Unlocking Structural Insights into CO2 Absorption with Ionic Liquids: A Comparative Study of Machine Learning, Association Rules, and Meta-Learning for Modeling Henry’s Law Constant

The discovery of novel solvents is crucial for advancing green chemistry, and chemoinformatics can accelerate this process. Ionic liquids (ILs) have diverse applications, but understanding the link between their structure and properties is key to optimizing their performance. In this study, the use of emerging algorithms in chemoinformatics was explored, such as the CN2 rule algorithm and model-agnostic meta-learning (MAML), to improve the predictive power of quantitative structure–property relationship (QSPR) models for ILs. The data set on carbon dioxide solubility was leveraged to develop predictive artificial intelligence (AI) models, as these techniques are particularly well-suited for addressing complex problems of high scientific interest, such as mitigating CO₂-related environmental issues. The effectiveness of using molecular fingerprints (MFs) and molecular descriptors (MDs) to represent ILs’ molecular structures was evaluated to find that MFs are more suitable for rule mining. By incorporating fractional free volume (FFV) alongside MDs, nonlinear QSPR models with higher predictive power were obtained. Henry’s law constant prediction utilized a data set composed of 76 ILs evaluated under the same temperature conditions. Model training utilized 80% of the data, while the rest was used for testing. Moreover, integrating COSMO-RS simulated data with MAML allowed for enhanced neural network performance. The gradient boosting model utilizing FFV was found to be the best performing. The findings on chemical data interpretation with rules can inform the development of more efficient solvents. MAML algorithm was further evaluated on a data set regarding solubility expressed in mole fraction for over 6000 data points for meta-training on tasks different than carbon dioxide solubility and fine-tuned on a fraction of over 9000 data points regarding CO₂ mole fraction in a two-component system with ILs. MAML allowed us to obtain a stable model even while utilizing 128 data points for fine-tuning and about 1800 data points for testing.

AI improves understanding of ionic liquid structure, enabling sustainable solvent tailoring and carbon dioxide capture in environment-friendly industrial processes.