DESC0025193

Chemical absorption of exhaust CO2 by amine solvents is a proven technology suitable for carbon capture and storage in industrial applications. The process involves the absorption of CO2 from flue gas by aminebased absorbents to form a carbamate complex, which is then regenerated to release concentrated CO2 during the stripping process. However, the continuously regenerated and recycled solvent undergoes thermal and oxidative degradation by reacting with O2, SO2 and NO2 impurities present in the flue gas. This leads to harmful emissions and increased corrosion rates in addition to loss of solvent. Hence, the solvent degradation becomes a hindrance to large scale deployment of the technology. Therefore, it is important to understand and predict the rate of solvent degradation and resulting emissions in advance in order to take steps to mitigate the harmful emissions and improve the CO2 capture process. Physics-based modeling of the CO2 capture is challenging due to the complex nature of the absorption and desorption process coupled with rate-controlled solvent degradation. Moreover, the chemical kinetics of solvent degradation is not well understood. Therefore, in this work, process modeling tool based on a Machine Learning approach will be developed to forecast solvent performance. The AI-based forecasting tools will be developed using Support Vector Regression (SVR) in conjunction with a deep learning approach such as Physics-Constrained Neural Network (PCNN) for the dynamic modeling of the CO2 capture process. The proposed dynamic model for real-time predictions harnesses the power of SVR and PCNN to forecast CO2 capture efficiency and emissions from solvent degradation in real-time. Commercial Applications and Other Benefits: The technology benefits the advancement of carbon capture, utilization, and storage. The modeling tool will be of interest to solvent technology developers, instrument manufacturers, and pilot plant operators as well as those monitoring and regulating carbon capture. The machine learning algorithms developed in this project will also benefit process modeling of other types of carbon capture applications.