Machine learning classification of stratified wakes using dynamic mode decomposition and decision trees

Previous work has shown that stratified wakes can be sorted into known regimes based on numerical criteria derived from the results of dynamic mode decomposition (DMD) [Ohh & Spedding, APS DFD 2019; OS19]. Here, we extend that work by applying methods from machine learning. As before, we compute features for each candidate wake using DMD: a dominant DMD mode is identified and characteristics of that mode, such as its symmetry in each cross-stream direction, are computed. Those features are then used to train a decision tree classifier, which labels candidate wakes using a series of if-then statements. This mirrors the general structure of the model developed previously in OS19, except that here, the criterion used for each if-then statement is determined automatically by the decision tree training algorithm, rather than using human expertise. We find that our model is able to achieve high accuracy while maintaining interpretability, a common challenge in machine learning. Furthermore, the decision tree utilizes many of the same features that were chosen in OS19.