A highly interpretable materials informatics approach for predicting microstructure-property relationship in fabric composites

Abstract Multiscale properties of fabric-reinforced composites are commonly modeled via numerical and experimental methods, which are often highly time-consuming and complex. In this paper, a materials informatics-based approach has been developed to link the micro/meso-level features of woven fabric composites, obtained via nondestructive micro-CT images, to their effective Young's moduli. To this end, a reduced-order quantification of a typical glass fiber/polypropylene lamina's microstructure is established using two-point spatial correlations and the principal component analysis. Next, a machine-learning model is implemented to predict the material microstructure-property (modulus) relationship via the captured images. Despite the limited number of samples, the presented data-driven techniques led to a model with highly interpretable components and excellent accuracy for different ply orientations, regardless of apparent uncertainties such as waviness. The findings appear to be a promising step forward for the potential use of materials informatics for smart design and optimization of woven fabric composites in prominent industries, including aerospace and transportation.