Large scale molecular dynamics (MD) simulations rely on interatomic potentials that are pre-constructed using ab initio data as well as both empirical and physical considerations. Specifically, machine-learned interatomic potentials (MLIAPs) are fitted to data available from expensive quantum chemistry computations. These MLIAPs encapsulate the functional relationship between the atomic configuration and the potential energy of an atomic system, and are trained in a supervised machine learning context. Training such potentials is a laborious process involving expert creation and curation of training datasets. Aiming to automate this process, active learning methods have been increasingly used for the automated creation or selection of training data, and refinement of interatomic potentials. Besides data selection, the optimal tradeoff between accuracy and complexity is also sought after to enable model selection.
Uncertainty quantification (UQ) for MLIAPs is critical for both training data selection and model selection. Furthermore, MLIAPs equipped with UQ enable the propagation of uncertainty through MD simulations, thereby providing uncertainty estimates on MD simulation outputs and allowing for meaningful comparisons with experimentally observed macroscale quantities of interest.
In this talk, we will discuss our work on a range of UQ approaches for MLIAPs from the perspective of active learning and model error estimation. This includes Bayesian inference of MLIAP parameters via both sampling and approximate methods, as well as embedded model error estimation, which augments MLIAP parameters with statistical error terms to be inferred within the Bayesian framework. We will also explore ensemble methods such as query-by-committee, as a means of extracting estimates of MLIAP predictive uncertainties. We will demonstrate the results on material systems of relevance in fusion energy science applications.