N-Body networks for molecular simulation

We introduce a novel deep learning architecture for learning molecular force fields. Our architecture, called CGnet, “bakes in” the underlying rotational invariance of the problem using the Clebsch-Gordan operator. We apply CGnet to the molecular N-Body problem of learning force-fields from ab-initio molecular dynamics simulations. We find state-of-the-art results when applied to the MD-17 dataset. We also train CGnets on QM9 and find competitive results. Finally, we provide a set of tools, called FastCG, along with an interface with PyTorch and Tensorflow, to allow for fast and efficient calculation of CG products, along with their analytical gradients.