xnn.dnn.models.base#
Shared base for descriptor + per-element-network potentials (HDNNP / ANI).
A DescriptorPotential composes an invariant per-atom Featurizer
(symmetry functions, AEV) with per-element atomic MLPs. HDNNP and ANI differ
mainly in the featurizer they pass in, so the model body lives here and is
reused by both hdnnp.py and ani.py.
The per-element networks support the flexibility ANI needs: a distinct hidden
architecture per element, a choice of activation (ANI uses CELU; the paper’s
ANI-1 used a Gaussian; SchNet-style models use SiLU), and optional
per-species self atomic energies added to each atom’s contribution (ANI’s
energy shift, in the same spirit as MACE/NequIP atomic_energies).
Classes
|
Shared body for descriptor-based potentials (HDNNP, ANI). |