(130g) Systematically Building and Optimizing a Generalized Hydrofluorocarbon Refrigerant Force Field

The phase-out of high global warming potential (GWP) hydrofluorocarbon (HFC) refrigerants has been mandated by the Kigali amendment and estimates suggest that this will represent a $270 billion dollar industry by 2050 [1]. However, common refrigerants are often near azeotropic mixtures, which makes it difficult to separate them into their high and low GWP components. Current work focuses on deep eutectic solvents (DESs) [2] as entrainers for extractive distillation to perform this separation. This requires accurate thermophysical properties of the refrigerant and DES systems for which experimental data is often limited [2]. Molecular simulation (MS) methods can be used to determine these properties, but this requires accurate system energy (force field (FF)) models which are difficult to generate. Generalized FFs which are applicable to multiple molecules are ideal as they reduce the number of FFs that must be calibrated, but these models are often inaccurate [3,4] and require years or decades to develop [5,6]. Alternatively, single-use FFs trained by expert-guided trial-and-error are more accurate, but still suboptimal and computationally wasteful to generate [7,8]. Recent work has shown that Gaussian process (GP)-based approaches for FF calibration, in which GP models are trained to reproduce the property predictions estimated by MS, are particularly effective [3,4].

Therefore, in this work, we used data-science techniques to guide the creation of a generalized FF for one- and two-carbon HFC refrigerants and used nonlinear regression (NLR) to automatically optimize its Lennard-Jones (LJ) parameters. Eigen-decomposition and estimability analysis [9] provided mathematical insight into which parameters were most important to objective minimization and into the reliability with which each parameter could be estimated to iteratively improve the structure of the generalized FF. GP surrogate models were used to reduce the computational effort of NLR from a timescale of years to hours. Our work suggested that six optimized LJ pairs recommended via data science techniques significantly improved on the predictions by the commonly-used Generalized Amber Force Field (GAFF). Parameter optimization also allowed simpler models (four LJ pairs) to achieve superior accuracy to GAFF. Lastly, our work demonstrated that adding more LJ pairs ad-hoc (as is typically done) can potentially lead to overfit FF models. Thus, the development of this method is a useful tool to quickly, accurately, and systematically create and calibrate generalized FFs.

References

[1] EPA. (2021). Draft Regulatory Impact Analysis for Phasing Down Production and Consumption of Hydrofluorocarbons (HFCs).

[2] Hansen, B. B., Spittle, S., Chen, B., Poe, D., Zhang, Y., Klein, J. M., Horton, A., Adhikari, L., Zelovich, T., Doherty, B. W., Gurkan, B., Maginn, E. J., Ragauskas, A., Dadmun, M., Zawodzinski, T. A., Baker, G. A., Tuckerman, M. E., Savinell, R. F., & Sangoro, J. R. (2021). Deep Eutectic Solvents: A Review of Fundamentals and Applications. Chem. Rev., 121(3), 1232–1285. https://doi.org/10.1021/acs.chemrev.0c00385

[3] Befort, B. J., Defever, R. S., Tow, G. M., Dowling, A. W., & Maginn, E. J. (2021). Machine Learning Directed Optimization of Classical Molecular Modeling Force Fields. J. Chem. Inf. Model, 19, 28. https://doi.org/10.1021/acs.jcim.1c00448

[4] Wang, N., Carlozo, M. N., Marin-Rimoldi, E., Befort, B. J., Dowling, A. W., & Maginn, E. J. (2023). Machine Learning-Enabled Development of Accurate Force Fields for Refrigerants. J. Chem. Theory Comput. https://doi.org/10.1021/acs.jctc.3c00338

[5] Jorgensen, W. L., Maxwell, D. S., & Tirado-Rives, J. (1996). Development and Testing of the OPLS All-Atom Force Field on Conformational Energetics and Properties of Organic Liquids. J. Am. Chem. Soc., 118, 11225–11236. https://pubs.acs.org/sharingguidelines

[6] Wang, J., Wolf, R. M., Caldwell, J. W., Kollman, P. A., & Case, D. A. (2004). Development and Testing of a General Amber Force Field. J. Comput. Chem., 25(9), 1157–1174. https://doi.org/10.1002/jcc.20035

[7] Peguin, R. P. S., Kamath, G., Potoff, J. J., & Da Rocha, S. R. P. (2009). All-Atom Force Field for the Prediction of Vapor-Liquid Equilibria and Interfacial Properties of HFA134a. J. Phys. Chem. B, 113(1), 178–187. https://doi.org/10.1021/jp806213w

[8] Raabe, G., & Maginn, E. J. (2010). A Force Field for 3,3,3-Fluoro-1-propenes, Including HFO-1234yf. J. Phys. Chem. B, 114(31), 10133–10142. https://doi.org/10.1021/jp102534z

[9] Yao, K. Z., Shaw, B. M., Kou, B., McAuley, K. B., & Bacon, D. W. (2003). Modeling Ethylene/Butene Copolymerization with Multi-Site Catalysts: Parameter Estimability and Experimental Design. Polym. React. Eng., 11(3), 563–588. https://doi.org/10.1081/PRE-120024426