Process Synthesis and Machine Learning for CO2 / H2 Membrane Separation

We propose a system design procedure for membrane gas separation based on a superstructure-based optimization strategy. We define a mathematical programming model of the system representing the superstructure. Instead of using the “classical” discretization of differential equations, we represent the membrane functioning using a surrogate model built by training a neural network. The resulting optimization problem is a non-convex non-linear programming problem, which is solved by a global optimization technique. We consider hydrogen purification from a pre-combustion gas mixture as a case study.

We model two different commercially available membranes (UBE BH and Polaris from MTR) building a machine model for each one of them based on a data set obtained using simulation. Training an accurate model for each membrane type can require a large set of data points and hence a high computational time. To overcome this issue, we propose a data augmentation technique that starts from a small-size data set and improves it sequentially by adding points. Instead of using standard augmentation techniques, we devised a strategy that is based on the optimization itself which ensures an accurate model in the regions of interest for the case study.

The results obtained for H2 purity of 90 and 99%, with different recovery requirements are detailed and discussed. A clear advantage emerges in using the surrogate model since it makes the optimization process numerically more stable, and much faster. All the solutions produced are validated and refined by a simulation approach.