• Citer

Quantifying Parametric Uncertainty in Turbulence Model for the Numerical Simulation of Atmospheric Boundary Layer Flows Using Dimensionality Reduction and Bayesian Regression

par Li, Haoyan ;Amaduzzi, Ruggero ;Mosca, Gabriele M.G.;Raghunathan Srikumar, Sampath Kumar;Gambale, Alessandro ;Parente, Alessandro
Référence Journal of wind engineering and industrial aerodynamics, 272, page (106397)
Publication Publié, 2026-02-26

Article révisé par les pairs

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

The Reynolds-Averaged Navier–Stokes (RANS) equations are widely used to study atmospheric boundary layer (ABL) flows due to their low computational cost; however, the predictive accuracy is limited by uncertainties in empirical turbulence parameters. This study applies sensitivity analysis within an uncertainty quantification framework to identify the most influential parameters of the k–ω Shear Stress Transport (SST) model for ABL flows. The selected parameters were calibrated using the least-squares method, improving the predictive performance of the RANS. To reduce computational cost, a surrogate model combining Proper Orthogonal Decomposition (POD) with Gaussian Process Regression (GPR) was developed. Latin Hypercube Sampling generated parameter combinations for CFD simulations, POD extracted dominant modes, and GPR mapped POD coefficients to turbulence model parameters. The POD/GPR surrogate reproduced flow fields with high accuracy, enabling efficient sensitivity analysis and calibration. The benchmark case of ABL flow around an isolated building was analysed across 4 critical regions: upstream recirculation, rooftop shear, side shear, and near-wake recirculation. Sensitivity analysis of 6656 parameter combinations revealed strong dependence of near-wake and ground-level upstream flows on turbulence model parameters. 3 dominant contributors were identified from 12 independent parameters, including the von-Karman constant, the eddy viscosity coefficient, and a parameter controlling building-influenced area. Despite calibration, persistent near-wake discrepancies indicated irreducible model-form error. Calibrated parameters reduced discrepancies with experiments by ∼30% for stream-wise velocity and ∼40% for turbulent kinetic energy (TKE) in this isolated-building case, and by ∼40% and ∼27% in another building case, demonstrating improved predictive capability across different configurations.