Development of a Machine Learning Interatomic Potential for Zirconium and Its Verification in Molecular Dynamics

Molecular dynamics (MD) can dynamically reveal the structural evolution and mechanical response of Zirconium (Zr) at the atomic scale under complex service conditions such as high temperature, stress, and irradiation. However, traditional empirical potentials are limited by their fixed function forms and parameters, making it difficult to accurately describe the multi-body interactions of Zr under conditions such as multi-phase structures and strong nonlinear deformation, thereby limiting the accuracy and generalization ability of simulation results. This paper combines high-throughput first-principles calculations (DFT) with the machine learning method to develop the Deep Potential (DP) for Zr. The developed DP of Zr was verified by performing molecular dynamic simulations on lattice constants, surface energies, grain boundary energies, melting point, elastic constants, and tensile responses. The results show that the DP model achieves high consistency with DFT in predicting multiple key physical properties, such as lattice constants and melting point. Also, it can accurately capture atomic migration, local structural evolution, and crystal structural transformations of Zr under thermal excitation. In addition, the DP model can accurately capture plastic deformation and stress softening behavior in Zr under large strains, reproducing the characteristics of yielding and structural rearrangement during tensile loading, as well as the stress-induced phase transition of Zr from HCP to FCC, demonstrating its strong physical fidelity and numerical stability.