Practical framework for data-driven RANS modeling with data augmentation

Inspired by the iterative procedure of computing mean fields with known Reynolds stresses (Guo et al., Theor Appl Mech Lett, 2021), we proposed a way to achieve data augmentation by utilizing the intermediate mean fields after proper selections. We also proposed modifications to the Tensor Basis Neural Network (Ling et al., J Fluid Mech, 2016) model. With the modification of the learning targets and the inclusions of wall distance and logarithm of normalized eddy viscosity in the model inputs, the modified version of the model with augmented training datasets shows better performance on Reynolds stress predictions for two dimensional incompressible flow over periodic hills under different geometries. Furthermore, better propagated mean velocity fields can be achieved, showing better agreements with the direct numerical simulations (DNS) results.

Automated summary

Inspired by the iterative procedure of computing mean fields with known Reynolds stresses, a method of achieving data augmentation was proposed by utilizing intermediate mean fields after proper selections. Modifications to the Tensor Basis Neural Network model were made, resulting in better performance on Reynolds stress predictions in two-dimensional incompressible flow. The modified model with augmented training datasets also showed improved agreements with direct numerical simulations for mean velocity fields.