Application of machine learning assisted turbulence modeling in flow separation prediction

Data-driven turbulence modeling has been considered as an effective method for improving the prediction accuracy of the Reynolds-averaged Navier-Stokes equations. By using machine learning algorithms such as the artificial neural network, features can be automatically extracted from high-fidelity data, and accurate prediction models from the mean flow characteristics to the Reynolds stress can be established. In this study, focused on the ice-accretion airfoil with the typical complex flow separation phenomenon under high Reynolds numbers, efforts are made in two aspects, i.e., the input and output feature selection and the data distribution characteristics of flow around an airfoil, to improve the smoothness and accuracy of the machine learning assisted prediction results. An input feature selection criterion based on the Reynolds stress tensor analysis and the flow characteristics identification is proposed, and a zonal modeling method together with a hybrid approach of the benchmark model and the machine learning assisted prediction model is also proposed. Results obtained from the improved method show that the Reynolds stress on both the training dataset and the prediction dataset can be accurately predicted, and the flow separation and pressure distribution on the airfoil agree well with the real situation.