Identification of Flowfield Regions by Machine Learning

Abstract

A machine learning algorithm is here proposed with the objective to identify homogeneous flow regions in computational fluid dynamics solutions. Given a numerical compressible viscous steady solution around a body at high Reynolds numbers, the task is to select the grid cells belonging to the boundary layer, shock waves, and external inviscid flow. The Gaussian mixture algorithm demonstrated to overcome some of the limitations and drawback of the currently adopted deterministic region selection methods, which require the adoption of case-dependent cutoff inputs, topological information, and final human check. This paper shows an example of application of this selection method performing an accurate breakdown of the aerodynamic drag in viscous and wave contributions by a classical far-field method. The new algorithm essentially leads to the same results of the reference method in terms of drag decomposition; slight differences could only be found in the shock-wave/boundary-layer interaction zone, where the drag breakdown is inherently ambiguous.