Colloquium: Alexander Blauw (AWEP)
24 september 2019 10:00 - Locatie: Lecture Room F, Faculty of Aerospace Engineering, Kluyverweg 1, Delft.
Bayesian Additive Regression Trees for data-driven RANS turbulence modelling
Turbulent flows are encountered in scientific research or engineering applications, which are governed by the Navier-Stokes equations. The Reynolds averaged form (RANS) is most commonly solved. The most recent efforts to close the RANS equations have been based around machine learning. Uncertainty quantifications is an important aspect for these models, as they can become inaccurate outside the range of the training data. Therefore, the aim of this research was to develop a data-driven turbulence model of the RANS equations that is able to predict and quantify the uncertainty.
Uncertainty is introduced in the anisotropic Reynolds stress tensor, which is predicted using Bayesian Additive Regression Tree model. The results of this model show great resemblance to DNS/LES data for square duct and backwards facing stp flows. Furthermore, this model shows increasing levels of uncertainty ,when square duct training data is extrapolated to test cases of higher aspect ratio ducts.