Unsupervised machine learning study on structural signature of glass transition in metallic glass-forming liquids

Abstract

As a liquid is quenched into glassy state, characterizing the glassy order responsible for glass transition in disordered liquid structures is a longstanding challenge. Here, an unsupervised machine learning method, affinity propagation clustering was developed, which is able to automatically classify disordered structures by using purely atomic distances in disordered structures at a given temperature in training set, without any prior input of local symmetry, atomic packing, or dynamical information. Surprisingly, the clustering models successfully identify the liquid- and glass-like atoms in metallic glass-forming liquids and predict their temperature evolution in whole cooling process. This demonstrates that a liquid or glass structure possesses all atomic structure information from high-temperature liquid states to low-temperature glassy states. Moreover, the glassy order formed by glass-like atoms increases with a power-law form followed by a linear one below a crossover point which is quite close to glass transition temperature. Meanwhile, the atomic structures of glassy order percolate in glass transition. These results manifest the structural signature of glass transition. Furthermore, the liquid- and glass-like structural characteristics show excellent correlation with properties in metallic liquids and glasses, such as structural relaxation, shear modulus and dynamic propensity. Our results provide a new machine-learning strategy for characterizing disordered structures and unraveling structure-property relationship in glass-forming systems.