Screening Environmentally Benign Ionic Liquids for CO₂ Absorption Using Representation Uncertainty-Based Machine...

Area of Science:

• Materials Science
• Computational Chemistry
• Environmental Science

Background:

• Machine learning (ML) models are vital for screening ionic liquids (ILs) with desirable properties like low viscosity, low toxicity, and high CO2 absorption for climate change mitigation.
• Predictions from ML models can be unreliable when candidate ILs are outside the model's training data range (extrapolation zone), leading to inefficient material discovery.

Purpose of the Study:

• To develop and validate a "representation uncertainty" (RU) approach to quantify prediction uncertainty in ML models for IL screening.
• To improve the reliability of identifying promising IL candidates for CO2 capture by addressing model extrapolation issues.

Main Methods:

• Employed four distinct IL representations: molecular fingerprint, descriptor, image, and graph, each feeding into a separate ML model.
• Quantified prediction uncertainty using the RU approach, calculated as the standard deviation of predictions across the four diverse ML models.
• Developed ensemble ML models combining predictions from the four representation-based models.

Main Results:

• The RU approach demonstrated superior performance over traditional model uncertainty (MU) in identifying unreliable predictions across viscosity, toxicity, refractive index, and CO2 absorption datasets.
• Ensemble models exhibited enhanced predictive performance compared to individual ML models based on single representations.
• Screened 1420 ILs using the RU approach, identifying 37 promising candidates with desired properties for CO2 absorption.
• Experimental validation confirmed the predictive accuracy of the ensemble model and the effectiveness of the RU approach for CO2 absorption.

Conclusions:

• The "representation uncertainty" approach provides a more reliable method for screening and designing ionic liquids, accelerating the discovery of materials for carbon capture.
• Ensemble modeling, guided by the RU approach, significantly improves predictive accuracy and aids in identifying high-performance ILs.
• This work offers a new perspective for developing robust ML models and enhances the discovery pipeline for functional materials.