(272b) Successive Global and Local Optimization Techniques Utilized to Generate New Accurate VLE Force Fields

The practical applicability of atomistic soft matter simulations requires the construction of appropriate molecular models for a wide range of chemicals. The key to a quantitative property prediction is the accuracy of the simulation’s basis, i.e. the force field. This is particularly true for the intermolecular interactions. Manual adjustment and optimization of the respective parameters is, at best, extremely time consuming. Hence, semi-automated parameterization schemes are essential in the pursuit to create tailor-made models for specific molecules in a timely fashion.

In this contribution, we want to introduce and assess global optimization techniques as new crucial components for the intermolecular part of our force field calibration workflows. Two new Vapor--Liquid--Equilibrium (VLE) force fields for ethylene oxide are presented as a test case. Both of them were developed by a combination of highly efficient global and local optimization algorithms. Thereby, a weighted quadratic loss function between simulated and experimental VLE properties at different temperatures, i.e. saturated liquid density, enthalpy of vaporization, and vapor pressure, is minimized. The global optimization techniques considered here are based on the collocation of meta models of either the loss function itself or of each of the VLE properties. We will present, how the application of existing methods led us to our own algorithm, called COSMoS („Calibration/Optimization by Simultaneous Modeling of Simulated data“), and how it performs.

The resulting force fields reproduce the saturated liquid density with a maximum error of 0.5%, the enthalpy of vaporization with a maximum error of 1%, and the vapor pressure with a maximum error of 5%. Furthermore, the models are capable to predict other thermo-physical properties quite well.