We fitted a new method allowing to use some of the knowledge involved in chemico-physical models with neural networks. This method, we call model-based neural networks, needs the disposal of experimental measures, a fully determined model (even approximate), and access to the state variables of the system. Owing this, we are able to fundamentally include mathematical models, such as physico-chemical ones, in the learning phase of the network, in order to improve its performances, although relying on experimental data. The method was successfully tested by using theoretical examples. It occurs to be especially useful when experimental data are badly determined.