Theoretical understanding of pharmaceutics solubility in supercritical CO2: Thermodynamic modeling and machine learning study

Both experimental and theoretical studies of drug solubility in supercritical CO₂ are essential for optimizing supercritical techniques in pharmaceutical formulation. This study explores the solubility of Chlorothiazide and Chloroquine, two drugs with low solubility and bioavailability, in supercritical CO₂ using various theoretical models, including empirical models, a non-cubic equation of state (PC-SAFT), a recently developed expanded liquid model proposed by Sodeifian, the regular solution model, and artificial neural network method (ANN). The reliability of these models in predicting and analyzing the solubility of the desired drugs in supercritical CO₂ is validated by comparing their results with experimental data previously recorded at temperatures between 308 K to 338 K and pressures ranging from 130 bar to 290 bar for Chlorothiazide and 120 bar to 400 bar for Chloroquine. The study found that all empirical and thermodynamic models provided satisfactory accuracy in correlating the solubility of both drugs, with AARD values below 10 %. For Chlorothiazide, the most accurate models were the empirical models (AARD%= 1.78–2.72), followed by PC-SAFT (AARD% = 5.23), the Sodeifian model (AARD% = 8.257), and the regular solution model (AARD% = 9.77–10.9). For Chloroquine, the ranking was slightly different, with PC-SAFT (AARD% = 4.15) performing the best, followed by the empirical models (AARD% = 7.1–8.3), the Sodeifian model (AARD% = 7.94), and the regular solution model (AARD% = 8.03–9.74). Moreover, the ANN-based multilayer perceptron (MLP), trained using Bayesian Regularization and Levenberg-Marquardt backpropagation, achieved exceptional accuracy, with over 99 % of predictions closely matching the experimental solubility data in supercritical CO₂.