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

Area of Science:

• Computational Chemistry
• Chemical Engineering
• Machine Learning Applications

Background:

• Accurate solubility prediction in supercritical carbon dioxide (scCO2) is essential for optimizing experimental design and reducing costs in chemical processes.
• A comprehensive database of 31,975 solubility records has been compiled, supporting the development of predictive models for various chemical compounds, especially drug-like substances.

Purpose of the Study:

• To develop and evaluate a domain-aware machine learning approach for predicting solubility in scCO2.
• To incorporate thermodynamic properties governing phase transitions into solubility prediction models.
• To compare the effectiveness of the CatBoost algorithm against a graph-based architecture with directed message passing.

Main Methods:

• A domain-aware machine learning strategy was employed, integrating thermodynamic features relevant to phase transitions.
• Predictive models were built using the CatBoost algorithm and a graph-based architecture with directed message passing.
• Auxiliary solute properties such as melting point, critical parameters, enthalpy of vaporization, and Gibbs free energy of solvation were also predicted.

Main Results:

• The study demonstrated the effectiveness of incorporating domain-specific thermodynamic features for enhancing scCO2 solubility prediction accuracy.
• Both CatBoost and graph-based models showed promise, with domain-specific features significantly improving predictive performance.
• Predicted auxiliary properties provided further insights into solute behavior in scCO2.

Conclusions:

• Integrating domain-specific thermodynamic features significantly enhances the predictive accuracy of solubility in scCO2.
• The developed models show good generalization capabilities, even for compounds outside the initial training domain.
• This approach streamlines experimental design by providing reliable solubility predictions for drug-like substances and other compounds.