For a complex reaction such as glucose production from maltose by hydrolysis in subcritical water, neural network analysis was applied to determine the optimum combination of the reaction conditions. As input parameters for neural network analysis, the preset temperature of the reactor, the residence time, the initial concentration of maltose, and the pressure of the subcritical water were selected. As the output parameter, the reactive index
I, the product of the yield and selectivity of glucose, was selected. In the first demonstration, the optimum reactive condition of residence time and temperature was determined at constant concentration and pressure. The neural network model was built from the various experimental data, and the reactive indices
IP at unknown conditions were predicted from the model. The neural network model was corrected by the addition of experimental data around the lowest
IP. The addition of data and the rebuild of the model were repeated. The final model was completed when the reactive conditions giving the lowest
IP and the experimental
IE were identical. To determine the optimum reactive condition, 1 or 3 rebuilds of the neural network model were required. The errors between the predicted and experimental conditions were 2°C in the reactor temperature and 1.89 min in the residence time. Also, in the expansion to a 4-dimensional reactive condition, that is, temperature, residence time, concentration and pressure, the optimum reactive condition was well presented by neural network analysis.
View full abstract