Accurately Predicting Solubility Curves via a Thermodynamic Cycle, Machine Learning, and Solvent Ensembles.

Area of Science:

• Physical Chemistry
• Computational Chemistry
• Chemical Engineering

Background:

• Accurate solubility prediction is vital for pharmaceuticals, agrochemicals, and environmental science.
• Current methods struggle with diverse solvents and temperatures, hindering process innovation.
• Estimating solubility limits is essential for drug formulation, synthesis, purification, and crystallization.

Purpose of the Study:

• To develop a fast and general method for predicting neutral organic molecule solubilities.
• To improve the design of chemical processes by overcoming solubility estimation challenges.
• To enhance the accuracy and applicability of solubility predictions across various conditions.

Main Methods:

• Utilized a thermodynamic fusion cycle approach.
• Combined machine learning predictions for activity coefficient, fusion enthalpy, and melting point.
• Introduced reference ensembling to leverage existing experimental solubility data.

Main Results:

• The method demonstrated high performance on over 100,000 experimental solubility values.
• Achieved comparable or superior results to existing methods, even at elevated temperatures.
• Reference ensembling improved model robustness and accuracy for solubility predictions.

Conclusions:

• The proposed method offers a significant advancement in predicting organic molecule solubility.
• This approach facilitates more efficient and cost-effective process design in chemistry and related fields.
• The thermodynamic fusion cycle and reference ensembling provide a robust framework for solubility estimation.