Machine learning–based prediction of paracetamol solubility and CO₂ density in supercritical systems using artificial rabbits optimization

The accurate prediction of solubility and solvent properties in supercritical CO₂ systems remains a critical challenge in pharmaceutical process design due to the nonlinear and coupled effects of temperature and pressure. This study proposes a novel artificial intelligence-based modeling framework for predicting solvent density and paracetamol mole fraction under supercritical conditions using temperature and pressure as input variables. Three regression models—Multilayer Perceptron (MLP), Support Vector Regression (SVR), and Tweedie Regression (TDR)—were developed and systematically optimized using the Artificial Rabbits Optimization (ARO) algorithm. Unlike conventional single-model studies, this study provides a comparative and optimization-driven evaluation of both nonlinear machine learning models and statistically grounded regression methods under identical conditions. The results demonstrate that the ARO-optimized MLP model achieves superior predictive performance for both solvent density (R² = 0.99898) and mole fraction (R² = 0.96555), outperforming SVR and TDR models across all evaluation metrics. The study further reveals clear nonlinear dependencies of solubility and density on pressure and temperature, which are effectively captured through data-driven modeling and visualized via contour-based response surfaces. The main innovation of this work lies in the integration of a metaheuristic optimization strategy (ARO) with multiple regression paradigms to establish a unified and systematic framework for supercritical solubility prediction. This approach provides both high predictive accuracy and interpretable process insights, supporting early-stage optimization of pharmaceutical manufacturing in supercritical CO₂ environments.