Machine Learning-Based Multiobjective Optimization of Pressure Swing Adsorption

Abstract

The transient, cyclic nature and flexibility in process design make the optimization of pressure swing adsorption (PSA) computationally intensive. Two hybrid approaches incorporating machine learning methods into optimization routines are described. The first optimization approach uses artificial neural networks as surrogate models for function evaluations. The surrogates are constructed in the course of the initial optimization and utilized for function evaluations in subsequent optimization. In the second optimization approach, important design variables are identified to reduce the high-dimensional search space to a lower dimension based on partial least squares regression. The accuracy, robustness, and reliability of these approaches are demonstrated by considering a complex eight-step PSA process for precombustion CO2 capture as a case study. The machine learning-based optimization offers ∼10× reduction in computational efforts while achieving the same performance as that of the detailed models.