Accelerating material melting temperature predictions by implementing machine learning potentials in the SLUSCHI package

Abstract

The SLUSCHI (Solid and Liquid in Ultra Small Coexistence with Hovering Interfaces) automated package, with interface to the first-principles code VASP (Vienna Ab initio Simulation Package), was developed by us for efficiently determining the melting temperatures of various materials. However, performing many molecular dynamics simulations for small liquid-solid coexisting supercells to predict the melting temperature of a material is still computationally expensive, often requiring weeks and tens to hundreds of thousands of CPU hours to complete. In the present paper, we made an attempt to interface the SLUSCHI package with the highly efficient molecular dynamics LAMMPS code and demonstrated that it achieves a much faster melting temperature determination, outperforming the original VASP-based approach by at least one order of magnitude. In our melting temperature calculations, the LAMMPS simulations were performed based on the LASP (Large-scale Atomic Simulation) machine learning potentials which are pre-built using first-principles data. Besides the dramatic CPU time reduction for melting temperature predictions the calculated melting temperatures for various materials (simple and transition metals, alloys, oxides and carbide) are reasonably accurate. Analysis of the calculated results shows that 60% of the melting temperatures are within 200 K of experimental values with the RMSE value of 187 K which is slightly worse than the first-principles DFT RMSE value of 151 K. Therefore, interfacing SLUSCHI with LAMMPS molecular dynamics simulations makes it possible to quickly screen out the best candidates from numerous materials in a much more efficient way, and facilitate the rational design of materials within the framework of the materials genome paradigm.