Direct deterministic neural predictors outperformed two generative pipelines, flow matching (FM) and diffusion, on the study's additional test of cylinder-flow fields. Across the complete additional cylinder evaluation set, every MLP variant beat both generators on E6, the study's descriptive equal-weight summary of six field errors.
The size of the gap was clear in the reported averages. Across five training seeds, FM's mean E6 was 0.0365 plus or minus 0.0017 and diffusion's was 0.0531 plus or minus 0.0037; the MLP variants ranged from 0.0288 to 0.0311. The plus-or-minus figures were population standard deviations across the five training seeds.
The study also reported a separate interpolation result. For out-of-sample conditions, the frozen pipelines reached cavity kinetic relative L1 errors at the 10^-5 level. On the cylinder cases, area-weighted RMSE was 0.038 for density, 0.041 for temperature and below 0.01 for velocity. The comparison used frozen models, with no refitting of the predictor, chart, decoder or normalization statistics on the evaluated cases.
A modular test across two flow problems
The system separates how a field is represented from how a condition-specific prediction is generated. A neural-field auto-decoder assigns each condition a lookup code without receiving physical-condition labels, while a train-only latent transformation creates the chart used by the conditional generators.
Neither generative pipeline contains a deterministic condition-to-latent predictor. The direct regressor was kept as a controlled baseline; it did not initialize, supervise or correct the generated latent samples.
The cavity benchmark and its generators used 16 listed training conditions, with three cases reserved for validation and four additional cases used for the main comparisons. The cylinder study used a 64-case tensor grid divided into 48 training, eight validation and eight internal-test conditions.
For out-of-sample comparisons, the direct predictors and generative pipelines were evaluated without refitting the predictor, chart, decoder or normalization statistics on the cases being assessed. That kept the reported comparison on a frozen footing.
Leaving the training range exposed a weakness
The lower-side cavity extrapolation was much less accurate than interpolation. At the test point marked Kn = 0.03, frozen FM had macroscopic and kinetic errors of 5.01 x 10^-3 and 5.26 x 10^-3. Frozen diffusion was lower at 3.10 x 10^-3 and 3.13 x 10^-3. Structured adapters changed these values by less than 10^-6, and the analysis placed the main limitation at the extrapolated chart endpoint.
At the upper-side point, Kn = 1.00, diffusion was more accurate than FM. Its macroscopic and kinetic errors were 7.19 x 10^-5 and 8.26 x 10^-5, compared with 2.91 x 10^-4 and 9.23 x 10^-4 for FM. Diffusion adaptation produced a kinetic error of 7.81 x 10^-5. FM adaptation changed the BGK diagnostic by 12.20 percent without a material change in kinetic error.
Physics checks improved selectively
The physics adapter showed a sharper effect in a paired cavity audit. On held-out development cases, normalized FM adaptation moved the kinetic error score, Ekin, from 5.659 x 10^-5 to 5.652 x 10^-5, while the BGK diagnostic fell 28.65 percent and a composite-physics diagnostic fell 22.06 percent relative to frozen FM. Independent-probe conservation and moment ratios were both 1.0574.
The cylinder result came with an important qualification. The analytic wall map gave exact no-penetration, but the reported inlet and global-balance improvements came from the map deployed jointly with the learned FM adapter. In that combined configuration, inlet violation was 0.277 of the frozen value and the global mass-balance ratio was 0.963, with negligible change in field error.
Those diagnostic changes did not translate into a better aggregate cylinder field score. On additional conditions, constrained FM changed E6 from 0.03655 to 0.03670, while constrained diffusion changed it from 0.04748 to 0.05008. Neither adapted model improved aggregate field accuracy.
A result bounded by the data
The result comes with a basic uncertainty warning. Because the study had one steady reference for each condition, the spread of generated endpoints was treated as transport sensitivity rather than calibrated physical uncertainty.
Model selection also was not fully fixed in advance. Architecture candidates were developed iteratively on development data rather than drawn from an exhaustively preregistered search space.
The document is an arXiv preprint identified in its front matter as arXiv:2608.25454v1 [physics.flu-dyn], dated 26 August 2026. The work reports support from the Science Foundation for Young Scientists of the State Key Laboratory of High Temperature Gas Dynamics, Chinese Academy of Sciences programs, the National Natural Science Foundation of China and the Beijing Natural Science Foundation.
Paper data and sources
Original title: Physics-Guided Generative Surrogates for Parametric Rarefied Flows with Neural-Field Auto-Decoders: A Pipeline-Level Study of Flow Matching and Diffusion
Authors: Yiming Qi, Guan Zhang, Xu Wang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text