Fig. 5
Quantitatively benchmarking individual feature sets. a Quantitatively benchmarked predictive powers of individual short-range order (SRO) feature sets and the feature sets further augmented by the medium-range order (MRO) features generated following the coarse-graining technique described in this work. Please refer to Methods and Supplementary Tables 4 and 6 for a full description of the feature sets. Results for Cu65Zr35, Cu50Zr50, Cu80Zr20, Ni62Nb38, Al90Sm10, and Fe80P20 MGs (quenched under 5 × 1010 K s−1) are shown. In the box plots, bounds of the box spans from 25% to 75% percentile, dashed line represents median, and whiskers show minima and maxima of data points. b Projecting <0,0,12,0,0>, <0,0,12,4,0>, and Frank–Kasper clusters to the two-dimensional (2-D) partial-dependence plot (PDP) of the top two features of the ML model. Conventional descriptors only capture a small fraction of the possible input space, whereas our features and ML form a more complete description of feature space.
