This study introduces machine learning (ML) models to predict CO₂ solubility in mixtures of a deep eutectic solvent (tetrabutylammonium bromide: ethylene glycol, 1:4 molar ratio) and methyldiethanolamine (MDEA). To ensure robust extrapolation capabilities and prevent potential data leakage, a rigorous grouped validation strategy, known as “Leave-One-Composition- Out,” was implemented, replacing conventional random data splitting. Four ML algorithms, Gradient Boosting (GBoost), Gaussian Process Regression (GPR), AdaBoost-SVR, and Multilayer Perceptron (MLP), were evaluated using temperature, pressure, and MDEA weight% as input features. Results revealed that the GPR model outperformed all others across all statistical metrics, achieving an R2 value of 0.996 and a mean absolute error (MAE) of 0.2839%. Furthermore, SHAP (SHapley Additive exPlanations) analysis identified pressure as the most significant driving factor for CO2 solubility, while temperature exhibited an inverse correlation. Crucially, this perfect alignment with fundamental principles clearly demonstrates that the model successfully learned the underlying thermodynamic relationships rather than merely memorizing numerical data. Finally, leverage validation confirmed that 99.16% of the data points safely reside within the model’s valid applicability domain. This work establishes the developed GPR framework as a highly reliable and physically consistent approach for predicting CO₂ solubility and optimizing complex multi-component capture systems.

Machine learning prediction of CO₂ solubility in deep eutectic solvent-MDEA mixtures
    Abstract
    1 Introduction
    2 Theory and methodology
        2.1 Dataset construction and description
        2.2 Machine learning models
            2.2.1 Predictive evaluation
            2.2.2 Rationale for model selection
            2.2.3 Multilayer perceptron (MLP)
            2.2.4 Gradient boosting (GBoost)
            2.2.5 Gaussian process regression (GPR)
            2.2.6 Adaptive boosting support vector regression (AdaBoost-SVR)
    2.3 Statistical error analysis
    3 Results and discussion
        3.1 Graphical error analysis
        3.2 Model trend analysis
        3.3 SHAP analysis
        3.4 Implementation of the leverage method
        3.5 Comparative sustainability perspective
    4 Conclusion
    References

• Amir Hossein Sheikhshoaei(Petroleum and Petrochemical Engineering School, Hakim Sabzevari University, Sabzevar, Iran)
• Ali Sanati(Petroleum and Petrochemical Engineering School, Hakim Sabzevari University, Sabzevar, Iran) Corresponding author