Correlating structure topological metrics with bulk composite properties via neural network analysis

Abstract

Given a database of any quantifiable set of cause and effect, machine learning methods can be trained to predict future effects based upon an assumed set of causes. In this paper, neural networks are trained to predict the bulk Young’s modulus and electrical conductivity of a two-phase composite with high material property contrast, based upon a sample’s microstructure. Various structure metrics are used to quantify the topological connectivity and disorder of analytically generated heterogeneous samples. The neural network is trained to predict the Young’s modulus and coefficient of electrical conductivity based upon values calculated for a training set of samples using a finite element model. By repeating the process with various subset of structure metrics we can determine which metrics—or combination of metrics—have the strongest influence in accurately predicting bulk material properties. Not only are neural net predictions of bulk properties in good agreement with calculated values for the 2D two-phase composites, but the insights into which metrics most strongly correlate with these properties (in this case, the connectivity metrics) may help focus the development of improved structure–property relations.