Molecular Dynamics Simulation of the Immiscibility in Bi–Ga Melts

Abstract—We are the first to analyze the immiscibility in Bi–Ga melts by molecular dynamics simulation. Interatomic interaction is specified using ab initio data parameterized neural network potential (DeePMD model). The DeePMD potential is parameterized using an active machine learning algorithm. In the course of molecular dynamics simulation, melts of the GaxBi100 – x (x = 0, 10, …, 90, 100) compositions are cooled from 800 to 300 K. Separation is detected using a changes in the temperature behavior of the partial radial distribution function of the Ga–Bi pair. The DeePMD potential, the initial training set of which has no configurations corresponding to a separated state, is still able to reproduce the miscibility gap in the Bi–Ga system. The concentration range of the miscibility gap determined by molecular dynamics simulation with the DeePMD potential coincides with the experimental data. The shift in the maximum of the immiscibility dome toward melts rich in gallium is also correctly described. Nevertheless, the maximum of the immiscibility dome is not determined correctly enough, specifically, Ga80Bi20 instead of experimental Ga70Bi30. In addition, the determined temperature range of the immiscibility dome is wider than the experimental one. Nevertheless, the use of neural network potentials in atomistic simulation is shown to be effectively used to predict a miscibility gap in binary metal systems.