Machine Learning-Enabled Development of Accurate Force Fields for Refrigerants.

Area of Science:

• Thermodynamics
• Computational Chemistry
• Materials Science

Background:

• Hydrofluorocarbons (HFCs) are widely used refrigerants replacing ozone-depleting substances.
• Some HFCs possess high global warming potential, necessitating phase-out and recycling technologies.
• Accurate thermophysical property data is essential for developing effective HFC management strategies.

Purpose of the Study:

• To refine a machine learning-based workflow for optimizing Lennard-Jones parameters in classical force fields for HFCs.
• To enhance the predictive accuracy of molecular simulations for key thermophysical properties of refrigerants.
• To develop accurate force fields for HFC-143a, HFC-134a, R-50, R-170, and R-14.

Main Methods:

• Employed a machine learning workflow integrating molecular dynamics (MD) and Gibbs ensemble Monte Carlo (GEMC) simulations.
• Utilized support vector machine classifiers and Gaussian process surrogate models to efficiently screen parameter sets.
• Iteratively optimized Lennard-Jones parameters based on liquid density and vapor-liquid equilibrium (VLE) data.

Main Results:

• Achieved excellent agreement between simulated and experimental thermophysical properties for optimized HFC force fields.
• Reported low mean absolute percent errors (MAPEs) for liquid density (0.3–3.4%), vapor density (1.4–2.6%), vapor pressure (1.3–2.8%), and enthalpy of vaporization (0.5–2.7%).
• The developed force fields demonstrated superior or comparable performance to existing literature models.

Conclusions:

• The machine learning-driven workflow effectively optimizes HFC force fields for accurate thermophysical property prediction.
• Accurate force fields are critical for advancing molecular simulation capabilities in refrigerant management and development.
• This approach significantly accelerates the discovery of optimized parameters, saving substantial computational resources.