How important is microstructural feature selection for data-driven structure-property mapping?

Abstract: Data-driven approaches now allow for systematic mapping of microstructure to properties. In particular, we now have diverse approaches to “featurize” microstructures, creating a large pool of machine-readable descriptors for subsequent structure-property analysis. We explore three questions in this work: (a) Can a small subset of features be selected to train a good structure-property predictive model? (b) Is this subset agnostic to the choice of feature selection algorithm? And (c) can the addition of expert-identified features improve model performance? Using a canonical dataset, we answer in the affirmative for all three questions. Graphical abstract: [Figure not available: see fulltext.]