Separated wake prediction of reference bluff bodies based on a tensor basis neural network-based turbulence model
Abstract
Conventional turbulence models based on the linear eddy-viscosity hypothesis cannot adequately represent turbulence anisotropy, which limits their predictive accuracy for three-dimensional separated flows around the Ahmed body. Within the Reynolds-averaged Navier–Stokes framework, this study develops a nonlinear Reynolds-stress anisotropy correction model for Ahmed-body wakes based on a tensor-basis neural network (TBNN), with a newly introduced canonical-flow training dataset and a wall-distance-based blending treatment. A posteriori evaluations of the Ahmed body flow demonstrate that the enriched training dataset improves the predictive accuracy of the TBNN model. At the same time, the wall-distance-based blending enhances numerical convergence and robustness. Compared with the baseline RANS model, the proposed model more accurately reproduces the major wake structures, including the recirculation region and C-pillar vortices, and yields flow-field predictions closer to those obtained from detached-eddy simulation (DES). Across rear-slant angles from 0° to 40°, the model exhibits favorable generalization for configurations with wake regimes similar to those in the training dataset. For the 5° configuration, whose wake regime lies outside the range covered by the training data, the improvement is comparatively limited. The proposed model also reduces computational time by 42% relative to DES, indicating its potential as an efficient turbulence modeling approach for automotive aerodynamic simulations.