Rationally design the ionic liquid-based absorbents for CO2 absorption using machine learning

Abstract

Ionic liquids (ILs) have shown great potential as CO₂ absorbents, but traditional experimental methods for screening ILs are both time-consuming and labor-intensive. To identify the most effective ILs with desirable properties, the computer-aided design approaches are essential. This study manifested a rapid screening method for ILs using graph neural network (GNN). Approximate 40,000 experimental data points were collected, including CO₂ solubility, viscosity, melting point, and toxicity, to train high-precision GNN regression models and XGBoost classification models. Various strategies were examined for constructing molecular graphs based on IL structures and several GNN models were compared, ultimately proposing an effective GNN architecture that strongly enhances prediction accuracy. To realize high-throughput screening, a database containing 200,000 IL structures was built, further established screening criteria: melting point below 298.15 K, viscosity under 100 MPa·s, and the toxicity with log₁₀EC₅₀, is greater than 3.4 μM. Through this screening process, 12 ILs were successfully identified that showed the highest CO₂ solubility. Density functional theory (DFT) calculations indicated that the main interaction mechanism between CO₂ and ILs involves the C atom of CO₂ being attracted by the O atom of the anion, while the O atoms of CO₂ are attracted by the H atoms of the cation. This data-driven approach simplifies the selection and design of IL absorbents, thereby accelerating the development of CO₂ capture technologies.