Presentation
Exploring Differentiable Physics Based Sub-Grid Models for Large Eddy Simulation of Turbulent Flows
Presenter
DescriptionThe well known (text book) definition of large eddy simulation (LES) is one where the large energy containing scales are evolved explicitly while the effects of the small scales of motion, that are lost due to the filtering operation, are modeled as the sub-grid stress (SGS) term in the LES equations. This suggests that the extent to which the SGS term would need to compensate for the lost scales of motion would depend on the cutoff wavenumber. In the practice of LES, however, it is assumed that the filtering operation is in the inertial range, and most SGS models, which are dissipative in nature, use an approximate measure of the filter width to control the effective turbulent viscosity. A recent paper by Shankar et al. [1] applied the differentiable physics paradigm to the 1D Burgers equation to formulate data-driven SGS models for various Reynolds numbers. The purpose of this talk is to explore the use the same neural ODE framework to study the effect of the cutoff wavenumber on the effectiveness of such data-driven SGS models, and its implications for LES of more general turbulent flows.
[1] Varun Shankar et al. 2023 Mach. Learn.: Sci. Technol. 4 015017, DOI 10.1088/2632-2153/acb19c
[1] Varun Shankar et al. 2023 Mach. Learn.: Sci. Technol. 4 015017, DOI 10.1088/2632-2153/acb19c
TimeMonday, June 2618:00 - 18:30 CEST
LocationDischma
Event Type
Minisymposium
Computer Science, Machine Learning, and Applied Mathematics