Vortex-SegFormer: a transformer-based vortex detection method

Abstract

The identification of vortices plays a critical role in various domains, such as fluid dynamics, weather forecasting, and engineering systems. Existing vortex detection methods are typically categorized into global, local, and machine learning-based methods. Global methods provide high accuracy but require high computational power, while local detection methods output results quickly at the expense of being less reliable. Machine learning-based methods aim to combine both speed and accuracy, but they tend to accept input patches, which discards some global information. To mitigate the drawbacks, we propose Vortex-SegFormer, a Transformer-based model, to detect vortices in 2D flows reliably without the need for patch sampling. Additionally, it can generalize well to unseen data, even with differing resolutions than the training data. The proposed method is then compared against existing vortex detection methods on simulated flow fields. The results show that Vortex-SegFormer outperforms existing methods and can capture more vortices.