Predictive Modeling in Characterizing Localization Relationships

Predictive modeling in data mining is able to extract information from materials data in order to characterize the "processing-structure-property" relationship. In a localization problem where the purpose is to determine the spatial distribution of the response at the microscale for an imposed loading condition at the macroscale, while executing Finite-Element Models on fully coupled scales is excessive and infeasible, a data mining model that incorporates the bridging knowledge from the microstructure to the response is of great value. In this work, we propose an regression tree-based classifier ensemble method that builds multiple models on selective data volumes. We leverage feature selection methods to rank the localized neighboring features, so that not only redundant features are removed, but hidden values are preserved and included. With 2500 volume samples, we take calibration/validation splits of 500/500, 500/2000, 1000/1000, 1000/1500, and are able to achieve an MASE of 3.74%, 5.86%, 3.25%, 4.22%, relatively.