Does fluid structure encode predictions of glassy dynamics?

Understanding and predicting the failure, flow, and rearrangement dynamics of amorphous solids has greatly benefited from data-driven methods for correlating local structures with dynamical features. Some of these approaches, such as the "Softness" method based on linear Support Vector Machines, have uncovered combinations of local structural characteristics that predict energy barriers to particle rearrangements in supercooled fluids based on a particle's local environment. The Softness method also was shown to predict the onset temperature of dynamical heterogeneity by estimating the temperature above which local structures are no longer predictive of dynamical activity. In this talk we implement a transfer learning approach and first show that simple classifiers can be trained to predict dynamical activity even well above the onset temperature. We then demonstrate that applying these classifiers to data from the supercooled phase yields results that are nearly identical to those obtained by softness in terms of the physical information about the relationship between local structures and energy barriers. We further show that the predicted onset temperature is independent of the training temperature, for training temperatures both above and below the onset temperature itself.