Reliable frameworks to predict solubility of NO in deep eutectic solvents via machine learning models

This study develops machine learning (ML) frameworks to accurately predict nitric oxide (NO) solubility in deep eutectic solvents (DESs) using key physicochemical parameters, including HBA and HBD densities, DES density and viscosity, HBA mole number, temperature, and pressure. A dataset containing 292 experimental data points was employed and screened for anomalies using a Monte Carlo-based outlier detection approach prior to model development. Multiple ML techniques, including ANN, CNN, Gaussian Process (GP), Random Forest, XGBoost, LightGBM, SVR, and regression-based methods, were evaluated. Among all investigated models, the Gaussian Process framework achieved the highest predictive accuracy, with R² values of 0.9991, 0.9992, and 0.9973 for training, validation, and testing datasets, respectively, together with very low mean squared errors (0.0092–0.0504). ANN and CNN models also demonstrated strong predictive capability with testing R² values above 0.985. SHAP-based feature importance analysis revealed that HBA density, temperature, and pressure were the most influential variables governing NO solubility behavior in DESs. The results demonstrate that advanced nonlinear ML models can provide reliable and interpretable predictions of NO solubility, offering a practical tool for accelerating DES screening and optimization in gas capture applications.