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, Tianbai Xiao · 2026
Read TransportBench: A Comprehensive Benchmark for Non-Equilibrium Flow Transport by Xu Wang, Minghao Li, Qizhen Hong, Yang Liu, Chen-an Zhang, Shuai Zhang, Wenhao Li, Yonghao Zhang, Tianbai Xiao (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-02997.txt.
Introduction
Wang and co-authors present TransportBench, a dataset and benchmark for non-equilibrium flow transport. Scientific machine learning is changing fluid mechanics research. Existing benchmarks, though, focus on continuum fluids. They offer little support for non-equilibrium phenomena such as rarefied or hypersonic flows.
TransportBench spans a broad physical spectrum. It covers continuum and rarefied regimes, low-speed and hypersonic flows, inert and reactive gases, and translational and internal-energy non-equilibrium. The team benchmarks representative neural architectures under unified protocols. They probe robustness to shocks, multi-scale effects, and generalization across geometry and parameters. No single architecture won across all tasks. Different design biases helped with smooth fields, shock discontinuities, or high-order statistics. The dataset and code are released under the MIT license. The authors offer the benchmark as a testbed for learning flows beyond Navier-Stokes hydrodynamics.
Catalog record
| CREATOR | Xu Wang, Minghao Li, Qizhen Hong, Yang Liu, Chen-an Zhang, Shuai Zhang, Wenhao Li, Yonghao Zhang, Tianbai Xiao |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 89 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.02997
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
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, Tianbai Xiao 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 1 1 1 https://huggingface.co/datasets/CFDML/TransportBench and implementation codes 2 2 2 https://github.com/CFDML/TransportBench are available under the MIT license. Table 1 : Nomenclature. \mathbf{x} space variable (x,y,z) \mathbf{v} particle velocity variable (u,v,w) t time variable f particle distribution function s species index l internal state index \mathcal{Q} collision operator in kinetic equation \rho macroscopic density \mathbf{u} macroscopic velocity E macroscopic energy density \mathbf{U} conservative variables of flow field \mathbf{P} pressure tensor \mathbf{q} heat flux p macroscopic pressure \varphi mass fraction h enthalpy S source term N_{s} total number of chemical species V_{s} number of vibrational states associated with the s-th species D_{s} diffusion coefficient of the s-th species \eta thermal conductivity associated with translational-rotational energy \mu shear viscosity \mathcal{A} operating conditions \mathcal{U} solution field \mathcal{G} machine learning model \mathcal{L} loss function \mathbf{b} class-shape transformation (CST) coefficient \mathscr{U} uniform distribution r random variable sampled from \mathscr{U} Ma Mach number Kn Knudsen number Re Reynolds number \omega viscosity index d_{\mathrm{ref}} diameter at reference temperature \mathcal{T} non-linear logarithmic transformation \mathbf{M} binary geometry mask a scalar field N number of spatial grid points MSE mean squared error MAE mean absolute error 1 Introduction Accurate simulation of non-equilibrium flow transport has become increasingly important as fluid systems are pushed toward more extreme spatial and thermodynamic regimes. In hypersonic flight and atmospheric re-entry 18 , gases may experience strong compression, shock heating, rarefaction, and internal-energy excitation; in micro/nano-electro-mechanical systems (MEMS/NEMS) 5 , the characteristic length can become comparable to the molecular mean free path. These conditions give rise to pronounced non-equilibrium effects, under which the continuum assumption may break down, and the classical Navier-Stokes equations become insufficient for accurately describing the underlying transport phenomena. A more faithful description requires statistical mechanics and gas-kinetic formulations such as the Boltzmann equation 4 , or direct-modeling particle methods such as the direct simulation Monte Carlo (DSMC) method 1 . Although these approaches provide reliable descriptions of non-equilibrium fluid dynamics, their high computational cost, slow statistical convergence, and inherent stochastic noise make them difficult to use for rapid design, uncertainty analysis, and repeated parametric studies. This creates a strong demand for efficient surrogate modeling and simulation acceleration methods tailored to non-equilibrium flow transport. Scientific machine learning is emerging as a promising paradigm for addressing these computational challenges by combining data-driven approximation with physical modeling 3 , 9 . However, the systematic evaluation of such methods remains largely constrained by benchmarks designed for continuum and near-equilibrium flow problems. Widely used datasets, such as PDEBench 19 and FlowBench 20 , primarily consider the Navier-Stokes-level dynamics close to thermodynamic equilibrium, in which the flow fields are often smooth, the output variables are mostly low-order macroscopic quantities, and the computational domains are regular or only mildly complex. They do not fully reflect the defining challenges of non-equilibrium transport, including rarefaction effects, Knudsen layers, high-order moment quantities, strong shock discontinuities, and multi-scale kinetic-to-continuum behavior. Consequently, strong performance on existing continuum-flow benchmarks does not necessarily indicate that a model can robustly predict rarefied or hypersonic non-equilibrium flows. Beyond the issue of physical coverage, existing evaluation methods also leave room for more systematic comparison in non-equilibrium flow learning. In many studies, neural architectures are assessed under different parameter budgets, grid resolutions, preprocessing strategies, and training protocols. Such differences can make it difficult to distinguish the effect of architectural inductive bias from that of model capacity or implementation details 22 . Moreover, different neural architectures encode distinct assumptions about locality, global coupling, mesh structure, and coordinate representation. These assumptions can strongly affect their ability to resolve sharp gradients, high-order non-equilibrium statistics, and localized flow structures. In particular, coordinate-based and attention-based models may exhibit spectral bias and tend to learn smoother, low-frequency mappings, while spectral or convolutional models may respond differently to high-frequency feature injection 15 . Therefore, a benchmark for non-equilibrium flow transport should not only report prediction errors, but also provide controlled settings for diagnosing how different architectures behave under specific physical challenges 25 , 26 . To address these issues, we introduce TransportBench, a benchmark framework with two closely connected objectives. First, it provides a high-fidelity dataset for non-equilibrium flow transport, covering rarefied and continuum regimes, low-speed and hypersonic flows, inert and reacting gases, and both translational and internal-energy non-equilibrium. The dataset is organized into four representative flow scenarios, including airfoil, cylinder, cavity, and double-cone configurations, which are designed to examine geometric generalization, physical-parameter generalization, high-order moment prediction, and shock-dominated hypersonic transport. Second, TransportBench establishes a standardized model benchmark for evaluating representative scientific machine learning architectures under controlled settings. By comparing operator, spectral, convolutional, attention-based, geometric, and latent models with unified preprocessing, training protocols, parameter budgets, and evaluation metrics, the benchmark enables a more systematic assessment of how different architectural inductive biases perform across distinct flow regimes. Through this dual design, TransportBench aims to serve not only as a reusable dataset for non-equilibrium fluid dynamics, but also as a diagnostic testbed for developing and evaluating learning-based methods for flow transport beyond the classical Navier-Stokes level. 2 Related Work 2.1 Neural architectures for learning fluid dynamics Scientific machine learning has introduced a broad range of neural architectures for approximating fluid dynamics across different spatial representations and physical scenarios. These models differ not only in the network design, but also in the inductive biases they impose on the learned solution fields. Understanding these biases is essential for evaluating their suitability for non-equilibrium transport problems, where the target fields may contain smooth regions, sharp discontinuities, multi-scale structures, and high-order kinetic statistics. Convolutional and latent-representation models , such as convolutional neural networks, U-Net 16 , and autoencoders 29 , have been widely used for learning solution fields represented on structured grids. Convolutional models exploit local receptive fields, hierarchical feature extraction, and translational equivariance, making them