Analysis and Clustering of Flow Phenomena Using Machine Learning Methods

Abstract

The present work investigates how unsupervised machine learning methods can be applied to detect complex phenomena in flow simulation data. For this purpose, a variational autoencoder (VAE) is implemented. It allows for non-linear dimensionality reduction of a data set, resulting in the so-called feature space. Suitable parameters for a VAE are determined, and a number of algorithms for finding clusters in the feature space are introduced and tested. The resulting clusters are investigated in a 3D plotting software, showing that the model can detect complex phenomena, e.g., a shock or a shear layer.