Quantifying Structure Evolution in Unsteady Turbulent Boundary Layers

Transportation vehicles lose much of their energy to turbulence. Nearly 15% of the world's energy consumption goes to overcoming turbulent drag on vehicle surfaces. Reducing this loss requires a deeper understanding of turbulence, one of the unsolved problems in physics. Hidden within the apparent randomness that characterizes turbulent motion, are organized patterns that could be key to controlling it, if only we could identify them.

This project uses explainable deep learning (DL), techniques that reveal how a DL model reaches its predictions, to pinpoint which regions of a turbulent flow matter most for its behavior. Building on a method inspired by game theory, originally used to value players' contributions in games, we train a DL model to reproduce the flow's dynamics, then rank the importance of every point, without presupposing what should count as important.

Working with Professor Beverley McKeon's group at Stanford, we will apply this approach to real measurements taken under changing pressure conditions, similar to what an accelerating vehicle experiences. Tracking how the most influential features shift, we aim to better understand and help control the turbulence that drains energy from transportation, continuing a long partnership between France and Stanford in fluid mechanics.