A multi-head attention-enhanced fully connected neural network for predicting the thermodynamic properties of low-GWP refrigerant R450A

Abstract This study presents a novel computational framework for predicting the thermodynamic properties of R450A, a promising refrigerant that has a low-global warming potential (GWP) of 547. A fully connected neural network (FCNN) integrated with a multi-head attention mechanism has been developed to accurately model enthalpy, entropy, and specific volume under a wide range of temperature and pressure conditions. This study presents the first integration of a multi-head attention mechanism into an FCNN to predict refrigerant thermodynamic properties. The model employs dense layers to extract and transform input features, followed by a multi-head attention layer that dynamically prioritizes important interactions. The model concludes with layer normalization and a final output layer for continuous value estimation. Grid search is used to optimize the number of hidden layer neurons and head numbers in the multi-head attention mechanism of the FCNN model. The proposed FCNN-attention model demonstrates high predictive capability, with a coefficient of determination ( R 2 ) above 0.998 for enthalpy, 0.981 for entropy, and 0.993 for specific volume. The model’s predictions indicate an enthalpy mean absolute percent error (MAPE) of approximately 0.2% in vapor and 1.6% in liquid, an entropy MAPE of around 0.4% in vapor and 1.3% in liquid, and a specific volume root mean square error (RMSE) of less than 0.06 m 3 kg −1 . It has been determined that these errors are well within the established engineering tolerances. Consequently, the design of environmentally friendly refrigeration is ensured to be accurate.