Abstract:
The aim of this thesis is to apply the machine learning methodologies for knowledge extraction from experimental data for water gas shift reaction over noble metal catalysts from the papers published in the literature to understand the general trends and improve the catalyst design. First, the experimental data were extracted from 85 articles in the literature between 2002 and 2012 and a database containing 4372 data points with 87 variables was constructed. Then the methodology for knowledge extraction was selected; the neural networks and support vector machines, which are the two most effective machine learning tools, were compared using a dataset that was analyzed in our group before, and the neural network method showed better performance on prediction the unseen data. Hence the water gas shift reaction database was analyzed by using neural networks. An optimal network structure was determined by constructing various neural network topologies and comparing the testing RMSE, which is a measure of generalization ability of the model (prediction ability for the unseen data); two hidden layer network containing 21 nodes in each layer was found to be optimum to represent the database. Then, the experimental CO conversions in each article were predicted by the neural networks trained using the data from the remaining articles. The results in 29 articles out of 85 were predicted with a R2 value of higher than 0.5, which can be considered as successful. The effects and relative significances of the catalyst design and operation variables were also analyzed and found to be generally in agreement with the literature.
