A physics-guided smoothing method for material modeling with digital image correlation (DIC) measurements

Chung-Hao Lee; Jihong Wang; William Richardson; Yue Yu

REVIEW 3 major objections 5 minor 25 references

A physics-guided smoothing method for material modeling with digital image correlation (DIC) measurements

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that adding a positivity constraint on normal strains during reproducing-kernel smoothing of digital image correlation measurements yields physically consistent displacement and strain fields and improves downstream…

desk verdict The PGS smoothing idea is sensible and clearly presented, but the evaluation does not establish that it recovers true deformation fields because the test labels are the same smoothed fields used for training. read the letter →

arxiv 2505.18784 v1 pith:JBBPEUIH submitted 2025-05-24 eess.IV cond-mat.mtrl-scics.LG

Jihong Wang , Chung-Hao Lee , William Richardson , Yue Yu This is my paper · ORCID

classification eess.IVcond-mat.mtrl-scics.LG

keywords digitalimagecorrelationphysics-guidedsmoothingreproducingkernelstrainfielddenoisingconstitutiveoperatorlearningperidynamicneuralbiaxialstretchingbiologicaltissuemechanics

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Digital image correlation (DIC) records full-field displacements of a specimen during biaxial stretching, but the measured fields contain sensor noise and smooth non-physical patterns from operating errors, and these artifacts corrupt strain fields used to learn material behavior. This paper introduces physics-guided smoothing (PGS), an optimization step that reconstructs the displacement field with reproducing-kernel basis functions subject to a penalty on negative normal strains. The paper aims to show that enforcing this partial physical knowledge at the data-processing stage produces displacement and strain fields that are both more physical and more useful for downstream tasks, concretely for learning a nonlocal constitutive law and the underlying fiber microstructure of biological tissue. On two real DIC datasets, a nitrile glove and a porcine tricuspid-valve anterior leaflet, PGS lowers the test relative $L^2$ displacement error of the learned constitutive operator compared with using raw or kernel-smoothed data. A sympathetic reader would care because the preprocessing step is cheap, independent of the learning model, and addresses a gap between DIC measurement and physics-informed machine learning.

What carries the argument

The central object is the PGS optimization problem: reproduce the displacement field as a linear combination of reproducing-kernel (RK) basis functions, then choose the RK coefficients to minimize a hybrid loss, $\text{loss}_u(\bar{u}_I) + \beta \, \text{loss}_E(\bar{u}_I)$. Here $\text{loss}_u$ measures agreement with the DIC measurements and $\text{loss}_E = \sum_J \bigl|\mathrm{ReLU}(-E_{11}[\mathbf{u}](x_J)) + \mathrm{ReLU}(-E_{22}[\mathbf{u}](x_J))\bigr|^2$ penalizes negative Green-Lagrange normal strains; the penalty weight $\beta$ is rescaled by the initial ratio of the two losses so its value is length-scale independent. The RK basis acts as a low-pass filter for random sensor noise, while the strain-positivity term is the physics-informed component that suppresses smooth, non-physical compressive artifacts from operating errors. In the anisotropic valve dataset the same machinery is used protocol-by-protocol: only $E_{11}$ is penalized for stretch ratios $F_x:F_y = 1:1,\,1:0.75,\,1:0.5,\,1:0.25$ and only $E_{22}$ for the three reversed ratios. The cleaned fields then feed a peridynamic neural operator that learns the nonlocal constitutive map.

What would settle it

Run PGS on a synthetic biaxial dataset whose ground-truth deformation is known and includes genuine compressive strains (e.g., from Poisson contraction under unequal tension), then compare the smoothed fields with the ground truth. If PGS systematically erases the true compressive regions or worsens the downstream learned operator compared with unconstrained smoothing, the sign rule is removing real physics rather than artifacts.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that adding a physics-based loss term that penalizes negative normal strains while reconstructing displacements from DIC measurements removes non-physical compressive patterns, and that training a peridynamic neural operator on these cleaned fields reduces errors in the learned constitutive law. In the glove dataset the test relative $L^2$ displacement error drops from 10.01% for the rough DIC data and 9.14% for RK smoothing to 7.35% for PGS; in the heart-valve dataset it drops from 6.21% to 5.11%, and on the challenging subset of originally non-physical samples it drops from 17.85% to 14.68%. The paper also reports that models trained on PGS fields generalize better to unseen loadings and to different mesh resolutions, and that the learned kernel functions reveal interpretable isotropic or anisotropic, heterogeneous microstructure.

Load-bearing premise

The load-bearing premise is that a negative normal strain along the direction of the dominant biaxial stretch is always a non-physical artifact in these experiments, so penalizing it cannot remove real deformation.

Editorial extensions

If this is right

Because the positivity penalty is applied during data preprocessing rather than inside the learning model, PGS can be dropped into existing DIC-to-material-modeling pipelines without changing the downstream architecture.
On the glove dataset, PGS lifts the benefit of smoothing beyond plain RK smoothing: the test error improves by roughly 20% relative to the smoothed baseline and about 27% relative to the rough data.
On the heart-valve dataset, the largest gain appears on the STest subset of originally non-physical samples, where the learned operator trained on PGS data reduces error from 17.85% to 14.68%.
Models trained on PGS-cleaned fields are less sensitive to test-resolution changes, retaining much of their accuracy on 16x16 and 31x31 grids.
The learned peridynamic kernels remain interpretable: isotropic on the glove and anisotropic with spatially varying fiber orientation on the valve leaflet.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural next test is to apply the same positivity-constrained smoothing to synthetic biaxial datasets with known ground-truth strain fields, separating how much of the gain comes from removing noise versus from biasing labels toward positive strain.
The protocol-specific sign rule (penalize $E_{11}$ for one group of stretch ratios and $E_{22}$ for the other) suggests the method can be extended to more general loading protocols, but the sign selection would need to be automated or replaced by inequality constraints informed by an estimated Poisson effect.
Because the penalty weight $\tilde{\beta}$ controls a fidelity-versus-physicality tradeoff, the PGS output is effectively a one-parameter family of datasets; scanning $\tilde{\beta}$ could be used for uncertainty quantification in the downstream constitutive operator.
The same physics-based-loss-during-smoothing idea could transfer to other imaging modalities or to smoothing bases other than reproducing kernels, so long as the constraint is a trustworthy physical invariant.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The method is straightforward to grasp and the paper is honest about its design: RK-based smoothing with a ReLU penalty on negative Green-Lagrange strains, applied as a pre-processing step for downstream constitutive-operator learning. That specific recipe is new, and the ablation over the penalty weight and kernel support is useful. The two real DIC datasets, including the heterogeneous valve tissue, give the paper a practical feel. If the method works as advertised, it would be a cheap, useful fix for a real problem in biomechanics data pipelines.

The soft spot is the evaluation. The headline numbers in Tables 3 and 4 compare downstream errors against PGS-smoothed fields as ground truth — the same kind of field used to train the operator. So a model trained on PGS data naturally fits PGS-shaped targets better than rough or RK-smoothed ones. That is circular. The STest subset on Dataset 2 is meant to fix this, but the ground truth there is not specified. If STest labels are also PGS-processed, the circularity remains. If they are the original noisy DIC fields, then a lower error against noise is not evidence of physical accuracy. Without a synthetic benchmark where the true deformation is known, the central claim that PGS 'extracts' more accurate fields is not actually supported.

The positivity constraint itself is a reasonable assumption for these biaxial protocols, and the paper does acknowledge that true compressive strains can appear under unequal tension. But the protocol-specific sign rule (only E11 for some, only E22 for others) is asserted, not validated. That is a secondary concern after the circularity. The lack of error bars is also notable because the Dataset 2 gain on the overall test set (5.11% vs 5.13%) is within noise. No code or data are provided, which makes it harder to check.

Who is this for? DIC users and researchers working on data-driven constitutive models. The idea deserves a serious look, but the empirical case needs a ground-truth benchmark and error bars. I would send it to peer review, but with the expectation that the authors add synthetic tests with known deformation fields and clarify what STest actually measures.

Referee Report

3 major / 5 minor

Summary. The paper proposes a physics-guided smoothing (PGS) method for digital image correlation (DIC) displacement measurements. The method fits a reproducing-kernel displacement field to the raw DIC data while adding a soft penalty that suppresses negative normal strains (Section 3, Eq. (9)-(10)). The smoothed fields are then used as training labels for a peridynamic neural operator that learns a constitutive law and a fiber orientation field (Section 3, Eqs. (12), (22)-(25)). The authors evaluate PGS on two experimental datasets—a nitrile glove (Dataset 1) and a porcine tricuspid valve anterior leaflet (Dataset 2)—and report that downstream displacement-prediction errors are lower when the operator is trained on PGS-processed data than on raw or RK-smoothed data (Tables 3 and 4).

Significance. If the central claim is established, PGS would be a useful pre-processing tool for DIC-based material modeling, because it combines kernel smoothing with targeted physical constraints and is designed to feed into a downstream operator-learning pipeline. The paper includes useful ablations (Table 2), resolution generalization experiments (Table 3), and interpretable kernel/fiber-orientation visualizations (Figs. 3 and 5). However, the current evaluation does not validate that PGS recovers the true deformation field: the reported test errors are computed against the same pre-processed field type used as training labels, and the only independent physical anchor is a scalar average stress term in the training loss. The study lacks a synthetic ground-truth benchmark, which is needed to support the claim that PGS extracts physically accurate displacement and strain fields.

major comments (3)

[Section 4, Tables 3 and 4] The evaluation is self-referential: for each data type, the relative L2 displacement error is computed against the same type of processed field used as training labels. A model trained on PGS-smoothed fields will naturally fit PGS-smoothed targets better than a model trained on raw or RK-smoothed fields fits its own targets, so the reported reductions (10.01% to 7.35% on Dataset 1; 6.21% to 5.11% on Dataset 2) do not establish that PGS extracts the true deformation. The paper does not specify the ground truth for the STest subset in Table 4; if STest labels are PGS-processed, the comparison remains circular, and if they are the original noisy measurements, a lower error against noise is not evidence of physical accuracy. The authors should validate PGS on synthetic data with known deformation fields (e.g., from finite element simulations with added sensor noise and operator artifacts) and report displacement/strain errors against the true field.
[Section 3, Eq. (9) and Section 4.2] The positivity constraint encodes a strong physical assumption: that negative normal strains, or negative strains in the direction of dominant tension, are always non-physical artifacts. The paper itself acknowledges in Section 2 that physical compressive strains can occur through Poisson contraction when applied tensions differ (references [22,23]). The protocol-specific rule in Section 4.2—penalizing E11 only for the first four loadings and E22 only for the last three—may be reasonable for a homogeneous, orthotropic specimen, but it is not justified for a heterogeneous biological tissue where local deformation can deviate from the global loading direction. If this assumption is violated for any sample, PGS will remove genuine compressive deformation and bias the training labels. The authors should test the sign rule on synthetic or known-field examples that include physically compressive zones, or provide direct experimental evidence that the penalized negative-strain regions are indeed artifacts.
[Tables 1-4 (general reporting)] No error bars, confidence intervals, or repeated random seeds are reported. This is particularly important for Dataset 2, where the overall test improvement of PGS over RK smoothing is marginal (5.13% vs. 5.11%, Table 4) and may not be statistically significant. Since hyperparameters (penalty weight, kernel support size) are selected on the same data, the authors should report mean and standard deviation over multiple training runs and data splits to demonstrate that the reported gains are robust.

minor comments (5)

[Algorithm 1, step 3 and Eq. (11)] Step 3 says 'using the analytical expression in (9)', but Eq. (9) is the physics-based loss; the reference should be to the least-squares solution in Eq. (6) or Eq. (11).
[Eq. (9)] The expression |ReLU(-E11) + ReLU(-E22)|^2 is redundant because both terms are nonnegative; if the intention was to penalize each component independently, the loss should be written as ReLU(-E11)^2 + ReLU(-E22)^2.
[General presentation] The manuscript contains numerous missing spaces and corrupted mathematical symbols (e.g., the abstract reads 'Inthiswork,wepresentanovelapproachtoprocesstheDIC'). A thorough proofread and cleanup of the LaTeX source is needed.
[Tables 1 and 2] The captions do not state that the reported values are relative L2 displacement errors; adding this to the captions would improve clarity.
[Section 4.1, Fig. 2] The claim that PGS 'successfully eliminated negative strains' is illustrated on a representative sample; a quantitative summary (e.g., the percentage of nodes with negative strain before and after PGS across all samples) would make the claim more precise.

Circularity Check

1 steps flagged · score 6.0 of 10

Reported downstream gains are measured against the same PGS-smoothed fields used as training labels, so Tables 3 and 4 do not establish that PGS extracts true deformation.

fitted input called prediction [Section 3 (Downstream constitutive operator learning, Eq. 27) and Tables 3/4 in Section 4]
"Consequently, PGS can effectively extract displacement and strain fields from DIC measurements, inducing smaller errors in downstream constitutive law learning tasks. Given S_tr numbers of displacement/loading function pair samples {(u^s_PGS(x), b^s(x))} ... parameters in t_NN and omega_NN are obtained by minimizing loss= ... ||G^{-1}[b^s]-u^s_PGS||_{L2(Omega)}/||u^s_PGS||_{L2(Omega)} ... Table 3: Dataset 1: Comparison of the relative L2 error of displacements using three types of datasets. Table 4: Dataset 2: Comparison of the relative L2 error of displacements using three types of datasets."

The downstream network is trained to output the PGS-smoothed field u_PGS, and the reported displacement errors are computed on held-out samples of the same dataset types. Since the only target defined in Eq. 27 is the PGS-processed field, the test error measures how well the model reproduces PGS's own smoothed labels, not whether PGS recovered the true deformation. A model trained on PGS labels naturally fits PGS test labels better than models trained on rougher labels, and smoother targets are easier to approximate; thus the cross-method comparison in Tables 3 and 4 conflates the processing method with the ground truth.

full rationale

The PGS optimization itself (Eqs. 5-10) is self-contained: it minimizes data fidelity plus a ReLU penalty on negative strains, and the positivity of the output is a direct consequence of that loss, not a circular derivation. However, the central empirical claim that PGS improves downstream constitutive-law learning is supported by Tables 3 and 4, where the displacement error is computed against the same PGS-smoothed fields used as training labels in Eq. 27. A model trained to output u_PGS will naturally score better on u_PGS test targets than models trained on rougher labels, so the reported error reduction is partly forced by construction. This is a property of the evaluation metric, not a claim about author intent. The STest subset, intended to break this circularity, is selected by a sign criterion but its reference field is not specified, so as written it does not provide an independent anchor. The resolution-transfer and unseen-loading checks are informative but do not compare against known true deformation. Self-citations to the PNO architecture [5,24,25] are not load-bearing for the smoothing claim, and no uniqueness theorem is invoked. The sign-rule assumption about compressive strains is a correctness risk, not a circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on two classes of assumptions: approximation-theoretic assumptions of reproducing-kernel smoothing, and domain assumptions about which strain components are non-physical under biaxial protocols. No new physical entities such as particles, forces, or fields are introduced; the peridynamic kernel and fiber-orientation field are learned representations, not new physical postulates.

free parameters (3)

Penalty weight tilde beta = 100
Chosen from ablation in Table 2 and Figs. 6-7 to balance positivity enforcement against displacement deviation; directly controls the strength of the central PGS correction.
Reproducing kernel support size a = 3.1 dx
Chosen among 2.1, 3.1, and 4.1 dx in Table 1; controls the smoothing level of the reconstructed displacement field.
RK polynomial order n = 1
Set as first-order polynomials in Appendix C.1; affects reproduction accuracy and smoothing behavior.

assumptions (5)

domain assumption Negative normal strains in the direction of dominant biaxial tension are non-physical artifacts.
Invoked in Section 3 lossE and Section 4.2 where E11 is penalized for the first four protocols and E22 for the last three; the paper cites [22,23] for physical Poisson-compressive strains in the opposite component, making this a protocol-specific sign rule that the central claim depends on.
domain assumption DIC measurement noise is high-frequency and can be separated from the true field by reproducing-kernel low-pass filtering.
Section 3.1 uses RK basis and polynomial reproduction conditions (Eq. 4) as a low-pass filter; this is standard in the cited meshfree literature [8].
standard math Reproducing kernel with n-th order polynomial reproduction conditions can represent the displacement field to order n.
Eq. (4); standard approximation theory, not verified in this paper.
domain assumption The peridynamic ordinary-mobile material model (Eqs. 18-20) with the PNO parameterization is expressive enough to represent the true constitutive law of the tested tissues.
Appendix 1.1 and 1.2; the downstream claim that PGS improves constitutive learning inherits this modeling assumption from prior PNO work [5,24].
domain assumption The learned kernel orientation through the rotation matrix in Eq. 23 corresponds to collagen fiber orientation.
Section 4.2 and Fig. 5; used to claim microstructure discovery.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A physics-guided smoothing method for material modeling with digital image correlation (DIC) measurements." pith.science (2026). https://pith.science/paper/JBBPEUIH

@misc{pith2026250518784,
  author       = {Pith},
  title        = {Pith review of: A physics-guided smoothing method for material modeling with digital image correlation (DIC) measurements},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JBBPEUIH}},
  note         = {Machine review of arXiv:2505.18784}
}

read the original abstract

In this work, we present a novel approach to process the DIC measurements of multiple biaxial stretching protocols. In particular, we develop a optimization-based approach, which calculates the smoothed nodal displacements using a moving least-squares algorithm subject to positive strain constraints. As such, physically consistent displacement and strain fields are obtained. Then, we further deploy a data-driven workflow to heterogeneous material modeling from these physically consistent DIC measurements, by estimating a nonlocal constitutive law together with the material microstructure. To demonstrate the applicability of our approach, we apply it in learning a material model and fiber orientation field from DIC measurements of a porcine tricuspid valve anterior leaflet. Our results demonstrate that the proposed DIC data processing approach can significantly improve the accuracy of modeling biological materials.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 22 canonical work pages

[1]

A new damage factor for seismic assess- ment of deficient bare and frp-retrofitted rc structures,

GeorgeMarkou,ReyesGarcia,ChristosMourlas,Mau- rizio Guadagnini, Kypros Pilakoutas, and Manolis Pa- padrakakis, “A new damage factor for seismic assess- ment of deficient bare and frp-retrofitted rc structures,” Engineering Structures, vol. 248, pp. 113152, 2021

work page 2021
[2]

Digital image cor- relation,

Nick McCormick and Jerry Lord, “Digital image cor- relation,” Materials today, vol. 13, no. 12, pp. 52–54, 2010

work page 2010
[3]

Numerical simulationofheterogeneousrockusingdiscreteelement model based on digital image processing,

Xin Tan, Heinz Konietzky, and Wei Chen, “Numerical simulationofheterogeneousrockusingdiscreteelement model based on digital image processing,”Rock Me- chanicsandRockEngineering ,vol.49,pp.4957–4964, 2016

work page 2016
[4]

Predicting the effec- tivemechanicalpropertyofheterogeneousmaterialsby image based modeling and deep learning,

Xiang Li, Zhanli Liu, Shaoqing Cui, Chengcheng Luo, Chenfeng Li, and Zhuo Zhuang, “Predicting the effec- tivemechanicalpropertyofheterogeneousmaterialsby image based modeling and deep learning,”Computer Methods in Applied Mechanics and Engineering, vol. 347, pp. 735–753, 2019

work page 2019
[5]

Heterogeneous Peridynamic Neural Operators: Discover Biotissue Constitutive Law and Microstructure From Digital Image Correlation Measurements

Siavash Jafarzadeh, Stewart Silling, Lu Zhang, Colton Ross, Chung-Hao Lee, SM Rahman, Shuodao Wang, and Yue Yu, “Heterogeneous peridynamic neural op- erators: Discover biotissue constitutive law and mi- crostructure from digital image correlation measure- ments,” arXiv preprint arXiv:2403.18597, 2024

work page Pith review arXiv 2024
[6]

Brief review of image denoising techniques,

Linwei Fan, Fan Zhang, Hui Fan, and Caiming Zhang, “Brief review of image denoising techniques,”Visual Computing for Industry, Biomedicine, and Art, vol. 2, no. 1, pp. 7, 2019

work page 2019
[7]

Strain field de- noisingfordigitalimagecorrelationusingaregularized cost-function,

Bing Pan, Jieyu Yuan, and Yong Xia, “Strain field de- noisingfordigitalimagecorrelationusingaregularized cost-function,” Optics and Lasers in Engineering, vol. 65, pp. 9–17, 2015

work page 2015
[8]

Filters, reproducing kernel, and adaptive meshfree method,

Yang You, J-S Chen, and Hongsheng Lu, “Filters, reproducing kernel, and adaptive meshfree method,” Computational mechanics, vol. 31, no. 3, pp. 316–326, 2003

work page 2003

Show all 25 references

[9]

Knowledge-based modeling of material behavior with neural networks,

J Ghaboussi, JH Garrett Jr, and Xiping Wu, “Knowledge-based modeling of material behavior with neural networks,” Journal of engineering mechanics, vol. 117, no. 1, pp. 132–153, 1991

1991
[10]

Physics- informed machine learning,

George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang, “Physics- informed machine learning,”Nature Reviews Physics, vol. 3, no. 6, pp. 422–440, 2021

2021
[11]

Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equa- tions,

Maziar Raissi, Paris Perdikaris, and George E Karni- adakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equa- tions,” JournalofComputationalPhysics ,vol.378,pp. 686–707, 2019

2019
[12]

Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors,

Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shi- mada, Vladislav Golyanik, Christian Theobalt, and Feng Xu, “Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors,” inProceedings of the IEEE/CVF conference on computer vision...

2022
[13]

High-resolution single- photon imaging with physics-informed deep learning,

LihengBian,HaozeSong,LintaoPeng,XuyangChang, XiYang,RoarkeHorstmeyer,LinYe,ChunliZhu,Tong Qin, Dezhi Zheng, et al., “High-resolution single- photon imaging with physics-informed deep learning,” NatureCommunications,vol.14,no.1,pp.5902,2023

2023
[14]

Neural operator: Graph kernelnetworkforpartialdifferentialequations,

Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar, “Neural operator: Graph kernelnetworkforpartialdifferentialequations,” arXiv preprint arXiv:2003.03485, 2020

2003 arXiv
[15]

Fourier NeuralOperator for Parametric Partial Differential Equations,

Zongyi Li, Nikola Borislavov Kovachki, Kamyar Az- izzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar, “Fourier NeuralOperator for Parametric Partial Differential Equations,” inInterna- tional Conference on Learning Representations, 2020

2020
[16]

Benchmark- ingphysics-informedframeworksfordata-drivenhyper- elasticity,

VahidullahTaç,KevinLinka,FranciscoSahli-Costabal, Ellen Kuhl, and Adrian Buganza Tepole, “Benchmark- ingphysics-informedframeworksfordata-drivenhyper- elasticity,”Computational Mechanics, pp. 1–17, 2023

2023
[17]

Geo- metric deep learning for computational mechanics part i: Anisotropic hyperelasticity,

NikolaosNVlassis,RanMa,andWaiChingSun, “Geo- metric deep learning for computational mechanics part i: Anisotropic hyperelasticity,” Computer Methods in Applied Mechanics and Engineering, vol. 371, pp. 113299, 2020

2020
[18]

Learninghyperelasticanisotropyfromdataviaatensor basis neural network,

Jan N Fuhg, Nikolaos Bouklas, and Reese E Jones, “Learninghyperelasticanisotropyfromdataviaatensor basis neural network,”Journal of the Mechanics and Physics of Solids, vol. 168, pp. 105022, 2022

2022
[19]

A re- view paper: noise models in digital image processing,

Ajay Kumar Boyat and Brĳendra Kumar Joshi, “A re- view paper: noise models in digital image processing,” arXiv preprint arXiv:1505.03489, 2015

2015 arXiv
[20]

Digital image correlation using newton-raphson methodofpartialdifferentialcorrection,

HABruck,SRMcNeill,MichaelASutton,andWHPe- ters, “Digital image correlation using newton-raphson methodofpartialdifferentialcorrection,” Experimental mechanics, vol. 29, pp. 261–267, 1989

1989
[21]

Learningdeepimplicitfourierneural operators (IFNOs) with applications to heterogeneous materialmodeling,

HuaiqianYou,QuinnZhang,ColtonJRoss,Chung-Hao Lee,andYueYu, “Learningdeepimplicitfourierneural operators (IFNOs) with applications to heterogeneous materialmodeling,” ComputerMethodsinAppliedMe- chanics and Engineering, vol. 398, pp. 115296, 2022

2022
[22]

Biaxial me- chanical properties of the natural and glutaraldehyde treatedaorticvalvecusp—parti: experimentalresults,

Kristen L Billiar and Michael S Sacks, “Biaxial me- chanical properties of the natural and glutaraldehyde treatedaorticvalvecusp—parti: experimentalresults,” J. Biomech. Eng., vol. 122, no. 1, pp. 23–30, 2000

2000
[23]

Mechanical characterisation of porcine rectus sheath under uniaxial and biaxial tension,

Mathew Lyons, Des C Winter, and Ciaran K Simms, “Mechanical characterisation of porcine rectus sheath under uniaxial and biaxial tension,”Journal of biome- chanics, vol. 47, no. 8, pp. 1876–1884, 2014

2014
[24]

Peridynamic neural operators: A data-driven nonlocal constitutive model for complex material responses,

Siavash Jafarzadeh, Stewart Silling, Ning Liu, Zhongqiang Zhang, and Yue Yu, “Peridynamic neural operators: A data-driven nonlocal constitutive model for complex material responses,”Computer Methods in Applied Mechanics and Engineering, vol. 425, pp. 116914, 2024

2024
[25]

Monotone peridynamic neural operator for nonlinear materialmodelingwithconditionallyuniquesolutions,

Jihong Wang, Xiaochuan Tian, Zhongqiang Zhang, Stewart Silling, Siavash Jafarzadeh, and Yue Yu, “Monotone peridynamic neural operator for nonlinear materialmodelingwithconditionallyuniquesolutions,” arXiv preprint arXiv:2505.01060, 2025. Appendix of “A physics-guided smoothing...

2025 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

A new damage factor for seismic assess- ment of deficient bare and frp-retrofitted rc structures,

GeorgeMarkou,ReyesGarcia,ChristosMourlas,Mau- rizio Guadagnini, Kypros Pilakoutas, and Manolis Pa- padrakakis, “A new damage factor for seismic assess- ment of deficient bare and frp-retrofitted rc structures,” Engineering Structures, vol. 248, pp. 113152, 2021

work page 2021

[2] [2]

Digital image cor- relation,

Nick McCormick and Jerry Lord, “Digital image cor- relation,” Materials today, vol. 13, no. 12, pp. 52–54, 2010

work page 2010

[3] [3]

Numerical simulationofheterogeneousrockusingdiscreteelement model based on digital image processing,

Xin Tan, Heinz Konietzky, and Wei Chen, “Numerical simulationofheterogeneousrockusingdiscreteelement model based on digital image processing,”Rock Me- chanicsandRockEngineering ,vol.49,pp.4957–4964, 2016

work page 2016

[4] [4]

Predicting the effec- tivemechanicalpropertyofheterogeneousmaterialsby image based modeling and deep learning,

Xiang Li, Zhanli Liu, Shaoqing Cui, Chengcheng Luo, Chenfeng Li, and Zhuo Zhuang, “Predicting the effec- tivemechanicalpropertyofheterogeneousmaterialsby image based modeling and deep learning,”Computer Methods in Applied Mechanics and Engineering, vol. 347, pp. 735–753, 2019

work page 2019

[5] [5]

Heterogeneous Peridynamic Neural Operators: Discover Biotissue Constitutive Law and Microstructure From Digital Image Correlation Measurements

Siavash Jafarzadeh, Stewart Silling, Lu Zhang, Colton Ross, Chung-Hao Lee, SM Rahman, Shuodao Wang, and Yue Yu, “Heterogeneous peridynamic neural op- erators: Discover biotissue constitutive law and mi- crostructure from digital image correlation measure- ments,” arXiv preprint arXiv:2403.18597, 2024

work page Pith review arXiv 2024

[6] [6]

Brief review of image denoising techniques,

Linwei Fan, Fan Zhang, Hui Fan, and Caiming Zhang, “Brief review of image denoising techniques,”Visual Computing for Industry, Biomedicine, and Art, vol. 2, no. 1, pp. 7, 2019

work page 2019

[7] [7]

Strain field de- noisingfordigitalimagecorrelationusingaregularized cost-function,

Bing Pan, Jieyu Yuan, and Yong Xia, “Strain field de- noisingfordigitalimagecorrelationusingaregularized cost-function,” Optics and Lasers in Engineering, vol. 65, pp. 9–17, 2015

work page 2015

[8] [8]

Filters, reproducing kernel, and adaptive meshfree method,

Yang You, J-S Chen, and Hongsheng Lu, “Filters, reproducing kernel, and adaptive meshfree method,” Computational mechanics, vol. 31, no. 3, pp. 316–326, 2003

work page 2003