API Reference#

Generated from the docstrings in src/xnn. Start from the subpackage matching what you need:

  • xnn.common: data pipeline and hub, featurizer base, configuration, model registry, hub and add-ons (LES, D3, D4), fine-tuning, training, benchmarking, deployment, CLI

  • xnn.gnn: SchNet, DimeNet, DimeNet++, GemNet, PaiNN, NequIP, MACE, Allegro, CACE, AIMNet2, their featurizers and fast paths (NequIP, MACE and Allegro need e3nn)

  • xnn.cnn: the SE(3) steerable CNN and the conventional 3D CNN over voxelized environments, the spherical CNN over spherical signals, and their featurizers

  • xnn.dnn: HDNNP (generations 1 to 4), ANI, PhysNet, the symmetry-function / AEV featurizers and the RuNNer model files

  • xnn.ffnn: ReaxFF, OPLS, DREIDING and the .frc force-field reader

  • xnn.transformer: edge attention and the exponential-normal basis

  • xnn.hybrid: BAMBOO, SpookyNet

xnn.common

Shared abstractions used across all model families (cnn / dnn / gnn).

xnn.gnn

Graph-network family: SchNet, DimeNet, DimeNet++, GemNet, PaiNN, and NequIP / MACE / Allegro / CACE / AIMNet2 (these need e3nn).

xnn.cnn

Convolutional family: 3D CNNs over voxelized atomic environments (the conventional CNN3D and the SE(3)-equivariant 3D steerable CNN of Weiler et al., NeurIPS 2018) and the spherical CNN of Cohen et al. (ICLR 2018) over spherical signals of the environments.

xnn.dnn

Dense neural networks (dnn): HDNNP (generations 1-4), ANI and PhysNet.

xnn.ffnn

ffnn: learnable classical force fields.

xnn.transformer

Transformer building blocks shared by attention-based potentials.

xnn.hybrid

Hybrid GNN + transformer interatomic potentials.
Show Source