Capability of Digital Twins for Representing a Complex Process—Prediction of Multicomponent Adsorption Breakthrough Curves and Times

This work tries to elucidate the reliability of artificial neural networks (ANN) to predict complex processes. This way, breakthrough curves and breakthrough times corresponding to 243 different scenarios of the multicomponent adsorption of H2, CO and CO2 in a fixed bed from a large set of runs (rather than a single run, which is the majority situation reported in the literature) generated through Aspen AdsorptionTM, were fitted to 600 ANNs configurations through a homemade software running in Fortran and 8 additional algorithms contained in the Scikit-Learn, a Python module for machine learning. To generate a consistent ANN, data obtained through Aspen AdsorptionTM were randomly divided into two groups: training (80% of the breakthrough curves and breakthrough times) and validation (20% of them). This procedure was able to properly predict single breakthrough curves. However, the capacity of the ANN for predicting a set of breakthrough curves was not so good as expected although the trends followed by the prediction curves could be used to make a good estimation of the dynamic behaviour of adsorption process. Finally, it was observed a good agreement between the values of the breakthrough times corresponding to the reduction of the H2 concentration in the outlet stream of 2% computed by Aspen AdsorptionTM and used for validation and those predicted by the best ANN model. The general procedure here followed could be equally used for analyzing real set of adsorption experiments or other different complex processes as described here.