 |
Center for Nonlinear Studies |
Predicting drag over complex bodies plays a crucial part in designing high performance engineering designs such as aircrafts. Current turbulence models are known to given erroneous onset of separation and thus needs improvements. Recent years has seen an increase in available high fidelity data sets and thus data driven modelling is now being tested as a potential tool to improve turbulence closure models. In-line with this goal, the present study aims to evaluate if machine learning can be used to augment turbulence modelling.Empirical data are obtained for a series of increasingly high bumps by Large Eddy Simulation. A patch of high turbulent kinetic energy forms in the lee of the bump and extends into the wake. It originates near the surface and has a significant influence on flow development. The highest bumps create a small separation bubble. Over the bump the log-law is absent, evidencing strong disequilibrium. The dataset is created to be used in data-driven modelling. An optimization method is used to extract fields of variables that are used in turbulence closure models. From this, it is shown how these models fail because they predict near-wall eddy viscosity erroneously. Machine learning is used to generalize the optimized field variables such that existing turbulence models can produce more accurate results on different test cases. It is shown that these machine learning augmented closure models result in a modest improvement in test cases.
Follow up discussions will continue (after lunch) in the Study Center, JRO-5 Room-12pm-4pm.
Host: Michael Chertkov