Deep learning models to predict CO₂ solubility in imidazolium-based ionic liquids

Area of Science:

• Chemical Engineering
• Computational Chemistry

Background:

• Accurate prediction of CO2 solubility in ionic liquids is crucial for carbon capture technologies.
• Imidazolium-based ionic liquids are promising solvents, but their thermodynamic properties require robust predictive models.

Purpose of the Study:

• To develop and compare deep learning models for predicting CO2 solubility in imidazolium-based ionic liquids.
• To evaluate the performance of various machine learning algorithms against established physical models.

Main Methods:

• Utilized deep learning models including Bayesian Neural Networks (BNN), Deep Neural Networks (DNN), Gradient Boosting Neural Networks (GrowNet), Tabular Neural Networks (TabNet), Random Forest (RF), and Support Vector Regression (SVR).
• Input parameters included critical pressure, critical temperature, molecular weight, and acentric factor.
• Compared model performance against two PC-SAFT models (cQC-PC-SAFT-MSA (1) and cQC-PC-SAFT-MSA (2)).

Main Results:

• Deep learning models demonstrated superior performance compared to PC-SAFT models.
• The GrowNet model achieved the lowest error, with a root mean square error (RMSE) of 0.0073 and a coefficient of determination (R²) of 0.9962.
• Shapley additive description (SHAP) and Pearson correlation coefficient (PCC) analyses identified pressure (P) as the most significant parameter influencing CO2 solubility.

Conclusions:

• Deep learning, particularly the GrowNet model, offers a highly accurate approach for predicting CO2 solubility in imidazolium-based ionic liquids.
• Understanding the impact of specific parameters like pressure is vital for optimizing CO2 capture processes using ionic liquids.