Mapping local atomic structure of metallic glasses using machine learning aided 4D-STEM

Amorphous materials such as polymers, metallic and oxidic glasses consist of heterogeneous atomic/molecular packing at the nanoscale. Spatial variation of the local structure plays an important role in determining material properties. Experimentally probing the local atomic structure within the amorphous phase has been one of the main challenges for material research. Here, we present a new approach to characterize the local atomic structure and map structural variants in the amorphous phase using machine learning (ML) aided four dimensional-scanning transmission electron microscopy (4D-STEM). We utilized nonnegative matrix factorization (NMF) to identify the local structural types of metallic glasses in 4D-STEM datasets. Using Fe-based metallic glasses as a model system, we demonstrate that two basic structural types, one with a more liquid-like and another with a more solid-like structure, are distributed throughout the glass with a characteristic length scale of a few nanometers. Thermal annealing induces a change in their distribution and relative population but without the appearance of any additional phase. This provides new insights into the relaxation phenomena of metallic glasses and solid experimental evidence for the theoretical hypothesis on atomic packing in glassy structures.