A robust data-driven subgrid scale (SGS) stress model using a graph neural network (GNN) architecture with spatial convolution is developed and trained over a comprehensive direct numerical simulation (DNS) dataset of turbulent incompressible flows. The DNS data is filtered at two different scales to represent varying Large Eddy Simulation (LES) resolutions. The grid-scale model inputs are the Leonard stress tensor and velocity fluctuations, calculated from test-filtering at multiple scales to capture a wider range of the energy spectrum. Galilean equivariance is ensured by the input feature definition. Rotational equivariance is ensured by data augmentation. The model is fully generalizable to nonuniform unstructured computational grids in both the formulation of its input features and convolution by GNN message passing layers. A nondimensionalization and scaling scheme restricts the dynamic range of the inputs, ensuring generalizability over a wide range of flow fields and turbulent conditions. The model is tested for a multitude of incompressible flows at Reynolds numbers unseen during training, as well as compressible reactive flows. The model exhibits very high prediction accuracy for unbounded flows, regardless of compressibility. The accuracy is reduced for wall-bounded flows due to the complexity of near-wall dynamics, but remains satisfactory.
 

Turbulence,Subgrid scale modeling,Machine learning,Graph neural network