Accurate Machine Learning for Predicting the Viscosities of Deep Eutectic Solvents

Area of Science:

• Materials Science
• Chemical Engineering
• Computational Chemistry

Background:

• Deep eutectic solvents (DESs) are promising eco-friendly solvents for industrial mass and heat transfer.
• Predicting DES viscosity is challenging due to the vast number of possible formulations and environmental factors.
• Experimental viscosity measurements for all potential DESs are infeasible, hindering their widespread application.

Purpose of the Study:

• To develop accurate and rapid machine learning (ML) models for predicting DES viscosity.
• To facilitate the rational design of DESs for specific industrial applications.
• To provide a computational tool for understanding DES properties.

Main Methods:

• Developed and compared three ML models: Support Vector Regression (SVR), Feed Forward Neural Networks (FFNNs), and Categorical Boosting (CatBoost).
• Utilized a comprehensive dataset of over 670 DESs across a wide temperature range (278.15–385.25 K).
• Employed Quantum chemistry-based, COSMO-RS-derived sigma profile (σ-profile) features as inputs for ML models.
• Interpreted ML model predictions using SHapley Additive Explanation (SHAP) analysis.

Main Results:

• The CatBoost model demonstrated excellent predictive performance with R² of 0.99, low RMSE, and AARD of 5.22% on external test sets.
• 98% of data points fell within 15% of average absolute relative deviations, indicating high accuracy.
• ML models significantly outperformed baseline regression methods (multilinear and two-factor polynomial regression).

Conclusions:

• Machine learning models, particularly CatBoost, provide accurate and rapid prediction of DES viscosity.
• These models, informed by quantum chemistry features, can accelerate the discovery and design of novel DESs.
• The developed approach aids in overcoming experimental limitations and promotes the industrial adoption of designer solvents.