Machine Learning

These documents develop the machine-learning framework used by ELFES. An atomistic system is described as an attributed geometric graph, and a model maps its geometry and features to node-, edge-, or graph-level predictions.

The graph is a common language for atoms, atomic pairs, and message passing, not a requirement that every implementation take the form of a conventional graph neural network. Most ELFES models nevertheless do use graph-based architectures.

The presentation separates message-passing Topology from the e3nn-style equivariant feature calculus as clearly as their interaction allows. The former describes which graph states communicate and how messages are aggregated; the latter describes how geometric features transform and which operations may combine them.

The physical definitions of model inputs and targets come from Electronic Structure, and their transformation laws and irrep structure come from Equivariance.

Documents