Predicting dynamic heterogeneity in glass-forming liquids by physics-informed machine learning

Abstract

We introduce GlassMLP, a machine learning framework using physics-informed structural input to predict the long-time dynamics in deeply supercooled liquids. We apply this deep neural network to atomistic models in 2D and 3D. Its performance is better than the state of the art while being more parsimonious in terms of training data and fitting parameters. GlassMLP quantitatively predicts four-point dynamic correlations and the geometry of dynamic heterogeneity. Its transferability from small to large system sizes allows us to probe the temperature evolution of spatial dynamic correlations, revealing a profound change with temperature in the geometry of rearranging regions.

Improving the prediction of glassy dynamics by pinpointing the local cage

“…This lower performance of machine-learned algorithms for predicting propensity in the caging regime is not unique to the results of Fig. 1, but has been observed in a variety of studies involving different machine learning methods and different ways to describe the system 13,16,17,32 , although recent, more advanced machine learning methods have improved significantly the correlation in this regime 18,19 .…”

“…A number of recent studies have made significant progress in predicting the dynamic propensity of particles in glassy fluids based on local structural information using a variety of machine learning algorithms 13,[16][17][18][19] . The accuracy of such predictions is typically evaluated using the Pearson correlation coefficient 35 between the predicted and measured dynamic propensities.…”

“…A common approach in glass literature for predicting dynamical behavior in a glassy system is to quench the system to its inherent state (IS) (see e.g. 12,17,36,37 ). The inherent state of a configuration is defined as the local potential energy minimum that one obtains via a rapid energy minimization 38 .…”

“…One of the approaches to probe the apparent discrepancy between structure and dynamics has been the use of machine learning [9][10][11][12][13][14][15][16][17][18][19][20] . By capturing the local structure of particles in terms of parameters and training algorithms to predict the mobility of particles based on these parameters, the idea is that we can learn what aspects of the structure influence the heterogeneous dynamics.…”

“…Since then, several works have improved on this feat, by e.g. using physics-informed parameters as input for a deep neural network 17 , by modifying the loss function of a GNN to also consider relative displacements between pairs of particles 18 , or by designing GNNs that preserve roto-translation equivariance 19 . These works clearly demonstrate that careful consideration of the physics involved can aid in improving the predictive accuracy of these advanced methods.…”