Margaux Boxho presents work on Gaussian NN @ NeurIPS 2021

"Direct Numerical Simulations (DNS) and even Large Eddy Simulations (LES) remain computationally expensive to handle turbulent boundary layers at high Reynolds numbers. To further reduce the computational cost, wall-model LES (wmLES) models the turbulence in the near-wall region such that no more turbulent vortices are resolved there. Many developed wall models assume that the boundary layer is turbulent, at equilibrium, and attached. However, these assumptions are not valid for flows with high-pressure gradients or separation. To address the separation problem, we develop a data-driven wall model trained on the well-known case of the doubly periodic hill. In an a priori validation, encouraging results are obtained with the Gaussian Mixture Neural Network (GMNN), which predicts the distribution of wall shear stress instead of instantaneous values."

 

Workshop neurIPS, December 13, 2021, https://ml4physicalsciences.github.io/2021/