Prediction of busulfan solubility in supercritical CO2 using tree-based and neural network-based methods

Determination of morphine sulfate anti-pain drug solubility in supercritical CO2 with machine learning method

Accurate solute solubility measuring and modeling in supercritical carbon dioxide (ScCO2) would address the best working conditions and thermodynamic boundaries for material processing with this type of fluid. Theory- and data-driven methods are two general modeling approaches. Using theory-driven methods, the solubility is estimated based on the principles of thermodynamics, while data-driven methods are developed by training the algorithms. Despite acceptance of each of these methods, more experimental solubility data are still needed to promote modeling performances. In this study, for the first time, solubility of morphine sulfate is determined and modeled by a set of 13 semi-empirical (theory-driven) and random forest (data-driven) models. Using a laboratory system with an ultraviolet-visible (UV-Vis) spectroscopy, the experimental solubilities including 48 data points were obtained at different temperatures (308-338 K) and pressures (12-27 MPa). The minimum (0.806 × 10-5) and maximum (5.902 × 10-5) equilibrium mole fractions were observed at working pressures of 12 and 27 MPa, respectively, both at the same temperature of 338 K. It was indicated that random forest model (with AARD% of 1.29%) had an excellent predictive performance against semi-empirical models (with AARD% from 9.33 to 19.76%). The results showed that solute molecular weight had the highest effect on random forest modeling. Using modeling results from Chrastil and Bartle models, total and vaporization enthalpies of dissolution of morphine sulfate in ScCO2 were found to be 35.12 and 59.04 kJ/mole, respectively.