Pith. sign in

REVIEW 3 major objections 5 minor 28 references

MeshMask: Physics-Based Simulations with Masked Graph Neural Networks

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Hiding up to 40% of mesh nodes during pre-training and reconstructing them with a small decoder makes graph-network fluid simulators far more accurate over long autoregressive rollouts, cutting all-rollout error by up to 60%.

desk verdict Masked pre-training genuinely helps GNN CFD rollouts on the datasets with the same-architecture control, but the headline 60% claim mixes in architecture and capacity, and missing controls/code keep it from being fully convincing. read the letter →

arxiv 2501.08738 v3 pith:ILWWHORL submitted 2025-01-15 cs.LG physics.flu-dyn

Paul Garnier , Vincent Lannelongue , Jonathan Viquerat , Elie Hachem This is my paper · ORCID
classification cs.LGphysics.flu-dyn
keywords maskedpre-traininggraphneuralnetworkscomputationalfluiddynamicsmesh-basedsimulationautoregressiverollouttransferlearningmultigridarchitectureaneurysmflow
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

Masked pre-training, in which up to 40% of a mesh's nodes are randomly hidden while a graph encoder-decoder learns to reconstruct them, sharply improves the long-term accuracy of learned fluid simulators. The paper reports that this scheme, combined with gated multi-layer perceptrons and a multigrid architecture, cuts all-rollout error by up to 60% compared with earlier models on seven CFD datasets, while keeping inference cost similar. The gain appears in long autoregressive rollouts rather than one-step predictions, suggesting the model learns to suppress error accumulation. The same pre-training works on several datasets at once, so one pretrained encoder can be fine-tuned for new tasks with less time and data.

What carries the argument

The load-bearing mechanism is a masked graph autoencoder: an asymmetric encoder-decoder where up to 40% of mesh nodes and their incident edges are randomly removed, replaced by a shared learnable [MASKED] token before a lightweight decoder reconstructs the hidden nodes. K-hop connections are added to preserve long-range information when the graph is sparse. Pre-training loss is computed only on masked nodes; afterwards only the encoder is fine-tuned. This is paired with gated multi-layer perceptrons and a multigrid W-cycle message-passing scheme, which together carry the error-suppression behaviour over long rollouts.

What would settle it

Train the final model twice on the 3D aneurysm dataset, once with masked pre-training and once without, keeping all other settings equal; the masked version should stay near the reported error of 8,772 and not drift toward the 13,747 baseline. Doing the same matched comparison on all seven datasets would settle whether masking or extra model size explains the results.

Watch

Extended reading notes

Core claim

The central claim is that masking part of the mesh during pre-training forces a GNN surrogate to build fluid representations that degrade more slowly over long rollouts, and that a lightweight decoder is enough to reconstruct the hidden nodes. On the same architecture, masking pre-training yields a consistent reduction in all-rollout RMSE of roughly 25%; with the full gated-MLP multigrid model the paper reports all-rollout RMSE of 29 versus 71.4 for MeshGraphNet on CYLINDER and 8,772 versus 13,747 on 3D-ANEURYSM, which it summarises as up to 60% better than previous best models at similar computational cost.

Load-bearing premise

The claim rests on masking being the cause of the gains, but the biggest reported gains come from a model that is also larger and trained with more compute than the baselines, and the same-size comparison is only shown for some of the datasets.

Editorial extensions

If this is right

  • Masked pre-training becomes a drop-in addition to existing GNN simulators: the same architecture, parameter count, and total training steps give roughly 25% better long-rollout RMSE.
  • The best configuration reported, with gated MLP and multigrid, puts all-rollout RMSE at 29 versus 71.4 on CYLINDER and 8,772 versus 13,747 on 3D-ANEURYSM against MeshGraphNet baselines.
  • One encoder pretrained on CYLINDER and BEZIER together improves both datasets by 15-23% over no pretraining and trains 33% faster.
  • A model trained on coarser meshes transfers to much finer out-of-distribution meshes (10k to 250k nodes) with only 75% worse performance than a model trained directly on fine meshes.
  • Inference with the trained GNN becomes cheaper than a classical finite-element solver after roughly 250 CYLINDER simulations or 20 3D-ANEURYSM simulations, counting training time.

Reading between the lines

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

  • An editorial inference: the headline 60% figure compares against smaller baselines, so the paper leaves open whether masking alone, rather than added capacity, drives the gains on PLATE, AIRFOIL, 2D-ANEURYSM, and 3D-ANEURYSM.
  • The masking procedure resembles a graph-specific data augmentation that may transfer to other mesh-based PDE surrogates (elasticity, electromagnetics) whenever finite-element locality dominates error growth.
  • Because one-step RMSE is flat while all-rollout RMSE improves, the mechanism is likely error-accumulation suppression; a direct measurement of rollout error growth rates would test this.
  • Multi-dataset pre-training works only when inputs and outputs align; extending to genuinely different physics would require a shared feature encoding, which the paper does not address.
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. MeshMask proposes a masked-autoencoder pre-training scheme for graph neural network (GNN) surrogates of CFD simulations. During pre-training, up to 40% of mesh nodes are randomly removed, the resulting masked graph is processed by an encoder, and a lightweight decoder reconstructs the missing nodes; the encoder is then fine-tuned for next-step prediction. The method is paired with gated MLPs and a multigrid processor, and is evaluated on seven datasets, including a new 3D intracranial aneurysm dataset with meshes of over 250,000 nodes. The paper reports up to 60% improvement in all-rollout RMSE over previous models and states that masked pre-training gives consistent gains of about 25% at equal architecture and training budget, with additional benefits from pre-training on multiple datasets.

Significance. If the central attribution holds, the paper provides a simple, low-cost pre-training recipe that could improve long-term rollout accuracy across physics-based graph simulators, and the new 3D aneurysm dataset would be a useful large-scale benchmark. The paper's strongest evidence is the same-architecture control: MGN with masking versus MGN at identical parameter count and total training steps, which yields clear all-rollout improvements on CYLINDER (71.4 to 46.5) and BEZIER (335 to 281), with smaller gains on PLATE and 2D-ANEURYSM. The multi-dataset pre-training experiments are also suggestive. However, the evidence as reported is incomplete: the masking-only control is missing on three of the seven datasets, the headline 60% figure compares a 9.2M-parameter gated-multigrid model against 2.1--3.5M-parameter baselines, and one transfer-learning statement is contradicted by the paper's own table. The manuscript does not currently release code or trained models, so the reported tables are the only reproducibility evidence.

major comments (3)
  1. [§4.2, Table 2] The claim that masked pre-training 'leads to a 25% improvement consistently' is not supported by the data reported in Table 2. The same-architecture control (MGN w/ masking versus MGN) is missing for FLAGSIMPLE, AIRFOIL, and 3D-ANEURYSM; on the four datasets where it appears, the all-rollout improvements are 34.9% (CYLINDER), 27.8% (PLATE), 16.1% (BEZIER), and 5.3% (2D-ANEURYSM), and the 1-step RMSE on PLATE degrades (0.09 versus 0.07). These missing controls are load-bearing because they are the only evidence that separates the effect of masking from the effect of the architecture and capacity changes in the 'Ours' model.
  2. [§4.2, Table 3; Abstract] The headline 'up to 60%' improvement compares the proposed 'Ours' model, which has 9.2M parameters, a gated-MLP processor, and multigrid structure, against baselines with 2.1--3.5M parameters and standard MLP processors. This does not isolate the contribution of masked pre-training. The paper needs a same-architecture ablation with 'Ours w/o masking' at the same parameter count and training budget; without it, the reported gain is equally consistent with the hypothesis that the larger, gated multigrid architecture is responsible for most of the improvement.
  3. [§4.3, Table 4] The transfer-learning claim that pre-training 'always leads to better performances in terms of all-rollout RMSE than no pretraining at all' is contradicted by the table's own numbers: fine-tuning on CYLINDER after pre-training on BEZIER is 7.5% worse than no pre-training (61.2 versus 56.9), and fine-tuning on BEZIER after pre-training on CYLINDER is 14.5% worse (315 versus 275). The statement is accurate only for the combined CYLINDER+BEZIER pre-training rows, so the claim should be restricted or the discrepancy explained.
minor comments (5)
  1. [§3.1, Table 2] The dataset is called FLAGSIMPLE in the table but SIMPLEFLAG in the Table 2 header, and the text alternates between the two names; please standardize.
  2. [§3.1] The reference 'see Table 3.1' appears to point to a table with no number in the manuscript; please fix the cross-reference.
  3. [§4.1] The sentence 'results can be seen here' contains a placeholder 'here' with no link or table reference; please replace it with a concrete pointer.
  4. [§A.1.1, §A.3] There are typos in the appendix, including 'ans was was introduced by' and 'presentend'; these should be corrected.
  5. [§A.2] The sub-mesh partitioning ablation for 3D-ANEURYSM is described only verbally ('we conducted an extensive study... no meaningful difference'); please report the actual comparison, since the 3D results rely on this partitioning strategy.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: masked pre-training gains are measured on held-out test trajectories via a same-architecture ablation, not derived from fitted parameters or loaded self-citations.

full rationale

The paper's derivation chain is self-contained and empirically grounded. Masked pre-training is a self-supervised reconstruction/next-step task on randomly masked nodes; fine-tuning is standard autoregressive next-step prediction; evaluation is on held-out test trajectories (100 train / 20 test) with no overlap. The claimed gains are supported by a same-architecture control (MGN w/ masking vs MGN) in Table 2 on four datasets, plus ablations (masking ratio, decoder depth, gated MLP) in Table 1. No equation defines the target metric in terms of a fitted parameter: the L2 loss, masking ratio, noise, and architectural choices are hyperparameters tuned on validation or one-step error, not on all-rollout test RMSE. Self-citations to Garnier et al. (2024) and Goetz et al. (2024b) supply baselines or datasets but are not used to define or force the masking result; the controlled MGN comparison does not rely on them. The incompleteness of the MGN w/ masking control on FLAGSIMPLE, AIRFOIL, and 3D-ANEURYSM, and the confound between masking and added capacity in the 'Ours' model, are threats to external validity, but they are not circularity: the reported numbers are measurements, not reductions to inputs.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The method's contribution is a training procedure, not a physical derivation, so the ledger records the tuned hyperparameters and the domain assumptions the procedure depends on. The two numbers that carry the most weight are the masking ratio (40%, chosen from a single-dataset ablation) and the per-dataset noise levels (chosen by inspecting one-step prediction error), since both directly affect the reported rollout accuracies. The gated-MLP expansion factor is listed because it triples the parameter count and confounds the masking attribution. No new physical entities are postulated; the [MASKED] token is a learned embedding whose benefit is marginal in the paper's own ablation.

free parameters (4)
  • Node masking ratio = 40%
    Default pre-training masking ratio; chosen from an ablation on CYLINDER only, where the best range is 25-40% (Figure 8). Central to the method but validated on a single dataset and transferred to all others.
  • Per-dataset input noise sigma = CYLINDER 0.02, PLATE 0.003, FLAGSIMPLE 0.001, AIRFOIL 10, BEZIER 0.02, 2D-ANEURYSM 10, 3D-ANEURYSM 10/0.5
    Noise levels 'selected by looking at average one-step error in predictions' (Appendix A.1.2); noise magnitude strongly affects autoregressive rollout accuracy, so this is tuned to the target metric.
  • Gated MLP expansion factor e = 3
    Multiplies parameters to 9.2M versus 2.8M for MGN and inflates VRAM from 7 to 16 GB (Table 3); this confounds the attribution of the headline gains to masking.
  • Sub-mesh size and partition count for 3D-ANEURYSM = 100,000 nodes; 7-15 partitions
    Needed to fit GPU memory (Appendix A.2); the paper states an ablation showed 'no meaningful difference' but provides no details of that ablation.
assumptions (4)
  • domain assumption Reconstructing randomly masked mesh nodes during pre-training forces the model to learn robust physics representations that transfer to autoregressive next-step prediction.
    Central hypothesis of Section 2.2 ('Theoretical justification'); supported only empirically by the ablations, not derived.
  • domain assumption Input noise N(0, sigma) is a valid training regularizer and its dataset-specific magnitude does not bias the comparison against baselines.
    Adopted from Sanchez-Gonzalez et al. (2020) and Pfaff et al. (2021), but the sigma values are 'selected by looking at average one-step error in predictions' (Appendix A.1.2), i.e., tuned on the target metric.
  • domain assumption All-rollout RMSE over autoregressive rollouts is the correct measure of long-term accuracy, and the reproduced baselines faithfully match their published configurations.
    Rollout error is the headline metric; baselines are numbers from prior papers or the authors' own reproduction (Table 2 caption), with no code-level verification possible.
  • standard math The multigrid W-cycle GNN (Garnier et al. 2024) is a sound base architecture and its implementation details carry over unchanged.
    Self-cited prior work by the same group, used as the Encoder and Decoder backbone; details are not re-derived in this paper.
invented entities (1)
  • [MASKED] token (shared learnable embedding for hidden nodes)
    purpose: Replaces masked node features before the Decoder reconstructs the full mesh (Section 2.2).
    Standard MAE device (Devlin et al. 2019; He et al. 2021) adopted here; the paper's own ablation (Table 1c) shows it slightly hurts all-rollout RMSE (54.3 with the token versus 52.0 without) and is 2.8 times slower, so its benefit is not independently established.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MeshMask: Physics-Based Simulations with Masked Graph Neural Networks." pith.science (2026). https://pith.science/paper/ILWWHORL

@misc{pith2026250108738,
  author       = {Pith},
  title        = {Pith review of: MeshMask: Physics-Based Simulations with Masked Graph Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ILWWHORL}},
  note         = {Machine review of arXiv:2501.08738}
}
read the original abstract

We introduce a novel masked pre-training technique for graph neural networks (GNNs) applied to computational fluid dynamics (CFD) problems. By randomly masking up to 40\% of input mesh nodes during pre-training, we force the model to learn robust representations of complex fluid dynamics. We pair this masking strategy with an asymmetric encoder-decoder architecture and gated multi-layer perceptrons to further enhance performance. The proposed method achieves state-of-the-art results on seven CFD datasets, including a new challenging dataset of 3D intracranial aneurysm simulations with over 250,000 nodes per mesh. Moreover, it significantly improves model performance and training efficiency across such diverse range of fluid simulation tasks. We demonstrate improvements of up to 60\% in long-term prediction accuracy compared to previous best models, while maintaining similar computational costs. Notably, our approach enables effective pre-training on multiple datasets simultaneously, significantly reducing the time and data required to achieve high performance on new tasks. Through extensive ablation studies, we provide insights into the optimal masking ratio, architectural choices, and training strategies.

Figures

Figures reproduced from arXiv: 2501.08738 by the authors.

Figure 1
Figure 1. The proposed Masked Graph Neural Network architecture. During pre-training, we remove [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Examples from test meshes in 3 different datasets: [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The overall AutoEncoder architecture [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Predictions on the CYLINDER with different Masking Ratio. Even with a high ratio (85%), the model is able to generalize well. 2.2 MASKING At each training step, we randomly sample a fraction of the existing nodes in a given mesh. We then proceed to completely remove th…
Figure 5
Figure 5. Figure 5: (top) Each edge is updated from its own feature, and the one from the nodes connected to it. (bottom) Each node is updated from its own features, and from the edges its connected to. (right) Information flow from a node perspec￾tive after one step of message passing. T…
Figure 6
Figure 6. Figure 6: Sample of our datasets, in the same order as in Table 3.1. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: (top-left) Overview of the 101 aneurysms. (top-right) Velocity inflow profile imposed at the inlet boundary as a parabolic flow. (bottom-left) Presentation of the vy flow in 3 aneurysms. (bottom-right) Example of a mesh. another 500k training steps, with the same strat…
Figure 8
Figure 8. Figure 8: A masking ratio between 25% and 40% leads to the best finetuning results on the [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Overview of our unstructured mesh. (top) Mesh from the Bezier dataset with around 30k nodes. (bottom) Mesh from the 3D-Aneurysm dataset with around 250k nodes. Dataset Inputs Outputs History tgnn (ms/step) tgt (ms/step) CYLINDER n, vx, vy vx, vy 0 33 820 PLATE n, x, y,…
Figure 10
Figure 10. Figure 10: Error distribution on different timesteps in a trajectory predicted from the 3D-Aneurysm [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Different sub-mesh generated at each batch update. [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]
Figure 12
Figure 12. Figure 12: (left) Ablation study of the number of message passing steps. We encounter a plateau starting from m = 10. (right) Impact of the number of neurons per layer. We encounter a plateau starting from n = 128. A.3 ABLATION STUDY We also conducted an ablation study on the Cy…
Figure 13
Figure 13. Figure 13: Uncurated random shapes from the validation cylinder and bezier shapes. [PITH_FULL_IMAGE:figures/full_fig_p017_13.png]
Figure 14
Figure 14. Figure 14: Uncurated random shapes from the validation aneurysm and bezier shapes. [PITH_FULL_IMAGE:figures/full_fig_p017_14.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 8 canonical work pages

  1. [10]

    Proposal for Numerical Benchmarking of Fluid-Structure Interaction in Cerebral Aneurysms

    URL https://arxiv.org/abs/2308.08301. Aur`ele Goetz, Pablo Jeken-Rico, Yves Chau, Jacques S ´edat, Aur´elien Larcher, and Elie Hachem. Analysis of intracranial aneurysm haemodynamics altered by wall movement. Bioengineering, 11(3), 2024a. ISSN 2306-5354. doi: 10.3390/bioengineering11030269. URL https://www. mdpi.com/2306-5354/11/3/269. 11 Published as a c...

    work page Pith review arXiv

  2. [11]

    (2018) using between 50000 and 100000 random edges

    Sub-meshes were generated using two strategies: a random neighbor sampling strategy from Hamilton et al. (2018) using between 50000 and 100000 random edges. The second strategy used the METIS Padua et al. (2011) algorithm to generate between 7 and 15 disjoint sub-meshes. We conducted an extensive study to compare results on model trained with and without ...

    work page 2018

  3. [12]

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll ´ar, and Ross Girshick

    URL https://arxiv.org/abs/1706.02216. Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll ´ar, and Ross Girshick. Masked autoencoders are scalable vision learners,

    arXiv

  4. [14]

    Ziniu Hu, Yuxiao Dong, Kuansan Wang, Kai-Wei Chang, and Yizhou Sun

    URL http://arxiv.org/abs/1905.12265. Ziniu Hu, Yuxiao Dong, Kuansan Wang, Kai-Wei Chang, and Yizhou Sun. GPT-GNN: generative pre-training of graph neural networks. CoRR, abs/2006.15437,

    arXiv 1905

  5. [17]

    Mario Lino, Chris Cantwell, Anil A

    URL https://arxiv.org/abs/2212.12794. Mario Lino, Chris Cantwell, Anil A. Bharath, and Stathi Fotiadis. Simulating continuum mechanics with multi-scale graph neural networks,

    arXiv

  6. [22]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever

    URL https://arxiv.org/abs/2312.15796. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Lan- guage models are unsupervised multitask learners

    arXiv

  7. [23]

    doi: https:// doi.org/10.1016/j.jcp.2018.10.045

    ISSN 0021-9991. doi: https:// doi.org/10.1016/j.jcp.2018.10.045. URL https://www.sciencedirect.com/science/ article/pii/S0021999118307125. Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter W. Battaglia. Learning to simulate complex physics with graph networks,

    doi:10.1016/j.jcp.2018.10.045 2018

  8. [25]

    cloze procedure

    URL https: //arxiv.org/abs/2201.02534. Wilson L Taylor. “cloze procedure”: A new tool for measuring readability. Journalism quarterly, 30 (4):415–433,

    arXiv

Show all 28 references
  1. [27]

    Jonathan Tompson, Kristofer Schlachter, Pablo Sprechmann, and Ken Perlin

    URL http://arxiv.org/abs/1810.08217. Jonathan Tompson, Kristofer Schlachter, Pablo Sprechmann, and Ken Perlin. Accelerating eulerian fluid simulation with convolutional networks. CoRR, abs/1607.03597,

    arXiv

  2. [28]

    Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li `o, and Yoshua Bengio

    URL http:// arxiv.org/abs/1607.03597. Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li `o, and Yoshua Bengio. Graph attention networks,

    arXiv

  3. [29]

    Anthony Zhou and Amir Barati Farimani

    doi: 10.1080/09540091.2022.2131737. Anthony Zhou and Amir Barati Farimani. Masked autoencoders are pde learners,

    2022

  4. [30]

    A A PPENDIX A.1 D ATASETS A.1.1 D ETAILS We give details below about the inputs and outputs used for each dataset (see Table A.1.1).CYLINDER , PLATE were generated with COMSOL multiphysics ® (2020) and were introduced by Pfaff et al. (2021). FLAGSIMPLE was generated with ArcSi...

    2020

  5. [32]

    (left) masked, (middle) predicted and (right) original

    16 Published as a conference paper at ICLR 2025 Figure 13: Uncurated random shapes from the validation cylinder and bezier shapes. (left) masked, (middle) predicted and (right) original. Figure 14: Uncurated random shapes from the validation aneurysm and bezier shapes. (left) ...

    2025

  6. [500]

    Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter W

    URL http: //dx.doi.org/10.1007/978-0-387-09766-4_500 . Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter W. Battaglia. Learning mesh- based simulation with graph networks,

    doi

  7. [2002]

    doi: 10.1002/fld.377

    ISSN 1097-0363. doi: 10.1002/fld.377. URL http://dx.doi.org/ 10.1002/fld.377. Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization,

    doi

  8. [2007]

    Thomas D

    doi: 10.1063/1.2740823. Thomas D. Economon, Francisco Palacios, Sean R. Copeland, Trent W. Lukaczyk, and Juan J. Alonso. Su2: An open-source suite for multiphysics simulation and design. AIAA Journal, 54(3):828–846, March

    doi

  9. [2010]

    doi: https://doi.org/10.1016/j.jcp.2010.07.030

    ISSN 0021-9991. doi: https://doi.org/10.1016/j.jcp.2010.07.030. URL https: //www.sciencedirect.com/science/article/pii/S0021999110004237. William L. Hamilton, Rex Ying, and Jure Leskovec. Inductive representation learning on large graphs,

    2010 doi

  10. [2011]

    doi: 10.1007/978-0-387-09766-4

    ISBN 9780387097664. doi: 10.1007/978-0-387-09766-4

    doi

  11. [2012]

    Proceedings of ACM SIGGRAPH Asia 2012, Singapore

    URL http: //graphics.berkeley.edu/papers/Narain-AAR-2012-11/ . Proceedings of ACM SIGGRAPH Asia 2012, Singapore. David Padua, Amol Ghoting, John A. Gunnels, Mark S. Squillante, Jos ´e Meseguer, James H. Cownie, Duncan Roweth, Sarita V . Adve, Hans J. Boehm, Sally A. McKee, Rob...

    2012

  12. [2016]

    doi: 10.2514/1.j053813

    ISSN 1533-385X. doi: 10.2514/1.j053813. URL http://dx.doi.org/10. 2514/1.J053813. Meire Fortunato, Tobias Pfaff, Peter Wirnsberger, Alexander Pritzel, and Peter Battaglia. MultiScale MeshGraphNets. arXiv e-prints, art. arXiv:2210.00612, October

    arXiv

  13. [2017]

    doi: 10.1145/3072959.3073643

    ISSN 1557-7368. doi: 10.1145/3072959.3073643. URL http://dx.doi.org/10.1145/3072959.3073643. Yann N. Dauphin, Angela Fan, Michael Auli, and David Grangier. Language modeling with gated convolutional networks,

  14. [2018]

    Yadi Cao, Menglei Chai, Minchen Li, and Chenfanfu Jiang

    URL http://arxiv.org/abs/1806.01261. Yadi Cao, Menglei Chai, Minchen Li, and Chenfanfu Jiang. Efficient learning of mesh-based physical simulation with bsms-gnn,

    arXiv

  15. [2019]

    Mengyu Chu and Nils Thuerey

    doi: 10.48550/ arXiv.1910.13532. Mengyu Chu and Nils Thuerey. Data-driven synthesis of smoke flows with cnn-based feature descriptors. ACM Transactions on Graphics , 36(4):1–14, July

    1910 doi arXiv

  16. [2020]

    org/abs/2006.15437

    URL https://arxiv. org/abs/2006.15437. V olker John. Higher order finite element methods and multigrid solvers in a benchmark problem for the 3d navier–stokes equations. International Journal for Numerical Methods in Fluids , 40(6): 775–798, October

    2006 arXiv

  17. [2021]

    Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, Remi Lam, and Matthew Willson

    12 Published as a conference paper at ICLR 2025 Ilan Price, Alvaro Sanchez-Gonzalez, Ferran Alet, Tom R. Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, Remi Lam, and Matthew Willson. Gencast: Diffusion-based ensemble f...

    2025

  18. [2022]

    doi: 10.48550/arXiv.2210. 00612. Paul Garnier, Jonathan Viquerat, and Elie Hachem. Multi-grid graph neural networks with self- attention for computational mechanics. arXiv preprint arXiv:2409.11899 ,

    doi arXiv

  19. [2023]

    Junfeng Chen, Jonathan Viquerat, and Elie Hachem

    URL https://arxiv.org/abs/2210.02573. Junfeng Chen, Jonathan Viquerat, and Elie Hachem. U-net architectures for fast prediction of incompressible laminar flows. arXiv e-prints, art. arXiv:1910.13532, October

    1910 arXiv

  20. [2024]

    Aur`ele Goetz, Pablo Jeken Rico, Yves Chau, Jacques S ´edat, Aur´elien Larcher, and Elie Hachem

    doi: https: //doi.org/10.48550/arXiv.2409.11899. Aur`ele Goetz, Pablo Jeken Rico, Yves Chau, Jacques S ´edat, Aur´elien Larcher, and Elie Hachem. Proposal for numerical benchmarking of fluid-structure interaction in cerebral aneurysms,

    doi arXiv

Pith tools

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