We investigate non-local features of granular flows in nearly jammed configurations. We identify the presence of contact networks and crystallization as the reasons for such a behavior and perform Discrete Element Simulations on a high shear disk impeller granulator. The association of particle properties (e.g. the coefficients of restitution and friction) with the observed flow structures is studied using multivariate data analysis and multilinear optimization. The effect of polydispersity on the structure of the contact networks is studied. A singularity in the momentum flux is observed when approaching a monosized particle assembly, implying the onset of crystallization. Macroscopic quantities, such as the velocity profiles and momentum flux, are linked to microscopic data from DEM simulations through several statistical methods: the spectral analysis (giving dominant frequencies of the flow), coordination numbers that map the local flow regimes and the pair correlation function that is weakly associated with the variations of the flow structure. (C) 2017 Elsevier B.V. All rights reserved.