Modeling CO Loading Capacity of Diethanolamine (DEA) Aqueous Solutions Using Advanced Deep Learning and Machine Learning Algorithms: Application to Carbon Capture

Several carbon capture techniques have been developed in response to the notable rise of atmospheric carbon dioxide (CO) levels. The utilization of diethanolamine (DEA) as an absorption method is prevalent in various industries due to its high reactivity and cost-efficiency. Hence, comprehending the equilibrium solubility of CO in DEA solutions is an essential step in developing and optimizing absorption procedures. In order to predict the CO loading capacity in the DEA solutions, four advanced deep learning and machine learning models were developed: recurrent neural networks (RNN), deep neural networks (DNN), random forest (RF), and adaBoost-support vector regression (AdaBoost-SVR). The models predict the capacity of CO loading as a function of temperature, CO partial pressure, and the concentration of DEA in the solution. Intelligent models were developed employing an extensive database which includes new experimental data points published within recent years, which were not considered in the previous studies. The RNN model was found to outperform other models based on graphical and statistical assessments, as evidenced by its lower root mean square error (RMSE=0.285) and standard deviation (SD=0.032), and higher determination coefficient (R=0.992). While the RNN model resulted in the highest accuracy in predicting CO absorption, the DNN, RF, and AdaBoost-SVR models also demonstrated satisfactory accuracy in predicting CO solubility, placed in the following ranking. A sensitivity analysis was performed on the four developed models, revealing that the CO partial pressure has the strongest effect on the CO loading capacity. Furthermore, a trend analysis was performed on the RNN model, demonstrating that the developed model has a high degree of accuracy in following physical trends. The binary interaction analysis was conducted with two varying parameters and one constant parameter in the RNN model through 3-D image plots, which illustrated the simultaneous effect of two independent parameters on CO loading. Finally, outlier detection was conducted by employing the Leverage method to find outlier data points in the data bank, demonstrating the applicability domain of intelligent models.