Atomistic uncertainty estimation in ANAKIN-ME neural network potentials

Nick Terrel, Adrian Roitberg

University of Florida

The Artificial NeurAl networK engINe for Molecular Energies (ANI) methodology has proven useful in producing general, scalable, and transferable neural network potentials (NNPs) capable of predicting molecular energy within 1 kcal/mol of the reference quantum mechanical (QM) level of theory at a greatly reduced computational cost. Estimation of reliability in predicted potential energy surfaces by machine learning models is of great interest for scaling the applicability of NNPs to systems larger than those included in training sets. ANI NNPs are trained using data generated by an automated active learning process, which utilizes a query by committee (QBC) approach to sampling chemical conformational space. This approach imposes reasonable agreement of molecular energy prediction by an ensemble of neural network models. ANI-2x is additionally trained to forces, which are predicted as analytical derivatives of the energy with respect to atomic positions. The ANI methodology lacks a means of identifying specific local chemical environments in which the forces acting on an atom are predicted with high variance by an ensemble of NNP models. Recognizing patterns in ensemble uncertainty in predicted atomic force components are hypothesized to give insight on the ability of NNPs to understand different atomic environments of varying complexity, and could be used to improve the process of generating new structures via active learning.