Statistical methods for resolving poor uncertainty quantification in machine learning interatomic potentials

Abstract: Machine learning interatomic potentials (MLIPs) are promising surrogates for quantum mechanics evaluations in ab-initio molecular dynamics simulations due to their ability to reproduce the energy and force landscape within chemical accuracy at four orders of magnitude less cost. While developing uncertainty quantification (UQ) tools for MLIPs is critical to build production MLIP datasets using active learning, only limited progress has been made and the most robust method, ensembling, still shows low correlation between high error and high uncertainty predictions. Here we develop a rigorous method rooted in statistics for determining an error cutoff that distinguishes regions of high and low UQ performance. The statistical cutoff illuminates that a main cause of the poor UQ performance is due to the machine learning model already describing the entire dataset and not having any datapoints with error greater than the statistical error distribution. Second, we extend the statistical analysis to create an interpretable connection between the error and uncertainty distributions to predict an uncertainty cutoff separating high and low errors. We showcase the statistical cutoff in active learning benchmarks on two datasets of varying chemical complexity for three common UQ methods: ensembling, sparse Gaussian processes, and latent distance metrics and compare them to the true error and random sampling, showing that the statistical cutoff is generalizable to a variety of different UQ methods and protocols and performs similarly to using the true error. Importantly, we conclude that utilizing this uncertainty cutoff enables using significantly lower cost uncertainty quantification tools such as sparse gaussian processes and latent distances compared to ensembling approaches for generating MLIP datasets at a fraction of the cost.