Fig. 3

Distribution of machine-learnt quench-in softness in metallic glasses. a Wide spread in the distribution of quench-in softness (QS) in Cu₆₅Zr₃₅ defines a spectrum of atomic environments within a glass. b The probability that an atom rearranges as a function of QS, P(plastic|QS), is a strong function of QS, increasing by several orders of magnitude, from hardest atoms at low QS end to softest atoms with high QS. c Fractal dimensionality sampling in a log–log plot for the hardest (QS < 0.1) and softest atoms (QS > 0.7) as well as all atoms, based on the power-law scaling of the mass distribution M(r) ~ r^D, where M(r) denotes the number of atoms within radius r centered by an atom and slope D is the dimensionality. The inset shows the pair correlation function functions g(r) of the hardest and softest atoms. d Violin plots of representative interstice distributions of the hardest and softest atoms, manifesting strong site environment contrasts between these two groups of characteristic atoms. e QS of the plastic atoms activated under a certain strain. The entire deformation range up to the strain of 10% with an interval of 0.125% is analyzed. The crossover between nominal elastic stage (plasticity highly dependent on high-QS atoms), shear banding and shear band propagation stages (plastic rearrangement extending across all levels of QS) is clearly shown and marked. f QS distributions of Cu₅₀Zr₅₀ @ 5 × 10¹⁰ K s⁻¹, Cu₈₀Zr₂₀ @ 5 × 10¹⁰ K s⁻¹, Cu₆₅Zr₃₅ @ 5 × 10¹¹ K s⁻¹, and Cu₆₅Zr₃₅ @ 5 × ¹0¹² K s⁻¹ to reveal the effects of composition or thermal history. By varying these parameters, one can in principle tailor the site-specific plastic response of a MG.