Learning Structure–Dynamics Relationships in Metallic Glasses from Atomic Configurations Using Machine Learning and DeepSets

Abstract Understanding the relationship between atomic structure and relaxation dynamics is essential for revealing the physical behavior of metallic glasses. In this work, we integrate molecular dynamics simulations, Elastically Collective Nonlinear Langevin Equation theory, and artificial intelligence to investigate both the forward and inverse structure–dynamics relationships in four Zr-based metallic glasses. For the forward prediction, conventional machine learning models and a DeepSets neural network were developed to predict the structural relaxation time from atomic configurations with and without temperature information. Atomic structure alone provides strong predictive capability, while incorporating temperature consistently improves the performance. Among the conventional models, Gradient Boosting achieved the best accuracy with a test R2 of 94.50%, whereas DeepSets further increased the accuracy to 96.35% by directly learning from atom-wise representations. For the inverse problem, machine learning models successfully reconstructed radial distribution functions from structural relaxation time and temperature, reproducing the major structural characteristics of metallic glasses. These results demonstrate the complementary roles of atomic structure and temperature in governing relaxation dynamics and establish an effective bidirectional artificial intelligence framework for studying and inversely designing amorphous materials.