Self-Consistent Determination of Long-Range Electrostatics in Neural Network Potentials (2109.13074v1)

Abstract: Machine learning has the potential to revolutionize the field of molecular simulation through the development of efficient and accurate models of interatomic interactions. In particular, neural network models can describe interactions at the level of accuracy of quantum mechanics-based calculations, but with a fraction of the cost, enabling the simulation of large systems over long timescales with ab initio accuracy. However, implicit in the construction of neural network potentials is an assumption of locality, wherein atomic arrangements on the scale of about a nanometer are used to learn interatomic interactions. Because of this assumption, the resulting neural network models cannot describe long-range interactions that play critical roles in dielectric screening and chemical reactivity. To address this issue, we introduce the self-consistent field neural network (SCFNN) model -- a general approach for learning the long-range response of molecular systems in neural network potentials. The SCFNN model relies on a physically meaningful separation of the interatomic interactions into short- and long-range components, with a separate network to handle each component. We demonstrate the success of the SCFNN approach in modeling the dielectric properties of bulk liquid water, and show that the SCFNN model accurately predicts long-range polarization correlations and the response of water to applied electrostatic fields. Importantly, because of the separation of interactions inherent in our approach, the SCFNN model can be combined with many existing approaches for building neural network potentials. Therefore, we expect the SCFNN model to facilitate the proper description of long-range interactions in a wide-variety of machine learning-based force fields.