Hybrid LES/RANS for flows including separation: A new wall function using Machine Learning based on binary search trees (2410.17767v2)

Abstract: Machine Learning (ML) is used for developing wall functions for Improved Delayed Detached Eddy Simulations (IDDES). The ML model is based on KDtree which essentially is a fast look-up table. It searches the nearest target datapoint(s) for which y+ and U+ are closest to the CFD y+ and U+ cells. The target y+ value gives the friction velocity which is used for setting the wall shear stress for the wall-parallel velocity and for fixing k and epsilon at the wall-adjacent cells. Two target databases are created from time-averaged data of low-Reynolds number (i.e. wall-resolved) IDDES: diffuser flow with opening angle alpha=15 degrees and hump flow. The new ML wall function is used to predict five test cases: diffuser flow with opening angles alpha=15 degrees and alpha=10 degrees the hump flow, channel flow at $Re_tau=16 000 and flat-plate boundary layer. A novel grid strategy is used. The wall-adjacent cells are large (20 < y+ < 60 in attached boundary layers). But further away from the wall, the wall-normal cell distribution is identical to that of a low-Re number grid. This new grid is found to improve the predictions compared to a standard wall-function grid. It is found that the number of cells for a wall-resolved IDDES grid (grid stretching 15%) is a factor of 0.2ln(Re_tau) larger than that of a standard wall-functions mesh (constant wall-normal grid cells). The new ML wall function is found to perform well compared to the low-Re IDDES and better than the Reichardt's wall function