xnn.hybrid

Hybrid GNN + transformer interatomic potentials.

This family holds models that combine graph message passing with a transformer (attention) update and a physics-based split of the energy:

• BAMBOO (Gong et al. 2024), a graph equivariant transformer whose energy is the sum of a semi-local neural-network term, a charge-equilibrium electrostatic term, and an optional D3(CSO) dispersion term;

• SpookyNet (Unke et al. 2021), message passing with s-, p- and d-orbital-like interactions, charge and spin embeddings and self-attention over all atoms, with ZBL repulsion, damped electrostatics and D4 dispersion added to the energy.

BAMBOO reuses the shared transformer primitives in xnn.transformer; SpookyNet’s radial basis lives in xnn.hybrid.featurizers. Both follow the common InteratomicPotential contract, so they plug into the same training/deploy pipeline as every other xnn model.

Modules

featurizers	Featurizers of the hybrid family.
models