Heterogeneous, nonlinear deformation constitutes a complex system, in the mathematical sense. When models are used to predict the failure of metals or proxies (maximum local strain, fatigue indicator parameters, etc.), it is emergent phenomena that are being predicted [1], [2]. I.e., non-linear interaction of top-down (boundary condition) and bottom-up (nonlinear constitutive response and heterogeneous microstructure) sources of information via a minimum energy pathway. Early models of material failure were empirical. In contrast, multiscale modelling and simulation have been employed to capture the physics that drive metal failure. Importantly, the finite element simulations within these multiscale approaches can exhibit emergent behaviour, albeit as numerical ersatz of real materials. In typical analysis, much multiscale modelling effort is focused on determining parameter values or measuring statistical trends in results. Rarely are simulation results studied as emergent from complex systems. This work argues two points: (1) the origins of material performance-determining emergent phenomena can be traced to independent random sub-scale events, and (2) the emergent weak points may be categorized by their local structure through a local sensitivity analysis. A new application of simulations in exploring materials failure is proposed: identifying critical sub-regions of material and the local boundary conditions associated with them. This will be examined by simulating a simplified microstructure in hundreds of configurations, measuring extrema, and studying the number and location of random events upon which they depend. A simplified two-phase isotropic media is studied using the Chaboche viscoplastic constitutive model in finite element simulations. Maximum strain is studied in grain-wise variations of phase. The sensitivity of the maximum strain to the phase identity in each grain of the simulated volume is measured. From these sensitive regions, critical microstructure arrangements and local boundary conditions can be derived for use in lower-length-scale simulations of strain localization.