Elemental augmentation of machine learning interatomic potentials

Machine learning interatomic potentials (MLIPs) bridge the gap between the accuracy of ab initio methods and the computational efficiency needed for large-scale simulations. However, custom-trained MLIPs are often limited to specific materials and lack flexibility for incorporating additional elements, while universal potentials (UPots), despite covering a wide range of chemical elements, may sacrifice accuracy for generalization. In this work, we propose an elemental augmentation strategy to efficiently expand MLIPs by incorporating new elements into pre-trained models. Using a Bayesian optimization driven active learning framework, we target the configuration space of new elements where the current MLIPs exhibit high uncertainty and demonstrate the addition of up to 10 elements to a pre-trained UPot. The results demonstrate a high tendency for sampling new structures composed of these elements, minimizing sampling requirements. It reduces computational costs by over an order of magnitude compared to training an MLIP from scratch, while preserving accuracy. This strategy offers a scalable pathway to extend MLIP applicability across diverse chemical spaces.