Small-data-based machine learning interatomic potentials for graphene grain boundaries enabled by structural unit model

Machine learning interatomic potentials (MLIPs) are emerging as a powerful tool to achieve efficient atomistic simulations with DFT-level accuracy, which can greatly enhance the capability of modeling more realistic systems. However, training high-quality MLIPs often requires a big database typically built by DFT calculations, which is very computationally expensive and even prohibitive for complex systems. Here, we propose an efficient strategy for developing MLIPs of grain boundaries (GBs) using a small database determined by the structural unit model. Using graphene GBs as an example, we show that the trained MLIP can achieve DFT-level accuracy in predicting a broad range of atomic structures and properties, particularly phonon properties, for the entire GB configuration space and exhibits high transferability. The proposed strategy will facilitate the development of high-fidelity MLIPs of general GBs and potentially other complex systems, greatly promoting the understanding of complex thermal issues by atomistic simulations with DFT-level accuracy.