A systematic approach to understanding the role of microstructure on the mechanical response of complex multiphase materials is presented based on the "microstructure reconstruction" method of Yeong and Torquato. Low-order statistical correlation functions from an experimental microstructure are used to generate, using a simulated annealing algorithm, a spectrum of statistically similar microstructures that are then tested numerically. To illustrate the method, a ductile iron microstructure is studied. All simulated structures have elastic behavior and average stress-strain behavior nearly identical to that predicted for the experimental microstructure. The structures differ in the strain at which nucleation of localization begins. To identify the characteristic critical "hot spot" driving localization in each specific microstructure, a new variant of the microstructure reconstruction method is proposed in which a potential "hot spot" region from a particularly weak microstructure is embedded into new statistical microstructures, which are then tested. If the selected "hot spot" region is large enough, the new microstructures fail in the same "hot spot" region and at comparable applied stress and strain levels. This method thus shows promise for capturing local failure initiation sites in complex microstructures. (C) 2005 Elsevier Ltd. All rights reserved.