Data-driven modeling of CO2 capture in rotating packed beds enhanced by carbonic anhydrase using explainable artificial intelligence methods
Theofilos Xenitopoulos, Αθανάσιος Ι. Παπαδόπουλος, Panos Seferlis
Abstract
• LightGBM accurately models enzymatic CO 2 absorption in rotating packed beds. • AI model predictions align with both experimental and kinetic model results. • SHAP explains key features influencing LightGBM model predictions. • This study pioneers XAI in Carbon Capture for model tuning and system insights. Reactive absorption in aqueous solutions is a widely applied CO 2 capture technology, but its efficiency can be significantly enhanced through process intensification. Rotating Packed Bed (RPB) technology offers a promising solution by intensifying mass transfer and enabling substantial equipment size reduction compared to packed columns. Incorporating biocatalysts, such as the enzyme carbonic anhydrase (CA), further boosts the efficiency of the CO 2 absorption process. This study employs a data-driven approach to model enzyme-enhanced CO 2 absorption in an RPB system using LightGBM, a gradient boosting framework that builds decision trees in a sequential manner, utilizing histogram-based learning and leaf-wise tree growth for enhanced accuracy and efficiency. The model is trained and validated based on experimental data collected from CO 2 absorption experiments with a 30 wt% N-methyldiethanolamine (MDEA) solution, with and without CA across various gas and liquid flow rates. The LightGBM model achieved a high mean cross-validation R 2 score (0.98) and low root mean squared error value (0.3) in predicting absorption efficiency, indicating high predictive accuracy. Shapely Additive Explanations (SHAP) are employed to analyze feature importance and understand the key parameters influencing absorption efficiency. The validated model is then used for operational analysis, offering insights into system performance optimization. Results reveal that enzyme-enhanced absorption can improve CO 2 absorption efficiency by up to 245.87% compared to the solvent without the enzyme, underscoring the potential of combining high-gravity technology with biocatalysts and machine learning techniques for next generation carbon capture systems.