Machine learning revolutionizes design of green solvents for carbon capture: A new era for ionic liquid development

With climate change posing an unprecedented global challenge, the demand for environmentally friendly solvents in green chemical processes and carbon dioxide capture has surged. Ionic liquids (ILs) have emerged as promising "designer solvents" due to their negligible volatility, broad liquid temperature range, and exceptional thermal stability. However, the immense chemical space of ILs—with theoretically up to 10¹⁸ possible cation-anion combinations—has created a critical bottleneck in efficient screening and design. Traditional experimental methods are costly and time-consuming, while theoretical calculations like molecular dynamics and quantum chemistry remain computationally prohibitive for large-scale screening. This urgent need for accelerated discovery has set the stage for a transformative technological leap.

A comprehensive review published in ENGINEERING Energy (formerly Frontiers in Energy) by researchers from Nanjing Tech University systematically evaluates the application of machine learning (ML) in ionic liquid design. The study maps the evolution from traditional trial-and-error methods to artificial intelligence-driven quantitative structure-property relationship (QSPR) modeling. The team examined cutting-edge ML techniques including neural networks, random forests, support vector machines, and Gaussian process regression, while emphasizing the critical role of molecular descriptor selection—from simple group contributions to complex quantum chemistry-based features. A key innovation highlighted is the integration of molecular dynamics (MD) and density functional theory (DFT) calculations with ML models to create interpretable "white-box" predictions that reveal underlying physical mechanisms.

Key research findings include:

1. High-accuracy property prediction 

1. For viscosity prediction, deep neural networks achieved R² values exceeding 0.99 on datasets with over 8,600 data points.
2. In CO₂ solubility prediction, graph neural networks reached R² = 0.9884.
3. An integrated CatBoost model further improved performance to R² = 0.9925 by combining group contribution and molecular structure descriptors.

1. Efficient multi-objective optimization 
   1. A multi-objective optimization framework successfully balanced viscosity, toxicity, and absorption capacity.
   2. This approach identified 37 optimal ionic liquids from a pool of 1,420 candidates.
2. Dramatic acceleration of computational screening 
   1. Machine learning reduced process optimization time from 1,218 seconds to just 4.12 seconds, compared with conventional simulation methods.
3. Integration of physics-based and data-driven methods 
   1. The combination of MD and DFT calculations with ML models enables more interpretable predictions.
   2. This “white-box” approach provides insights into the underlying structure–property relationships at the molecular level.

This work establishes a systematic framework that fundamentally transforms IL development from serendipitous discovery to predictive design. By bridging experimental data, theoretical computations, and AI algorithms, the approach enables rapid screening of millions of candidate solvents while providing atomic-level insights into structure-property relationships. The implications are profound: accelerating the deployment of sustainable solvents for carbon capture, advancing green chemistry applications, and facilitating the design of safer electrolytes for next-generation batteries. As the world races toward carbon neutrality, this ML-driven paradigm offers a powerful toolkit to unlock the full potential of ionic liquids in building a sustainable chemical industry, potentially reducing development cycles from years to months while minimizing environmental impact through computational prioritization.

Journal: ENGINEERING Energy (formerly Frontiers in Energy)

Read the full article for free: https://rdcu.be/eQgK7

Cite this article: Shao, Y., Wang, Z., Wang, L. et al. Machine learning-based structure—property modeling for ionic liquids design and screening: A state-of-the-art review. Front. Energy 19, 815–838 (2025). https://doi.org/10.1007/s11708-025-1011-7