Improving the prediction of glassy dynamics by pinpointing the local cage

Abstract

The relationship between structure and dynamics in glassy fluids remains an intriguing open question. Recent work has shown impressive advances in our ability to predict local dynamics using structural features, most notably due to the use of advanced machine learning techniques. Here, we explore whether a simple linear regression algorithm combined with intelligently chosen structural order parameters can reach the accuracy of the current, most advanced machine learning approaches for predicting dynamic propensity. To achieve this, we introduce a method to pinpoint the cage state of the initial configuration—i.e., the configuration consisting of the average particle positions when particle rearrangement is forbidden. We find that, in comparison to both the initial state and the inherent state, the structure of the cage state is highly predictive of the long-time dynamics of the system. Moreover, by combining the cage state information with the initial state, we are able to predict dynamic propensities with unprecedentedly high accuracy over a broad regime of time scales, including the caging regime.

Rotation-equivariant graph neural networks for learning glassy liquids representations

“…We can compare these observations with the findings of Alkemade et al [40]. They identify three physical quantities, each being relevant in a given time range: 1.…”

“…It has been observed several times that pre-processing the input positions by quenching them to their corresponding Inherent Structures (IS) helps most Machine Learning models in predicting long-time mobility measures [26,39,40]. Such a quench is performed using the FIRE algorithm: temperature is set to 0 (velocities set to 0), and positions adjust gradually so as to converge to a local minimum of the potential energy, typically close to the original configuration.…”

“…The features consist for a part of node-wise rotation-invariant descriptors, and for another, of these elementary features but averaged over neighboring nodes, over a small radius (mimicking a GNN aggregation step). There is currently some debate [15,[37][38][39][40][41] over whether Deep Learning is the right way to do better than such expert features, and part of the community leans on this expert side.…”

Dead or alive: Distinguishing active from passive particles using supervised learning (a)

“…structural indicators that predict dynamical properties in densely disordered (near-)equilbrium systems [6][7][8][9][10][11][12][13][14][15][16][17][18][19][20][21][22][23][24][25]. These findings firmly establish a correlation between local structure and the propensity of passive particles to move in a crowded environment.…”

Toward Interpreting the Thermally Activated β Dynamics in Metallic Glass Using the Structural Constraint Neural Network

“…In the past five years, both supervised and unsupervised algorithms have been extensively applied to predict dynamic properties in amorphous alloys from the structure, covering aspects such as athermal local structural deformation, − activation energy, long-time diffusion, , and so on. − These models take geometric features as input, such as radial symmetry functions, interstice distribution, or solely atom position information. − After training the model with the dynamic label, we can establish a direct relationship between structure and dynamics. Other works attempt to incorporate complementary physical parameters such as the thermodynamic vibrational entropy and kinetic features ,− to improve the prediction performance. In these approaches, machine learning algorithms such as Support Vector Machine (SVM), Gradient-boosted Trees Model (GBDT), Linear regression, and Graph Neural Network (GNN) are commonly employed for building correlations in a supervised manner.…”