Accelerating melting temperature predictions by leveraging LASP machine learning potentials in the SLUSCHI package

A Campbell and LG Wang and Q- Hong, JOURNAL OF THE AMERICAN CERAMIC SOCIETY, 109, e70398 (2025).

DOI: 10.1111/jace.70398

The automated computational package SLUSCHI, originally interfaced with the first-principles package VASP, has demonstrated effectiveness but remains computationally demanding for accurately calculating melting temperatures. This study leverages machine learning potentials via the efficient molecular dynamics simulator LAMMPS, utilizing pre-trained LASP neural network potentials derived from first-principles data. Tests on 30 diverse material systems-including simple metals, transition metals, alloys, oxides, and carbides-demonstrate that this approach significantly cuts computational costs, often by more than one order of magnitude compared to conventional DFT simulations. Approximately 60% of the calculated melting temperatures, prior to applying any DFT-based correction, fall within 200 K of experimental values. Focusing specifically on single-element systems, where direct comparison with DFT is possible, the percentage of melting temperatures within 200 K of experimental data improves from 53% to 82% following a DFT correction. This substantial improvement in computational efficiency, without sacrificing accuracy, facilitates high-throughput materials screening and accelerates material design consistent with the materials genome paradigm.

Return to Publications page