Generalization enhancement of artificial neural network for turbulence closure by feature selection

Abstract Feature selection targets for selecting relevant and useful features, is a vital challenge in turbulence modeling by machine learning methods. In this paper, new posterior feature method based on validation dataset proposed, which an efficient universal complex systems including turbulence. Different from the priori importance ranking of filter exhaustive search subset wrapper method, proposed ranks features according to model performance dataset, generates subsets order importance. Using black-box built artificial neural network (ANN) reproduce behavior Spalart-Allmaras (S-A) high Reynolds number ( Re ) airfoil flows aeronautical engineering. The results show that compared with without selection, generalization ability after significantly improved. To some extent, it also demonstrated although can be reflected parameters during training process, still very necessary.