Prior-guided adaptive local refinement of foundation interatomic potentials
Foundation interatomic potentials enable simulations across diverse materials, but task-specific inaccuracies can require additional first-principles data. Obtaining these data is challenging when the relevant local environments arise in large simulation cells. Here we introduce the Prior-guided Adaptive Refinement Scheme (PARS), which identifies weakly supported environments during molecular dynamics by comparing their frozen representations with prototypes of pretraining environments, without requiring initial task labels or model ensembles. Selected environments are reconstructed as compact periodic cells for first-principles labelling and iterative refinement. Applications to phase stability, surface migration and vacancy clustering show improvements in target properties, including an 81% reduction in the mean absolute error of Ti vacancy-cluster energies per vacancy relative to the pretrained model. For Si/Ge, a foundation potential refined using 800 labelled cells containing 10-128 atoms is subsequently distilled into a compact potential without additional first-principles training labels. In simulations containing approximately 362,880 atoms, the distilled model predicts an interfacial thermal resistance consistent with that of a potential trained directly on density-functional theory data. These results demonstrate a route for adapting foundation potentials to target applications through local reference calculations.