Joint incremental learning network for flexible modeling of carbon dioxide solubility in aqueous mixtures of amines

• Published in: Separation and purification technology 2024-02, Vol.330, p.125299, Article 125299
• Main Authors: Hsiao, Yu-Da, Chang, Chuei-Tin
• Format: Article
• Language: English
• Subjects: Aqueous amine mixtures; Artificial neural networks; CO2 solubility; Incremental learning; Vapor–liquid equilibrium
• ISSN: 1383-5866; 1873-3794
• DOI: 10.1016/j.seppur.2023.125299

•Innovative joint incremental learning method for flexible modeling of the CO2 solubility in aqueous amine mixtures.•Reusing existing mono-solvent models for modeling of multi-solvent systems.•31–68% error reductions were achieved if compared with conventional ANN counterparts. Vapor-liquid equilibrium (VLE) modeling is one of the most essential tasks for rigorous estimations of thermodynamic properties in chemical process design and analysis. To realize any commercially feasible carbon dioxide (CO2) capture scheme using a carefully chosen aqueous amine solution, the accurate model of CO2 solubility in this water-amine system plays a vital role. However, the experimental data of various mixtures of amine solvents are insufficient and expensive to acquire. In view of this problem, a novel joint incremental learning network (JILN) structure is proposed for modeling equilibrium solubility of CO2 in various bi-solvent aqueous amine mixtures. With the proposed method, the knowledge embedded in the well-trained mono-solvent models can be extracted and effectively transferred to their related multi-solvent models. By adopting the proposed modeling method, the numerical experimental results showed that the mean absolute percentage errors (MAPEs) for various bi-solvent models were 2.1–4.9%, which indicates a maximum of 68% reduction in prediction blunders if compared with their ANN counterparts.