Addressing the Elephant in the Room: Uncertainties in Physical Predictions From Machine-Learned Force Fields

Learning molecular force-fields (FF) has played a leading role in the path towards reliable molecular dynamics simulations in biology, chemistry, and materials science[1,2]. However, simulation’s predictive power is only as good as the underlying interatomic potential. Although it is common practice to evaluate the reliability of trained FF models based on typical error measures, this only quantifies the error on the database given a set of training points. The relevant question to ask is how well a learned FF reproduce the actual physical properties a system. Here, we present an analysis of the uncertainties in properties derived from learned-FFs, such as vibrational spectrum and thermodynamics. A clear correlation is found between learning errors and the derived properties' uncertainty. The robustness of the symmetric gradient-domain machine learning (sGDML) framework[1] against such problem is evinced by its fast uncertainty minimization with the training set size. These results will serve as reference for the developing of robust and predictive learned physical models.

[1] Chmiela et al. Sci. Adv. 3 (5), e1603015 (2017); Nat. Commun. 9 (1), 3887 (2018); Comput. Phys. Commun. 240, 38 (2019).
[2] Sauceda et al. J. Chem. Phys. 150 (11), 114102 (2019); arXiv:1909.08565 (2019).