Identifying regions of importance in wall-bounded turbulence through explainable deep learning.

Understanding the mechanisms underlying turbulent flows has been a challenge for more than a century. Traditionally, the study of these flows has relied on analyzing so-called coherent structures. One of the most common definitions of coherent structures is the intense Reynolds-stress events or Q events. This work uses an explainable-deep-learning methodology to assess the importance of the different Q events for two different databases. The first one is a simulated three-dimensional turbulent channel with a friction Reynolds number of 125. The second is an experimental turbulent boundary layer at a friction Reynolds number of 1377. The analysis reinforces that ejections (low-speed flow moving upstream and away from the wall) and sweeps (high-speed moving downstream and towards the wall) dominate the evolution of the flow, containing, those structures of higher importance, a higher Reynolds stress. However, when both magnitudes are divided by the volume of the structure, the direct relationship disappears, evidencing that the regions of higher importance in the domain are different from those regions containing higher values of the Reynolds stress.