Development of a Co_MgO Catalyst for High-pressure Dry Reforming of Methane Based on Design of Experiment, Artificial Neural Network and Grid Search

Combinatorial catalysis is now commonly used as a promising tool for catalyst research and development. Various tools were demonstrated in special issues of Catalysis Today1) and Applied Catalysis A2) in 2003. Combinatorial and rapid library synthesis with highthroughput screening (HTS) is the main technology accelerating every step of the R&D of catalysts. Considering the complexity of advanced materials, however, the two tools are not sufficient. If 76 nonradioactive elements can be applied to the catalysts, there are 2,850 binary but 18,474,840 quinary combinations of these elements3). Such a number of combinations is clearly too much for even the most advanced HTS system at present. Therefore, data mining tools such as the artificial neural network (ANN) and genetic algorithm (GA) are used to compensate. We reported optimization of the Cu_Zn oxide catalyst for methanol synthesis using HTS with 96 parallel reaction lines4), ANN to correlate catalyst parameters with activities, and GA5)~8) or a grid search (formerly reported as all-encompassing calculation9)~11)) to find the global maximum on the artificial neural network. Random parameters for the catalyst were used in these previous experiments to obtain training data for the ANN because such dispersed data were thought to give better networks. Rather large numbers of datasets are necessary in this method and additional data are sometimes required for re-training of the ANN for more precise prediction. A better and well-designed dataset for effective training of ANN is necessary. Statistical design of the experiment was applied for a few catalysts to identify the important factors among catalyst composition, preparation parameters and reaction conditions12)~15). The quantitative influence of the parameters was estimated using a polynomial model and a regression technique14). The apparent complexity of the statistical method can be simplified by the Taguchi method13),15),16). The orthogonal array can reduce the number of necessary experiments, while maximizing the amount of information derived from the reduced experiment set. Such arranged datasets seem suitable for ANN training17)~19). The resulting ANN gives good predictions for the output parameter of the ANN such as activity and selectivity. The global optimum on the ANN was found rapidly with the assistance of a grid search18)~20). More precise predictions are of course possible by ANNs trained by a large