Machine learning models coupled with ionic fragment σ-profiles to predict ammonia solubility in ionic liquids

Emitting NH 3 into the atmosphere leads to significant air pollution, while NH 3 itself serves as an essential component for fertilizers and refrigerants in industry. Thus, recovering and reusing NH 3 is highly valuable. Ionic liquids (ILs) have shown great potential for NH 3 capture, where the accurate prediction of solubility is a critical point for selecting ILs and designing a separation process. This work combined the Ionic Fragment Contribution (IFC) strategy with machine learning (ML) to develop four models (IFC-ML) to predict NH 3 solubility in ILs. A dataset containing 785 solubility data points, covering 10 cations and 10 anions, was collected. From this dataset, the S 1– S 6 descriptors based on the IFC method were used as inputs for the ML models, together with temperature ( T ) and pressure ( P ). Among the models, the IFC-GBR model was recommended for predicting NH 3 solubility in ILs due to its higher coefficient of determination (R 2 ) of 0.9945 and lower mean squared error (MSE) of 0.0003 than the others. Additionally, in comparison with previous conductor-like screening model for real solvents (COSMO-RS) and extreme learning machine (ELM) methods, the IFC-GBR method showed a more accurate prediction of the NH 3 solubility in ILs over a wider range of temperatures and pressures, providing additional chemical insights into IL-NH 3 system that cations played a more important role for NH 3 solubility. These results highlighted the developed IFC-GBR (gradient boosting regressor) model offered valuable insights for helping guide the process design of absorbing NH 3 through IL-based technology. • This work constructed six new kinds of descriptors for the input of IFC-ML models. • A richer database containing 785 data for NH 3 solubility in various ILs was collected. • IFC-GBR and IFC-RF models demonstrated cations played a more important role in nonmetallic IL-NH 3 systems. • IFC-GBR showed better performance than previously reported models.