A Benchmark Studyof Machine-Learning Regressors andSemiempirical Quantum Methods for CO2 Capture in Amine-BasedSolvents

Abstract The capture of CO2 using amine-based solvents is being continuously improved in an ongoing effort to mitigate anthropogenic CO2 emissions. In this study, the 488 CO2 capture reactions involving amine-based solvents were optimized, and their vibrational frequencies were calculated at the density functional theory (DFT) level by using the CAM-B3LYP/6–311++G(d,p) method. Amine-based solvents were selected, including primary, secondary, tertiary, cyclic, and aromatic types containing a variety of functional groups. Due to the computational cost associated with this level of theory, a benchmark study of 17 semiempirical methods was conducted to identify an approach that reproduces the trends predicted by the DFT calculations. The performance of these semiempirical methods was evaluated in terms of geometrical, electronic, and thermodynamic descriptors. Among them, the PM7 method showed the highest overall accuracy compared with DFT results. However, none of the tested methods provided satisfactory correlations for thermodynamic properties such as enthalpy and Gibbs free energy variations. Given the limited accuracy of the semiempirical methods in reproducing thermochemical parameters, 31 regression-based machine learning (ML) models were tested using both cross-validation and train/test strategies. Additionally, regression models were supplied by using features derived from electronic and structural descriptors. Five postprocessing techniques were employed to detect and remove outliers, with efficiency assessed as the ratio of data rejection to the improvement in correlation after treatment. Finally, the Extra Tree method, trained on descriptors calculated at the PM7 level and combined with the anomaly detection robust covariance, exhibited the best correlation and a lower mean absolute error (MAE), (1.26 kcal/mol) with the reference results obtained using DFT. The proposed workflow provides a robust approach to reproducing reference method results using a low-level method augmented by machine learning, significantly reducing the computational cost associated with modeling CO2 capture systems and thereby enabling a broader and more efficient exploration of potential sorbent materials.