Reviews: Cormorant: Covariant Molecular Neural Networks

The paper is well-written and clearly draws the connection between physical interactions, tensors and the proposed neural network Cormorant. The proposed network is related to earlier work on tensor field networks [Thomas et al] and covariant compositional networks, but presents architectural changes that lead to improved results on the QM9 and MD-17 benchmarks. Confusingly, while the introduction motivates the work by prediction of atomic force fields, only scalar values are predicted in the experiments. This is also part of the definition of Cormorant: "C3. The type of each output neuron is [...] a scalar". This seems not to be compatible with force field predictions, and also some other important chemical properties are vectors (e.g.