A Machine Learning Study on the Underlying Structural Origin of Glassy Dynamics

Finding an underlying structural origin of glassy dynamics is a challenging problem in the field of the glass transition. In this study, we construct a convolutional neural network (CNN) machine learning model that classifies liquids and glasses of two-dimensional (2D) colloidal suspensions only with structural information. We employ the machine learning model to investigate whether any structural origin would exist for the glass transition. 2D colloidal suspensions are an excellent testbed because the hexatic medium-range crystalline order (MRCO) exists for the 2D polydisperse colloids (2DPCs) while the MRCO is not observed for the 2D binary colloids (2DBCs). We find that when the machine learning model is given only snapshots of 2D colloidal suspensions, the machine learning classifies the suspensions into liquids and glasses, successfully. This indicates that one can employ only structural information to understand the states of the suspensions. More interestingly, the machine learning model trained with only 2DBCs that lacks MRCO can also classify the suspensions of 2DPCs with MRCO structures. This shows that the hexatic MRCO would not be the structural origin of 2DPC.