Pith. sign in

REVIEW 3 major objections 5 minor 26 references

Neural Operators for Immersed-Boundary Soft Swimmers Locomotion

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read A Fourier neural operator predicts the full velocity, vorticity, and pressure fields of immersed-boundary eel swimmers one time step ahead, with 3.51% planar and 3.44% volumetric velocity field error.

desk verdict Honest feasibility study of FNOs for swimmer flow fields; the one-step numbers need a persistence baseline before they carry weight. read the letter →

arxiv 2608.07722 v1 pith:C27WBIXX submitted 2026-08-07 cs.LG physics.flu-dyn

Mohammad Sadegh Eshaghi , Yizheng Wang , Navid Valizadeh , Xiaoying Zhuang , Timon Rabczuk This is my paper · ORCID
classification cs.LGphysics.flu-dyn
keywords NeuraloperatorFourierimmersed-boundarymethodfluid–structureinteractionbio-inspiredswimmersurrogatemodelingcomputationalfluiddynamicsscientificmachinelearning
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

The paper asks whether a Fourier neural operator (FNO) can replace repeated high-fidelity fluid–structure simulations of a swimming eel by directly predicting the surrounding velocity, vorticity, and pressure fields one time step ahead. Its central claim is yes: a planar FNO jointly predicts all four scalar fields with a 3.51% full-domain global relative $L^2$ error on held-out high-Reynolds-number trajectories, and volumetric models reach 3.44% for velocity, 5.58% for vorticity, and 19.2% for pressure on within-range trajectories. The authors stress that these numbers are for teacher-forced one-step prediction, not recursively unrolled trajectories. If the claim holds, a trained surrogate could make design sweeps, parameter studies, and control loops much cheaper by returning entire distributed fields rather than scalar performance measures.

What carries the argument

The load-bearing object is the Fourier neural operator (FNO): a network whose layers mix a local pointwise map with a truncated spectral convolution, allowing it to learn maps between functions sampled on regular grids. Each Fourier block executes $v_{\ell+1} = \sigma_\ell\left(W_\ell v_\ell + M_\ell \mathcal{F}^{-1}(R_\ell \odot \mathcal{F}(v_\ell))\right)$, where $R_\ell$ is a learned complex multiplier on retained modes. Conditioning on the moving swimmer comes from a binary occupancy mask and a broadcast Reynolds number, with phase features $\sin\phi, \cos\phi$ added in the volumetric case, and the temporal task is the reference-state map from fields at $t_n$ to fields at $t_{n+1}$. The volumetric formulation is three target-specific FNOs sharing one multichannel input rather than a joint seven-output network.

What would settle it

Run the published frozen checkpoints recursively: start from one held-out state, feed each prediction back as the next input for dozens of steps, and compare the unrolled fields with the simulation; if global relative error grows well above the reported 3.5% / 3.4% levels or the divergence and vorticity-curl residuals amplify within a few steps, the temporal-prediction claim as a practical surrogate would be falsified.

Watch

Extended reading notes

Core claim

On its own terms, the paper demonstrates that a Fourier neural operator can act as a field-resolved surrogate for the hydrodynamic fields of an immersed-boundary swimmer. A planar joint model maps the current velocity, vorticity, pressure, geometry mask, Reynolds number, and time step to the next velocity, vorticity, and pressure fields, and three volumetric models each predict one target group from a common 15-channel input. Across five held-out planar trajectories at Reynolds numbers 7300–8500, the full-domain global relative $L^2$ error of the joint prediction is 3.51%. Across five held-out volumetric trajectories at Reynolds numbers 750–6200, the errors are 3.44% for velocity, 5.58% for vorticity, and 19.2% for pressure. Pressure is the least accurate field, and the planar model's predicted divergence and vorticity-curl inconsistency are larger than the reference fields', so the authors identify physical consistency and pressure accuracy as the main development priorities.

Load-bearing premise

The load-bearing premise is that teacher-forced one-step evaluation, where the true fields at time $t_n$ are given as input and the model predicts $t_{n+1}$, is a meaningful measure of temporal prediction; the paper does not demonstrate stable recursive rollout.

Editorial extensions

If this is right

  • A single trained surrogate returns distributed velocity, vorticity, and pressure fields for a new Reynolds number at the cost of one forward pass, which is the ingredient a design-optimization or control loop would need to replace many repeated FSI solves.
  • The volumetric models generalize across the tested Reynolds range (750–6200) and the planar model extrapolates mildly to 7300–8500, with error growing gradually rather than collapsing.
  • Velocity and vorticity are accurate enough for wake-structure analysis, while the 19.2% volumetric pressure error means pressure-derived quantities such as loads cannot be taken from the surrogate without dedicated treatment.
  • The reported accuracy is the accuracy of a one-step transition from a known state; it does not by itself certify multi-step temporal prediction.
  • Global field accuracy does not imply differential consistency, since the planar surrogate's divergence and vorticity-curl residuals exceed those of the reference fields.

Reading between the lines

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

  • If the one-step map is stable under feedback, an obvious extension is to train the same architecture with a rollout loss or a recurrent head so it can be unrolled over many time steps; the paper neither claims nor tests that.
  • Because removing the spatial mean barely changes the pressure error, the 19% volumetric pressure error is likely spatial structure rather than a gauge offset, so force-relevant integrals such as thrust or pressure drag might be far more accurate than the pointwise norm suggests.
  • A cheap testable modification is to add an incompressibility penalty or a curl–vorticity consistency term to the training loss; the paper's own diagnostics show the surrogate's divergence and vorticity-curl residuals are the main physics violations.
  • The conditioning recipe (geometry mask plus Reynolds number) could be pushed toward unseen gaits and body shapes with a latent geometry encoder, since the current study varies only Reynolds number for one eel geometry.
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 / 5 minor

Summary. The paper develops Fourier neural operator (FNO) surrogates for one-step temporal prediction of hydrodynamic fields around immersed-boundary eel swimmers. A planar model jointly predicts two velocity components, scalar vorticity, and pressure, while the volumetric formulation uses three separate target-specific FNOs for velocity, vorticity, and pressure. The models are trained on regular-grid fields exported from adaptive IBAMR simulations, conditioned on swimmer geometry masks and Reynolds number, and evaluated on held-out trajectories. Reported full-domain global relative L2 errors are 3.51% for the planar joint model and 3.44%, 5.58%, and 19.2% for volumetric velocity, vorticity, and pressure, respectively. The paper also reports fluid-region errors, component-wise errors, and physical diagnostics showing that predicted fields violate incompressibility and vorticity-velocity consistency more than the target fields. The authors explicitly frame the work as a feasibility study and list several limitations, including one-step-only evaluation and single-seed training.

Significance. If the evaluation protocol is made sound, this is a useful application-scale benchmark: it brings neural operators to moving-boundary FSI data with multi-channel outputs, parameter conditioning, and genuinely volumetric fields. The paper is honest in reporting pressure weakness, physical inconsistency, and the lack of recursive stability, and its error definitions are clear and reproducible in principle. However, the current evidence does not fully support the central claim of temporal prediction skill, because the headline errors come from a checkpoint-selection protocol that uses the held-out trajectories and because no persistence baseline or rollout experiment is reported. The contribution is best read as a data-generation and baseline benchmark that needs strengthened evaluation before the quantitative claims can be accepted.

major comments (3)
  1. [Section 5.2 and Figure 2] The checkpoint-selection protocol uses the held-out trajectories. Section 5.2 states that selection "minimized the held-out mean per-sample joint full-domain relative L2 loss," and Figure 2 shows the monitored held-out curves with the frozen checkpoints marked. The planar split is even labeled "held-out selection/evaluation" in Appendix A. Consequently, the errors in Tables 3 and 5 are minima over training epochs on the same set that is later used for reporting, so they are not independent test estimates and are likely optimistic. The authors list the correct protocol in Section 7.5 ("reserve trajectory-level validation and final test sets"), but the current paper still presents these numbers as held-out generalization results. This should be fixed by retraining with a separate validation set and reporting final-test errors, or by clearly relabeling the reported numbers as validation-selected estimates and acknowledging the bias quantitatively.
  2. [Sections 3.2, 6, and 7.4] The evaluation is teacher-forced one-step prediction only, and the executed time steps are very small (planar t=0.71 to 0.72; volumetric t=0.97384 to 0.98827). Since the input is the reference field at the previous time step, a trivial persistence predictor that simply copies the input would also achieve low relative L2 error on these slowly evolving flows. The paper reports no persistence baseline, no linear-extrapolation baseline, and no recursive rollout. Therefore the reported 3.44%–3.51% velocity and vorticity errors do not by themselves separate learned dynamics from temporal autocorrelation, which is load-bearing for the claim that the surrogate can "temporal[ly] predict" the hydrodynamic fields and eventually replace repeated simulation calls. I request a persistence baseline and at least one short recursive rollout experiment, or a clear statement of why persistence is not a meaningful comparator for this protocol.
  3. [Section 7.4 and Table 5] All experiments use a single training seed, and Figure 6 explicitly notes that each point is not an uncertainty estimate across seeds. The quantitative claims in Tables 3 and 5 are point estimates from one run. Given that the paper emphasizes reproducible evaluation artifacts, the sensitivity of the reported errors to seed and initialization should be quantified, at least for the smaller planar model or for a subset of volumetric runs. Without this, the reported precision (e.g., 3.51% versus 3.44%) cannot be distinguished from training stochasticity.
minor comments (5)
  1. [Section 6.1] In the sentence beginning "Foo all 1000 held-out temporal pairs," the word "Foo" appears to be a typo and should read "For."
  2. [Abstract and Table 5] The abstract reports the volumetric vorticity error as 5.58% and pressure as 19.2%, while Table 5 lists 5.586% and 19.203%; the rounding is fine, but the abstract's "5.59%" in Section 8 should be checked for consistency with Table 5.
  3. [Table 8] The planar forward time is reported at batch size 8 while the volumetric times are at batch size 1; the table caption should state this explicitly to avoid an apples-to-oranges comparison.
  4. [Data and Code Availability] The repository is announced as "will be made publicly available soon," but the paper's conclusion claims that "frozen models and reproducible evaluation artifacts" provide a concrete basis for further work; please clarify the current availability status or provide a permanent snapshot identifier.
  5. [Section 7.3] The sentence "The present masks provide useful geometry conditioning and fluid-region reporting, but a signed-distance field or Lagrangian geometry encoder could better resolve subcell interface location" is clear, but it would help to state whether the binary occupancy proxy M_e is used only as an input channel or also in the loss weighting.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular dependency: the FNO generalization claim rests on held-out trajectory evaluation, and self-citations are related-work only.

full rationale

The derivation chain in this paper is empirical, not analytic: a Fourier neural operator is trained on IBAMR-simulated fields and evaluated on trajectories excluded from training. The reported global relative L2 errors (3.51% planar; 3.44%/5.58%/19.2% volumetric) are computed from Eq. (12) on held-out data, so they are not equal by construction to any training loss or fitted parameter. The teacher-forced one-step protocol is explicitly disclosed in Sections 3.2 and 7.4; it limits the temporal claim but does not make the measured error an input to the model. Checkpoint selection on the monitored held-out set (Section 5.2, Figure 2) is a mild selection-on-validation issue, not circularity, because the reported numbers still quantify generalization to unseen trajectories within the stated protocol. Self-citations [22]-[24] appear only as related work and future directions, not as load-bearing justification for the FNO architecture or the error measurements. No uniqueness theorem, ansatz, or definitional equivalence is invoked. The paper's own limitations section explicitly flags missing experimental validation, recursive stability, and geometry/gait transfer, which are scope restrictions rather than circular steps.

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

No new physical entities are introduced. The central claim rests on the fidelity of the IBAMR simulation data, the choice of fixed-grid masks and normalizers, and the FNO hyperparameters. The learned network weights are the primary fitted quantities.

free parameters (5)
  • Planar FNO weights = 4,208,452 real-scalar parameters
    Learned spectral multipliers, pointwise maps, and projection, optimized with Adam on the joint relative L2 loss (Section 5.2).
  • Volumetric FNO weights (three models) = 254,814,243 each
    Separate learned weights for velocity, vorticity, and pressure models (Section 5.1, Table 2).
  • Channelwise normalizer statistics = not reported in text
    Gaussian mean and standard deviation fit on 8,955 training pairs (planar) and used in affine scaling for volumetric inputs and outputs (Section 5.2).
  • Phase period T = 1
    Chosen encoding period for the sin(phi) and cos(phi) input channels in the volumetric model (Section 3.2), not fitted.
  • Architecture hyperparameters = planar: 16 modes, 32 width, 4 blocks; 3D: 24 modes, 24 width, 4 blocks
    Hand-chosen configuration; no ablation study is reported (Table 2).
assumptions (5)
  • domain assumption The Peskin immersed-boundary formulation (equations 1 to 4) provides a valid model of swimmer hydrodynamics.
    Used to generate the reference fields; if the simulation model is inaccurate, the surrogate inherits the error (Section 3.1).
  • domain assumption IBAMR adaptive simulations are accurate enough to serve as ground truth for training and evaluation.
    All target fields are solver outputs; no experimental validation is provided (Section 7.4).
  • domain assumption The fixed-grid cropped fields and binary mask proxies adequately represent the flow for the surrogate.
    Data are exported on a regular grid with mask proxies, not exact sharp-interface geometry (Section 4, Appendix A).
  • ad hoc to paper The FNO architecture can approximate the one-step field-to-field map on this distribution.
    The paper provides empirical evidence from one seed but no approximation guarantees (Section 5.1).
  • domain assumption Held-out Reynolds numbers (planar extrapolation, volumetric interpolation) are representative of deployment conditions.
    Generalization is assessed only along the Reynolds-number axis (Sections 6.1, 6.2, 7.4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neural Operators for Immersed-Boundary Soft Swimmers Locomotion." pith.science (2026). https://pith.science/paper/C27WBIXX

@misc{pith2026260807722,
  author       = {Pith},
  title        = {Pith review of: Neural Operators for Immersed-Boundary Soft Swimmers Locomotion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C27WBIXX}},
  note         = {Machine review of arXiv:2608.07722}
}
read the original abstract

High-fidelity immersed-boundary simulation resolves the coupled motion of a deforming swimmer and its surrounding flow, but the resulting cost limits repeated evaluations for engineering design, parameter studies, and control. We develop neural-operator surrogates for temporal prediction of the hydrodynamic fields generated by planar and volumetric eel swimmers. The surrogates are trained on regular-grid fields exported from adaptive fluid--structure simulations and are conditioned on swimmer geometry and Reynolds number. The planar model jointly predicts two velocity components, scalar vorticity, and pressure. On five held-out high-Reynolds-number trajectories, its full-domain global relative L^2 error is 3.51 %. The volumetric formulation uses three target-specific models with a common multichannel input: one model predicts three-dimensional velocity, one predicts vorticity, and one predicts pressure. Their full-domain global relative L^2 errors on five held-out within-range trajectories are 3.44 %, 5.58 %, and 19.2 %. Together, the results demonstrate the feasibility of field-resolved neural surrogates for moving-boundary swimmer flows while identifying pressure accuracy and physical consistency as priorities for further development.

Figures

Figures reproduced from arXiv: 2608.07722 by the authors.

Figure 1
Figure 1. Planar swimmer field prediction for held-out run 45 (Re = 7300), from retained input [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Training and monitored held-out relative [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Velocity prediction on an xy slice through the volumetric swimmer field. Rows show ux, uy, uz; columns show target, FNO prediction, and signed error. 5.5 6.0 6.5 x −0.75 −0.50 −0.25 0.00 0.25 0.50 0.75 ωx y Target 5.5 6.0 6.5 x −0.75 −0.50 −0.25 0.00 0.25 0.50 0.75 y FNO prediction 5.5 6.0 6.5 x −0.75 −0.50 −0.25 0.00 0.25 0.50 0.75 y Signed error 5.5 6.0 6.5 x −0.75 −0.50 −0.25 0.00 0.25 0.50 0.75 ωy y Target 5.5 6… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Vorticity prediction on the same xy slice through the volumetric field. Rows show Ωx, Ωy, Ωz; columns show target, FNO prediction, and signed error. 9 [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Pressure target, FNO prediction, and signed error on the same [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Global relative L 2 error versus Reynolds number for the target-specific volumetric models. Full-domain and target-time fluid-mask ratios are reported separately. Each point represents one held-out trajectory, not an uncertainty estimate across seeds. 10 [PITH_FULL_IM…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 14 canonical work pages

  1. [1]

    Charles S. Peskin. The immersed boundary method.Acta Numerica, 11:479–517, 2002. doi: 10.1017/S0962492902000077

    doi:10.1017/s0962492902000077 2002

  2. [2]

    Griffith, Richard D

    Boyce E. Griffith, Richard D. Hornung, David M. McQueen, and Charles S. Peskin. An adaptive, formally second order accurate version of the immersed boundary method.Journal of Computational Physics, 223(1):10–49, 2007. doi: 10.1016/j.jcp.2006.08.019

    work page doi:10.1016/j.jcp.2006.08.019 2007

  3. [3]

    Griffith and Xiaoyu Luo

    Boyce E. Griffith and Xiaoyu Luo. Hybrid finite difference/finite element immersed boundary method.International Journal for Numerical Methods in Biomedical Engineering, 33(12): e2888, 2017. doi: 10.1002/cnm.2888

    work page doi:10.1002/cnm.2888 2017

  4. [4]

    IBAMR: Immersed boundary adaptive mesh refinement software infras- tructure.https://ibamr.github.io/, n.d

    IBAMR Developers. IBAMR: Immersed boundary adaptive mesh refinement software infras- tructure.https://ibamr.github.io/, n.d. Software project, accessed 14 July 2026

    work page 2026

  5. [5]

    Fourier neural operator for parametric partial differential equations

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. InInternational Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=c8P9NQVtmnO

    2021

  6. [6]

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

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators. Nature Machine Intelligence, 3(3):218–229, 2021. doi: 10.1038/s42256-021-00302-5

    doi:10.1038/s42256-021-00302-5 2021

  7. [7]

    Physics-informed neural operator for learn- ing partial differential equations.ACM/IMS Journal of Data Science, 1(3):1–27, 2024

    Zongyi Li, Hongkai Zheng, Nikola Kovachki, David Jin, Haoxuan Chen, Burigede Liu, Kam- yar Azizzadenesheli, and Anima Anandkumar. Physics-informed neural operator for learn- ing partial differential equations.ACM/IMS Journal of Data Science, 1(3):1–27, 2024. doi: 10.1145/3648506. 14

    doi:10.1145/3648506 2024

  8. [8]

    Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew M

    Nikola B. Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew M. Stuart, and Anima Anandkumar. Neural operator: Learning maps between func- tion spaces with applications to PDEs.Journal of Machine Learning Research, 24(89):1–97,

    work page

Show all 26 references
  1. [9]

    Katzschmann, Joseph DelPreto, Robert MacCurdy, and Daniela Rus

    Robert K. Katzschmann, Joseph DelPreto, Robert MacCurdy, and Daniela Rus. Exploration of underwater life with an acoustically controlled soft robotic fish.Science Robotics, 3(16): eaar3449, 2018. doi: 10.1126/scirobotics.aar3449

    2018 doi

  2. [10]

    Zhang, Kui Wu, Andrew Spielberg, Robert K

    Pingchuan Ma, Tao Du, John Z. Zhang, Kui Wu, Andrew Spielberg, Robert K. Katzschmann, and Wojciech Matusik. DiffAqua: A differentiable computational design pipeline for soft underwater swimmers with shape interpolation.ACM Transactions on Graphics, 40(4):132:1– 132:14, 2021. d...

    2021

  3. [11]

    Zhang, Yu Zhang, Pingchuan Ma, Elvis Nava, Tao Du, Philip Arm, Wojciech Matusik, and Robert K

    John Z. Zhang, Yu Zhang, Pingchuan Ma, Elvis Nava, Tao Du, Philip Arm, Wojciech Matusik, and Robert K. Katzschmann. Sim2real for soft robotic fish via differentiable simulation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 12598–126...

    2022

  4. [12]

    Zhang, Mike Yan Michelis, Tao Du, Pingchuan Ma, Benjamin F

    Elvis Nava, John Z. Zhang, Mike Yan Michelis, Tao Du, Pingchuan Ma, Benjamin F. Grewe, Wojciech Matusik, and Robert Kevin Katzschmann. Fast aquatic swimmer opti- mization with differentiable projective dynamics and neural network hydrodynamic mod- els. InProceedings of the 39t...

    2022

  5. [13]

    Michelis, Robert K

    Jeong Hun Lee, Mike Y. Michelis, Robert K. Katzschmann, and Zachary Manchester. Aquar- ium: A fully differentiable fluid–structure interaction solver for robotics applications. In2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11272–11279. IEEE, 202...

    2023

  6. [14]

    Michelis, Lewis S

    Oncay Yasa, Yasunori Toshimitsu, Mike Y. Michelis, Lewis S. Jones, Miriam Filippi, Thomas Buchner, and Robert K. Katzschmann. An overview of soft robotics.An- nual Review of Control, Robotics, and Autonomous Systems, 6:1–29, 2023. doi: 10.1146/ annurev-control-062322-100607

    2023

  7. [15]

    Griffith, and Neelesh A

    Amneet Pal Singh Bhalla, Rahul Bale, Boyce E. Griffith, and Neelesh A. Patankar. A unified mathematical framework and an adaptive numerical method for fluid–structure interaction with rigid, deforming, and elastic bodies.Journal of Computational Physics, 250:446–476,

  8. [16]

    Battaglia

    Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter W. Battaglia. Learning mesh-based simulation with graph networks. InInternational Conference on Learning Repre- sentations, 2021. URLhttps://openreview.net/forum?id=roNqYL0_XP

    2021

  9. [17]

    Battaglia

    Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Pe- ter W. Battaglia. Learning to simulate complex physics with graph networks. InProceedings of the 37th International Conference on Machine Learning, volume 119 ofProceedings of Ma- chine Le...

    2020

  10. [18]

    Fourier neural operator with learned deformations for PDEs on general geometries.Journal of Machine Learning Research, 24(388):1–26, 2023

    Zongyi Li, Daniel Zhengyu Huang, Burigede Liu, and Anima Anandkumar. Fourier neural operator with learned deformations for PDEs on general geometries.Journal of Machine Learning Research, 24(388):1–26, 2023. URLhttps://jmlr.org/papers/v24/23-0064.html

    2023

  11. [19]

    Geometry-informed neural operator for large-scale 3d PDEs

    Zongyi Li, Nikola Kovachki, Chris Choy, Boyi Li, Jean Kossaifi, Shourya Otta, Moham- mad Amin Nabian, Maximilian Stadler, Christian Hundt, Kamyar Azizzadenesheli, and An- imashree Anandkumar. Geometry-informed neural operator for large-scale 3d PDEs. In Advances in Neural Info...

    2023 doi

  12. [20]

    Gege Wen, Zongyi Li, Kamyar Azizzadenesheli, Anima Anandkumar, and Sally M. Benson. U-FNO—an enhanced fourier neural operator-based deep-learning model for multiphase flow. Advances in Water Resources, 163:104180, 2022. doi: 10.1016/j.advwatres.2022.104180

    2022

  13. [21]

    Smith, Ayya Alieva, Qing Wang, Michael P

    Dmitrii Kochkov, Jamie A. Smith, Ayya Alieva, Qing Wang, Michael P. Brenner, and Stephan Hoyer. Machine learning–accelerated computational fluid dynamics.Proceedings of the Na- tional Academy of Sciences, 118(21):e2101784118, 2021. doi: 10.1073/pnas.2101784118

    2021 doi

  14. [22]

    Variational physics-informed neural operator (VINO) for solving partial differential equations.Computer Methods in Applied Mechanics and Engineering, 437: 117785, 2025

    Mohammad Sadegh Eshaghi, Cosmin Anitescu, Manish Thombre, Yizheng Wang, Xiaoying Zhuang, and Timon Rabczuk. Variational physics-informed neural operator (VINO) for solving partial differential equations.Computer Methods in Applied Mechanics and Engineering, 437: 117785, 2025. ...

    2025

  15. [23]

    Multi-head neural operator for modelling interfacial dynamics

    Mohammad Sadegh Eshaghi, Navid Valizadeh, Cosmin Anitescu, Yizheng Wang, Xiaoying Zhuang, and Timon Rabczuk. Multi-head neural operator for modelling interfacial dynamics. International Journal of Mechanical Sciences, 2026. doi: 10.1016/j.ijmecsci.2026.111363

    2026

  16. [24]

    NOWS: Neural operator warm starts for accelerating iterative solvers.Computer Methods in Applied Mechanics and Engineering, 458:118989, 2026

    Mohammad Sadegh Eshaghi, Cosmin Anitescu, Navid Valizadeh, Yizheng Wang, Xiaoying Zhuang, and Timon Rabczuk. NOWS: Neural operator warm starts for accelerating iterative solvers.Computer Methods in Applied Mechanics and Engineering, 458:118989, 2026. doi: 10.1016/j.cma.2026.118989. 16

    2026

  17. [2013]

    doi: 10.1016/j.jcp.2013.04.033

    2013 doi

  18. [2023]

    URLhttps://www.jmlr.org/papers/v24/21-1524.html

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.