Materials informatics of woven fabric composites: Effect of different dimensionality reduction and learning methods

Abstract

A microscale data-driven framework, under the emerging Materials Informatics paradigm, is developed to build structure-property linkages of woven composites. To show the approach, the microstructure of a set of glass fiber-reinforced polymer samples are imaged using micro-computed tomography, with the property of interest being the elastic modulus obtained from tensile tests. After data collection, fingerprints of microstructural images are calculated by means of 2-point spatial correlations. Different dimensionality reduction algorithms are next employed and compared to build reduced-order representations of the ensuing heatmaps and subsequently identify the dominant features for each composite sample through the Least Absolute Shrinkage and Selection Operator (LASSO) regularization. The tested dimensionality reduction algorithms include Principal Component Analysis (PCA), Non-negative Matrix Factorization (NMF), and Partial Least Squares (PLS). Using the dominant features, different structure-property surrogate models are built and compared via a set of common machine learning techniques: Multiple Linear Regression (MLR), Random Forest Regression (RFR), Support Vector Regression (SVR), and Gaussian Process Regression (GPR). In addition, a Linear Mixed-effect Model (LMM) is developed to account for potential random effects taking place during the fabrication of samples. In this case study, the best performing models were RFR and PLS, with the latter containing components well correlated with the material geometrical attributes such as fiber orientation and fiber volume fraction. This preliminary Composites Informatics case study would furnish the path for a non-destructive property-microstructure prediction of woven fabrics, as well as data-driven discovery and design of new reinforcements.