Data-driven Augmentation of RANS Turbulence Models for Improved Prediction of Separation in Wall-bounded Flows

摘要

The limited ability of conventional RANS (Reynolds-averaged Navier-Stokes) models to accurately predict the separation and reattachment in wall-bounded flows is improved by data-driven augmentation for a two-equation eddy-viscosity model. A field inversion and machine learning approach is applied on a k-ω RANS model to take into account for not only calibration errors but primarily structural errors and uncertainties by enhancing the functional form of the transport equation for the specific turbulent dissipation rate ω. The field inversion step identifies a spatial distribution of the model correction for several flow configurations by solving inverse problems regarding available, limited flow data from high-fidelity simulations and experiments. The determined spatial correction for the modeling discrepancies is reconstructed via machine learning by means of a neural network as a function of non-dimensional local flow features. The trained neural network implemented into a flow solver augments the baseline RANS model resulting in an improved prediction of the separation, reattachment and the overall flow solution.