The MB-pol many-body potential has been shown to be an accurate molecular model for water simulations across different phases and environments. Its success is based on the synergy between interatomic potentials that are formulated in terms of a rigorous many-body expansion, machine learning techniques that accurately represent short-range interaction energy terms, and highly accurate reference data from CCSD(T) electronic structure calculations. Here we employ the MB-pol framework in combination with reference data obtained with a range of popular density functionals for the development of analytical many-body potential energy functions. We discuss the performance of permutationally invariant polynomials and neural networks in representing water two-body and three-body interaction energies in this setting. We also provide a detailed analysis of individual DFT many-body contributions to the interaction energy and various physical properties of water across different phases. By means of disentangling errors that are introduced at different orders of the many-body expansion we uncover the extent of error compensation that is responsible for the success and failure of different density functionals in reproducing the properties of water.

DFT a la carte – many-body potentials employing machine learning techniques for water

256th National Meeting (2018)

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