Abstract
Neural network (NN) interatomic potentials provide fast prediction of potential energy surfaces, closely matching the accuracy of the electronic structure methods used to produce the training data. However, NN predictions are only reliable within well-learned training domains, and show volatile behavior when extrapolating. Uncertainty quantification methods can flag atomic configurations for which prediction confidence is low, but arriving at such uncertain regions requires expensive sampling of the NN phase space, often using atomistic simulations. Here, we exploit automatic differentiation to drive atomistic systems towards high-likelihood, high-uncertainty configurations without the need for molecular dynamics simulations. By performing adversarial attacks on an uncertainty metric, informative geometries that expand the training domain of NNs are sampled. When combined with an active learning loop, this approach bootstraps and improves NN potentials while decreasing the number of calls to the ground truth method. This efficiency is demonstrated on sampling of kinetic barriers, collective variables in molecules, and supramolecular chemistry in zeolite-molecule interactions, and can be extended to any NN potential architecture and materials system.
Highlights
Neural network (NN) interatomic potentials provide fast prediction of potential energy surfaces, closely matching the accuracy of the electronic structure methods used to produce the training data
Several models have combined different representations and NN architectures to predict potential energy surfaces (PESes) with increasing accuracy[11,12,13,14]. They have been applied to predict molecular systems[15,16], solids[17], interfaces[18], chemical reactions[19,20], kinetic events[21], phase transitions[22], and many more[4,6]. Despite their remarkable capacity to interpolate between data points, NNs are known to perform poorly outside of their training domain[23,24] and may fail catastrophically for rare events, such as those occurring in atomistic simulations with large sizes or time scales not explored in the training data
molecular dynamics (MD) trajectories can be unstable when executed with an NN potential and sample unrealistic events that are irrelevant to the true PES, especially in early stages of the active learning (AL) cycle when the NN training set is not representative of the overall configuration space
Summary
Similar strategies have been used in classifiers, graph-structured data[45,46], and physical models[47], no work has yet connected these strategies to sample multidimensional potential energy landscapes. In this framework, an adversarial attack maximizes the uncertainty in the property under prediction (Fig. 1a). Ground-truth properties are generated for the adversarial example This could correspond to obtaining energies and forces for a given conformation with density functional theory (DFT) or force field approaches. We can estimate that the probability p that a state Xδ with predicted energy EðXδÞ will be sampled is proportional to pðX δ Þ
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
