Machine learning building-block-flow wall model for large-eddy simulation

A wall model for large-eddy simulation (LES) is proposed by devising the flow as a combination of building blocks. The core assumption that finite set simple canonical flows contains essential physics to predict wall-shear stress in more complex scenarios. constructed zero/favourable/adverse mean pressure gradient turbulence, separation, statistically unsteady turbulence with three-dimensionality, and laminar flow. approach implemented using two types artificial neural networks: classifier, which identifies contribution each block flow, predictor, estimates via building-block flows. training data are directly obtained from wall-modelled LES (WMLES) optimised reproduce correct quantities. This guarantees consistency numerical discretisation gridding strategy solver. output accompanied confidence score prediction aids detection regions where underperforms. validated (e.g. laminar/turbulent boundary layers, turbulent channels, Poiseuille-Couette pipe) realistic aircraft configurations: NASA Common Research Model High-lift Juncture Flow experiment. It shown building-block-flow outperforms (or matches) predictions an equilibrium model. also concluded further improvements WMLES should incorporate advances subgrid-scale modelling minimise error propagation