Fast and Scalable Uncertainty Estimates in Deep Learning Interatomic Potentials

Deep Learning has emerged as a promising and computationally-efficient tool for achieving highly accurate predictions of molecular and materials properties. However, a common short-coming shared by deep learning approaches is that they only obtain point estimates of their predictions without any associated predictive uncertainties, which are important for detecting rare events in molecular dynamics simulations. Existing attempts to quantify uncertainty in deep learning models have focused primarily on the use of ensembles of independently trained neural networks, the sample distribution of which is used to estimate model uncertainty. This process, however, results in a large computational overhead in both training and prediction, often orders of magnitude more expensive than a single deep learning model. In this project, we propose a method to estimate predictive uncertainty using only a single neural network along with a computationally-inexpensive Gaussian Mixture Model, eliminating the need for an ensemble. We compare the quality of the uncertainty estimates obtained with our method to those of ensembles. Furthermore, we study the effectiveness of our method in active learning and discover the results to be comparable to active learning with ensembles.