Torch operations and model components
elfes.modules provides the Torch operations used to build electronic-structure models and interpret their predictions. Its components work with atomic geometry, equivariant features, and physical quantities, following the basis and transformation conventions defined by elfes.physics.
Geometry and rotations provide neighbor search, spherical harmonics, and rotation matrices. Orbital transforms change the coordinates or scales of orbital data and transform cell-shift-indexed matrices to k-points. These operations connect the physical representation of a quantity to the form needed by a model or calculation.
Neural network components supply the reusable operations within a model. Decoders map learned features to outputs with a specified physical structure, and loss functions compare those outputs with reference quantities. Model architectures choose how these pieces fit together; the training engine manages their optimization and evaluation.