(326a) Developing Data-Driven Dynamic Models for Predicting Emissions from an Amine-Based CO2 Capture Process

Carbon capture, utilization, and storage (CCUS) technologies are being significantly evaluated for reducing greenhouse gas (GHG) emissions [1]. Post-combustion CO₂ capture using aqueous amine-based solvents—such as monoethanolamine (MEA) and CESAR1 (27 wt% 2-amino-2-methyl-1-propanol (AMP) and 13 wt% piperazine (PZ) [2])—is one of the most widely studied and commercially viable CCUS technologies. In this process, CO₂ is chemically absorbed from flue gas into the amine solvent, forming a carbamate intermediate. The solvent is then thermally regenerated at elevated temperatures to release high-purity CO₂ for compression and storage or utilization. However, continuous cycling of the solvent leads to its solvent degradation, primarily through three mechanisms: thermal degradation, carbamate polymerization, and oxidative degradation caused by reactions with flue gas impurities such as O₂, SO₂, and NO₂ [3]. These degradation pathways result in the formation of various byproducts, including ammonia, aldehydes, and carboxylic acids (e.g., formic, acetic, and nitric acids). Such byproducts not only contribute to harmful emissions but also lead to operational challenges like corrosion, foaming, increased solvent viscosity, and equipment fouling [4]. To enable cost-effective and environmentally benign deployment of amine-based CO₂ capture at scale, it is critical to develop predictive models capable of estimating solvent degradation and associated emissions. These models should account for both current and historical operating conditions, enabling the design of proactive mitigation and operational strategies that support long-term performance and environmental compliance. However, degradation mechanisms, thermodynamics and kinetics of degradation products under absorber and stripper conditions are highly complex and currently not well understood. This motivates development of data-driven models.

In this study, a novel hybrid modeling framework is developed by integrating series and parallel all-nonlinear dynamic-static artificial neural network (ANN) architectures [5]. Specifically, the hybrid model combines a nonlinear static feedforward neural network with a nonlinear dynamic feedback network, arranged in both series and parallel configurations. This hybridization enables synthesizing comparably smaller network sizes compared to deep neural networks even for complex transient data. To address challenges related to data scarcity and model interpretability, a physics-constrained neural network (PCNN) model is developed. PCNNs can constrain the model to satisfy fundamental physical laws including chemistry, thermodynamics, reaction kinetics. A sequential parameter estimation algorithm is employed, allowing the static and dynamic components to be trained independently. This modular training approach significantly improves computational efficiency, achieving convergence rates 50–100 times faster than if the entire network is trained by using an off-the-shelf optimizer for solving the constrained optimization problem. To develop and validate the model, a dataset for the CESAR1 solvent obtained from the Technology Center Mongstad, the world’s largest pilot plant, is used. Preprocessing steps include the application of a Savitzky-Golay filter for noise reduction and data smoothing, followed by principal component analysis (PCA) to identify and retain dominant input features. The processed dataset is segmented into multiple multi-hour time windows for training to ensure that the system dynamics and impact of the initial state on system dynamics are represented in the data. K-fold cross-validation is then applied to minimize bias in data selection. Results show that the proposed hybrid PCNN model outperform several existing state-of-the-art deep neural network models such as LSTM and GRU, demonstrating superior prediction accuracy and reliability for predicting amine degradation.

References:

1. US DOE. DOE Industrial Decarbonization Roadmap. United States Department of Energy, 2022. https://www.energy.gov/industrial-technologies/doe-industrial-decarbonization-roadmap. [Accessed on April 5, 2025]
2. Hume, S.A.; McMaster, B.; Drageset, A.; Shah, M.I.; Kleppe, E.R.; Campbell, M. Results from Cesar1 Testing at the CO₂ Technology Centre Mongstad. Verification of Residual Fluid Catalytic Cracker (RFCC) Baseline Results. 16^th International Conference on Greenhouse Gas Control Technologies, GHGT-16, 2022, Lyon, France.
3. Vega, F.; Sanna, A.; Navarrete, B.; Maroto-Valer, M.M.; Cortes, V.J. Degradation of amine-based solvents in CO₂ capture process by chemical absorption. Gas Sci. Tech., 2014, 4(6), 707-733.
4. Veltman, K.; Singh, B.; Hertwich, E,G. Human and Environmental Impact Assessment of Postcombustion CO₂ Capture Focusing on Emissions from Amine-based Scrubbing Solvents to Air. Sci. Tech., 2010, 44(4), 1496-1502.
5. Mukherjee, A.; Bhattacharyya, D. Hybrid Series/Parallel All-Nonlinear Dynamic-Static Neural Networks: Development, Training, and Application to Chemical Processes. Eng. Chem. Res., 2023, 62(7), 3221-3237.