Characterization and prediction of extreme events in turbulence

Extreme events in Nature such as tornadoes, large floods and strong earthquakes are rare but can have devastating consequences. The predictability of these events is very limited at present. Extreme events in turbulence are the very large events in small scales that are intermittent in character. We examine events in energy dissipation rate and enstrophy which are several tens to hundreds to thousands of times the mean value. To this end we use our DNS database of homogeneous and isotropic turbulence with Taylor Reynolds numbers spanning a decade, computed with different small scale resolutions and different box sizes, and study the predictability of these events using machine learning. We start with an aggressive data augmentation to virtually increase the number of these rare events by two orders of magnitude and train a deep convolutional neural network to predict their occurrence in an independent data set. The goal of the work is to explore whether extreme events can be predicted with greater assurance than can be done by conventional methods (e.g., D.A. Donzis & K.R. Sreenivasan, J. Fluid Mech. 647, 13-26, 2010).