Using recurrent multilayer neural network for simulating batch reactors

The authors investigate the potential of multilayer artificial neural networks for simulating the dynamic behavior of batch reactors, i.e., chemical reactors or bioreactors. The databases used for training the neural networks consist of several types of data: initial conditions, command parameters, and observations on the process state at different steps during operation. Several architectures of multilayer neural networks have been studied but the emphasis of this investigation has been placed on recurrent multilayer neural networks. After training such a network, one is able to predict the dynamic behavior of the batch reactor when new initial conditions or new command parameters are given. Two applications are discussed: the first concerns data derived by simulating two successive chemical reactions; the second involves experimental data on alcohol fermentation