Developing amine chemical structure–activity relationships for carbon dioxide (CO2) absorption and desorption of amine solvents using machine learning models

Selecting amine solvents for industrial post combustion CO2 capture has never been easy. Usually, this process requires conducting a thorough experiment in order to determine the CO2 absorption and desorption activities of the amine solvents such as equilibrium CO2 rich loading, lean loading, initial CO2 absorption and desorption rates, heat of CO2 desorption, and cyclic capacity which depend so much on the structural, physical, and chemical properties of the amine solvents. This experimental process is not only stressful but also expensive, and time consuming. To address these problems, several machine learning models have been developed in this study to predict the CO2 capture activities of amine solvents such as equilibrium CO2 rich loading, lean loading, initial CO2 absorption and desorption rates, heat of CO2 desorption, and cyclic capacity based on their structural, physical, and chemical properties. This was achieved first, by conducting CO2 absorption and desorption experiment for forty (40) different amines strategically selected to cover the various classes of amines such as primary, secondary, tertiary, multi, and cyclic amines. The concentration of the amine solvents was fixed at 2 M/L, the volume of the solvent was also fixed at 100 ml, the absorption and desorption temperatures were fixed at 40 oC and 90 oC respectively, the concentration of CO2 in the feed gas was maintained at 4% while the feed gas flow rate was fixed at 200 ml/min throughout the experiment. The chemical properties, of the amines such as acid dissociation constant (pka) was determined experimentally, the steric hindrance, and electron donating and withdrawing potentials were determined using some suitable correlations developed by Muchan et al. (2024), while physical properties such as density and viscosity were determined by measurement using the anton paar density meter and viscometer respectively. In order to establish the structure – activity, and property – activity relationships, the physical, chemical, and structural properties of the amines were correlated with their CO2 capture activities. Based on the results of these findings, it could be concluded that pka has a strong positive effect on equilibrium CO2 rich loading of amine solvent, viscosity has a weak negative effect on equilibrium CO2 rich loading, the electron donating potential amines has a moderate positive effect on equilibrium CO2 rich loading, the electron withdrawing potential of amines has a strong negative effect on equilibrium CO2 rich loading, the steric hindrance of primary amines has a strong negative effect on the initial CO2 absorption rate, the pka of primary amines has a moderate positive effect on their initial CO2 absorption rate. This trend was found to be consistent with the secondary, tertiary, and cyclic amines apart from the multi amines which might be due to the variations in the chemical structures within this class of amines. However, since each multi amine has multiple pKa depending on the number of amino functional group present, it is suggested that taking a geometric mean of all the pKa for each multi amine might result in a significant correlation with equilibrium CO2 rich loading. At the end of this work, several machine learning models were developed for predicting the CO2 capture activities of amine solvents and the result shows that the equilibrium CO2 rich loading of amine solvents could be predicted using Catboost algorithm with AAD of 0.17% on the test data, the initial CO2 absorption rate of amine solvents could be predicted using Catboost algorithm with AAD of 0.15% on the test data, the CO2 lean loading of amine solvent could be predicted using Catboost algorithm with AAD of 0.28% on the test data, the initial CO2 desorption rate of amine solvents could be predicted using Catboost algorithm with AAD of 0.39% on the test data, the heat of CO2 desorption could be predicted using Catboost with AAD of 0.59% on the test data, while the cyclic capacity of amine solvents could be predicted using Catboost algorithm with AAD of 0.3% on the test data. To validate the equilibrium CO2 rich loading predictive model developed, some amines solvent of 2M concentration loaded at 4kPa CO2 partial pressure which were neither part of the train data nor test data were selected from the literature and fitted into the model. The results of this exercise were discussed in two cases: case1 with R2 value of 0.92 and AAD of 7% suggesting a good fit, and case2 with R2 value of 0.75 and AAD of 50% suggesting lack of good fit. The results from Case 1 show that the model accurately predicted CO₂ rich loading for new amines like HEEAB, 4A1B, DAMP, and TMDA with an AAD of 7%, likely due to their structural similarity to the training and testing data. In contrast, Case 2 revealed poor model performance (R² = 0.75, AAD = 50%) for structurally dissimilar amines, indicating limited generalizability.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.