Finite-volume residual formulations for Reynolds-averaged Navier–Stokes-constrained physics-informed neural network reconstruction of a hydrofoil tip-clearance flow from two-dimensional three-component particle image velocimetry
Physics of Fluids · 01.11.2026 · 38 (11)
DOI: 10.1063/5.0359797 ↗Kurz gefasst
Fachartikel in der Zeitschrift Physics of Fluids, online erschienen am 06.11.2026. Verfasst von Zhiyuan Geng und 5 weiteren Autor:innen. Die Arbeit nennt 36 Literaturangaben.
Abstract (Originalsprache)
Particle image velocimetry (PIV) provides spatially resolved velocity measurements, whereas pressure remains unobserved. Physics-informed neural networks (PINNs) can infer hidden quantities by coupling observations with governing equations, but the effects of residual discretization and internal-face flux organization remain unclear. In this paper, a hydrofoil tip-clearance flow is reconstructed from single-plane two-dimensional three-component PIV using a Reynolds-averaged Navier–Stokes-constrained PINN framework. Pointwise strong-form, independent local finite-volume method (IL-FVM), and shared-face (SF)-FVM residuals are compared. Finite-volume integration is applied to in-plane fluxes, with temporal terms and out-of-plane derivatives evaluated by automatic differentiation. At α = 8° and t = 5 mm, all models yield time-averaged velocity relative L2 differences of 4.0%–4.3% under a common training budget. IL-q1 matches the strong-form baseline in velocity accuracy while reducing training time from 9.7 to 5.1 h. For hidden pressure, IL-q2 agrees most closely with the large-eddy simulation reference, with a mean squared error of 10.0 × 10−3 and a relative L2 difference of 67.9% for the zero-mean time-averaged pressure coefficient. An independent control-volume test shows that SF-q2 reduces continuity and momentum residuals by ∼40% and 32%, respectively, relative to the strong-form baseline. However, improved local physics consistency does not improve velocity or pressure reconstruction and incurs greater computational cost. An additional test at α = 4° and the same tip clearance yields similar velocity errors for Vanilla and IL-q2, with ∼34% less training time for IL-q2. Thus, stronger SF discrete conservation improves local physics consistency without necessarily improving flow reconstruction.
Zitieren
APA
Geng, Z., Deng, R., Liu, H., Yu, X., Persoons, T., & Zhao, W. (2026). Finite-volume residual formulations for Reynolds-averaged Navier–Stokes-constrained physics-informed neural network reconstruction of a hydrofoil tip-clearance flow from two-dimensional three-component particle image velocimetry. Physics of Fluids, 38(11). https://doi.org/10.1063/5.0359797
Vancouver
Geng Z, Deng R, Liu H, Yu X, Persoons T, Zhao W. Finite-volume residual formulations for Reynolds-averaged Navier–Stokes-constrained physics-informed neural network reconstruction of a hydrofoil tip-clearance flow from two-dimensional three-component particle image velocimetry. Physics of Fluids. 2026;38(11). doi:10.1063/5.0359797
BibTeX
@article{Geng2026,
author = {Geng, Zhiyuan and Deng, Rui and Liu, Hailong and Yu, Xiang and Persoons, Tim and Zhao, Wenguang},
title = {Finite-volume residual formulations for Reynolds-averaged Navier–Stokes-constrained physics-informed neural network reconstruction of a hydrofoil tip-clearance flow from two-dimensional three-component particle image velocimetry},
journal = {Physics of Fluids},
year = {2026},
volume = {38},
number = {11},
doi = {10.1063/5.0359797}
}
EUVIMED