Pith. sign in

REVIEW 3 major objections 7 minor 2 cited by

Geometric Operator Learning with Optimal Transport

T0 review · 3 major / 7 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read The paper claims that geometry embedding in neural operators for PDEs can be posed as optimal transport, with each surface mesh mapped instance-by-instance onto a uniform latent grid, and that this makes surface-bound flow prediction both…

desk verdict OT-based geometry embedding for neural operators is a fresh idea, but the implementation snaps to nearest vertices after transport, so the central OT claim is not actually tested. read the letter →

arxiv 2507.20065 v1 pith:46A4TZE4 submitted 2025-07-26 cs.LG

Xinyi Li , Zongyi Li , Nikola Kovachki , Anima Anandkumar This is my paper · ORCID
classification cs.LG
keywords optimaltransportneuraloperatorsgeometryembeddingRANSequationssurfacesub-manifoldSinkhornalgorithminstance-dependentdeformationcomputationalfluiddynamics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to show that the geometry-encoding step in neural operators for PDEs can be redefined as an optimal transport problem: instead of deforming all shapes with one shared map or interpolating through a fixed graph, each input mesh is treated as a density of mass and transported to a uniform grid in a latent reference space. The authors build OTNO, which computes this instance-dependent transport either through a Kantorovich plan solved with Sinkhorn or through a Monge map solved with projection-pursuit Monge maps, and then applies a Fourier neural operator on the resulting regular 2D grid. For surface-bound solutions of the RANS equations on cars, this means the 3D surface mesh is embedded into a 2D parameterized latent space, so the heavy computation happens in 2D. On ShapeNet-Car and DrivAerNet-Car, OTNO matches or slightly beats existing machine-learning baselines in accuracy while using roughly 2x-8x less time and memory, and it shows a large accuracy gain on FlowBench, whose shapes are more diverse. The point of the paper is that optimal transport gives a more flexible, per-instance geometry embedding than the shared-deformation and graph-interpolation alternatives.

What carries the argument

The load-bearing object is the optimal transport problem between the physical mesh density $d\mu = f(x)\,dx$ and a uniform reference density $\lambda$, with squared Euclidean distance as the cost. Two solvers carry the implementation: the Sinkhorn algorithm produces an entropy-regularized Kantorovich coupling matrix $P$, whose marginal map $X' = PX$ defines the transported mesh, and projection-pursuit Monge maps produce a bijective transport map $T$ by iterating one-dimensional OT along informative projection directions. Around this, the encoder and decoder snap transported points to nearest original vertices, the latent representation adds the cross product of latent and physical surface normals, and a Fourier neural operator acts on the resulting regular 2D grid. This combination is what makes computations 2D for surface problems and instance-dependent for each geometry.

What would settle it

Re-mesh each car in a test set so that all shapes have roughly uniform vertex density while their surfaces are unchanged, then train OTNO on the re-meshed data. If accuracy collapses, the reported gains depend on density inhomogeneities; if accuracy holds, the density aspect of the transport is not actually load-bearing.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that geometry embedding for operator learning generalizes naturally from discretized meshes to mesh density functions, and that the right embedding is the optimal transport between that density and a uniform density on a latent reference manifold. This subsumes earlier methods as special cases: the Monge transport map generalizes the shared deformation map of Geo-FNO, and the Kantorovich transport plan generalizes the graph coupling of GINO. The resulting OTNO solves each instance's transport problem independently, projects the transported points back onto the original mesh by nearest-neighbour matching, and learns the latent solution operator with FNO on the regular grid. When the PDE output lives on a surface, the framework restricts the operator to the boundary sub-manifold and computes in one dimension fewer, which the paper reports as both more accurate and substantially cheaper in time and memory on RANS car datasets, and markedly more accurate on the diverse-geometry FlowBench dataset.

Load-bearing premise

The method assumes that a car's shape is adequately encoded by the spatial density of its mesh vertices plus surface normals; if density and normals miss geometry that controls the flow, the transport embedding cannot recover it.

Editorial extensions

If this is right

  • Surface-bound RANS predictions on car geometries can be performed in a 2D latent representation, cutting total time by about 2x-8x and GPU memory by about 2x-8x compared with Geo-FNO and GINO while matching or slightly beating their accuracy.
  • Because the OT embedding is solved per shape, datasets with highly variable geometries gain the most; on FlowBench, OTNO reduces relative L2 error substantially below FNO and GINO.
  • The framework unifies map-type and plan-type geometry encoders, so improvements in OT solvers will transfer directly to the geometric embedding used by the neural operator.
  • The reported convergence rate with respect to mesh resolution is faster for OTNO (1.85) than for GINO (1.37) or Geo-FNO (1.32), suggesting that the instance-dependent embedding makes better use of added resolution.

Reading between the lines

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

  • An implicit consequence is that OTNO's representation of a shape is only as rich as its vertex density plus surface normals; geometry not visible in either, such as intended crease placement or thin features, is invisible to the encoder.
  • The nearest-neighbour snapping after transport discards the soft coupling weights, so one could test whether keeping a weighted average of neighbouring vertices, rather than one nearest vertex, recovers information lost by the Sinkhorn approximation.
  • If the vertex density is re-meshed to be more uniform while geometry is held fixed, the OT map changes even though the surface does not; measuring how much OTNO's error changes under such re-meshing would isolate how much of the improvement is genuinely due to density-driven transport rather than to the FNO backbone.
  • The same sub-manifold trick should transfer to other surface-bound solution operators, such as wall shear stress or heat flux on airfoils and other designs, whenever the output of interest lives on the boundary.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The paper proposes OTNO, a neural operator that embeds surface geometries into a 2D regular latent grid by solving an optimal transport problem between a mesh density and a uniform density on a canonical latent shape. Two variants are presented: OTNO(Plan), based on Sinkhorn's entropy-regularized Kantorovich formulation, and OTNO(Map), based on the Projection Pursuit Monge Map. After transporting latent points onto the physical mesh, Algorithm 1 snaps the transported points to the nearest original vertex to form encoder/decoder index maps and then applies a 2D FNO on the latent grid, using the original coordinates and cross-product normals as features. Experiments on ShapeNet-Car and DrivAerNet-Car report lower error and lower cost than Geo-FNO and GINO, and on FlowBench the method shows large accuracy gains. The paper also claims that Geo-FNO and GINO are special cases of the proposed OT framework.

Significance. If the central claim were established, instance-dependent OT-based embedding would be a meaningful alternative to the shared deformation of Geo-FNO and the graph interpolation of GINO, and the sub-manifold dimension reduction from 3D to 2D is practically valuable for automotive aerodynamic surrogate modeling. The paper is clearly specified, releases code, and includes extensive ablations on two large 3D datasets and one diverse 2D dataset. The accuracy gains on FlowBench are particularly striking. However, the central scientific claim that optimal transport explains the empirical improvement is not currently isolated from the resampling mechanism, and several key design choices are selected on test error without error bars. The significance of the contribution is therefore real but conditional on additional control experiments and tightened evaluation.

major comments (3)
  1. [4.1 (Algorithm 1); 6.1.1 (Table 5)] The implemented encoder does not actually feed the transport map or plan to the operator. In Algorithm 1 (lines 4-10), the transport map/plan is used only to define index maps E and D; the FNO input is T_j = (Ξ_j, M_j, H_j × N_j(E)) with M_j = X_j(E), i.e., the nearest original mesh vertices, and the decoder is a nearest-neighbor lookup on the latent grid. Thus the OT solution enters only as an instance-dependent resampling of the original vertices onto a 2D grid, and the coupling P is discarded. Since the paper does not compare against any non-OT density-adaptive resampling baseline (e.g., farthest-point sampling, PCA-aligned parameterization, or a learnable grid-to-vertex assignment), the central claim that 'formulating geometry embedding as an OT problem' explains the reported gains is untested. The accuracy gains over Geo-FNO and GINO could plausibly come from the instance-dependent submanifold parameterization plus the 2D FNO rather than from optimal transport as such. Please add such control baselines and, if they match OTNO, revise the contribution claim accordingly.
  2. [3.2 (Eq. 21)] The claimed generalization of Geo-FNO as a special case of the OT framework relies on a cost function c(x,T^{-1}(x)) := G(T^{-1}(x)) - u_sub(x) dμ(x) that contains the unknown solution operator G and the target u_sub. This cost is not a function of the source and target points alone, as required by the Monge/Kantorovich problems in Section 2.4, and it is chosen post hoc to force Geo-FNO into the framework. The argument is therefore circular and does not establish a mathematical unification. I recommend either removing the claim that Geo-FNO is a special case, or rephrasing it as a high-level analogy that does not rely on an invented, solution-dependent cost.
  3. [Sections 5-6 (Tables 1-10, Figs. 8-9)] Key design decisions -- latent mesh shape (Table 8), expansion factor (Fig. 8), voxel downsampling size (Fig. 9), normal feature type (Table 7), and PPMM iteration count (Table 9) -- are selected based on test-set errors, and all tables report single runs without error bars or standard deviations. Because the reported margins over the strongest baseline are small (ShapeNet 6.70% vs 7.21%; DrivAerNet MSE 3.28e-5 vs 3.33e-5), the 'slightly enhancing accuracy' claim can be fully explained by selection bias. Please report performance on a held-out validation set for model selection and provide means and standard deviations over multiple random seeds for the final comparisons.
minor comments (7)
  1. [Section 2.2] The paragraph ends with the incomplete sentence 'In this work, we propose to'; please complete or remove it.
  2. [Section 1 and elsewhere] There are several typos, including 'densitys' (Section 1), 'sparity' (Section 2.4.2), 'closet point' (Section 4.1), and 'with with' (Section 1); a careful proofread is needed.
  3. [Section 3 (Eq. 12)] Since Ω denotes the (d−1)-dimensional submanifold ∂Ω_f in this section, the notation dμ = f(x)dx should be stated with the appropriate surface measure dS rather than the Lebesgue measure on R^d.
  4. [Figure 2 caption] The claim that OT 'preserves the global measure, which is essential for computing integral operators' is not reflected in Algorithm 1, which uses the transport only for index selection; please reconcile the caption with the implementation.
  5. [Section 7.1 / Table 11] The elasticity example in Table 11 lacks any description of the PDE, dataset, or baseline setup; please add the experimental details or remove the table.
  6. [Section 1] The claim of being 'approximately 7,000 times faster than traditional approaches' is not backed by a direct comparison in the paper; please provide the underlying runtime figures or a reference.
  7. [Section 4.2.3] No ablation is reported for the Sinkhorn regularization parameter β; since the authors state it controls the accuracy-cost trade-off, a brief sensitivity check would strengthen Section 6.

Circularity Check

1 steps flagged · score 1.0 of 10

No load-bearing circularity; one post-hoc definitional framing in Section 3.2, where Geo-FNO is made a special case of OT by a cost that contains the target solution.

  1. self definitional [Section 3.2, Eq. (21), 'Method Generalization' (page 13)]
    "∫_Ω c(x,T^{-1}(x)) dμ(x) := ∫_Ω G(T^{-1}(x)) − u_sub(x) dμ(x) ... However, this is a non-standard, generalized cost function, since it depends not on (x,T^{-1}(x)), but the global solution operator G and domain Ω."

    The generalized 'cost' in Eq. (21) is defined as the discrepancy between the transported feature and the target solution u_sub. Because the cost itself contains the object the method is supposed to predict, any supervised deformation trained to reduce solution error can be relabeled as an OT map for this contrived cost. Thus the claim that Geo-FNO is a special case of the OT framework is true by definition rather than by the mathematical structure of optimal transport. This is a post-hoc reframing step, not a derivation that supports OTNO; it does not enter the implemented OT encoders/decoders, which use squared Euclidean cost and Sinkhorn/PPMM, and it does not affect the external benchmark comparisons.

full rationale

I find no load-bearing circularity in OTNO. The central encoder is computed per instance from the spatial distribution of mesh vertices using squared Euclidean cost and either Sinkhorn (Kantorovich) or PPMM (Monge), with no use of target pressure or drag values in constructing the transport map/plan. The learned operator is then trained and evaluated on external CFD datasets (ShapeNet-Car, DrivAerNet-Car, FlowBench) against standard baselines, so the experimental claims are not forced by construction. The one definitionally circular passage is Section 3.2, where the authors define a generalized OT cost containing the target solution u_sub to argue that Geo-FNO is a special case of their framework. This makes that unification claim true by construction, but it is a non-experimental framing device and does not feed the OTNO algorithm or the reported accuracy/efficiency numbers. The nearest-neighbor snapping in Algorithm 1 discards the coupling after forming an index mapping; this is a gap between the theoretical OT story and the implemented encoder, not a circular reduction, because the OT computation still determines the indices. Self-citations to Geo-FNO and GINO are used as baselines and prior work rather than as externally imported uniqueness theorems, and no fitted parameter is renamed as a prediction. Overall, the derivation chain for the paper's actual contributions is self-contained against external data, so the appropriate score is a low non-zero value reflecting only the minor Section 3.2 definitional framing.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central method rests on standard optimal transport mathematics plus several domain assumptions. The main new burden is treating vertex density as a sufficient geometry signal and assuming the surface solution operator is learnable on the OT-embedded latent manifold. Benchmark-specific free parameters (latent shape, expansion factor, voxel size, Sinkhorn β, PPMM iterations, normal features) were tuned on test error. No new physical entities are introduced.

free parameters (6)
  • Expansion factor α (latent grid size multiplier) = α = 3
    Sets n2 = ceil(sqrt(α*n1))^2; α=3 chosen by lowest test error in Fig 8 for both car datasets, with no validation split described.
  • Voxel downsampling size r = r = 0.05 for DrivAerNet
    Controls density normalization and number of points; r=0.05 chosen from test MSE in Sec 6.2.2 and Fig 9a.
  • Sinkhorn entropy regularization β = β = 1e6
    Chosen by hand in Sec 4.2.3 to approximate exact OT without excessive cost; no sensitivity study is reported.
  • Latent mesh shape = Torus
    Selected by test error in Table 8; the explanation that the torus aligns with periodic Fourier modes is post hoc.
  • PPMM iteration count K = K up to 2000, scaling with voxel size
    Set from ablation in Table 9; the paper infers K proportional to n^(1/2), but this is an empirical fit, not a guarantee.
  • Normal feature type = Cross product of normals
    Selected by test error in Table 7; the gain over no normals on ShapeNet is small (6.70 vs 7.19 percent).
assumptions (5)
  • domain assumption For the nonlinear RANS boundary problem, the surface solution operator can be approximated by a learned composition P∘G*∘Q acting on an OT-embedded 2D latent grid.
    Section 2.1 and 3.1 assume this learnability without providing an approximation theorem or error bounds.
  • domain assumption The surface mesh vertex density, normalized to a probability measure, is a sufficient representation of geometry for the solution operator.
    Invoked in Eq (12) and in the choice of uniform discrete masses in Sec 4.2.3; not proven or ablated.
  • standard math Brenier's theorem and Kantorovich duality guarantee existence and uniqueness of the OT maps and plans used.
    Standard results cited in Sec 2.4; used to justify the Monge and Kantorovich formulations.
  • domain assumption The PDEs considered have a unique solution for every input density in F_PDE_c.
    Section 2.1 defines the data class via unique solvability; no verification is given for the RANS cases.
  • domain assumption Pressure drag dominates wall shear drag at the DrivAerNet Reynolds number, so Cd can be computed from the pressure term alone.
    Appendix B, Eqs (27) and (28); stated without quantitative evidence for these car shapes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Geometric Operator Learning with Optimal Transport." pith.science (2026). https://pith.science/paper/46A4TZE4

@misc{pith2026250720065,
  author       = {Pith},
  title        = {Pith review of: Geometric Operator Learning with Optimal Transport},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/46A4TZE4}},
  note         = {Machine review of arXiv:2507.20065}
}
read the original abstract

We propose integrating optimal transport (OT) into operator learning for partial differential equations (PDEs) on complex geometries. Classical geometric learning methods typically represent domains as meshes, graphs, or point clouds. Our approach generalizes discretized meshes to mesh density functions, formulating geometry embedding as an OT problem that maps these functions to a uniform density in a reference space. Compared to previous methods relying on interpolation or shared deformation, our OT-based method employs instance-dependent deformation, offering enhanced flexibility and effectiveness. For 3D simulations focused on surfaces, our OT-based neural operator embeds the surface geometry into a 2D parameterized latent space. By performing computations directly on this 2D representation of the surface manifold, it achieves significant computational efficiency gains compared to volumetric simulation. Experiments with Reynolds-averaged Navier-Stokes equations (RANS) on the ShapeNet-Car and DrivAerNet-Car datasets show that our method achieves better accuracy and also reduces computational expenses in terms of both time and memory usage compared to existing machine learning models. Additionally, our model demonstrates significantly improved accuracy on the FlowBench dataset, underscoring the benefits of employing instance-dependent deformation for datasets with highly variable geometries.

Figures

Figures reproduced from arXiv: 2507.20065 by the authors.

Figure 1
Figure 1. Illustration of the optimal transport neural operator (OTNO). (a) [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. OT plans can be viewed as bi-partite graphs. In the figure, the green nodes [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Convergence Plot: This figure presents a comparison of convergence rates among [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Illustration of Sub-Manifold Solution Operator for RANS Equation on Automotive [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: A simple illustration for OT encoder and OT decoder. The curve represents the [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Results visualization for OTNO on Car Datasets [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]
Figure 7
Figure 7. Figure 7: Results Visualization for OTNO on FlowBench Dataset [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]
Figure 8
Figure 8. Figure 8: Ablation Studies of expand factor for OTNO(Plan) [PITH_FULL_IMAGE:figures/full_fig_p025_8.png]
Figure 9
Figure 9. Figure 9: Ablation Studies of Sampling Mesh Size for OTNO(Map) and OTNO(Plan) on [PITH_FULL_IMAGE:figures/full_fig_p027_9.png]
Figure 10
Figure 10. Figure 10: Results Visualization of Flow Dataset under M1 metric (full space) [PITH_FULL_IMAGE:figures/full_fig_p035_10.png]
Figure 11
Figure 11. Figure 11: Results Visualization of Flow Dataset under M2 metric (boundary) [PITH_FULL_IMAGE:figures/full_fig_p036_11.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. MoNo: Multiscale Optimal Transport Neural Operator for Solving PDEs on General Geometries

    cs.LG 2026-08 conditional novelty 5.0 of 10

    MoNo uses entropy-regularized optimal transport to build balanced, stable latent-space projections in a multiscale neural operator, achieving lower relative L2 errors and GFLOPs than prior neural operators.

  2. Latent Dynamics Graph Convolutional Networks for model order reduction of parameterized time-dependent PDEs

    cs.LG 2026-01 conditional novelty 5.0 of 10

    LD-GCN couples an encoder-free latent-space neural ODE with a graph convolutional decoder, achieving accurate reduced-order modeling of time-dependent parameterized PDEs and detecting bifurcations from the latent traj...

Reference graph

Works this paper leans on

51 extracted references · 27 canonical work pages · cited by 2 Pith papers

  1. [1]

    Diffeomorphic latent neural operators for data-efficient learning of solutions to partial differential equations, 2024

    Zan Ahmad, Shiyi Chen, Minglang Yin, Avisha Kumar, Nicolas Charon, Natalia Trayanova, and Mauro Maggioni. Diffeomorphic latent neural operators for data-efficient learning of solutions to partial differential equations, 2024. URL https://arxiv.org/abs/2411.18014

    arXiv 2024

  2. [2]

    Universal physics transformers

    Benedikt Alkin, Andreas F \"u rst, Simon Schmid, Lukas Gruber, Markus Holzleitner, and Johannes Brandstetter. Universal physics transformers. arXiv preprint arXiv:2402.12365, 2024

    arXiv 2024

  3. [3]

    Prediction of aerodynamic flow fields using convolutional neural networks

    Saakaar Bhatnagar, Yaser Afshar, Shaowu Pan, Karthik Duraisamy, and Shailendra Kaushik. Prediction of aerodynamic flow fields using convolutional neural networks. Computational Mechanics, 64: 0 525--545, 2019

    work page 2019

  4. [4]

    Spherical fourier neural operators: Learning stable dynamics on the sphere

    Boris Bonev, Thorsten Kurth, Christian Hundt, Jaideep Pathak, Maximilian Baust, Karthik Kashinath, and Anima Anandkumar. Spherical fourier neural operators: Learning stable dynamics on the sphere. In International conference on machine learning, pages 2806--2823. PMLR, 2023

    work page 2023

  5. [5]

    Polar factorization and monotone rearrangement of vector-valued functions

    Yann Brenier. Polar factorization and monotone rearrangement of vector-valued functions. Communications on pure and applied mathematics, 44 0 (4): 0 375--417, 1991

    work page 1991

  6. [6]

    The geometry of r-adaptive meshes generated using optimal transport methods

    CJ Budd, Robert D Russell, and E Walsh. The geometry of r-adaptive meshes generated using optimal transport methods. Journal of Computational Physics, 282: 0 113--137, 2015

    work page 2015

  7. [7]

    Shapenet: An information-rich 3d model repository

    Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. In arXiv:1512.03012 [cs.GR], 2015

    arXiv 2015

  8. [8]

    Implicit neural spatial representations for time-dependent pdes

    Honglin Chen, Rundi Wu, Eitan Grinspun, Changxi Zheng, and Peter Yichen Chen. Implicit neural spatial representations for time-dependent pdes. In International Conference on Machine Learning, pages 5162--5177. PMLR, 2023

    work page 2023

Show all 51 references
  1. [9]

    Crom: Continuous reduced-order modeling of pdes using implicit neural representations

    Peter Yichen Chen, Jinxu Xiang, Dong Heon Cho, Yue Chang, GA Pershing, Henrique Teles Maia, Maurizio M Chiaramonte, Kevin Carlberg, and Eitan Grinspun. Crom: Continuous reduced-order modeling of pdes using implicit neural representations. arXiv preprint arXiv:2206.02607, 2022

    2022 arXiv

  2. [10]

    Flash: Fast landmark aligned spherical harmonic parameterization for genus-0 closed brain surfaces

    Pui Tung Choi, Ka Chun Lam, and Lok Ming Lui. Flash: Fast landmark aligned spherical harmonic parameterization for genus-0 closed brain surfaces. SIAM Journal on Imaging Sciences, 8 0 (1): 0 67--94, 2015

    2015

  3. [11]

    Factorized implicit global convolution for automotive computational fluid dynamics prediction

    Chris Choy, Alexey Kamenev, Jean Kossaifi, Max Rietmann, Jan Kautz, and Kamyar Azizzadenesheli. Factorized implicit global convolution for automotive computational fluid dynamics prediction. arXiv preprint arXiv:2502.04317, 2025

    2025 arXiv

  4. [12]

    Shift-suv sample: High-fidelity computational fluid dynamics dataset for suv external aerodynamics, 2025

    Luminary Cloud. Shift-suv sample: High-fidelity computational fluid dynamics dataset for suv external aerodynamics, 2025. URL https://huggingface.co/datasets/LuminaryCloud/shift-suv-samples

    2025

  5. [13]

    An algorithm for the machine calculation of complex fourier series

    James W Cooley and John W Tukey. An algorithm for the machine calculation of complex fourier series. Mathematics of computation, 19 0 (90): 0 297--301, 1965

    1965

  6. [14]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013

    2013

  7. [15]

    Drivaernet: A parametric car dataset for data-driven aerodynamic design and graph-based drag prediction, 2024 a

    Mohamed Elrefaie, Angela Dai, and Faez Ahmed. Drivaernet: A parametric car dataset for data-driven aerodynamic design and graph-based drag prediction, 2024 a . URL https://arxiv.org/abs/2403.08055

    2024 arXiv

  8. [16]

    Drivaernet++: A large-scale multimodal car dataset with computational fluid dynamics simulations and deep learning benchmarks

    Mohamed Elrefaie, Florin Morar, Angela Dai, and Faez Ahmed. Drivaernet++: A large-scale multimodal car dataset with computational fluid dynamics simulations and deep learning benchmarks. arXiv preprint arXiv:2406.09624, 2024 b

    2024 arXiv

  9. [17]

    Alaya, Aur \'e lie Boisbunon, Stanislas Chambon, Laetitia Chapel, Adrien Corenflos, Kilian Fatras, Nemo Fournier, L \'e o Gautheron, Nathalie T.H

    R \'e mi Flamary, Nicolas Courty, Alexandre Gramfort, Mokhtar Z. Alaya, Aur \'e lie Boisbunon, Stanislas Chambon, Laetitia Chapel, Adrien Corenflos, Kilian Fatras, Nemo Fournier, L \'e o Gautheron, Nathalie T.H. Gayraud, Hicham Janati, Alain Rakotomamonjy, Ievgen Redko, Antoin...

    2021

  10. [18]

    Genus zero surface conformal mapping and its application to brain surface mapping

    Xianfeng Gu, Yalin Wang, Tony F Chan, Paul M Thompson, and Shing-Tung Yau. Genus zero surface conformal mapping and its application to brain surface mapping. IEEE transactions on medical imaging, 23 0 (8): 0 949--958, 2004

    2004

  11. [19]

    Gnot: A general neural operator transformer for operator learning

    Zhongkai Hao, Zhengyi Wang, Hang Su, Chengyang Ying, Yinpeng Dong, Songming Liu, Ze Cheng, Jian Song, and Jun Zhu. Gnot: A general neural operator transformer for operator learning. In International Conference on Machine Learning, pages 12556--12569. PMLR, 2023

    2023

  12. [20]

    Nvidia simnet™: An ai-accelerated multi-physics simulation framework

    Oliver Hennigh, Susheela Narasimhan, Mohammad Amin Nabian, Akshay Subramaniam, Kaustubh Tangsali, Zhiwei Fang, Max Rietmann, Wonmin Byeon, and Sanjay Choudhry. Nvidia simnet™: An ai-accelerated multi-physics simulation framework. In International conference on computational sc...

    2021

  13. [21]

    Darcy's law and the field equations of the flow of underground fluids

    M King Hubbert. Darcy's law and the field equations of the flow of underground fluids. Transactions of the AIME, 207 0 (01): 0 222--239, 1956

    1956

  14. [22]

    Finite element solution of the helmholtz equation with high wave number part i: The h-version of the fem

    Frank Ihlenburg and Ivo Babu s ka. Finite element solution of the helmholtz equation with high wave number part i: The h-version of the fem. Computers & Mathematics with Applications, 30 0 (9): 0 9--37, 1995

    1995

  15. [23]

    On a problem of monge

    LK Kantorovich. On a problem of monge. Journal of Mathematical Sciences, 133 0 (4), 2006

    2006

  16. [24]

    Neural operator: Learning maps between function spaces with applications to pdes

    Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces with applications to pdes. Journal of Machine Learning Research, 24 0 (89): 0 1--97, 2023

    2023

  17. [25]

    Deepgcns: Can gcns go as deep as cnns? In Proceedings of the IEEE/CVF international conference on computer vision, pages 9267--9276, 2019

    Guohao Li, Matthias Muller, Ali Thabet, and Bernard Ghanem. Deepgcns: Can gcns go as deep as cnns? In Proceedings of the IEEE/CVF international conference on computer vision, pages 9267--9276, 2019

    2019

  18. [26]

    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. arXiv preprint arXiv:2010.08895, 2020 a

    2010 arXiv

  19. [27]

    Neural operator: Graph kernel network for partial differential equations

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Graph kernel network for partial differential equations. arXiv preprint arXiv:2003.03485, 2020 b

    2003 arXiv

  20. [28]

    Neural operator: Graph kernel network for partial differential equations

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Graph kernel network for partial differential equations. ICLR, 2021

    2021

  21. [29]

    Fourier neural operator with learned deformations for pdes on general geometries

    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 0 (388): 0 1--26, 2023 a

    2023

  22. [30]

    Geometry-informed neural operator for large-scale 3d PDE s

    Zongyi Li, Nikola Borislavov Kovachki, Chris Choy, Boyi Li, Jean Kossaifi, Shourya Prakash Otta, Mohammad Amin Nabian, Maximilian Stadler, Christian Hundt, Kamyar Azizzadenesheli, and Anima Anandkumar. Geometry-informed neural operator for large-scale 3d PDE s. NIPS, 2023 b

    2023

  23. [31]

    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 0 (3): 0 218--229, 2021

    2021

  24. [32]

    Large-scale optimal transport map estimation using projection pursuit

    Cheng Meng, Yuan Ke, Jingyi Zhang, Mengrui Zhang, Wenxuan Zhong, and Ping Ma. Large-scale optimal transport map estimation using projection pursuit. Advances in Neural Information Processing Systems, 32, 2019

    2019

  25. [33]

    A double fourier sphere method for d-dimensional manifolds

    Sophie Mildenberger and Michael Quellmalz. A double fourier sphere method for d-dimensional manifolds. Sampling Theory, Signal Processing, and Data Analysis, 21 0 (2): 0 23, 2023

    2023

  26. [34]

    M \'e moire sur la th \'e orie des d \'e blais et des remblais

    Gaspard Monge. M \'e moire sur la th \'e orie des d \'e blais et des remblais. Mem. Math. Phys. Acad. Royale Sci., pages 666--704, 1781

  27. [35]

    Computational optimal transport: With applications to data science

    Gabriel Peyr \'e , Marco Cuturi, et al. Computational optimal transport: With applications to data science. Foundations and Trends in Machine Learning , 11 0 (5-6): 0 355--607, 2019

    2019

  28. [36]

    Learning mesh-based simulation with graph networks

    Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter W Battaglia. Learning mesh-based simulation with graph networks. ICLR, 2021

    2021

  29. [37]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space

    Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017

    2017

  30. [38]

    Assanet: An anisotropic separable set abstraction for efficient point cloud representation learning

    Guocheng Qian, Hasan Hammoud, Guohao Li, Ali Thabet, and Bernard Ghanem. Assanet: An anisotropic separable set abstraction for efficient point cloud representation learning. Advances in Neural Information Processing Systems, 34: 0 28119--28130, 2021

    2021

  31. [39]

    Pointnext: Revisiting pointnet++ with improved training and scaling strategies

    Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in neural information processing systems, 35: 0 23192--23204, 2022

    2022

  32. [40]

    Optimal transport for applied mathematicians, volume 87

    Filippo Santambrogio. Optimal transport for applied mathematicians, volume 87. Springer, 2015

    2015

  33. [41]

    Koupa \

    Louis Serrano, Lise Le Boudec, Armand Kassa \" Koupa \" , Thomas X Wang, Yuan Yin, Jean-No \"e l Vittaut, and Patrick Gallinari. Operator learning with neural fields: Tackling pdes on general geometries. Advances in Neural Information Processing Systems, 36: 0 70581--70611, 2023

    2023

  34. [42]

    Elias M. Stein. Singular Integrals and Differentiability Properties of Functions (PMS-30). Princeton University Press, 1970

    1970

  35. [43]

    Flowbench: A large scale benchmark for flow simulation over complex geometries

    Ronak Tali, Ali Rabeh, Cheng-Hau Yang, Mehdi Shadkhah, Samundra Karki, Abhisek Upadhyaya, Suriya Dhakshinamoorthy, Marjan Saadati, Soumik Sarkar, Adarsh Krishnamurthy, et al. Flowbench: A large scale benchmark for flow simulation over complex geometries. arXiv preprint arXiv:2...

    2024 arXiv

  36. [44]

    Deep learning methods for reynolds-averaged navier--stokes simulations of airfoil flows

    Nils Thuerey, Konstantin Wei enow, Lukas Prantl, and Xiangyu Hu. Deep learning methods for reynolds-averaged navier--stokes simulations of airfoil flows. AIAA Journal, 58 0 (1): 0 25--36, 2020

    2020

  37. [45]

    An overview of naca 6-digit airfoil series characteristics with reference to airfoils for large wind turbine blades

    W Timmer. An overview of naca 6-digit airfoil series characteristics with reference to airfoils for large wind turbine blades. In 47th AIAA aerospace sciences meeting including the new horizons forum and aerospace exposition, page 268, 2009

    2009

  38. [46]

    Learning three-dimensional flow for interactive aerodynamic design

    Nobuyuki Umetani and Bernd Bickel. Learning three-dimensional flow for interactive aerodynamic design. ACM Transactions on Graphics (TOG), 37 0 (4): 0 1--10, 2018

    2018

  39. [47]

    Transolver: A fast transformer solver for pdes on general geometries

    Haixu Wu, Huakun Luo, Haowen Wang, Jianmin Wang, and Mingsheng Long. Transolver: A fast transformer solver for pdes on general geometries. arXiv preprint arXiv:2402.02366, 2024

    2024 arXiv

  40. [48]

    Dimon: Learning solution operators of partial differential equations on a diffeomorphic family of domains

    Minglang Yin, Nicolas Charon, Ryan Brody, Lu Lu, Natalia Trayanova, and Mauro Maggioni. Dimon: Learning solution operators of partial differential equations on a diffeomorphic family of domains. arXiv preprint arXiv:2402.07250, 2024

    2024 arXiv

  41. [49]

    Continuous pde dynamics forecasting with implicit neural representations

    Yuan Yin, Matthieu Kirchmeyer, Jean-Yves Franceschi, Alain Rakotomamonjy, and Patrick Gallinari. Continuous pde dynamics forecasting with implicit neural representations. arXiv preprint arXiv:2209.14855, 2022

    2022 arXiv

  42. [50]

    Point-bert: Pre-training 3d point cloud transformers with masked point modeling

    Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19313--19322, 2022

    2022

  43. [51]

    Diffeomorphism neural operator for various domains and parameters of partial differential equations

    Zhiwei Zhao, Changqing Liu, Yingguang Li, Zhibin Chen, and Xu Liu. Diffeomorphism neural operator for various domains and parameters of partial differential equations. Communications Physics, 8 0 (1): 0 15, 2025

    2025

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.