Speaker: Alec Linot, Ph.D.
IDRE Fellow
Mechanical and Aerospace Engineering
University of California Los Angeles

Location:Zoom (Registration required)

Time: 11:30 AM – 12:30 PM (PST)
Registration Link: https://ucla.zoom.us/meeting/register/tJMrdeirqDkpGt1TLOylwUolqcjX6tMy1dMH

Abstract: Forecasting the dynamics of fluid flows plays a crucial role in our understanding of processes such as the swimming of fish, turbulence on a plane, and hurricane formation. Unfortunately, simulating these systems can be prohibitively expensive even though we often know the equations of motion. Due to this high computational cost, major effort has gone toward developing reduced-order models (ROMs) of fluid flows both from first principles and in a data-driven manner. Various ROMs using Galerkin methods and neural networks, for example, have been shown to accurately predict the dynamics of fluid systems with far fewer degrees of freedom than needed in high-resolution simulations. However, these ROMs typically apply to very specific systems with a fixed state size (e.g. grid size or latent space size). In this work, we present a data-driven ROM method for discovering vortex dynamics that overcomes the challenge of a fixed state size by using a hierarchy of graph neural networks (GNNs). This method allows us to consider a fluid flow as a graph of the vortices within a flow. Then, by grouping clusters of vortices, we construct a hierarchy of graphs with which we train GNNs to predict vortex dynamics. Notably, this hierarchal approach mirrors our intuition on how groups of vortices often cluster to act as a cohesive unit. We show that this hierarchical method is both more accurate and faster than constructing a fully connected GNN, and we show that this approach allows us to predict vortex dynamics with state sizes (i.e. the number of vortices) outside of our training data.

About the speaker: Dr. Alec Linot is a postdoctoral researcher with Prof. Kunihiko (Sam) Taira in the Mechanical and Aerospace Engineering Department at UCLA. He received a BS in Chemical Engineering from Kansas State University. His Ph.D is in Chemical and Biological Engineering from the University of Wisconsin – Madison. In his Ph.D., he developed machine learning techniques for modeling and controlling turbulent flows. His current research is in modeling, control, and stability of chaotic dynamical systems. Chaotic dynamical systems are deterministic systems where small perturbations to the system result in dramatically different dynamics over time.