(311c) Validation of New Bayesian-Optimization Tuned Interatomic Potentials for Hydrofluorocarbons (HFCs), Perfluoroalkanes and Simple Alkanes.

Better interatomic potentials, commonly referred to as force fields (FFs), are needed for reliable and more fundamental understanding of matter through molecular simulations. This work focuses on the development of accurate FFs for a few important subsets of molecules namely, hydrofluorocarbons (HFCs), perfluoroalkanes and simple alkanes. The HFCs studied in this work are R32, R125, R134a and R143a. The perfluoroalkane studied as a model for other perfluoroalkanes is tetrafluoromethane (R14) while methane (R50) and ethane (R170) are studied as model examples of simple alkanes.

This work builds on recent work in our group, where multiple high-performing, molecule specific, FFs were obtained by tuning the Lennard-Jones parameters of the General Amber Force Field (GAFF), using the machine learning tool of Bayesian Optimization (BO), for the molecules studied in this work. The obtained FFs, henceforth referred to as BO-tuned FFs, have been shown to perform better than previously available FFs based on prediction of vapor-liquid equilibria (VLE). In this work, we use molecular dynamics (MD) simulations to investigate the hypothesis that these BO-tuned FFs developed and tuned using only VLE data can capture much of the physics of the molecules they were designed for. We investigate this hypothesis by subjecting the FFs to validation using other properties not used in tuning them.

Thermal conductivities, viscosities, diffusivities, isobaric and isochoric heat capacities, isothermal compressibility, thermal expansion coefficient, thermal pressure coefficient, Joule-Thomson coefficient, sonic speed and surface tension were calculated using the recommended BO-tuned FFs. MD results were compared with data from the National Institute of Standards and Technology (NIST).

The LAMMPS code was used for all MD simulations. Non-equilibrium molecular dynamics (NEMD) methods were used to compute thermal conductivity and viscosity. Thermal expansion coefficients were computed using multiple simulations in the Isothermal-Isobaric (NPT) ensemble. The Isobaric heat capacities were calculated by decomposing the total heat capacity into excess and ideal components. The excess components were calculated using MD simulations in the NPT ensemble. The ideal heat capacities were calculated from quantum mechanical (QM) calculations using the Psi4 QM package. Diffusivities were calculated using appropriate Einstein relations in the canonical ensemble with the PyLAT package. The canonical ensemble was also used for calculating the thermal pressure coefficients and the isochoric heat capacities. All other properties were obtained using standard thermodynamic relationships.

Results show generally excellent agreement with experiments for all the recommended BO-tuned FFs, thereby confirming the hypothesis. The average of the mean absolute percentage errors (MAPEs) over multiple state points for the properties used in the analysis was within 10 % for most of the BO-tuned FFs with several properties accurately captured to within 5 % from experimental data across multiple state points.

We then proceed to compare these BO-tuned FFs with expert tuned FFs as well as off-the-shelf generalized FFs like GAFF for R32 and R125 as case studies. We computed several thermodynamic and transport properties using these expert-tuned and/or generalized FFs for R32 and R125, as was done for the BO-tuned FFs. We demonstrate that the BO-tuned FFs are superior to the expert-tuned and/or generalized FFs in terms of their ability to accurately capture and model the properties of the molecules they were designed for.

Furthermore, we also validated the transferability of the FFs to state points outside the temperature and pressure ranges for which they were tuned using Bayesian Optimization. We performed this validation using R14, R50, R134a, R143a and R170 as case studies. We found that the BO-tuned FFs provided similarly accurate predictive capabilities of properties at the state points outside their tuning ranges as they did for predictions within their tuning ranges. This completes the successful and rigorous validation of these newly developed FFs for the molecules studied in this work.

All the molecules studied in this work are of key importance in the refrigeration industry. Some of these refrigerant molecules contribute excessively to global warming. The Kigali agreement of 2016 seeks to globally phase out these environmentally harmful refrigerant components. One of the challenges of this phase out is the need to reclaim and repurpose these refrigerant components and thus prevent them from being flared. A first technical step for achieving this is in the separation of these refrigerant mixtures into their components. Thus, a near future application of the outcomes of this work is in the use of the validated FFs to simulate systems such as refrigerant mixtures, refrigerants with ionic liquids (ILs) and refrigerants in zeolites. This will help to design optimal separation technologies of refrigerant mixtures.

This work provides a comprehensive set of simulation data for FF validation on a wide range of thermophysical, transport and structural properties across multiple state points for seven industrially relevant molecules. It provides a foundation for similar studies on other molecules.