Extracting Atomic Environments for Training Machine-learned Interatomic Potentials

In order to appropriately capture large-scale material features and emergent phenomena via atomistic simulations, such as Molecular Dynamics (MD), the system scale can range up to hundreds of millions of atoms—depending on the application. However, the force-field models that drive those simulations are generally trained with Density Functional Theory (DFT) reference data, limited to relatively small configurations on the order of 100s or 1000s of atoms. To get DFT forces on atoms in regions of interest, one needs to extract a small set of atoms from the large simulation box; however, methods to select the shape and size for this set, as well as to generate a potentially necessary passivating envelope, have not been systematically analyzed. In this work, we present a benchmark of various techniques to extract atomic environments from large, bulk configurations and embed them into smaller configurations suitable for DFT calculations with periodic boundary conditions. We demonstrate that a notably simple procedure yields superior performance over many previously proposed methods. Finally, we apply this technique to incorporate grain boundaries into the training of a metal-oxide force field.