Reviews: SchNet: A continuous-filter convolutional neural network for modeling quantum interactions

Summary: In order to design new molecules, one needs to predict if the newly designed molecules could reach an equilibrium state. All possible configurations of atoms do not lead to such state. To evaluate that, the authors propose to learn predictors of such equilibrium from annotated datasets, using convolutional networks. Different to previous works, that might lead to inaccurate predictions from convolutions computed in a discretized space, the authors propose continuous filter convolution layers. They compare their approach with two previous ones on 2 datasets and also introduce a third dataset.