Machine learning based Reynolds averaged simulation of backward-facing step flows at different Reynolds numbers

Abstract

To address the closure problem of Reynolds-averaged Navier–Stokes in numerical simulations of turbulence, the method of solving Reynolds-averaged Navier–Stokes equations based on artificial neural network is introduced in this paper. We establish the nonlinear mapping relationship between the average flow field and the steady-state eddy viscosity field. The machine learning (ML) surrogate model for the shear stress transport turbulence model is constructed. The solution process of replacing the original turbulence model equations with the predicted field variables is realized by coupling the ML algorithm with the CFD solver. The classical backward facing step problem is selected in our study to verify the simulation accuracy of the surrogate model. The comparative analysis is carried out on the six backward facing step flows simulations at different Reynolds numbers. The results of simulations show that the testing flows with the Reynolds numbers closest training datasets Reynolds numbers can obtain the best simulation accuracy. Then for the Reynolds number that is lower than the training datasets, the simulation accuracy will decrease as the Reynolds number decreases. On the contrary, the simulation accuracy of the test flow will increase as the Reynolds number increases. These results indicate the feasibility of the ML surrogate model to simulate at higher Reynolds number. It shows the great potential of applying ML algorithms to Reynolds-averaged Navier–Stokes simulation (RANS) turbulence model and also provides a new idea for industrial simulations of turbulent flows.