Pith. sign in

REVIEW 5 major objections 5 minor 53 references

Multi-Resolution Training-Enhanced Kolmogorov-Arnold Networks for Multi-Scale PDE Problems

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read MR-PIKAN claims that alternating between low- and high-resolution collocation batches trains physics-informed Kolmogorov-Arnold networks faster on multi-scale PDEs without sacrificing accuracy.

desk verdict A simple, plausible recipe for cheaper PIKAN training on multi-scale PDEs, with honest but under-powered evidence; worth a proper review, not a desk reject. read the letter →

arxiv 2507.19888 v1 pith:FLOJUTNB submitted 2025-07-26 physics.comp-ph

Yu-Sen Yang , Ling Guo , Xiaodan Ren This is my paper · ORCID
classification physics.comp-ph
keywords multi-resolutiontrainingKolmogorov-Arnoldnetworksphysics-informedneuralmulti-scalePDEsinverseproblemsinformationbottleneckChebyshevKANsamplingresolution
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 proposes MR-PIKAN, a training schedule for physics-informed Kolmogorov-Arnold networks (PIKANs) that alternates, or sequentially switches, between low- and high-resolution batches of sampling points when solving multi-scale PDEs. The central claim is that this multi-resolution schedule cuts training time on both forward and inverse problems while matching or slightly improving the accuracy of training purely at high resolution. The paper supports the claim with an information-bottleneck analysis of the gradient signal, showing that higher-resolution sampling raises the signal-to-noise ratio of the optimization, and with three numerical benchmarks: a two-dimensional Poisson equation, a Gray-Scott reaction-diffusion inverse problem, and a hidden-elasticity problem for two-phase random materials. If correct, the result offers a simple way to reduce the dominant computational cost of physics-informed deep learning on problems with widely separated scales.

What carries the argument

The central object is the training resolution function $N(t)$, which sets the number of residual collocation points used to estimate the gradient at iteration $t$. Instead of holding the batch size fixed, MR-PIKAN switches $N(t)$ between a low-resolution value and a high-resolution value, either sequentially (coarse first, then fine) or alternately on a short cycle. The argument that this switching preserves useful signal rests on the batch-wise signal-to-noise ratio $\mathrm{SNR} = \|\mu\|_2 / \|\sigma\|_2$ of the gradient and the related gradient-homogeneity metric $\mathrm{SRR}_B$, which the paper uses to show that higher-resolution batches give higher SNR during the total-diffusion phase of training. The representation model is a Chebyshev-polynomial KAN, whose univariate activations are weighted Chebyshev polynomials up to degree 6 or 15, chosen because KANs show reduced spectral bias on high-frequency components.

What would settle it

Train the same MR-PIKAN schedule on a multiscale PDE with a frequency gap wider than the benchmarks, such as a Poisson problem whose solution mixes wavelengths differing by more than an order of magnitude, and compare final test error against the high-resolution-only run across several random seeds. If the alternating schedule's final error is consistently above the single-resolution baseline, the claimed resolution-switching efficiency does not generalize.

Watch

Extended reading notes

Core claim

The paper establishes that the resolution of the residual sampling grid is a first-order factor in how accurately a physics-informed KAN learns multi-scale fields, and that this cost can be exploited rather than merely paid. Training alternately with a small and a large collocation batch—so that cheaper low-resolution steps do the early and repeated coarse work—reaches the same or lower test error than single-resolution high-resolution training in roughly one-sixth less wall-clock time in the forward Poisson benchmark, about 20% less in the Gray-Scott inverse problem, and less than half the time to reach a given accuracy in the two-phase material inverse problem. The mechanism is framed through the signal-to-noise ratio of the batch gradient: higher resolution raises that ratio, and the alternating schedule keeps it high while lowering the per-step cost.

Load-bearing premise

The load-bearing premise is that switching between small and large batches of sampling points preserves the useful progress of gradient descent: the cheap coarse steps must not push the network into a region the fine steps cannot escape.

Editorial extensions

If this is right

  • For forward multiscale PDEs, alternating every 20 to 80 epochs between a low and a high collocation batch can cut training time by roughly one-sixth while achieving slightly lower test error than the high-resolution-only baseline.
  • For inverse problems with mini-batch training, the same alternating schedule reduces training time by about 20% in the Gray-Scott example and by over 30% in time-to-acceptable-accuracy for the two-phase material, without degrading the inferred parameters or fields.
  • The strategy works with both full-batch and mini-batch sampling, so it remains usable when memory limits prevent simply raising the resolution further.
  • Because the method only changes how existing random sampling batches are scheduled, it can be combined with adaptive sampling algorithms that focus points in high-residual regions.

Reading between the lines

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

  • The paper's SNR analysis suggests a testable design rule it does not pursue: the optimal low-resolution batch size and switching interval should track the degree of scale separation, with wider frequency gaps likely needing longer coarse phases.
  • If the benefit is genuinely about gradient SNR rather than the KAN architecture, the same alternating schedule should transfer to other low-spectral-bias representations; the paper's own Fourier-feature PINN experiment shows no clear time savings, so the boundary of where the trick pays off is left open.
  • For very large problems where full-batch high resolution exceeds memory, alternating training offers a practical fallback: the low-resolution phase still supplies useful global signal, and the high-resolution phase can be shortened.
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.

Referee Report

5 major / 5 minor

Summary. The manuscript proposes MR-PIKAN, a training schedule that alternates or sequentially switches the collocation-point batch size between low and high resolution for physics-informed Kolmogorov-Arnold networks. After an information-bottleneck-style analysis of single-resolution training dynamics using SNR and gradient-homogeneity metrics, the authors validate MR-PIKAN on three multi-scale problems: a 2D Poisson forward problem, a Gray-Scott inverse parameter-identification problem, and an inverse elasticity problem for two-phase random materials. In each case they report that MR-PIKAN matches or slightly improves the accuracy of a single-resolution high-resolution PIKAN while reducing wall-clock training time.

Significance. If the efficiency claim is confirmed, the method is attractive because it is simple, requires no changes to the PDE loss, and is easy to layer onto existing PIKAN implementations. The paper has real strengths: it tests against exact and FEM reference solutions, uses three problems with different forward/inverse structures, gives detailed implementation tables, and ships a complete algorithmic description. However, the current evidence does not yet support the headline claim, because the comparisons are equal-iteration rather than compute-matched, all runs are single-seed, and several statements about training dynamics are asserted rather than demonstrated.

major comments (5)
  1. [§4.1.1, Fig. 8; Table 2] The reported time savings are computed by comparing MR-PIKAN with SR-PIKAN after an equal number of iterations (40,000, 120,000, and 400,000). Since MR-PIKAN deliberately replaces part of the high-resolution workload with cheaper low-resolution batches, the wall-clock reduction is to some extent built into the comparison. The paper never shows that MR-PIKAN reaches a target accuracy in less total compute than the best single-resolution schedule; a compute-matched SR run stopped at the MR wall-clock budget could in principle match or beat the MR final error. Only Fig. 15 provides a time-to-target comparison, and only for the TRM problem. Please add time-to-accuracy curves for all three examples and a matched-budget SR baseline.
  2. [§4 (all examples)] All experiments in Section 4 are single-seed, with no error bars in Figs. 8, 9, 11, 15 or Tables 1–2. The claim that MR-PIKAN does not sacrifice accuracy is therefore not distinguished from run-to-run variation. Please report seed-averaged means and standard deviations (at least 3–5 seeds) for final L2 errors, inferred parameters, and training times, and state whether the observed differences are statistically meaningful.
  3. [§3.2, Fig. 6 caption] The statement that the gradient descent process trained with these strategies maintains consistent expected information across resolutions is not established. Equation (21) only shows that the stochastic gradient is an unbiased estimate of the full-batch gradient at a fixed resolution; it does not imply that low-resolution updates produce equal progress or that switching resolutions preserves optimizer state. The plateau in Fig. 8(a) in fact suggests low-resolution epochs contribute little progress. This premise should be presented as a heuristic motivation, not as a theoretical guarantee.
  4. [§4.2, Table 2] For the Gray-Scott inverse problem, the paper reports only the inferred coefficients ε1 and ε2 and their relative errors; no quantitative field-wise L2 errors are given for the predicted concentration fields u and v shown in Fig. 10. Since the central claim is that accuracy is not sacrificed, the field-error comparison between SR-PIKAN and MR-PIKAN should be reported.
  5. [§4 (overall)] The related work cites adaptive sampling methods (RAR, residual redistribution, importance sampling, refs. [22–27]) as the main alternative for reducing sampling cost, but no such baseline is included in any experiment. A comparison against at least one adaptive-sampling or importance-sampling method is needed to substantiate the claim that MR-PIKAN effectively reduces computational costs relative to existing cost-reduction techniques rather than only relative to a fixed high-resolution schedule.
minor comments (5)
  1. [Eq. (25)] The expanded loss omits the weights ωru, ωrv, ωdu, ωdv that appear in the first line of the equation; please reconcile the notation.
  2. [Appendix A, Tables A1–A3] The tables mix 'epochs' and 'iterations' (e.g., Table A3 lists 'Number of epochs 4000' while the text says 400,000 iterations); please specify the intended unit consistently across all tables.
  3. [Figs. 8 and 15] The 'Time Saved' annotation is not defined; state how the time saving is computed, relative to which baseline, and in what units.
  4. [Abstract and §3.1] The phrase 'information bottleneck theory' is used only through the SNR and SRR_B metrics from ref. [39]; either cite the underlying IB formulation or soften the terminology to 'gradient statistics analysis.'
  5. [Various captions] A light language edit is needed for subject-verb agreement and article usage in several captions (e.g., Fig. 4: 'model trained with the single-resolution setting' vs. 'models trained'); this does not affect the technical content.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central accuracy and efficiency claims rest on independent test-error and reference-solution comparisons; the only self-citation is non-load-bearing data generation.

full rationale

The paper's central claim is that alternating or sequential low/high-resolution training reduces wall-clock time without sacrificing accuracy. The efficiency part is an empirical timing comparison at equal iteration counts, and the accuracy part is checked against dense test sets and exact/FEM reference solutions (Section 4.1, 4.2, 4.3; Eq. 22). These claims do not reduce to a fitted parameter renamed as a prediction, nor are they derived from a self-citation chain. The information-bottleneck/SNR analysis in Section 3.1 follows the external framework of reference [39] and is used only as motivation, while the multi-resolution idea is explicitly credited to multigrid methods [40] and prior computer-vision work [32,33]. The only self-citation is reference [50] (Ren and Lyu, by the corresponding author) for generating the random-field-level-cut microstructure in Section 4.3; this supplies a synthetic test case but does not carry the efficiency or accuracy conclusion. A possible fairness concern is that the time saving is partly built into comparing equal iteration counts with cheaper low-resolution batches, but that is a benchmarking/correctness issue, not a circularity of the kind defined here. Accordingly, no circular step meets the evidence bar, and the score reflects only the minor non-load-bearing self-citation.

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

The central claim rests on a small number of hand-chosen hyperparameters (resolution levels, switch steps, loss weights) and on several domain assumptions: the applicability of IB analysis to PIKAN training, the accuracy of reference-generated data, and the key empirical premise that a coarse-to-fine schedule does not hurt final accuracy. The paper introduces no invented physical entities.

free parameters (4)
  • Low-resolution batch size (e.g., Nr=1280 or 1000) = 1280 (Poisson), 1000 (Gray-Scott and TRM)
    Chosen by hand for the multi-resolution schedule; the method's time savings depend on this choice.
  • Alternating step size (epochs per resolution) = 20/40/60/80 (Poisson), 10-80 (Gray-Scott)
    Hyperparameter scanned in experiments; performance depends on it without a principled selection rule.
  • Loss weights = omega_r=0.1, omega_b=100 (Poisson); adaptive weights elsewhere
    Selected for training stability; typical hyperparameters, not fitted to the target result.
  • Initial log-diffusion values alpha1, alpha2 = -10
    Initialization for the Gray-Scott inverse problem; a chosen starting point, not a fitted parameter.
assumptions (4)
  • standard math Kolmogorov-Arnold representation theorem permits KAN to approximate the solutions of the considered PDEs.
    Section 2.3; the representational basis of the PIKAN model.
  • domain assumption The information bottleneck and total-diffusion analysis from ref [39] applies to PIKAN training dynamics.
    Section 3.1, Eqs. 16-18; the paper uses SNR and SRR_B to claim that higher resolution improves accuracy through higher SNR, but this transfer is assumed, not derived.
  • ad hoc to paper Low-resolution training followed by high-resolution fine-tuning does not permanently bias the optimizer toward a poor basin.
    Section 3.2 and Figure 8; the entire acceleration claim relies on this empirical premise.
  • domain assumption Generated reference solutions (Chebfun for Gray-Scott, FEM for TRM) are accurate enough to serve as ground truth.
    Sections 4.2 and 4.3; no convergence study of the reference data is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Resolution Training-Enhanced Kolmogorov-Arnold Networks for Multi-Scale PDE Problems." pith.science (2026). https://pith.science/paper/FLOJUTNB

@misc{pith2026250719888,
  author       = {Pith},
  title        = {Pith review of: Multi-Resolution Training-Enhanced Kolmogorov-Arnold Networks for Multi-Scale PDE Problems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FLOJUTNB}},
  note         = {Machine review of arXiv:2507.19888}
}
read the original abstract

Multi-scale PDE problems present significant challenges in scientific computing. While conventional MLP-based deep learning methods exhibit spectral bias in resolving multi-scale features, the physics-informed Kolmogorov-Arnold network (PIKAN) mitigates this issue through its novel architecture, demonstrating certain advantages. On the other hand, insights from the information bottleneck theory suggest that high-resolution training points are essential for these hybrid methods to accurately capture multi-scale behavior, although this requirement often leads to longer training times. To address this challenge, we propose a simple yet effective multi-resolution training-enhanced PIKAN framework, termed MR-PIKAN, which trains the data-physics hybrid model either sequentially or alternately across different resolutions. The proposed MR-PIKAN is validated on various multi-scale forward and inverse PDE problems. Numerical results indicate that this new training strategy effectively reduces computational costs without sacrificing accuracy, thereby enabling efficient solutions of complex multi-scale PDEs in both forward and inverse settings.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 21 canonical work pages

  1. [1]

    Communications in Numerical Methods in Engineering 10(3), 195–201 (1994) https://doi.org/10.1002/cnm.1640100303

    Dissanayake, M.W.M.G., Phan-Thien, N.: Neural-network-based approximations for solving partial differential equations. Communications in Numerical Methods in Engineering 10(3), 195–201 (1994) https://doi.org/10.1002/cnm.1640100303

    doi:10.1002/cnm.1640100303 1994

  2. [2]

    IEEE Transactions on Neural Networks 9(5), 987–1000 (1998) https://doi.org/10.1109/72.712178

    Lagaris, I.E., Likas, A., Fotiadis, D.I.: Artificial neural networks for solving ordi- nary and partial differential equations. IEEE Transactions on Neural Networks 9(5), 987–1000 (1998) https://doi.org/10.1109/72.712178

    doi:10.1109/72.712178 1998

  3. [3]

    Communications in Mathematics and Statistics 6(1), 1–12 (2018) https://doi.org/10.1007/s40304-018-0127-z

    Weinan, E., Yu, B.: The deep ritz method: A deep learning-based numerical algorithm for solving variational problems. Communications in Mathematics and Statistics 6(1), 1–12 (2018) https://doi.org/10.1007/s40304-018-0127-z

    doi:10.1007/s40304-018-0127-z 2018

  4. [4]

    Nature Reviews Physics 3(6), 422–440 (2021) https://doi.org/10.1038/s42254-021-00314-5

    Karniadakis, G.E., Kevrekidis, I.G., Lu, L., Perdikaris, P., Wang, S., Yang, L.: Physics-informed machine learning. Nature Reviews Physics 3(6), 422–440 (2021) https://doi.org/10.1038/s42254-021-00314-5

    doi:10.1038/s42254-021-00314-5 2021

  5. [5]

    arXiv preprint arXiv:1711.10561 (2017)

    Raissi, M., Perdikaris, P., Karniadakis, G.E.: Physics informed deep learning (part i): Data-driven solutions of nonlinear partial differential equations. arXiv preprint arXiv:1711.10561 (2017)

    arXiv 2017

  6. [6]

    Raissi, M., Perdikaris, P., Karniadakis, G.E.: Physics informed deep learning (part ii): Data-driven discovery of nonlinear partial differential equations. arXiv preprint arXiv:1711.10566 (2017) 26 Implementation details: Inverse Gray-Scott Model Network approximation [ uθ (x, y, t) , vθ (x, y, t)] Network architecture KAN Number of hidden layers 2 Neuron...

    arXiv 2017

  7. [7]

    arXiv preprint arXiv:1912.00873 (2019)

    Kharazmi, E., Zhang, Z., Karniadakis, G.E.: Variational physics-informed neural networks for solving partial differential equations. arXiv preprint arXiv:1912.00873 (2019)

    arXiv 2019

  8. [8]

    Computer Methods in Applied Mechanics and Engineering 362, 112790 (2020) https://doi.org/10.1016/j.cma.2019.112790

    Samaniego, E., Anitescu, C., Goswami, S., Nguyen-Thanh, V.M., Guo, H., Ham- dia, K., Zhuang, X., Rabczuk, T.: An energy approach to the solution of pdes via ml. Computer Methods in Applied Mechanics and Engineering 362, 112790 (2020) https://doi.org/10.1016/j.cma.2019.112790

    work page arXiv 2020

Show all 53 references
  1. [9]

    IEEE Transactions on Neural Networks 4(6), 910–918 (1993) https://doi.org/10.1109/72.286886

    Chen, T.: Approximations of continuous functionals by neural networks with application to dynamic systems. IEEE Transactions on Neural Networks 4(6), 910–918 (1993) https://doi.org/10.1109/72.286886

    1993 doi

  2. [10]

    In: Proceedings of the 36th International Conference on Machine Learning, vol

    Rahaman, N., Baratin, A., Arpit, D., Draxler, F., Lin, M., Hamprecht, F., Bengio, Y., Courville, A.: On the spectral bias of neural networks. In: Proceedings of the 36th International Conference on Machine Learning, vol. 97, pp. 5301–5310. PMLR, (2019)

    2019

  3. [11]

    Computer Methods in Applied Mechanics and Engineering384, 113938 (2021) https://doi.org/10.1016/j.cma.2021.113938

    Wang, S., Wang, H., Perdikaris, P.: On the eigenvector bias of fourier feature networks. Computer Methods in Applied Mechanics and Engineering384, 113938 (2021) https://doi.org/10.1016/j.cma.2021.113938

    2021

  4. [12]

    Journal of Computational Physics 404, 109136 (2020) https://doi.org/10.1016/j.jcp.2019

    Jagtap, A.D., Kawaguchi, K., Karniadakis, G.E.: Adaptive activation functions accelerate convergence in deep and physics-informed neural networks. Journal of Computational Physics 404, 109136 (2020) https://doi.org/10.1016/j.jcp.2019. 109136 27 Implementation details: Hidden e...

    2020 doi

  5. [13]

    arXiv preprint arXiv:2009.14597 (2020)

    Li, X.-A., Xu, Z.-Q.J., Zhang, L.: A multi-scale dnn algorithm for nonlinear elliptic equations with multiple scales. arXiv preprint arXiv:2009.14597 (2020)

    2020 arXiv

  6. [14]

    arXiv preprint arXiv:2410.00053 (2024)

    Huang, J., You, R., Zhou, T.: Frequency-adaptive multi-scale deep neural networks. arXiv preprint arXiv:2410.00053 (2024)

    2024 arXiv

  7. [15]

    arXiv preprint arXiv:2404.19756 (2024)

    Liu, Z., Wang, Y., Vaidya, S., Ruehle, F., Halverson, J., Soljaˇ ci´ c, M., Hou, T.Y., Tegmark, M.: Kan: Kolmogorov-arnold networks. arXiv preprint arXiv:2404.19756 (2024)

    2024 arXiv

  8. [16]

    arXiv preprint arXiv:2408.10205 (2024)

    Liu, Z., Ma, P., Wang, Y., Matusik, W., Tegmark, M.: Kan 2.0: Kolmogorov- arnold networks meet science. arXiv preprint arXiv:2408.10205 (2024)

    2024 arXiv

  9. [17]

    Computer Methods in Applied Mechanics and Engineering 431, 117290 (2024) https://doi.org/10.1016/j.cma.2024.117290

    Shukla, K., Toscano, J.D., Wang, Z., Zou, Z., Karniadakis, G.E.: A comprehen- sive and fair comparison between mlp and kan representations for differential equations and operator networks. Computer Methods in Applied Mechanics and Engineering 431, 117290 (2024) https://doi.org...

    2024

  10. [18]

    arXiv preprint arXiv:2405.07200 (2024)

    Sidharth, S., Keerthana, A., Gokul, R., Anas, K.: Chebyshev polynomial-based kolmogorov-arnold networks: An efficient architecture for nonlinear function approximation. arXiv preprint arXiv:2405.07200 (2024)

    2024 arXiv

  11. [19]

    arXiv preprint arXiv:2405.06721 (2024)

    Li, Z.: Kolmogorov-arnold networks are radial basis function networks. arXiv preprint arXiv:2405.06721 (2024)

    2024 arXiv

  12. [20]

    arXiv 28 preprint arXiv:2405.12832 (2024)

    Bozorgasl, Z., Chen, H.: Wav-kan: Wavelet kolmogorov-arnold networks. arXiv 28 preprint arXiv:2405.12832 (2024)

    2024 arXiv

  13. [21]

    arXiv preprint arXiv:2406.11045 (2024)

    Wang, Y., Sun, J., Bai, J., Anitescu, C., Eshaghi, M.S., Zhuang, X., Rabczuk, T., Liu, Y.: Kolmogorov arnold informed neural network: A physics-informed deep learning framework for solving pdes based on kolmogorov arnold networks. arXiv preprint arXiv:2406.11045 (2024)

    2024 arXiv

  14. [22]

    SIAM Review 63(1), 208–228 (2021) https:// doi.org/10.1137/19M1274067

    Lu, L., Meng, X., Mao, Z., Karniadakis, G.E.: Deepxde: A deep learning library for solving differential equations. SIAM Review 63(1), 208–228 (2021) https:// doi.org/10.1137/19M1274067

    2021 doi

  15. [23]

    Computer Methods in Applied Mechanics and Engineering403, 115671 (2023) https://doi.org/10.1016/j.cma.2022.115671

    Wu, C., Zhu, M., Tan, Q., Kartha, Y., Lu, L.: A comprehensive study of non-adaptive and residual-based adaptive sampling for physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering403, 115671 (2023) https://doi.org/10.1016/j.cma.2022.115671

    2023

  16. [24]

    Applied Mathematics and Mechanics 44(7), 1069–1084 (2023) https://doi.org/10.1007/s10483-023-2994-7

    Mao, Z., Meng, X.: Physics-informed neural networks with residual/gradient- based adaptive sampling methods for solving partial differential equations with sharp solutions. Applied Mathematics and Mechanics 44(7), 1069–1084 (2023) https://doi.org/10.1007/s10483-023-2994-7

    2023 doi

  17. [25]

    SIAM Journal on Scientific Computing 45(4), 1971–1994 (2023) https://doi.org/10

    Gao, Z., Yan, L., Zhou, T.: Failure-informed adaptive sampling for pinns. SIAM Journal on Scientific Computing 45(4), 1971–1994 (2023) https://doi.org/10. 1137/22M1527763

    2023

  18. [26]

    arXiv preprint arXiv:2407.17611 (2024)

    Rigas, S., Papachristou, M., Papadopoulos, T., Anagnostopoulos, F., Alexan- dridis, G.: Adaptive training of grid-dependent physics-informed kolmogorov- arnold networks. arXiv preprint arXiv:2407.17611 (2024)

    2024 arXiv

  19. [27]

    Computer-Aided Civil and Infrastruc- ture Engineering 36(8), 962–977 (2021) https://doi.org/10.1111/mice.12685

    Nabian, M.A., Gladstone, R.J., Meidani, H.: Efficient training of physics-informed neural networks via importance sampling. Computer-Aided Civil and Infrastruc- ture Engineering 36(8), 962–977 (2021) https://doi.org/10.1111/mice.12685

    2021 doi

  20. [28]

    Journal of Computational Physics 476, 111868 (2023) https://doi.org/10.1016/j.jcp.2022.111868

    Tang, K., Wan, X., Yang, C.: Das-pinns: A deep adaptive sampling method for solving high-dimensional partial differential equations. Journal of Computational Physics 476, 111868 (2023) https://doi.org/10.1016/j.jcp.2022.111868

    2023

  21. [29]

    Computer Methods in Applied Mechanics and Engineering 419, 116666 (2024) https://doi.org/10.1016/ j.cma.2023.116666

    Aldirany, Z., Cottereau, R., Laforest, M., Prudhomme, S.: Multi-level neural net- works for accurate solutions of boundary-value problems. Computer Methods in Applied Mechanics and Engineering 419, 116666 (2024) https://doi.org/10.1016/ j.cma.2023.116666

    2024

  22. [30]

    Journal of Computational Physics 504, 112865 (2024) https: //doi.org/10.1016/j.jcp.2024.112865

    Wang, Y., Lai, C.-Y.: Multi-stage neural networks: Function approximator of machine precision. Journal of Computational Physics 504, 112865 (2024) https: //doi.org/10.1016/j.jcp.2024.112865

    2024

  23. [31]

    Journal of Computational Physics 520, 113438 (2025) https://doi.org/10.1016/j.jcp.2024.113438

    Song, S., Mukerji, T., Zhang, D.: Physics-informed multi-grid neural operator: 29 Theory and an application to porous flow simulation. Journal of Computational Physics 520, 113438 (2025) https://doi.org/10.1016/j.jcp.2024.113438

    2025

  24. [32]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Wu, C.-Y., Girshick, R., He, K., Feichtenhofer, C., Krahenbuhl, P.: A multigrid method for efficiently training video models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 153–162 (2020). https://doi.org/10.1109/CVPR42600.2020.00023

    2020

  25. [33]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Lin, L., Wang, X., Qi, Z., Shan, Y.: Accelerating the training of video super-resolution models. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37(2), pp. 1595–1603 (2023)

    2023

  26. [34]

    SIAM Journal on Scientific Computing 43(5), 3055–3081 (2021) https://doi.org/10.1137/20M1318043

    Wang, S., Teng, Y., Perdikaris, P.: Understanding and mitigating gradient pathologies in pinns. SIAM Journal on Scientific Computing 43(5), 3055–3081 (2021) https://doi.org/10.1137/20M1318043

    2021 doi

  27. [35]

    Journal of Computational Physics 449 (2022) https: //doi.org/10.1016/j.jcp.2021.110768

    Wang, S., Yu, X., Perdikaris, P.: When and why pinns fail to train: A neural tangent kernel perspective. Journal of Computational Physics 449 (2022) https: //doi.org/10.1016/j.jcp.2021.110768

    2022

  28. [36]

    Journal of Computational Physics 474, 111722 (2023) https://doi.org/10

    McClenny, L.D., Braga-Neto, U.M.: Self-adaptive physics-informed neural net- works. Journal of Computational Physics 474, 111722 (2023) https://doi.org/10. 1016/j.jcp.2022.111722

    2023

  29. [37]

    Computer Meth- ods in Applied Mechanics and Engineering 421, 116805 (2024) https://doi.org/ 10.1016/j.cma.2024.116805

    Anagnostopoulos, S.J., Toscano, J.D., Stergiopulos, N., Karniadakis, G.E.: Residual-based attention in physics-informed neural networks. Computer Meth- ods in Applied Mechanics and Engineering 421, 116805 (2024) https://doi.org/ 10.1016/j.cma.2024.116805

    2024

  30. [38]

    In: Doklady Akademii Nauk SSSR, vol

    Kolmogorov, A.N.: On the representation of continuous functions of many vari- ables by superposition of continuous functions of one variable and addition. In: Doklady Akademii Nauk SSSR, vol. 114(5), pp. 953–956 (1957). Russian Academy of Sciences

    1957

  31. [39]

    arXiv preprint arXiv:2403.18494 (2024)

    Anagnostopoulos, S.J., Toscano, J.D., Stergiopulos, N., Karniadakis, G.E.: Learn- ing in pinns: Phase transition, total diffusion, and generalization. arXiv preprint arXiv:2403.18494 (2024)

    2024 arXiv

  32. [40]

    Springer Series in Com- putational Mathematics, vol

    Hackbusch, W.: Multi-Grid Methods and Applications. Springer Series in Com- putational Mathematics, vol. 4. Springer, (1985). https://doi.org/10.1007/ 978-3-662-02427-0

    1985

  33. [41]

    In: Teh, Y.W., Titterington, M

    Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedfor- ward neural networks. In: Teh, Y.W., Titterington, M. (eds.) Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics. Proceedings of Machine Learning Researc...

    2010

  34. [42]

    In: Advances in Neural Information Processing Systems, vol

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chin- tala, S.: Pytorch: An imperat...

    2019

  35. [43]

    In: Inter- national Conference on Learning Representations (ICLR2015)

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: Inter- national Conference on Learning Representations (ICLR2015). Ithaca, NY, (2015)

    2015

  36. [44]

    Oxford University Press, (1990)

    Gray, P., Scott, S.K.: Chemical Oscillations and Instabilities: Non-linear Chemical Kinetics. Oxford University Press, (1990)

    1990

  37. [45]

    Pafnuty Publications, Oxford (2014)

    Driscoll, T.A., Hale, N., Trefethen, L.N.: Chebfun guide. Pafnuty Publications, Oxford (2014)

    2014

  38. [46]

    Journal of Computational Physics 176(2), 430–455 (2002)

    Cox, S.M., Matthews, P.C.: Exponential time differencing for stiff systems. Journal of Computational Physics 176(2), 430–455 (2002)

    2002

  39. [47]

    Renewable Energy 220, 119565 (2024) https://doi.org/10.1016/j.renene.2023.119565

    Liu, B., Wang, Y., Rabczuk, T., Olofsson, T., Lu, W.: Multi-scale modeling in thermal conductivity of polyurethane incorporated with phase change materials using physics-informed neural networks. Renewable Energy 220, 119565 (2024) https://doi.org/10.1016/j.renene.2023.119565

    2024

  40. [48]

    Annual Review of Materials Research 45(1), 391–431 (2015) https://doi.org/10.1146/annurev-matsci-070214-021103

    Tasan, C.C., Diehl, M., Yan, D., Bechtold, M., Roters, F., Schemmann, L., Zheng, C., Peranio, N., Ponge, D., Koyama, M.: An overview of dual-phase steels: advances in microstructure-oriented processing and micromechanically guided design. Annual Review of Materials Research 45...

    2015 doi

  41. [49]

    In: International Design Engineering Technical Conferences and Computers and Information in Engineering Conference, vol

    Iyer, A., Zhang, Y., Prasad, A., Tao, S., Wang, Y., Schadler, L., Brinson, L.C., Chen, W.: Data-centric mixed-variable bayesian optimization for materials design. In: International Design Engineering Technical Conferences and Computers and Information in Engineering Conference...

    2019

  42. [50]

    Engineering Applications of Artificial Intelligence 127 (2024) https://doi.org/10.1016/j.engappai.2023

    Ren, X., Lyu, X.: Mixed form based physics-informed neural networks for per- formance evaluation of two-phase random materials. Engineering Applications of Artificial Intelligence 127 (2024) https://doi.org/10.1016/j.engappai.2023. 107250

    2024 doi

  43. [51]

    Experimental Mechanics 55(6), 1105–1122 (2015) https: //doi.org/10.1007/s11340-015-0009-1

    Blaber, J., Adair, B., Antoniou, A.: Ncorr: Open-source 2d digital image corre- lation matlab software. Experimental Mechanics 55(6), 1105–1122 (2015) https: //doi.org/10.1007/s11340-015-0009-1

    2015 doi

  44. [52]

    Proceedings of the National Academy of Sciences 118(31), 2102721118 (2021) 31 https://doi.org/10.1073/pnas.2102721118

    Chen, C.-T., Gu, G.X.: Learning hidden elasticity with deep neural networks. Proceedings of the National Academy of Sciences 118(31), 2102721118 (2021) 31 https://doi.org/10.1073/pnas.2102721118

    2021 doi

  45. [53]

    Springer, (2016) 32

    Langtangen, H.P., Pedersen, G.K.: Scaling of Differential Equations. Springer, (2016) 32

    2016

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.