Differentiable sampling of molecular geometries with uncertainty-based adversarial attacks

Neural network (NN) force fields can predict potential energy surfaces with high accuracy and speed compared to electronic structure methods typically used to generate their training data. However, NN predictions are well-defined only for points close to the training domains, and may exhibit poor results during extrapolation. Uncertainty quantification methods can detect geometries for which predicted errors are high, but sampling regions of high uncertainty requires a thorough exploration of the phase space, often using expensive simulations. Our work uses automatic differentiation to sample atomistic configurations by balancing thermodynamic accessibility and uncertainty quantification without using molecular dynamics simulations. This dataset provides the atomistic data used to train the NN potentials for the ammonia, alanine dipeptide, and zeolite-molecule systems. For all materials, geometries, energies, and forces are provided. The ammonia and zeolite systems were computed using density functional theory calculations, while the alanine dipeptide dataset was generated using molecular dynamics simulations with the OPLS force field.