Fig. 2. Molecular convolution block (A) illustration of self interaction and equivariant activation layers. Self interaction layers are similar to traditional 1 × 1 convolutions, while equivariant activation layers ensure higher order tensor field feature arrays are activated across all representation indices. The shifted softplus activation from SchNet is used: ssp(x) = ln(0.5ex + 0.5). (B) Illustration of the complete molecular convolution block contain all layers for scalar and vector field inputs. These blocks are analogous to traditional convolutional layers in their mathematical function, with the added benefit of equivariant afforded through the spherical harmonics.
