The successful design of an observer for inferring the outlet composition from a chemical reactor heavily relies on the goodness of the adopted kinetic rate model (Baratti et al., 1993). On the other hand, often, it is difficult to dispose of a simple, but, exhaustive kinetic model because of the complexity of the reaction scheme one has to deal with. In this work, we explore the possibility to represent global (lumped) reaction rate laws by the use of neural network models. The aim is to develop a nonlinear observer (extended Kalman filter, EKF) of an heterogeneous gas-solid reactor that relies on a gray model where the "neural reaction rate" law is integrated within a first principles model. The procedure is outlined for the case of the catalytic oxidation of carbon monoxide over Pt-alumina catalyst. The results show that neural networks (NN) can be effectively used in representing lumped reaction rates since NN are able to capture the essential characteristics of the functional relationship relating the state variables.