Exploration of transferable and uniformly accurate neural network interatomic potentials using optimal experimental design

Machine learning has been proven to have the potential to bridge the gap between the accuracy of ab initio methods and the efficiency of empirical force fields. Neural networks are one of the most frequently used approaches to construct high-dimensional potential energy surfaces. Unfortunately, they lack an inherent uncertainty estimation which is necessary for efficient and automated sampling through the chemical and conformational space to find extrapolative configurations. The identification of the latter is needed for the construction of transferable and uniformly accurate potential energy surfaces. In this paper, we propose an active learning approach that uses the estimated model's output variance derived in the framework of the optimal experimental design. This method has several advantages compared to the established active learning approaches, e.g. Query-by-Committee, Monte Carlo dropout, feature and latent distances, in terms of the predictive power and computational efficiency. We have shown that the application of the proposed active learning scheme leads to transferable and uniformly accurate potential energy surfaces constructed using only a small fraction of data points. Additionally, it is possible to define a natural threshold value for the proposed uncertainty metric which offers the possibility to generate highly informative training data on-the-fly.