Orateur
Description
Turbulence is encountered in most industrial flows, yet simulating turbulent flows remains computationally challenging due to the wide range of spatial and temporal scales involved. This breadth of scales induces very high computational costs for Direct Numerical Simulations (DNS), driving the development of more affordable approaches such as Large-Eddy Simulations (LES). However, classical LES closure models lack the flexibility to accurately capture subgrid-scale physics across diverse flow regimes.
In recent years, machine learning (ML) has emerged as a promising avenue for data-driven closure models trained on high-fidelity simulation data. In this work, we propose a graph neural network-based LES closure model, where the graph representation naturally handles anisotropic and unstructured user-defined meshes. The model is designed to be equivariant to rotations and reflections, ensuring physically consistent predictions regardless of mesh orientation while reducing the amount of training data required. Special attention is given to non-dimensionalization and normalization of input features, which prove critical to the model's stability and generalizability.