On-the-Fly Bayesian Active Learning of Interpretable Force-Fields for Atomistic Rare Events

3 Apr 2019  ·  Jonathan Vandermause, Steven B. Torrisi, Simon Batzner, Alexie M. Kolpak, Boris Kozinsky ·

Machine learning based interatomic potentials currently require manual construction of training sets consisting of thousands of first principles calculations and are often restricted to single-component and nonreactive systems. This severely limits the practical application of these models due to both low training efficiency and limited accuracy in treating important rare events such as reactions and diffusion. We present an adaptive Bayesian inference method for automating and accelerating the on-the-fly construction of accurate interatomic force fields using structures drawn from molecular dynamics simulations. Within an online active learning algorithm, the internal uncertainty of a Gaussian process regression model is used to decide whether to accept the model prediction or to perform a first principles calculation to augment the training set of the model. The method is applied to a range of single- and multi-component systems and shown to achieve state-of-the-art accuracy with minimal ab initio data.

PDF Abstract

Categories


Computational Physics Materials Science