Applied Machine Learning for Prediction and Control of Fluid Flows

Author: Renn, Peter Ian James

Year: 2023

Degree: Dissertation (Ph.D.)

Advisor: Gharib, Morteza

Committee Members: Bae, H. Jane; Anandkumar, Anima; Dabiri, John O.; Gharib, Morteza

Option: Aeronautics

DOI: 10.7907/smnv-tz73

Abstract

Modern aerodynamic technologies such as unmanned aerial systems and horizontal axis wind turbines must regularly contend with forces from highly stochastic and turbulent atmospheric gusts. Conventional methods for modeling and controlling fluid flows are limited in their ability to mitigate these aerodynamic forces in real-time. By applying modern machine learning techniques in an experimental setting, this thesis demonstrates the utility of machine learning in addressing these important problems. We follow two complementary approaches towards this goal.

First, we find an end-to-end solution for control in a gusty environment with model-free reinforcement learning. We deploy state-of-the-art reinforcement learning algorithms on a generalized aerodynamic test-bed consisting of an airfoil with motorized trailing edge flaps. The system features embedded flow sensors, enabling the inclusion of flow measurements in state observations. We place this system in a highly irregular wake behind a bluff-body, dynamically mounted on elastic bands and therefore free to oscillate, and train reinforcement learning agents to minimize the net lifting force on the system by controlling the position of the trailing edge flaps. We find that model-free reinforcement learning agents can outperform basic linear controllers in this gusty, turbulent environment. We also show that augmenting state observations with flow measurements can lead to more consistent learning of the system dynamics.

Next, we explore Fourier neural operators (FNOs) as a method for forecasting the time evolution of turbulent fluid flows. FNOs are capable of learning underlying operator solutions to families of partial differential equations and can be evaluated in just milliseconds. We specifically focus on training FNOs with experimentally measured velocity fields of bluff body wakes in the subcritical regime. To the best of our knowledge, this is the first application of operator learning for fluid mechanics that features experimental measurements. We find that FNOs can accurately predict the evolution of these turbulent wakes even when trained with imperfect measurements. We then show that FNOs can quickly adapt to unseen conditions with minimal data and training through transfer learning. Finally, we consider the performance of FNOs over longer prediction horizons. This approach could enable real-time gust prediction capabilities and monitoring for applied aerodynamic systems.

Files