Neural network-based eddy viscosity prediction for bluff body vehicle wake flow at high Reynolds number

Abstract

Improvements in RANS model accuracy and computational efficiency can substantially impact automotive aerodynamic optimization. In this study, a deep neural network–based turbulence model is developed to predict the flow around bluff body vehicles. Using an Ahmed body dataset generated by the SST $k-\omega$ model, a mapping between flow-field physical quantities and eddy viscosity is constructed. The input features are extended to capture three-dimensional flow characteristics, with a random forest algorithm employed for feature selection. Analysis reveals that the orthogonality between velocity and its gradient, previously less significant in two-dimensional flows, becomes critical for predicting three-dimensional Ahmed body turbulence. The proposed model fully replaces the conventional SST $k-\omega$ model and is coupled with the CFD solver. Results show that it accurately reproduces the velocity and pressure fields, closely matching baseline RANS predictions. The predicted drag coefficient deviates by less than 6% from experimental measurements. For off-training conditions at the different yaw angle, the model exhibits slight underprediction in the wake core region and minor discrepancies in capturing upper vortices. Moreover, the model achieves a 30% reduction in computational time, demonstrating its potential for efficient, high-fidelity aerodynamic simulations in industrial applications.