Expecting Turbulence: Machine Learning for Prediction and Control of Fluid Flows

ABSTRACT

Modern autonomous aerodynamic technologies such as unmanned aerial systems and wind turbines must regularly contend with highly stochastic and turbulent flow conditions. Conventional methods for modeling and controlling these fluid flows are limited in their ability to mitigate the resulting forces in real time. Modern machine learning techniques offer a solution, with the potential to revolutionize prediction and control of flow systems across scales.

In this talk, I will present three vignettes demonstrating the utility of artificial intelligence for applied flow prediction and control. First, we explore end-to-end solutions for control in a gusty environment with model-free and model-based reinforcement learning on a generalized aerodynamic test bed; this work represents the first successful application of reinforcement learning algorithms in a turbulent aerodynamic environment. Second, we explore a novel application of supervised learning methods for the prediction of turbulent flows, finding that state-of-the-art methods (e.g., Fourier neural operations) can effectively predict the physical time evolution of the flow field faster than real-time over a range of Reynolds numbers. Finally, we predict the flow fields in urban environments using easily attainable environmental information (e.g., building geometry) to enable flow-informed path planning for unmanned aerial systems. Given the utility of these methods over diverse flow conditions, I will discuss the broad applications of this work and plans for future contributions to making physically aware, intelligent autonomous systems a reality.

BIOGRAPHY

Peter I. Renn received his Ph.D. in Aeronautics at the California Institute of Technology in 2023 under the mentorship of Professor Morteza Gharib. He is interested in developing physics-informed solutions for machine learning and autonomous systems, drawing on expertise in experimental fluid mechanics, control theory, and aerial robotics. His current research focuses on physics-based machine learning solutions for planning and control in turbulent environments. Peter earned his M.S. from Caltech in Aeronautics in 2019, his B.S. from Caltech in Mechanical Engineering in 2018, and also holds a B.A. in Physics from Occidental College; he currently works as a quantitative researcher at a high-frequency trading firm. Peter received the William F. Ballhaus Prize for Best Dissertation in Aeronautics in 2023, and his work has received support from NSF, RTX, Verizon, Amazon, and the Technology Innovation Institute, among other partners.

For questions about the speaker, contact Prof. Nicholaus Parziale