arxiv: 2604.11403 · v1 · submitted 2026-04-13 · 💻 cs.CE · cs.AI· physics.flu-dyn

Recognition: unknown

One Scale at a Time: Scale-Autoregressive Modeling for Fluid Flow Distributions

Mario Lino , Nils Thuerey

Authors on Pith no claims yet

Pith reviewed 2026-05-10 16:03 UTC · model grok-4.3

classification 💻 cs.CE cs.AIphysics.flu-dyn

keywords scale-autoregressive modelingfluid flow distributionsunstructured meshesunsteady flowsgenerative modelingdiffusion modelsflow matchingturbulent statistics

0 comments

The pith

Scale-autoregressive modeling generates fluid flow distributions by sampling hierarchically from coarse to fine scales.

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

The paper introduces scale-autoregressive modeling (SAR) to produce full distributions of possible states in unsteady fluid flows on unstructured meshes. It first generates a low-resolution field, then refines successive higher-resolution versions while conditioning each level only on the preceding coarser prediction. This structure concentrates most computation where uncertainty is largest and trims the number of refinement steps required at fine scales. On benchmarks of varying complexity, SAR reduces distributional error and improves per-sample accuracy relative to multi-scale graph neural network diffusion models, while matching or exceeding a linear-time flow-matching transformer yet completing inference two to seven times faster depending on the task. The result is a practical route to estimating quantities such as turbulent kinetic energy and spatial correlations without the prohibitive cost of full PDE solvers or the error buildup of learned time-steppers.

Core claim

Scale-autoregressive modeling generates fluid flow fields by first producing a low-resolution version and then refining it progressively at higher resolutions, with each finer level conditioned on the coarser prediction. This factorization allows the model to focus computational resources where uncertainty is highest while using fewer operations at detailed scales. Across multiple unsteady flow benchmarks on unstructured meshes, this yields lower distributional error and higher sample accuracy than multi-scale graph neural network diffusion models, and matches or exceeds the performance of a flow-matching transformer solver at two to seven times the speed.

What carries the argument

The hierarchical coarse-to-fine autoregressive sampling process that generates flow distributions on unstructured meshes by conditioning each finer resolution on the preceding coarser output.

If this is right

Statistical flow quantities such as turbulent kinetic energy and two-point correlations become feasible to estimate quickly and accurately without full PDE solves.
Independent sampling of entire states avoids the compounding rollout error that limits learned time-stepping surrogates over long horizons.
Most uncertainty is resolved at coarse scales, so computation is reduced at fine scales while still preserving mesh-level detail.
The method outperforms diffusion models based on multi-scale GNNs in both error metrics and sample quality while delivering clear speed gains over flow-matching transformers.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same coarse-to-fine conditioning structure could be tested on other multi-scale physical systems where uncertainty varies strongly with resolution.
Adaptive choice of scale progression based on local flow features might further cut unnecessary computation on heterogeneous meshes.
The efficiency profile suggests direct use in engineering workflows that require many independent realizations for uncertainty quantification.

Load-bearing premise

Conditioning each finer-scale sample only on the coarser-scale prediction recovers accurate fine-scale statistics and physical correlations without systematic bias or loss of important dependencies in complex unsteady flows.

What would settle it

A direct comparison on an unsteady flow benchmark showing that two-point correlations or energy spectra at the finest scale deviate systematically from ground-truth references when produced by the hierarchical model but remain consistent when produced by non-hierarchical baselines.

Figures

Figures reproduced from arXiv: 2604.11403 by Mario Lino, Nils Thuerey.

**Figure 2.** Figure 2: (a) Speed/distributional-accuracy trade-off on [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: Speed and distributional/sample-accuracy trade-off on the [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: (a) Performance comparison between the latent SAR and non-latent SAR models across [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

**Figure 5.** Figure 5: (a) Impact of ablation variants compared to our default configuration (striped bars), mea [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗

**Figure 6.** Figure 6: Horizontal component of the velocity field at each of the three resolution scales ( [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗

**Figure 7.** Figure 7: Diffusion GNN (Lino et al., 2025) and our SAR model both partition the node set V into resolution scales but differ in their processing approach. Diffusion GNNs apply the same number of denoising steps across all scales using local message passing and local unpooling. In contrast, SAR allows fewer denoising steps at finer scales, making it feasible to use otherwise expensive attention. Upsampling is perfor… view at source ↗

**Figure 8.** Figure 8: (a) Impact of the number of scales, measured using the Wasserstein-2 distance. Bars [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗

**Figure 9.** Figure 9: Samples from LDGN (Lino et al., 2025), FMT-8, and SAR for (a) a simulation from the ELLIPSEFLOW-HIGHRE dataset, and (b) a simulation from the ELLIPSEFLOW-AOA10 dataset. SAR produces the most accurate samples across both settings. Reynols shear stress TKE Inference time: 1.7s 25s 4.2s [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗

**Figure 10.** Figure 10: Turbulent kinetic energy (top row), Reynolds shear stress (middle row), and inference [PITH_FULL_IMAGE:figures/full_fig_p020_10.png] view at source ↗

**Figure 11.** Figure 11: Speed/sample-accuracy trade-off on the ELLIPSEFLOW-INDIST, ELLIPSEFLOW-HIGHRE, and ELLIPSEFLOW-AOA10 datasets. Curves for LDGN and LFM-GNN are obtained using 3, 5, 10, and 25 denoising steps. FMT curves use 3, 5, 10, 15, and 20 steps. The yellow SAR curve corresponds to using 2, 3, 5, and 10 denoising steps across all scales. The red SAR curve uses a different number of steps for each of the three scales:… view at source ↗

**Figure 12.** Figure 12: (a) Standard deviation of pressure on a wing geometry unseen during training ( [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗

read the original abstract

Analyzing unsteady fluid flows often requires access to the full distribution of possible temporal states, yet conventional PDE solvers are computationally prohibitive and learned time-stepping surrogates quickly accumulate error over long rollouts. Generative models avoid compounding error by sampling states independently, but diffusion and flow-matching methods, while accurate, are limited by the cost of many evaluations over the entire mesh. We introduce scale-autoregressive modeling (SAR) for sampling flows on unstructured meshes hierarchically from coarse to fine: it first generates a low-resolution field, then refines it by progressively sampling higher resolutions conditioned on coarser predictions. This coarse-to-fine factorization improves efficiency by concentrating computation at coarser scales, where uncertainty is greatest, while requiring fewer steps at finer scales. Across unsteady-flow benchmarks of varying complexity, SAR attains substantially lower distributional error and higher per-sample accuracy than state-of-the-art diffusion models based on multi-scale GNNs, while matching or surpassing a flow-matching Transolver (a linear-time transformer) yet running 2-7x faster than this depending on the task. Overall, SAR provides a practical tool for fast and accurate estimation of statistical flow quantities (e.g., turbulent kinetic energy and two-point correlations) in real-world settings.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

SAR introduces a practical coarse-to-fine autoregressive factorization for fluid flow distributions that trades some modeling assumptions for clear speed gains over diffusion baselines.

read the letter

SAR introduces a hierarchical generative approach for fluid flows on unstructured meshes: it samples a coarse field first, then progressively refines finer scales conditioned on the coarser predictions. This factorization concentrates compute where uncertainty is highest and reduces the number of steps needed at fine resolutions compared to full diffusion or flow-matching runs over the entire mesh. The abstract shows this yields lower distributional error and better per-sample accuracy than multi-scale GNN diffusion models, while matching or beating a flow-matching Transolver at 2-7x lower cost on unsteady benchmarks. The motivation for fast statistical estimates like turbulent kinetic energy and two-point correlations without long rollouts is straightforward and useful for engineering settings. The work is distinct from prior diffusion and flow-matching lines in the fluid domain. The main soft spot is that the central assumption—coarse conditioning suffices to recover accurate fine-scale statistics and correlations—rests on limited evidence from the abstract alone. If the coarse aggregation discards high-wavenumber information that correlates with retained modes, the finer samples could still show under-dispersion or biased spectra even when aggregate metrics look good. Without seeing the exact aggregation operator, ablations, error bars, or data splits, the performance edge is hard to judge fully. The stress-test concern about possible bias in fine-scale recovery is plausible and worth checking against the full results. This paper is for researchers in ML for CFD who need fast generative sampling of flow distributions rather than deterministic surrogates. A reader focused on practical generative tools for physics would get value from the benchmarks and the architectural idea. It deserves peer review so the experimental details and the conditioning claim can be examined properly.

Referee Report

2 major / 1 minor

Summary. The paper introduces scale-autoregressive modeling (SAR) for sampling distributions of unsteady fluid flows on unstructured meshes. It factorizes the joint distribution hierarchically as p(x_L | x_{L-1}) ... p(x_1), first generating a low-resolution field then progressively refining higher resolutions conditioned on coarser predictions. This is presented as improving efficiency by concentrating computation at coarser scales. Across unsteady-flow benchmarks, SAR is claimed to achieve substantially lower distributional error and higher per-sample accuracy than state-of-the-art multi-scale GNN diffusion models, while matching or surpassing a flow-matching Transolver yet running 2-7x faster.

Significance. If the empirical results hold under scrutiny, SAR offers a practical efficiency improvement for generative modeling in computational fluid dynamics, enabling faster estimation of statistical quantities such as turbulent kinetic energy and two-point correlations without the full cost of many-step diffusion or flow-matching on high-resolution meshes. The hierarchical approach could extend to other multi-scale physical simulation tasks.

major comments (2)

[Abstract and §3] Abstract and §3 (Modeling Approach): The claim that the coarse-to-fine factorization recovers the full joint distribution without systematic bias is load-bearing for the performance claims. On unstructured meshes the coarse field is obtained via unspecified aggregation; if this discards high-wavenumber information that is statistically dependent on retained modes, conditioning finer-scale GNN/transformer samples on the coarse prediction cannot restore correct two-point statistics or energy spectra. The reported lower distributional error metric can be satisfied even with conditional under-dispersion at fine scales, so explicit validation (e.g., spectra or correlation comparisons on the benchmarks) is required.
[§5] §5 (Experiments): The abstract reports 2-7x speedups and accuracy gains versus the flow-matching Transolver, but without access to the precise data splits, error bars, ablation studies on the number of scales, or the exact aggregation operator used for coarsening, the support for these quantitative claims cannot be fully verified. Reproducibility artifacts (code, seeds, mesh details) would be needed to confirm the gains are not sensitive to implementation choices.

minor comments (1)

Notation for the scale indices (L, L-1, …, 1) and the precise definition of the conditioning mechanism should be introduced earlier and used consistently to aid readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major comment below and describe the revisions we will make to strengthen the manuscript.

read point-by-point responses

Referee: [Abstract and §3] Abstract and §3 (Modeling Approach): The claim that the coarse-to-fine factorization recovers the full joint distribution without systematic bias is load-bearing for the performance claims. On unstructured meshes the coarse field is obtained via unspecified aggregation; if this discards high-wavenumber information that is statistically dependent on retained modes, conditioning finer-scale GNN/transformer samples on the coarse prediction cannot restore correct two-point statistics or energy spectra. The reported lower distributional error metric can be satisfied even with conditional under-dispersion at fine scales, so explicit validation (e.g., spectra or correlation comparisons on the benchmarks) is required.

Authors: We appreciate the referee's emphasis on this foundational aspect. The SAR factorization is constructed as an exact decomposition p(x) = p(x_1) ∏_{l=2}^L p(x_l | x_{<l}), where x_1 denotes the coarsest scale; this guarantees that the marginal distribution at every scale is recovered without bias whenever the conditional models are accurate. On unstructured meshes we employ a conservative averaging aggregation that preserves integral invariants such as total momentum and kinetic energy. In the revised §3 we will explicitly define this operator together with a short proof sketch showing that the hierarchical conditioning does not introduce systematic bias in the two-point statistics. To directly validate fine-scale fidelity we will add, in the revised §5, side-by-side comparisons of kinetic-energy spectra and two-point correlation functions for all benchmarks, confirming that SAR reproduces the ground-truth distributions without conditional under-dispersion. revision: yes
Referee: [§5] §5 (Experiments): The abstract reports 2-7x speedups and accuracy gains versus the flow-matching Transolver, but without access to the precise data splits, error bars, ablation studies on the number of scales, or the exact aggregation operator used for coarsening, the support for these quantitative claims cannot be fully verified. Reproducibility artifacts (code, seeds, mesh details) would be needed to confirm the gains are not sensitive to implementation choices.

Authors: We fully agree that reproducibility details are essential for verifying the reported speed-ups and accuracy gains. In the revised manuscript we will (i) report all metrics with error bars computed over at least five independent random seeds, (ii) include an ablation study on the number of scales, (iii) specify the exact data splits, mesh resolutions, and aggregation operator (as noted in the response to the first comment), and (iv) release the complete code base, trained checkpoints, random seeds, and mesh files upon acceptance. These additions will allow independent verification that the 2–7× speed-ups and distributional improvements are robust. revision: yes

Circularity Check

0 steps flagged

No circularity: architectural choice with external benchmark validation

full rationale

The provided abstract and description present SAR as an independent modeling architecture: a coarse-to-fine hierarchical factorization for sampling on unstructured meshes, with direct empirical comparisons to multi-scale GNN diffusion models and a flow-matching Transolver on unsteady-flow benchmarks. No equations, derivations, or claims are shown that reduce performance metrics, distributional error, or statistical quantities (e.g., TKE, correlations) to quantities defined or fitted within the paper itself. The factorization p(x_L | x_{L-1}) ... is introduced as a design decision for efficiency, not derived from self-referential inputs. External baselines and benchmark results supply independent content, satisfying the self-contained criterion for a non-circular finding.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Based on the abstract, the central claim rests on standard assumptions from generative modeling and mesh-based fluid representations rather than new free parameters or invented entities. No specific fitted constants or ad-hoc postulates are mentioned.

pith-pipeline@v0.9.0 · 5522 in / 1278 out tokens · 52916 ms · 2026-05-10T16:03:08.496484+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

23 extracted references · 11 canonical work pages · 2 internal anchors

[1]

AB-UPT: Scaling Neural CFD Surrogates for High- Fidelity Automotive Aerodynamics Simulations via Anchored-Branched Universal Physics Transformers,

Benedikt Alkin, Maurits Bleeker, Richard Kurle, Tobias Kronlachner, Reinhard Sonnleitner, Matthias Dorfer, and Johannes Brandstetter. Ab-upt: Scaling neural cfd surrogates for high- fidelity automotive aerodynamics simulations via anchored-branched universal physics transform- ers.arXiv preprint arXiv:2502.09692,

work page arXiv
[2]

Layer Normalization

Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization.arXiv preprint arXiv:1607.06450,

work page internal anchor Pith review Pith/arXiv arXiv
[3]

Flow matching meets PDE s: A unified framework for physics-constrained generation

Giacomo Baldan, Qiang Liu, Alberto Guardone, and Nils Thuerey. Flow matching meets pdes: A unified framework for physics-constrained generation.arXiv preprint arXiv:2506.08604,

work page arXiv
[4]

Relational inductive biases, deep learning, and graph networks

Peter W Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, et al. Relational inductive biases, deep learning, and graph networks.arXiv:1806.01261,

work page internal anchor Pith review arXiv
[5]

Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman

doi: 10.2514/1.J061454. Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 11315–11325,

work page doi:10.2514/1.j061454
[6]

Diffusion models beat gans on image synthesis.Advances in neural information processing systems, 34:8780–8794, 2021a

Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis.Advances in neural information processing systems, 34:8780–8794, 2021a. Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. InAdvances in Neural Information Processing Systems 34, pp. 8780–8794, 2021b. URL https: //proceedings.neu...

2021
[7]

Multiscale MeshGraphNets

Meire Fortunato, Tobias Pfaff, Peter Wirnsberger, Alexander Pritzel, and Peter Battaglia. Multiscale MeshGraphNets. InICML 2022 Workshop on AI for Science,

2022
[8]

Energy-based trans- formers are scalable learners and thinkers.arXiv preprint arXiv:2507.02092, 2025

Alexi Gladstone, Ganesh Nanduru, Md Mofijul Islam, Peixuan Han, Hyeonjeong Ha, Aman Chadha, Yilun Du, Heng Ji, Jundong Li, and Tariq Iqbal. Energy-Based Transformers are Scalable Learn- ers and Thinkers.arXiv preprint arXiv:2507.02092,

work page arXiv
[9]

Kingma and Max Welling

11 Diederik P. Kingma and Max Welling. Auto-Encoding Variational Bayes. In2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings,

2014
[10]

Turbulent flow simulation using autoregressive condi- tional diffusion models.arXiv preprint arXiv:2309.01745,

Georg Kohl, Li-Wei Chen, and Nils Thuerey. Turbulent flow simulation using autoregressive condi- tional diffusion models.arXiv preprint arXiv:2309.01745,

work page arXiv
[11]

Transformer for partial differential equations’ operator learning.arXiv preprint arXiv:2205.13671, 2022

Tianyi Li, Alessandra S Lanotte, Michele Buzzicotti, Fabio Bonaccorso, and Luca Biferale. Multi- scale reconstruction of turbulent rotating flows with generative diffusion models.Atmosphere, 15 (1):60, 2023b. Zijie Li, Kazem Meidani, and Amir Barati Farimani. Transformer for partial differential equations’ operator learning.arXiv:2205.13671,

work page arXiv
[12]

Learning Distributions of Complex Fluid Simulations with Diffusion Graph Networks

Mario Lino, Tobias Pfaff, and Nils Thuerey. Learning Distributions of Complex Fluid Simulations with Diffusion Graph Networks. In13th International Conference on Learning Representations (ICLR 2025),

2025
[13]

Transolver++: An accurate neural solver for pdes on million-scale geometries.arXiv preprint arXiv:2502.02414, 2025

Huakun Luo, Haixu Wu, Hang Zhou, Lanxiang Xing, Yichen Di, Jianmin Wang, and Mingsheng Long. Transolver++: An accurate neural solver for pdes on million-scale geometries.arXiv preprint arXiv:2502.02414,

work page arXiv
[14]

Battaglia

Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter W. Battaglia. Learning mesh- based simulation with graph networks. In9th International Conference on Learning Representa- tions (ICLR 2021),

2021
[15]

Battaglia

Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter W. Battaglia. Learning to simulate complex physics with graph networks. InProceedings of the 37th International Conference on Machine Learning (ICML 2020), volume 119, pp. 8459–8468,

2020
[16]

arXiv preprint arXiv:2112.15275 , year=

Kimberly Stachenfeld, Drummond B Fielding, Dmitrii Kochkov, Miles Cranmer, Tobias Pfaff, Jonathan Godwin, Can Cui, Shirley Ho, Peter Battaglia, and Alvaro Sanchez-Gonzalez. Learned coarse models for efficient turbulence simulation.arXiv:2112.15275,

work page arXiv
[17]

Geometry-aware operator transformer as an efficient and accurate neural surrogate for PDEs on arbitrary domains.arXiv preprint arXiv:2505.18781, 2025

Shizheng Wen, Arsh Kumbhat, Levi Lingsch, Sepehr Mousavi, Yizhou Zhao, Praveen Chan- drashekar, and Siddhartha Mishra. Geometry aware operator transformer as an efficient and accurate neural surrogate for pdes on arbitrary domains.arXiv preprint arXiv:2505.18781,

work page arXiv
[18]

A.2 ARCHITECTUREDETAILS The condition encoder, autoregressive module, and sampler components of our SAR model are all based on the Transolver architecture proposed in Wu et al

for a sample from theELLIPSEFLOW-INDISTdataset. A.2 ARCHITECTUREDETAILS The condition encoder, autoregressive module, and sampler components of our SAR model are all based on the Transolver architecture proposed in Wu et al. (2024), enhanced with adaptive temper- ature as introduced by Luo et al. (2025). This design provides a global receptive field effic...

2024
[19]

It is conditioned on each node’s spatial locationx j, scale one-hot vector γj, and latent representationsy j andz j

applied, at autoregressive stepk, to nodesj∈ S k. It is conditioned on each node’s spatial locationx j, scale one-hot vector γj, and latent representationsy j andz j. The input feature vectors to the Transolver are defined as vj ←[MLP([s j,r,z j,MLP([x j,γ j]) +MLP(y j)]),r],∀j∈ S k,(19) wheres j,r is the intermediate solution at denoising timer, andr∈R F...

2016
[20]

Diffusion GNNs apply the same number of denoising steps across all scales using local message passing and local unpooling

and our SAR model both partition the node setVinto resolution scales but differ in their processing approach. Diffusion GNNs apply the same number of denoising steps across all scales using local message passing and local unpooling. In contrast, SAR allows fewer denoising steps at finer scales, making it feasible to use otherwise expensive attention. Upsa...

2022
[21]

2025)0.31 ± 0.18 0.24 ± 0.18 0.46 ± 0.22 0.14 ± 0.03 0.68 ± 0.17 0.64 ± 0.09 10 LFM-GNN (Lino et al

0.23 ± 0.12 0.17 ± 0.08 0.42 ± 0.180.10 ± 0.020.57 ± 0.15 0.59 ± 0.11 50 FM-GNN (Lino et al. 2025)0.31 ± 0.18 0.24 ± 0.18 0.46 ± 0.22 0.14 ± 0.03 0.68 ± 0.17 0.64 ± 0.09 10 LFM-GNN (Lino et al. 2025)0.26 ± 0.14 0.19 ± 0.09 0.44 ± 0.18 0.12 ± 0.02 0.63 ± 0.17 0.61 ± 0.11 10 FMT-4 0.29 ± 0.23 0.19 ± 0.15 0.44 ± 0.23 0.11 ± 0.03 0.67 ± 0.21 0.79 ± 0.14 20 SA...

2025
[22]

2025)0.995 ± 0.0070.998 ± 0.0020.986 ± 0.0190.997 ± 0.0010.991 ± 0.0090.966 ± 0.028 50 FM-GNN (Lino et al

0.994 ± 0.0060.997 ± 0.0010.988 ± 0.0150.994 ± 0.0020.992 ± 0.0070.968 ± 0.026 50 LDGN (Lino et al. 2025)0.995 ± 0.0070.998 ± 0.0020.986 ± 0.0190.997 ± 0.0010.991 ± 0.0090.966 ± 0.028 50 FM-GNN (Lino et al. 2025)0.995 ± 0.0070.997 ± 0.0020.987 ± 0.0150.996 ± 0.0030.991 ± 0.0090.966 ± 0.029 10 LFM-GNN (Lino et al. 2025)0.995 ± 0.0080.998 ± 0.0020.985 ± 0.0...

2025
[23]

SAR produces the most accurate samples across both settings

19 (a) OOD test case with Re = 1083 (b) OOD test case with AoA = 10 deg FMT-8LDGN SARFMT-8LDGN SAR Figure 9: Samples from LDGN (Lino et al., 2025), FMT-8, and SAR for (a) a simulation from the ELLIPSEFLOW-HIGHREdataset, and (b) a simulation from theELLIPSEFLOW-AOA10dataset. SAR produces the most accurate samples across both settings. Reynols shear stress ...

2025