Graph neural network-driven inverse design of relaxation dynamics in metallic-glass formers

RQ Zhao and MZ Li and ZL Ning and JF Sun and YJ Huang, JOURNAL OF NON- CRYSTALLINE SOLIDS, 668, 123773 (2025).

DOI: 10.1016/j.jnoncrysol.2025.123773

Despite the development of numerous theories to explain the structural origins of long-time relaxation dynamics in amorphous materials, a unified theory remains elusive. Unlike crystalline materials, glasses and supercooled liquids lack long-range orders or structural units that are closely correlated with their properties. Researchers face challenges in pinpointing specific features and the intricate dynamic relaxation behaviors spanning diverse time scales. In this study, we focused on investigating the structural origin of relaxation dynamics with two advanced methods: graph neural network (GNN) and swap Monte Carlo. The swap Monte Carlo method facilitated extensive sampling of supercooled liquid configurations, while GNN was employed to evaluate changes in their dynamic propensities induced by structural changes. Our observations revealed a short-to-medium-range ordered structure within CuZr supercooled liquids. Notably, we found that configurations exhibiting lower peaks and higher valleys in the radial distribution function, (r), within a certain distance range corresponded to greater dynamic propensities. Among the numerous structural descriptors, those that represent local five-fold symmetry in CuZr supercooled liquids exhibit the strongest correlation with relaxation dynamics. This targeted adjustment paradigm not only serves as a versatile tool for tailoring the structure of supercooled liquids but also provides interpretable insights into the complex dynamics of relaxation in glass formers.

Return to Publications page