Expandable neural networks for efficient modeling of various amine scrubbing configurations for CO2 capture

Abstract

Modeling of improved amine scrubbers using artificial neural networks (ANNs) were carried out in this study. Instead of training models from scratch with the case-by-case method, the expandable neural networks were utilized to progressively increase the number of model parameters and change the model input/output structures in a step-wise fashion. Efficient and rapid transformation of any existing model to a new one can be made realizable. This proposed strategy has been successfully validated in several process modification scenarios. From the experimental results, the required sampling sizes to achieve the similar prediction accuracy of the corresponding baseline model were considerably smaller, and, furthermore, over 47% of total data acquisition time can be saved. Finally, the corresponding sensitivity analyses showed that the proposed models were physically interpretable and able to extract the correct process mechanisms in the sense that the gain scales and signs were consistent with those of their rigorous counterparts.