Partial differentiation of neural network for the analysis of factors controlling catalytic activity

In order to examine the possibility for identifying the factors controlling catalytic activity by neural network, the numerical partial differentiation of trained neural network was applied to several examples of experimentally established correlations of catalytic activities with primary factors: oxidation of propene on oxide catalysts, oxidation of butane on lanthanide oxides, decomposition of formic acid on metal catalysts, oxidation of methane on lanthanide oxides, and support and additive effects on low temperature combustion of propane over Pt catalyst. The relative importance of the given factors including dummy parameters were estimated from the numerical differentiation of trained artificial neural network, and they were compared with those obtained by previously proposed methods using the weightings of connecting links of trained neural network. In all the examples, the primary factors that had been proposed in experimental studies were successfully identified by the numerical differentiation of trained neural network. As for the connecting weight-methods examined for the comparison, only the method proposed by Olden et al. and us gave satisfactory results to identify the primary factors. Further, it was demonstrated that the partial differentiation method could be used to obtain local information, that is, the partial derivatives for individual catalyst, which would enable us to know the method how each catalyst can be improved.

It was attempted to identify factors controlling catalytic activity by applying the numerical partial differentiation of trained neural network to several examples of experimentally established correlations of catalytic activities with properties of catalyst components. In all the examples, the primary factors that had been proposed in experimental studies were successfully identified by the method. Figure optionsDownload as PowerPoint slide