Development of a neuroevolution machine learning potential of Pd-Cu-Ni-P alloys

Abstract

Pd-Cu-Ni-P alloy is an ideal model system of metallic glass known for its exceptional glass-forming ability. However, few correlation of structures with properties was systematically investigated owing to a lack of interatomic potential. In this work, a neuroevolution machine learning potential (NEP) with efficiency close to embedded atom method (EAM) potentials is developed. Its accuracy has been compared to density functional theory (DFT) calculations. For energy, force and virial, the training errors are 6.0 meV/atom, 111.1 meV/Å and 21.5 meV/atom, respectively. By means of this NEP, several thermodynamic parameters such as glass transition temperatures and pair distribution functions of Pd40Cu30Ni10P20 and Pd40Ni40P20 liquid and glassy alloys as well as their short-range orders, tensile and compression strengths, transport properties etc. have been evaluated by a series of molecular dynamics simulations. A good agreement with DFT calculations and previous experiments indicates this NEP provides an accurate and efficient scheme in the analysis and exploration of microstructures, thermodynamic and kinetic properties of Pd-Cu-Ni-P alloys.