Machine-Learned Interatomic Potential for the Mo-S-Al-O System

Abstract

We develop a machine-learned neuroevolution potential (NEP) for the Mo-S-Al-O system, targeting MoS$_2$ heteroepitaxy on c-plane sapphire, $α$-Al$_2$O$_3$(0001). The potential is trained on 63,611 PBE+D3 structures distilled from about 1.1 million curated density-functional-theory (DFT) configurations, using an energy-difference loss function that directly targets interfacial and orientation-dependent energies. On held-out data, the energy and force errors are 29 meV/atom and 0.23 eV/$\mathring{A}$. Validated against DFT on a suite of physical properties, the NEP reproduces the lattice constants within 0.3 %, the elastic constants within 17 %, the phonon dispersions with mean deviations below 2.5%, and the surface energies within 1.3 %. At the MoS$_2$/sapphire interface, it reproduces the adhesion energy minimum within 0.8 %, the interfacial energy within 2.2 % and the most stable orientation of the monolayer. It preserves the DFT-predicted energetic ordering of edges, point defects (Spearman $ρ$ = 0.995) and mirror-twin boundaries ($ρ$ = 1.0), and reproduces tilt grain-boundary energies within 8 % ($ρ$ = 0.98) and stacking-fault energy barriers within 14 %. The resulting potential enables future studies of the atomic-scale mechanisms of defect formation in two-dimensional heteroepitaxy.