Atomic graph-based symmetry recovery for machine learning force fields.

Machine-learning force fields (MLFF) based on kernel ridge regression show high accuracy and efficiency for molecules, materials, and interfaces [1,2]. However, the performance of MLFFs greatly depends upon incorporating the physical symmetries of the considered systems. Finding all relevant symmetries becomes a challenging task for large system sizes. Here we develop a data-driven symmetry search method based on molecular graphs for revealing relevant symmetries in molecules and materials. We demonstrate that our approach allows distinguishing atoms with different chemical environments, as well as controlling the accuracy of the MLFF by adjusting the level of symmetry. Effective MLFFs were constructed for complex periodic systems allowing the efficient investigation of defects behavior in CsPbBr₃ and the study of the free energy landscape for graphene interface with ethanol, 1,8-naphthyridine, D-histidine, D-alanine, and D-proline.

1. S. Chmiela, H. Sauceda, K.-R. Müller, A. Tkatchenko, Nat. Commun., 9(1), 3887 (2018)

2. H. Sauceda, L.E Gálvez-González, S. Chmiela, et al. Nat. Commun., 13, 3733 (2022)