Coarse-grained graph architectures for all-atom force predictions
Abstract
We introduce a coarse-grained all-atom force field (CGAA-FF) that integrates coarse-grained message passing into all-atom force fields using equivariant graph models. By embedding atomistic coordinates into grain-level nodes, CGAA-FF predicts grain-level energies and atom-level forces. Tested on EC/EMC electrolytes and crystalline and disordered RDX phases, it achieves 0.201 and 0.253 eV Å −1 force errors, with ~10-fold speed and ~5-fold memory improvements over conventional MLIPs, enabling efficient soft-matter simulations.
Fields of Study & Subjects
Machine Learning in Materials Science
Fuel Cells and Related Materials
Boron and Carbon Nanomaterials Research
Publication details
- Journal / source
- npj Computational Materials
- Publisher
- Springer Science and Business Media LLC
- Publication date
- 29 July 2026
- Language
- EN