Designing solvent systems using self-evolving solubility databases and graph neural networks.

Area of Science:

• Computational chemistry
• Machine learning
• Chemical engineering

Background:

• Accurate solubility prediction is crucial for designing efficient chemical synthesis and separation processes.
• Current machine learning models for solubility prediction suffer from limitations in accuracy and generalizability.
• Integrating experimental and computational data is key to overcoming these limitations, but requires reconciling discrepancies.

Purpose of the Study:

• To develop a novel approach for creating self-evolving solubility databases.
• To enhance the predictive accuracy and generalizability of solubility models by integrating heterogeneous data.
• To apply the improved model to practical problems in solvent selection and property prediction.

Main Methods:

• Implemented graph neural networks with semi-supervised self-training for solubility prediction.
• Created integrated databases by augmenting experimental data with quantum-mechanical calculations, correcting for discrepancies.
• Dataset augmentation expanded from 11,637 to over 900,000 data points.

Main Results:

• Achieved high accuracy in solubility prediction (mean absolute error ~0.2 kcal mol⁻¹ on the test set).
• Successfully applied the model to solvent selection in organic reactions and separation processes.
• Accurately predicted partition coefficients for lignin-derived monomers and drug-like molecules, demonstrating quantitative utility in Linear Free Energy Relationships.

Conclusions:

• The self-evolving solubility database and graph neural network approach significantly improves solubility prediction.
• This method offers a powerful tool for solvent selection and property prediction in chemistry and materials science.
• Future work can expand predictions to more complex chemical species and systems.