Pith. sign in

REVIEW 4 major objections 5 minor 98 references

This paper introduces interval and fuzzy physics-augmented neural networks that learn lower, mean, and upper free-energy branches whose stresses enclose noisy stress observations, offering a distribution-free, deterministic route to aleator

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 10:06 UTC pith:KRPUTW4J

load-bearing objection Interesting framework, but the central enclosure theorem has a sign error that invalidates the claim as written; worth a serious referee, but major revision needed. the 4 major comments →

arxiv 2607.20339 v1 pith:KRPUTW4J submitted 2026-07-22 cs.LG physics.comp-ph

Interval and fuzzy physics-augmented neural networks (iPANN and fPANN) for uncertainty quantification and propagation in constitutive modeling

Somesh Pratap Singh , Govinda Anantha Padmanabha , Jingye Tan , Steven Yang , Reese E. Jones , D. Thomas Seidl , Nikolaos Bouklas This is my paper
classification cs.LG physics.comp-ph MSC 74B2068T0774S0574A05
keywords uncertainty quantificationhyperelasticityinterval neural networksfuzzy setsinput convex neural networkssparse regularizationconstitutive modelingfinite element propagation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that noisy stress–deformation data can be turned into a deterministic, inspectable interval of admissible constitutive responses without Bayesian inference or any assumed noise distribution. It does so by learning three sparse, convex free-energy branches (lower, mean, upper) whose automatic-differentiation stresses are forced, through a bound-aware loss, to enclose the observed stresses. If correct, practitioners could obtain certified worst- and best-case stress bounds directly from experiments and plug those bounds into downstream finite element solvers for uncertainty propagation. The paper also proposes a fuzzy extension, in which the three branches are interpolated through alpha-cuts to yield a nested family of responses, letting a user tune conservatism between the mean and the extreme bounds. A sympathetic reader would care because the method is lightweight at inference, needs no sampling, and produces compact closed-form energy expressions.

Core claim

The central claim is that an interval-valued free-energy representation, learned by an input-convex neural network with smoothed L0 sparsification and a two-stage transfer-learning procedure, yields principal-stress bounds that enclose noisy stress observations while preserving objectivity, consistency, and promoting polyconvexity. The key identity is the sign-flipped principal-stress enclosure: because the sign of a principal Green–Lagrange strain flips which branch bounds the stress from above or below, the interval can be anchored to the free-energy ordering and then differentiated to give stress enclosures. On synthetic isotropic hyperelastic data with heteroscedastic noise, the learned

What carries the argument

The central object is the interval free-energy density with its induced principal-stress enclosure, formalized in Proposition 1: if the sign-weighted principal stresses are ordered between lower, mean, and upper branches, then the free energies are ordered, and the interval can be written with a sign flip (Eq. 38) so that positive strains are bounded above by the upper branch and negative strains by the lower branch. This is built into an input-convex neural network (ICNN) that enforces convexity, combined with smoothed L0 sparsification to obtain compact closed-form energy expressions, and a two-stage transfer-learning scheme that trains the mean branch first and then fine-tunes the lower a

Load-bearing premise

The whole training and evaluation pipeline reduces every deformation state to a canonical diagonal right Cauchy–Green tensor, so the learned enclosure is only demonstrated for coaxial, diagonal stress observations and may not enclose general tensor-valued stress data.

What would settle it

Take a trained iPANN and evaluate its bounds on a set of non-diagonal deformation states, e.g., simple shear with F = I + γ e1⊗e2, where the right Cauchy–Green tensor is not diagonal; compare the componentwise predicted stress interval against noisy observations of the full second Piola–Kirchhoff stress tensor. If any off-diagonal stress component or any rotation-dependent linear combination of stresses falls outside the predicted interval, the enclosure claim fails for general deformation states.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Practitioners can obtain certified stress bounds from limited, noisy stress–strain data without distributional assumptions or posterior sampling, making the approach suitable for sparse experimental datasets.
  • The learned closed-form potentials can be inserted directly into finite element solvers in place of a single constitutive model, propagating uncertainty to boundary-value-problem outputs as upper/lower solution fields.
  • The fuzzy alpha-cut representation allows tuning conservatism continuously between the mean and extreme bounds without retraining, and the empirical membership fraction provides a data-driven way to choose alpha.
  • Because the bound width adapts to local noise magnitude, the resulting enclosures are heteroscedastic by construction, reflecting multiplicative or state-dependent noise.
  • Sparsification yields interpretable energy expressions (exponentials, logs, polynomials) that can be inspected and possibly reused in classical hyperelastic model forms.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Since the enclosure is constructed on principal stresses with sign flipping, for non-coaxial stress states (where the stress tensor has off-diagonal components) the componentwise diagonal bounds are not a full tensor enclosure; a spectral or rotationally consistent interval would be needed to extend the guarantee.
  • The empirical-membership-fraction curves suggest a calibration procedure: given a target coverage level, a practitioner could pick the alpha that yields that coverage on a validation set, similar to conformal prediction but without distributional assumptions.
  • The framework could be extended to anisotropic materials by adding structural tensors as inputs, but the sign-flip proposition would need re-derivation because the principal-strain sign ordering no longer holds in an arbitrary material symmetry frame.
  • Testing the trained bounds on a simple shear or other non-diagonal deformation would reveal whether the learned interval generalizes beyond the canonical diagonal representatives used in training.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes interval and fuzzy physics-augmented neural networks (iPANN and fPANN) for uncertainty quantification in hyperelastic constitutive modeling. Three input-convex free-energy branches (lower, mean, upper) are trained from noisy stress data using a combined MSE, smoothed L0 sparsity, and bound-violation loss; the stresses are obtained by automatic differentiation. Fuzzy membership is then introduced by alpha-cut interpolation between the three learned energy branches. Numerical experiments cover synthetic heteroscedastic noise, seed/mean/std perturbations, fuzzy membership analysis, and propagation of the learned potentials in a finite element simulation.

Significance. If the central enclosure claim were correct, the framework would offer a distribution-free, deterministic alternative to Bayesian UQ for constitutive modeling, with sparse, interpretable energy potentials and direct deployment in finite element simulations. The paper has clear strengths: the use of input-convex architectures, smoothed L0 sparsification, a two-stage transfer-learning procedure, a crisp fuzzy interpolation construction, and an FE demonstration. However, the formal stress-enclosure definition contains a sign/order error, and the experiments are restricted to diagonal deformation and stress states. These issues affect the central claim as stated and require correction before the paper can be considered sound.

major comments (4)
  1. [Section 2.4, Eq. (30); also Eq. (38) and Eq. (52)] The interval [sign(E_i) S_i, sign(E_i) \bar{S}_i] is not an enclosure of the observed principal stress S_i when E_i < 0. Under Eq. (19), for E_i<0 one obtains \bar{S}_i ≤ S_i ≤ S_i, so the containing interval is [\bar{S}_i, S_i] (equivalently [min(S_i,\bar{S}_i), max(S_i,\bar{S}_i)]), not the sign-flipped positive interval displayed. Example: E_i=-1, S_i=-2, \bar{S}_i=-4, observed S_i=-3; Eq. (30) gives [2,4], which does not contain -3. The bound loss in Eq. (44) appears to enforce sorted containment, so this may be a misstatement of an otherwise correct implementation, but as written the theorem and the enclosures in Eqs. (30), (38), and (52) are incorrect and must be revised.
  2. [Proposition 1 proof, case ii around Eqs. (25)-(26)] The proof of Proposition 1 is internally inconsistent. From the stated assumption (19) with E_i < 0, the ordering is \bar{S}_i ≤ S_i ≤ S_i, not S_i ≤ S_i ≤ \bar{S}_i ≤ 0 as written. Consequently the differential inequalities in Eq. (26) have the wrong direction, and the conclusion that Ψ ≤ Ψ ≤ \bar{Ψ} does not follow from the argument as presented. The case split needs to be redone with the correct reversed ordering, and the final enclosure formula must be adjusted accordingly.
  3. [Section 4.1, data generation and evaluation] The entire training and evaluation pipeline is restricted to diagonal (coaxial) deformation states. The manuscript states that each selected invariant triple is replaced by a canonical diagonal representative of C, and that evaluating the Gent-Gent model on this diagonal C yields S = diag(S11,S22,S33) with off-diagonal components vanishing identically by construction. The central claim that the learned bounds 'enclose noisy stress observations' is therefore only demonstrated for diagonal stress observables. Eq. (38) is a principal-value interval, not a component-wise enclosure for a general non-diagonal second Piola-Kirchhoff stress tensor. The claims should be scoped to principal/diagonal responses, or new experiments with non-coaxial stress states should be provided.
  4. [Section 4.3.3, empirical membership fraction] The training-set empirical membership fraction is not independent evidence of enclosure, because the bound loss in Eqs. (44)-(45) directly penalizes the same violations that are counted in b_emf(α). The test-set generalization claim rests on visual inspection of Figures 8 and 9, with no numerical coverage values reported. The paper should provide a table of training and test empirical membership fractions at selected α levels (at least α=0 and an intermediate α) for experiments E1-E4, with a quantitative train-test gap, to substantiate the generalization claim.
minor comments (5)
  1. [Abstract and Section 3.1] The language 'promoting polyconvexity' is stronger than what the architecture provides. Section 3.1 explicitly notes that ICNN convexity in the invariants is not full polyconvexity and that additional structural considerations are required. Please qualify the wording in the abstract and introduction.
  2. [Eq. (43)] The smooth Macaulay approximation g(x) may take small negative values near x=0 because of the sigmoid gating. State whether this can cause the bound loss to slightly reward violations and whether it affects the certified-enclosure interpretation.
  3. [Appendix A.2, Table A.2] The hyperparameters differ substantially between E1 (250/150/150 epochs) and E2-E4 (50/50/50 epochs), and E4 uses λ_bound=1000 while E1-E3 use λ_bound=100. Explain whether these differences were tuned separately for each experiment or fixed a priori.
  4. [Figures 8 and 9] The horizontal axis label |a| is used without a definition in the captions. Define |a| in the text or in the caption, and clarify the mapping α = 1 - |a| in the figure itself.
  5. [Appendix C, Table C.3] The notation for the elementary function forms (sp, gsp, φ_k, ψ_ℓ) is terse. Spell out the exact functional forms, e.g., gsp(aI1 - bJ) = exp(aI1 - bJ) + 1, to make the learned energy expressions self-contained.

Circularity Check

1 steps flagged

Training-set enclosure is imposed by the bound loss rather than independently shown; test-set emf and FEM propagation remain out-of-sample, so circularity is partial.

specific steps
  1. fitted input called prediction [Section 3.3, Eqs. (44)–(45) and Section 4.3.3, Eqs. (58)–(59)]
    "The violation variable x is defined for the upper and lower bounds as x = sign(E_i)(S_i − S*_i), upper bound; sign(E_i)(S*_i − S_i), lower bound (Eq. 44). ... b_ii(α) = (1/N) Σ 1{ S^{(n)}_{ii,noisy} ∈ S*_{ii,α}(F^{(n)}) } (Eq. 58)."

    The bound loss L_bound (Eq. 45) penalizes exactly the violation events that the empirical membership fraction b_ii(α) later counts as successes. On the training set, the near-100% emf reported in Figs. 8–9 is therefore produced by the optimizer (λ_bound = 100 in Table A.2) rather than being an independent confirmation of the enclosure claim. The test-set emf and FEM propagation are unaffected by this training loss, so the circularity is partial.

full rationale

The core construction is not definitionally circular: lower, mean, and upper energy branches are trained from stress data, and the test-set membership fraction and FEM propagation are out-of-sample quantities. However, the paper's central evidence that "the learned bounds enclose noisy stress observations" on the training set reduces to the bound-penalty objective: Eq. 44 defines violations of the same sign(E_i)-flipped interval that Eq. 58 counts, so high training emf is enforced rather than discovered. This is a fitted-input-called-prediction pattern, but it is confined to the training portion of the evaluation. No load-bearing self-citation chain was found; citations to [78,64,30,40] concern sparsification and sampling conventions, not the enclosure result. The sign inconsistency in Eq. (30)/Proposition 1 (for E_i<0 the printed interval is positive and cannot contain negative observed stresses) is a correctness defect, not a circularity, and is therefore noted here but not factored into the circularity score.

Axiom & Free-Parameter Ledger

6 free parameters · 8 axioms · 0 invented entities

The central construction rests on a small number of standard mathematical facts and a larger number of modeling choices. The mathematical facts — ICNN convexity, interval arithmetic, alpha-cut reconstruction, path independence of hyperelastic work — are standard. The modeling choices — canonical diagonal data, sign-weighted principal stress ordering, soft Macaulay bound penalty, triangular fuzzy interpolation, two-stage transfer learning — are introduced by the paper and mostly unvalidated against external benchmarks. The learned branch coefficients are fit to data and are therefore not independent contributions.

free parameters (6)
  • Learned free-energy branch coefficients (LB/M/UB) = Tables C.5–C.8, e.g., E1 LB: c1=0.452959, c2=0.457589, c3=1.7295, ...
    All neural network weights and the reported closed-form coefficients are fitted to the noisy stress data; the bound branches are explicitly trained to contain observations via Lbound in Eqs. (44)–(45).
  • Loss weights lambda_mse, lambda_bound, lambda_sparse and epoch counts = Table A.2, e.g., E1: lambda_mse=1, lambda_bound=100, lambda_sparse=5e-3 (mean) / 5e-6 (bounds), epochs 250/150/150
    Selected by grid search on E1 (Appendix A.1); they control the trade-off between fit, sparsity, and enclosure tightness.
  • Smooth Macaulay parameters epsilon and m = Not reported numerically; Fig. 1 illustrates their effect
    Hand-chosen shape parameters in the differentiable bound-penalty g(x) of Eq. (43); they determine how violations are weighted and therefore how tightly bounds can enclose data.
  • Hard-concrete L0 constants gamma, zeta, beta = gamma=-0.1, zeta=1.1, beta=2/3
    Taken from [78,64]; not fitted here, but they set the sparsification schedule and determine which coefficients survive.
  • Network architecture and optimizer = ICNN, 2 hidden layers x 30 neurons, softplus, Adam lr=1e-3; delta=0.2; 500 training points; 1000 test points
    Fixed by the authors; the capacity and sampling range determine what the bounds can represent.
  • Noise controls (mu, sigma, seeds) per experiment = E1: mu=0, sigma=0.1; E2/E3/E4 vary seed/mean/std
    Chosen data-generation inputs, not fitted, but the claimed robustness is only demonstrated for these synthetic perturbations.
axioms (8)
  • standard math ICNN with non-negative W(z) weights and convex non-decreasing activations is convex in its inputs (Amos et al. [28])
    Invoked in Section 3.1 to make each energy branch convex in the chosen invariants.
  • standard math Interval arithmetic for monotone functions and alpha-cut reconstruction of convex fuzzy sets (Sections 2.1–2.2)
    Provides the formal link between interval enclosures and fuzzy membership used throughout the paper.
  • domain assumption Isotropic hyperelastic free energy is objective and expressible via principal invariants; stress-free reference and common energy datum (Eqs. 16–18)
    Needed to define the energy normalization and the stress enclosure at the reference state.
  • domain assumption For every admissible strain state, sign(E_i) S_i <= sign(E_i) S_i <= sign(E_i) S_i (Prop. 1, Eq. 19)
    This ordering is assumed to make the stress interval meaningful and to imply energy ordering; it is enforced only approximately through Lbound, not verified for all states.
  • domain assumption Training and test data can be reduced to canonical diagonal C with diagonal S (Section 4.1)
    The entire numerical study uses diagonal stress tensors; the method is not tested on non-coaxial, non-diagonal stress states.
  • ad hoc to paper Piecewise-linear alpha-interpolation Eqs. (50)–(51) yields a convex fuzzy set with meaningful membership
    No data or theory connects alpha to observed frequency; it is a modeling choice that defines the fPANN.
  • ad hoc to paper Soft Macaulay penalty (Eq. 43) with finite epsilon and m provides a certified enclosure
    Because the bound is a soft penalty, violations can remain; calling the result a certification overstates what the loss guarantees.
  • ad hoc to paper Two-stage transfer learning from mean to bounds improves optimization and preserves sparsity (Section 4.2)
    The warm-start protocol is a heuristic; its benefit is asserted but not compared against joint training.

pith-pipeline@v1.3.0-alltime-deepseek · 24703 in / 17902 out tokens · 140550 ms · 2026-08-01T10:06:22.175694+00:00 · methodology

0 comments
read the original abstract

Constitutive modeling under uncertainty remains a central challenge for reliable mechanics simulations, particularly when the available stress-deformation data are sparse, noisy, or heterogeneous. We propose interval and fuzzy physics-augmented neural networks (iPANNs and fPANNs) for uncertainty-aware hyperelastic constitutive modeling. iPANNs learn sparse lower, mean, and upper free energy density branches whose stresses, obtained by automatic differentiation, ultimately enclose noisy stress observations. In contrast to this deterministic interval description, fPANNs embed the learned iPANN branches into a fuzzy-set representation through alpha-cut interpolation, yielding a nested family of admissible responses. iPANNs and fPANNs encode mechanistic constraints - preserving objectivity, consistency and promoting polyconvexity - and smoothed L0 regularization promotes interpretable energy representations. The bound models are trained through a two-stage transfer-learning procedure in which a sparse mean constitutive response is learned first and then fine-tuned into lower and upper energy branches. We evaluate the framework on synthetic isotropic hyperelastic data with heteroscedastic noise, varying random realizations, shifted noise means, and varying noise magnitudes. The results show that the learned bounds enclose noisy stress observations while generalizing to the test set. Further, we examine the propagation of uncertainty through the mean, upper and lower bound predictions of the learned iPANN models in a finite element setting. The proposed framework provides a compact, physics-consistent route for distribution-free aleatoric uncertainty quantification in hyperelastic constitutive modeling, and propagation in downstream finite element simulations.

Figures

Figures reproduced from arXiv: 2607.20339 by D. Thomas Seidl, Govinda Anantha Padmanabha, Jingye Tan, Nikolaos Bouklas, Reese E. Jones, Somesh Pratap Singh, Steven Yang.

Figure 1
Figure 1. Figure 1: We formulate the constraint in terms of the magnitude of the observables |Si | and of the learned representation of the interval from Eq. 38. The violation variable x is defined for the upper and lower bounds as x =    sign(Ei)(Si − S ∗ i ), upper bound, sign(Ei)(S ∗ i − Si), lower bound, (44) and the smooth Macaulay approximation g(x) is applied in both cases. The bound term for the loss function can b… view at source ↗
Figure 1
Figure 1. Figure 1: Smooth approximation of the Macaulay bracket used in the bound loss. [PITH_FULL_IMAGE:figures/full_fig_p015_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Triangular fuzzy membership function for the learned free energy densities. [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: E1: Stress components (S11, S22, S33) with mean and bounds on training (left) and test (right) data. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: E1: Free energy density Ψ with mean and bounds on training (left) and test data(right) multiple independent noise realizations obtained by changing the random seed of the noise generator (E2), a shift in the noise mean (E3), and a change in the noise standard deviation (E4). These synthetic perturbations are motivated by scenarios encountered in mechanical testing of materials, namely (i) random variations… view at source ↗
Figure 5
Figure 5. Figure 5: E2: Stress components (S11, S22, S33) on training (left) and test (right) data. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: E3: Stress components (S11, S22, S33) on training (left) and test (right) data. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: E4: Stress components (S11, S22, S33) on training (left) and test (right) data. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: E1: Membership analysis. (a) Test S11 data and learned α-cut bounds. (b)–(c) Training and test empirical membership fractions. (d) Absolute train–test gap. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: E2: Membership analysis with five noise realizations. (a) Test S11 data and learned α-cut bounds. (b)–(c) Training and test empirical membership fractions with K = 5. (d) Absolute train–test gap. 4.3.4. Uncertainty propagation in a FEM setting To demonstrate that the learned closed-form free energy density expressions are directly usable in a downstream mechanics simulation, we deploy the three learned pot… view at source ↗
Figure 10
Figure 10. Figure 10: E1: FEM demonstration of learned free energy density. 5. Conclusion This work introduced interval and fuzzy physics-augmented neural network framework for uncertainty-aware hyperelastic constitutive modeling. Relative to widely used Bayesian neu￾ral network approaches, the proposed formulation offers a complementary non-probabilistic route to uncertainty quantification that is comparatively lightweight at… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

98 extracted references · 3 canonical work pages

  1. [1]

    R. S. Rivlin, Large elastic deformations of isotropic materials. i. fundamental concepts, Philosophical Transactions of the Royal Society of London, Series A: Mathematical and Physical Sciences 240 (822) (1948) 459–490

    1948

  2. [2]

    Mooney, A theory of large elastic deformation, Journal of applied physics 11 (9) (1940) 582–592

    M. Mooney, A theory of large elastic deformation, Journal of applied physics 11 (9) (1940) 582–592

    1940

  3. [3]

    R. W. Ogden, Large deformation isotropic elasticity–on the correlation of theory and experiment for incompressible rubberlike solids, Proceedings of the Royal Society of London. A. Mathematical and Physical Sciences 326 (1567) (1972) 565–584

    1972

  4. [4]

    M. E. Gurtin, An introduction to continuum mechanics, Vol. 158, Academic press, 1982

    1982

  5. [5]

    G. A. Holzapfel, Nonlinear solid mechanics: a continuum approach for engineering science (2002)

    2002

  6. [6]

    M. A. Sutton, Computer vision-based, noncontacting deformation measurements in mechanics: A generational transformation, Applied Mechanics Reviews 65 (2013).doi: 10.1115/1.4024984

    work page doi:10.1115/1.4024984 2013

  7. [7]

    M. A. Sutton, F. Hild, Recent advances and perspectives in digital image correlation, Experimental Mechanics 55 (2015) 1–8.doi:10.1007/s11340-015-9991-6

    work page doi:10.1007/s11340-015-9991-6 2015

  8. [8]

    Leclerc, J.-N

    H. Leclerc, J.-N. Périé, S. Roux, F. Hild, Voxel-scale digital volume correlation, Exper- imental mechanics 51 (4) (2011) 479–490

    2011

  9. [9]

    P. J. Withers, C. Bouman, S. Carmignato, V. Cnudde, D. Grimaldi, C. K. Hagen, E. Maire, M. Manley, A. D. Plessis, S. R. Stock, X-ray computed tomography (12 2021).doi:10.1038/s43586-021-00015-4

    doi:10.1038/s43586-021-00015-4 2021

  10. [10]

    Miller, P

    M. Miller, P. Dawson, Understanding local deformation in metallic polycrystals using high energy x-rays and finite elements, Current opinion in solid state and materials science 18 (5) (2014) 286–299. 30

    2014

  11. [11]

    Mathieu, H

    F. Mathieu, H. Leclerc, F. Hild, S. Roux, Estimation of elastoplastic parameters via weighted femu and integrated-dic, Experimental Mechanics 55 (1) (2015) 105–119

    2015

  12. [12]

    Kumar, D

    S. Kumar, D. T. Seidl, B. N. Granzow, J. Yang, J. N. Fuhg, A comparative study of calibration techniques for finite strain elastoplasticity: Numerically-exact sensitivities for femu and vfm, Computer Methods in Applied Mechanics and Engineering 444 (2025) 118159

    2025

  13. [13]

    Pierron, M

    F. Pierron, M. Grédiac, The virtual fields method: extracting constitutive mechani- cal parameters from full-field deformation measurements, Springer Science & Business Media, 2012

    2012

  14. [14]

    J. N. Fuhg, G. A. Padmanabha, N. Bouklas, B. Bahmani, W. C. Sun, N. N. Vlassis, M. Flaschel, P. Carrara, L. D. Lorenzis, A review on data-driven constitutive laws for solids (4 2025).doi:10.1007/s11831-024-10196-2

    doi:10.1007/s11831-024-10196-2 2025

  15. [15]

    Ghaboussi, J

    J. Ghaboussi, J. Garrett Jr, X. Wu, Knowledge-based modeling of material behavior with neural networks, Journal of engineering mechanics 117 (1) (1991) 132–153

    1991

  16. [16]

    Carrara, M

    P. Carrara, M. Ortiz, L. De Lorenzis, Model-free fracture mechanics and fatigue, in: Current trends and open problems in computational mechanics, Springer, 2022, pp. 75–82

    2022

  17. [17]

    Goodfellow, Deep learning (2016)

    I. Goodfellow, Deep learning (2016)

    2016

  18. [18]

    Cybenko, Approximation by superpositions of a sigmoidal function, Mathematics of control, signals and systems 2 (4) (1989) 303–314

    G. Cybenko, Approximation by superpositions of a sigmoidal function, Mathematics of control, signals and systems 2 (4) (1989) 303–314

    1989

  19. [19]

    Hornik, M

    K. Hornik, M. Stinchcombe, H. White, Multilayer feedforward networks are universal approximators, Neural networks 2 (5) (1989) 359–366

    1989

  20. [20]

    J. Park, I. W. Sandberg, Universal approximation using radial-basis-function networks, Neural computation 3 (2) (1991) 246–257

    1991

  21. [21]

    Hornik, Approximation capabilities of multilayer feedforward networks, Neural net- works 4 (2) (1991) 251–257

    K. Hornik, Approximation capabilities of multilayer feedforward networks, Neural net- works 4 (2) (1991) 251–257

    1991

  22. [22]

    M.Schmidt, H.Lipson, Distillingfree-formnaturallawsfromexperimentaldata, science 324 (5923) (2009) 81–85

    2009

  23. [23]

    J. R. Koza, Genetic programming as a means for programming computers by natural selection, Statistics and computing 4 (2) (1994) 87–112

    1994

  24. [24]

    Y. Wang, N. Wagner, J. M. Rondinelli, Symbolic regression in materials science, MRS communications 9 (3) (2019) 793–805. 31

    2019

  25. [25]

    S. L. Brunton, J. L. Proctor, J. N. Kutz, Discovering governing equations from data by sparse identification of nonlinear dynamical systems, Proceedings of the national academy of sciences 113 (15) (2016) 3932–3937

    2016

  26. [26]

    Tibshirani, Regression shrinkage and selection via the lasso, Journal of the Royal Statistical Society Series B: Statistical Methodology 58 (1) (1996) 267–288

    R. Tibshirani, Regression shrinkage and selection via the lasso, Journal of the Royal Statistical Society Series B: Statistical Methodology 58 (1) (1996) 267–288

    1996

  27. [27]

    Raissi, P

    M. Raissi, P. Perdikaris, G. E. Karniadakis, Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations, Journal of Computational physics 378 (2019) 686–707

    2019

  28. [28]

    B.Amos, L.Xu, J.Z.Kolter, Inputconvexneuralnetworks, in: Internationalconference on machine learning, PMLR, 2017, pp. 146–155

    2017

  29. [29]

    Linka, E

    K. Linka, E. Kuhl, A new family of constitutive artificial neural networks towards automated model discovery, Computer Methods in Applied Mechanics and Engineering 403 (2023) 115731

    2023

  30. [30]

    J. N. Fuhg, N. Bouklas, On physics-informed data-driven isotropic and anisotropic constitutive models through probabilistic machine learning and space-filling sampling, Computer Methods in Applied Mechanics and Engineering 394 (2022) 114915

    2022

  31. [31]

    R. G. Patel, R. E. Jones, B. N. Granzow, D. T. Seidl, J. N. Fuhg, A general, auto- mated method for building structural tensors of arbitrary order for anisotropic function representations, Mathematics and Mechanics of Solids (2025) 10812865261429943

    2025

  32. [32]

    D. K. Klein, M. Fernández, R. J. Martin, P. Neff, O. Weeger, Polyconvex anisotropic hyperelasticity with neural networks, Journal of the Mechanics and Physics of Solids 159 (2022) 104703

    2022

  33. [33]

    J. N. Fuhg, A. Jadoon, O. Weeger, D. T. Seidl, R. E. Jones, Polyconvex neural network models of thermoelasticity, Journal of the Mechanics and Physics of Solids 192 (2024) 105837

    2024

  34. [34]

    D. K. Klein, M. Hossain, K. Kikinov, M. Kannapinn, S. Rudykh, A. J. Gil, Neural networks meet hyperelasticity: A monotonic approach, European Journal of Mechanics- A/Solids (2025) 105900

    2025

  35. [35]

    J. N. Fuhg, C. M. Hamel, K. Johnson, R. Jones, N. Bouklas, Modular machine learning- basedelastoplasticity: Generalizationinthecontextoflimiteddata, ComputerMethods in Applied Mechanics and Engineering 407 (2023) 115930

    2023

  36. [36]

    R. E. Jones, A. L. Frankel, K. L. Johnson, A neural ordinary differential equation framework for modeling inelastic stress response via internal state variables, Journal of Machine Learning for Modeling and Computing 3 (3) (2022) 1–35.doi:10.1615/ JMachLearnModelComput.2022042917. URLhttps://doi.org/10.1615/JMachLearnModelComput.2022042917 32

    work page doi:10.1615/jmachlearnmodelcomput.2022042917 2022

  37. [37]

    Steinmann, Geometrical foundations of continuum mechanics, Lecture Notes in Ap- plied Mathematics and Mechanics 2 (2015)

    P. Steinmann, Geometrical foundations of continuum mechanics, Lecture Notes in Ap- plied Mathematics and Mechanics 2 (2015)

    2015

  38. [38]

    R. E. Jones, J. N. Fuhg, A hierarchy of thermodynamics learning frameworks for in- elastic constitutive modeling, arXiv preprint arXiv:2603.02645 (2026)

    Pith/arXiv arXiv 2026

  39. [39]

    Flaschel, S

    M. Flaschel, S. Kumar, L. De Lorenzis, Unsupervised discovery of interpretable hypere- lastic constitutive laws, Computer Methods in Applied Mechanics and Engineering 381 (2021) 113852

    2021

  40. [40]

    J. Tan, G. A. Padmanabha, S. J. Yang, N. Bouklas, Towards rapid constitutive model discovery from multi-modal data: Physics augmented finite element model updating (pafemu), arXiv preprint arXiv:2604.07746 (2026)

    Pith/arXiv arXiv 2026

  41. [41]

    Nemani, L

    V. Nemani, L. Biggio, X. Huan, Z. Hu, O. Fink, A. Tran, Y. Wang, X. Zhang, C. Hu, Uncertainty quantification in machine learning for engineering design and health prog- nostics: A tutorial, Mechanical Systems and Signal Processing 205 (2023) 110796

    2023

  42. [42]

    Fernandez, M

    J. Fernandez, M. Chiachio, J. Chiachio, R. Munoz, F. Herrera, Uncertainty quantifica- tion in neural networks by approximate bayesian computation: Application to fatigue in composite materials, Engineering Applications of Artificial Intelligence 107 (2022) 104511

    2022

  43. [43]

    J. T. Fong, J. J. Filliben, R. dewit, R. J. Fields, B. Bernstein, P. V. Marcal, Uncertainty in finite element modeling and failure analysis: A metrology-based approach (2006)

    2006

  44. [44]

    Cheng, F

    C.-H. Cheng, F. Diehl, G. Hinz, Y. Hamza, G. Nührenberg, M. Rickert, H. Ruess, M. Truong-Le, Neural networks for safety-critical applications—challenges, experiments and perspectives, in: 2018 Design, Automation & Test in Europe Conference & Exhi- bition (DATE), IEEE, 2018, pp. 1005–1006

    2018

  45. [45]

    Abulawi, R

    Z. Abulawi, R. Hu, P. Balaprakash, Y. Liu, Bayesian optimized deep ensemble for un- certainty quantification of deep neural networks: a system safety case study on sodium fast reactor thermal stratification modeling, Reliability Engineering & System Safety 264 (2025) 111353

    2025

  46. [46]

    Begoli, T

    E. Begoli, T. Bhattacharya, D. Kusnezov, The need for uncertainty quantification in machine-assisted medical decision making, Nature Machine Intelligence 1 (1) (2019) 20–23

    2019

  47. [47]

    E.Hüllermeier, W.Waegeman, Aleatoricandepistemicuncertaintyinmachinelearning: An introduction to concepts and methods, Machine learning 110 (3) (2021) 457–506

    2021

  48. [48]

    Ostoja-Starzewski, Random field models of heterogeneous materials, International Journal of Solids and Structures 35 (19) (1998) 2429–2455

    M. Ostoja-Starzewski, Random field models of heterogeneous materials, International Journal of Solids and Structures 35 (19) (1998) 2429–2455. 33

    1998

  49. [49]

    Ostoja-Starzewski, Material spatial randomness: From statistical to representative volume element, Probabilistic engineering mechanics 21 (2) (2006) 112–132

    M. Ostoja-Starzewski, Material spatial randomness: From statistical to representative volume element, Probabilistic engineering mechanics 21 (2) (2006) 112–132

    2006

  50. [50]

    Der Kiureghian, J.-B

    A. Der Kiureghian, J.-B. Ke, The stochastic finite element method in structural relia- bility, Probabilistic engineering mechanics 3 (2) (1988) 83–91

    1988

  51. [51]

    Stefanou, The stochastic finite element method: past, present and future, Computer methods in applied mechanics and engineering 198 (9-12) (2009) 1031–1051

    G. Stefanou, The stochastic finite element method: past, present and future, Computer methods in applied mechanics and engineering 198 (9-12) (2009) 1031–1051

    2009

  52. [52]

    D. J. C. MacKay, A practical bayesian framework for backpropagation networks, Neural Computation 4 (3) (1992) 448–472.doi:10.1162/neco.1992.4.3.448

    doi:10.1162/neco.1992.4.3.448 1992

  53. [53]

    R. M. Neal, Bayesian learning for neural networks, Ph.D. thesis, University of Toronto (1996)

    1996

  54. [54]

    Blundell, J

    C. Blundell, J. Cornebise, K. Kavukcuoglu, D. Wierstra, Weight uncertainty in neural networks, in: Proceedings of the 32nd International Conference on Machine Learning, Vol. 37 of Proceedings of Machine Learning Research, 2015, pp. 1613–1622

    2015

  55. [55]

    Y. Zhu, N. Zabaras, Bayesian deep convolutional encoder–decoder networks for surro- gate modeling and uncertainty quantification, Journal of Computational Physics 366 (2018) 415–447

    2018

  56. [56]

    Gelman, J

    A. Gelman, J. B. Carlin, H. S. Stern, D. B. Rubin, Bayesian data analysis, Chapman and Hall/CRC, 1995

    1995

  57. [57]

    L. V. Jospin, H. Laga, F. Boussaid, W. Buntine, M. Bennamoun, Hands-on bayesian neural networks—a tutorial for deep learning users, IEEE Computational Intelligence Magazine 17 (2) (2022) 29–48.doi:10.1109/MCI.2022.3155327

    arXiv 2022

  58. [58]

    Kendall, Y

    A. Kendall, Y. Gal, What uncertainties do we need in bayesian deep learning for com- puter vision?, in: Advances in Neural Information Processing Systems, Vol. 30, 2017

    2017

  59. [59]

    Graves, Practical variational inference for neural networks, in: Advances in Neural Information Processing Systems, Vol

    A. Graves, Practical variational inference for neural networks, in: Advances in Neural Information Processing Systems, Vol. 24, 2011

    2011

  60. [60]

    Y. Gal, Z. Ghahramani, Dropout as a bayesian approximation: Representing model uncertainty in deep learning, in: Proceedings of the 33rd International Conference on Machine Learning, Vol. 48 of Proceedings of Machine Learning Research, 2016, pp. 1050–1059

    2016

  61. [61]

    Q. Liu, D. Wang, Stein variational gradient descent: A general purpose bayesian infer- ence algorithm, Advances in neural information processing systems 29 (2016)

    2016

  62. [62]

    G. A. Padmanabha, C. Safta, N. Bouklas, R. E. Jones, Concurrent, condensed stein variational gradient descent for uncertainty quantification of neural networks, Journal of Machine Learning for Modeling and Computing 6 (3) (2025). 34

    2025

  63. [63]

    Abdar, F

    M. Abdar, F. Pourpanah, S. Hussain, D. Rezazadegan, L. Liu, M. Ghavamzadeh, P. Fieguth, X. Cao, A. Khosravi, U. R. Acharya, et al., A review of uncertainty quan- tification in deep learning: Techniques, applications and challenges, Information fusion 76 (2021) 243–297

    2021

  64. [64]

    G. A. Padmanabha, J. N. Fuhg, C. Safta, R. E. Jones, N. Bouklas, Improving the performance of stein variational inference through extreme sparsification of physically- constrained neural network models, Computer Methods in Applied Mechanics and En- gineering 432 (2024) 117359

    2024

  65. [65]

    M. C. Kennedy, A. O’Hagan, Bayesian calibration of computer models, Jour- nal of the Royal Statistical Society: Series B (Statistical Methodology) 63 (3) (2001) 425–464.arXiv:https://rss.onlinelibrary.wiley.com/doi/pdf/10.1111/ 1467-9868.00294,doi:https://doi.org/10.1111/1467-9868.00294. URLhttps://rss.onlinelibrary.wiley.com/doi/abs/10.1111/1467-9868.00294

    arXiv 2001

  66. [66]

    Tonini, T

    A. Tonini, T. Bui-Thanh, F. Regazzoni, L. Dede’, A. Quarteroni, Improvements on uncertainty quantification with variational autoencoders, Mathematical Models and Methods in Applied Sciences 36 (04) (2026) 787–823

    2026

  67. [67]

    J. Yi, B. P. Ferreira, M. A. Bessa, Single-to-multi-fidelity history-dependent learning withuncertaintyquantificationanddisentanglement: Applicationtodata-drivenconsti- tutive modeling, Computer Methods in Applied Mechanics and Engineering 448 (2026) 118479

    2026

  68. [68]

    Bahmani, Conformal quantile regression for neural probabilistic constitutive model- ing, Computer Methods in Applied Mechanics and Engineering 457 (2026) 118981

    B. Bahmani, Conformal quantile regression for neural probabilistic constitutive model- ing, Computer Methods in Applied Mechanics and Engineering 457 (2026) 118981

    2026

  69. [69]

    Alefeld, Introduction to interval analysis, SIAM Review 53 (2) (2011) 380

    G. Alefeld, Introduction to interval analysis, SIAM Review 53 (2) (2011) 380

    2011

  70. [70]

    Neumaier, Interval methods for systems of equations, no

    A. Neumaier, Interval methods for systems of equations, no. 37, Cambridge university press, 1990

    1990

  71. [71]

    G. Klir, B. Yuan, Fuzzy sets and fuzzy logic, Vol. 4, Prentice hall New Jersey, 1995

    1995

  72. [72]

    Zimmermann, Fuzzy set theory—and its applications, Springer Science & Business Media, 2011

    H.-J. Zimmermann, Fuzzy set theory—and its applications, Springer Science & Business Media, 2011

    2011

  73. [73]

    Chen, X.-W

    S.-H. Chen, X.-W. Yang, Interval finite element method for beam structures, Finite elements in analysis and design 34 (1) (2000) 75–88

    2000

  74. [74]

    Möller, W

    B. Möller, W. Graf, M. Beer, Fuzzy structural analysis usingα-level optimization, Computational mechanics 26 (6) (2000) 547–565

    2000

  75. [75]

    W. Graf, S. Freitag, J.-U. Sickert, M. Kaliske, Structural analysis with fuzzy data and neural network based material description, Computer-Aided Civil and Infrastructure Engineering 27 (9) (2012) 640–654. 35

    2012

  76. [76]

    Freitag, W

    S. Freitag, W. Graf, M. Kaliske, Recurrent neural networks for fuzzy data, Integrated Computer-Aided Engineering 18 (3) (2011) 265–280

    2011

  77. [77]

    Harazin, J

    F. Harazin, J. Platen, F. N. Schietzold, W. Graf, M. Kaliske, Multiscale polymor- phic uncertainty quantification based on physics-augmented neural networks, Computer Methods in Applied Mechanics and Engineering 452 (2026) 118726

    2026

  78. [78]

    N.Fuhg, R.E

    J. N.Fuhg, R.E. Jones, N.Bouklas, Extremesparsification ofphysics-augmentedneural networks for interpretable model discovery in mechanics, Computer Methods in Applied Mechanics and Engineering 426 (2024) 116973

    2024

  79. [79]

    S. Yang, M. Levin, G. A. Padmanabha, M. Borshevsky, O. Cohen, D. T. Seidl, R. E. Jones, N. Bouklas, N. Cohen, Physics augmented machine learning discovery of composition-dependent constitutive laws for 3d printed digital materials, International Journal of Engineering Science 217 (2025) 104381

    2025

  80. [80]

    J. N. Fuhg, I. Kalogeris, A. Fau, N. Bouklas, Interval and fuzzy physics-informed neural networks for uncertain fields, Probabilistic Engineering Mechanics 68 (2022) 103240

    2022

Showing first 80 references.