Data-efficient construction of high-fidelity graph deep learning interatomic potentials

Abstract Machine learning potentials (MLPs) have become an indispensable tool in large-scale atomistic simulations. However, most MLPs today are trained on data computed using relatively cheap density functional theory (DFT) methods such as the Perdew-Burke-Ernzerhof (PBE) generalized gradient approximation (GGA) functional. While meta-GGAs such as the strongly constrained and appropriately normed (SCAN) functional have been shown to yield significantly improved descriptions of atomic interactions for diversely bonded systems, their higher computational cost remains an impediment to their use in MLP development. In this work, we outline a data-efficient multi-fidelity approach to constructing Materials 3-body Graph Network (M3GNet) interatomic potentials that integrate different levels of theory within a single model. Using silicon and water as examples, we show that a multi-fidelity M3GNet model trained on a combined dataset of low-fidelity GGA calculations with 10% of high-fidelity SCAN calculations can achieve accuracies comparable to a single-fidelity M3GNet model trained on a dataset comprising 8 × the number of SCAN calculations. This work provides a pathway to the development of high-fidelity MLPs in a cost-effective manner by leveraging existing low-fidelity datasets.