Machine Learning-Enabled Development of Accurate Force Fields for Refrigerants

Abstract

Hydrofluorocarbon (HFC) refrigerants with zero ozone-depleting potential have replaced chlorofluorocarbons and are now ubiquitous. However, some HFCs have high global warming potential, which has led to calls by governments to phase out these HFCs. Technologies to recycle and repurpose these HFCs need to be developed. Therefore, thermophysical properties of HFCs are needed over a wide range of conditions. Molecular simulations can help understand and predict the thermophysical properties of HFCs. The prediction capability of a molecular simulation is directly tied to the accuracy of the force field. In this work, we applied and refined a machine learning-based workflow to optimize the Lennard-Jones parameters of classical HFC force fields for HFC-143a (CF3CH3), HFC-134a (CH2FCF3), R-50 (CH4), R-170 (C2H6), and R-14 (CF4). Our workflow involves liquid density iterations with molecular dynamics simulations and vapor–liquid equilibrium (VLE) iterations with Gibbs ensemble Monte Carlo simulations. Support vector machine classifiers and Gaussian process surrogate models save months of simulation time and can efficiently select optimal parameters from half a million distinct parameter sets. Excellent agreement as evidenced by low mean absolute percent errors (MAPEs) of simulated liquid density (ranging from 0.3% to 3.4%), vapor density (ranging from 1.4% to 2.6%), vapor pressure (ranging from 1.3% to 2.8%), and enthalpy of vaporization (ranging from 0.5% to 2.7%) relative to experiments was obtained for the recommended parameter set of each refrigerant. The performance of each new parameter set was superior or similar to the best force field in the literature.

Ionic Liquids for the Separation of Fluorocarbon Refrigerant Mixtures

“…A Gaussian process surrogate model was used to reduce the number of simulations required to efficiently search and screen promising parameter sets from half a million sets generated from Latin hypercube sampling. Multiple high-quality parameter sets were found for each HFC, and each had improved performance relative to the original GAFF model as well as Raabe’s hand-tuned HFC-32 model. , The workflow has been applied to other refrigerants including HFC-143a (CF 3 CH 3 ), HFC-134a (CH 2 FCF 3 ), HC-50 (CH 4 ), HC-170 (C 2 H 6 ), and PFC-14 (CF 4 ) and can be generalized to more complex molecules. The recommended models have the transferability to accurately predict other properties that were not used during force field optimization.…”

“…The development of refrigerant force fields has progressed from simplified united atom models to more precise all-atom models. In certain cases, , machine-learning-assisted models have been employed to better align with experimental data. Studied properties mainly involve VLE properties, dynamic properties, and structural properties.…”

Accelerated Bayesian Inference for Molecular Simulations using Local Gaussian Process Surrogate Models

“…Perhaps the most widely adopted application of GP surrogate models in computational chemistry is for model optimization. In the past decade, GP surrogates of simple thermophysical properties including density, heat of vaporization, enthalpy, diffusivity and pressure have been used for force field design. − However, to our knowledge, there are no Bayesian optimization studies that apply GP surrogate models to thermophysical properties with many independent variables, such as structural correlation functions or electromagnetic spectra. In this work, independent variables (IVs) are defined as the fixed quantities over which a measurement is made (e.g., frequencies along a spectrum or radial positions along a radial distribution function) and the outcomes of those measurements are referred to as quantities-of-interest (QoIs).…”

Assessment and Ranking of Difluoromethane (R32) and Pentafluoroethane (R125) Interatomic Potentials Using Several Thermophysical and Transport Properties Across Multiple State Points

“…The proposed MLD method produced twenty-six and forty-five distinct FF parameter sets for R32 and R125, respectively, that outperformed the best available FFs . Recently, Wang et al applied this machine learning-driven FF optimization procedure to five other refrigerants: R143a, R134a, R50, R-170, and R-14. Other molecule-specific FFs for HFCs include R134a, R152a, and R161 .…”