A machine learning technique based on group contributions to calculate the solubility of dye molecules in supercritical CO2

Abstract

A new modeling approach is proposed for the estimation of dye solubility in supercritical carbon dioxide. This approach is based on the synergy obtained from a group contribution approach and an artificial neural network. 33 functional groups have been proposed to create the dye molecules, and a database containing 1917 experimental solubility points of 66 molecules was utilized to establish the best architecture of the artificial neural network. The results showed that this new approach could calculate the dye solubility in this supercritical fluid with modeling errors lower than 10 % under different pressure and temperature conditions. Specifically, it showed a competitive performance with R2 > 0.99, AARDlny = 0.91 ± 1.2 % and AARDy = 10.4 ± 17.9 %. The minimum and maximum values of ARDlny and ARDy were 4.0E-04 – 16.8 % and 5.0E-03 – 386 %. This contribution group – artificial neural network approach outperformed classical empirical and semi-empirical equations, and other surrogate models based on artificial intelligence, even those using more detailed information of tested compounds (e.g., molecular descriptors that should be obtained from computational chemistry calculations). It is an interesting option to be incorporated into process simulators to improve the process design of dyeing technologies using supercritical carbon dioxide.