Pith. sign in

REVIEW 5 major objections 5 minor 30 references

A Hybrid Virtual Element Method and Deep Learning Approach for Solving One-Dimensional Euler-Bernoulli Beams

T0 review · 5 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A VEM-trained two-branch neural network can predict beam displacement fields from just 80 samples, with errors that shrink under mesh refinement until an unexpected rise at 384 elements.

desk verdict A plausible but underspecified neural-VEM surrogate whose headline claims can't be checked because key losses and the variable-mesh mechanism are missing. read the letter →

arxiv 2501.06925 v1 pith:RDXKAWBJ submitted 2025-01-12 cs.LG

Paulo Akira F. Enabe , Rodrigo Provasi This is my paper · ORCID
classification cs.LG MSC 65N3074K1068T07
keywords VirtualElementMethoddeeplearningsurrogateEuler-BernoullibeamSobolevtrainingGradNormdisplacementfieldpredictionstructuralmechanicsporticoframe
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 establish that a surrogate model combining the Virtual Element Method (VEM) with a two-branch neural network can predict displacement fields of one-dimensional Euler-Bernoulli beams and portico frames across varying material and geometric parameters. The network keeps nodal geometry and material properties in separate sub-networks, then trains on only 80 VEM solutions using Sobolev training, which matches derivatives, and GradNorm, which balances the loss terms. Reported H1 errors against VEM references generally decrease as the mesh refines, with the order-4 model slightly outperforming the order-5 model, and both models showing an unexpected error increase at 384 elements. If the claim holds, structural surrogates could be built from small datasets and evaluated quickly enough for iterative design and analysis.

What carries the argument

The carrying mechanism is the two-branch neural network: a node sub-network of dense layers receives coordinates, boundary conditions, and geometry while a parallel material sub-network receives Young's modulus, area, and moment of inertia, and the concatenated outputs pass through final dense layers to predict the deflection field. Training is driven by a composite loss of displacement error, a Sobolev term penalizing mismatches in projected derivatives, and a material-penalty term, with the three contributions balanced adaptively by GradNorm. The VEM side supplies the reference solutions through a general-order projection formulation that introduces internal moment degrees of freedom for higher-order beam elements.

What would settle it

Evaluate the trained network on a resolution not seen in training, such as 768 elements, and check whether it even accepts the input and whether the H1 error continues a smooth trend; if it cannot predict or the error jumps far off the VEM convergence curve, the claimed mesh-agnostic surrogate fails.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that a neural network can learn the parameter-to-displacement map of Euler-Bernoulli beam problems from VEM-generated examples, provided the architecture separates node data from material data and the training loss includes derivative information with dynamically balanced task weights. The experiments show a mean H1 error that decreases from 24 to 192 elements, indicating the learned surrogate tracks VEM's refinement behavior, and that 80 training samples suffice for this tracking. The paper also discovers that the tracking fails at 384 elements, where the error rises for both quadratic and cubic formulations, a deviation it attributes to interference between the network's learned non-linear approximation space and VEM's expected polynomial convergence.

Load-bearing premise

The results assume one dense node sub-network can handle meshes from 24 to 384 elements, but no mechanism is stated for how a fixed-size network input represents a variable-size mesh.

Editorial extensions

If this is right

  • If the surrogate generalizes as reported, structural analysis can precompute a small set of VEM solutions and then run fast inference for other parameter values.
  • The two-branch design is a reusable pattern for problems where one part of the input is geometric or nodal and another part is scalar material data.
  • The Sobolev-plus-GradNorm training recipe is presented as a way to keep multi-term physics losses stable when training data is scarce.
  • Fast inference would support real-time design iteration, optimization, and uncertainty quantification in one-dimensional frame analysis.
  • The error rise at 384 elements implies the current hybrid does not inherit VEM's asymptotic convergence and needs architectural changes before it can be trusted on fine meshes.

Reading between the lines

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

  • The paper leaves open how a fixed-size dense node sub-network accommodates element counts from 24 to 384; an editor-level inference is that a genuinely mesh-invariant encoding, such as aggregated per-element features, is needed before the convergence results can be read as a single model.
  • The plateau and rise at fine meshes suggest network capacity, not VEM order, is the bottleneck; scaling network width or depth with element order is a direct testable extension.
  • Because material parameters enter as global scalars, the same architecture could plausibly predict at extrapolated E, A, and I values, but the paper does not test extrapolation, so that experiment remains open.
  • The hybrid recipe could extend to two- and three-dimensional elasticity if the node branch is replaced with a geometrically aware encoder, since the material branch would still operate on scalars.
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

5 major / 5 minor

Summary. The manuscript proposes a hybrid surrogate model that combines a one-dimensional Virtual Element Method (VEM) formulation for Euler–Bernoulli beams with a deep neural network split into node-specific and material-specific sub-networks. Training uses Sobolev derivative information and GradNorm loss balancing. The paper reports H1 errors for quadratic (order 4) and cubic (order 5) VEM formulations on a portico geometry across mesh sizes from 24 to 384 elements, using an 80-sample training set, and claims that the approach achieves accurate displacement predictions with very little data and faster inference than traditional numerical methods.

Significance. The idea of pairing a VEM solver with a neural-network surrogate for parametric beam analysis is potentially useful, particularly because the reported training set is small and the architecture separates nodal from material features. The manuscript also gives a clear, compact derivation of the VEM beam formulation and includes a pseudocode description of Sobolev training with GradNorm. However, as written, the central numerical claims cannot be checked: the primary result figure is absent, the material penalty loss L3 is never defined, the handling of variable mesh sizes is unspecified, validation is performed only against the same VEM generator that produced the training data, and the advertised inference-time advantage is never measured. At this stage the paper does not establish its main contributions.

major comments (5)
  1. [§4, Figure 3] The H1-error results that constitute the paper's central numerical evidence appear only as a reference to Figure 3, but no figure or data table with the mean and standard deviation values is included in the manuscript. Without the actual numbers or plots, the claimed convergence behavior and the differences between the order 4 and order 5 models cannot be verified or reproduced.
  2. [§3.3, Eq. (39) and Algorithm 1] The 'material penalty loss' L3, which appears in the total loss Ltotal = θ1L1 + θ2L2 + θ3L3 and in the GradNorm gradient norms, is never given a concrete definition. Algorithm 1 states only 'Compute the material penalty loss: L3 (problem-specific regularization)' and Section 3.3 says it 'penalizes deviations in material properties,' but no functional form, target quantity, or weighting is specified, making the training procedure irreproducible.
  3. [§3 and §4] The node sub-network is described as a series of dense layers taking nodal coordinates, boundary conditions, and geometric factors as inputs, yet the experiments sweep 24, 48, 96, 192, and 384 elements. Dense layers require fixed input dimensions, and the manuscript never states whether separate models are trained per mesh count or whether a variable-size encoding (padding, masking, aggregation, weight sharing) is used. Either way, the reported convergence study and the claim of a mesh-flexible surrogate are not interpretable as stated.
  4. [§5 and §4.2] The conclusion lists 'the efficiency of performing inference compared to traditional methods' among the demonstrated strengths, but Section 4 contains no timing measurements, no runtime comparison with VEM or FEM, and no complexity analysis. The efficiency claim is therefore unsupported by any experimental evidence in the manuscript.
  5. [§4] All reported H1 errors are computed against 'reference solutions obtained using the VEM formulation' that also generated the 80 training samples, and no analytical beam solution or independent solver is used for comparison. Because the training and test targets come from the same VEM implementation, the reported accuracy cannot distinguish genuine generalization error from systematic errors in the data generator. In addition, the unexpected increase in error at 384 elements is reported without a diagnostic analysis; Section 4.1 offers only qualitative speculation about nonlinearity, which does not establish the cause or demonstrate that the result is not an artifact.
minor comments (5)
  1. [Figures 1 and 2] The captions for Figures 1 and 2 appear without the corresponding images, and Figure 1's caption literally reads 'Sample figure caption,' indicating an unresolved placeholder.
  2. [§4] The terms 'quadratic (order 4)' and 'cubic (order 5)' should be defined in relation to the polynomial degree n used in the VEM formulation of Section 2, because the current terminology is ambiguous.
  3. [Abstract and §4] The abstract and introduction frame the contribution as solving one-dimensional Euler–Bernoulli beams, while Section 4 evaluates a portico (two-dimensional frame) geometry; this scope mismatch should be clarified.
  4. [Appendix A, Eq. (41)–(42)] Equation (42) writes the square of the L2 norm as equal to the L2 norm itself, and Eq. (41) then combines these quantities in a way that is dimensionally inconsistent; the standard definitions of \(\|u-u_h\|_{L^2}^2\) and \(\|\nabla(u-u_h)\|_{L^2}^2\) should be used.
  5. [§1] There is a typo in 'Physical-Informed Neural Networks' in the paragraph after the description of PINNs; it should read 'Physics-Informed.'

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the VEM formulation, neural surrogate, and training losses are stated independently, and the reported H1 errors measure generalization to held-out VEM samples rather than reducing to the training objective by construction.

full rationale

The paper's claimed derivation chain is a standard VEM stiffness construction followed by a supervised neural-network surrogate. Section 2 derives the VEM projection and stiffness matrix from the Euler-Bernoulli functional without importing the network's outputs; Section 3 defines the surrogate as a dense two-sub-network model with Sobolev and GradNorm losses; Section 4 evaluates the trained model on a 20-sample test set against VEM reference solutions. Nothing in the equations makes the predicted displacement equal to the training target by construction: the loss L1 = ||m(x_i;θ) - f(x_i)||^2 in Algorithm 1 is a supervised fitting objective, and the reported H1 error is a held-out generalization metric, not a fitted value. The VEM formulation itself is supported by external references [1, 9, 27, 28], and no load-bearing premise is justified only by the present authors' prior work. A limitation worth noting is that the reference solutions and training labels both come from the same VEM implementation, so the experiments validate approximation of that generator rather than physical ground truth; the paper itself acknowledges the need for pure-VEM convergence isolation in Section 5. That self-referential validation is an external-validity concern, not a circular derivation, so the circularity score is 0.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the validity of Euler-Bernoulli theory, on the VEM formulation and its projection properties, and on the assumption that VEM solutions are accurate references. The neural network component inherits the universal approximation theorem and the empirical assumptions behind Sobolev training and GradNorm. Many training hyperparameters and the input parameter ranges are not reported, and the Sobolev derivative order is unspecified, so the reader must accept several unstated choices to reproduce the claimed accuracy.

free parameters (4)
  • Neural network hyperparameters (layers, neurons, activation, learning rates eta and eta_theta, epochs, batch size) = not reported
    These determine the reported H1 errors; without them the central accuracy claim is not reproducible.
  • GradNorm balancing coefficient alpha and weight normalization target T = not reported
    Equations (35) and (37) introduce these hyperparameters but no values are given; the loss-balancing behavior depends on them.
  • Sobolev derivative order K and projection vector distribution = not reported
    Equation (30) sums over K derivative orders; the paper never states K or the distribution of the random vectors v_j, so the Sobolev loss is underspecified.
  • Parameter ranges for E, I, and cross-sectional area in the 80-sample training set = not reported
    The generalization claim depends on the variability of these inputs, which is never quantified.
assumptions (6)
  • standard math Universal Approximation Theorem (Theorem A.1)
    Used to justify that a feedforward network can approximate the displacement function; stated in Appendix A.
  • domain assumption Euler-Bernoulli beam theory (Eq. 1)
    The target PDE EI w'''' = q is assumed valid for the beams and frame analyzed; no shear deformation or large-deflection effects are considered.
  • domain assumption VEM projection consistency (Eqs. 7-9)
    The paper assumes the projection operator reproduces curvature, gradient, and displacement moments so that the stiffness matrix is exact on polynomial spaces; this is standard VEM theory cited from [1] but not proven here.
  • domain assumption VEM reference solutions are accurate enough to serve as ground truth
    The H1 error is measured against VEM solutions, but the VEM code is never validated against analytical or converged FEM solutions; errors in VEM could dominate the reported surrogate errors.
  • standard math Johnson-Lindenstrauss Lemma (Theorem A.2)
    Cited to justify random projections in the Sobolev loss, though the lemma concerns distance preservation rather than derivative preservation; the applicability to derivative loss is not established.
  • domain assumption Sobolev training and GradNorm improve generalization
    The paper assumes that adding derivative losses and gradient balancing improves accuracy with 80 samples; this is empirical folklore from [27] and [28], not demonstrated here.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Hybrid Virtual Element Method and Deep Learning Approach for Solving One-Dimensional Euler-Bernoulli Beams." pith.science (2026). https://pith.science/paper/RDXKAWBJ

@misc{pith2026250106925,
  author       = {Pith},
  title        = {Pith review of: A Hybrid Virtual Element Method and Deep Learning Approach for Solving One-Dimensional Euler-Bernoulli Beams},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RDXKAWBJ}},
  note         = {Machine review of arXiv:2501.06925}
}
read the original abstract

A hybrid framework integrating the Virtual Element Method (VEM) with deep learning is presented as an initial step toward developing efficient and flexible numerical models for one-dimensional Euler-Bernoulli beams. The primary aim is to explore a data-driven surrogate model capable of predicting displacement fields across varying material and geometric parameters while maintaining computational efficiency. Building upon VEM's ability to handle higher-order polynomials and non-conforming discretizations, the method offers a robust numerical foundation for structural mechanics. A neural network architecture is introduced to separately process nodal and material-specific data, effectively capturing complex interactions with minimal reliance on large datasets. To address challenges in training, the model incorporates Sobolev training and GradNorm techniques, ensuring balanced loss contributions and enhanced generalization. While this framework is in its early stages, it demonstrates the potential for further refinement and development into a scalable alternative to traditional methods. The proposed approach lays the groundwork for advancing numerical and data-driven techniques in beam modeling, offering a foundation for future research in structural mechanics.

Figures

Figures reproduced from arXiv: 2501.06925 by the authors.

Figure 1
Figure 1. Sample figure caption. 5 [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Portico geometry. The results of the hybrid deep learning and VEM approach for quadratic (order 4) and cubic (order 5) formulations, shown in [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. H1 -error comparison of order 4 and order 5 with standard deviation. 4.1 A brief discussion about nonlinearity in deep learning Deep learning models are inherently non-linear approximators. Unlike the Virtual Element Method (VEM), which uses predefined polynomial spaces (quadratic or cubic) for basis functions, neural networks learn their approximation space from the training data. While this non-linear nature enabl… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 9 canonical work pages

  1. [1]

    Beirão da Veiga, F

    L. Beirão da Veiga, F. Brezzi, A. Cangiani, G. Manzini, L. D. Marini, and A. Russo. Basic principles of virtual element methods. In Mathematical Models and Methods in Applied Sciences, V olume 23, Issue 1, pages 199–214,

    work page

  2. [2]

    Beirão da Veiga, C

    L. Beirão da Veiga, C. Lovadina, and D. Mora. A Virtual Element Method for elastic and inelastic problems on polytope meshes. In Computer Methods in Applied Mechanics and Engineering , V olume 295, pages 327–346,

    work page

  3. [3]

    Virtual Element Methods for hyperbolic problems on polygonal meshes

    Giuseppe Vacca. Virtual Element Methods for hyperbolic problems on polygonal meshes. In Computers & Mathematics with Applications, V olume 74, Issue 5, pages 882–898, 2017. ISSN 0898-1221. https://doi.org/ 10.1016/j.camwa.2016.04.029

    work page doi:10.1016/j.camwa.2016.04.029 2017

  4. [4]

    Artioli, L

    E. Artioli, L. Beirão da Veiga, and F. Dassi. Curvilinear Virtual Elements for 2D solid mechanics applications. In Computer Methods in Applied Mechanics and Engineering, V olume 359, page 112667, 2020. ISSN 0045-7825. https://doi.org/10.1016/j.cma.2019.112667

    work page arXiv 2020

  5. [5]

    Wriggers, B

    P. Wriggers, B. Hudobivnik, and F. Aldakheel. A virtual element formulation for general element shapes. In Computational Mechanics, V olume 66, Number 4, pages 963–977, 2020. https://doi.org/10.1007/ s00466-020-01891-5

    work page 2020

  6. [6]

    A low order 3D virtual element formulation for finite elasto–plastic deformations

    Blaž Hudobivnik, Fadi Aldakheel, and Peter Wriggers. A low order 3D virtual element formulation for finite elasto–plastic deformations. In Computational Mechanics, V olume 63, Number 2, pages 253–269, 2019.https: //doi.org/10.1007/s00466-018-1593-6

    doi:10.1007/s00466-018-1593-6 2019

  7. [7]

    A virtual element method for 3D contact problems with non-conforming meshes

    Mertcan Cihan, Blaž Hudobivnik, Jože Korelc, and Peter Wriggers. A virtual element method for 3D contact problems with non-conforming meshes. In Computer Methods in Applied Mechanics and Engineering, V olume 402, page 115385, 2022. ISSN 0045-7825. https://doi.org/10.1016/j.cma.2022.115385

    arXiv 2022

  8. [8]

    High-order 3D virtual element method for linear and nonlinear elasticity

    Bing-Bing Xu, Wei-Long Fan, and Peter Wriggers. High-order 3D virtual element method for linear and nonlinear elasticity. In Computer Methods in Applied Mechanics and Engineering, V olume 431, page 117258, 2024. ISSN 0045-7825. https://doi.org/10.1016/j.cma.2024.117258

    arXiv 2024

Show all 30 references
  1. [9]

    Wriggers

    P. Wriggers. On a virtual element formulation for trusses and beams. In Archive of Applied Mechanics, V olume 92, pages 1655–1678, 2022. https://doi.org/10.1007/s00419-022-02113-5

    2022 doi

  2. [10]

    G. Cybenko. Approximation by superpositions of a sigmoidal function. In Mathematics of Control, Signals, and Systems, V olume 2, pages 303–314, 1989.https://doi.org/10.1007/BF02551274

    1989 doi

  3. [11]

    Multilayer feedforward networks are universal ap- proximators

    Kurt Hornik, Maxwell Stinchcombe, and Halbert White. Multilayer feedforward networks are universal ap- proximators. In Neural Networks , V olume 2, Issue 5, pages 359–366, 1989. ISSN 0893-6080. https: //doi.org/10.1016/0893-6080(89)90020-8

    1989 doi

  4. [12]

    Approximation capabilities of multilayer feedforward networks

    Kurt Hornik. Approximation capabilities of multilayer feedforward networks. In Neural Networks, V olume 4, Issue 2, pages 251–257, 1991. ISSN 0893-6080. https://doi.org/10.1016/0893-6080(91)90009-T

    1991 doi

  5. [13]

    I. E. Lagaris, A. Likas, and D. I. Fotiadis. Artificial neural networks for solving ordinary and partial differential equations. In IEEE Transactions on Neural Networks, V olume 9, Number 5, pages 987–1000, September 1998. https://doi.org/10.1109/72.712178

    1998 doi

  6. [14]

    Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations

    Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations. Preprint, 2017. https://arxiv.org/abs/1711.10561

    2017 arXiv

  7. [15]

    Physics Informed Deep Learning (Part II): Data- driven Discovery of Nonlinear Partial Differential Equations

    Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics Informed Deep Learning (Part II): Data- driven Discovery of Nonlinear Partial Differential Equations. Preprint, 2017. https://arxiv.org/abs/1711. 10566

    2017

  8. [16]

    Physics-informed neural networks for approximating dynamic (hyperbolic) PDEs of second order in time: Error analysis and algorithms

    Yanxia Qian, Yongchao Zhang, Yunqing Huang, and Suchuan Dong. Physics-informed neural networks for approximating dynamic (hyperbolic) PDEs of second order in time: Error analysis and algorithms. In Journal of Computational Physics, V olume 495, page 112527, 2023. ISSN 0021-999...

    2023

  9. [17]

    Sharma, L

    P. Sharma, L. Evans, M. Tindall, and others. Stiff-PDEs and Physics-Informed Neural Networks. In Archives of Computational Methods in Engineering, V olume 30, pages 2929–2958, 2023.https://doi.org/10.1007/ s11831-023-09890-4

    2023

  10. [18]

    Jagtap, Shandian Zhe, George Em Karniadakis, and Robert M

    Michael Penwarden, Ameya D. Jagtap, Shandian Zhe, George Em Karniadakis, and Robert M. Kirby. A unified scalable framework for causal sweeping strategies for Physics-Informed Neural Networks (PINNs) and their temporal decompositions. In Journal of Computational Physics, V olum...

    2023

  11. [19]

    Anagnostopoulos, Juan Diego Toscano, Nikolaos Stergiopulos, and George Em Karniadakis

    Sokratis J. Anagnostopoulos, Juan Diego Toscano, Nikolaos Stergiopulos, and George Em Karniadakis. Residual- based attention in physics-informed neural networks. In Computer Methods in Applied Mechanics and Engineering, V olume 421, page 116805, 2024. ISSN 0045-7825.https://do...

    2024

  12. [20]

    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. In Journal of Computational Physics, V olume 449, page 110768, 2022. ISSN 0021-9991. https: //doi.org/10.1016/j.jcp.2021.110768

    2022

  13. [21]

    Challenges in Training PINNs: A Loss Landscape Perspective

    Pratik Rathore, Weimu Lei, Zachary Frangella, Lu Lu, and Madeleine Udell. Challenges in Training PINNs: A Loss Landscape Perspective. Preprint, 2024. https://arxiv.org/abs/2402.01868

    2024 arXiv

  14. [22]

    Deep learned finite elements

    Jaeho Jung, Kyungho Yoon, and Phill-Seung Lee. Deep learned finite elements. In Computer Methods in Applied Mechanics and Engineering, V olume 372, page 113401, 2020. ISSN 0045-7825. https://doi.org/10.1016/j. cma.2020.113401

    2020

  15. [23]

    Self-updated four-node finite element using deep learning

    Jaeho Jung, Hyunok Jun, and Phill-Seung Lee. Self-updated four-node finite element using deep learning. In Com- putational Mechanics, V olume 69, pages 23–44, 2022.https://doi.org/10.1007/s00466-021-02081-7

    2022 doi

  16. [24]

    A deep energy method for finite deformation hyperelasticity

    Vien Minh Nguyen-Thanh, Xiaoying Zhuang, and Timon Rabczuk. A deep energy method for finite deformation hyperelasticity. In European Journal of Mechanics - A/Solids, V olume 80, page 103874, 2020. ISSN 0997-7538. https://doi.org/10.1016/j.euromechsol.2019.103874

    2020

  17. [25]

    Abueidda, Seid Koric, Rashid Abu Al-Rub, Corey M

    Diab W. Abueidda, Seid Koric, Rashid Abu Al-Rub, Corey M. Parrott, Kai A. James, and Nahil A. Sobh. A deep learning energy method for hyperelasticity and viscoelasticity. InEuropean Journal of Mechanics - A/Solids, V olume 95, page 104639, 2022. ISSN 0997-7538. https://doi.org...

    2022

  18. [26]

    R. E. Meethal, A. Kodakkal, M. Khalil, and others. Finite element method-enhanced neural network for forward and inverse problems. In Advances in Modeling and Simulation in Engineering Sciences, V olume 10, page 6, 2023. https://doi.org/10.1186/s40323-023-00243-1

    2023 doi

  19. [27]

    Sobolev training for neural networks

    Wojciech Marian Czarnecki, Simon Osindero, Max Jaderberg, Grzegorz ´Swirszcz, and Razvan Pascanu. Sobolev training for neural networks. arXiv preprint arXiv:1706.04859, 2017. https://arxiv.org/abs/1706.04859

    2017 arXiv

  20. [28]

    GradNorm: Gradient normalization for adaptive loss balancing in deep multitask networks

    Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. GradNorm: Gradient normalization for adaptive loss balancing in deep multitask networks. arXiv preprint arXiv:1711.02257, 2018. https://arxiv. org/abs/1711.02257. 15 JANUARY 14, 2025 A Functional analysis def...

    2018 arXiv

  21. [2013]

    https://doi.org/10.1142/S0218202512500492

    doi

  22. [2015]

    https://doi.org/10.1016/j.cma.2015.07.013

    ISSN 0045-7825. https://doi.org/10.1016/j.cma.2015.07.013

    2015 doi

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.