Past Event: Oden Institute Seminar

Uncertainty Quantification for Reynolds-Averaged Navier-Stokes Predictions

Todd Oliver, PECOS, ICES

Abstract

In many applications in science and engineering, solutions of very high fidelity models---i.e., first principles models that are accepted as truth within a well-known domain---are unavailable, either because such models are unknown or because their solution is too computationally expensive for the application of interest. In these cases, important decisions must be informed using predictions from less complete models. Thus, it is critical to be able to estimate the uncertainty in the predictions of such models and, ultimately, to be able to assess the validity of these predictions. This ICES Forum will explore the uncertainty quantification (UQ) process in the context of Reynolds-averaged Navier-Stokes (RANS) turbulence modeling. While the Navier-Stokes equations are a sufficiently high fidelity model for a wide range of problems in fluid mechanics, for many turbulent flows of technical interest, the computational costs required to achieve well-resolved numerical solutions of the Navier-Stokes equations are prohibitive. Thus, the typical procedure in engineering applications is to solve the RANS equations coupled with a semi-empirical model of the unclosed terms introduced by the averaging procedure. This closure model represents the effects of turbulence on the mean flow and typically includes a number of uncertain parameters. More importantly, the effects of turbulence are difficult to model, and the standard closure models are notoriously unreliable. Thus, it is important to quantify the uncertainty due to both uncertain parameters and model inadequacy. In this work, a Bayesian probabilistic approach is used to capture both effects. The approach involves stochastic modeling, calibration against data, and finally prediction. In contrast to typical deterministic RANS simulations, the output of the process is a probability density function which characterizes both the best prediction and the uncertainty for the quantity of interest. The talk will show this UQ process applied to RANS predictions for two simple wall-bounded flows. Finally, recent research towards a validation procedure for such models will be briefly described. Coffee and cookies will be provided. We hope to see you there.