Interatomic potential construction with self-learning and adaptive database

Abstract

We have developed an automatic machine learning potential (MLP) construction scheme, the self-learning and adaptive database (SLAD). The sample structures for training are collected by the molecular dynamics simulations using the MLP itself with the aid of the spilling factor for simultaneous error estimation. The utility of the SLAD is demonstrated by applying it for the solid-state ionic conductor, ${\mathrm{Li}}_{2}{\mathrm{B}}_{12}{\mathrm{H}}_{12}$. Starting from the low-temperature $\ensuremath{\alpha}$-phase structure, the MLP is successfully developed, for which the total number of density functional calculations required is only 84. The constructed MLP reproduces well the phase transition to the high-temperature $\ensuremath{\beta}$ phase accompanied by an abrupt lattice expansion, promotion of Li diffusion, and orientational disordering of ${\mathrm{B}}_{12}{\mathrm{H}}_{12}$ complexes. The SLAD allows us the efficient construction of MLPs for complicated systems.