Skip to main content
NASA Advanced Supercomputing (NAS) Division World-class modeling, simulation, and supercomputing in support of the NASA mission

Discretization-Consistent LES Closure Strategies

AMS Seminar Series

Speaker: Andrea Beck, University of Stuttgart

Presentation

Seminar Slide Deck (PDF-8.4MB)

Abstract

The vast majority of LES formulation is based on the concept of grid-filtering. In this so-called implicitly filtered LES, the induced filter kernel and thus the closure terms are determined by the properties of the grid and the discretization operator, leading to additional unclosed computational subgrid terms that are generally unknown in a priori analysis. This interaction of the discretized PDE and the model terms contributes to the overall uncertainties of LES and the observed inconsistent closure model performance across schemes and codes.

In this talk, I will discuss how to find optimal, discretization-specific LES closure strategies. For this purpose, the task is formulated as a Markov decision process and solved by Reinforcement Learning (RL). This allows to adjust the model to the actual discretization as it also incorporates the interaction between the discretization and the model itself. Thus, this approach evaluates and optimizes the model performance in an a posteriori manner.

I will briefly discuss our RL strategy and methods employed as well as present our framework for a solver-in-the-loop on hybrid HPC architectures as well as its application to closure strategies for a discontinuous Galerkin (DG) spectral element scheme for canonical turbulence. Our results indicate that an RL-trained eddy viscosity model can adapt to the inhomogeneous dissipation properties of the baseline scheme and give excellent results compared to state of the art models. I will also explore a novel implicit modelling approach based on a Finite Volume / DG with RL-governed blending. Overall, the results demonstrate that the proposed RL optimization can provide discretization-consistent closures for state of the art high order schemes. While this is work in progress, the presented approach can help reduce uncertainty in LES.

Biography

Photo of Andrea Beck
Andrea Beck is a professor for Numerical Methods in Fluid Mechanics at the faculty of Aerospace Engineering and Geodesy and the Center for Simulation Science at the University of Stuttgart. She holds an M.Sc. in aerospace engineering from Georgia Tech and a Ph.D. in numerical methods from the University of Stuttgart. She is a co-developer and maintainer of the open source high order Discontinuous Galerkin framework FLEXI for multiscale, multiphase and Multiphysics flows. Her main research is numerical methods for compressible turbulent flow, HPC and the integration of data-driven methods into CFD.