摘要
Accurate prediction of amine pK a values is critical for the rational design of CO2 capture solvents, as basicity is one of the fundamental properties strongly influencing absorption kinetics and capacity. This work advances predictive modeling by refining implicit solvation protocols and introducing a quantum-informed delta-learning framework. The optimized implicit model yields consistent accuracy for primary and secondary amines but systematically overestimates the occurrence of tertiary amines. To address this limitation, the delta-learning approach integrates first-principles predictions with machine-learning corrections, achieving a mean absolute error of 0.36 pK a across 116 diverse amines while maintaining low uncertainty, outperforming both first-principles and purely data-driven models. It delivers reliable performance even with limited training data and is insensitive to the choice of the machine learning model. These results demonstrate that coupling physics-based solvation models with data-driven corrections improves both the accuracy and generalizability. The proposed framework provides a scalable and data-efficient tool for CO2 capture solvent screening.