A comparison of deep learning models applied to Water Gas Shift catalysts for hydrogen purification

Abstract

As a consequence of the renewed interest in the Water Gas Shift reaction a great volume of information was produced. Since a traditional method like the reaction kinetics or mechanism are not capable of dealing with all this information, a deep learning model is convenient to explore to make useful predictions of catalysts performance. In the present work some novel features were included, a measure of reducibility, the crystal size, and the catalysts cost. The Principal Component Analysis indicated that the chosen features of the dataset were not redundant and the suggested novel features strongly influenced the most important components. A Random Forest Regressor was optimized and then trained in order to obtain the feature importance. An Artificial Neural Network was employed after a Grid Search optimization. This model was fed with different sizes of datasets in order to determine its effect on the accuracy of the predictions.