Uncertainty quantification of machine learning potentials for atomistic simulation

Atomistic simulation plays a growing important role in material design and analysis, owing to accurate first principles methods that are free from empirical parameters and phenomelogical models. As the successful implementation of atomistic material simulations largely depends on the availability of efficient and accurate force field potentials, machine learning interpolation of interatomic potential energy surfaces has been developed in recent years due to its ourstanding performance in enabling automatic construction of highly accurate atomic interaction potentials. These methods circumvent the parameterization problem in conventional empirical potentials by employing machine learning techniques to interpolate the first principle potential energy surface based on a set of reference calculations. Despite the advantages such as a well automated process and high flexibility of model functions, the uncertainty involved in the machine learning interpolated atomic potential energy surface has not been investigated. In this paper, the uncertainty quantification study is performened for the machine learning interpolated atomic potential for titanium dioxide (TiO2), an industrially relevant and well-studies material.