Abstract
Abstract
Conventional Reynolds–Averaged Navier–Stokes (RANS) turbulence models often exhibit systematic errors in predicting the dynamic wall-shear response of pulsatile turbulent flows, whereas Direct Numerical Simulation (DNS) remains prohibitively expensive for routine engineering applications. This work presents a sparse DNS-informed parametric Dynamic Mode Decomposition (pDMD) framework that enhances the predictive capability of reduced-order models through parametric discrepancy learning without modifying the underlying turbulence model. A reduced-order representation of phase-resolved RANS simulations is constructed using Proper Orthogonal Decomposition and complex Dynamic Mode Decomposition. Sparse DNS observations are then used to learn a complex wall-shear transfer-function discrepancy, which is interpolated over the parameter space and applied to previously unseen operating conditions to simultaneously correct the amplitude and phase of the dominant oscillatory response. Using only four DNS calibration cases, the proposed framework reduces wall-shear amplitude errors from approximately 10–15% to agreement with the available DNS calibration data while maintaining smooth parametric predictions across the investigated operating range. Monte Carlo uncertainty propagation further demonstrates that a 3% uncertainty in the governing pulsation parameter results in narrow wall-shear prediction intervals, while sensitivity analysis shows that the turbulence-production response is most sensitive in the intermediate attenuation regime rather than in the strongly amplified regime. The proposed methodology provides a general framework for enhancing reduced-order CFD models using sparse high-fidelity information through discrepancy learning, offering an efficient alternative to direct turbulence-model correction.