A physics-informed machine learning perspective to present the structures and properties of titanium matrixes and nanoclusters through atomic modeling

Abstract

A machine learning potential within the framework of an artificial neural network model was developed to describe the interactions among the atoms of titanium bulks and nanoclusters, wherein atomic simulations were used to present their structures and mechanical properties. The developed machine learning potential, which was trained on extensive first-principles datasets, demonstrated remarkable accuracy in predicting various lattices, elastic constants, and defect properties, along with high-temperature characteristics, including α–β structural transition, thermal expansion, and melting point for titanium matrixes. The generalised stacking fault energy lines and surfaces on multiple slip planes were used to compare the MLP performance with other potential models in assessing the material mechanical properties. The atomic-level stress maps were used to describe the atomic stress characteristics of five twin boundaries. Molecular dynamics simulations were used to present the lattice evolution of Ti bulks under high pressure at room temperature and the structural transition of titanium nanoclusters during heating. The pair analysis technique was used to describe the local packing of the atoms in the titanium nanoclusters.