Abstract:
The objective of this dissertation is to apply various data mining techniques for knowledge extraction from experimental and computational data to improve catalyst design and testing conditions for selective CO oxidation and water gas shift reactions. First, the experimental data produced in our laboratory were analyzed in three parts: selective CO oxidation over Pt catalysts by neural networks, selective CO oxidation over Au catalysts by decision trees and neural networks, and water gas shift reaction over Pt catalysts by neural networks. In all these works, the models successfully established the effects and relative significances of the catalyst variables in a reasonable agreement with the literature. Then, the data generated by density functional theory for selective CO oxidation over Au nanoparticles were modeled using neural networks and logistic regression, both of which demonstrated that the relations between structural properties of the Au nanoparticles and adsorption energies of CO and O2 on these particles can be successfully established. Finally, the published data by the other investigators were modeled to assess the possibility of predicting the outcome of an experiment, which was not done yet, from the knowledge accumulated in the literature. First, the publications on selective CO oxidation over Cu based catalysts were modeled by neural networks. Then, the publications on the same reaction for noble metal-gold based catalysts were modeled by genetic algorithm supported clustering together with neural networks. In both works, the experimental CO conversions reported in each publication were successfully predicted by the neural networks trained using the data from the remaining publications. The relative significances of the input variables and the major trends associated with these variables were also successfully estimated. It was then concluded that the knowledge accumulated in the published data through years can be extracted using some suitable data mining tools to help the researchers in planning their future experimental works more effectively.
