Improved, Reliable Uncertainty Quantification of Interatomic Models using Sloppy Model Analysis

Interatomic models (IMs) are widely used in molecular modeling to circumvent the computational cost of quantum calculations. These IMs are often designed for specific applications of interest, and they are used to predict other materials properties that are not used in the development process. Uncertainty quantification (UQ) is relevant for assessing the reliability of these out-of-sample predictions. Previous studies have shown that many IMs are insensitive to large, coordinated changes in many of their parameters, a phenomenon known as sloppiness. Furthermore, our previous work has shown that sloppiness poses challenges both for the implementation and interpretation of traditional UQ analysis. We propose a systematic UQ process for sloppy models, utilizing the Manifold Boundary Approximation Method (MBAM) to identify sloppy parameters and find the reduced, less-sloppy model. We demonstrate this process using the Stillinger-Weber (SW) potential, calibrating it to the atomic forces of a molybdenum disulfide system. We find that the parametric uncertainty of the reduced model is less sensitive to the choice of confidence level and leads to non-diverging uncertainty in higher statistical confidence level.