Optimization of transferable molecular step potentials

Abstract

To study thermodynamic properties for hydrocarbons and oxygenated compounds, discontinuous molecular dynamics and thermodynamic perturbation theory have been used. Accurate thermodynamic predictions require realistic and efficient transferable potential models for intermolecular interactions. Necessary parameters for these models can be obtained through reliable optimization methods. A stochastic search algorithm, namely Recursive Random Search, combined with a local search algorithm, namely Levenberg-Marquardt, is used together to optimize the parameters of 12 different potential models. These models are then compared using 5 statistical testing methods, which are cross-validation, Akaike7s Information Criterion, Bayesian Information Criterion, Mallow's Information Criterion and F-Test. The optimum model is then selected to be the StepYukawa-Universal model (SYU) because of its predictive power on validation sets. Besides SYU, parameters for 3 additional models, Lennard Jones model (LJ), Yukawa-Universal model (YU) and Linear 2580 model, are also documented for their potential use and wide acceptance. The root mean squared percentage vapor pressure errors evaluated at reduced temperatures approximately between 0.480 and 0.900 for 102 compounds from 6 organic families, namely n-alkanes, branched alkanes, aromatics, naphthenics, alcohols and phenols, are 13.30, 13.23, 13.28 and 13.42 per cent for SYU, YU, LJ and Linear 2580 model, respectively. In addition to the accuracy in vapor pressure predictions, similar united atoms exhibit high similarities in terms of the parameters for molecular potentials which support the consistency and accuracy of the optimized parameters as well.