Graph Attention Networks for Accurate and Explainable Prediction of CO₂ Solubility in Diverse Physical Solvents

The rational design and high-throughput screening of sustainable CO₂ capture solvents are crucial for industrial decarbonization. Recent advances in artificial intelligence provide reliable and interpretable tools to accelerate this process. Here, we developed four neural network models trained on a curated dataset of 1,671 experimentally measured CO₂ solubility values spanning diverse solvent structures and operating conditions. Model development employed Bayesian hyperparameter optimization and 5-fold cross-validation to ensure consistent and robust performance. Among these models, the graph attention network (GAT) achieves the highest predictive accuracy (R ² = 0.9837, RMSE = 0.0342) and exhibits strong robustness to perturbations in temperature, pressure, and solubility values. To enhance interpretability, we introduced a perturbation-based sensitivity analysis specifically tailored for graph-based models, enabling atomic-level attribution of solubility predictions. This analysis revealed polarizability as the dominant positive contributor, while heavy atom degree and hydrogen-bonding features act as primary negative contributors, consistent with physical absorption mechanisms. These findings demonstrate the potential of graph-based neural networks as accurate, robust, and interpretable tools for the data-driven design of green solvents for efficient CO₂ capture.