Mechanical Properties of Nanoporous Graphenes: Transferability of Graph Machine‐Learned Force Fields Compared to Local and Reactive Potentials

Abstract

Nanoporous and chemically‐bridged graphene nanosheets span a wide chemical space with a broad set of applications in sensing and electronics. Modeling the structure and dynamics of such nanosheets is challenging, as chemical bond making and breaking as well as non‐covalent interactions must be captured accurately and on equal footing. Here it is showed that recent graph‐based machine‐learned force field (MLFF) SO3krates [J. T. Frank et al., Nat. Commun. 15, 6539 (2024)] is able to reliably model the dynamics and mechanical response for a broad class of nanoporous graphenes when trained on accurate density functional theory data that includes long‐range many‐body dispersion (MBD) interactions. In contrast, local moment tensor potentials and empirical reactive potentials are much less accurate. It is also found that recent MLFFs trained on solid‐state datasets must be used with care, since even empirical potentials occasionally yield more accurate results. These findings highlight the potential of properly‐trained graph MLFFs in modeling the properties of whole chemical spaces of complex functional materials.