Research on Cu-Sn machine learning interatomic potential with active learning strategy

Abstract

Cu-Sn alloy materials are widely used in electronic industry, aerospace and 3D printing. When studying the structure and properties of materials, a contradiction between arithmetic and accuracy is encountered by Molecular dynamics (MD). Molecular dynamics is a general theoretical calculation method for studying the mechanical properties of alloy materials. However, molecular dynamics simulations of alloy materials are limited to simple systems because the construction of traditional interatomic potentials is replicative and inefficient. In this study, the Cu-Sn material machine learning interatomic potential was constructed by using a deep neural network model, and the first-principles calculation results were used as the training data set to ensure the accuracy of the quantum mechanical interatomic potential. This process features an “active learning” process, data generation and model training methods with minimal human intervention. Molecular dynamics using machine learning interatomic potentials (MLIP) and first-principles calculations show good consistency. It was concluded that MLIP can accurately predict energy, force and mechanical properties. The root mean square error (RMSEs) of the energy and force per atom is approximately 10 meV/atom and 100 meV/Å. It shows good advantages in energy-volume curve, phase transition temperature and elastic modulus, laying the foundation for the wide application of the MD method in the design and development of Cu-Sn alloy materials.