Generalizable data-driven turbulence closure modeling on unstructured grids with differentiable physics

The recent study on data-driven turbulence closure modeling marks a significant advancement in computational fluid dynamics (CFD) by embedding deep learning models within finite element solvers for the Navier-Stokes equations. This innovative approach utilizes differentiable physical simulators to facilitate end-to-end training of machine learning models, merging the strengths of physics-based simulations with the adaptability of deep learning. The research initially validates this framework through a two-dimensional backward-facing step flow, establishing a proof of concept that leads to stable and meaningful turbulence closures. Expanding on this, the study successfully applies the method to a three-dimensional turbulent flow scenario, where the graph neural network (GNN) based closure not only minimizes prediction errors but also accurately recovers essential turbulence statistics. This work underscores the potential of combining machine learning with traditional physics-based model…