Constructing Interpretable Machine Learning Models for Predicting Reaction Equilibrium Constant of Carbonyl Sulfide with Organic Amines

The sustainable utilization of natural gas is crucial for the low-carbon transformation of the global energy mix. Carbonyl sulfide (COS), commonly found in natural gas reservoirs, poses significant hazards and is typically removed via amine-based absorption methods. However, the development of efficient solvents remains challenging due to the unclear reaction characteristics between organic amines and COS involved. This study investigates the reaction equilibrium between organic amines and COS, focusing on substituent effects on the two-step zwitterion mechanism. Using density functional theory (DFT), we analyzed the Gibbs free energy and equilibrium constants for linear aliphatic, cyclic aliphatic, and aromatic amines. Key descriptors, such as the average local ionization energy (ALIE) and Hirshfeld charge, are identified to quantitatively correlate with reaction spontaneity. Cyclic aliphatic amines exhibit the highest reactivity due to favorable electron distribution, while aromatic amines show lower reactivity due to conjugation effects. Furthermore, we developed an interpretable machine learning (ML) model to predict the first-step equilibrium constant for 535 amines, achieving high accuracy (<i>R</i>² = 0.978, <i>Q</i>² = 0.834). This work integrates first-principles calculations with interpretable ML to guide the design of efficient COS capture solvents, advancing sustainable natural gas purification.