Data-driven and Physics-constrained Uncertainty Quantification for Turbulence Models PDF Download

Author: Jan Felix Heyse
Publisher:
ISBN:
Category :
Languages : en
Pages :

Get Book Here

Book Description
Numerical simulations are an important tool for prediction of turbulent flows. Today, most simulations in real-world applications are Reynolds-averaged Navier-Stokes (RANS) simulations, which average the governing equations to solve for the mean flow quantities. RANS simulations require modeling of an unknown quantity, the Reynolds stress tensor, using turbulence models. These models are limited in their accuracy for many complex flows, such as those involving strong stream-line curvature or adverse pressure gradients, making RANS predictions less reliable for design decisions. For RANS predictions to be useful in engineering design practice, it is therefore important to quantify the uncertainty in the predictions. More specifically, in this dissertation the focus is on quantifying the model-form uncertainty associated with the turbulence model. A data-free eigenperturbation framework introduced in the past few years, allows to make quantitative uncertainty estimates for all quantities of interest. It relies on a linear mapping from the eigenvalues of the Reynolds stress into the barycentric domain. In this framework, perturbations are added to the eigenvalues in that barycentric domain by perturbing them towards limiting states of 1 component, 2 component, and 3 component turbulence. Eigenvectors are permuted to find the extreme states of the turbulence kinetic energy production term. These eigenperturbations allow to explore a range of shapes and alignments of the Reynolds stress tensor within constraints of physical realizability of the resulting Reynolds stresses. However, this framework is limited by the introduction of a uniform amount of perturbation throughout the domain and by the need to specify a parameter governing the amount of perturbation. Data-driven eigenvalue perturbations are therefore introduced in this work to address those limitations. They are built on the eigenperturbation framework, but use a data-driven approach to determine how much perturbation to impose locally at every cell. The target amount of perturbation is the expected distance between the RANS prediction and the true solution in the barycentric domain. A general set of features is introduced, computed from the RANS mean flow quantities. The periodic flow over a wavy wall (for which also a detailed high-fidelity simulation dataset is available) serves as training case. A random forest machine learning model is trained to predict the target distance from the features. A hyperparameter study is carried out to find the most appropriate hyperparameters for the random forest. Random forest feature importance estimates confirm general expectations from physical intuition. The framework is applied to two test cases, the flow over a backward-facing step and the flow in an asymmetric diffuser. Both test cases and the training case exhibit a flow separation where the cross sectional area increases. The distribution of key features is studied for these cases and compared against the one from the training case. It is found that the random forest is not extrapolating. The results on the two test cases show uncertainty estimates that are characteristic of the true error in the predictions and give more representative bounds than the data-free framework does. The sets of eigenvectors from the RANS prediction and the true solution can be connected through a rotation. The idea of data-driven eigenvector rotations as a data-driven extension to the eigenvectors is studied. However, continuousness of the prediction targets is not generally achievable because of the ambiguity of the eigenvector direction. The lack of smoothness prevents the machine learning models from learning the relationship between the features and the targets, making data-driven eigenvector rotations in the discussed setup not practical. The last chapter of this dissertation introduces a data-driven baseline simulation, which corresponds to the expected value in the data-driven eigenvalue perturbation framework. The Reynolds stress is a weighted sum of the Reynolds stresses from the extreme states. A random classification forest trained to predict which extreme state is closest to the true Reynolds stress is used to compute these weights. It does so by giving a probabilistic meaning to the raw predictions of the constituent decision trees. On the test cases, the data-driven baseline predictions are similar but not equal to the data-free baseline. They complement the uncertainty estimates from the data-driven eigenvalue perturbations.