PIL-Net: a physics-informed graph convolutional network for predicting atomic multipoles

Abstract

We introduce PIL-Net, a physics-informed graph convolutional network capable of predicting molecular properties quickly and with low error, using only basic information about each molecule's atoms and bonds. The PIL-Net model combines the representational power of graph neural networks with domain knowledge to predefine a set of constraints that force the network to make physically consistent predictions; this leads to faster model convergence. We apply PIL-Net to the task of predicting atomic multipoles, which describe the charge distribution within an atom. Atomic multipoles have several applications, including their incorporation into force fields for molecular dynamics simulations. We emphasize our model's ability to predict atomic octupoles, a higher-order atomic multipole property, with a mean absolute error of only 0.0013 eų, more than an order of magnitude less than results reported in the literature. Moreover, our framework can approximate molecular multipole moments post-training with little additional cost. Finally, we elaborate on how our network can be used for greater model interpretability, reconstruction of the molecular electrostatic surface potential, and prediction on out-of-domain datasets.