Modeling cosolvent effects on solubility in supercritical CO2 using data-driven approaches

Supercritical CO₂ (scCO₂) is an environmentally friendly solvent, but its low polarity limits the solubility of polar compounds. Cosolvents are commonly used to enhance solvation capability, yet comprehensive data-driven studies are scarce. We compiled the largest dataset to date — 4401 experimental solubility records with 22 cosolvents for 93 nonionic solutes, plus 4855 records in pure scCO₂ for the same solutes. Machine learning models (Random Forest, LightGBM, CatBoost, TabPFN) were trained using melting point, enthalpy of vaporization, Abraham parameters, RDKit descriptors, and solvent properties. CatBoost and TabPFN showed the best results, particularly in strict cross-validation. Including solubility data in pure scCO₂ improved predictions significantly, reducing RMSE up to 36%. High-throughput screening of 1958 solute-cosolvent pairs across 12 chemical classes revealed the strongest enhancement for polyphenols, nitrogen heterocycles, aromatic acids, and sulfonamides, with minimal effect for nonpolar compounds. Polar protic and aprotic cosolvents were the most effective, while water often showed negligible or negative effect. The results provide quantitative guidelines for cosolvent selection and demonstrate the value of integrated datasets and interpretable ML for designing supercritical fluid processes.

Subjects

Cosolvent

Machine learning

Solubility

Supercritical carbon dioxide

DDC Class

006.31: Machine Learning

Options

Modeling cosolvent effects on solubility in supercritical CO2 using data-driven approaches