E(n) Equivariant Normalizing Flows for Molecule Generation in 3D

Mapping molecules to hidden vector space representations is important, since people can utilize this representation to do many tasks, e.g. discovering new molecules, find molecules with similar property etc. There are already a series of methods to encode molecule into numerical vectors, using e.g. GAN and GCN. Despite these generative networks, there are a special kind of generative networks which can be utilized to generate molecule structures starting from random hidden representations, i.e. Normalizing Flow in this paper, which can be represented in .