AI-driven predictive modeling of supercritical CO2 solubility using cheetah-optimized machine learning

Abstract

Artificial intelligence (AI) and machine-learning models provide powerful data-driven tools for modeling complex nonlinear systems and predicting process behavior from limited experimental data. In this study, an AI-driven predictive framework is developed for modeling paracetamol solubility and solvent density in supercritical carbon dioxide (SC-CO 2 ). A dataset comprising 40 experimental observations was employed, with pressure and temperature used as input variables and paracetamol mole fraction and SC-CO 2 density considered as target outputs. Three machine-learning regression algorithms—Gaussian Process Regression (GPR), Tweedie Regression (TDR), and K-Nearest Neighbors (KNN)—were developed and comparatively evaluated. To enhance predictive performance and automate model configuration, the Cheetah Optimizer (CO), a nature-inspired metaheuristic optimization algorithm, was integrated into the modeling framework for hyperparameter optimization. Model performance was assessed using the coefficient of determination ( R 2 ), mean squared error (MSE), and mean absolute error (MAE). The results demonstrated that TDR provided the highest predictive accuracy for paracetamol mole fraction, achieving an R 2 of 0.985, whereas GPR exhibited the best performance for SC-CO 2 density prediction, with an R 2 of 0.991. The developed framework successfully captured the nonlinear relationships between operating conditions and the predicted properties while demonstrating that the optimal learning algorithm depends on the target variable. These findings highlight the potential of AI-assisted optimization and machine-learning approaches for data-driven prediction and intelligent modeling of complex supercritical-fluid processes.