Dynamic heterogeneity at the experimental glass transition predicted by transferable machine learning

Dynamic heterogeneities represent a critical aspect of glassy dynamics, and their analysis constitutes a fundamental step in understanding the essence of the glass transition. The authors employ physics-inspired machine learning to predict the microscopic structural relaxation from amorphous configurations of deeply supercooled liquids. Leveraging the transferability in temperature of the trained networks enables the prediction of dynamic heterogeneities at the experimental glass transition temperature, thereby circumventing the prohibitively large computational cost associated with conventional simulation methods at such low temperatures.