Experimentally validated machine learning frameworks for accelerated prediction of cyclic steady state and optimization of pressure swing adsorption processes

• Published in: Separation and purification technology 2020-06, Vol.241, p.116651, Article 116651
• Main Authors: Pai, Kasturi Nagesh, Prasad, Vinay, Rajendran, Arvind
• Format: Article
• Language: English
• Subjects: Machine learning; Multi-objective optimization; Neural networks; Post combustion carbon capture; Surrogate modelling and optimization; Vacuum swing adsorption
• ISSN: 1383-5866; 1873-3794
• DOI: 10.1016/j.seppur.2020.116651

•Machine-learning models for rapid pressure swing adsorption optimization.•Models for predicting process performance and cyclic-steady-state profiles.•Upto 25× reduction in computational loads for multi-objective optimization.•Models validated experimentally for CO2 capture from dry flue gas (15%CO2+85%). Machine learning-based surrogate models are presented to accelerate the optimization of pressure swing adsorption processes. Various supervised machine learning algorithms, such as decision trees, random forests, support vector machines, Gaussian process regression, and artificial neural networks, are tested for their ability to predict key performance indicators for a given set of operating conditions. Among the algorithms studied, Gaussian process regression-based surrogate models were found to be the best at predicting process outputs, with minimal training effort. The adjusted coefficient of determination for predictions using the surrogate model is greater than 0.98 using a sampled training set of 400 operating conditions. A surrogate model based on artificial neural networks is also presented to predict the bed profiles of the intensive variables at cyclic steady state. The surrogate models show very good agreement with the detailed model simulations. Experiments performed on a lab-scale two-column rig, for the concentration of CO2 from a mixture of CO2+N2 on Zeolite-13X, confirm performance indicators such as purity, recovery and axial profiles predicted by the surrogate models. Two new optimization frameworks are presented: Surrogate optimization in which the trained surrogate model is used to provide the process performance; and cyclic steady state optimization in which the predicted cyclic steady state profiles are provided as an initial condition for detailed model in order to accelerate convergence. Both techniques are shown to accurately predict Pareto fronts of Purity-Recovery and Energy-Productivity calculated from optimization that uses a detailed process model. The Surrogate optimization and the cyclic steady state accelerated Detailed optimization show ≈23× and 6× reduction in computational load, respectively, when compared to the traditional optimization using detailed models.