Surrogate model development for hydrogen separation via pressure swing adsorption processes: selection and evaluation of machine learning algorithms

Abstract

Within complex chemical engineering applications, subsystems or technologies have to be selected and evaluated, without being an expert in each technology. Here, a surrogate model can help, if it can be set up with easily available reliable tools and data from publications. A surrogate model for pressure swing adsorption processes to separate hydrogen from gaseous mixtures was developed here, using 90 published data sets for training and testing five different machine learning algorithms. The resulting specific surrogate model is valuable, as also the procedure of its development and analysis, which is transferable to other scientific questions.
For these data sets a random forest regression yielded the best results, in terms of high coefficient of variation, low mean absolute error and low root mean square error, when 80 % of the data was used for training and 20 % for testing. The predicted hydrogen recovery deviated from the true value by 6.6 %. A subsequent global sensitivity analysis revealed that the hydrogen recovery is mainly dependent on the number of adsorption beds and adsorption time. The purity depends on the adsorption pressure and the purge to feed ratio but should be investigated further by increasing the number of data sets, as soon as more publications become available. In the future, the surrogate model shall be implemented in a subordinate process concept model for testing the suitability of PSA processes for the separation of hydrogen from exhaust gas mixtures ordinating from fuel-rich operated HCCI engines.