Deep representation and unsupervised learning for the experimental analysis of rigid particle margination in red blood cell flows

In continuation of recent experimental efforts by Coutinho et al. (2023, Phys. Rev. Fluids), we provide further evidence on the margination mechanism of rigid particles (RPs) in microcirculation, which is in contrast to predictions from previous numerical works (Crowl and Fogelson, 2011, J. Fluid. Mech.; Závodszky et al., 2019, Phys. Fluids). We depart from single particle trajectories (SPT) of the RPs in red blood cell (RBC) flows obtained from defocus particle tracking (DPT) measurements in straight microchannels and employ an unsupervised learning pipeline for the analysis of SPT data patterns. The combination of feature extraction, dimensionality reduction, and HDBSCAN clustering allows an analysis of physical quantities beyond traditional approaches while minimizing user influence on hidden structures in the SPT data. The optimal clustering solution in RBC flows showed the presence of two clusters: one located in the core flow region and another sitting closer to the cell-free layer (CFL). The RPs in the latter region exhibited a wavy pattern, with those moving close to the CFL being re-directed towards the center of the flow. The absence of external forces acting on the system points to the RBCs flowing at the boundary with the CFL as key elements in delaying margination.