Pith. sign in

REVIEW 3 major objections 5 minor 79 references

Graph Transformers for inverse physics: reconstructing flows around arbitrary 2D airfoils

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 This paper introduces a hybrid Graph Transformer network that reconstructs the full pressure and velocity fields around arbitrary 2D airfoils from surface pressure measurements alone, reporting test R2 values of 0.997 for pressure, 0.982…

desk verdict Useful open dataset and architecture, but the blanket accuracy claim is undercut by worse pressure RMSE and the 'arbitrary airfoils' claim is untested beyond one meshing strategy. read the letter →

arxiv 2501.17081 v1 pith:7B7QMUCX submitted 2025-01-28 cs.LG cs.AIcs.CE

Gregory Duthé , Imad Abdallah , Eleni Chatzi This is my paper · ORCID
classification cs.LGcs.AIcs.CE MSC 68T0768R1076-08 PACS 47.11.-j47.85.Gj
keywords graphtransformersinversephysicsflowreconstructionairfoilaerodynamicsmessage-passingneuralnetworkslinearattentionRANSsimulationssparsesensor
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 single learned model can invert the physics of external aerodynamics: given only sparse pressure readings on the surface of a 2D airfoil, can it reconstruct the complete pressure and velocity field around the body? The authors argue yes, and they build a network, the Flow Reconstruction Graph Transformer (FRGT), that combines local message-passing on the CFD mesh with a global linear-attention layer. Tested on hundreds of unseen airfoil geometries and inflow conditions, the model recovers fields with high accuracy and runs roughly four times faster than a pure message-passing baseline. The wider claim is that this hybrid local-global design is a general pattern for mesh-based inverse physics problems, not just a trick for airfoils.

What carries the argument

The Flow Reconstruction Graph Transformer (FRGT) is an Encode-Process-Decode network whose processor stacks two operations on the mesh graph: GEN message-passing layers, which aggregate information across neighboring cells using edge attributes such as relative direction, edge length, and shared boundary length, and Galerkin linear-attention layers, which replace softmax self-attention with a normalized linear projection Q(K^T V)/n so that all pairs of nodes can exchange information at linear cost. Feature Propagation initializes the unknown fluid nodes by radiating surface pressure values outward before encoding. The decoder maps latent node features back to pressure and velocity. The study finds that the stacked variant, with L message-passing layers followed by a single Transformer layer, performs best, and that local geometric layers dominate accuracy, with x-velocity RMSE dropping from 15.9 to 3.1 as L increases from 1 to 10.

What would settle it

Take the trained FRGT-S model and apply it to the same test airfoils re-meshed with a different strategy, such as a structured O-grid or a different cell density or Delaunay point set, while keeping the physical solutions identical; if the test R2 for velocity drops substantially below the values reported for the consistent meshing, the model was keyed to discretization artifacts rather than flow physics.

Watch

Extended reading notes

Core claim

The central discovery is that stacking many local message-passing layers before a single global linear-attention Transformer layer lets the network propagate sparse surface pressure information across a 50,000-node unstructured mesh and reconstruct the full flow field. The reported test performance is R2 = 0.997 for pressure, 0.982 for x-velocity, and 0.988 for y-velocity, with inference around 200 ms on a single GPU. The architecture also degrades gracefully when sensor coverage is reduced: velocity fields remain usable at 20% of the nominal surface coverage. The ablation studies show that the local message-passing layers carry most of the reconstruction burden, while the Transformer layers supply long-range corrections that pure message-passing models miss, particularly in separated wakes and shear layers.

Load-bearing premise

The model is trained and tested on meshes produced by one consistent unstructured meshing pipeline, so the accuracy may reflect patterns in how the meshes are drawn rather than genuine physical relationships; if that is the case, the performance will not transfer to differently meshed or truly arbitrary geometries.

Editorial extensions

If this is right

  • A single trained model can reconstruct full 2D steady RANS flow fields around unseen airfoils from surface pressure taps alone, with high accuracy in both pressure and velocity.
  • The hybrid local-global design runs at about 200 ms per case on consumer hardware, roughly four times faster than the pure message-passing alternative, making near-real-time flow estimation feasible.
  • Velocity field reconstruction stays reasonable down to 20% surface sensor coverage, which matters for MEMS pressure-tap systems on wind turbine blades where full coverage is impractical.
  • More local message-passing layers (up to L = 10) consistently improve accuracy, so the geometric, edge-aware processing is the load-bearing part of the architecture, with global attention supplying long-range correction.

Reading between the lines

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

  • The success of the pressure-to-field mapping suggests the network implicitly learns a low-dimensional manifold of RANS solutions parameterized by airfoil geometry and inflow conditions; if so, the same architecture could reconstruct other boundary-observed fields, such as temperature or stress, on meshes with minimal changes.
  • The consistent-mesh caveat in the paper implies a direct stress test: re-meshing the same geometries differently and measuring the accuracy drop would reveal whether the model encodes physics or discretization patterns, and a mesh-agnostic version would likely need multiple meshing strategies in training or geometric features beyond the signed distance field.
  • The robustness to reduced coverage hints at a sensor-placement optimization: the model could be used to select which 20% of surface locations carry the most information, informing where physical pressure taps should be installed.
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 introduces a hybrid Graph Transformer architecture (FRGT) for reconstructing steady RANS pressure and velocity fields around 2D airfoils from sparse surface pressure measurements. The model combines feature propagation, edge-feature-aware message passing (GEN), and a linear Galerkin Transformer for global information exchange, following an encode-process-decode structure. The authors generate a dataset of 2,907 converged OpenFOAM simulations around UIUC airfoil geometries, convert CFD meshes into cell-centric graphs, and report held-out test metrics: pressure R2 of 0.997, x-velocity R2 of 0.982, and y-velocity R2 of 0.988, with inference times around 200 ms. Additional experiments compare stacked and interleaved layer designs, ablate layer/head configurations, and test robustness to reduced sensor coverage.

Significance. If the central claims hold, the paper provides a useful demonstration of graph-based inverse flow reconstruction with a public dataset and code, three-seed metrics on the main comparison, and a clean architectural ablation. The practical value lies in fast reconstruction from surface-only pressure measurements and in the robustness experiments for partial sensor coverage. However, the significance is tempered by two load-bearing gaps: the generalization claim to 'arbitrary 2D airfoils' and 'general inverse physics engine' is only tested within a single meshing family, and the claimed accuracy advantage over the message-passing baseline is not supported for pressure reconstruction. These issues prevent the paper from currently supporting its strongest conclusions.

major comments (3)
  1. [Section 4.1 and Section 7] All graphs in the dataset are generated with a single Gmsh O-grid unstructured triangular meshing strategy with fixed near-wall sizing and global refinement controls. The authors themselves state in Section 7 that 'the consistent meshing strategy may allow the model to learn implicit patterns in the spatial discretization rather than the purely physical relationships.' This is load-bearing because the abstract and title claim a general inverse physics engine on meshes and arbitrary 2D airfoils. The reported R2 values demonstrate interpolation within one discretization family, not mesh-agnostic physical reconstruction. Please provide a cross-mesh experiment, for example evaluating the trained model on test airfoils re-meshed with a different generator, refinement strategy, or element type, or training on multiple meshing families; alternatively, the generalization claims should be explicitly restricted to the single-mesh-family setting.
  2. [Section 6.1.1, Table 2, and Section 8] Table 2 shows the Reversible GAT baseline achieving lower pressure RMSE (75.42 ± 16.9 Pa) than both FRGT-Stacked (86.79 ± 4.5 Pa) and FRGT-Interleaved (89.44 ± 4.7 Pa), with identical pressure R2 of 0.997. The conclusion in Section 8 that the hybrid 'outperforms pure message-passing approaches in both accuracy and computational efficiency' is therefore not supported for pressure accuracy. The claim should be restricted to velocity accuracy and inference time, or the pressure discrepancy should be explained, for example by reporting the absolute pressure scale and showing that the RMSE differences are physically negligible.
  3. [Section 6.1.2 and Figure 9; Section 6.3 and Table 3] Figure 9 presents ablation results without error bars, despite the paper reporting three training runs for the main comparison in Section 6.1.1. The design recommendation that 'more, smaller heads' are preferable rests on x-velocity RMSE differences of roughly 3.0–3.5 m/s, which may be within seed-to-seed variability. Similarly, Table 3 reports partial-coverage degradation percentages with no measure of variability, making it unclear whether the observed monotone trends are significant. Please provide per-configuration mean ± standard deviation or confidence intervals for these figures, or state explicitly which numbers come from single runs.
minor comments (5)
  1. [Section 5.2] The text contains a typo: 'Transfomer' should be 'Transformer'.
  2. [Figure 4 caption] The caption says 'UIC database' but the text and references refer to the 'UIUC airfoil database'; please correct the abbreviation for consistency.
  3. [Figure 8 caption] The caption uses 'FRGT-Intertwined' while the text and Table 2 use 'FRGT-Interleaved'; please make the naming consistent.
  4. [Section 4.2, Eq. (5)] The definitions of p0 and p_Va in the sentence following Eq. (5) are unclear: p0 is said to be the total pressure estimated as max(p_Va), but p_Va is not previously defined. Please define the surface pressure variable explicitly.
  5. [Table 3] Table 3 has no column headers identifying the coverage levels; the reader must infer the ordering 80%, 60%, 40%, 20% from the text. Please add explicit headers to the table.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reconstruction results are not forced by definition, by a fitted parameter renamed as a prediction, or by a load-bearing self-citation chain; the main caveat (single meshing family) is a generalization risk, not a circular reduction.

full rationale

I walked the paper's derivation chain from the problem statement (Eqs. 1-2) through dataset generation (Section 4), normalization (Section 4.2), architecture (Section 5), and evaluation (Section 6). No step reduces to its own inputs by construction. The input surface pressure is normalized by its own mean and standard deviation, and the velocity targets are rescaled by the Bernoulli-derived inflow estimate U∮ = sqrt(2 p0 / ρ) from the maximum surface pressure. These are physically motivated preprocessing statistics derived from observable surface data, not fitted parameters that are later relabeled as predictions, and they do not by themselves determine the target field values. The high test R2 is measured against independently generated CFD ground truth on held-out airfoils and flow conditions, so the central empirical claim is not equivalent to the training objective or to the normalization scheme. The only same-author citation used in a comparative role is the Reversible GAT baseline [17], but that baseline is implemented and trained by the present authors under matched conditions (Table 2) rather than invoked as an authority; it is an experimental comparison, not a load-bearing premise. The paper's own stated limitation, 'the consistent meshing strategy may allow the model to learn implicit patterns in the spatial discretization rather than the purely physical relationships' (Section 7), is a real threat to cross-mesh generalization and to the title's 'arbitrary 2D airfoils' claim, but it is a validity/external-generalization concern, not circularity: the model is not defined in terms of the target fields, and no result is forced by a self-citation chain. The CFD pipeline is also checked against external experimental and numerical benchmarks (Figure 3), providing independent support. I therefore find no circular step and assign score 0.

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

The central contribution is a trained neural mapping, so the load-bearing fitted quantities are the model weights, optimized on the training set. No new physical entities are introduced. The listed hyperparameters are hand-chosen settings that affect performance; the remaining assumptions are domain and data-generation choices inherited from the CFD pipeline.

free parameters (4)
  • Neural network weights (encoder, processor, decoder) = ~1.4M parameters
    Trained on 2469 RANS graphs with L2 loss (Section 6.1). These are the central fitted quantities of the method; they are openly disclosed but not interpretable as physics constants.
  • Feature propagation iterations = 30
    Hand-chosen in Section 5.1; controls how far surface pressure is radiated into the field before learning.
  • Architecture hyperparameters (d, L, T, heads) = d=160, L=10, T=1
    Chosen to fit the 1.4M parameter budget; the L:T ratio strongly affects performance (Section 6.1.2).
  • Optimization settings = lr=5e-4, weight decay=1e-4, 500 epochs
    Reported in Section 6.1; standard choices not derived from first principles.
assumptions (4)
  • domain assumption Steady 2D RANS with k-omega SST closure yields sufficiently accurate ground truth fields for the reconstruction task.
    Dataset labels come from OpenFOAM solving Eq. 3-4; the CFD pipeline is only validated against NACA 0012 pressure coefficients (Figure 3), not against velocity or separated-flow measurements.
  • domain assumption The maximum surface pressure approximates the stagnation total pressure p0 used for inflow velocity normalization.
    Section 4.2, Eq. 5: U-hat_infinity = sqrt(2 p0 / rho), with p0 = max surface pressure. This approximation may break down for separated or high-loss flows.
  • domain assumption Mesh-cell adjacency in the generated unstructured triangular O-grids captures the spatial locality needed for the inverse mapping.
    The graph is built from Gmsh cell connectivity (Section 4.2); the model has no alternative spatial representation, so any mesh artifact can be learned as a proxy.
  • domain assumption A single consistent meshing strategy does not introduce a discretization-specific bias that the model exploits.
    Section 7 explicitly warns that the consistent meshing strategy may allow the model to learn implicit patterns in the spatial discretization rather than purely physical relationships.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Graph Transformers for inverse physics: reconstructing flows around arbitrary 2D airfoils." pith.science (2026). https://pith.science/paper/7B7QMUCX

@misc{pith2026250117081,
  author       = {Pith},
  title        = {Pith review of: Graph Transformers for inverse physics: reconstructing flows around arbitrary 2D airfoils},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7B7QMUCX}},
  note         = {Machine review of arXiv:2501.17081}
}
read the original abstract

We introduce a Graph Transformer framework that serves as a general inverse physics engine on meshes, demonstrated through the challenging task of reconstructing aerodynamic flow fields from sparse surface measurements. While deep learning has shown promising results in forward physics simulation, inverse problems remain particularly challenging due to their ill-posed nature and the difficulty of propagating information from limited boundary observations. Our approach addresses these challenges by combining the geometric expressiveness of message-passing neural networks with the global reasoning of Transformers, enabling efficient learning of inverse mappings from boundary conditions to complete states. We evaluate this framework on a comprehensive dataset of steady-state RANS simulations around diverse airfoil geometries, where the task is to reconstruct full pressure and velocity fields from surface pressure measurements alone. The architecture achieves high reconstruction accuracy while maintaining fast inference times. We conduct experiments and provide insights into the relative importance of local geometric processing and global attention mechanisms in mesh-based inverse problems. We also find that the framework is robust to reduced sensor coverage. These results suggest that Graph Transformers can serve as effective inverse physics engines across a broader range of applications where complete system states must be reconstructed from limited boundary observations.

Figures

Figures reproduced from arXiv: 2501.17081 by the authors.

Figure 1
Figure 1. Depiction of an inverse physics problem on a graph. Given a set of measurements on the sensed nodes in our [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Distribution of airfoil geometries visualized through their first two principal components. The training set [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Validation of the CFD pipeline against experimental measurements [ [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Illustration of the simulation pipeline. Airfoil shapes are selected at random from subsets of the UIC database, [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Geometrical features used as inputs to the model. Overall the signed distance field is plotted, with a zoom on [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Architecture of the FRGT-S model. The pipeline consists of: (1) initial feature propagation on the graph, [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Illustration of the two proposed approaches for combining local and global layers. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Comparing the reconstructed pressure and velocity magnitude of the FRGT-Stacked, FRGT-Intertwined and [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Study comparing model performance (x- and y-velocity RMSE) across message-passing vs. Transformer [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Comparison of flow field predictions between the CFD ground truth and the FRGT-S reconstructions for [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Evaluation of the FRGT-S model’s extrapolative performance comparing ground truth CFD solutions with [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: Comparison of flow field predictions (velocity magnitude and pressure) for two representative airfoils using [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

79 extracted references · 57 canonical work pages

  1. [1]

    Abdallah, I., Duthé, G., Barber, S., and Chatzi, E. (2022). Identifying evolving leading edge erosion by tracking clusters of lift coefficients. In Journal of Physics: Conference Series, volume 2265, page 032089. IOP Publishing

    work page 2022

  2. [2]

    and Yahav, E

    Alon, U. and Yahav, E. (2020). On the bottleneck of graph neural networks and its practical implications. arXiv preprint arXiv:2006.05205

    arXiv 2020

  3. [3]

    Apte, J. S. and Manchanda, C. (2024). High-resolution urban air pollution mapping. Science, 385(6707):380–385

    work page 2024

  4. [4]

    Barber, S., Deparday, J., Marykovskiy, Y ., Chatzi, E., Abdallah, I., Duthé, G., Magno, M., Polonelli, T., Fischer, R., and Müller, H. (2022). Development of a wireless, non-intrusive, mems-based pressure and acoustic measurement system for large-scale operating wind turbine blades. Wind Energy Science, 7(4):1383–1398

    work page 2022

  5. [5]

    Berkooz, G., Holmes, P., and Lumley, J. L. (1993). The proper orthogonal decomposition in the analysis of turbulent flows. Annual review of fluid mechanics, 25(1):539–575

    work page 1993

  6. [6]

    M., Bruna, J., Cohen, T., and Veliˇckovi´c, P

    Bronstein, M. M., Bruna, J., Cohen, T., and Veliˇckovi´c, P. (2021). Geometric deep learning: Grids, groups, graphs, geodesics, and gauges. arXiv preprint arXiv:2104.13478. 17

    arXiv 2021

  7. [7]

    Brown, T. B. (2020). Language models are few-shot learners. arXiv preprint arXiv:2005.14165

    arXiv 2020

  8. [8]

    Cao, S. (2021). Choose a transformer: Fourier or galerkin. Advances in neural information processing systems, 34:24924–24940

    work page 2021

Show all 79 references
  1. [9]

    Chen, J., Hachem, E., and Viquerat, J. (2021). Graph neural networks for laminar flow prediction around random 2d shapes. arXiv preprint arXiv:2107.11529

    2021 arXiv

  2. [10]

    Child, R., Gray, S., Radford, A., and Sutskever, I. (2019). Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509

    2019 arXiv

  3. [11]

    Chinesta, F., Keunings, R., and Leygue, A. (2013). The proper generalized decomposition for advanced numerical simulations: a primer. Springer Science & Business Media

    2013

  4. [12]

    A., Pagone, V

    Danciu, B. A., Pagone, V . A., Böhm, B., Schmidt, M., and Frouzakis, C. E. (2024). Flow reconstruction in time-varying geometries using graph neural networks. arXiv e-prints, pages arXiv–2411

    2024

  5. [13]

    de N Santos, F., Duthé, G., Abdallah, I., Réthoré, P.-É., Weijtjens, W., Chatzi, E., and Devriendt, C. (2024). Multivariate prediction on wake-affected wind turbines using graph neural networks. In Journal of Physics: Conference Series, volume 2647, page 112006. IOP Publishing

    2024

  6. [14]

    Dehnhardt, G., Mauck, B., Hanke, W., and Bleckmann, H. (2001). Hydrodynamic trail-following in harbor seals (phoca vitulina). Science, 293(5527):102–104

    2001

  7. [15]

    Dosovitskiy, A. (2020). An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929

    2020 arXiv

  8. [16]

    Duthé, G., Abdallah, I., Barber, S., and Chatzi, E. (2021). Modeling and monitoring erosion of the leading edge of wind turbine blades. Energies, 14(21):7262

    2021

  9. [17]

    Duthé, G., Abdallah, I., Barber, S., and Chatzi, E. (2023). Graph neural networks for aerodynamic flow reconstruction from sparse sensing. arXiv preprint arXiv:2301.03228

    2023 arXiv

  10. [18]

    Duthé, G., de N Santos, F., Abdallah, I., Weijtjens, W., Devriendt, C., and Chatzi, E. (2024). Flexible multi-fidelity framework for load estimation of wind farms through graph neural networks and transfer learning. Data-Centric Engineering, 5:e29

    2024

  11. [19]

    Dwivedi, V . P. and Bresson, X. (2020). A generalization of transformer networks to graphs. arXiv preprint arXiv:2012.09699

    2020 arXiv

  12. [20]

    B., Mathelin, L., Yao, Z., Brunton, S

    Erichson, N. B., Mathelin, L., Yao, Z., Brunton, S. L., Mahoney, M. W., and Kutz, J. N. (2020). Shallow neural networks for fluid flow reconstruction with limited sensors.Proceedings of the Royal Society A, 476(2238):20200097

    2020

  13. [21]

    and Remacle, J.-F

    Geuzaine, C. and Remacle, J.-F. (2020). A three-dimensional finite element mesh generator with built-in pre-and post-processing facilities. International Journal for Numerical Methods in Engineering, (11):79

    2020

  14. [22]

    S., Riley, P

    Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., and Dahl, G. E. (2017). Neural message passing for quantum chemistry. In International conference on machine learning, pages 1263–1272. PMLR

    2017

  15. [23]

    and O’reilly, C

    Gregory, N. and O’reilly, C. (1970). Low-speed aerodynamic characteristics of naca 0012 aerofoil section, including the effects of upper-surface roughness simulating hoar frost

    1970

  16. [24]

    Hadamard, J. (1923). Lectures on Cauchy’s Problem in Linear Partial Differential Equations. Mrs. Hepsa Ely Silliman memorial lectures. Yale University Press

    1923

  17. [25]

    Haywood-Alexander, M., Liu, W., Bacsa, K., Lai, Z., and Chatzi, E. (2024). Discussing the spectrum of physics-enhanced machine learning: a survey on structural mechanics applications. Data-Centric Engineering, 5:e30

    2024

  18. [26]

    He, X., Wang, Y ., and Li, J. (2022). Flow completion network: Inferring the fluid dynamics from incomplete flow information using graph neural networks. Physics of Fluids, 34(8)

    2022

  19. [27]

    Hernández, Q., Badías, A., Chinesta, F., and Cueto, E. (2022). Thermodynamics-informed graph neural networks. IEEE Transactions on Artificial Intelligence, 5(3):967–976

    2022

  20. [28]

    and Haller, G

    Jain, S. and Haller, G. (2022). How to compute invariant manifolds and their reduced dynamics in high-dimensional finite element models. Nonlinear dynamics, 107(2):1417–1450

    2022

  21. [29]

    Jasak, H., Jemcov, A., Tukovic, Z., et al. (2007). Openfoam: A c++ library for complex physics simulations. In International workshop on coupled methods in numerical dynamics , volume 1000, pages 1–20. IUC Dubrovnik Croatia

    2007

  22. [30]

    Jiang, Y ., Fu, J., Zhang, D., and Zhao, Y . (2016). Investigation on the lateral line systems of two cavefish: Sinocy- clocheilus macrophthalmus and s. microphthalmus (cypriniformes: Cyprinidae). Journal of Bionic Engineering, 13(1):108–114. 18

    2016

  23. [31]

    Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., Žídek, A., Potapenko, A., et al. (2021). Highly accurate protein structure prediction with alphafold. nature, 596(7873):583–589

    2021

  24. [32]

    Katharopoulos, A., Vyas, A., Pappas, N., and Fleuret, F. (2020). Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156–5165. PMLR

    2020

  25. [33]

    Kovachki, N., Li, Z., Liu, B., Azizzadenesheli, K., Bhattacharya, K., Stuart, A., and Anandkumar, A. (2023). Neural operator: Learning maps between function spaces with applications to pdes. Journal of Machine Learning Research, 24(89):1–97

    2023

  26. [34]

    Kreuzer, D., Beaini, D., Hamilton, W., Létourneau, V ., and Tossou, P. (2021). Rethinking graph transformers with spectral attention. Advances in Neural Information Processing Systems, 34:21618–21629

    2021

  27. [35]

    Krist, S. L. (1998). CFL3D user’s manual (version 5.0). National Aeronautics and Space Administration, Langley Research Center

    1998

  28. [36]

    LeCun, Y ., Bottou, L., Bengio, Y ., and Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324

    1998

  29. [37]

    Li, G., Xiong, C., Thabet, A., and Ghanem, B. (2020a). Deepergcn: All you need to train deeper gcns. arXiv preprint arXiv:2006.07739

    2020 arXiv

  30. [38]

    Li, Q., Li, X., Chen, X., and Yao, W. (2024). A novel graph modeling method for gnn-based hypersonic aircraft flow field reconstruction. Engineering Applications of Computational Fluid Mechanics, 18(1):2394177

    2024

  31. [39]

    Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., and Anandkumar, A. (2020b). Fourier neural operator for parametric partial differential equations. arXiv preprint arXiv:2010.08895

    2020 arXiv

  32. [40]

    Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., and Anandkumar, A. (2020c). Neural operator: Graph kernel network for partial differential equations. arXiv preprint arXiv:2003.03485

    2020 arXiv

  33. [41]

    Ling, J., Kurzawski, A., and Templeton, J. (2016). Reynolds averaged turbulence modelling using deep neural networks with embedded invariance. Journal of Fluid Mechanics, 807:155–166

    2016

  34. [42]

    Loshchilov, I. (2017). Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101

    2017 arXiv

  35. [43]

    and Hutter, F

    Loshchilov, I. and Hutter, F. (2016). Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983

    2016 arXiv

  36. [44]

    Lu, L., Jin, P., and Karniadakis, G. E. (2019). Deeponet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators. arXiv preprint arXiv:1910.03193

    2019 arXiv

  37. [45]

    Marykovskiy, Y ., Deparday, J., Abdallah, I., Duthé, G., Barber, S., and Chatzi, E. (2023). Hybrid model for inflow conditions inference on airfoils under uncertainty. AIAA Journal, 61(11):4913–4925

    2023

  38. [46]

    and Esch, T

    Menter, F. and Esch, T. (2001). Elements of industrial heat transfer predictions. In 16th Brazilian Congress of Mechanical Engineering (COBEM), volume 109, page 650

    2001

  39. [47]

    Oettershagen, P., Müller, B., Achermann, F., and Siegwart, R. (2019). Real-time 3d wind field prediction onboard uavs for safe flight in complex terrain. In 2019 IEEE Aerospace Conference, pages 1–10. IEEE

    2019

  40. [48]

    O’Sullivan, F. (1986). A statistical perspective on ill-posed inverse problems. Statistical science, pages 502–518

    1986

  41. [49]

    Pfaff, T., Fortunato, M., Sanchez-Gonzalez, A., and Battaglia, P. W. (2020). Learning mesh-based simulation with graph networks. arXiv preprint arXiv:2010.03409

    2020 arXiv

  42. [50]

    Podvin, B., Fraigneau, Y ., Lusseyran, F., and Gougat, P. (2005). A reconstruction method for the flow past an open cavity. Journal of Fluids Engineering, 128(3):531–540

    2005

  43. [51]

    Polonelli, T., Deparday, J., Abdallah, I., Barber, S., Chatzi, E., and Magno, M. (2023). Instrumentation and measurement systems: Aerosense: A wireless, non-intrusive, flexible, and mems-based aerodynamic and acoustic measurement system for operating wind turbines. IEEE Instru...

    2023

  44. [52]

    and Rohde-Brandenburger, K

    Raab, C. and Rohde-Brandenburger, K. (2021). Dynamic flight load measurements with mems pressure sensors. CEAS aeronautical journal, 12(4):737–753

    2021

  45. [53]

    P., Luu, A

    Rampášek, L., Galkin, M., Dwivedi, V . P., Luu, A. T., Wolf, G., and Beaini, D. (2022). Recipe for a general, powerful, scalable graph transformer. Advances in Neural Information Processing Systems, 35:14501–14515

    2022

  46. [54]

    I., Chamberlain, B

    Rossi, E., Kenlay, H., Gorinova, M. I., Chamberlain, B. P., Dong, X., and Bronstein, M. (2021). On the unreasonable effectiveness of feature propagation in learning on graphs with missing node features. arXiv preprint arXiv:2111.12128. 19

    2021 arXiv

  47. [55]

    K., Bronstein, M

    Rusch, T. K., Bronstein, M. M., and Mishra, S. (2023). A survey on oversmoothing in graph neural networks. arXiv preprint arXiv:2303.10993

    2023 arXiv

  48. [56]

    Sanchez-Gonzalez, A., Godwin, J., Pfaff, T., Ying, R., Leskovec, J., and Battaglia, P. (2020). Learning to simulate complex physics with graph networks. In International Conference on Machine Learning, pages 8459–8468. PMLR

    2020

  49. [57]

    T., Merel, J., Riedmiller, M., Hadsell, R., and Battaglia, P

    Sanchez-Gonzalez, A., Heess, N., Springenberg, J. T., Merel, J., Riedmiller, M., Hadsell, R., and Battaglia, P. (2018). Graph networks as learnable physics engines for inference and control. In International Conference on Machine Learning, pages 4470–4479. PMLR

    2018

  50. [58]

    C., Hagenbuchner, M., and Monfardini, G

    Scarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M., and Monfardini, G. (2008). The graph neural network model. IEEE transactions on neural networks, 20(1):61–80

    2008

  51. [59]

    Scharer, N., Polonelli, T., Deparday, J., and Magno, M. (2024). Towards a non-invasive monitoring system for wind turbine blades. In 2024 IEEE International Instrumentation and Measurement Technology Conference (I2MTC), pages 1–6. IEEE

    2024

  52. [60]

    Schmid, P. J. (2010). Dynamic mode decomposition of numerical and experimental data. Journal of fluid mechanics, 656:5–28

    2010

  53. [61]

    Selig, M. S. (1996). Uiuc airfoil data site

    1996

  54. [62]

    J., and Sinop, A

    Shirzad, H., Velingker, A., Venkatachalam, B., Sutherland, D. J., and Sinop, A. K. (2023). Exphormer: Sparse transformers for graphs. In International Conference on Machine Learning, pages 31613–31632. PMLR

    2023

  55. [63]

    Tanaka, A., Tomiya, A., and Hashimoto, K. (2021). Inverse Problems in Physics, pages 129–138. Springer Singapore, Singapore

    2021

  56. [64]

    Thuerey, N., Weißenow, K., Prantl, L., and Hu, X. (2020). Deep learning methods for reynolds-averaged navier–stokes simulations of airfoil flows. AIAA Journal, 58(1):25–36

    2020

  57. [65]

    Tobin, J., Fong, R., Ray, A., Schneider, J., Zaremba, W., and Abbeel, P. (2017). Domain randomization for transferring deep neural networks from simulation to the real world. In 2017 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 23–30. IEEE

    2017

  58. [66]

    P., Dong, X., and Bronstein, M

    Topping, J., Di Giovanni, F., Chamberlain, B. P., Dong, X., and Bronstein, M. M. (2021). Understanding over-squashing and bottlenecks on graphs via curvature. arXiv preprint arXiv:2111.14522

    2021 arXiv

  59. [67]

    Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.-A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al. (2023). Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971

    2023 arXiv

  60. [68]

    W., and Nikoueeyan, P

    Vandervoort, B., Naughton, J. W., and Nikoueeyan, P. (2023). The development of pressure gloves for unsteady pressure measurements on varying airfoil geometry. In AIAA AVIATION 2023 Forum, page 3678

    2023

  61. [69]

    Vaswani, A. (2017). Attention is all you need. Advances in Neural Information Processing Systems

    2017

  62. [70]

    Veliˇckovi´c, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., and Bengio, Y . (2017). Graph attention networks. arXiv preprint arXiv:1710.10903

    2017 arXiv

  63. [71]

    Wandel, N., Weinmann, M., and Klein, R. (2020). Learning incompressible fluid dynamics from scratch–towards fast, differentiable fluid models that generalize. arXiv preprint arXiv:2006.08762

    2020 arXiv

  64. [72]

    Wandel, N., Weinmann, M., Neidlin, M., and Klein, R. (2022). Spline-pinn: Approaching pdes without data using fast, physics-informed hermite-spline cnns. In Proceedings of the AAAI conference on artificial intelligence, volume 36, pages 8529–8538

    2022

  65. [73]

    and Kipf, T

    Welling, M. and Kipf, T. N. (2016). Semi-supervised classification with graph convolutional networks. In J. International Conference on Learning Representations (ICLR 2017)

    2016

  66. [74]

    Xu, K., Hu, W., Leskovec, J., and Jegelka, S. (2018). How powerful are graph neural networks? arXiv preprint arXiv:1810.00826

    2018 arXiv

  67. [75]

    Yeh, P.-K., Chen, H.-W., and Chen, M.-S. (2023). Random walk conformer: learning graph representation from long and short range. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 37, pages 10936–10944

    2023

  68. [76]

    M., and Guo, J

    Yen, W.-K., Sierra, D. M., and Guo, J. (2018). Controlling a robotic fish to swim along a wall using hydrodynamic pressure feedback. IEEE Journal of Oceanic Engineering, 43(2):369–380

    2018

  69. [77]

    Ying, C., Cai, T., Luo, S., Zheng, S., Ke, G., He, D., Shen, Y ., and Liu, T.-Y . (2021). Do transformers really perform badly for graph representation? Advances in neural information processing systems, 34:28877–28888

    2021

  70. [78]

    Zhao, Z., Yang, Q., Li, R., Yang, J., Liu, Q., Zhu, B., Weng, C., Liu, W., Hu, P., Ma, L., et al. (2024). A comprehensive review on the evolution of bio-inspired sensors from aquatic creatures.Cell Reports Physical Science, 5(7). 20

    2024

  71. [79]

    H., Shen, L., Ma, Z., Wang, G., and Xu, H

    Zhou, H., Hu, T., Low, K. H., Shen, L., Ma, Z., Wang, G., and Xu, H. (2015). Bio-inspired flow sensing and prediction for fish-like undulating locomotion: A cfd-aided approach. Journal of Bionic Engineering, 12(3):406–417. 21

    2015

Pith tools

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