Data-driven modeling of supercritical CO₂ processing with optimized machine learning: prediction of clobetasol propionate solubility.

Machine learning can be used to support data-driven modeling of supercritical CO₂ processing. The method of machine learning modeling is applied in this work for evaluation of small-molecule processing under supercritical conditions. As a necessary step, the solute solubility in the solvent is evaluated via different models. The purpose of this investigation is to develop models for the solubility of Clobetasol Propionate (CP) based on the two parameters of temperature and pressure as the model's features. Adaptive Boosting (AdaBoost) is utilized in this study on top of three core models: Decision Tree (DT), Lasso, and Gaussian Process Regression (GPR). The models are tuned through the firefly algorithm (FA) to determine their unknown coefficients. The final models are called FB-DT (FA-optimized Boosted DT), FB-GPR (FA-optimized Boosted GPR), and FB-LASSO (FA-optimized Boosted LASSO) and have R2 scores of 0.934, 0.977, and 0.813, respectively, for fitting the solubility data of CP. Accordingly, FB-GPR is the most accurate model, with an RMSE of 1.11 × 10⁻2. The developed models demonstrate great performance in estimating CP solubility in supercritical CO₂ under different temperature and pressure conditions.