Machine learning model and optimization of a PSA unit for methane-nitrogen separation

In this work we study the separation of N2/CH4 in a bed packed with silicalite. Pressure swing adsorption (PSA) is a competitive technology for this task. Predicting PSA performance is a time consuming computational intensive problem. Direct optimization of the system of differential algebraic equations (DAE) describing the phenomena takes an impractical amount of time. We then analyze the suitability of using artificial neural networks (ANN) as a surrogate model to predict and optimize the PSA performance. Using the ANN surrogate model, optimization time decreased from 15.7 h to 50 s. We demonstrate that the PSA cycle proposed can achieve an optimized 99.5% nitrogen purity stream from an 85% inlet stream and a 50% purity stream from a 10% inlet stream. We also show that nitrogen recovery can be at most 90%. We further carry out a multi-objective optimization to demonstrate the tradeoff curve between nitrogen purity and recovery.