Abstract
In this work, different types of neural networks and modeling methodologies are used and compared: feedforward and recurrent networks, stack neural networks and a hybrid model composed from a simplified phenomenological model and a neural network. For each situation, the performance of the networks was evaluated through mean squared error and correlation between training data and neural network predictions. Accurate results were obtained with different types of neural models, but our approach recommends feedforward neural networks which are simple to train and use. The well known free radical polymerization of methyl methacrylate, accompanied of gel and glass effects and achieved in a batch bulk process is considered as example. In a hybrid methodology, the kinetic model used until the onset of the gel effect is associated with a neural model which replaces the diffusional effects representing the difficult part to model in the process. Keywords: neural networks, free radical polymerization modeling, polymethyl methacrylate, hybrid model