Data science for thermodynamic modeling: Case study for ionic liquid and hydrofluorocarbon refrigerant mixtures

Abstract

The sustainable phaseout of high global warming potential hydrofluorocarbon (HFC) refrigerant mixtures requires novel solvents, such as ionic liquids (ILs), for new HFC reuse and recycle technologies. Accurate, predictive, and interpretable thermodynamic models for HFC/IL mixtures are essential for multiscale design schemes aiding this phaseout. Still, there is limited guidance regarding the best thermodynamic model for an HFC/IL system. We propose a rigorous thermodynamic model selection and analysis workflow for HFC/IL mixtures which harnesses data science tools – visualization, nonlinear regression, Akaike information criteria, Fischer information matrix (FIM)-based identifiability and uncertainty analyses, and model-based design of experiments methods – to evaluate the accuracy, predictive capability, and interpretability of a thermodynamic model. The open-source IDAES™ platform facilitates training and comparison of sixteen candidate HFC/IL thermodynamic models, including two cubic equations of state, Peng–Robinson and Soave–Redlich–Kwong, and eight variations on temperature dependence within a classical van der Waals mixing rule. We apply this analysis to models for three HFC/IL systems: HFC-32/[emim][TF2N], HFC-125/[emim][TF2N], and HFC-32/[bmim][PF6]. For these mixtures, we observe that models with a temperature dependent mixing rule are consistently ranked higher by Akaike information criteria for model selection. However, these models may still have high parameter uncertainty and correlation, indicating that data at multiple temperatures should be obtained. This result differs from the current practice of generating single isotherm dataset for most new HFC/IL mixtures. Additionally, we find that the most valuable experiments are taken at the bounds of composition, temperature (e.g., 273 and 348 K), and pressure (e.g., 1 MPa) measurements. This analysis guides data generation efforts, showing that optimally selected measurements across multiple temperatures are adequate for regressing thermodynamic models for multiscale process design.