Prediction of CO2 absorbing performance of amine aqueous solution using random forest models

Abstract

Designing an amine(s) solution for the chemical absorption of CO2 for carbon capture, utilization and storage is a significant challenge. To accelerate the development of high-performance amine solutions, a method for predicting the CO2 loading of amine(s) solution is studied. Available literature data are curated (45 amines, 3151 data points), and random forest regression is conducted using the data with descriptors calculated using HSPiP or RDkit software. The CO2 loadings of single amine aqueous solutions are regressed with high accuracy (R2 = 0.943, RMSE = 0.072–0.073 for the validation data). Feature importance analysis suggests that the partial charge of atoms in the amine is an important descriptor, as well as process parameters such as CO2 partial pressure and temperature. Based on the analysis, a simpler descriptor list is developed with temperature, CO2 partial pressure, and partial charges of N atoms calculated by density functional theory. CO2 loading of a single amine solution is regressed with an accuracy comparable to that of the above models (R2 = 0.931, RMSE = 0.079) with only 5 descriptors. Moreover, data including both single and blended amine solutions are regressed with high accuracy (R2 = 0.944, RMSE = 0.073) with only 8 descriptors. The generalization performances of the models are evaluated using the leave-one-group-out procedure. Although some outliers exist, the overall trend of almost all amines can be predicted using the procedure in this study.