In modern fixed-charge force fields, the van der Waals dispersion-repulsion interactions are almost universally represented by the simple Lennard-Jones (LJ) potential. These interactions are critical drivers of many biomolecular phenomena studied using atomistic molecular dynamics, but require good parameterization to produce useful and quantitative results. The traditional paradigm for training the LJ potential is fitting against experimental measurements of pure densities and enthalpies of vaporization.<br/>However, enthalpies of vaporization are a problematic target, as fixed-charge molecular dynamics simulations cannot capture changes in molecular polarization that affect energetics. There are also relatively few modern measurements of enthalpy of vaporization, with many older measurements lacking uncertainty estimates or detailed provenance information. By contrast, binary enthalpies of mixing are readily available in the modern NIST ThermoML archive and provide a similar enthalpic constraint. Parameterizing against mixture properties also has other benefits, allowing for easy inclusion of interactions between different functional groups and varying training targets by composition.<br/>In this work, we re-trained key LJ parameters from the OpenFF 1.0.0 (Parsley) force field against a set of training molecules composed of alkanes, alcohols, ethers, esters, ketones. For this set of molecules, we selected training datasets composed of a) only pure density and enthalpy of vaporization, b) only binary enthalpies of mixing and densities of binary mixtures, and c) combinations of a) and b). We show through benchmarking against a larger test set that mixture properties are able to replace the traditional pure properties as an LJ training target. Training against mixture properties also corrects systematic errors in the test set enthalpies of mixing, as well as improving correlation in test enthalpies of vaporization. This LJ training technique, applied to the parameterization of the OpenFF 2.0.0 (Sage) force field, improves enthalpic interactions and allows us to leverage the well-curated data in the NIST ThermoML database to produce improved LJ parameters.

Improving Lennard-Jones parameters by training against condensed phase mixture data

ACS Spring 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.