Physics-informed COSMO-SAC-machine learning framework for predicting ammonia solubility in choline chloride-based deep eutectic solvents

A physics-informed modeling framework was developed by integrating COSMO-SAC thermodynamic calculations with machine learning to predict ammonia (NH 3 ) solubility in choline chloride (ChCl)-based deep eutectic solvents (DESs). Six machine learning models, including simple Multi-Layer Perceptron (MLP), Multiple Linear Regression (MLR), k-Nearest Neighbors Regression (kNN), Support Vector Regression (SVR), Random Forest Regression (RFR), and eXtreme Gradient Boosting Regression (XGB), were evaluated using COSMO-SAC-predicted solubility, DES density, and pressure as input features. The XGB model exhibits the highest predictive accuracy (R 2 = 0.995) across 611 experimental data points, effectively capturing nonlinear patterns while maintaining interpretability through feature importance analysis. COSMO-SAC emerges as the dominant predictor, with DES density and pressure serving as corrective parameters for free volume and operating conditions. Sensitivity analysis demonstrated the utilization of the model in assessing the effects of hydrogen bond donor (HBD) structure, molar ratio, and operating conditions, with containing high HBD content and hydroxyl-rich species showing the highest NH 3 uptake. The framework provides reliable solubility predictions and offers an interpretable tool for solvent screening and process design in sustainable NH 3 recovery. Physics-informed COSMO-SAC-machine learning framework for predicting ammonia solubility in deep eutectic solvents • COSMO-SAC-VLE model tends to overpredict NH 3 solubility in ChCl-based DESs. • Adding physical parameters improves model accuracy over using pure COSMO-SAC-VLE. • Group-out validation shows XGB model generalizes well to common ChCl-based DESs. • XGBoost outperforms other models by leveraging physics-informed parameters. • XGBoost sensitivity shows high HBD and hydroxyl-rich DESs enhance NH 3 solubility.