Physics-constrained uncertainty quantification of Reynolds-averaged Navier–Stokes model in a centrifugal compressor with adaptive non-uniform perturbations

The performance prediction of centrifugal compressors using Reynolds-averaged Navier–Stokes models often show significant deviations from experimental results, mainly due to structural uncertainties arising from the linear eddy viscosity assumption. This study develops a physics-constrained framework for turbulence model uncertainty quantification (UQ) in the numerical simulation of a shrouded centrifugal compressor. The framework introduces an adaptive non-uniform perturbation strategy within the barycentric map, enabling simultaneous modification of eigenvalues and eigenvectors of the Reynolds stress anisotropy tensor. Compared with conventional uniform perturbation methods, the adaptive approach not only improves numerical convergence under large perturbations of turbulent kinetic energy production but also ensures more consistent performance predictions across the entire operating range. The results indicate that, after introducing eigenvector perturbations, the predicted uncertainty interval covers most of the experimental data at three different rotational speeds, despite obvious discrepancies between baseline shear stress transport model (k–ω SST) predictions and measurements. Furthermore, the investigation of flow structure and entropy generation is conducted to elucidate how different perturbations influence internal flow characteristics and loss mechanisms. Overall, this study presents the first application of a physics-constrained UQ framework with adaptive non-uniform perturbations to centrifugal compressors, establishing reliable uncertainty bounds for performance prediction and clarifying the role of anisotropic turbulence limiting states in flow separation mechanisms.