Machine-Learning-Assisted Design of Deep Eutectic Solvents Based on Uncovered Hydrogen Bond Patterns.

Area of Science:

• Materials Science
• Computational Chemistry
• Chemical Engineering

Background:

• Non-ionic deep eutectic solvents (DESs) are versatile designer solvents with broad applications.
• Current methods for discovering new DES candidates rely on intuition or trial-and-error, leading to inefficiencies.
• Hydrogen bonds (HBs) are critical for DES formation, but distinguishing features between DES and non-DES systems are not well-defined.

Purpose of the Study:

• To identify specific hydrogen bond (HB) features that differentiate DES from non-DES systems.
• To develop and validate machine learning (ML) models for predicting DES formation.
• To facilitate the discovery of novel DES candidates.

Main Methods:

• Analysis of hydrogen bond properties from molecular dynamics (MD) simulations of 38 known DES and 111 non-DES systems.
• Development of 30 ML models using ten algorithms and three types of HB-based descriptors.
• Benchmarking model performance using receiver operating characteristic (ROC)-area under the curve (AUC) and feature importance analysis.
• Validation of ML models using experimental data from 34 additional systems.

Main Results:

• DES systems exhibit a greater imbalance in intra-component HBs and more numerous, stronger inter-component HBs compared to non-DES systems.
• ML models effectively captured these HB features, with performance consistent between simulation analysis and feature importance.
• The extra trees forest model achieved a high ROC-AUC of 0.88 in experimental validation, demonstrating predictive power.

Conclusions:

• Hydrogen bond characteristics are crucial determinants of deep eutectic solvent formation.
• Machine learning models, informed by HB analysis, offer a powerful and efficient tool for discovering new DES.
• This approach significantly enhances the prediction of DES formation, overcoming limitations of traditional methods.