Molecular dynamics (MD) simulations are often done using machine-learned interatomic potentials (MLIAPs) that are constructed from empirical and physical considerations, and fitted to data available from expensive ab initio quantum chemistry computations. These MLIAPs encapsulate the functional relationship between atomic configuration and potential energy of an atomic system, and are trained in a supervised machine learning context.
Uncertainty quantification (UQ) for MLIAPs is useful for both training data selection in an active learning context, and the selection of MLIAP models of optimal complexity. Furthermore, MLIAPs equipped with UQ enable the propagation of uncertainty through MD simulations, thereby providing uncertainty estimates on MD simulation outputs.
In this talk, we will discuss our work on a range of UQ approaches for MLIAPs and subsequent propagation of uncertainties through MD simulations. This includes Bayesian inference of MLIAP parameters via Markov chain Monte Carlo sampling, as well as approximate versions including variational inference and approximate Bayesian computation to help in the handling of highly overparameterized MLIAPs, such as those based on neural network forms. We will also explore ensemble methods such as query-by-committee, as a means of extracting MLIAP predictive uncertainties. We will demonstrate the results on material systems of interest, driven by fusion energy science applications.