Advances in deep learning have sparked interest in using this emerging technology in the domain of physics simulations. Models derived from computer vision architectures, as well as physics simulation specific architectures like PINNs and Neural Operators have shown promising results. Recent works have demonstrated that generative models can be used to model the distribution of states of turbulent flows. In this work I will model the distribution of instantaneous flow fields from a wall-modelled LES simulation on curvilinear grids. This grid type allows for representing curved surfaces and enables locally varying grid resolutions, essential capabilities for accurate simulation of turbulent flow around real-world objects.
In the first part of this work, the GPU-based differentiable CFD solver PICT will be extended with functionality for efficiently regridding data on curvilinear grids.
The second part will describe an effort to model turbulent flow through diffusers similar to the Buice-Eaton diffuser. For this purpose, a vision transformer derived model will be trained with a flow matching objective. Generalization to diffuser designs not present in the training data will be evaluated. New data augmentation techniques made possible by the curvilinear grid formulation and the regridding functionality from the first part will be used in an attempt to make the model more robust towards different grid geometries.