Improved Solubility Predictions in scCO2 Using Thermodynamics-Informed Machine Learning Models

Abstract

Accurate solubility prediction in supercritical carbon dioxide (scCO₂) is crucial for optimizing experimental design by eliminating unnecessary and costly trials at an early stage, thereby streamlining the workflow. A comprehensive solubility database containing 31,975 records has been compiled, providing a foundation for developing predictive models applicable to a diverse class of chemical compounds, with a particular focus on drug-like substances. In this study, we propose a domain-aware machine learning approach that incorporates thermodynamic properties governing phase transitions to solubility predictions in scCO₂. Predictive models were developed using the CatBoost algorithm and a graph-based architecture employing directed message passing to identify the most effective approach. Furthermore, auxiliary properties of the solute, including melting point, critical parameters, enthalpy of vaporization, and Gibbs free energy of solvation, were predicted as part of this work. The findings underscore the efficacy of incorporating domain-specific thermodynamic features to enhance the predictive accuracy of scCO₂ solubility modeling. The interpretation and the applicability domain assessment have confirmed the qualitative selection of the employed descriptors, demonstrating their ability to generalize to unique compounds that fall outside the defined domain.