Title: Predicting precursors of slow glassy dynamics with high precision by machine learning relative motion

Abstract: The recent developments of machine learning have enabled accurate predictions of heterogeneous dynamics of glassy dynamics. However, existing machine-learning models are designed so that they learn a single dynamic quantity and relate it to the structural features of glassy liquids. In this Letter, we propose a new graph neural network model which learns relative motion between neighboring pairs of particles, in addition to self motion of particles. By relating the structural features to these two different dynamical variables, the model autonomously acquires the ability to discern how two different types of dynamic processes, strain fluctuations and particle rearrangements, affect the self motion of particles undergoing fast beta relaxation. Our model affords a methodology to treat these two at the same time and realizes highly precise predictions of rearranging hotspots in glassy systems.