Often, in real applications it is difficult to dispose of a simple, yet, representative kinetic model because of the complexity of the reactions taking place. To overcome this limitation a hybrid modelling approach is proposed for the identification of the dynamic behaviour of chemical reactors. In particular, the tools of neural network modelling have been exploited to represent the kinetic reaction data. The "neural reaction rate model" is integrated within a first principles model that constitutes the basis of a nonlinear observer extended Kalman filter (EKF) for an heterogeneous gas-solid reactor where the catalytic oxidation of carbon monoxide takes place. The outlined procedure shows that artificial neural networks (ANN) can be effectively used to formulate lumped reaction rates because of their capability in capturing the essential characteristics of the functional relationship among the state variables.