Constructing Robust Machine Learning Potential through Enhanced Sampling in UMLIP Latent Feature Space

Recent advances in machine learned interatomic potentials (MLIPs) have enabled accurate, large-scale atomic simulations for a wide range of materials and molecules. The robustness and accuracy of MLIPs critically depend on the coverage of the configuration space in the training dataset. A common strategy involves sampling diverse configurations generated by finite-temperature molecular dynamics (MD) for the initial dataset, followed by iterative expansion of the training set through active learning. However, certain configurations, such as transition states, are inherently rare events that may remain undersampled, even in long-timescale MD simulations. Here, we propose a method that leverages recent advances in universal MLIPs with biased MD simulations to enhance the sampling of such configurations. We demonstrate that the resulting MLIPs are significantly improved in robustness and predictive accuracy using several application domains, including lithium diffusion in battery materials.