Data-Driven Discovery of Composition-Dependent Constitutive Models for Hyperelasticity and Viscoelasticity of Digital Materials
Josué García-Ávila, Beijun Shen, Manuel K. Rausch, Mary C. Boyce, Adrián Buganza-Tepole
cs.AI · cond-mat.mtrl-sci · cond-mat.soft
Quick verdict
Generalizes Bergström–Boyce viscoelasticity via composition-conditioned NODEs; thermodynamically consistent, benchmarked, but only uniaxial compression tested.
Summary
The paper presents a physics-augmented machine-learning framework for finite-strain viscoelasticity of 3D-printed digital materials — voxel-scale mixtures of Agilus 30 Black and Vero Ultra White spanning apparent stiffness over an order of magnitude. It generalizes the Bergström–Boyce model: multiplicative F = FeFv kinematics with an equilibrium branch and two nonequilibrium branches, invariant-based strain energies, and a scalar flow rule directed along the normalized nonequilibrium deviatoric stress. A composition descriptor Z conditions either closed-form models (neo-Hookean, Fung–Demiray, Arruda–Boyce with MLP-predicted parameters) or neural ODE energies, while the flow law is either a generalized power law or a softplus-bounded positive network. Polyconvexity and non-negative dissipation are enforced architecturally, without penalty terms. Trained on 20 uniaxial compression curves across 5 compositions and 4 strain rates (0.001–0.5 s⁻¹), the NODE and chain-stretch-regularized Arruda–Boyce perform best in leave-one-material-out tests (~10% average validation NRMSE). The learned model emergently separates the two nonequilibrium branches into a fast dissipative mode and a slow energy-storing mode. Main limitations: validation is restricted to uniaxial compression of a single material family; multiaxial generalization is architectural, not experimental; semi-closed flow-law parameters are non-identifiable; and no code or data availability is stated.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A single composition-conditioned framework retaining Bergström–Boyce structure (multiplicative split, invariant energies, stress-direction flow) captures the rate- and composition-dependent response of five digital materials across four strain rates (0.001–0.5 s⁻¹).
- Polyconvexity and non-negative dissipation are guaranteed by construction — NODE monotone flows with softplus offsets and a positive flow-rule network — requiring no penalty terms.
- In leave-one-material-out tests, the NODE gives the lowest mean NRMSE and the chain-stretch-regularized Arruda–Boyce is comparable, with errors averaging about ten percent on the validation set.
- Two nonequilibrium branches with distinct roles — one fast and dissipative, one slow and energy-storing — emerge without imposition and persist across both closed-form and NODE energy representations, suggesting a real constitutive feature.
- Semi-closed power-law flow laws fit the data accurately while their parameters remain non-identifiable, demonstrating that functional-form choice imposes strong assumptions on any physical interpretation.
Strengths
Thermodynamic consistency by architecture, not penalties
NODE non-crossing trajectories plus softplus offsets force ∂ψ/∂I1 ≥ 0 and monotonicity, giving convexity/polyconvexity for the I1-only setting (Sec. 3.1.2); γ̇ = τeff·softplus(·) guarantees γ̇ ≥ 0 hence Dint ≥ 0 (Eqs. 2.18–2.19, 3.21–3.22) without loss terms, unlike ICNN/CANN variants discussed in Sec. 1.
Disciplined model comparison and transfer testing
Leave-one-material-out over all five compositions (Fig. 5A–C), a held-out A60 composition for the full viscoelastic model (Fig. 7), and 2-D loss-landscape sensitivity scans (Fig. 5D, Appendix E); the paper candidly reports that neo-Hookean is the best predictor for A50 and that unregularized Arruda–Boyce and Fung–Demiray extrapolate poorly toward A30.
Emergent, cross-checked branch structure
Candid identifiability treatment
Weaknesses
Uniaxial-only experimental evidence
Frozen equilibrium branch may be biased
No code or data availability
No uncertainty quantification
'Discovery' claim leans on a shared architecture
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.04541v2, García-Ávila et al., Columbia/UT Austin) targets constitutive modeling of 'digital materials' — voxel-scale mixtures of Agilus 30 Black and Vero Ultra White printed on a Stratasys PolyJet J826 (Sec. 5.1, Table 2: A30–A95, Vero volume fraction 0–36.2%). The central claim is a unified, composition-conditioned framework generalizing the Bergström–Boyce finite viscoelasticity model (Secs. 2–3): multiplicative F = FeFv kinematics per branch (equilibrium A plus nonequilibrium B and C), invariant-based strain energies, and a scalar flow γ̇ directed along the normalized nonequilibrium deviatoric stress (Eq. 2.14). Composition Z = [P] conditions either closed-form models (neo-Hookean, Fung–Demiray, Arruda–Boyce with MLP-predicted parameters, Table 1) or NODE energies (Sec. 3.1.2); the flow law is either a generalized power law (Eq. 3.20) or a softplus-bounded positive network (Eqs. 3.21–3.22). The authors claim architectural guarantees of polyconvexity and non-negative dissipation, accurate capture across 5 compositions × 4 strain rates, and emergent discovery of two distinct dissipative mechanisms.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What are 'digital materials' in this paper?
They are voxel-scale mixtures of a soft elastomer (Agilus 30 Black) and a stiff polymer (Vero Ultra White) jetted together on a Stratasys PolyJet J826 printer. By varying the mixing ratio (five mixtures labeled A30 to A95), the effective stiffness and dissipation of the printed part change continuously, spanning more than an order of magnitude in apparent stiffness.
What data does the model train on?
Twenty uniaxial compression load–unload curves: five material compositions (A30, A50, A70, A85, A95) at four engineering strain rates (0.001, 0.01, 0.1, and 0.5 s⁻¹), each cycle reaching about 67% engineering strain. A fresh specimen was printed for each test.
What is the advantage over classical viscoelastic models like Bergström–Boyce?
Instead of calibrating a separate model with fixed closed-form parameters for each material, one shared model is conditioned on the composition vector, so it interpolates to unseen mixtures and rates. Thermodynamic requirements — polyconvex strain energy and non-negative dissipation — are guaranteed by the network architecture itself rather than by penalty terms or manual parameter bounds.
What are the main limitations of the study?
Validation is confined to uniaxial compression of one printer's material family, so multiaxial validity is architectural rather than experimentally demonstrated; parameters of the semi-closed flow laws are shown to be non-identifiable; the equilibrium branch is extracted from the lowest-rate cycle, which still contains some viscous response for the stiffer mixtures; and no code or data release is stated.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Appendices A–E and the supplementary material (Tables S1–S2, Figs. S1–S12) are referenced throughout but absent from the provided text, which ends at the reference list; details stated to reside there (χchain penalty form, training pseudocode, Table 4 variants) could not be verified.
- Figures could not be inspected; assessments of Figs. 3–9 rest on captions and in-text descriptions only.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.