APPLICATION OF MACHINE LEARNING METHODS TO CHANNEL FLOW MODELLING

Machine-learning methods to enhance an approximation for the Reynolds-stress anisotropy tensor are presented. The approach of tensor basis random forest is applied for this. The set of input features and tensors in the basis are discussed. Different ways to propagate the Reynoldsstress anisotropy tensor into the Reynolds-averaged Navier–Stokes equation solver are explored. It is demonstrated that the conventional expression for Reynolds-stress anisotropy based on the linear eddy-viscosity model is not able to reproduce a secondary flow in the square duct cross-section, whereas the machine-learning modifications can fix such a disadvantage and have potentials for further improvements of turbulence models.