xnn.cnn.featurizers.spherical#
Spherical signals of atomic environments: the input of the spherical CNN.
The molecular representation of Cohen et al. (ICLR 2018, Sec. 5.4): a
sphere of radius r is placed around every atom i and, for every
species z, the potential of the atoms of that species is sampled on it,
U_z(x) = sum_{j != i, Z_j = z} Z_i Z_j / |x - p_j|^exponent,
on the Driscoll-Healy grid of bandwidth b (2b x 2b points, axes
beta, alpha). The signals are invariant to translations, rotate with the
molecule, and are differentiable in the positions, so autograd forces work
unchanged. The paper’s potential has exponent = 1; its data script
uses the inverse square distance (exponent = 2).
Module Attributes
grid points per block of edges (bounds the |
Classes
|
Per-atom spherical signals of the species-resolved neighbor potential. |