Detailed exploration of potential energy surface (PES) plays crucial role in design of chemical systems and adjoint theoretical/experimental study. Machine learning interatomic potentials (MLIPs) constructed in terms of local atomic environment are system-size independent, i.e. they could be trained on stoichiometry Xn and then applied to Xn+m systems. We propose to investigate in detail what are the limits of the Xn→Xn+m gap. Can we systematically skip some stoichiometries in the dataset and approach larger systems with acceptable accuracy in MLIPs predictions? When exploring PES, we are mostly interested in low-lying isomers, therefore, we applied MLIPs trained on Xn clusters to assess relative stabilities of Xn+m systems. We tested our idea on boron clusters for which comprehensive data is accumulated. Our preliminary results show that MLIP easily selects relative isomers of B16 when trained only on B10, B12, and B14 dataset of 85k entries. Inclusion of B16 clusters lead to successful determination of relevant B18 isomers. The same dataset was used to systematic exploration of more complex B20 PES on which MLIP correctly ranked global minimum and lowest-lying isomers among 20k of random structures. Current insight shows that gap of 4 atoms is feasible and dramatically accelerates PES search. This is an ongoing project, and we aim to systematically utilize Xn→Xn+m strategy to investigate much larger systems with greater m increment. We hope this would allow us to broaden boron chemistry which is currently limited by B42 PES. Moreover, this strategy is general and could be applied to any other system. To facilitate further studies, we also present our code STRENDEN which is designed for fast and automated generation of custom datasets, as well as handling batch submissions and file operations on computational clusters.

Size does not matter: Machine learning potentials trained on Xn stoichiometries are applicable for Xn+m systems

ACS Spring Meeting 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.