Interpretable Feedforward Neural Network and XGBoost-Based Algorithms to Predict CO₂ Solubility in Ionic Liquids

This study investigates the efficacy of feedforward neural network and XGBoost models in screening ionic liquid solvents for CO2 capture. Both models were integrated with either group contribution (GC), molecular structure descriptors (MSD), or hybrid GC–MSD, to enable performance comparisons. It was demonstrated that the XGBoost models performed better over feedforward neural network models, irrespective of descriptor types. Notably, the XGBoost–GC–MSD model outperformed the artificial neural network with group contribution (ANN–GC) and structure encoding multilayer perceptron (SE-MLP) models from previous work, demonstrating an R2 value of 0.98963, MAE of 0.01480, and RMSE of 0.02369. Even the least-performing XGBoost–GC model surpassed earlier ANN–GC and SE-MLP models, showcasing an R2 value of 0.98891. Lastly, the Shapley additive explanation analysis identified the top five influential input features, including pressure, temperature, Chi2v, Chi0n, and BertzCT. These findings provide valuable insights into the molecular determinants affecting CO2 solubility in ionic liquids.