A hybrid computer vision and machine learning approach for robust vortex core detection in fluid mechanics applications

Vortex core detection remains a challenging topic within the field of computational fluid dynamics (CFD). Local methods, such as the Q, delta, or swirling-strength criterion, are commonly used to detect vortices and these methods are entirely based on the local velocity gradient tensor. Results have shown that reasonable estimates can be obtained with these methods, however, at the same time, these methods produce a significant number of false positives and negatives. User-defined tuning parameters are introduced to keep the number of false positives and negatives in balance, but this requires knowledge of the vortices and thus does not present a robust and self-contained approach. We recently proposed a novel computer vision approach where we have trained a convolutional neural network (CNN) to look at line integral convolution (LIC)-based streamline plots. We showed that this approach is capable of accurately predicting the regions where vortices reside, and we were able to reduce the false positives and negatives to zero. Furthermore, we showed the universality of this approach by successfully applying our trained CNN to a different test case for which it has not been trained and which featured different vortical structures (generated through a different physical process). The CNN-based approach is limited in the sense that it is only able to predict the bounding boxes of vortex cores, but not the exact location of the vortex core itself. Therefore, we propose a hybrid machine learning and computer vision approach in this study, where we first identify areas of vortical structures using computer vision to which we add a layer of machine learning to find the vortex core within the vortex region. We test different sets of input parameters for both the hybrid and pure machine learning approach, starting with just the primitive variables (velocity and pressure), and adding more derived quantities (velocity gradients, pressure gradients, Q-criterion, vorticity, and magnitude of vector quantities). Comparing the hybrid with the pure machine learning approach applied to the full flow field, we show that the hybrid approach reduces the training time for all tested cases up to a factor of 2. We also find that using the primitive variables along with their derivatives provide fewer false positives and negatives using the hybrid approach. At the same time, using the variable set with all possible inputs does not provide a more accurate prediction of vortex cores and thus we demonstrate that our hybrid computer vision and machine learning approach is an effective way to reduce false positives and negatives entirely using just the primitive variables and their derivatives.