falsifier":"Apply the trained network to a 3D boundary value problem whose local deformation states fall outside the invariant envelope covered by the training tests — for example, combined torsion and inflation of a thick cylinder where the invariant pair exceeds $I_1 > 50$ or $I_2 > 625$ — and compare the resulting stress, energy, and damage fields against the Ogden-with-Mullins reference solution. If the relative errors grow substantially beyond the ~1% seen inside the envelope, the claim of full 3D recovery would be restricted to interpolation of the training data.","tokens_in":22150,"feed_emoji":"🧪","tokens_out":5034,"duration_ms":45119,"temperature":0.7,"pith_summary":"The paper claims that a small feed-forward neural network can act as a complete constitutive model for incompressible hyperelastic materials with isotropic Mullins-type damage, provided its architecture is built around physical constraints rather than trained as a black box. The network is trained only on stress data generated from three plane-stress tests — uniaxial, equibiaxial, and planar tension — for an Ogden material with a standard exponential damage law. After training, it reproduces not just those curves but the full three-dimensional stress and energy response under far more complex loading, including the evolution of the damage variable. A sympathetic reader would care because this suggests that simple, inexpensive test data may be enough to obtain a ready-to-use material model for finite-element simulation of rubber-like parts with damage.","feed_headline":"Three simple tests teach a small neural net the full Mullins effect","feed_subtitle":"Uniaxial, equibiaxial, and planar tension data reproduce stress, energy, and damage in complex 3D simulations.","key_machinery":"The load-bearing object is an invariant-based feed-forward network with two shared-weight blocks and no biases. The hidden layer uses the linear-exponential activation $g(x)=e^{\\alpha x}-1$, so the energy starts at zero in the undeformed state and, with suitable non-negative weights, is automatically non-negative and polyconvex; objectivity and material symmetry enter by construction because the inputs are the invariants $I_1,I_2$ of the right Cauchy-Green tensor. The same block is evaluated at the current invariants and at the invariants of the state where the undamaged energy was maximal, giving $\\psi_0$ and $\\psi_{0,\\max}$, and the Mullins damage factor $1-\\zeta$ multiplies the output. Stress is obtained by differentiating the network output with respect to the invariants, which is what enforces thermodynamic consistency, and the incompressibility pressure is fixed from the plane-stress condition during training and from the finite-element formulation in the 3D examples.","core_discovery":"The central claim is that the damaged strain energy $\\psi = (1-\\zeta)\\psi_0$ can be learned by a single feed-forward network whose undamaged block computes $\\psi_0(I_1,I_2)$ from exponential activations $g(x)=e^{\\alpha x}-1$ acting on weighted combinations of $I_1-3$ and $I_2-3$, while the same block, with shared weights, computes the historical maximum $\\psi_{0,\\max}$ that drives the damage parameter $\\zeta = \\zeta_\\infty(1-e^{-\\psi_{0,\\max}/\\iota})$. The network is trained on stresses only, obtained by automatic differentiation of the energy and the plane-stress Lagrange multiplier, so it never sees energy values during training and yet recovers energy and damage evolution. In the authors' tests the unconstrained variant (no polyconvexity constraints) reproduced the reference Ogden-with-Mullins model with median relative errors around 0.4% in the simple verification tests, about 0.9–1% in the rolling-disc boundary value problem, and about 0.4% in the diabolo tension–torsion problem.","pith_inferences":["Editorial inference: the three training curves trace a two-dimensional surface in the $(I_1,I_2)$ plane, so the method should be understood as strong interpolation inside that envelope; its success on the disc and diabolo examples suggests the tested 3D states stayed within or near the trained region.","Editorial inference: a direct test of the method's boundary would be to run a 3D example whose deformation states clearly fall outside the $I_1 \\le 50$, $I_2 \\le 625$ envelope used for validation; if errors stay small there, the network is genuinely recovering the energy surface rather than interpolating it.","Editorial inference: the same architecture could be adapted to anisotropic damage by adding pseudo-invariants, which the authors note as a future direction; the subnetwork version of the damage law is the piece that makes this straightforward.","Editorial inference: because the network is trained on stress derivatives, its energy accuracy is only pinned at the undeformed state; a user who needs accurate total energies in large deformations should verify the recovered $\\psi_0$ against calorimetric or independent data."],"forward_implications":["If the claim holds, a constitutive model for a new rubber-like material with Mullins damage could be obtained from three standard laboratory tests, without postulating or calibrating a closed-form strain energy function.","The same network can be dropped into existing finite-element software through a UHYPER-style subroutine, since only energy derivatives with respect to invariants are needed.","The damage variable and the undamaged energy are recovered even though the network never sees energy or damage data during training, so the method extracts more from stress curves than a direct stress fit.","Dropping polyconvexity constraints gave the most accurate model in these tests, implying that strict convexity is not required for stable, accurate recovery on the tested domain.","The approach extends naturally to damage laws of unknown form by replacing the exponential damage expression with a small subnetwork whose output is bounded between 0 and 1."],"supporting_citations":[{"why":"Supplies the three-term Ogden model and its constants that generate the training data and serve as the reference solution.","marker":"[27]"},{"why":"Provides the linear-exponential activation function and the constitutive artificial neural network idea that the architecture builds on.","marker":"[32]"},{"why":"Earlier machine-learning treatment of the Mullins effect that this work extends with a single shared-weight network.","marker":"[13]"},{"why":"Documents that strict polyconvexity constraints worsen approximation quality, motivating the paper's unconstrained case 4.","marker":"[18]"},{"why":"Supplies the rolling solid rubber disc benchmark problem used to test generalisation to complex 3D loading.","marker":"[37]"},{"why":"Supplies the diabolo tension-torsion boundary value problem used as the second complex 3D test.","marker":"[38]"}],"fun_headline_variants":["Three tests teach small NN full Mullins damage in 3D","Physics-augmented net learns Mullins from stresses alone","Compact neural net nails Mullins effect across load cases","One NN, three tests: full Mullins hyperelastic damage","Tiny net learns Mullins damage, predicts complex 3D paths"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The weakest link is the assumption that stress data from three plane-stress tests determine the two-invariant strain energy and the damage law over every deformation state the network will later see, including the 3D states in the validation examples; the paper checks accuracy inside its training envelope but gives no proof that states outside it are recovered rather than extrapolated.","fun_headline_variants_meta":{"raw":{"variants":["Three tests teach small NN full Mullins damage in 3D","Physics-augmented net learns Mullins from stresses alone","Compact neural net nails Mullins effect across load cases","One NN, three tests: full Mullins hyperelastic damage","Tiny net learns Mullins damage, predicts complex 3D paths"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1360,"prompt_tokens":936,"completion_tokens":424,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":338}},"tokens_in":552,"tokens_out":424,"duration_ms":4513,"temperature":1.0,"reasoning_tokens":338,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:45:29.959585+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply the trained network to a 3D boundary value problem whose local deformation states fall outside the invariant envelope covered by the training tests — for example, combined torsion and inflation of a thick cylinder where the invariant pair exceeds $I_1 > 50$ or $I_2 > 625$ — and compare the resulting stress, energy, and damage fields against the Ogden-with-Mullins reference solution. If the relative errors grow substantially beyond the ~1% seen inside the envelope, the claim of full 3D recovery would be restricted to interpolation of the training data.","supporting_citations":[{"cited_title":"Ghaderi, V","cited_arxiv_id":null,"evidence_quote":"Earlier machine-learning treatment of the Mullins effect that this work extends with a single shared-weight network."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the rolling solid rubber disc benchmark problem used to test generalisation to complex 3D loading."},{"cited_title":"Chagnon, E","cited_arxiv_id":null,"evidence_quote":"Supplies the diabolo tension-torsion boundary value problem used as the second complex 3D test."}],"review_version":1}