“…However, to effectively train an ML model, one has to design proper 'fingerprints' or descriptors for their data (data featurization), as shown in figure 3(b). At the present time, data featurization is mostly based on the chemical composition of alloys [78,87,91] or by translating a chemical composition into empirical parameters guided by the aforementioned empirical rules [66, 71, 74-76, 88-90, 92], such as mean atomic size [74,75,79,87], atomic size difference [71,74,79,90], mean atomic volume [74,90], mixing enthalpy [74,75,79,87], ideal mixing entropy [71,74,75,79,87], mean electronegativity [75,79,87,90], electronegativity difference [71,79], valence electron concentration [71,75,79,87] and calculated density [71,75,87]. If one considers all individual and collective attributes of constituent elements, the number of the data descriptors designed based on the empirical rules could reach 186 [76], which suggests the intrinsic complexity of the ML based design of BMGs.…”