Machine learning-assisted computation of water activity for ionic liquid-based aqueous ternary elements

Binghui Xu, Kassem Jummaa, Farag M.A. Altalbawy, R. Roopashree, Shelesh Krishna Saraswat, Subhashree Ray, Prakhar Tomar, Sameh Naem, Hadil Hussain Hamza, Aseel Smerat, Ahmad Khalid,
Machine learning-assisted computation of water activity for ionic liquid-based aqueous ternary elements,
Desalination and Water Treatment,
Volume 324,
2025,
101484,
ISSN 1944-3986,
https://doi.org/10.1016/j.dwt.2025.101484.
(https://www.sciencedirect.com/science/article/pii/S1944398625005004)
Abstract: This study develops machine learning models to predict water activity in ionic liquid-based aqueous ternary systems, focusing on systems containing imidazolium-, ammonium-, and phosphonium-based ionic liquids paired with components such as amino acids, salts, water-soluble polymers, or carbohydrates. Using a dataset of 1829 experimental data points, models were trained on eight input parameters: pressure (kPa), temperature (K), ionic liquid critical pressure (kPa), ionic liquid critical temperature (K), ionic liquid acentric factor, second component molecular weight (g/mol), ionic liquid molality (mol/kg), and second component molality (mol/kg). A diverse set of machine learning methods, including Ridge Regression, Lasso Regression, Random Forests, Support Vector Machines, Linear Regression, K-Nearest Neighbors, Decision Trees, Gradient Boosting Machines, Elastic Net, Convolutional Neural Networks, Artificial Neural Networks, LightGBM, CatBoost, Gaussian Processes, and XGBoost, were employed to capture complex relationships between these inputs and water activity. Data reliability was ensured using a Monte Carlo-based outlier detection method. Evaluation metrics revealed that Gradient Boosting, XGBoost, and Random Forest outperformed other methods, achieving R² values of 0.9617, 0.9391, and 0.9617, respectively, on the test set, with mean relative deviation percentages (MRD%) below 0.48 %. SHAP analysis identified ionic liquid molality and second component molality as the primary influencers of water activity, reflecting their role in modulating intermolecular interactions. These results highlight the effectiveness of machine learning in providing accurate and interpretable predictions for optimizing processes like CO2 capture and biomass conversion involving ionic liquid systems.
Keywords: Water Activity Prediction; Machine Learning Models; SHAP Analysis; Monte Carlo Outlier Detection