Committee Disagreement Biased Active Learning of Interatomic Potentials

Committee models are well known to improve generalizability of machine-learned models and neural-network models in general. Moreover, the disagreement between the predictions of the individual models can be used as a proxy for overall model uncertainty quantification. By exploiting the differentiability of interatomic potential models, atomic structures can be driven into regions of high uncertainty to find new training structures as part of an adversarial attack scheme. We explore several ways of incorporating this adversarial attack scheme into practical structure generation schemes like molecular dynamics and our contour exploration scheme [1] for efficient active learning of interatomic potentials. We showcase the performance of this approach using transition metals and their oxides as benchmark systems.

[1] Michael J Waters and James M Rondinelli 2021 J. Phys.: Condens. Matter 33 445901 DOI: 10.1088/1361-648x/ac1af0