Development of accurate transferable hydrofluorocarbon refrigerant force fields using a machine learning and optimization approach

Developing a transferable classical force field (FF) has historically been a lengthy, expert-informed process. In this work, we integrate optimization, machine learning, and data science techniques to accelerate the systematic design and parameterization of transferable FF models. As a demonstration, we create transferable FF models for one- and two-carbon hydrofluorocarbons to accurately model diverse thermophysical properties including saturated liquid and vapor densities, vapor pressure, and enthalpy of vaporization. Estimability analysis and eigen-decomposition of the Fisher information matrix inform the number an identity of atom types in the final FF model. This model obtains an average mean absolute percent deviation (MAPD) between 2.92% (liquid density) and 31.5% (vapor density) on molecules not considered in the training set. This model (MAPD = 18.37%) also achieves a lower overall MAPD than an optimized version of the generalized AMBER FF (MAPD = 19.95%). Gaussian process surrogate models reduce the evaluation time associated with model selection and optimization from an order of months to minutes. This work suggests that the use of surrogate models combined with data science and optimization methods can greatly accelerate the development of accurate transferable force fields.