Data-driven sub-grid model development for large eddy simulations of turbulence

Romit Maulik
Lecture

Researchers continue to search for modeling strategies that improve the representation of high-wavenumber content in practical computational fluid dynamics applications. In this research, we utilize physics-informed machine learning to reconstruct the effect of unresolved frequencies on grid-resolved variables obtained through LES. The proposed methodologies require the synthesis of labeled data from direct numerical simulations of our target phenomenon as well as the development of stability preserving modifications instead of a direct deployment of predictions. These stability preserving techniques may be through prediction modulation – where learning outputs are statistical truncated. They may also be through the utilization of classifiers where a categorical cross-entropy loss flags for the most appropriate model at a grid-point. In this thesis, we outline several investigations utilizing the aforementioned philosophies and come to the conclusion that sub-grid turbulence models built through the utilization of machine learning are capable of recovering viable statistical trends in stabilized a-posteriori deployments for Kraichnan and Kolmogorov turbulence.