Open Force Field

@openforcefield.org

97 Followers 19 Following

Overview of the AshGC charge model. A molecule is first checked against a look-up table comprised of molecules with three or fewer heavy atoms. If found, the partial charges in that look-up table are returned. If not found, the molecule is converted into a featurized graph and passed through the neural network model. In the first stage, a graph convolutional neural network generates atom embeddings, which are then pooled and passed to a multi-layer perceptron (MLP) to predict an initial charge, electronegativity, and hardness value for each atom. From these, the final partial charges are computed as the analytical minimum of the electrostatic energy.
Left to right: Mike Henry, Alyssa Travitz, Irfan Alibay, Jennifer Clark, Jeffrey Wagner, Hugo McDermott-Opeskin
14 people standing in a group on a patio, looking happy to be together.