Accumulative coarse-graining of simple structural descriptors enables accurately predicting the dynamics of metallic liquids

Abstract

The typical dynamic mode of liquids is often marked by a long-time α relaxation which is accompanied by heterogeneous dynamics. However, the explicit structural origin of the α relaxation is ongoing debated. Recent advances in graph neural networks (GNN) have shown promise in this area, with an intuitive approach that can be simplified as an iterative coarse-graining (CG) to update atomic features based on neighboring atoms. Here, we propose a new CG strategy called accumulative CG, which aggregates the spherically averaged and deviated features to generate a robust representation of atomic arrangement over a broad range. This can enhance the direct visibility of the center atom among all its surrounding atoms, as opposed to being limited to its nearest-neighbors and relying on iterative message passing to indirectly observe further neighbors. Starting with two descriptors (local energy and volume), the augmented features derived from accumulative CG along with machine learning (ML) can accurately predict the atomic-level dynamics of multi-component liquids, as indicated by the inverse of α-relaxation time. We performed atomistic simulations of multi-component Al–Fe-(Si, Y) liquids with a set of deep-learning interatomic potentials to demonstrate the feasibility of the accumulative CG approach. By injecting physical inspirations from liquid dynamics theories into data-driven ML model, we interpret the structural origin of dynamic heterogeneity, particularly the bimodal behavior of Si atoms in Al–Fe–Si liquids.