BPCPPDYSOE21 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

DY: Fachverband Dynamik und Statistische Physik

DY 8: Fluid Physics 2 - organized by Stephan Weiss and Michael Wilczek (Göttingen)

DY 8.1: Vortrag

Montag, 22. März 2021, 11:00–11:20, DYa

Interpreted machine learning: Explaining relaminarisation events in wall-bounded shear flows — Martin Lellep¹, Jonathan Prexl², Bruno Eckhardt³, and •Moritz Linkmann⁴ — ¹School of Physics and Astronomy, University of Edinburgh, UK — ²Dept. of Civil, Geo and Environmental Engineering, Technical University of Munich, Germany — ³Dept. Physics, Philipps-University of Marburg, Germany — ⁴School of Mathematics and Maxwell Institute for Mathematical Sciences, University of Edinburgh, UK

Machine Learning (ML) is becoming increasingly popular in fluid dynamics. Powerful ML algorithms such as neural networks or ensemble methods are notoriously difficult to interpret. Here, we use the novel Shapley Additive Explanations (SHAP) algorithm (Lundberg & Lee, 2017), a game-theoretic approach that explains the output of a given ML model, to ascertain which physical processes are significant in the prediction of relaminarisation events in wall-bounded parallel shear flows. The flow is described by an established low-dimensional model whose variables have a clear physical and dynamical interpretation in terms of known representative features of the near-wall dynamics, i.e. streamwise vortices, streaks and linear streak instabilities. We consistently find only three modes to play a major role in the prediction: the laminar profile, the streamwise vortex, and a specific streak instability. SHAP thus distinguishes representative from significant features, hence we demonstrate that it is an explainable AI method which can provide useful and human-interpretable insight for fluid dynamics.