ISSN Online: 2379-1748
ISBN Flash Drive: 978-1-56700-518-9
5th Thermal and Fluids Engineering Conference (TFEC)
DEVELOPMENT OF A DEEP-LEARNING MODEL TO IMPROVE LARGE EDDY SIMULATIONS OF TURBULENT FLOWS
Abstract
Turbulence is the most dominant characteristic of a turbulent flow. Therefore, successful modeling of turbulence can significantly improve the results of numerical simulation. Large Eddy Simulation (LES) computation of turbulent flows has been achieved a great attention recently since post-processing of LES results yields information of both mean flow and statistics of resolved fluctuations which is unique to LES and hence can model flows where persistent large-scale vortices results in flow development, e.g. flow behind bluff bodies, tumble swirl in engine combustion chambers and prediction of noise from high-speed flows (Versteeg 2009). However, this requires to address some issues to control the error sources and generate robust LES methodology for industrial applications.
In this study, a deep-learning approach is used to augment existing LES models using the relevant flow features. To this end, Random Forest Regression is developed to map relevant statistical flow-features within the LES solution to errors in calculated statistics such as sub-grid scale stresses. In this context, the exact solution is given by Direct Numerical Simulation (DNS) data. The capability of the proposed framework is examined by posteriori tests.
In this study, a deep-learning approach is used to augment existing LES models using the relevant flow features. To this end, Random Forest Regression is developed to map relevant statistical flow-features within the LES solution to errors in calculated statistics such as sub-grid scale stresses. In this context, the exact solution is given by Direct Numerical Simulation (DNS) data. The capability of the proposed framework is examined by posteriori tests.