Identifying the structural origin of soft spots in prism-dominated Mg₆₅Cu₂₅Y₁₀ metallic glasses via SOAP-driven machine learning

Mg-based metallic glasses (MGs) are typically brittle and exhibit prism-dominated local order, unlike the more ductile Cu-Zr MGs built on icosahedral backbones. Elucidating the structural origin of plastic susceptibility in such glasses remains challenging because conventional low-dimensional descriptors cannot capture the coupled effects of local topology and chemical environment. Here we develop a Smooth Overlap of Atomic Positions (SOAP)-driven machine-learning framework to predict plasticity-prone “soft spots” in modeled Mg₆₅Cu₂₅Y₁₀ MG directly from the undeformed atomic structure. Using optimized high-dimensional SOAP descriptors, a sigmoid-calibrated logistic-regression classifier achieves strong predictive performance (ROC-AUC = 0.913), clearly outperforming representative low-dimensional descriptors based on local chemistry and fivefold-symmetry-related order. These results indicate that plastic susceptibility in the modeled Mg-Cu-Y glass is governed by coupled topological-chemical features, which are more effectively encoded in the high-dimensional SOAP representation. This SOAP-driven framework provides a robust structural fingerprint for locating shear-transformation-susceptible regions in amorphous alloys.