Predicting overall mass transfer coefficients of CO2 capture into monoethanolamine in spray columns with hybrid machine learning. (April 2023)

• Record Type: Journal Article
• Authors: Di Caprio, Ulderico; Wu, Min; Vermeire, Florence; Van Gerven, Tom; Hellinckx, Peter; Waldherr, Steffen; Kayahan, Emine; Leblebici, M. Enis
• Is Part Of: Journal of CO₂ utilization. Volume 70(2023)
• Publication Date: 2023-04
• Subjects: Hybrid modelling -- CO2 capture -- Machine learning -- Process intensification; Carbon dioxide -- Periodicals; Carbon dioxide -- Environmental aspects -- Periodicals; Carbon dioxide mitigation -- Periodicals; Carbon dioxide; Carbon dioxide -- Environmental aspects; Carbon dioxide mitigation; Periodicals; 628.53205
• DOI: 10.1016/j.jcou.2023.102452
• Languages: English
• ISSNs: 2212-9820

Abstract

Abstract: In order to avoid the catastrophic effects of global warming, we need to reduce CO2 emissions. Currently, the most mature technology to reduce large industrial CO2 emissions is the absorption of CO2 into aqueous monoethanolamine (MEA) solutions. The process is mostly studied in packed columns, for which many correlations have been offered to predict overall mass transfer coefficients ( K G a ). Spray columns are less prone to corrosion and were shown to enhance K G a . However, to the best of our knowledge, there are no models to predict K G a in spray columns. Hybrid modelling tools, a combination of machine learning techniques and first-principle information, showed remarkable capabilities in modelling complex systems. In this work, we applied hybrid modelling techniques benchmarking performances of four regressors: Ridge regression, decision tree regressor (DTr), support vector machine regressor (SVMr) and fully connected artificial neural network (ANN). We compared the performances of these modelling techniques with a model developed using the Buckingham Π-theorem, which is the most used state of the art technique to model K G a based on dimensionless numbers. SVMr and DTr showed higher accuracies among the trained models on the test set. SVMr can predict K G a within 6.4% error on the test set, whereas the Buckingham modelling approach resulted in 83 % error. The use of machine learning techniques resulted in predictive models with higher accuracies compared to the Buckingham Π-theorem. Predicting K G a with a higher accuracy allows more control over operational parameters and better column designs. Highlights: ML techniques are applied to predict KG a for the CO2 capture in spray columns. All the ML models, except the ANN, perform better than the Buckingham Π-theorem model. The employed ANN has similar performances of the Buckingham Π-theorem model. DTr and SVMr have higher accuracies than ANN and Ridge regression. SVMr has the best accuracy and generalisation performances in this study.