Reduced-Rank Multi-Fidelity Modeling of Complex Aerodynamic Flows

• Main Author: Skinner, Ryan W
• Format: Dissertation
• Language: English
• Published: ProQuest Dissertations & Theses 01-01-2021
• Subjects: Aerospace engineering; Applied Mathematics; Fluid mechanics; Mechanical engineering
• ISBN: 9798738631375

This thesis explores the efficacy of a reduced-rank bi-fidelity modeling framework when applied to aerodynamic flows characterized by high Reynolds numbers, complex geometries, interacting cross-flows, and highly unsteady turbulent separation. It asks the question, "how can we speed up parameter space exploration of aerodynamic flows whose complexity prohibits practitioners from carrying out more than a handful of costly high-fidelity simulations?"The bi-fidelity method's viability is initially demonstrated on a two-dimensional NACA airfoil subject to geometric variation. The bi-fidelity model is able to predict pressure coefficient along the airfoil to within RANS validation error at a cost that is roughly 20x less than a full suite of high-fidelity simulations.Two extensions to the bi-fidelity framework are then proposed and evaluated. The first is a general, formal procedure for assessing the model's viability for application to a novel flow. The second is a method for expanding the parameter space after a bi-fidelity model has already been constructed. Below a certain initial bi-fidelity model rank, the high-fidelity solutions used to construct the initial bi-fidelity model are equally useful for the expanded-space model, meaning computational resources expended on those initial high-fidelity solutions are not wasted and can be re-used.The remainder of the thesis concerns the aggressive diffuser, which is a complex three-dimensional system representing an industrially-relevant flow to which the bi-fidelity modeling framework can be applied. A high-fidelity DDES model of the diffuser is constructed and extensively validated against time- and phase-averaged experimental data provided by Prof. Amitay's group at Rensselaer Polytechnic Institute (RPI). The dynamics of this flow are documented in detail using the DDES simulation's phase-averaged fields, and potential avenues for improving the flow control are suggested based on our enhanced understanding of the dynamics.Finally, the reduced-rank bi-fidelity modeling framework is applied to a closely-related aggressive diffuser that uses a segmented tangential blower and corner suction to control separation. On average, one high-fidelity model evaluation requires over 800,000 core-hours on an IBM Blue Gene/Q supercomputer, which is over 350 times higher than the cost of one low-fidelity model evaluation. Bi-fidelity models of ranks 10 through 19 achieve qualitatively favorable predictions of both time- and ensemble-averaged velocity magnitude at the diffuser exit plane on this complex, industrially-relevant three-dimensional system at a cost reduction over a traditional brute-force approach of one to two orders of magnitude.In summary, this thesis demonstrates that the reduced-rank bi-fidelity modeling framework can achieve favorable predictive accuracy at over an order of magnitude reduction in simulation cost when applied not only to simple two-dimensional aerodynamic test cases, but also to complex aerodynamic flows at the forefront of our modeling capabilities. Most importantly, this method is applicable to such systems where only a handful of high-fidelity model evaluations can be carried out due to their outsize computational cost. These results motivate further study of this bi-fidelity modeling framework and its application to industrially-relevant systems of interest to practitioners.