Skip to content

Neural network components

elfes.modules.nn contains reusable operations for constructing and updating equivariant network features. These components can be combined within different model architectures; the reference follows their mathematical roles.

Linear operations mix feature channels, while tensor products couple features to form new equivariant features. Normalization controls feature scale while preserving the transformation rules.

Frames and spherical transforms cover local-frame rotations, component ordering, and conversion between spherical coefficients and grid values. Radial functions turn interatomic distances into scalar features and provide smooth cutoff envelopes.

Geometric operations such as spherical harmonics and Wigner rotations are documented under elfes.modules. Decoders map model features to physical predictions, and losses compare those predictions with reference quantities.