We employ a recent technique for building complex networks from time series data to construct a directed network embodying time structure to collate the predictive properties of individual granular sensors in a series of biaxial compression tests. For each grain, we reconstruct a static predictive model. This combines a subset selection algorithm and an information theory fitting criterion that selects which other grains in the assembly are best placed to predict a given grain's local stress throughout loading history. The local stress of a grain at each time step is summarized by the magnitude of its particle load vector. A directed network is constructed by representing each grain as a node, and assigning an in-link to a grain from another grain if the latter is selected within the best predictive model of the first grain. The grains with atypically large out-degree are thus the most responsible for predicting the stress history of the other grains: These turn out to be only a few grains which reside inside shear bands. Moreover, these “smart grains” prove to be strongly linked to the mechanism of force chain buckling and intermittent rattler events. That only a small number of grain sensors situated in the shear band are required to accurately capture the rheological response of all other grains in the assembly underlines the crucial importance of nonlocal interactions, espoused by extended continuum theories which posit nonlocal evolution laws. Findings here cast the spotlight on two specific mechanisms as being key to the formulation of robust evolution laws in deforming granular materials under compression and shear: the long held mechanism for energy dissipation of force chain buckling and the sudden switch in roles that a rattler plays as it enters in and out of force chains.

• Received 22 November 2012

DOI:https://doi.org/10.1103/PhysRevE.87.032203