Local cluster-based identification of flow dynamics

A novel data-driven framework is proposed for the analysis and reduced-order modelling of complex fluid flows based on experimental velocity-field measurements. The methodology combines manifold learning and cluster-based network modelling to identify spatially localised flow dynamics that are often hidden in traditional global analyses. First, a lowdimensional spatial manifold is constructed from the temporal evolution of the flow field using ISOMAP, allowing the automatic partitioning of the domain into kinematically coherent subregions through clustering. Subsequently, local cluster-based network models are developed for each subdomain to describe their individual dynamics. The framework is validated using time-resolved PIV measurements of a transitional water jet, where it successfully identifies physically meaningful regions associated with the potential core, vortex shedding, or vortex pairing. The results demonstrate that local modelling provides a more interpretable and accurate representation of multi-regime flow behaviour than conventional global reduced-order approaches, offering a promising tool for flow analysis, reconstruction, and future data-assimilation applications.