Seeking insight from the performance of machine learning classifiers in determining the generating mechanism of vortex patterns

Machine learning algorithms have been applied with impressive success in Silicon Valley, helping solve problems that are complex and often without a governing equation, and doing so using enormous datasets. As such, there has been a rapidly growing interest in whether or not these tools can be applied with similar success in fluid mechanics, where nonlinearities, complex dynamics, and large datasets are common. In this work, we consider the vortex patterns generated by the motion of rigid flat plates at low Reynolds numbers, a common model flow for bio-inspired propulsors. We numerically simulate prescribed pitching and plunging motions, respectively, choosing parameters such that 2S vortex streets are observed. Using machine learning methods such as linear discriminant analysis, we are able to use downstream measurements to accurately identify which upstream motion has generated the observed vortex pattern. In addition, we evaluate the performance of our machine learning classifiers as various hyperparameters are changed; these include the number of sensors, the measured flow variables, and the sensor locations. Observed changes in classifier performance provide insight into the underlying flow physics and suggest best practices for sensor design.