CPGN: A Multi-Scale GNN for Crystal Property Prediction
A new graph neural network (GNN) called Coordination Polyhedron Graph Network (CPGN) has been proposed for predicting crystal properties. Unlike existing models like CGCNN, MEGNet, ALIGNN, and SchNet that primarily learn atomic-level representations, CPGN jointly learns atomic, bond, and coordination-polyhedron representations. It constructs three coupled graphs: an atom graph for elemental and bonding information, a line graph for angular interactions, and a coordination polyhedron graph for Voronoi-derived local environments. This multi-scale approach addresses the limitation that many material properties are governed by coordination polyhedra. The work is published on arXiv with ID 2607.24818.
Key facts
- CPGN is a multi-scale GNN for crystal property prediction.
- It jointly learns atomic, bond, and coordination-polyhedron representations.
- Three coupled graphs: atom graph, line graph, coordination polyhedron graph.
- Coordination polyhedron graph uses Voronoi-derived local environments.
- Existing GNNs like CGCNN, MEGNet, ALIGNN, SchNet are primarily atomic-level.
- Many material properties are governed by coordination polyhedra.
- Published on arXiv with ID 2607.24818.
- Addresses limitation of implicit learning of local chemical environments.
Entities
Institutions
- arXiv