Data-Driven Approach for Turbulence Modeling in Rotating Flows and Stratified Flows [electronic resource]

다국어 초록 (Multilingual Abstract)

Turbulence modeling, including wall models in large-eddy simulations (LESs) and RANS models in Reynolds-averaged Navier-Stokes (RANS) simulations, is usually considered for traditional canonical flows, such as a channel flow or a shear layer flow. Consequently, the simulations with turbulence modeling have difficulty in handling unconventional flows, including rotating flows and stratified flows. Some of the main difficulties lie in the fact that these flows have multiple flow controlling parameters (FCPs), and thus, the flow behavior is hard to explore, let alone get accurate modeling.The data-driven approach is considered a possible solution to this. The increasing computational resources and shared turbulence data allow another way to utilize the data other than pure human analyses of the physics. However, pure data-driven methods are often criticized for their weak interpretability and generalizability.In this dissertation, multiple data-driven techniques are applied to some persistent problems in turbulence modeling under the circumstances of rotating flows and stratified flows. The problems include not only the accurate modeling of the flow but also the efficient FCP space exploration, model selection, and uncertainty quantification, etc. Both the dataset and existing knowledge of physics are utilized, and then data-driven approach shows the interpretability and generalizability. They show how these traditionally difficult problems can be tackled through physics-informed data-driven approach, which significantly saves human labor.In chapter 1, a detailed literature review of physical problems, difficulties in turbulence modeling, and data-driven approach provide a brief overview of the current research and the objectives of this dissertation.In chapters 2 and 3, wall-modeled large-eddy simulations (WMLESs) are explored. For a spanwise rotating channel, the mean flow shows a linear profile and wall models can be developed in both a physics-based approach and a data-driven approach. The data-driven approach shows better accuracy and capability to generalize, which makes it a more appealing choice to save human labor in developing wall models. For an arbitrarily directional rotating channel, the mean flow does not have a known profile so it is currently impossible to find a wall model through physical understanding. To handle a large number of FCPs, the FCP space is explored by Bayesian optimization and a wall model is developed through a surrogate model, namely Gaussian progress regression. In summary, the capability of wall modeling is extended to flows with rotation.In chapters 3, 4 and 5, RANS simulations are explored. For flows controlled by different physical processes, a recommender system is developed to automate the process of model selection. Meanwhile, the feature vectors from the recommender system align with existing experiences of the dominating physical processes in a quantity of interest (QoI) and the ability of a RANS model. Therefore, the prediction from a recommender system can be physically interpreted since it is consistent with human experiences. For a stratified wake, the multi-stage behavior of the flow requires a switch of modeling as the flow develops. A linear logistic regression classifies the flow into weakly stratified turbulence (WST) regime and strongly stratified turbulence (SST) regime accurately when the dataset of only one flow condition is fed into the classifier during training. The classifier will find the dominating physical processes which is shared among different flow conditions, which is also consistent with how the regimes are physically identified. The applicable range of a RANS model is then identified through a global epistemic uncertainty quantification (UQ) method for a stratified shear layer. This method allows the exploration of dominating terms in a RANS model and determining a priori if a calibration can generalize. They are quantified through effectiveness and inconsistency, which are factors that calibration will consider.In general, a data-driven approach has been used for multiple applications in turbulence modeling. The involvement of data shows power in improving prediction accuracy and saving human labor, and the consideration of the underlying physics enables its interpretability and generalizability. More work can be done in the future for multiple aspects of turbulence modeling to realize accurate prediction in real-world flow conditions.