The ability to compute turbulent flows around aerospace geometries has revolutionized our industry's capabilities: when turbulent flow solvers produce accurate results, they have found their way…

This is a cooperative agreement awarded by the National Aeronautics and Space Administration (NASA) to the Regents of the University of Michigan to develop improved turbulence modeling capabilities for computational fluid dynamics (CFD) solvers. The $713,987 Phase II award builds on prior Phase I work to create data-driven turbulence models that can better predict transitional and separated turbulent flows. The goal is to incorporate these models into CFD codes to enable more accurate and robust simulations of aerospace applications. No set-aside was used for this award. The University of Michigan has extensive experience conducting research and providing services to federal agencies like the Department of Defense and NASA under both prime contracts and subcontracts.

The ability to compute turbulent flows around aerospace geometries has revolutionized our industry's capabilities: when turbulent flow solvers produce accurate results, they have found their way into analysis, certification, and design processes. However, our predictive capabilities are still far from perfect: progress in rans turbulence modeling has largely stagnated over the past decade and the confidence with which we can predict transitional and separated flows, as well as various mixing processes, is still lacking. In this proposal, we leverage advances in the theory of turbulence modeling and our recent ability to produce and process data (from high-fidelity simulations and experiments) to create vastly improved models that can be embedded into existing CFD solvers. Such a capability would have a transformational effect on many aspects of the research being conducted at armd, NASA, and the aerospace industry as a whole. Our efforts in phase i answered several fundamental questions, resulting in the development of new tools and techniques that are already proving to be valuable to the community. We discovered an innovative formulation for data-driven turbulence modeling, and showed in very clear terms how IT can have great value. Specifically, we developed a novel method of inverse modeling to infer functional discrepancies in turbulence models using data from high-fidelity simulations and experiments. This approach provides direct and unprecedented information about qualitative and quantitative deficiencies in the closure terms. Machine learning techniques were developed, to reconstruct closure terms as functions of parametric inputs. In essence, we developed a framework that extracts information from data, reconstructs information to create knowledge, and embeds IT into a CFD solver to enable predictive calculations. This framework was demonstrated via canonical and practical problems, involving the prediction of transitional and turbulent flows. Phase II, will narrow the application area, focusing on improved modeling of turbulent separated flows. Towards this end, linear and non-linear eddy viscosity closures will be targeted and the data-driven framework will be applied to two and three-dimensional turbulent flow separation. The tools and intuition that we developed in phase i will be expanded into a complete data-to-prediction cycle. Machine learning is just one tool to enable the information-knowledge step; also, we will introduce physics-based constraints to ensure that our tool does not overfit the data and lose robustness in predictive capability. The machine learning and inversion models developed in phase i were exploratory and not production-ready. Based on the lessons learnt, we have a clear way forward and will work towards improving scalability and efficiency. While we seek to radically improve rans models, we recognize the fact that rans models have a ceiling in terms of predictive capability. With an eye towards the future, we will begin exploring the application of the aforementioned ideas to hybrid rans-LES methods. Specifically, we seek to optimize and inform near-wall length-scale parameters and blending functions. No model will ever be perfect, and IT is pertinent to ask whether we can gain information about uncertainties in the model development process. Towards this end, we will adopt a bayesian approach, such that bounds can be predicted. This may be an expensive proposition, as a number of realizations of the solver may need to be run, but IT may be a useful capability for robust analysis and design. As part of the legacy of this program, our team is committed to providing, to the broad community, a data-enabled turbulence model that incorporates the results of our work in phase II. Our vision is to maintain a continuously augmented website that provides improved closure models in the form of modular software that can be seamlessly embedded into CFD codes.