Pith. sign in

REVIEW 3 major objections 7 minor 45 references

Surrogate Modeling of Landau Damping with Deep Operator Networks

T0 review · 3 major / 7 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 A Deep Operator Network trained on Vlasov-Poisson data can act as a surrogate for Landau damping, predicting electric field energy time series with sub-percent mean relative L2 errors in both linear and nonlinear regimes.

desk verdict A clean but narrow interpolation study: DeepONets fit Landau damping energy over temperature, but the headline nonlinear-accuracy claim is not backed by the reported error metric. read the letter →

arxiv 2507.16960 v2 pith:DAA5IERK submitted 2025-07-22 physics.plasm-ph astro-ph.HEastro-ph.IMastro-ph.SRphysics.comp-phphysics.space-ph

Simin Shekarpaz , Chuanfei Dong , Ziyu Huang This is my paper · ORCID
classification physics.plasm-phastro-ph.HEastro-ph.IMastro-ph.SRphysics.comp-phphysics.space-ph
keywords LandaudampingDeepoperatornetworkslearningVlasov-Poissonequationssurrogatemodelingelectricfieldenergynonlinearplasmadynamicskineticsimulation
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

Landau damping is a basic collisionless plasma process whose kinetic simulations are computationally expensive. This paper argues that a trained Deep Operator Network can serve as a surrogate: given a plasma temperature, it returns the time evolution of the electric field energy without solving the Vlasov-Poisson equations. The claim is demonstrated for a single damped mode and for a five-mode mixture whose large perturbations push the system into nonlinear Landau damping, with electron phase-space holes and oscillatory energy exchange. If the claim holds, temperature parameter studies that currently require a kinetic run per sample could be replaced by a single trained model that evaluates new cases in milliseconds at mean relative L2 errors near or below one percent.

What carries the argument

The central object is the DeepONet $G_\omega$, an operator network with a branch sub-network that encodes the temperature input $T$ at sensor points and a trunk sub-network that encodes time coordinates $t_j$; the outputs combine to approximate $E(t)=\int |E_x(x,t)|^2\,dx$, the electric field energy. Training minimizes the mean squared error between the network output and reference time series, using Adam with an exponentially decaying learning rate, tanh activations, and fully connected hidden layers of width 200 and depth 6. The reference data are generated by a continuum Vlasov-Poisson solver with fixed perturbation modes (Table 1), and the trained network is evaluated on temperatures in $[0.5,1.5]$ not seen during training. The operator-learning construction is what allows a single model to predict entire energy time series for new temperature values without re-running a kinetic simulation.

What would settle it

Take the trained single-mode model and evaluate it at a temperature outside the training range, say $T=0.3$ or $T=2.0$, or at a different perturbation amplitude with the same wavenumber; compare the prediction with a fresh Vlasov-Poisson run at that parameter. If the relative L2 error rises far above the reported test mean of 0.0083, the surrogate is interpolating its training domain rather than having learned the damping law itself.

Watch

Extended reading notes

Core claim

The central claim is that the operator mapping temperature to electric field energy is learnable by DeepONets with accuracy comparable to fully kinetic first-principles simulations. In the single-mode case, 200 training samples produce a mean relative L2 error of 0.0078 on training data and 0.0083 on unseen test data. In the five-mode case, scaling the training set from 50 to 800 samples lowers the test error from 0.0312 to 0.0043; with 400 training samples the test error is 0.0049. The five-mode setup includes nonlinear Landau damping, so the paper is establishing that a surrogate can follow not just the monotonic linear decay but the oscillatory, phase-space-hole-driven energy evolution that follows large initial perturbations.

Load-bearing premise

The load-bearing premise is that the reference Vlasov-Poisson data faithfully represent Landau damping, and that with the fixed perturbation modes and amplitudes, the scalar temperature in $[0.5,1.5]$ determines the electric field energy time series smoothly enough for the network to interpolate it.

Editorial extensions

If this is right

  • A trained model predicts electric field energy time series for previously unseen temperatures in $[0.5,1.5]$ with mean relative L2 errors below roughly one percent in both single- and five-mode tests.
  • Evaluating 100 five-mode test cases took 0.00148 seconds on one GPU, so a temperature sweep that would require many kinetic simulations can be performed nearly instantly once training is done.
  • The five-mode runs show that the surrogate preserves the nonlinear signature of Landau damping, including oscillatory energy exchange associated with electron phase-space holes, rather than only reproducing monotonic linear decay.
  • Test error in the five-mode case drops from 0.0312 with 50 training samples to 0.0043 with 800 samples, indicating that accuracy improves steadily as more reference data are supplied.

Reading between the lines

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

  • Beyond the paper, a natural next test is to feed the branch network a parameter vector containing temperatures, wavenumbers, and amplitudes; the same architecture should handle it if the energy functional depends smoothly on those parameters, but that remains to be shown.
  • Beyond the paper, the reported errors are in-range interpolation errors; out-of-range temperatures or initial conditions outside the fixed mode set could break the accuracy, so the conservative reading is that the surrogate is a fast interpolator on the explored parameter domain.
  • Beyond the paper, the operator formulation is not tied to Landau damping, so the same training recipe could transfer to other kinetic plasma phenomena, such as ion-acoustic waves or driven instabilities, provided reference data that resolve those dynamics exist.
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

3 major / 7 minor

Summary. This manuscript constructs DeepONet surrogates that map a scalar temperature parameter T in [0.5, 1.5] to the electric-field-energy time series E(t) for Vlasov-Poisson Landau damping. The training data are generated with the Gkeyll continuum Vlasov code for two configurations: a single mode (k lambda_e = 0.35, A = 0.05) evolved to t = 20 omega_pe^{-1}, and a five-mode superposition (Table 1) evolved to t = 40 omega_pe^{-1}. The authors report mean relative L2 errors (Eq. 9) below 1% on held-out temperature samples and interpret this as evidence that DeepONets capture both linear and nonlinear Landau damping. The paper also reports loss convergence and an error-versus-training-set-size study for the five-mode case.

Significance. If the accuracy claim were supported, this would be a useful proof-of-concept for replacing repeated kinetic Vlasov simulations with a cheap surrogate for parameter studies. The paper's strengths are its use of first-principles kinetic reference data, a clean train/test split, and a systematic study of training-set size. The caveat is that the reported error metric is dominated by the early linear phase, so the headline claim about nonlinear-regime accuracy is not yet substantiated; the scope of the 'generalization' claim also exceeds the experiments. With phase-resolved error reporting and more cautious wording, the result would be a solid, if narrow, contribution to surrogate modeling of kinetic plasma physics.

major comments (3)
  1. [Section 4, Eq. (9) and Tables 3-4] The relative L2 error is computed over the full time interval, but because E(t) decays substantially in the linear phase, both the numerator and denominator of Eq. (9) are dominated by the early, larger-amplitude portion of the time series. Late-time trapped-particle oscillations and phase-space-hole dynamics therefore contribute almost nothing to the reported mean errors. The claim in Section 5 that the model 'accurately captures the evolution of electric field energy in both linear and nonlinear regimes' is thus not tested by the tabulated metric. Please report time-windowed relative L2 errors restricted to the nonlinear epoch (e.g., after the linear decay has saturated), or normalized errors computed separately for the linear and nonlinear phases, and include pointwise error time series so the nonlinear phase is visible.
  2. [Section 5 (Conclusion) and Abstract] The statement that the model generalizes 'across varying initial conditions and perturbations' overstates what is demonstrated. The only varying parameter is temperature in [0.5, 1.5]; the initial perturbation amplitudes and wavenumbers are fixed (Table 1 and Eq. 8). The evidence supports interpolation over temperature for this specific fixed perturbation set, not generalization over perturbations. Please either qualify the conclusion and abstract or add experiments that vary A_i and k_i.
  3. [Section 4.2, final paragraph] The claimed 'significant speedup compared to the conventional numerical solver' is not quantified: no Gkeyll wall-clock time is reported for the test dataset, and no baseline surrogate (e.g., a linear-damping analytic model, principal-component regression, or a standard MLP) is compared. Without a baseline, the reader cannot tell whether the DeepONet architecture is essential to the reported accuracy or whether the small errors simply reflect a smooth, low-dimensional temperature-to-E(t) map. Please add a baseline comparison and report Gkeyll runtimes for the same test cases.
minor comments (7)
  1. [Section 3.2] Please specify the Gkeyll numerical configuration (domain length, velocity-grid resolution, time step, boundary conditions) and provide a convergence check for the five-mode nonlinear case, since the accuracy of the reference data is the basis for the reported errors.
  2. [Section 2 and Figure 1] The branch input is described as 'temperature values,' but temperature is a scalar parameter rather than a function over a spatial domain; please clarify how the branch sensors are defined (e.g., a constant function sampled at m points) so that the operator-learning formulation is precise.
  3. [Eq. (1)] The operator N in N(T,E) = 0 is not defined; please state explicitly that it denotes the Vlasov-Poisson residual operator, and define the function spaces for T and E.
  4. [Table 4] Unlike Table 3, Table 4 reports only mean errors for the five-mode case; please add standard deviations or min/max ranges for each training-set size, since mean errors alone may hide poor per-sample performance in the nonlinear regime.
  5. [Figures 3 and 5] Because the vertical axes are logarithmic, late-time differences are visually compressed; please add linear-scale insets or error-versus-time panels for the nonlinear phase so the reader can assess the late-time agreement directly.
  6. [Throughout] Please correct minor language issues: 'notable for by their ability' in Section 5, 'permitivity' in Section 3.1, and 'dynamical evolution' style repetitions.
  7. [Section 4.2] For reproducibility, please state the number of epochs, batch size, and hardware details for the reported per-epoch training time, and make the trained models and datasets available or specify a repository.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the surrogate is fitted to independent Gkeyll simulation data, and held-out temperature interpolation is standard validation, not a self-referential derivation.

full rationale

The paper's central claim is that a DeepONet trained on Gkeyll Vlasov-Poisson data can interpolate electric field energy over temperature. The reference data are external (Gkeyll, Juno et al. 2018), not produced by the neural network or by a self-citation. The word 'prediction' refers to test samples from the same temperature range, which is a standard train/test split, and the loss (Eq. 10) and error norm (Eq. 9) compare network output directly against reference simulations. No equation in the paper defines a target quantity in terms of the network output, and no fitted parameter is renamed as a prediction. The cited prior work by the authors, such as Huang et al. 2025 for phase-space holes, is contextual background rather than load-bearing evidence for the accuracy claim. The concern that the full-interval relative L2 error is dominated by the early linear damping phase is a legitimate metric-selection issue for the nonlinear-regime claim, but it is not circularity. The derivation chain is therefore self-contained with respect to the data and the validation protocol, and no significant circularity is present.

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

The central claim depends on the fidelity of Gkeyll Vlasov-Poisson data and on standard supervised-learning assumptions; no new physical entities, forces, or conserved quantities are introduced. The main free choices are neural network hyperparameters, time sampling, and sample counts.

free parameters (4)
  • Neural network hyperparameters (depth, width, activation) = depth=6, width=200, tanh
    Chosen by hand in Table 2; not derived from plasma physics.
  • Optimizer settings = Adam, initial learning rate 0.001, exponential decay, 1e6 iterations
    Chosen by hand in Section 4; standard choices for DeepONet training.
  • Time sampling step Delta t = 0.002 omega_pe^-1
    Chosen for output discretization in Sections 4.1 and 4.2; not physics-derived.
  • Training/test sample counts = single-mode 200/50; five-mode 400/100 (also 50/12, 200/50, 800/200)
    Selected by authors in Tables 2 and 4; affects reported errors.
assumptions (5)
  • domain assumption Vlasov-Poisson equations correctly describe collisionless electrostatic plasma dynamics
    Invoked in Section 3.1, Equations (2)-(4), as the basis for data generation.
  • domain assumption Gkeyll continuum Vlasov solver produces sufficiently accurate reference data
    Used as ground truth in Section 3.2 and all error calculations.
  • domain assumption Immobile ion neutralizing background approximation is valid
    Stated in Section 3.2 via n_i(x)=n_0; simplifies the problem to electron dynamics.
  • standard math Universal approximation theorem guarantees the DeepONet architecture can represent the T-to-E(t) mapping
    Cited in Section 2 (Hornik et al. 1989) as the foundation of DeepONets.
  • domain assumption Scalar temperature T in [0.5,1.5] is the only varying input and fully determines the output for fixed perturbations
    The branch network encodes only T; initial modes are fixed per case in Sections 3.2 and 4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Surrogate Modeling of Landau Damping with Deep Operator Networks." pith.science (2026). https://pith.science/paper/DAA5IERK

@misc{pith2026250716960,
  author       = {Pith},
  title        = {Pith review of: Surrogate Modeling of Landau Damping with Deep Operator Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DAA5IERK}},
  note         = {Machine review of arXiv:2507.16960}
}
read the original abstract

Kinetic simulations excel at capturing microscale plasma physics phenomena with high accuracy, but their computational demands make them impractical for modeling large-scale space and astrophysical systems. In this context, we build a surrogate model, using Deep Operator Networks (DeepONets), based upon the Vlasov-Poisson simulation data to model the dynamical evolution of plasmas, focusing on the Landau damping process - a fundamental kinetic phenomenon in space and astrophysical plasmas. The trained DeepONets are able to capture the evolution of electric field energy in both linear and nonlinear regimes under various conditions. Extensive validation highlights DeepONets' robust performance in reproducing complex plasma behaviors with high accuracy, paving the way for large-scale modeling of space and astrophysical plasmas.

Figures

Figures reproduced from arXiv: 2507.16960 by the authors.

Figure 1
Figure 1. The DeepONets architecture consists of two fully connected neural networks (FNN): the branch network, which encodes temperature values, T, and the trunk net￾work, which encodes the coordinates (here the coordinates are time, t). The outputs of these networks are combined to approximate the electric field energy, R |Ex| 2 dx. The final output is Gω(T)(t), where ω represents the model’s learned weights and T is a vect… view at source ↗
Figure 2
Figure 2. The dispersion relation, i.e., the real and imaginary frequencies versus wavenumber, of the least damped mode in a uniform, electrostatic plasma with immobile ion background and temperature T = 1. The blue dot identifies the wavenumber k = 0.35 used for our single-mode case. The frequencies and wavenumbers are normalized over electron plasma frequency, ωpe, and electron Debye length, λe, respectively [PITH_FULL_IMA… view at source ↗
Figure 3
Figure 3. Single-mode case: comparison of DeepONets predictions with reference solutions obtained from the Gkeyll Vlasov model for a range of temperature values (T). The vertical axis shows the electric field energy, R |Ex| 2 dx, on a logarithmic scale, and the horizontal axis represents time in ω −1 pe [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Single-mode case: training and test loss functions. 4.2. Five-mode Case Now, let us consider the initial condition consists of five modes, with the corresponding values listed in [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Five-mode case: comparison between DeepONet predictions and reference Gkeyll Vlasov solutions for varying temperatures (T). The vertical axis shows the electric field energy, R |Ex| 2 dx, on a logarithmic scale, and the horizontal axis represents time in ω −1 pe [PIT…
Figure 6
Figure 6. Figure 6: Five-mode case: training (left) and test (right) losses as a function of epochs for different training sample sizes (see [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Five-mode case: the left panel shows the error norms for training and testing as a function of the number of training samples. The right panel depicts the training and test losses vs. the number of training samples. The error norms and losses are calculated using Equat…
Figure 8
Figure 8. Figure 8: Five-mode case: histogram (left) and scatter plot (right) of the relative L 2 test errors for the case with 800 training samples and 200 test samples [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 28 canonical work pages

  1. [1]

    P., & Fiuza, F

    Alves, E. P., & Fiuza, F. 2022, Physical Review Research, 4, 033192, doi: 10.1103/PhysRevResearch.4.033192

    work page doi:10.1103/physrevresearch.4.033192 2022

  2. [2]

    Azizmalayeri, M., & Rohban, M. H. 2023, Machine Learning, 112, 3003, doi: 10.1007/s10994-023-06348-3

    work page doi:10.1007/s10994-023-06348-3 2023

  3. [3]

    A., & Karniadakis, G

    Cai, S., Wang, Z., Lu, L., Zaki, T. A., & Karniadakis, G. 2021, Journal of Computational Physics, 436, 110296, doi: 10.1016/j.jcp.2021.110296

    work page arXiv 2021

  4. [4]

    2023, Computer Physics Communications, 287, 108714, doi: 10.1016/j.cpc.2023.108714

    Chen, Y., T´ oth, G., Zhou, H., & Wang, X. 2023, Computer Physics Communications, 287, 108714, doi: 10.1016/j.cpc.2023.108714

    work page arXiv 2023

  5. [5]

    2017, Journal of Geophysical Research: Space Physics, 122, 10,318, doi: 10.1002/2017JA024186

    Chen, Y., T´ oth, G., Cassak, P., et al. 2017, Journal of Geophysical Research: Space Physics, 122, 10,318, doi: 10.1002/2017JA024186

    doi:10.1002/2017ja024186 2017

  6. [6]

    2023, Computer Physics Communications, 282, 108538, doi: https://doi.org/10.1016/j.cpc.2022.108538

    Cheng, W., Fu, H., Wang, L., et al. 2023, Computer Physics Communications, 282, 108538, doi: https://doi.org/10.1016/j.cpc.2022.108538

    work page arXiv 2023

  7. [7]

    Daldorff, L. K. S., T´ oth, G., Gombosi, T. I., et al. 2014, Journal of Computational Physics, 268, 236, doi: 10.1016/j.jcp.2014.03.009

    doi:10.1016/j.jcp.2014.03.009 2014

  8. [8]

    2023, Journal of Plasma Physics, 89, 895890105, doi: 10.1017/S0022377823000119

    Donaghy, J., & Germaschewski, K. 2023, Journal of Plasma Physics, 89, 895890105, doi: 10.1017/S0022377823000119

    work page doi:10.1017/s0022377823000119 2023

Show all 45 references
  1. [9]

    2019, Geophysical Research Letters, 46, 11,584, doi: 10.1029/2019GL083180

    Dong, C., Wang, L., Hakim, A., et al. 2019, Geophysical Research Letters, 46, 11,584, doi: 10.1029/2019GL083180

    2019 doi

  2. [10]

    2023, Engineering with Computers, 39, 2169–2186, doi: 10.1007/s00366-022-01603-y

    Hajimohammadi, Z., Shekarpaz, S., & Parand, K. 2023, Engineering with Computers, 39, 2169–2186, doi: 10.1007/s00366-022-01603-y

    2023 doi

  3. [11]

    W., & Perkins, F

    Hammett, G. W., & Perkins, F. W. 1990, Physical Review Letters, 64, 3019, doi: 10.1103/PhysRevLett.64.3019

    1990 doi

  4. [12]

    1989, Neural Networks, 2 (5), 359, doi: 10.1016/0893-6080(89)90020-8

    Hornik, K., Stinchcombe, M., & White, H. 1989, Neural Networks, 2 (5), 359, doi: 10.1016/0893-6080(89)90020-8

    1989 doi

  5. [13]

    2025, Proceedings of the National Academy of Science, 122, e2419073122, doi: 10.1073/pnas.2419073122

    Huang, Z., Dong, C., & Wang, L. 2025, Proceedings of the National Academy of Science, 122, e2419073122, doi: 10.1073/pnas.2419073122

    2025 doi

  6. [14]

    2025, Journal of Plasma Physics, 91, E64, doi: 10.1017/S0022377825000285

    Pusztai, I. 2025, Journal of Plasma Physics, 91, E64, doi: 10.1017/S0022377825000285

    2025 doi

  7. [15]

    2020, Acta Astronautica, 170, 6, doi: 10.1016/j.actaastro.2020.01.030

    Jarmak, S., Leonard, E., Akins, A., et al. 2020, Acta Astronautica, 170, 6, doi: 10.1016/j.actaastro.2020.01.030

    2020 doi

  8. [16]

    S., & Thomas, A

    Joglekar, A. S., & Thomas, A. G. R. 2023, Machine Learning: Science and Technology, 4, 035049, doi: 10.1088/2632-2153/acf81a

    2023 doi

  9. [17]

    2018, Journal of Computational Physics, 353, 110, doi: 10.1016/j.jcp.2017.10.009

    Juno, J., Hakim, A., TenBarge, J., Shi, E., & Dorland, W. 2018, Journal of Computational Physics, 353, 110, doi: 10.1016/j.jcp.2017.10.009

    2018 doi

  10. [18]

    2023, in Proceedings of the Platform for Advanced Scientific Computing Conference, PASC ’23 (New York, NY, USA: Association for Computing Machinery), doi: 10.1145/3592979.3593412

    Kurth, T., Subramanian, S., Harrington, P., et al. 2023, in Proceedings of the Platform for Advanced Scientific Computing Conference, PASC ’23 (New York, NY, USA: Association for Computing Machinery), doi: 10.1145/3592979.3593412

    2023

  11. [19]

    2022, Physics of Plasmas, 29, 032706, doi: 10.1063/5.0066397

    Laperre, B., Amaya, J., Jamal, S., & Lapenta, G. 2022, Physics of Plasmas, 29, 032706, doi: 10.1063/5.0066397

    2022 doi

  12. [20]

    2023, AIP Advances, 13, 015126, doi: 10.1063/5.0122087

    Li, D., Chen, Y., Dong, C., Wang, L., & Toth, G. 2023, AIP Advances, 13, 015126, doi: 10.1063/5.0122087

    2023 doi

  13. [21]

    2020, arXiv preprint, arXiv:2010.08895, doi: 10.48550/arXiv.2010.08895

    Li, Z., Kovachki, N., Azizzadenesheli, K., et al. 2020, arXiv preprint, arXiv:2010.08895, doi: 10.48550/arXiv.2010.08895

    2020 doi arXiv

  14. [22]

    2021, Nature Machine Intelligence, 3, 218, doi: 10.1038/s42256-021-00302-5

    Lu, L., Jin, P., Pang, G., Zhang, Z., & Karniadakis, G. 2021, Nature Machine Intelligence, 3, 218, doi: 10.1038/s42256-021-00302-5

    2021 doi

  15. [23]

    2020, Physics of Plasmas, 27, doi: /10.1063/1.5129158

    Ma, C., Zhu, B., Xu, X.-Q., & Wang, W. 2020, Physics of Plasmas, 27, doi: /10.1063/1.5129158

    2020 doi

  16. [24]

    2023, The Astrophysical Journal Letters, 950, L12, doi: 10.3847/2041-8213/acd77f

    Mao, S., Dong, R., Lu, L., et al. 2023, The Astrophysical Journal Letters, 950, L12, doi: 10.3847/2041-8213/acd77f

    2023 doi

  17. [25]

    2020, Journal of Computational Physics, 401, 109020, doi: 10.1016/j.jcp.2019.109020

    Meng, X., & Karniadakis, G. 2020, Journal of Computational Physics, 401, 109020, doi: 10.1016/j.jcp.2019.109020

    2020

  18. [26]

    2019, SIAM Journal on Scientific Computing, 41, A2603, doi: 10.1137/18M1229845

    Pang, G., Lu, L., & Karniadakis, G. 2019, SIAM Journal on Scientific Computing, 41, A2603, doi: 10.1137/18M1229845

    2019 doi

  19. [27]

    2020, Engineering with Computers, 36, 789, doi: 10.1007/s00366-020-01186-6

    Parand, K., Razzaghi, M., Sahleh, R., & Jani, M. 2020, Engineering with Computers, 36, 789, doi: 10.1007/s00366-020-01186-6

    2020 doi

  20. [28]

    2019, SeMA Journal, 76, 615, doi: 10.1007/s40324-019-00196-9

    Parand, K., Yari, H., Taheri, R., & Shekarpaz, S. 2019, SeMA Journal, 76, 615, doi: 10.1007/s40324-019-00196-9

    2019 doi

  21. [29]

    2023, Physical Review Research, 5, 033079, doi: 10.1103/PhysRevResearch.5.033079

    Qin, Y., Ma, J., Jiang, M., et al. 2023, Physical Review Research, 5, 033079, doi: 10.1103/PhysRevResearch.5.033079

    2023 doi

  22. [30]

    Raissi, M., Perdikaris, P., & Karniadakis, G. E. 2019, Journal of Computational Physics, 378, 686, doi: 10.1016/j.jcp.2018.10.045

    2019 doi

  23. [31]

    2023, Solving Ordinary Differential Equations by LS-SVM, ed

    Razzaghi, M., Shekarpaz, S., & Rajabi, A. 2023, Solving Ordinary Differential Equations by LS-SVM, ed. J. A

    2023

  24. [32]

    Parand, & S

    Rad, K. Parand, & S. Chakraverty (Singapore: Springer Nature Singapore), 147–170, doi: 10.1007/978-981-19-6553-1 7

    doi

  25. [33]

    2021, in AGU Fall Meeting Abstracts, Vol

    Rulke, E., Wang, L., & Dong, C. 2021, in AGU Fall Meeting Abstracts, Vol. 2021, SM53C–08

    2021

  26. [34]

    2024, International Journal of Heat and Mass Transfer, 219, 124813, doi: 10.1016/j.ijheatmasstransfer.2023.124813

    Sahin, I., Moya, C., Mollaali, A., Lin, G., & Paniagua, G. 2024, International Journal of Heat and Mass Transfer, 219, 124813, doi: 10.1016/j.ijheatmasstransfer.2023.124813

    2024

  27. [35]

    2020, Numerical Methods for Partial Differential Equations, 36, 654, doi: 10.1002/num.22447

    Shekarpaz, S., & Azari, H. 2020, Numerical Methods for Partial Differential Equations, 36, 654, doi: 10.1002/num.22447

    2020 doi

  28. [36]

    Shekarpaz, S., Azizmalayeri, M., & Rohban, M. H. 2022, https://arxiv.org/abs/2207.06647 10

    2022 arXiv

  29. [37]

    2024, Communications in Computational Physics, 35, 1–37, doi: 10.4208/cicp.oa-2023-0121

    Shekarpaz, S., Zeng, F., & Karniadakis, G. 2024, Communications in Computational Physics, 35, 1–37, doi: 10.4208/cicp.oa-2023-0121

    2024 doi

  30. [38]

    H., et al

    Shukla, K., Zou, Z., Chan, C. H., et al. 2025, Computer Methods in Applied Mechanics and Engineering, 433, 117498, doi: https://doi.org/10.1016/j.cma.2024.117498 T´ oth, G., Jia, X., Markidis, S., et al. 2016, Journal of Geophysical Research (Space Physics), 121, 1273, doi: 10...

    2025

  31. [39]

    2018, Journal of Geophysical Research (Space Physics), 123, 2815, doi: 10.1002/2017JA024761

    Wang, L., Germaschewski, K., Hakim, A., et al. 2018, Journal of Geophysical Research (Space Physics), 123, 2815, doi: 10.1002/2017JA024761

    2018 doi

  32. [40]

    2015, Physics of Plasmas, 22, 012108, doi: 10.1063/1.4906063

    Germaschewski, K. 2015, Physics of Plasmas, 22, 012108, doi: 10.1063/1.4906063

    2015 doi

  33. [41]

    2020, Journal of Computational Physics, 415, 109510, doi: 10.1016/j.jcp.2020.109510

    Germaschewski, K. 2020, Journal of Computational Physics, 415, 109510, doi: 10.1016/j.jcp.2020.109510

    2020

  34. [42]

    2022, Journal of Geophysical Research: Space Physics, 127, e2021JA030091, doi: 10.1029/2021JA030091

    Wang, X., Chen, Y., & T´ oth, G. 2022, Journal of Geophysical Research: Space Physics, 127, e2021JA030091, doi: 10.1029/2021JA030091

    2022 doi

  35. [43]

    2023, in 2023 International Applied Computational Electromagnetics Society Symposium (ACES-China), IEEE, 01–03, doi: 10.23919/ACES-China60289.2023.10249492

    Wei, S., Liu, Y., Fu, H., Dong, C., & Wang, L. 2023, in 2023 International Applied Computational Electromagnetics Society Symposium (ACES-China), IEEE, 01–03, doi: 10.23919/ACES-China60289.2023.10249492

    2023

  36. [44]

    2020, SIAM Journal on Scientific Computing, 42, A639, doi: 10.1137/19M1260141

    Zhang, D., Guo, L., & Karniadakis, G. 2020, SIAM Journal on Scientific Computing, 42, A639, doi: 10.1137/19M1260141

    2020 doi

  37. [45]

    2020, Journal of Geophysical Research (Space Physics), 125, e28162, doi: 10.1029/2020JA028162

    Zhou, H., T´ oth, G., Jia, X., & Chen, Y. 2020, Journal of Geophysical Research (Space Physics), 125, e28162, doi: 10.1029/2020JA028162

    2020 doi

Pith tools

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