Kevin Schuurman (TU Delft), Reduced state space approach for precursor identification of extremes in turbulent flows.
Abstract
Extreme events in turbulent flows, such as flashback in hydrogen combustion, pose significant challenges to the safety and reliability of engineering systems. Identifying their precursors can enable earlier warnings, improve control strategies, and provide insight into the mechanisms underlying these events. This talk explores how dynamical systems theory and deep learning can be combined to develop reduced-order methodologies for precursor identification. Viewing turbulent flow as a trajectory in state space provides a foundation for understanding the geometry of the underlying dynamics and the challenges of representing it in reduced dimensions. Using this view, we will see how to improve on the usual dimensionality reduction. The central idea is to exploit the system's underlying symmetries before learning a low-dimensional representation. By canonicalizing equivalent states, symmetry reduction can simplify the geometry of the learned manifold and potentially improve the representation's predictive capabilities.
The talk will not only focus on understanding when and why this canonicalization improves the learned representation. It will also examine how supervised learning objectives can further shape the reduced-order space with nonlinear manifold learners to distinguish precursor states from normal dynamics. Together, these approaches provide complementary perspectives on data-driven precursor identification. Their application to dissipation events in Kolmogorov flow and a minimal flow unit will illustrate how symmetry-aware representations and task-specific learning objectives can support the identification of significant dynamical states.
