Uncertainty-biased molecular dynamics for learning uniformly accurate interatomic potentials

Machine learning has transformed how we model atomic interactions in molecular and material systems, bridging the gap between the accuracy of first-principles methods, e.g., density functional theory, and the efficiency of empirical force fields. Nevertheless, the effectiveness of machine-learned interatomic potentials crucially relies on training data sufficiently covering the relevant configurational (atom positions) and compositional (atom types) spaces. Without such training data, machine-learned interatomic potentials cannot faithfully reproduce the underlying physics, resulting in unphysical force predictions.

Therefore, creating a comprehensive training data set poses an open challenge for generating uniformly accurate machine-learned interatomic potentials, complicated by our general limited knowledge about the atomic system of interest. Moreover, this objective should be realized while reducing the number of expensive first-principles calculations. To address this challenge, we introduce an uncertainty-driven active learning approach, which uses uncertainty-biased molecular dynamics simulations to explore the relevant phase space of molecular and material systems. We apply our method to learn interatomic potentials for alanine dipeptide—a widely used model for protein backbone structure—and MIL-53(Al)—a flexible metal-organic framework.

Work performed as a collaboration between NEC Laboratories Europe GmbH, SimTech, and University of Stuttgart. For more information, read the article “Uncertainty-biased molecular dynamics for learning uniformly accurate interatomic potentials" from Zaverkin et al. here: