A Graph Attention Neural Network for Subgrid Stress Modeling

Large Eddy Simulation (LES) subgrid stress modeling directly impacts the quality of the LES solution, and should be improved to help advance the accuracy of higher Reynolds number simulations of engineering interest. A novel graph neural network architecture is introduced that can scale to complex simulations, and is able to outperform existing neural network models and traditional models. A priori, the newly proposed graph neural network model is also able to extrapolate to various grid sizes not in the training set, and to non-cartesian meshes even though the training set only consists of cartesian meshes. A posteriori, the graph neural network architecture outperforms existing neural network and traditional subgrid stress models when evaluated on flows and grid resolutions in the training set, while avoiding catastrophic failure during extrapolation to grid resolutions coarser than the training set. Furthermore, the graph neural network is evaluated on flow over a cylinder, even though it is never trained on curvilinear meshes or a cylinder flow, and provides surface pressure measurements and coefficient of drag predictions that match experimental studies.