pith. sign in

arxiv: 2606.02997 · v1 · pith:HIYJDAH4new · submitted 2026-06-02 · ⚛️ physics.comp-ph

TransportBench: A Comprehensive Benchmark for Non-Equilibrium Flow Transport

Xu Wang , Minghao Li , Qizhen Hong , Yang Liu , Chen-an Zhang , Shuai Zhang , Wenhao Li , Yonghao Zhang
show 1 more author
Tianbai Xiao
This is my paper

Pith reviewed 2026-06-28 08:06 UTC · model grok-4.3

classification ⚛️ physics.comp-ph
keywords non-equilibrium flowsscientific machine learningbenchmark datasetneural architecturesrarefied gas dynamicsshock discontinuitiesfluid transport
0
0 comments X

The pith

Neural network performance on non-equilibrium flows varies sharply with flow regime and no single architecture dominates all tasks.

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

The paper introduces TransportBench, a high-fidelity dataset spanning continuum to rarefied regimes, inert to reactive gases, and translational to internal non-equilibrium effects. It applies unified evaluation protocols to representative neural architectures and reports that success depends on the specific flow features present. A sympathetic reader would care because the results indicate that architectural inductive biases confer targeted advantages for smooth fields, shock discontinuities, or high-order statistics rather than general superiority. This positions the benchmark as a diagnostic tool for developing models that handle transport beyond standard Navier-Stokes descriptions.

Core claim

TransportBench shows that model performance exhibits a pronounced dependence upon the specific flow characteristics. No single architecture consistently performs best for all the tasks. Instead, different architectural inductive biases provide distinct advantages in capturing smooth flow fields, shock-induced discontinuities, and high-order non-equilibrium statistics.

What carries the argument

The TransportBench dataset spanning continuum/rarefied, inert/reactive, and translational/internal non-equilibrium regimes, together with unified evaluation protocols that test robustness to discontinuities, multi-scale effects, and parameter generalization.

If this is right

  • Different neural architectures are better matched to smooth flow fields than to shock discontinuities or high-order statistics.
  • Model selection for non-equilibrium transport must account for the dominant flow features rather than assume general applicability.
  • The benchmark enables systematic diagnosis of where current scientific machine learning methods fail outside Navier-Stokes hydrodynamics.

Where Pith is reading between the lines

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

  • Hybrid models that combine multiple inductive biases might reduce the observed performance gaps across regimes.
  • Extending the benchmark to include experimental measurements would test whether the simulation-based advantages persist in real data.
  • The findings imply that architecture search for fluid problems should be conditioned on the target flow spectrum rather than performed in a single setting.

Load-bearing premise

The dataset built from the described physical spectrum combined with the unified protocols is sufficient to reveal the true strengths and limitations of the tested neural architectures.

What would settle it

A single neural architecture that achieves the best score on every task and every metric in the benchmark would falsify the reported dependence on flow characteristics.

Figures

Figures reproduced from arXiv: 2606.02997 by Chen-an Zhang, Minghao Li, Qizhen Hong, Shuai Zhang, Tianbai Xiao, Wenhao Li, Xu Wang, Yang Liu, Yonghao Zhang.

Figure 1
Figure 1. Figure 1: Representative ground-truth flow fields of the four physical scenarios in Transport￾Bench. From left to right: (a) rarefied airfoil flows with varying geometries for geometry-dependent pre￾dictions; (b) rarefied cylinder flow across a range of Mach and Knudsen numbers for operating-condition￾dependent predictions; (c) lid-driven cavity flows for distribution-function and high-order moment predic￾tions; and… view at source ↗
Figure 2
Figure 2. Figure 2: Baseline RAE2822 airfoil and CST-generated geometric variations. [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Representative temperature fields in the cylinder flow. The four cases correspond to different [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Representative non-equilibrium moment fields in the lid-driven cavity flow. The figure shows the [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Representative Mach-number contour for the high-enthalpy double-cone flow at [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Quantitative performance overview across the four TransportBench tasks. Subfigures (a)-(c) report [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison on Task I: rarefied airfoil flow. The rows show the DSMC reference solution, model [PITH_FULL_IMAGE:figures/full_fig_p024_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Representative predictions on Task II: rarefied cylinder flow. The rows show the DSMC reference [PITH_FULL_IMAGE:figures/full_fig_p026_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison on Task III: lid-driven cavity flow. The figure compares reference, predicted, and [PITH_FULL_IMAGE:figures/full_fig_p028_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Particle distribution function prediction in the lid-driven cavity flow. The figure compares the [PITH_FULL_IMAGE:figures/full_fig_p030_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: U-Net prediction on the double-cone task without Fourier feature injection. The figure compares [PITH_FULL_IMAGE:figures/full_fig_p031_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Representative predictions of DeepONet and FNO on the double-cone task. Panel (a) shows that [PITH_FULL_IMAGE:figures/full_fig_p033_12.png] view at source ↗
read the original abstract

Scientific machine learning models, as versatile tools for numerical simulation and analysis, are increasingly transforming the landscape of fluid mechanics research. However, existing datasets and benchmarks are primarily limited to continuum fluids and provide limited support for non-equilibrium transport phenomena. To address this gap, we present TransportBench, a high-fidelity dataset and standardized benchmark for non-equilibrium flow transport, designed to reveal the strengths and limitations of neural network models across diverse flow regimes. Specifically, the dataset encompasses a broad physical spectrum, covering continuum and rarefied regimes, low-speed and hypersonic flows, inert and chemically reactive gases, and both translational and internal-energy non-equilibrium effects. Built upon this dataset, we systematically benchmark representative neural architectures using unified evaluation protocols to probe key challenges in learning non-equilibrium flows, including robustness to shock-dominated discontinuities and multi-scale effects, as well as generalization across geometry and physical parameters. Numerical results demonstrate that model performance exhibits a pronounced dependence upon the specific flow characteristics. No single architecture consistently performs best for all the tasks. Instead, different architectural inductive biases provide distinct advantages in capturing smooth flow fields, shock-induced discontinuities, and high-order non-equilibrium statistics. By jointly providing the non-equilibrium flow dataset and model benchmark, TransportBench offers a new testbed for the development, evaluation, and diagnosis of scientific machine learning methods for fluid transport beyond the Navier-Stokes hydrodynamics. The benchmark datasets and implementation codes are available under the MIT license.

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.

Referee Report

0 major / 1 minor

Summary. The manuscript introduces TransportBench, a high-fidelity dataset and standardized benchmark for non-equilibrium flow transport in scientific machine learning. It spans a broad physical spectrum including continuum/rarefied regimes, low-speed/hypersonic flows, inert/reactive gases, and translational/internal non-equilibrium effects. Using unified evaluation protocols, the authors benchmark representative neural architectures and report that performance exhibits pronounced dependence on flow characteristics, with no single architecture best across all tasks and different inductive biases suiting smooth fields, shock discontinuities, and high-order non-equilibrium statistics. The dataset, protocols, and MIT-licensed code are released to serve as a testbed beyond Navier-Stokes hydrodynamics.

Significance. If the reported numerical results hold under the described construction and protocols, the work supplies a much-needed resource for diagnosing SciML model limitations in non-equilibrium transport. Strengths include the release of the dataset, unified protocols, and MIT-licensed code, which enable reproducible evaluation across the specified physical spectrum (continuum/rarefied, inert/reactive, translational/internal). This could accelerate development of architectures robust to multi-scale effects and discontinuities.

minor comments (1)
  1. [Abstract / §1] The abstract and introduction would benefit from an explicit statement of the number of flow cases, grid resolutions, and exact neural architectures tested to allow readers to assess coverage without consulting the supplementary material.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive evaluation of the manuscript and the recommendation to accept. We are pleased that the significance of the dataset release, unified protocols, and code availability was recognized as a resource for diagnosing SciML limitations in non-equilibrium transport.

Circularity Check

0 steps flagged

No significant circularity; benchmark paper with empirical results only

full rationale

The paper is a benchmark release that constructs a dataset spanning continuum/rarefied, inert/reactive, and non-equilibrium regimes, then reports empirical performance rankings of neural architectures under unified protocols. No derivation chain, parameter fitting, or predictive claim is present that reduces to its own inputs by construction. The central claim (performance dependence on flow characteristics, with architecture-specific inductive biases) follows directly from the supplied dataset and code without self-referential reduction or load-bearing self-citation. This is the expected non-finding for a dataset/benchmark paper.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is a benchmark and dataset paper with no central mathematical derivation, so the ledger contains no free parameters, axioms, or invented entities.

pith-pipeline@v0.9.1-grok · 5807 in / 1008 out tokens · 12337 ms · 2026-06-28T08:06:08.965248+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 5 canonical work pages · 2 internal anchors

  1. [1]

    Sziroczak, H

    D. Sziroczak, H. Smith, A review of design issues specific to hypersonic flight vehicles, Progress in Aerospace Sciences 84 (2016) 1–28

    2016

  2. [2]

    Gad-el Hak, The MEMS handbook, CRC press, 2001

    M. Gad-el Hak, The MEMS handbook, CRC press, 2001

    2001

  3. [3]

    Cercignani, The Boltzmann Equation and Its Applications, Springer, 1988

    C. Cercignani, The Boltzmann Equation and Its Applications, Springer, 1988

    1988

  4. [4]

    G. A. Bird, Molecular gas dynamics and the direct simulation of gas flows, Oxford university press, 1994

    1994

  5. [5]

    S. L. Brunton, B. R. Noack, P. Koumoutsakos, Machine learning for fluid mechanics, Annual review of fluid mechanics 52 (2020) 477–508

    2020

  6. [6]

    G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, L. Yang, Physics- informed machine learning, Nature Reviews Physics 3 (6) (2021) 422–440

    2021

  7. [7]

    Takamoto, T

    M. Takamoto, T. Praditia, R. Leiteritz, D. MacKinlay, F. Alesiani, D. Pflüger, M. Niepert, PDEBench: An Extensive Benchmark for Scientific Machine Learning, in: Advances in Neural Information Processing Systems (NeurIPS), Vol. 35, 2022, pp. 1596–1611

    2022

  8. [8]

    R. Tali, A. Rabeh, C.-H. Yang, M. Shadkhah, S. Karki, A. Upadhyaya, et al., Flow- Bench: A Large Scale Benchmark for Flow Simulation over Complex Geometries, arXiv preprint arXiv:2409.18285 (2024)

    work page arXiv 2024

  9. [9]

    H. Wang, R. Li, F. Xu, F. Sun, et al., FD-Bench: A Modular and Fair Benchmark for Data-driven Fluid Simulation, arXiv preprint (2025)

    2025

  10. [10]

    Rahaman, A

    N. Rahaman, A. Baratin, D. Arpit, F. Draxler, M. Lin, F. Hamprecht, Y. Bengio, A. Courville, On the spectral bias of neural networks, in: International Conference on Machine Learning (ICML), PMLR, 2019, pp. 5301–5310

    2019

  11. [11]

    T. Xiao, M. Frank, Using neural networks to accelerate the solution of the Boltzmann equation, Journal of Computational Physics 443 (2021) 110521

    2021

  12. [12]

    T. Xiao, M. Frank, RelaxNet: A structure-preserving neural network to approximate the Boltzmann collision operator, Journal of Computational Physics 490 (2023) 112317

    2023

  13. [13]

    Ronneberger, P

    O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks for biomedical image segmentation, in: International Conference on Medical image computing and computer-assisted intervention, Springer, 2015, pp. 234–241

    2015

  14. [14]

    J. Zhai, S. Zhang, J. Chen, Q. He, Autoencoder and its various variants, in: 2018 IEEE international conference on systems, man, and cybernetics (SMC), IEEE, 2018, pp. 415–419. 37

    2018

  15. [15]

    L. Lu, P. Jin, G. Pang, Z. Zhang, G. E. Karniadakis, Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators, Nature machine intelligence 3 (3) (2021) 218–229

    2021

  16. [16]

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

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  17. [17]

    K. Han, Y. Wang, H. Chen, X. Chen, J. Guo, Z. Liu, Y. Tang, A. Xiao, C. Xu, Y. Xu, et al., A survey on vision transformer, IEEE transactions on pattern analysis and machine intelligence 45 (1) (2022) 87–110

    2022

  18. [18]

    H. Zhao, L. Jiang, J. Jia, P. H. Torr, V. Koltun, Point transformer, in: Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 16259–16268

    2021

  19. [19]

    Hao, et al., PINNacle: A comprehensive benchmark of physics-informed neural net- works for solving PDEs, arXiv preprint arXiv:2306.08827 (2023)

    Z. Hao, et al., PINNacle: A comprehensive benchmark of physics-informed neural net- works for solving PDEs, arXiv preprint arXiv:2306.08827 (2023)

    work page arXiv 2023

  20. [20]

    Y. Luo, Y. Chen, Z. Zhang, Cfdbench: A comprehensive benchmark for machine learn- ing methods in fluid dynamics, arXiv preprint arXiv:2310.02672 (2023)

    work page arXiv 2023

  21. [21]

    31, 2018

    A.Jacot, F.Gabriel, C.Hongler, Neuraltangentkernel: Convergenceandgeneralization in neural networks, in: Advances in neural information processing systems (NeurIPS), Vol. 31, 2018

    2018

  22. [22]

    Tancik, P

    M. Tancik, P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J. Barron, R. Ng, Fourier features let networks learn high frequency functions in low dimensional domains, in: Advances in Neural Information Processing Systems (NeurIPS), Vol. 33, 2020, pp. 7537–7547

    2020

  23. [23]

    Nompelis, G

    I. Nompelis, G. Candler, M. Holden, Computational investigation of hypersonic vis- cous/inviscid interactions in high enthalpy flows, in: 36th AIAA Thermophysics Con- ference, 2003, p. 3642

    2003

  24. [24]

    X. Wang, J. Guo, Q. Hong, S. Li, High-fidelity state-to-state modeling of hypersonic flow over a double cone, Physics of Fluids 35 (11) (2023)

    2023

  25. [25]

    I. D. Boyd, T. E. Schwartzentruber, Nonequilibrium Gas Dynamics and Molecular Simulation, Cambridge Aerospace Series, Cambridge University Press, 2017

    2017

  26. [26]

    S. J. Plimpton, S. G. Moore, A. Borner, A. K. Stagg, T. P. Koehler, M. A. Gallis, Direct simulation Monte Carlo on petaflop supercomputers and beyond, Physics of Fluids 31 (8) (2019) 086101

    2019

  27. [27]

    T. Xiao, C. Liu, K. Xu, Q. Cai, A velocity-space adaptive unified gas kinetic scheme for continuum and rarefied flows, Journal of Computational Physics 415 (2020) 109535. 38

    2020

  28. [28]

    Xiao, Kinetic

    T. Xiao, Kinetic. jl: A portable finite volume toolbox for scientific and neural comput- ing, Journal of Open Source Software 6 (62) (2021) 3060

    2021

  29. [29]

    X. Wang, Q. Hong, Y. Hu, Q. Sun, On the accuracy of two-temperature models for hypersonic nonequilibrium flow, Acta Mechanica Sinica 39 (2) (2023) 122193

    2023

  30. [30]

    L. N. Smith, A disciplined approach to neural network hyper-parameters: Part 1–learning rate, batch size, momentum, and weight decay, arXiv preprint arXiv:1803.09820 (2018). 39 Appendix A. Numerical Implementation and Hyperparameters To ensure complete reproducibility of the TransportBench evaluation, we detail the ex- act hyperparameter configurations an...

    work page internal anchor Pith review Pith/arXiv arXiv 2018