Neural network-augmented SED-SL modeling of turbulent flows over airfoils

A novel modeling paradigm, named as SED-SL-RBF, enhances the structural ensemble dynamics-stress length (SED-SL) model of wall-bounded turbulence using limited aerodynamic data from NACA airfoils. It constructs a multi-layer structure (MLS) of the turbulent boundary layer (BL) over the airfoils and uses machine learning to reconstruct model parameters from experimental data. This approach has been applied to turbulent flows over nine distinct NACA airfoil types, with a broad spectrum of Reynolds numbers and angles of attack. The model parameters, l ^∞₀ and y ^∞_buf , are reconstructed using a radial basis function (RBF) neural network and applied to an SED-SL computational fluid dynamics (CFD) solver. This results in improved predictions of lift and drag coefficients for geometries and flow conditions previously calculated using the Menter shear stress transport (SST) turbulence model. The accuracy of the predictive lift coefficient C_L exceeded 95%, while the error in the predictive drag coefficient C_D was less than 6 counts. The neural network-augmented SED-SL model also demonstrated exceptional predictive accuracy for the pressure field. The MLS parameters for NACA 2421 exhibit similarities with angle of attack (AOA), which can be treated as functions of the Reynolds number. These findings suggested that the MLS parameters for NACA 2421 are independent of the AOA prior to stall. This similarity behavior provides a promising approach to model airfoil flows under various physical conditions. The broader vision is to integrate data to reveal innate model discrepancies in terms of model parameters, thereby extending the applicability of the SED-SL-RBF model to a wider range of flow scenarios.