Performance limits of neural networks for optimizing an adsorption process for hydrogen purification and CO2 capture

Abstract

In this work, we use surrogate models to accelerate the optimization of an adsorption process for H2 purification and CO2 capture within the context of fossil-based low-carbon H2 production. A one dimensional column model was used to generate a training set for different feed compositions with up to four impurities and for varying process conditions. Subsequently, an artificial neural network (ANN) surrogate model was trained for six key performance indicators, achieving adjusted R2 values over 0.999 for all indicators. Finally, the ANN was used for the constrained optimization of the H2 separation performance and of the process performance (energy consumption and productivity), and the results were compared to full model optimization results. The agreement is very good for the H2 separation performance, and in the case of low H2 purity targets (99%) also for the process performance. For higher H2 purities, however, the energy-productivity Pareto front features a very high sensitivity toward the H2 purity constraint, and even small deviations in H2 purity between full model and ANN surrogate model can translate to big deviations in the energy-productivity Pareto front. This is particularly pronounced for poorly sampled input regions at the edge of the sampling domain, for example for a binary H2-CO2 feed. If the ANN surrogate model is used at such edges of the sampling domain, additional sampling is required in this region to increase its accuracy. For all other cases, the deviations between the full model and the ANN surrogate model regarding the minimum energy consumption (or the maximum productivity) were well below 5%, showing that the ANN can be used with good accuracy instead of the full model.