Identifying flow units by machine learning in a model metallic glass

Characterizing and predicting the flow units directly from the atomic structure are longstanding challenges in metallic glasses. We report the successful identification of flow units in the model Zr₅₀Cu₅₀ metallic glass above and below its glass transition temperature by machine learning methods. We find that the differences of the structural characteristics between flow units and the rest of the system are beyond short-range order, and further confirmed by the local structural entropy. Our study demonstrates that machine learning provides an unconventional tool to understand the intrinsic heterogeneities in metallic glasses, and sheds light on the structural indicator of heterogeneous dynamic behaviors in amorphous solids.