“…Although a number of options exist for dimensionality reduction, PCA was chosen here because it offers the following benefits: (i) it is a distance-preserving transformation, which allows a highly accurate and low-cost computation of a difference measure between any two microstructures using just the low-dimensional representations, (ii) it provides an orthogonal basis for representing the microstructure statistics which should lead to robust representations of process-structure-property (PSP) linkages, (iii) easy access to highly efficient computational toolsets for computing PCA on large datasets [67,71,72], (iv) a remarkable ability to recover the original high-dimensional microstructure statistics with only a handful of PC scores as long as the eigenvectors found in the PCA are stored [47], and (v) prior success in establishing robust PSP linkages in a wide range of multiscale materials phenomena [47,73,74]. Consequently, for each microstructure indexed by m, its feature vector (set of three chord length distributions representing a total of 503 chord length statistics) denoted by CLD m can be approximately decomposed into a linear combination of basis vectors (called principal components) and weights (i.e., PC scores) [74] such that…”