Applied Math Seminar: Jahrul Alum (Memorial University of Newfoundland)

Title: Data-driven large eddy simulation of atmospheric turbulence

Abstract: Over the last few years, machine learning has been critical in science and engineering and emerged as a data-driven turbulence model. However, machine learning depends on training data from previous experiments on turbulent flows. Typically, training data capture only a fraction of the active scales of turbulence. Despite decades of research, the best turbulence theory has yet to emerge, which limits the training of supervised machine learning models. Reinforcement learning is one way to alleviate these challenges. A reinforcement learning model interacts directly with the dynamical system itself. In this talk, I will use the Burgers equation to illustrate data-driven learning of dynamical systems. Then, I use simulations of a NACA airfoil and a wind farm to outline the reinforcement learning framework. Finally, the talk presents a proof of concept for optimizing large eddy simulation through reinforcement learning.