Title: Unraveling intricate processes of glassy dynamics from static structure by machine learning relative motion

Abstract: Glasses, a rigid state realized by rapidly quenching a liquid, are characterized by their random and featureless structure, which is markedly distinct from crystals. Identifying how the structural changeover in a glass-forming liquid proceeds in space and time is one of the major problems. Recent progress in machine-learning technologies provides us highly accurate prediction of such glassy dynamics by using the data-driven approach. In this article, we show that predictions of intricate dynamic processes are made possible by incorporating relative motion between particles as the target feature of learning in a graph neural network. Consequently, our model can achieve unprecedented precision in predicting glassy dynamics in the short time, implicating the underlying mechanism of glassy structural relaxation. Our model may open new paths to versatile applications in molecular simulations.