Regime classification for stratified wakes from planar velocity field snapshots using convolutional neural networks

Previous studies have shown that stratified wakes can be classified into various regimes from topological features present in the velocity field, which in turn depend on the Reynolds number and Froude number. In this study we use a machine learning-based technique to develop a wake regime classifier from very limited data: single 2D-2C velocity snapshots in the vertical plane. Specifically, we use convolutional neural networks (CNNs) which are often used for image classification due to their ability to "learn" distinguishing features in the images and generate translation-equivariant feature maps. This also makes CNNs ideal for the present application. We train the CNN on a labelled dataset of velocity field snapshots available from direct numerical simulations. Classification accuracy is then evaluated using an experimental dataset that does not always have the same field of view as the numerical dataset that the CNN has been trained on. The resulting accuracy is therefore a measure of the robustness of the classifier to real-world measurements. This study complements previous work on the development of an expert-defined decision-tree-based classification system and holds promise for the development of automated fluid pattern classifiers.