Predicting fluorinated gas solubility in deep eutectic solvents using machine learning and deep learning approaches

This paper introduces a data-driven framework by deploying machine learning (ML) and deep learning (DL) architectures to estimate fluorinated gas (F-gas) absorption capacities in deep eutectic solvents (DESs). The proposed approach aims to support the screening and design of DESs for gas capture applications by correlating thermodynamic and molecular descriptors with experimentally measured solubility data. Four predictive models, namely Gradient Boosting (GBoost) and Random Forest (RF) as ML techniques, and TabNet and GrowNet as DL architectures, were developed and compared using a dataset of 440 experimental data points, incorporating temperature, pressure, and critical molecular descriptors of both F-gases and DESs as input parameters. Further evaluation was conducted using SHapley Additive exPlanations (SHAP) analysis and the leverage approach. DL models, especially GrowNet, demonstrated high accuracy and strong generalization on test data. GrowNet outperformed all others, achieving the lowest test errors (MSE = 0.0046%, R² = 0.9985). SHAP analysis revealed pressure and temperature as the most important variables affecting the model predictions, with critical temperature and acentric factor of F-gases also playing significant secondary roles. The leverage method confirmed model reliability, with over 98% of data points lying within its applicability domain. The results demonstrate that ML- and DL-based methods can provide effective predictive tools for estimating F-gas solubility in DES systems.