Characterizing Structural Heterogeneity in Metallic Glasses: A Molecular Dynamics-Guided Machine Learning Approach

Abstract

The main objective of this research is to develop a robust Bayesian machine learning (ML) model capable of predicting and characterizing the structural heterogeneity in metallic glasses (MGs). The model is constructed using input data obtained from molecular dynamics simulations of CuZr MGs, encompassing a wide range of alloying compositions and simulation parameters. The ML model utilized crucial output variables: the 2D fractal dimension (with a fractal exponent ranging from 1.55 to 1.81) and correlation function (correlation length spanning from 1.1 to 4.05 nm), demonstrating inverse and direct relationships with the degree of heterogeneity, respectively. The results demonstrate the model's high predictive performance, with accuracy values of 0.9398 for the fractal dimension and 0.9639 for the correlation length. It is noteworthy that the correlation length proves to be a reliable indicator for low to intermediate levels of structural heterogeneity, while the fractal dimension effectively characterizes high-level heterogeneity in MGs. Moreover, the integration of both indicators complements each other in accurately predicting structural heterogeneity. Additionally, the developed ML model showcases its versatility in effectively characterizing MG samples exposed to diverse treatments, such as annealing and rejuvenation processes. © The Indian Institute of Metals - IIM 2023.