Date of Award
Doctor of Philosophy (PhD)
Peter E. Hamlington
Turbulence is pervasive throughout most thermal-fluid systems, yet the modeling of turbulence, and its effects on engineering systems, remains a persistent challenge. This dissertation brings a new set of analytical tools to bear on the turbulence problem, and in doing so reveals new insights into turbulence modeling and engineering design optimization. Data-driven machine learning and optimization techniques are employed in a new autonomic closure for coarse-grained turbulent flow simulations. Sparsity-inducing, multi-task learning, feature extraction, and kernel extensions of the autonomic closure are further explored. These techniques improve the speed, accuracy and interpretability of the closure. Additionally, efficient adjoint optimization techniques are used to improve the engineering design of wind turbine layouts. This novel application of adjoints brings groundbreaking model fidelity and high-dimensional gradient-based optimization algorithms to the challenging turbulent flow control problem found in designing wind farms.
The autonomic turbulence closure learns optimal flow-specific turbulence closures on the fly, and is fundamentally different from past attempts at turbulence modeling rooted in continuum mechanics. The autonomic closure uses test filtering in the inertial range to collect training data for a supervised learning problem that discovers an optimal turbulence closure at a test scale. The learned model is then applied at the grid scale by invoking scale similarity in the inertial range. The supervised learning problem is solved with Volterra series expansion, sparsity-inducing optimization, feature extraction, and Gaussian process regression techniques. The resulting models demonstrate high accuracy, automation, and low cost.
Additionally, high dimensional optimization techniques are applied to the turbulent flow control problem of optimizing wind turbine locations. The strong coupling between power production, mechanical loads, and atmospheric turbulence requires high-fidelity flow models and efficient optimization algorithms. Gradient-based optimization of a wind plant's annual energy production is demonstrated with gradients obtained from adjoint simulations of the wind flow model. The layouts and optimization strategies obtained with this approach utilize nonlinear flow curvature effects that are typically absent in engineering flow models used by industry. In both turbulence model development and wind plant design, the optimization and learning framework provides new insights into turbulence physics, data-driven modeling and high-dimensional engineering design optimization.
King, Ryan Nicholas, "Learning and Optimization for Turbulent Flows" (2016). Mechanical Engineering Graduate Theses & Dissertations. 135.