R. Pochampalli(Chair for Scientific Computing, University of Kaiserslautern)
hosted by Seminar Series on Scientific Computing
"New Approaches in Data Driven Turbulence Modeling"
RANS equations combined with turbulence models perform adequately at lower computational cost than high fidelity methods. One equation turbulence models such as Spalart-Allmaras(SA), produce good results for specific applications, such as flows past wings and airfoils. However, the latter suffers from inaccuracies on quantities such as lift and drag in certain regimes, such as separated flows. A direct means to improve existing models is by introducing correction functionals onto terms of the model. For a specific flow configuration, these correction functionals can be estimated by inverse problems based on experimental measurements. Data driven techniques generalize this process by learning the correction functional in terms of flow variables, using data obtained from inverse problems.
We present two new approaches that primarily address statistical aspects of turbulence modeling with machine learning. In particular, the correction functional varies depending on several factors including the flow conditions and topology. The modeling approaches taken up here reflect the circumstance that the data at hand does not conform to the classical assumptions of statistical learning theory. Consequently, the focus is on developing frameworks that can treat non-identically distributed data. The first approach introduces a means to reduce the dependency of data on flow domains. The second approach transforms the data into a high dimensional representation that captures the topology of the flow domain. Finally, the performance of the machine learning enhanced turbulence model is compared to the SA model.
|Time:||Thursday, 05.11.2020, 11:30|