Structural Signatures of Mechanically and Thermally Activated Dynamics in Glassy Materials

Structural relaxation in glassy materials is spatially and dynamically heterogeneous, with regions of fast particles coexisting near slow ones when the material evolves through spontaneous thermal activation. In sheared glasses at low temperature, where the activation is purely mechanical, deformation proceeds via localized shear transformations. Substantial efforts have been directed towards establishing whether such local dynamics also has a local structural origin. This talk will discuss two ML based approaches that contribute to this goal. In the first part, we employ the atomic cluster expansion (ACE) to obtain a systematic set of local descriptors, and use them to predict regions of high and low mechanical or thermal activity in the glass via linear regression. We find that for both modes of activation, most of the predictive power is already obtained at the two-body order, while higher order terms provide only modest improvements. Dimensionality reduction is then employed to construct minimal sets of descriptors.<br/><br/>In the second part, we turn to unsupervised learning and construct a Markov-State model by coarse-graining the molecular dynamics trajectories into a low-dimensional feature space using graph neural networks in combination with the variational principle for Markov processes (VAMP). The transition timescale between states is larger than the conventional structural relaxation time, but can be obtained from much shorter trajectories. The learned map of states assigned to the particles reveals a form of heterogeneity, and corresponds to local excess Voronoi volume. These results resonate with classic free volume theories of the glass transition, singling out local packing fluctuations as one of the dominant slowly relaxing features.