Improved training set sampling techniques for machine learned interatomic potentials

Machine learned interatomic potentials (MLIPs) offer the prospect of ab initio level accuracy and reliability as linear-scaling, surrogate models. Currently, much work is focused on improving spatial descriptors and model formulations, which are typically trained on a random structures and modulated structures derived from a representative set of known phases. Ab initio molecular dynamics is typically employed to sample the local potential energy surface of the training structures. Here, we explore the biases that molecular dynamics can impart in the performance MLIPs, using the solid-state Zr-O system as test system presenting structural and chemical complexity. To address these biases, we explore improvements in configuration space sampling, including a novel technique that may have applications beyond machine learning. To this end, our methods are implemented in our open source framework for rapidly generating training data on HPC clusters.