ELFES
ELFES (Equivariant Learning Framework for Electronic Structure) is a unified, model-agnostic framework for learning electronic structure at scale, built on PyTorch.
ELFES brings data from different DFT codes into a common physical description across spin configurations and boundary conditions. Different model architectures share the same physical data types, IO, datasets, and training infrastructure. This framework supports research on Hamiltonians, density matrices, electron and magnetization densities, energies, and forces, with consistent physical conventions from data preparation to the interpretation of predictions.
Architecture and workflow
ELFES covers the workflow from first-principles data to equivariant modeling, training, and physical analysis. Physical quantities retain consistent definitions across these stages, with distinct components handling source formats, dataset storage, and model computation.
Data and training at scale
ELFES is designed for growth in dataset size, model capacity, and computational resources. Efficient dataset access and distributed training extend the optimized model components into a complete workflow for large-scale electronic-structure learning.
Spin and periodicity
Spin and periodicity are part of the physical description throughout ELFES. Shared data types and conventions describe magnetic and nonmagnetic systems across finite and periodic geometries. Their physical meaning and transformation rules remain consistent through data processing, equivariant modeling, and analysis.
Modular design
Physical targets, model architectures, and learning methods are independent choices. ELFES separates physical definitions from model design and training execution, allowing data, scientific operations, and training infrastructure to be reused across research tasks. Components can be developed and optimized independently while retaining consistent physical conventions.
Optimized computation
ELFES optimizes the geometric and equivariant operations shared by many models, including tensor products, spherical transformations, and neighbor construction. Specialized CPU and GPU kernels make these operations available as reusable model components, so computational improvements carry across architectures.
Documentation
- Theory: electronic structure, equivariance, and their role in learning.
- Physical data and datasets: the common foundation for electronic-structure data.
- Equivariant components, decoders, and losses: model computation and physical outputs.
- Training engine: training and evaluation at scale.
- Source code: implementations, research models, and workflows.