Screening HFC/HFO and ionic liquid for absorption refrigeration at the atomic scale by the prediction model of machine learning

Abstract

Absorption refrigeration is a highly effective method for utilizing renewable energy, as it can be driven by low-grade heat sources such as industrial waste heat, solar energy, and geothermal energy. The development of new working pairs, particularly hydrofluorocarbon/hydrofluoroolefin refrigerants combined with ionic liquids, has been pivotal in enhancing the cooling efficiency of absorption refrigeration systems. These systems rely on the solubility difference between the generator and absorber, making solubility a crucial factor in determining their efficiency. In this context, we have established an advanced solubility estimation model. This model employs the Attention E(n)-equivariant Graph Neural Network (AEGNN) applied to disconnected graphs, enabling comprehensive learning from both topological and Euclidean structural information. Our atomic-scale model demonstrates significantly higher accuracy than traditional group contribution methods, with an average absolute deviation of 0.003 mol/mol from experimental data. Moreover, it encompasses a much broader range of working pairs. Through extensive screening, we have identified working pairs with high estimated solubility differences. Compared to the high-efficiency working pair identified in the literature, the best-screened working pairs exhibit an improvement in solubility differences by more than 0.3 mol/mol under common operating conditions.