Abstract:
This study is aimed to model design and reaction parameters of CO oxidation over CuOx-CoOx-CeO2 catalyst using neural and modular neural networks. Since there are many possible architectures for modular networks, the best network is searched using the measures of correlations such as R2, R2 adj, and RMSE. In all models, first a small network is constructed and enlarged without over fitting the data. Then the best networks were optimized with Quasi – Newton method using the MATLAB® Optimization Toolbox. The significance of input parameters and their effects on CO conversion were also analyzed. The data was also modeled using the multiple regression method for comparison. The effects of precipitation pH and temperature were modeled with a neural network of two neurons in the hidden layer with an R2 value of 0.970. The precipitation and reaction conditions were also modeled together with a neural network of 3-1 structure, and modular neural networks of 2-3-1 and 1-1-1structures with the R2 values of 0.981, 0.989 and 0.971 respectively indicating that both models can be used. However, it was found that the modular neural networks have some advantages; they feed the similar input parameters into the same module which enhances the modeling power, and decreases the possibility of over-fitting by lowering the number of connections for the same number of data points. Similar models were developed for various combinations of catalyst preparation conditions (temperature and pH), target metal loadings (Cu, Ce and Co weight per cent) and reaction parameters (temperature, W/F and time on stream) with considerable success. It was also observed that the effects of measured catalyst properties (actual metal loadings and total surface area) on CO conversion can also be modeled successfully using both neural and modular neural networks.
