Adversarial Sampling of Geometries Using Uncertainty Quantification from Single Neural Networks

In recent years, neural network (NN) interatomic potentials have been increasingly applied to construct potentials for a wide range of systems since they provide high numerical accuracy like in electronic structure methods and low computational cost comparable to that of analytical interatomic potentials. However, due to architectural limitations, NN interatomic potentials are still constrained by their inability to extrapolate well. In high-dimensional systems such as crystal or molecular structures, NN interatomic potentials often fail to perform robustly in regions outside of well-learned training domains. To expand this learning space, many methods in the realm of active learning and uncertainty quantification (UQ) have been proposed. Here, we apply multiple UQ methods ranging from ensembling, mean-variance estimation, and deep evidential regression, and exploit a differentiable sampling technique based on adversarial attacks to sample atomic configurations with high uncertainty. Evaluation of these methods allow us to benchmark performance of single deterministic networks against ensemble methods for UQ task, especially for large bulk material systems which are restricted by their expensive training costs. Coverage of sampled regions is also evaluated for these UQ methods using statistical analysis after several active learning loops. These models are benchmarked on kinetic barrier sampling in ammonia and bulk interactions in silica glass systems.