Numerical simulation of a highly bent intake duct using Machine learning algorithms to close the rans equations

This study intends to improve the numerical flow prediction of highly bent intake ducts by the specific use of machinelearning algorithms for the closure of the Reynolds-Averaged-Navier-Stokes (RANS) equations. A computational fluid dynamics (CFD)-driven machine learning approach utilizing an adapted Gene Expression Programming (GEP) algorithm is applied to approximate and iteratively improve an additional anisotropy stress tensor for the definition of Reynolds stresses. The determined turbulence model candidates are evaluated based on experimentally obtained data of the double s-shaped Military Engine Intake Research Duct (MEIRD). The newly generated non-linear stress models deliver a very promising improvement of the agreement with experimental data of MEIRD. This applies especially to the prediction of a large-scale flow separation and the corresponding position of the reattachment zone, which is significantly improved compared to the baseline k-ω SST model, closely matching the experimentally obtained data. Applying the model to non-trained operating points is shown to achieve similar improvements. Conclusively, the performance of the model is discussed in terms of remaining weaknesses and suggestions for further developments are presented.