Machine learning boosted Design of Ionic Liquid-based Deep Eutectic Solvents for CO2 Capture

Abstract

The escalating atmospheric CO2 concentration demands sustainable and scalable carbon capture technologies. Deep eutectic solvents (DESs), especially those based on ionic liquids (ILs), offer a green and tunable alternative to conventional absorbents. However, their virtually infinite chemical space makes exhaustive experimental screening impossible. Machine learning (ML) has emerged as a promising tool to bypass trial‑and‑error approaches by enabling high‑throughput virtual screening and the discovery of structure‑property relationships. This review provides a systematic analysis of recent advances from 2020 to 2025 in ML‑driven design of DESs for CO2 capture. A comprehensive database of CO2 absorption data for DESs reported during 2008 ~ 2025 is compiled, covering a wide range of chemical compositions and operating conditions. Feature engineering strategies are surveyed, ranging from basic molecular descriptors to advanced quantum chemical and COSMO‑RS derived features. A multi‑dimensional evaluation framework is introduced that assesses model performance not only by predictive accuracy but also by generalization capability and interpretability. The evolution of ML models from linear and kernel methods to ensemble trees and neural networks is critically discussed, with ensemble tree models (e.g., CatBoost, XGBoost) highlighted as the dominant approaches due to their generally high predictive accuracy and interpretability. Finally, key challenges such as data sparsity, model extrapolation, and integration with process simulation are identified, and future directions are proposed, including active learning platforms, physics‑informed models, and generative design across scales. This review offers a systematic technical reference and a methodology to guide the intelligent design of next-generation absorbents for CO2 capture.