(708a) Accelerating Force Field Parameterization to Improve the Quantitative Predictability of Thermophysical Properties

A plethora of force fields for organic compounds exist in the literature. These models differ in the underlying molecular assumptions and the data used in the parameterization process, which are typically liquid density, vapor pressure, and/or heat of vaporization. The traditional parameterization approach rarely considers the entire parameter space due to the relatively high computational cost for molecular simulations, although significant progress has been made in recent years. Instead, the force field parameters are usually optimized in a trial and error manner or with gradient based algorithms. With the traditional optimization approach it is difficult to assure that a global minimum is obtained, to assess the degree of nonlinear coupling between parameters, and to quantify the uncertainty in the optimal set of parameters. Furthermore, the optimization process typically includes only a small set of experimental data (to limit the number of molecular simulations required) for a few compounds (to reduce the number of nonbonded parameters that are optimized simultaneously). This is significant because the reliability and transferability of the resulting force field is intimately related to the quality and quantity of data to which it is parameterized. For this reason, we present a parameterization approach designed to overcome these deficiencies and, thereby, develop a more quantitatively predictive force field for thermophysical properties.

The novel methodology employed in this work reduces the computational cost of force field parameterization by up to three orders of magnitude. The optimization is accelerated by using simulated configurations for an initial parameter set to predict physical property values for unsampled parameter sets. This acceleration enables the simultaneous optimization of several parameters to the large amount of experimental data available at the National Institute of Standards and Technology (NIST) Thermodynamics Research Center (TRC). In addition, this approach facilitates the use of Bayesian inference to assign uncertainties to the force field parameters. This is important as meaningful uncertainties are essential for molecular simulation to be a quantitative tool in thermophysical property prediction. Furthermore, advanced Bayesian methodologies are employed to ensure that the parameters are not overfit to the training data set by simultaneously investigating several nonbonded potential functional forms (e.g. Lennard-Jones 12-6, Mie n-m, Buckingham exponential-6, etc.). This improves the transferability of the force field when compared to data at different state points.