Data augmentation assisted phase behavior modeling of refrigerant mixtures

İngilizce (OpenAlex)

In this study, vapor–liquid equilibrium (VLE) data, including vapor phase mole fraction and system pressure, are predicted at temperatures not covered by available experimental measurements. Due to the limited availability of experimental data for model development, VLE datasets for refrigerant mixtures obtained from the literature were augmented using curve-fitting techniques before training the model. The augmented data were then employed to train a deep neural network for representing the equilibrium behavior of refrigerant mixtures. To evaluate the prediction performance of the proposed model, experimental data of binary systems at 303.2 K were excluded from the training dataset. The root mean square error (RMSE) was calculated for both vapor-phase mole fraction and pressure to determine the agreement between experimental and predicted results. The equilibrium data predicted by the proposed model for all compositions within the specified temperature range of 250–350 K, including temperatures both below and above the experimental range, were visualized using three-dimensional plots. For the studied binary systems, RMSE values ranged from 0.0002 to 0.0030 for pressure and from 0.0010 to 0.0483 for vapor phase mole fraction. At 303.2 K, which was not included in the training set, the minimum RMSE values for pressure and vapor-phase mole fraction were obtained as 0.0004 and 0.0067, respectively, for the R161 + R1234yf system. The results indicate that the proposed modeling approach provides reliable VLE predictions for binary refrigerant mixtures under limited data conditions. Binary refrigerant mixtures are used as case studies due to the availability of relatively comprehensive experimental datasets. However, the proposed methodology is not restricted to these systems and is generally applicable to other binary vapor–liquid equilibrium systems, provided that suitable experimental data are available.