The development of accurate and transferable interatomic potentials for molecular simulations is a challenging task. The most promising potential energy functions are based on the many-body expansion of the interaction energy, combining a physically motivated representation of long-range interactions with machine learning techniques that accurately represent the complex quantum mechanical short-range interaction energy terms. Here we present an open source software platform for the routine development of such high-quality many-body potentials from electronic structure reference data. Our software integrates a centralized database storage with workflow tools for generating data sets of representative molecular configurations, automating electronic structure calculations on high-performance and distributed computing platforms, and supervised learning of many-body potentials. In order to generate robust training sets from the complex molecular configuration space we employ active learning frameworks that minimize the training data set size while maximizing accuracy. We demonstrate the effectiveness of our software with representative examples ranging from molecular liquids, such as formamide, to ionic liquids and solvation of small gas molecules like N<sub>2</sub>O<sub>5</sub> in water. In particular, we present the numerical accuracy of the many-body potentials with respect to reference electronic structure calculations as well as results for dynamic and thermodynamic properties from condensed phase molecular dynamics simulations.

Automated machine learning of accurate many-body potentials for molecular simulations

2019 Fall National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.