Explainable sequential unsupervised data driven clustering for feature identification in external flows

This paper introduces a novel unsupervised clustering algorithm tailored for feature identification in external fluid flows within computational fluid dynamics. By employing sequential clustering techniques, the proposed method extracts key flow features, such as shock waves, boundary layers and wakes, vortices, and laminar separation bubbles from compressible and incompressible flows. Unlike traditional approaches, this algorithm relies on primitive flow variables and their inherent physical properties, eliminating preprocessing requirements and enhancing generalization capabilities. Utilizing Gaussian mixture models and incorporating rescaling and binarization operations, the method partitions the flow domain into distinct regions of interest through transformations of primitive variables and gradients of thereof. The parameter-free design ensures ease of implementation and robustness across diverse flow scenarios, including two- and three-dimensional configurations. The algorithm's versatility and accuracy are demonstrated through its application to various cases, showcasing its potential as a powerful and accessible tool to gain deeper insight into complex fluid phenomena.