Reduced Order Machine Learning Models for Accurate Prediction of CO₂ Capture in Physical Solvents

Abstract

CO 2 sorption in physical solvents is one of the promising approaches for carbon capture from highly concentrated CO 2 streams at high pressures. Identifying an efficient solvent and evaluating its solubility data at different operating conditions are highly essential for effective capture, which generally involves expensive and time-consuming experimental procedures. This work presents a machine learning based ultrafast alternative for accurate prediction of CO 2 solubility in physical solvents using their physical, thermodynamic, and structural properties data. First, a database is established with which several linear, nonlinear, and ensemble models were trained through a systematic cross-validation and grid search method and found that kernel ridge regression (KRR) is the optimum model. Second, the descriptors are ranked based on their complete decomposition contributions derived using principal component analysis. Further, optimum key descriptors (KDs) are evaluated through an iterative sequential addition method with the objective of maximizing the prediction accuracy of the reduced order KRR (r-KRR) model. Finally, the study resulted in the r-KRR model with nine KDs exhibiting the highest prediction accuracy with a minimum root-mean-square error (0.0023), mean absolute error (0.0016), and maximum R 2 (0.999). Also, the validity of the database created and ML models developed is ensured through detailed statistical analysis.