Fig. 5: The CCNN architecture design flow.
From: Faithful novel machine learning for predicting quantum properties
A small cubic region surrounding one molecule of the crystal (Gd2O3 as an example) is converted into an antialiased voxel lattice. Each voxel encodes a user-configurable representation of an atom whose center is less than one voxel unit away from the voxel’s center. Classification is performed by augmenting a fully connected network (red) with a series of convolution layers (blue) that process per-voxel atomic embeddings.
