A comparative analysis of density-based and neural network modeling for predicting the solubility of organic compounds in supercritical carbon dioxide

This study investigates the estimation of solute solubility in supercritical carbon dioxide (SC-CO2) within a pressure and temperature range of 80 bar to 490.29 bar and 308 K to 423 K. We propose a novel empirical model that establishes a correlation between relevant parameters and the targeted solubility. A feature importance algorithm facilitated the development of this empirical model. The model's accuracy is comprehensively evaluated using 40 published experimental datasets, with an average absolute relative deviation (AARD) of 9.9 %. It demonstrates superior performance compared to 12 previously established models. Furthermore, a fine-tuned artificial neural network (ANN) is developed to harness the unique capabilities of machine learning techniques. The ANN outperforms the proposed model, achieving a significantly lower AARD% of 4.38. This outcome emphasizes the potential of machine learning techniques, particularly ANNs, for achieving superior accuracy.