Autonomously revealing hidden local structures in supercooled liquids

Abstract: Few questions in condensed matter science have proven as difficult to unravel as the interplay between structure and dynamics in supercooled liquids and glasses. The conundrum: close to the glass transition, the dynamics slow down dramatically and become heterogeneous while the structure appears largely unperturbed. Largely unperturbed, however, is not the same as unperturbed, and many studies have attempted to identify "slow" local structures by exploiting dynamical information. Nonetheless, the question remains open: is the key to the slow dynamics imprinted in purely structural information? And if so, is there a way to determine the relevant structures without any dynamical information? Here, we use a newly developed unsupervised machine learning (UML) algorithm to identify structural heterogeneities in three archetypical glass formers. In each system, the UML approach autonomously designs an order parameter based purely on structural variation within a single snapshot. Impressively, this order parameter strongly correlates with the dynamical heterogeneity. Moreover, the structural characteristics linked to slow particles disappear as we move away from the glass transition. Our results demonstrate the power of machine learning techniques to detect structural patterns even in disordered systems, and provide a new way forward for unraveling the structural origins of the slow dynamics of glassy materials.