On-the-fly Active Learning of ChIMES Force Fields for f-electron Materials

The creation of training data for molecular dynamics models of f-electron materials for the study of grain boundaries, surface chemistry, and corrosion properties poses a unique challenge in that calculations with Density Functional Theory (DFT) can be extraordinarily computationally intensive, making it difficult to impossible to generate sufficient training sets with standard sampling processes alone. In this work, we propose methods to overcome this difficulty through uncertainty quantification (UQ) for linearly parameterized MD models. Our methods can maximize the information content of the underlying DFT data, allowing for accurate MD model optimization with very sparse data sets. We combine our approach with an on-the-fly active learning (AL) scheme to the Chebyshev Interaction Model for Efficient Simulations (ChIMES) interatomic potential framework, which fits many-body potential energy surfaces with a linear combination of Chebyshev polynomials evaluated on atomic clusters. We find that our combined UQ/AL method is able to train a ChIMES model for a model metal-oxide system that yields accurate phonon dispersion curves and hydrogen absorption energies with as few as a twenty DFT evaluations.