Title:Chemical Properties from Graph Neural Network-Predicted Electron Densities

Abstract:According to density functional theory, any chemical property can be inferred from the electron density, making it the most informative attribute of an atomic structure. In this work, we demonstrate the use of established physical methods to obtain important chemical properties from model-predicted electron densities. We introduce graph neural network architectural choices that provide physically relevant and useful electron density predictions. Despite not training to predict atomic charges, the model is able to predict atomic charges with an order of magnitude lower error than a sum of atomic charge densities. Similarly, the model predicts dipole moments with half the error of the sum of atomic charge densities method. We demonstrate that larger data sets lead to more useful predictions in these tasks. These results pave the way for an alternative path in atomistic machine learning, where data-driven approaches and existing physical methods are used in tandem to obtain a variety of chemical properties in an explainable and self-consistent manner.