Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Equilibrium Conserving Neural Operators for Super-Resolution Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read This paper shows that a high-resolution neural surrogate for solid-mechanics fields can be trained entirely on low-resolution simulation data, with equilibrium and compatibility built into the architecture to supply the missing…

desk verdict ECO is a solid hard-constraint architecture with real super-resolution gains, but the 'low-resolution-only' claim is not actually demonstrated in the pore case and needs a low-frequency check in the poly case. read the letter →

arxiv 2504.13422 v1 pith:UJYDNJJP submitted 2025-04-18 cs.LG physics.comp-ph

Vivek Oommen , Andreas E. Robertson , Daniel Diaz , Coleman Alleman , Zhen Zhang , Anthony D. Rollett , George E. Karniadakis , Rémi Dingreville This is my paper · ORCID
classification cs.LGphysics.comp-ph
keywords physics-informedmachinelearninghardconstraintssuper-resolutionelasticityequilibriumconservationmicromechanicspolycrystallinematerialsneuraloperator
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

This paper tries to overturn a central limitation of neural surrogate solvers in solid mechanics: the need for expensive high-resolution simulation data. It claims that a high-resolution network can be trained using only low-resolution simulations if the governing conservation laws—stress equilibrium and strain compatibility—are enforced directly inside the network architecture rather than merely penalized in the loss. The missing high-frequency content, the argument runs, is not learned from data but supplied by the known physics evaluated on the high-resolution microstructure. If this is right, micromechanical surrogates become roughly two orders of magnitude cheaper to build, while still resolving sharp stress concentrations and extreme strain-energy values that matter for fatigue and damage predictions.

What carries the argument

The load-bearing object is the Equilibrium Conserving Operator (ECO) architecture wrapped around a UNet backbone. A compatibility block enforces kinematic compatibility by producing strains from displacement gradients, e.g. $\epsilon_{ij}=\frac{1}{2}(u_{i,j}+u_{j,i})$, so that any predicted strain field is realizable by a displacement field. An equilibrium block builds a symmetric, divergence-free Cauchy stress from an arbitrary vector field $P=(f,g,h)$ by assembling carefully chosen mixed second derivatives, so the divergence-free condition $\nabla \cdot \sigma = 0$ and the symmetry condition $\sigma = \sigma^T$ hold identically regardless of the network weights. The remaining physics—the constitutive law $\sigma = \mathbb{C}(m)\epsilon$—is left as a soft loss term evaluated at full resolution, while the supervised loss compares the downsampled prediction to the low-resolution simulation. This combination is what forces the network to extrapolate frequency content: the hard blocks rule out nonphysical fields, and the soft constitutive penalty anchors the high-frequency response using only the high-resolution microstructure and stiffness field.

What would settle it

Run s-ECO on a microstructure whose low-resolution simulation is generated by a genuinely under-resolved discretization (for example, a grain boundary thinner than the coarse voxel) rather than by downsampling a converged fine simulation; if the high-resolution prediction fails to match an independent fine DNS, the method depends on the low-resolution data being a faithful low-pass sample, not on physics alone.

Watch

Extended reading notes

Core claim

The central claim is that high-resolution stress and strain fields for elastic micromechanical problems can be learned by a neural surrogate trained exclusively on low-resolution direct numerical simulation data, provided the network architecture itself guarantees satisfaction of the conservation laws. The proposed Equilibrium Conserving Operator (ECO) decomposes the prediction into a compatibility block, which produces strain from a displacement-like vector field, and an equilibrium block, which maps an arbitrary vector field through mixed second derivatives into a symmetric, divergence-free stress tensor. Because equilibrium and compatibility are built in, the training loss only needs the low-resolution supervised terms plus a soft penalty on the constitutive law evaluated at full resolution; the physics supplies the high-frequency information the coarse data cannot. In two case studies (pores in a homogenized matrix and textured polycrystals) the strongly constrained s-ECO matches the high-resolution reference spectra beyond the training grid's Nyquist limit, reaches near-machine-precision stress divergence, and outperforms weakly constrained (w-ECO) and operator-learning baselines, with reported training-cost reductions of roughly two orders of magnitude.

Load-bearing premise

The load-bearing premise is that coarse simulations give an unbiased low-frequency picture of the true high-resolution solution, and that the high-resolution microstructure and stiffness field are accurate, so the physics term can correctly fill in the missing fine detail.

Editorial extensions

If this is right

  • Training a high-resolution surrogate for micromechanics no longer requires high-resolution simulation data: the reported pipeline uses only coarse DNS fields (e.g. $32^3$ or $64^3$) and reaches $64^3$ or $128^3$ predictions, cutting upfront data-collection cost by an estimated factor of 83.
  • Because equilibrium is hard-wired, s-ECO's stress predictions satisfy $\nabla \cdot \sigma = 0$ and $\sigma = \sigma^T$ to near machine precision, so downstream analyses that consume stresses do not inherit solver-level equilibrium error.
  • Frequency spectra of predicted fields follow the high-resolution ground truth beyond the Nyquist frequency of the training grid, meaning sharp features like grain-boundary stress concentrations and extreme strain-energy tails are recoverable rather than oversmoothed.
  • The framework remains stable when the training dataset is substantially reduced and when the resolution jump is increased from $64 \rightarrow 128$ to $32 \rightarrow 128$, suggesting that the physics constraint, not data volume, is the dominant source of accuracy.
  • Under the same conditions, a Tensorized Fourier Neural Operator trained in the standard way severely underestimates high-frequency content, consistent with the paper's argument that operator mesh-independence does not imply super-resolution for non-band-limited micromechanical fields.

Reading between the lines

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

  • Editorial: the method implicitly claims a conditional uniqueness result—that the low-frequency solution plus the constitutive law determine the high-frequency solution; a clean test would apply s-ECO to a problem with a known closed-form high-resolution solution and check whether the recovered spectrum matches exactly.
  • Editorial: if the argument holds, the same hard-constraint substitution should work for any conservation law expressible as a linear differential operator (heat flux balance, incompressibility, and similar), because the equilibrium block depends only on the operator structure, not on the constitutive model.
  • Editorial: the reported sensitivity to microstructure representation suggests the framework's guarantees are tied to how faithfully the input stiffness field represents the true material; one could test robustness by corrupting the input stiffness field with controlled noise and measuring how much high-resolution fidelity degrades.
  • Editorial: since the soft constitutive loss is the only high-frequency anchor, a material with an unknown or approximate constitutive law is a stress test; using a deliberately wrong stiffness tensor during training would quantify how much of the recovered high frequencies are physics-driven versus learned.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes ECO, a UNet-based super-resolution surrogate for micromechanical fields. Two variants are compared: s-ECO strongly enforces stress equilibrium and strain compatibility by construction and penalizes the constitutive law softly, while w-ECO enforces the constitutive law exactly and penalizes stress divergence softly. Training loss matches downsampled high-resolution predictions against low-resolution simulated fields, with an additional physics-based loss. The method is evaluated on embedded pores and polycrystalline microstructures. The authors claim that high-resolution surrogates can be trained with only low-resolution data, recover frequency content beyond the training grid's Nyquist limit, and reduce data collection cost by roughly 83x.

Significance. If substantiated, the result would be significant: it would break the need for high-resolution datasets in micromechanical surrogate modeling and provides a construction that exactly enforces conservation laws. The equilibrium block (Eq. 12) is mathematically clean, and the near-machine-precision stress divergence of s-ECO predictions (Fig. 4) is a clear strength. The comparison against a plain UNet and TFNO is informative. However, the evidence currently does not fully support the headline claim: the pore data are not true low-resolution simulations, and the polycrystalline experiment used high-resolution simulations for hyperparameter selection. The significance is therefore conditional on additional validation.

major comments (4)
  1. [Datasets and Problem Descriptions (embedded pores)] The pore experiment does not demonstrate training on true low-resolution data. The dataset section states that training and validation simulations were run at 128^3 and then downsampled to 64^3 for supervision, while test simulations were run at 256^3 and downsampled to 128^3. A downsampled 128^3 solution is a low-pass projection of the high-resolution solution by construction; a genuine 64^3 DNS has its own discretization error. The load-bearing premise of Eq. (1) is that D(f_theta(m)) is supervised against a field that is a faithful low-frequency representation of the high-resolution truth, and the manuscript never checks this for real coarse DNS. Please run a subset of true 64^3 simulations and compare their low-frequency content and effective stiffness with the downsampled 128^3 fields; if the biases are small, quantify and report them, and if they are not, state what the constitutive penalty in Eq. (13) must overcome.
  2. [Implementation Specifics (polycrystalline hyperparameter optimization)] The paper's central claim of training with only low-resolution data is not consistent with the hyperparameter procedure. In the polycrystalline implementation, the loss weights alpha, beta, gamma, dropout rate, peak learning rate, initial lift channels, channel factors, and weight decay were optimized with Optuna over a validation dataset of five high-resolution simulations. These high-resolution simulations were excluded from the main results, but they were still used to select the final model. The abstract's 'only low-resolution data' and 'eliminates reliance on high-fidelity data' therefore do not describe what was done. Either repeat the selection using only low-resolution validation data, or weaken the claim to low-resolution training data with high-resolution validation for hyperparameters.
  3. [Table 1 and Polycrystalline Microstructures section] The reported errors are too high to support 'high fidelity' or 'good fidelity' for the polycrystalline case. Table 1 lists nRMSE values of 0.18-0.49 for almost all components, e.g., sigma_11 0.44 (w-ECO) and 0.49 (s-ECO), sigma_23 0.28/0.30, and F^e_23 0.38/0.40. These numbers are not contextualized against a baseline: what is the nRMSE of the DNS:32 field itself when compared with DNS:64? If the low-resolution supervision already differs from the high-resolution reference by a comparable amount, then the method is doing no more than reproducing the coarse solution. Please report this baseline alongside Table 1 and state explicitly which components are 'high fidelity.'
  4. [Figure 6 and Datasets (polycrystalline)] No quantitative low-frequency validation is provided for the polycrystalline case. Figure 6 shows radially averaged power spectra and demonstrates high-frequency recovery, but it does not report a low-frequency error metric. In particular, there is no check that the low-frequency content of DNS:32 (|k| < 16) agrees with the corresponding low-passed DNS:64/DNS:128 fields. If coarse DNS is biased at low frequencies, Eq. (1) anchors the wrong modes and the hard equilibrium/compatibility blocks cannot fix that bias. Add, for a representative subset, the relative L2 difference between low-pass DNS:32 and DNS:64 fields and the low-frequency log-ratio of the power spectra; this is needed to support the claim that training on DNS:32 is equivalent to training on the low-frequency content of DNS:64.
minor comments (5)
  1. [Discussion] The opening sentence of the Discussion contains a typo: 'Equilibirium Conserving Operator' should be 'Equilibrium Conserving Operator.'
  2. [Results, Eq. (1) discussion] In the sentence beginning 'R is the dimensionality of the microstructure representation at each voxel,' the text says 'therefore H = 1'; H is undefined and should read R = 1.
  3. [Methods, polycrystalline implementation] The text refers to the tabulated quantities as elastic strains, but for the polycrystalline case the table lists the elastic deformation gradient F^e; please use consistent terminology throughout.
  4. [Ultra-High Resolution Evaluation] For the 128^3 evaluation, the text reports 'relative absolute errors generally below 20%' but no aggregate nRMSE or MAE for this experiment is given; a quantitative entry in Table 1 or a new table would allow comparison with the 64^3 results.
  5. [Fig. 6 caption and surrounding text] The notation DNS:HR, DNS:64, DNS:128, and TFNO:64 is used interchangeably; define these once, for example in a resolution table in Methods, to improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the physics blocks are derived from the PDEs, the constitutive penalty uses known material laws, and the high-resolution predictions are checked against held-out DNS.

full rationale

The derivation chain is not circular. Equation (1) supervises the network by downsampling the high-resolution prediction and comparing it with low-resolution fields; this is a standard low-pass constraint, not a restatement of the high-resolution target. Equation (13) adds a soft constitutive-law penalty at full resolution, and that law C(m)epsilon = sigma is a known material model supplied independently of the training data, so the high-frequency information is not being renamed as a prediction; it is being induced by the constitutive relation applied to the high-resolution microstructure input. The s-ECO equilibrium block (Eq. 12) constructs a symmetric, divergence-free stress from second derivatives of arbitrary potentials, which is a hard PDE constraint built into the architecture, and the compatibility block (Eq. 9) derives strain from displacement; neither block is fit to the target fields. The paper tests the resulting high-resolution fields against held-out DNS:HR and against external baselines (TFNO, plain UNet), so the central super-resolution claim is externally falsifiable rather than forced by construction. The embedded-pore dataset was generated by running 128^3 simulations and downsampling them to 64^3 for supervision, and a small set of high-resolution simulations was used for hyperparameter optimization; these are legitimate concerns about the strength of the "only low-resolution data" claim and about possible coarse-DNS bias, but they are not cases where a predicted quantity reduces to the fitting input by construction. Self-citations (e.g., Refs. 43, 48, 52) are used for context or background and are not load-bearing: the oversmoothing claim is also demonstrated empirically with the TFNO baseline. Therefore no circular step meeting the quotation-and-reduction bar was found.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on known continuum physics, not on newly invented physical entities. The main carried assumptions are representational completeness of the stress-function block, the reliability of coarse-simulation labels, the availability of accurate high-resolution microstructures and stiffness, and the sufficiency of a soft constitutive penalty. The fitted items are training and loss hyperparameters, which affect quantitative performance but do not by themselves generate the claimed physics.

free parameters (4)
  • s-ECO loss weights (alpha, beta, gamma) = alpha=4.7068, beta=1.3258, gamma=1.7297
    Chosen by Optuna on a validation set; they balance the supervised low-resolution losses and the soft constitutive penalty, and changes in them affect the reported accuracy.
  • w-ECO loss weights (alpha, beta, gamma) = alpha=4.7068, beta=100.0, gamma=1.7297
    Chosen by Optuna; the large beta weights the weakly enforced divergence penalty.
  • Kaiming initialization gain = 0.5 for s-ECO, 0.2 for w-ECO
    Tuned to match the marginal distribution of stress and deformation-gradient fields at initialization, which affects training stability.
  • Other training hyperparameters = dropout=0.1513, weight_decay=0.0165, batch_size=16, warmup=100 epochs, peak LR=5e-4, final LR=1e-5
    Tuned through Optuna; needed to reproduce the reported performance but not directly part of the physics claim.
assumptions (5)
  • domain assumption Maxwell stress-function representation in Eq. 12 yields all symmetric divergence-free stress fields needed for the solved problems.
    The s-ECO equilibrium block outputs only tensors of this form, so if some true stress field cannot be represented this way in the discrete periodic setting, the architecture cannot reproduce it. The paper does not prove completeness for its finite-difference implementation.
  • domain assumption Low-resolution DNS fields are unbiased low-frequency supervision for the high-resolution solution.
    Eq. 1 treats the low-resolution stress and strain fields as the target after downsampling; if coarse simulations are biased by discretization or by the different voxelization of microstructures, the supervised signal is wrong.
  • domain assumption The high-resolution microstructure m(x) and the stiffness field C(m(x)) are known and accurate.
    The network receives high-resolution microstructure as input and uses it to evaluate the constitutive penalty at full resolution. Errors in m(x) or C(m(x)) directly corrupt the recovered high-frequency content.
  • ad hoc to paper The soft constitutive penalty with weight beta is sufficient to make the predicted stress and strain physically consistent.
    s-ECO does not strongly enforce the constitutive law; the paper states that an architecture satisfying both conservation laws and constitutive relations is deferred to future work. The method assumes the soft penalty is strong enough for the target applications.
  • standard math Finite-difference stencils at the operating voxel resolution faithfully approximate the continuous divergence and compatibility operators.
    The equilibrium and compatibility blocks use finite differences on a voxel grid; the claimed 'machine precision' equilibrium is measured on that discrete operator, not on the continuous PDE.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Equilibrium Conserving Neural Operators for Super-Resolution Learning." pith.science (2026). https://pith.science/paper/UJYDNJJP

@misc{pith2026250413422,
  author       = {Pith},
  title        = {Pith review of: Equilibrium Conserving Neural Operators for Super-Resolution Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UJYDNJJP}},
  note         = {Machine review of arXiv:2504.13422}
}
read the original abstract

Neural surrogate solvers can estimate solutions to partial differential equations in physical problems more efficiently than standard numerical methods, but require extensive high-resolution training data. In this paper, we break this limitation; we introduce a framework for super-resolution learning in solid mechanics problems. Our approach allows one to train a high-resolution neural network using only low-resolution data. Our Equilibrium Conserving Operator (ECO) architecture embeds known physics directly into the network to make up for missing high-resolution information during training. We evaluate this ECO-based super-resolution framework that strongly enforces conservation-laws in the predicted solutions on two working examples: embedded pores in a homogenized matrix and randomly textured polycrystalline materials. ECO eliminates the reliance on high-fidelity data and reduces the upfront cost of data collection by two orders of magnitude, offering a robust pathway for resource-efficient surrogate modeling in materials modeling. ECO is readily generalizable to other physics-based problems.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Global Stress Generation and Spatiotemporal Super-Resolution Physics-Informed Operator under Dynamic Loading for Two-Phase Random Materials

    cs.LG 2025-04 conditional novelty 5.0 of 10

    Two machine learning stages, a space-time diffusion generator and a physics-informed super-resolution operator, produce and refine dynamic stress fields for two-phase random materials, with relative errors around 1 to...

Reference graph

Works this paper leans on

66 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [1]

    Overview of constitutive laws, kinematics, homogenization and multiscale methods in crystal plasticity finite-element modeling: Theory, experiments, applications

    Franz Roters, Philip Eisenlohr, Luc Hantcherli, Denny Dharmawan Tjahjanto, Thomas R Bieler, and Dierk Raabe. Overview of constitutive laws, kinematics, homogenization and multiscale methods in crystal plasticity finite-element modeling: Theory, experiments, applications. Acta Mater., 58(4):1152–1211, 2010

    work page 2010

  2. [2]

    Crystal plasticity simulation study on the influence of texture on earing in steel

    D Raabe, Y Wang, and F Roters. Crystal plasticity simulation study on the influence of texture on earing in steel. Computational Materials Science, 34(3):221–234, 2005

    work page 2005

  3. [3]

    The Finite Element Method: Linear Static and Dynamic Finite Element Analysis

    Thomas JR Hughes. The Finite Element Method: Linear Static and Dynamic Finite Element Analysis. Courier Corporation, 2003

    work page 2003

  4. [4]

    M. Little. NASA high-end computing (HEC) needs assessment for 2020. Technical report, NASA High-End Computing Program, 2020

    work page 2020

  5. [5]

    Green algorithms: quantifying the carbon footprint of computation

    Loïc Lannelongue, Jason Grealey, and Michael Inouye. Green algorithms: quantifying the carbon footprint of computation. Advanced Science, 8(12):2100707, 2021

    work page 2021

  6. [6]

    MPI+X: task-based parallelisation and dynamic load balance of finite element assembly

    Marta Garcia-Gasulla, Guillaume Houzeaux, Roger Ferrer, Antoni Artigues, Victor López, Jesús Labarta, and Mariano Vázquez. MPI+X: task-based parallelisation and dynamic load balance of finite element assembly. International Journal of Computational Fluid Dynamics, 33(3):115–136, 2019

    work page 2019

  7. [7]

    A heterogeneous parallel model of unstructured mesh finite element method based on CPU+GPU.Highlights in Science, Engineering and Technology, 77:173–178, 2023

    Yu Lei, Guoliang Peng, Yinjun Gao, Feng Han, and Dong Wang. A heterogeneous parallel model of unstructured mesh finite element method based on CPU+GPU.Highlights in Science, Engineering and Technology, 77:173–178, 2023

    work page 2023

  8. [8]

    Continuum Scale Simulation of Engineering Materials: Fundamentals – Microstructures – Process Applications

    Dierk Raabe, Franz Roters, Frédéric Barlat, and Long-Qing Chen. Continuum Scale Simulation of Engineering Materials: Fundamentals – Microstructures – Process Applications. John Wiley & Sons, 2004

    work page 2004

Show all 66 references
  1. [9]

    Microstructure-sensitive computational structure-property relations in materials design

    David L McDowell. Microstructure-sensitive computational structure-property relations in materials design. In D. Shin and J. Saal, editors, Computational Materials System Design. Springer, Cham, 2018

    2018

  2. [10]

    Microstructure sensitive design for performance optimization

    David T Fullwood, Stephen R Niezgoda, Brent L Adams, and Surya R Kalidindi. Microstructure sensitive design for performance optimization. Progress in Materials Science, 55(6):477–562, 2010

    2010

  3. [11]

    Neural operators for accelerating scientific simulations and design

    Kamyar Azizzadenesheli, Nikola Kovachki, Zongyi Li, Miguel Liu-Schiaffini, Jean Kossaifi, and Anima Anand- kumar. Neural operators for accelerating scientific simulations and design. Nature Reviews Physics, 6:320–328, 2024

    2024

  4. [12]

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

    Maziar Raissi, Paris Perdikaris, and George 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, 378:686–707, 2019

    2019

  5. [13]

    Fourier neural operator for parametric partial differential equations, 2020

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations, 2020. arXiv:2010.08895

    2020 arXiv

  6. [14]

    Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators. Nature Machine Intelligence, 3(3):218–229, 2021

    2021

  7. [15]

    Laplace neural operator for solving differential equations

    Qianying Cao, Somdatta Goswami, and George Em Karniadakis. Laplace neural operator for solving differential equations. Nature Machine Intelligence, 6(6):631–640, 2024

    2024

  8. [16]

    Wavelet neural operator for solving parametric partial differential equations in computational mechanics problems

    Tapas Tripura and Souvik Chakraborty. Wavelet neural operator for solving parametric partial differential equations in computational mechanics problems. Computer Methods in Applied Mechanics and Engineering, 404:115783, 2023

    2023

  9. [17]

    Towards multi-spatiotemporal-scale generalized PDE modeling, 2022

    Jayesh K Gupta and Johannes Brandstetter. Towards multi-spatiotemporal-scale generalized PDE modeling, 2022. arXiv:2209.15616

    2022 arXiv

  10. [18]

    Ditto: Diffusion-inspired temporal transformer operator, 2023

    Oded Ovadia, Eli Turkel, Adar Kahana, and George Em Karniadakis. Ditto: Diffusion-inspired temporal transformer operator, 2023. arXiv:2307.09072

    2023 arXiv

  11. [19]

    Thermodynamically-informed iterative neural operators for heterogeneous elastic localization, 2024

    Conlain Kelly and Surya R Kalidindi. Thermodynamically-informed iterative neural operators for heterogeneous elastic localization, 2024. arXiv:2411.06529

    2024 arXiv

  12. [20]

    Learning deep implicit Fourier neural operators (IFNOs) with applications to heterogeneous material modeling

    Huaiqian You, Quinn Zhang, Colton J Ross, Chung-Hao Lee, and Yue Yu. Learning deep implicit Fourier neural operators (IFNOs) with applications to heterogeneous material modeling. Computer Methods in Applied Mechanics and Engineering, 398:115296, 2022

    2022

  13. [21]

    Recurrent localization networks applied to the lippmann-schwinger equation

    Conlain Kelly and Surya R Kalidindi. Recurrent localization networks applied to the lippmann-schwinger equation. Computational Materials Science, 192:110356, 2021. 19 Equilibrium Conserving Neural Operators

    2021

  14. [22]

    Prediction of local elasto-plastic stress and strain fields in a two-phase composite microstructure using a deep convolutional neural network

    Indrashish Saha, Ashwini Gupta, and Lori Graham-Brady. Prediction of local elasto-plastic stress and strain fields in a two-phase composite microstructure using a deep convolutional neural network. Computer Methods in Applied Mechanics and Engineering, 421:116816, 2024

    2024

  15. [23]

    Computational Methods for Microstructure-Property Relationships, volume

    Somnath Ghosh and Dennis Dimiduk. Computational Methods for Microstructure-Property Relationships, volume

  16. [24]

    An artificial neural network for surrogate modeling of stress fields in viscoplastic polycrystalline materials

    Mohammad S Khorrami, Jaber R Mianroodi, Nima H Siboni, Pawan Goyal, Bob Svendsen, Peter Benner, and Dierk Raabe. An artificial neural network for surrogate modeling of stress fields in viscoplastic polycrystalline materials. npj Computational Materials, 9(1):37, 2023

    2023

  17. [25]

    Learning the solution operator of parametric partial differential equations with physics-informed DeepONets

    Sifan Wang, Hanwen Wang, and Paris Perdikaris. Learning the solution operator of parametric partial differential equations with physics-informed DeepONets. Science Advances, 7(40):eabi8605, 2021

    2021

  18. [26]

    Physics-informed neural operator for learning partial differential equations

    Zongyi Li, Hongkai Zheng, Nikola Kovachki, David Jin, Haoxuan Chen, Burigede Liu, Kamyar Azizzadenesheli, and Anima Anandkumar. Physics-informed neural operator for learning partial differential equations. ACM/JMS Journal of Data Science, 1(3):1–27, 2024

    2024

  19. [27]

    Physics-informed deep neural operator networks

    Somdatta Goswami, Aniruddha Bora, Yue Yu, and George Em Karniadakis. Physics-informed deep neural operator networks. In Machine Learning in Modeling and Simulation: Methods and Applications, pages 219–254. Springer, 2023

    2023

  20. [28]

    From PINNs to PIKANs: Recent advances in physics-informed machine learning, 2024

    Juan Diego Toscano, Vivek Oommen, Alan John Varghese, Zongren Zou, Nazanin Ahmadi Daryakenari, Chenxi Wu, and George Em Karniadakis. From PINNs to PIKANs: Recent advances in physics-informed machine learning, 2024. arXiv:2410.13228

    2024 arXiv

  21. [29]

    Multifidelity modeling for physics- informed neural networks (PINNs)

    Michael Penwarden, Shandian Zhe, Akil Narayan, and Robert M Kirby. Multifidelity modeling for physics- informed neural networks (PINNs). Journal of Computational Physics, 451:110844, 2022

    2022

  22. [30]

    Physics-informed deep learning for 3D modeling of light diffraction from optical metasurfaces

    Vlad Medvedev, Andreas Erdmann, and Andreas Rosskopf. Physics-informed deep learning for 3D modeling of light diffraction from optical metasurfaces. Optics Express, 33(1):1371–1384, 2025

    2025

  23. [31]

    On the eigenvector bias of Fourier feature networks: From regression to solving multi-scale PDEs with physics-informed neural networks

    Sifan Wang, Hanwen Wang, and Paris Perdikaris. On the eigenvector bias of Fourier feature networks: From regression to solving multi-scale PDEs with physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering, 384:113938, 2021

    2021

  24. [32]

    Learning a neural solver for parametric PDE to enhance physics-informed methods, 2024

    Lise Le Boudec, Emmanuel De Bézenac, Louis Serrano, Ramon Daniel Regueiro-Espino, Yuan Yin, and Patrick Gallinari. Learning a neural solver for parametric PDE to enhance physics-informed methods, 2024. arXiv:2410.06820

    2024

  25. [33]

    Physics- informed neural networks with hard constraints for inverse design

    Lu Lu, Raphael Pestourie, Wenjie Yao, Zhicheng Wang, Francesc Verdugo, and Steven G Johnson. Physics- informed neural networks with hard constraints for inverse design. SIAM Journal Scientific Computing , 43(6):B1105–B1132, 2021

    2021

  26. [34]

    Beyond finite layer neural networks: Bridging deep architectures and numerical differential equations

    Yiping Lu, Aoxiao Zhong, Quanzheng Li, and Bin Dong. Beyond finite layer neural networks: Bridging deep architectures and numerical differential equations. In International Conference on Machine Learning , pages 3276–3285. PMLR, 2018

    2018

  27. [35]

    Hamiltonian neural networks

    Samuel Greydanus, Misko Dzamba, and Jason Yosinski. Hamiltonian neural networks. Advances in Neural Information Processing Systems, 32, 2019

    2019

  28. [36]

    Machine learning structure preserving brackets for forecasting irreversible processes

    Kookjin Lee, Nathaniel Trask, and Panos Stinis. Machine learning structure preserving brackets for forecasting irreversible processes. Advances in Neural Information Processing Systems, 34:5696–5707, 2021

    2021

  29. [37]

    Structure-preserving sparse identification of nonlinear dynamics for data-driven modeling

    Kookjin Lee, Nathaniel Trask, and Panos Stinis. Structure-preserving sparse identification of nonlinear dynamics for data-driven modeling. In Mathematical and Scientific Machine Learning, pages 65–80. PMLR, 2022

    2022

  30. [38]

    GFINNs: GENERIC formalism informed neural networks for deterministic and stochastic dynamical systems

    Zhen Zhang, Yeonjong Shin, and George Em Karniadakis. GFINNs: GENERIC formalism informed neural networks for deterministic and stochastic dynamical systems. Philosophical Transactions of the Royal Society A, 380(2229):20210207, 2022

    2022

  31. [39]

    Structure preserving neural networks and applications to optimal control problems

    Zhen Zhang, Pengzhan Jin, and George Em Karniadakis. Structure preserving neural networks and applications to optimal control problems. In Fall Eastern Sectional Meeting. AMS, 2022

    2022

  32. [40]

    Convolutional neural operators

    Bogdan Raonic, Roberto Molinaro, Tobias Rohner, Siddhartha Mishra, and Emmanuel de Bezenac. Convolutional neural operators. In ICLR 2023 Workshop on Physics for Machine Learning, 2023

    2023

  33. [41]

    Universal physics transformers: A framework for efficiently scaling neural operators

    Benedikt Alkin, Andreas Fürst, Simon Schmid, Lukas Gruber, Markus Holzleitner, and Johannes Brandstetter. Universal physics transformers: A framework for efficiently scaling neural operators. Advances in Neural Information Processing Systems, 37:25152–25194, 2024. 20 Equilibri...

    2024

  34. [42]

    A spectral-based physics-informed finite operator learning for prediction of mechanical behavior of microstructures, 2024

    Ali Harandi, Hooman Danesh, Kevin Linka, Stefanie Reese, and Shahed Rezaei. A spectral-based physics-informed finite operator learning for prediction of mechanical behavior of microstructures, 2024. arXiv:2410.19027

    2024 arXiv

  35. [43]

    Integrating neural operators with diffusion models improves spectral representation in turbulence modeling, 2024

    Vivek Oommen, Aniruddha Bora, Zhen Zhang, and George Em Karniadakis. Integrating neural operators with diffusion models improves spectral representation in turbulence modeling, 2024. arXiv:2409.08477

    2024 arXiv

  36. [44]

    McDowell

    D.L. McDowell. Damage mechanics and metal fatigue: A discriminating perspective. International Journal of Fracture Mechanics, 8, 1999

    1999

  37. [45]

    When and why PINNs fail to train: A neural tangent kernel perspective

    Sifan Wang, Xinling Yu, and Paris Perdikaris. When and why PINNs fail to train: A neural tangent kernel perspective. Journal of Computational Physics, 449:110768, 2022

    2022

  38. [46]

    Multi-grid tensorized Fourier neural operator for high-resolution PDEs, 2023

    Jean Kossaifi, Nikola Kovachki, Kamyar Azizzadenesheli, and Anima Anandkumar. Multi-grid tensorized Fourier neural operator for high-resolution PDEs, 2023. arXiv:2310.00120

    2023 arXiv

  39. [47]

    Simulated microstructure-sensitive extreme value probabilities for high cycle fatigue of duplex Ti–6Al–4V

    Craig P Przybyla and David L McDowell. Simulated microstructure-sensitive extreme value probabilities for high cycle fatigue of duplex Ti–6Al–4V. International Journal of Plasticity, 27(12):1871–1895, 2011

    2011

  40. [48]

    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, 3(6):422–440, 2021

    2021

  41. [49]

    Embedding hard physical constraints in neural network coarse-graining of three-dimensional turbulence

    Arvind T Mohan, Nicholas Lubbers, Misha Chertkov, and Daniel Livescu. Embedding hard physical constraints in neural network coarse-graining of three-dimensional turbulence. Physical Review Fluids, 8(1):014604, 2023

    2023

  42. [50]

    Divergence-free neural operators for stress field modeling in polycrystalline materials, 2024

    Mohammad S Khorrami, Pawan Goyal, Jaber R Mianroodi, Bob Svendsen, Peter Benner, and Dierk Raabe. Divergence-free neural operators for stress field modeling in polycrystalline materials, 2024. arXiv:2408.15408

    2024 arXiv

  43. [51]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6840–6851. Curran Associates, Inc., 2020

    2020

  44. [52]

    Statistically conditioned polycrystal generation using denoising diffusion models

    Michael O Buzzy, Andreas E Robertson, and Surya R Kalidindi. Statistically conditioned polycrystal generation using denoising diffusion models. Acta Materialia, 267:119746, 2024

    2024

  45. [53]

    Adam: A method for stochastic optimization, 2014

    Diederik P Kingma. Adam: A method for stochastic optimization, 2014. arXiv:1412.6980

    2014 arXiv

  46. [54]

    Microstructure reconstruction of 2D/3D random materials via diffusion-based deep generative models

    Xianrui Lyu and Xiaodan Ren. Microstructure reconstruction of 2D/3D random materials via diffusion-based deep generative models. Scientific Reports, 14(1):5041, 2024

    2024

  47. [55]

    Generating 3D images of material microstructures from a single 2D image: a denoising diffusion approach

    Johan Phan, Muhammad Sarmad, Leonardo Ruspini, Gabriel Kiss, and Frank Lindseth. Generating 3D images of material microstructures from a single 2D image: a denoising diffusion approach. Scientific Reports, 14(1):6498, 2024

    2024

  48. [56]

    A review of activation function for artificial neural network

    Andrinandrasana David Rasamoelina, Fouzia Adjailia, and Peter Sin ˇcák. A review of activation function for artificial neural network. In 2020 IEEE 18th World Symposium on Applied Machine Intelligence and Informatics (SAMI), pages 281–286. IEEE, 2020

    2020

  49. [57]

    Decoupled weight decay regularization, 2017

    I Loshchilov. Decoupled weight decay regularization, 2017. arXiv:1711.05101

    2017 arXiv

  50. [58]

    High-performance large-scale image recognition without normalization

    Andy Brock, Soham De, Samuel L Smith, and Karen Simonyan. High-performance large-scale image recognition without normalization. In International Conference on Machine Learning, pages 1059–1071. PMLR, 2021

    2021

  51. [59]

    On the convergence of Adam and beyond, 2019

    Sashank J Reddi, Satyen Kale, and Sanjiv Kumar. On the convergence of Adam and beyond, 2019. arXiv:1904.09237

    2019 arXiv

  52. [60]

    Optuna: A next-generation hyperparameter optimization framework

    Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. Optuna: A next-generation hyperparameter optimization framework. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 2623–2631, 2019

    2019

  53. [61]

    Tree-structured Parzen estimator: Understanding its algorithm components and their roles for better empirical performance, 2023

    Shuhei Watanabe. Tree-structured Parzen estimator: Understanding its algorithm components and their roles for better empirical performance, 2023. arXiv:2304.111271904.09237

    2023 arXiv

  54. [62]

    An elasto-viscoplastic formulation based on fast Fourier transforms for the prediction of micromechanical fields in polycrystalline materials

    Ricardo A Lebensohn, Anand K Kanjarla, and Philip Eisenlohr. An elasto-viscoplastic formulation based on fast Fourier transforms for the prediction of micromechanical fields in polycrystalline materials. International Journal of Plasticity, 32:59–69, 2012

    2012

  55. [63]

    C. N. Alleman, J. W. Foulk, A. Mota, H. Lim, and D. J. Littlewood. Concurrent multiscale modeling of microstructural effects on localization behavior in finite deformation solid mechanics. Computational Mechanics, 61:207–218, 2018

    2018

  56. [64]

    Groeber and Michael A

    Michael A. Groeber and Michael A. Jackson. DREAM.3D: A digital representation environment for the analysis of microstructure in 3D. Integrating Materials and Manufacturing Innovation, 3(1):56–72, 2014. 21 Equilibrium Conserving Neural Operators

    2014

  57. [65]

    Robertson, Daniel Diaz, Coleman Alleman, Zhen Zhang, Anthony D

    Vivek Oommen, Andreas E. Robertson, Daniel Diaz, Coleman Alleman, Zhen Zhang, Anthony D. Rol- lett, George E. Karniadakis, and Rémi’ Dingreville. Equilibrium conserving neural operators, 2025. https://github.com/vivekoommen/PhysicsBased_NeuralSR

    2025

  58. [66]

    medium”|“high

    Ari Frankel, Kousuke Tachida, and Reese Jones. Prediction of the evolution of the stress field of polycrystals undergoing elastic-plastic deformation with a hybrid neural network model. Machine Learning: Science and Technology, 1(3):035005, 2020. 22 Equilibrium Conserving Neur...

    2020

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.