Progressive learning for surrogate modeling of amine scrubbing CO₂ capture processes

Amine scrubbing process is a promising approach for post-combustion CO₂ capture. To analyze and improve the process performance, various rigorous mathematical models have already been adopted for simulation purpose. If these high-fidelity models are utilized for more advanced applications, the required computational demand can be overwhelming. This drawback inevitably motivates the use of data-based surrogate models to relieve computation effort. However, due to the inherent complexities of any given amine scrubbing process, the task of taking enough simulation data for building an accurate model is still computationally expensive. Therefore, in this paper, an innovative modeling procedure using artificial neural networks is proposed to effectively alleviate the aforementioned data acquisition effort. To improve the data sampling efficiency while maintaining model accuracy, several concepts used in process synthesis and progressive learning have been adapted in this work. The proposed surrogate model was constructed for the specific purposes of predicting the CO₂ emission rate, the reboiler duty and the compression duty. In applying the proposed procedure, a large number of samples collected from repeated standalone simulation runs of absorber process were used to pre-train the corresponding surrogate model, and this model was then further fine-tuned and expanded according to samples collected from relatively few plantwide simulation runs. By using this modeling strategy, both plantwide sampling size and total data collection time can be effectively reduced. More specifically, a comparison between the conventional method and the current model-building strategy showed that over 23–64% of data acquisition time can be saved with the latter approach.